PhD Thesis Khairy Elsayed
Click here to load reader
-
Upload
carlos-antonio-ribeiro -
Category
Documents
-
view
182 -
download
39
Transcript of PhD Thesis Khairy Elsayed
FACULTY OF ENGINEERING
Department of Mechanical Engineering
Analysis and Optimization of
Cyclone Separators Geometry
Using RANS and LES
Methodologies
Thesis submitted in fulfillment of the requirements for the
award of the degree of Doctor in de Ingenieurswetenschappen
(Doctor in Engineering) by
Khairy Elsayed
Brussels, October 2011
Advisor: Prof. Dr. Ir. Chris Lacor
Analysis and Optimization of Cyclone Separators Geometry Using RANS and LES Methodologies
by
Khairy Elsayed
Submitted to the Department of Mechanical Engineering, in partial fulfillment of the requirements
for the degree ofDoctor in Engineering
Vrije Universiteit BrusselOctober 2011
Advisor: Prof. Dr. Ir. Chris Lacor
Analysis and Optimization of Cyclone Separators Geometry Using RANS and LES Methodologies
Khairy Elsayed
Department of Mechanical Engineering, Vrije Universiteit BrusselPleinlaan 2, B-1050 Brussels, Belgium
Thesis submitted in partial fulfillment of the requirements for the academic degree of Doctor in Engineering
Promoter:Prof. dr. ir. Chris LacorJury: Prof. dr. ir. Johan Deconinck, voorzitterProf. dr. ir. Rik Pintelon, vice-voorzitterProf. dr. ir. Gunther Steenackers, secretarisProf. dr. ir. Gert DesmetProf. dr. ir. Harry van den Akker (Delft University of Technology, Netherlands)Prof. dr. ir. Herman Deconinck (Von Karman Institute, Belgium)
© 2011 Khairy Elsayed
2011 Uitgeverij University PressLeegstraat 15 B-9060 Zelzate Tel +32 9 342 72 25E-mail: [email protected] www.universitypress.be
Vrije Universiteit Brussel – Faculteit Ingenieurswetenschappen Pleinlaan 2 – 1050 Brussel Contact: +32 (0)2 629 39 10 http://www.vub.ac.be/IR – [email protected]
ISBN 978-94-9069-594-1
All rights reserved. No parts of this book may be reproduced or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, without the prior written permission of the author.
Abstract
The gas-solids cyclone separator is an industrial equipment that has been
widely used for more than a century. Due to its industrial relevance, a
large number of experimental, theoretical and computational studies have
been reported in the literature aimed at understanding and predicting the
performance of cyclones in terms of pressure loss and collection efficiency
(cut-off diameter). The currently used mathematical models for the predic-
tion of cyclone performance, however, exhibit limited accuracy and gener-
ality. Moreover, the cyclone performance can be calculated using the artifi-
cial neural networks approach. An alternative approach is to simulate the
gas-particle flow field in a cyclone by computational fluid dynamics (CFD).
As a result of the recent progress of computational power and numerical
techniques, CFD has been widely applied to industrial flow problems. The
cyclone performance parameters are governed by many operational pa-
rameters (e.g., the gas flow rate and temperature) and geometrical param-
eters. This study focuses only on the effect of the geometrical parameters
on the flow field pattern and performance of the tangential inlet cyclone
separators using three different approaches, the most robust mathemat-
ical models, artificial neural networks and CFD approaches. The study
was limited to reverse-flow gas-solids cyclone separators operating at low
solids loading.
The objective of this study is four-fold. First, to determine the most sig-
nificant factors affecting the cyclone performance based on the previous
studies and statistical analysis of data using response surface methodol-
ogy. Second, to study each (significant) parameter separately to obtain
more details about its effect on the flow field pattern and the performance.
Third, to obtain the most efficient cyclone design for minimum pressure
drop (using the most robust mathematical models, artificial neural net-
works and CFD dataset). Finally, to obtain the most efficient cyclone de-
sign for best performance (minimum pressure drop and minimum cut-off
diameter) using multi-objective optimization techniques with two different
optimization techniques (both the Nelder-Mead with desirability function
and the genetic algorithms (NSGA-II)).
The response surface methodology has been performed using dataset ob-
tained from the Muschelknautz method of modeling (MM) to determine
the most significant parameters. Four geometrical factors have significant
effects on the cyclone performance viz., the vortex finder diameter, the in-
i
let width, the inlet height and the cyclone total height. There are strong
interactions between the effect of inlet dimensions and the vortex finder
diameter on the cyclone performance. The same investigation has been re-
peated using artificial neural network approach based on the experimental
pressure drop. A radial basis neural network (RBFNN) is developed and
employed to model the pressure drop for cyclone separators. The neural
network has been trained and tested by experimental data available in
literature. The result demonstrates that artificial neural networks can of-
fer an alternative and powerful approach to model the cyclone pressure
drop. The analysis indicates the significant effect of the vortex finder di-
ameter and the vortex finder length, the inlet width and the total height.
Furthermore, Four mathematical models (Muschelknautz method “MM”,
Stairmand, Ramachandran and Shepherd and Lapple) have been tested
against the experimental values. The residual error (the difference be-
tween the experimental value and the model value) of the MM model is
the lowest.
The numerical simulations of cyclone flow were carried out by solving
the unsteady-state, three-dimensional Reynolds averaged Navier-Stokes
(RANS) equations combined with a closure model for the turbulent stresses
and the large eddy simulation approach. The modeling of the cyclonic flow
by computational fluid dynamics (CFD) simulation has been reported be-
fore in the literature. Using the experimental data available in literature,
a generic assessment was carried out on a number of turbulence closure
models. Only the Reynolds stress turbulence model (RSM) and large eddy
simulation (LES) captured the cyclone flow field best compared to the ex-
perimental measurements.
The CFD model also predicted the collection efficiency, where the partic-
ulate phase was treated in a Lagrangian framework by tracking a large
number of particles of different size classes through the computational do-
main. The stochastic nature of the particle motion due to the fluid turbu-
lence was taken into account by a particle dispersion model. There was a
reasonable agreement between the calculated and measured cut-off diam-
eter for both the RSM and LES results.
The effect of the cone tip-diameter on the flow field and performance of
cyclone separators was investigated because of the discrepancies and un-
certainties in the literature about its influence. Three cyclones with dif-
ferent cone tip diameters were studied using large eddy simulation (LES).
The flow field pattern has been simulated and analyzed with the aid of ve-
locity components and static pressure contour plots. The obtained results
demonstrate that the cone tip-diameter has an insignificant effect on the
collection efficiency (the cut-off diameter) and the pressure drop. The sim-
ii
ulation results agree well with the published experimental results and the
mathematical models’ trend.
The effect of the cyclone inlet dimensions on the performance and the flow
field pattern has been investigated computationally using the Reynolds
stress turbulence model (RSM) for five cyclone separators. The maximum
tangential velocity in the cyclone decreases with increasing the cyclone in-
let dimensions. Increasing the cyclone inlet dimensions decreases the pres-
sure drop. The cyclone cut-off diameter increases with increasing cyclone
inlet dimension. Consequently, the cyclone overall efficiency decreases due
to weakness of the vortex strength. The effect of changing the inlet width
b is more significant than the inlet height a, especially for the cut-off di-
ameter. The optimum ratio of inlet width to inlet height b/a is from 0.5 to
0.7.
The effect of the vortex finder dimensions (both the diameter and length)
on the performance and flow field pattern has been investigated computa-
tionally using the large eddy simulation (LES) for nine cyclone separators.
The maximum tangential velocity in the cyclone decreases with increasing
the vortex finder diameter. Whereas, a negligible change is noticed with
increasing the vortex finder length. Increasing the vortex finder length
makes a small change in both the static pressure, axial and tangential ve-
locity profiles. However, decreasing the vortex finder diameter gradually
changes the axial velocity profile from the inverted W to the inverted V pro-
file. Decreasing the cyclone vortex finder diameter increases the maximum
tangential velocity. The maximum tangential velocity approaches asymp-
totically 1.589 times the inlet velocity when decreasing the vortex finder
diameter. The Euler number (dimensionless pressure drop) decreases with
increasing the vortex finder diameter. Increasing the vortex finder length
slightly increases the Euler number. The Stokes number increases with
increasing the vortex finder diameter and slightly increases as the vortex
finder length is increased.
The effect of the cyclone height (both the barrel and cone) on the perfor-
mance and flow field pattern has been investigated computationally for six
cyclone separators. The maximum tangential velocity in the cyclone de-
creases with increasing the cyclone (barrel or cone) height. Increasing the
barrel height, makes a small change in the axial velocity, whereas increas-
ing the cone height changes it considerably. Increasing the cyclone (barrel
or cone) height decreases both the pressure drop and the cut-off diameter.
The changes in the performance beyond h/D = 1.8 are small at constant
cone height, whereas the performance improvement stops after hc/D = 4.0(Ht/D = 5.5) at constant barrel height where h is the barrel height, hc is
the cone height, Ht is the total cyclone height and D is the barrel diameter.
iii
The effect of changing the cone height on the flow pattern and performance
is more significant than that of the barrel height.
The CFD model was used to predict the pressure drop and the collection
efficiency of a range of cyclone geometries based on Stairmand’s high-
efficiency design. These predictions were used to obtain an algebraic equa-
tion that relates the performance of a cyclone to its design and a limited
set of dimensionless quantities (Euler number and Stokes number). This
approach towards predicting cyclone performance by varying many geo-
metrical parameters has not been reported before.
To obtain new optimized cyclone separators, several optimization studies
have been conducted in this thesis. Both the response surface methodology
(RSM) and the radial basis function neural network (RBFNN) have been
used as meta-models. Three different sources of data have been used to fit
the second order polynomial in case of RSM and for training the RBFNN.
These data come from analytical models, experimental measurements and
CFD simulations. Two optimization techniques have been used to optimize
the cyclone geometry for minimum pressure drop, namely, the Nelder-
Mead and the genetic algorithms techniques. To handle the bi-objective
optimization problem (both the pressure drop and the cut-off diameter),
two approaches have been applied, the desirability function and NSGA-II
techniques.
All the new optimized cyclones obtained either for single objective or for
bi-objective problems exhibit better performance than the Stairmand de-
sign. Moreover, a new correlation between the Stokes number and the
Euler number is obtained. The new correlation can be used to estimate
the Stokes number if the Euler number is known.
iv
Acknowledgments
First and foremost, I would like to gratefully acknowledge the enthusiastic
supervision of my promoter Prof. Chris Lacor who gave me the opportunity
to do a PhD under his guidance. I particularity thank him for our weekly
technical discussions, which had a major influence on this thesis. I am in-
debted to him for showing great confidence in me and always pushing me
to achieve greater heights, as well as for granting me sufficient freedom to
pursue my own ideas. I can say for sure that the past years at VUB have
been the most productive days of my learning.
I also thank all the members of my thesis committee: Prof. Johan Decon-
inck, Prof. Rik Pintelon, Prof. Gunther Steenackers, Prof. Gert Desmet,
Prof. Harry van den Akker and Prof. Herman Deconinck, whose construc-
tive criticism and valuable suggestions improved the quality of this disser-
tation.
I warmly thank the IT support of our system administrator Alain Wery.
His support is invaluable for the research at our department. I greatly
appreciate him for his good mood and everlasting patience through the
perpetual stream of requests and computer problems coming towards him.
I am yet to meet someone who is so patient and always ready to help oth-
ers. Thank you very much, Alain!
The support of our secretary Jenny D’haes started even before I arrived in
Belgium. She was there for help, starting from filling down my admission
forms in Dutch, to organizing my PhD defense. Thanks a lot Jenny. A
word of thanks should also go to Birgit Buys and my Egyptian colleague
Mahmoud El-kafafy who helped me in printing the draft version.
I am pleased to acknowledge my colleagues, Ghader Ghorbaniasl. The
many discussions on mathematics and physics I have had with Ghader
were always fruitful. Santhosh Jayaraju and Kris van den Abeele gave
v
me the template that was used for this thesis, and in doing so, saved a lot
of much needed time for me. I also enjoyed the scientific discussion with
them. In this regard, I should mention Willem Deconinck as well. I taught
the students a basic techniques in computer simulation course with him.
He made it fun to do so with his pleasant mood and sense of humor, even
though I was under the pressure of writing my thesis at that time. I would
like to thank Willem once more for proofreading of some part of my the-
sis. At my first days at VUB during which we were office mates, I have
shared many laughs and a lot of joy with Mahdi Zakyani. I am pleased
to acknowledge my present and former colleagues Dean Vucinic, Matteo
Parsani, Patryk Widera, Xiadong Wang, Vivek Agantori, Florian Krause
and Dinesh Kumar. We have nice discussions and good fun at the coffee
corner.
In addition, I am sincerely thankful to Prof. Momtaz F. Sedrak, Prof.
Ahmed F. Helal, Prof. Mohammed M. Abdelrahman, Prof. Mohammed
Fatouh and Prof. Samira Elshereef who played a major role in my scien-
tific career. I have learned a lot from them during my Master and Bachelor
studies. I consider them as good examples for Egyptian professors. All my
thanks are given to the Egyptian community at Belgium for advices, sup-
port and continuous encouragement. Special thanks are given to Omar
Ellabban, Sameh Sorror, Romany Abskharon, Ehab Khatab and Wael Mo-
hammed for the help and advice they gave me during my stay here at
Brussel, especially at my first days at Belgium.
I cannot forget to give all thanks to the spirit of my late parents who I am
indebted with all my life. Lastly, and most importantly, my utmost grat-
itude is reserved to my dear wife and my two sons Omar and Ahmed for
their patience and encouragement.
vi
Jury Members
President Prof. dr. ir. Johan Deconinck
Vrije Universiteit Brussel
Vice-President Prof. dr. ir. Rik Pintelon
Vrije Universiteit Brussel
Secretary Prof. dr. ir. Gunther Steenackers
Vrije Universiteit Brussel
Internal Member Prof. dr. ir. Gert Desmet
Vrije Universiteit Brussel
External members Prof. dr. ir. Harry van den Akker
Delft University of Technology
Prof. dr. ir. Herman Deconinck
Von Karman Institute
Promoter Prof. dr. ir. Chris Lacor
Vrije Universiteit Brussel
vii
viii
Contents
1 Introduction 1
1.1 Overview of dust collectors . . . . . . . . . . . . . . . . . . . . 1
1.2 Cyclone separators: types and principals . . . . . . . . . . . 2
1.2.1 Advantages and disadvantages of cyclones . . . . . . 5
1.2.2 Applications . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.3 Principals of cyclonic separation . . . . . . . . . . . . 7
1.2.4 Factors affecting the cyclone performance . . . . . . . 10
1.3 Motivation of this work . . . . . . . . . . . . . . . . . . . . . . 11
1.4 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . 13
2 Literature Review 15
2.1 Classification of study approaches . . . . . . . . . . . . . . . 15
2.2 Mathematical models . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Experimental methods . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Computational fluid dynamics (CFD) . . . . . . . . . . . . . . 17
2.5 Discrepancy in the previous studies . . . . . . . . . . . . . . 18
2.5.1 The cone tip diameter . . . . . . . . . . . . . . . . . . . 18
2.5.2 The dust outlet geometry . . . . . . . . . . . . . . . . . 19
2.5.3 The inlet dimensions . . . . . . . . . . . . . . . . . . . 19
2.5.4 The vortex finder dimensions . . . . . . . . . . . . . . 20
2.5.5 The cyclone heights . . . . . . . . . . . . . . . . . . . . 21
2.5.6 Previous optimization studies . . . . . . . . . . . . . . 21
2.6 Summary and research plan . . . . . . . . . . . . . . . . . . . 23
3 Governing Equations 25
3.1 Turbulence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 The governing equations for the gas phase . . . . . . . . . . 28
3.2.1 Reynolds averaged Navier Stokes (RANS) . . . . . . . 29
3.2.2 Reynolds stress model (RSM) . . . . . . . . . . . . . . 31
3.2.3 Large eddy simulation (LES) . . . . . . . . . . . . . . 32
3.3 Discrete phase modeling . . . . . . . . . . . . . . . . . . . . . 37
ix
3.3.1 Governing equations for the particles . . . . . . . . . 37
3.3.2 Modeling the particle phase . . . . . . . . . . . . . . . 40
3.3.3 Stochastic trajectory approach . . . . . . . . . . . . . 41
4 Sensitivity Analysis of Geometrical Parameters 45
4.1 Sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . 45
4.1.1 Response surface methodology (RSM) . . . . . . . . . 46
4.1.2 Design of experiment (DOE) . . . . . . . . . . . . . . . 47
4.1.3 Analysis of response surfaces . . . . . . . . . . . . . . 48
4.1.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2 The cone-tip diameter . . . . . . . . . . . . . . . . . . . . . . . 52
4.2.1 Numerical simulation . . . . . . . . . . . . . . . . . . 54
4.2.2 Results and discussion . . . . . . . . . . . . . . . . . . 63
4.2.3 The flow pattern in the three cyclones . . . . . . . . . 64
4.2.4 Comparison with mathematical models . . . . . . . . 70
4.2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3 The dust outlet geometry . . . . . . . . . . . . . . . . . . . . . 72
4.3.1 Numerical simulation . . . . . . . . . . . . . . . . . . 73
4.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.3.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . 87
4.4 Closure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5 The Vortex Finder Dimensions 89
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.2 Numerical settings . . . . . . . . . . . . . . . . . . . . . . . . 92
5.2.1 Configuration of the tested cyclones . . . . . . . . . . 92
5.2.2 Solver settings . . . . . . . . . . . . . . . . . . . . . . . 92
5.2.3 Boundary conditions . . . . . . . . . . . . . . . . . . . 93
5.2.4 Grid independency study . . . . . . . . . . . . . . . . . 93
5.3 Results and discussions . . . . . . . . . . . . . . . . . . . . . 95
5.3.1 The axial variation . . . . . . . . . . . . . . . . . . . . 95
5.3.2 The flow pattern . . . . . . . . . . . . . . . . . . . . . . 96
5.3.3 The radial variation . . . . . . . . . . . . . . . . . . . . 99
5.3.4 The cyclone performance . . . . . . . . . . . . . . . . . 102
5.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6 The Inlet Dimensions 109
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.2 Numerical settings . . . . . . . . . . . . . . . . . . . . . . . . 111
6.2.1 Configuration of the five cyclones . . . . . . . . . . . . 111
6.2.2 Boundary conditions . . . . . . . . . . . . . . . . . . . 111
6.2.3 Selection of the time step . . . . . . . . . . . . . . . . . 111
6.2.4 CFD grid . . . . . . . . . . . . . . . . . . . . . . . . . . 112
x
6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.3.1 The axial variation of the flow properties . . . . . . . 113
6.3.2 The flow pattern . . . . . . . . . . . . . . . . . . . . . . 115
6.3.3 The cyclone performance . . . . . . . . . . . . . . . . . 120
6.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7 The Cyclone Height 127
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
7.2 Numerical settings . . . . . . . . . . . . . . . . . . . . . . . . 128
7.2.1 Configuration of the tested cyclones . . . . . . . . . . 128
7.2.2 Boundary conditions . . . . . . . . . . . . . . . . . . . 128
7.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.3.1 The axial variation . . . . . . . . . . . . . . . . . . . . 130
7.3.2 The radial variation . . . . . . . . . . . . . . . . . . . . 131
7.3.3 The flow pattern . . . . . . . . . . . . . . . . . . . . . . 135
7.3.4 The performance . . . . . . . . . . . . . . . . . . . . . 137
7.3.5 The cone height versus the barrel height . . . . . . . 141
7.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
8 Optimization 147
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
8.2 Single-objective using MM model . . . . . . . . . . . . . . . . 149
8.2.1 CFD comparison between the two designs . . . . . . . 150
8.2.2 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . 154
8.3 Single-objective using RBFNN . . . . . . . . . . . . . . . . . 158
8.3.1 Radial basis function neural networks (RBFNN) . . . 158
8.3.2 Evaluation of different mathematical models . . . . . 164
8.3.3 Design of experiment (DOE) . . . . . . . . . . . . . . . 169
8.3.4 CFD Comparison between the two designs . . . . . . 179
8.3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . 185
8.4 Multi-objective optimization using GA . . . . . . . . . . . . . 186
8.4.1 Artificial neural network (ANN) approach . . . . . . . 187
8.4.2 Single objective optimization . . . . . . . . . . . . . . 194
8.4.3 Optimal cyclone design for best performance . . . . . 202
8.4.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . 204
8.5 Multi-objective optimization using CFD data . . . . . . . . . 215
8.5.1 Design variables and approaches . . . . . . . . . . . . 215
8.5.2 The desirability function . . . . . . . . . . . . . . . . . 220
8.5.3 Artificial neural network (ANN) approach . . . . . . . 231
8.5.4 Optimization Using Genetic Algorithms . . . . . . . . 235
8.5.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . 239
xi
9 Conclusions and Future Directions 243
9.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
9.1.1 The most significant geometrical factors . . . . . . . . 243
9.1.2 The impact of geometry . . . . . . . . . . . . . . . . . 244
9.1.3 Optimization . . . . . . . . . . . . . . . . . . . . . . . . 245
9.1.4 Multi-objective optimization . . . . . . . . . . . . . . . 246
9.2 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . 248
A Mathematical models 251
A.1 General assumptions . . . . . . . . . . . . . . . . . . . . . . . 251
A.2 Barth model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
A.3 The Muschelknautz method of modeling (MM) . . . . . . . . 256
A.4 Stairmand model for pressure drop . . . . . . . . . . . . . . . 259
A.5 Purely empirical models for pressure drop . . . . . . . . . . . 260
A.6 Iozia and Leith model for the cut-off diameter . . . . . . . . 261
A.7 Rietema model for cut-off diameter . . . . . . . . . . . . . . . 262
B Optimization Techniques 263
B.1 Nelder-Mead . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
B.2 Genetic algorithms (GA) . . . . . . . . . . . . . . . . . . . . . 266
B.2.1 Description of the genetic algorithm process . . . . . 268
B.2.2 Genetic operators . . . . . . . . . . . . . . . . . . . . . 268
B.3 Multi-objective optimization . . . . . . . . . . . . . . . . . . . 272
Bibliography 281
xii
Nomenclature
Upper-Case Roman
Ai inlet cross sectional area [m]
AR total inside area of the cyclone [m2]
Bc cyclone cone-tip diameter [m]
CD particle drag coefficient
Cin inlet dust concentration, [kg/m3]
D cyclone barrel diameter [m]
Dx cyclone vortex finder diameter [m]
Eu Euler number [-]
Fdrag drag force [N]
Frx Froude number (inertia force / gravitational force) [-]
HCS length of the control surface [m]
Ht cyclone total height [m]
K fluctuating kinetic energy [m2/s2]
K vortex finder entrance factor
Li distance between the inlet section and cyclone center [m]
Le distance between the outlet section and the barrel top [m]
Ln cyclone natural length [m]
P fluctuating kinetic energy production [m2/s3]
P mean pressure [N/m2]
Pr Prandtl number [-]
Qin gas volume flow rate [m3/s]
R cyclone radius [m]
R2 coefficient of multiple determination
Rb dust outlet radius [m]
Rij Reynolds stress tensor [m2/s2]
Rep Reynolds number based on the relative particle velocity [-]
ReR cyclone body Reynolds number [-]
Rm geometric mean radius [m]
Rx vortex finder radius [m]
xiii
S vortex finder length [m]
Sg geometrical swirl number [-]
Stk50 Stokes number at the cut-off diameter [-]
V cyclone volume [m3]
Lower-Case Roman
a cyclone inlet height [m]
acceleration [m/s2]
b cyclone inlet width [m]
d50 cut-off diameter [µm]
dp particle diameter [m]
f total friction factor [-]
fair gas friction factor [-]
fr friction factor due to wall roughness [-]
fsm friction factor for smooth wall [-]
gi acceleration due to gravity in i direction [m/s2]
h barrel height [m]
hc cone height [m]
k turbulent kinetic energy [m2/s2]
ks wall relative roughness of the cyclone wall [m]
m mass [kg]
mp dust mass flow rate [kg/s]
p static pressure [N/m2]
t flow physical time [s]
tres flow average residence time [s]
ui flow velocity component in i direction [m/s]
ui mean velocity [m/s]
u′i fluctuating velocity component in i direction [m/s]
upi particle velocity in i direction [m/s]
vx mean gas velocity through the vortex finder [m/s]
vzw wall axial velocity [m/s]
x particle diameter [µ m]
x50 cut-off diameter [µm]
xi position [m]
Upper-Case Greek
∆ filter width [m]
∆P pressure drop in the cyclone [N/m2]
∆Pbody pressure drop in the cyclone body [N/m2]
∆Px pressure drop in the vortex finder [N/m2]
xiv
Lower-Case Greek
α gas moment-of-momentum ratio at inlet [-]
β ratio of inlet width to the cyclone radius [-]
δij Kronecker delta [-]
ε turbulence dissipation rate [m2/s3]
εν viscous dissipation [m2/s3]
µ dynamic viscosity [kg/(m s)]
µt turbulent (eddy) viscosity [kg/(m s)]
ν kinematic viscosity (ν = µ/ρ) [m2/s]
νt turbulent (eddy) kinematic viscosity (νt = µt/ρ) [m2/s]
ξ spatial separation [m]
ρ gas density [kg/m3]
ρbulk bulk density of the solid [kg/m3]
ρp particle density [kg/m3]
ρstr bulk density of the strand layer at the walls [kg/m3]
σij viscous stress tensor [N/m2]
τij subgrid scale stress tensor [N/m2]
Subscripts
CS at the control surface
g gas
in at the inlet surface
p particle properties
θ angular (tangential component)
w close to the wall
Abbreviations
ANOVA Analysis Of Variance
ANN Artificial Neural Networks
CFD Computational Fluid Dynamics
CFL Courant-Friedrichs-Lewy number
DNS Direct Numerical Simulation
DOE Design Of Experiment
DPM Discrete Phase Modeling
GA Genetic Algorithm
xv
GCI Grid Convergence Index
GEC Grade Efficiency Curve
LDA Laser Doppler Anemometry
LES Large Eddy Simulation
MM Muschelknautz Method of modeling
PIV Particle Image Velocimetry
RANS Reynolds Averaged Navier-Stokes
RBFNN Radial Basis Function Neural Network
RNG Renormalization Group
RSM Response Surface Methodology
RSM Reynolds Stress turbulence Model
SGS Subgrid Scale Model
xvi
Chapter 1
Introduction
1.1 Overview of dust collectors
There are four principal types of industrial dust collectors [77] namely,
inertial separators, fabric collectors, wet scrubbers and electrostatic pre-
cipitators. The selection of one type depends mainly on the particle size as
shown in Fig. 1.1. The classification of different dust collectors is shown in
Fig. 1.2.
The inertial separators separate dust from gas streams using a combina-
tion of forces, such as centrifugal, gravitational, and inertial. These forces
move the dust to an area where the forces exerted by the gas stream are
minimal. The separated dust is moved by gravity into a hopper, where
it is temporarily stored. The three primary types of inertial separators
are settling chambers, baffle chambers, and centrifugal collectors (e.g., cy-
clone separator). A settling chamber consists of a large box installed in the
ductwork. The sudden expansion at the chamber reduces the speed of the
dust-filled airstream and heavier particles settle down. Settling chambers
are simple in design and can be manufactured from almost any material.
However, they are seldom used as primary dust collectors because of their
large space requirements and low efficiency. A practical use is as preclean-
ers for more efficient collectors. Baffle chambers use a fixed baffle plate
that causes the conveying gas stream to make a sudden change of direc-
tion. Large-diameter particles do not follow the gas stream but continue
into a dead air space and settle. Baffle chambers are used as precleaners
for more efficient collectors.
Fabric collectors are commonly known as baghouses. Fabric collectors use
filtration to separate dust particulates from dusty gases. They are one
1
Chapter 1. Introduction
Particle size (micron)
10-3
10-2
10-1
100
101
102
103
Settling chambers
Cyclone separator
Liquid scrubbers
Filters
Electrostatic precipitators
Figure 1.1: Suitable methods for removing particles from a gas stream [128]
of the most efficient types of dust collectors available and can achieve a
collection efficiency of more than 99% for very fine particulates.
Dust collectors that use liquid are commonly known as wet scrubbers. In
these systems, the scrubbing liquid (usually water) comes into contact with
a gas stream containing dust particles. The greater the contact of the gas
and liquid streams, the higher the dust removal efficiency.
The electrostatic Precipitators use electrostatic forces to separate dust par-
ticles from exhaust gases. A number of high-voltage, direct-current dis-
charge electrodes are placed between grounded collecting electrodes. The
contaminated gases flow through the passage formed by the discharge and
collecting electrodes. The airborne particles receive a negative charge as
they pass through the ionized field between the electrodes. These charged
particles are then attracted to a grounded or positively charged electrode
and adhere to it [77].
1.2 Cyclone separators: types and principals
The gas cyclones belong to the type of centrifugal separators. A gas cyclone
is a stationary mechanical device that utilizes centrifugal force to separate
solid or liquid particles from a carrier gas. The flow enters near the top
through the tangential inlet, which gives rise to an axially descending spi-
ral of gas and a centrifugal force field that causes the incoming particles
to concentrate along, and spiral down, the inner walls of the cyclone sep-
arator. The collected particulates are allowed to exit out an underflow
2
1.2. Cyclone separators: types and principals
Figure 1.2: Classification of dust collectors
3
Chapter 1. Introduction
Figure 1.3: Typical cyclone separator
pipe while the gas phase reverses its axial direction of flow and exits out
through the vortex finder (gas outlet tube) [77]. Figure 1.3 shows a typical
cyclone separator.
The cyclone separator is one of the most efficient and robust dust sepa-
rators. Its robustness results from lack of moving parts and the ability
to withstand harsh operating environments. Moreover, cyclones are well
suited for high pressure and temperature applications.
Centrifugal collectors use cyclonic action to separate dust particles from
the gas stream. In a typical cyclone, the dust gas stream enters tangen-
tially forcing the flow into a spiral movement. The centrifugal force cre-
ated by the circular flow throws the dust particles toward the wall of the
cyclone. After striking the wall, the particles fall into a hopper located
underneath. The most common types of centrifugal, or inertial, collectors
in use today are single-cyclone separators and multiple-cyclone separators
(multiclone). Single-cyclone separators create a dual vortex to separate the
dust from the gas. The main vortex spirals downward and carries most of
the heavier particles. The inner vortex, created near the bottom of the
4
1.2. Cyclone separators: types and principals
cyclone, spirals upward and carries finer dust particles. Multiple-cyclone
separators consist of a number of small-diameter cyclones, operating in
parallel and having a common gas inlet and outlet. Multiclones operate on
the same principle as cyclones–creating a main downward vortex and an
ascending inner vortex.
1.2.1 Advantages and disadvantages of cyclones
Compared with the other separation devices, the cyclone separators ad-
vantages are [77]:
• the collected product remains dry and, normally useful.
• low capital investment and maintenance costs in most applications.
• very compact in most applications.
• can be used under extreme processing conditions, in particular at
high temperatures and pressures and with chemically aggressive feeds.
• no moving parts.
• very robust.
• can be constructed from most any material suitable for the intended
service including plate steel, casting metals, alloys, aluminum, plas-
tics, ceramics, etc.
• can be equipped with erosion or corrosion resistant or ‘particle re-
pelling type liners, such as Teflon. Internal surfaces may be electro
polished to help combat fouling.
• can be fabricated from plate metal or, in the case of smaller units,
cast in molds.
• can, in some processes, handle sticky or tacky solids with proper liq-
uid irrigation.
• can separate either solids or liquid particulates; sometimes both in
combination with proper design.
Some disadvantages of cyclones are [77]:
• low efficiency for particle sizes below their ‘cut-off diameter when op-
erated under low solids-loading conditions.
• usually higher pressure loss than other separator types, including
bag filters and low pressure drop scrubbers.
• subject to erosive wear and fouling if solids being processed are abra-
sive or ‘sticky.
• can operate below expectations if not designed and operated properly.
Although this problem, as well as the erosion and fouling problem
mentioned above, is not unique to cyclones.
5
Chapter 1. Introduction
Figure 1.4: Applications of cyclone separator
1.2.2 Applications
Due to the mentioned advantages, cyclones have found application in vir-
tually every industry where there is a need to remove particles from a gas
stream. Figure 1.4 presents some examples of cyclones industrial applica-
tions with wide range of sizes, locations and applications. Today, cyclone
separators are found in:
• ship unloading installations
• power stations
• spray dryers
• fluidized bed and reactor riser systems (such as catalytic crackers
and cockers)
• synthetic detergent production units
• food processing plants
• crushing, separation, grinding and calcining operations in the min-
eral and chemical industries
• fossil and wood-waste fired combustion units (normally upstream of
6
1.2. Cyclone separators: types and principals
a wet scrubber, electrostatic precipitator or fabric filter)
• vacuum cleaning machines
• dust sampling equipment
Cyclones have also been used to classify solids on the basis of their charac-
teristic such as their mass, density, size, or shape. Because of their simple
construction and high reliability, cyclones are also used very effectively
to separate two-phase gas-liquid mixtures, such as the entrained droplets
exiting a venturi scrubber or other types of scrubber. Other examples in-
clude the removal of water droplets from steam generators and coolers and
oil-mist from the discharge of air compressors. Likewise, they have been
widely applied in process machinery to remove entrained oil and hydrocar-
bon droplets generated from spraying, injection, distillation, or most any
process that results in the production of entrained droplets or a two-phase
mixture. They have even been used as inlet devices to prevent foaming in
gravity separation drums [77].
1.2.3 Principals of cyclonic separation
In centrifugal devices, the dust-laden gas is initially brought into a swirling
motion. The dust particles are slung outward to the wall, and transported
downward to the dust outlet by the downwardly directed gas flow near the
wall. A sketch of a standard reverse-flow, cylinder-on-cone cyclone with a
tangential, slot-type inlet is shown in Fig. 1.5.
For the standard, reverse-flow cyclone, (with a so-called slot type of entry)
the swirling motion is brought about by designing the inlet in such a man-
ner that it forces the gas to enter the unit on a tangent to the inner body
wall. The inlet is normally of rectangular cross section. As the gas swirls,
it moves axially downwards in the outer part of the separation space. In
the conical part of the cyclone, the gas is slowly forced into the inner re-
gion of the cyclone, where the axial movement is upwardly directed. This
flow pattern is often referred to as a double vortex: an outer vortex with
a downwardly directed axial flow and an inner one with an upwardly di-
rected flow. The gas exits the cyclone through the so-called vortex finder,
which extends downward from the center of the roof. This outlet pipe goes
by many different names, with vortex tube and dip-tube being the most
common, aside from the vortex finder [77]. The particles in the inlet gas
are slung outwards to the wall in the centrifugal field, and are transported
to the dust exit by the downwardly directed gas flow near the wall. Below
more details of the flow pattern in the separation space will be given.
The geometry of a cyclone with a slot type inlet is determined by the fol-
lowing eight dimensions as shown in Fig. 1.5:
7
Chapter 1. Introduction
aS
h
D
b
Ht
Dx
Bc
hc
Figure 1.5: Sketches of a reverse-flow, cylinder-on-cone cyclone with a tangential
inlet. The geometrical notation is indicated in the right sketch
1. the body diameter (barrel diameter) D2. the total height of the cyclone (from roof to dust exit) Ht
3. the vortex finder diameter Dx
4. the vortex finder length (from the roof of the separation space) S5. the inlet height a6. the inlet width b7. the height of the conical section hc or the height of the cylindrical
section h8. the cone-tip diameter (dust exit diameter) Bc
1.2.3.1 Real vortex flow
Swirling flow, or vortex flow, occurs in different types of equipment, such as
cyclones, hydrocyclones, spray dryers and vortex burners [77]. Two basic
types of swirling flows can be distinguished:
1. forced vortex flow, which is a swirling flow with the same tangential
velocity distribution as a rotating solid body
2. free vortex flow, which is the way a frictionless fluid would swirl.
8
1.2. Cyclone separators: types and principals
Tangential velocity
Distance from center
Forced vortex
Free vortexRankine vortex
Figure 1.6: The tangential velocity distribution in a real vortex [77]
The tangential velocity in such a swirl is such that the moment-of-
momentum of fluid elements is the same at all radii.
The tangential velocity distribution in a real swirling flow is intermediate
between these two extremes. Now imagine first that the swirling fluid has
an infinite viscosity (behaves like a solid body). Hence, no shearing motion
exists between fluid layers at different radii. In this case, the fluid ele-
ments at all radial positions are forced to have the same angular velocity
Ω which equals vθ/r where vθ is the tangential velocity. This is the forced
vortex flow or solid-body rotation:
vθ = Ωr (1.1)
In the other extreme, if the swirling fluid has no viscosity, the motion of
a given fluid element is not influenced by the neighboring elements at
smaller and larger radii. If in such a fluid, we bring an element to a
smaller radius, its tangential velocity will increase, since its moment-of-
momentum (mvθr) will be conserved. Such a vortex is called a free or
frictionless vortex. In such a flow, we have rvθ = C, with C a constant,
so that:
vθ =C
r(1.2)
This is the second basic swirl flow. A real swirling flow normally has a
core of near solid-body rotation surrounded by a region of near loss-free
rotation as sketched in Fig. 1.6. This is called a Rankine vortex.
The flow and pressure distribution within cyclones is more easily under-
stood if we make clear the relation between static and dynamic pressures;
p and 1/2ρv2, respectively, with ρ the density. The well-known Bernoulli
9
Chapter 1. Introduction
equation for steady flow of a frictionless, constant density fluid, which can
be derived from the Navier-Stokes equations, states that:
p
ρ+ gh+
1
2v2 = constant along a streamline (1.3)
In this equation, we recognize the static and dynamic pressures (the lat-
ter is often called the velocity head) as the first and third terms on the
left-hand side. They have been divided by the fluid density. This equa-
tion shows that static and dynamic pressures can be interchanged in the
flow field [77]. In areas where the velocity is high, the static pressure will
be low and vice versa. It is especially important to appreciate this interde-
pendence between static and dynamic pressure when dealing with swirling
flows [77]. The left-hand side of Eq. 1.3 is sometimes called Bernoulli’s tri-
nomial. The second term is unimportant relative to the two others when
discussing gas cyclones, since the fluid density is relatively low, and height
differences not very large. In an actual flow situation, the fluid is not fric-
tionless. Frictional dissipation of mechanical energy will therefore cause
Bernoullis trinomial to decrease in the flow direction, i.e. the trinomial
is no longer constant, but decreases along a streamline. Frictionless flow
is, nevertheless, a reasonably good approximation in the outer part of the
swirl in a cyclone; Bernoulli’s trinomial does not change very much there
[77].
1.2.4 Factors affecting the cyclone performance
Figure 1.7 indicates the possible factors affecting the cyclone performance
and flow pattern. These factors can be sub-classified as follows:
1. Cyclone dimensions
• Cyclone diameter
• Inlet height
• Inlet width
• Vortex finder diameter
• Vortex finder length
• Cylinder height
• Cyclone total height
• Cone tip diameter
2. Particle properties
• Density
• Shape
• Diameter and distribution
10
1.3. Motivation of this work
• Mass loading
3. Gas properties
• Velocity
• Density
• Viscosity
• Temperature
• Pressure
4. Other factors
• Wall roughness
• Shape of vortex finder
• Eccentricity of vortex finder
1.3 Motivation of this work
In spite of the fact that the use of cyclone separators is common in many
industrial applications, an accurate prediction tool for their behaviors is
still not available. The challenge of this work is therefore, a detailed study
of the flow phenomena in cyclones and the design of an optimum cyclone
separator (minimum pressure drop and maximum collection efficiency).
The cyclone performance parameters are governed by many operational
parameters (e.g. the gas flow rate and temperature) and geometrical pa-
rameters. This study focuses only on the effect of the geometrical param-
eters on the flow field pattern and on the performance of the tangential
inlet cyclone separators.
For that the specific goals of the work are the following:
• Determine the most significant factors affecting the cyclone perfor-
mance based on previous studies and statistical analysis of data.
• Study each (significant) parameter separately to obtain more details
on its effect on the flow field pattern.
• To obtain the most efficient cyclone design for minimum pressure
drop (using the most robust mathematical models, artificial neural
networks and CFD data sets).
• To obtain the most efficient cyclone design for best performance (min-
imum pressure drop and minimum cut-off diameter) using multi-
objective optimization techniques.
11
Ch
ap
ter
1.
Intro
du
ction
Figure 1.7: Cause and effect plot for cyclone separator
12
1.4. Outline of the thesis
The study is limited to reverse flow gas cyclone separators operating at low
mass loading.
1.4 Outline of the thesis
The thesis is organized in ten chapters. After an introduction and overview
given in the previous sections, chapter 2 deals with the different study
approaches of cyclone separators. In addition, an overview is presented of
work reported in the literature on the application of CFD in cyclone mod-
eling. Moreover, the discrepancy of the results from the previous studies
has been discussed in details. Furthermore, a summary of the previous op-
timization studies is given. Chapter 3 presents the governing equations
for the carrier (gas) and the discrete phase (solids) in detail. The applica-
tion of response surface methodology and design of experiment statistical
techniques to estimate the most significant geometrical parameters are
given in chapter 4. Furthermore, chapter 4 deals with the uncertainty
of the significant effect of cone tip diameter and the necessity of including
the dust outlet geometry in the simulation domain. Chapter 5 presents
the study of the effect of the vortex finder dimensions, whereas chapter
6 deals with the effect of the inlet dimensions. The effects of both the
cone height and the barrel height on the flow field and performance are
discussed in chapter 7. The new optimized cyclones for minimum pres-
sure drop and best performance using different techniques are analyzed in
chapter 8. The used dataset are collected from different sources; math-
ematical models calculations, artificial neural network models and CFD
simulations data. Chapter 9 summarizes the main conclusions and some
future directions. The thesis contains two appendices. Appendix A, deals
with the details of eight mathematical models used to predict the cyclone
performance parameters. The details of the two optimization techniques
used in this thesis are given in appendix B.
13
Chapter 1. Introduction
14
Chapter 2
Literature Review
The most important parameter that affects the cyclone performance and
flow pattern is the cyclone geometry. For reversed flow cyclones, there are
seven geometrical parameters, viz. the inlet height a, the inlet width b,the vortex finder diameter (gas outlet tube diameter) Dx, the vortex finder
length S, the cylindrical part height h, the cyclone total height Ht, and the
cone-tip diameter Bc. All these dimensions are expressed in terms of bar-
rel diameter D as shown in Fig. 2.1. The two performance indicators used
are the pressure drop and the particle separation (collection) efficiency.
The latter is normally expressed as a “grade efficiency curve” a graph of
the collection efficiency against the particle diameter. For low mass load-
ing cyclone separators, the cut-off diameter x50 is usually given instead of
grade efficiency curves.
2.1 Classification of study approaches
There is a widespread literature on the effect of cyclone geometry on per-
formance, using one or more of the four main approaches of study, which
are:
1. Analytical methods (mathematical models), which can be classified
into [194]:
(a) theoretical and semi-empirical models
(b) statistical models
2. Experimental measurements
3. Computational fluid dynamics (CFD) simulations
4. Artificial neural networks (ANN) approach
Recently, optimization studies based on data available from one of the
main four approaches have been performed. Also artificial neural net-
15
Chapter 2. Literature Review
Figure 2.1: The cyclone separator dimensions
works become a tool to study the effect of cyclone geometry on performance.
2.2 Mathematical models
The theoretical models were developed by many researchers e.g., Shep-
herd and Lapple [157], Alexander [1], First [58], Stairmand [166], Barth
[9], Avci and Karagoz [5], Zhao [192], Karagoz and Avci [90] and Chen and
Shi [22]. These models were derived from physical descriptions and math-
ematical equations. They require a very detailed understanding of gas
flow pattern and energy dissipation mechanisms in cyclones. In addition,
due to using different assumptions and simplifying conditions, different
theoretical or semi-empirical models can lead to significant differences be-
tween predicted and measured results. Predictions by some models are
sometimes twice the experimental values [172].
Since the first application of aerocyclones in 1886 [3], theories for the es-
timation of both particle collection efficiency and pressure drop of cyclone
have been developed by many contributors using different methods with
various simplifying assumptions. During the past 50 years, interest in
particle collection and pressure theories has steadily increased [196]. The
most widely used mathematical models for the cut-off diameter and pres-
sure drop estimation are:
• Barth model [9]
• The Muschelknautz method of modeling (MM) [29, 77, 114–116, 174,
175]
16
2.3. Experimental methods
• Stairmand model [165]
• Shepherd and Lapple model [157]
• Casal and Martinez-Bent model [21]
• Ramachandran model [139]
• Iozia and Leith model [84]
• Rietema model [142]
The interested reader can refer to appendix A for more details about these
eight models.
2.3 Experimental methods
There are numerous experimental investigations performed on the cyclone
separators. The majority of these studies used either laser doppler anemom-
etry (LDA) or particle image velocimetry (PIV) to obtain the flow field pat-
tern. Some of the studies only measured the pressure drop and collec-
tion efficiency without any details of the flow fields. For instance, Dirgo
and Leith [43] measured the collection efficiency and pressure drop for the
Stairmand high efficiency cyclone at different flow rates. Hoekstra et al.
[75] measured the mean and fluctuating velocity components for gas cy-
clones with different geometric swirl numbers (SG = πDxD4ab ) by means of
the laser doppler anemometry technique. The experimental data shows
the strong effect of the geometric swirl number on the mean flow charac-
teristics, in particular with respect to vortex core size and the magnitude of
the maximum tangential velocity. It is shown that the forced vortex region
of the flow is dominated by the so-called precessing vortex core.
Hoffmann et al. [76] investigated the effect of the cyclone length on the
separation efficiency and the pressure drop experimentally and theoret-
ically by varying the length of the cylindrical segment of a cylinder-on-
cone cyclone. They found for cyclone lengths from 2.65 to 6.15 cyclone
diameters, a marked improvement in cyclone performance with increas-
ing length up to 5.5 cyclone diameters; beyond this length the separation
efficiency was dramatically reduced. For the interested reader, other ex-
perimental results on cyclones can be found in [36, 43, 66, 74–76, 93, 102,
105, 121–124, 127, 138, 152, 163, 184].
2.4 Computational fluid dynamics (CFD)
Boysan et al. [14] presented the first CFD investigation in the field of cy-
clone separators. From that time, the CFD technique becomes a widely
used approach for the flow simulation and performance estimation for cy-
clone separators. For example, Griffiths and Boysan [68] computationally
17
Chapter 2. Literature Review
investigated three cyclone samplers. They reported that the CFD predicted
pressure drops are in excellent agreement with the measured data. The
CFD modeling approach is also able to predict the features of the cyclone
flow field in great details, which providing a better understanding of the
fluid dynamics in cyclone separators [68]. Consequently, CFD approach is
a reliable and relatively inexpensive method of examining the effects of a
number of design changes. Moreover, this makes the CFD methods rep-
resent a cost-effective route for geometry optimization in comparison with
the experimental approach. Another example, Gimbun et al. [64] success-
fully applied CFD to predict and to evaluate the effects of temperature
and inlet velocity on the pressure drop of gas cyclones [194]. The success-
ful application of CFD technique in different studies in cyclone separators
has been reported by many researchers [e.g., 6, 11, 23, 49, 50, 52, 62–
64, 146, 147, 186, 198]. Nevertheless, CFD is still more expansive in com-
parison with the mathematical models approach. The main reasons behind
the cost of the CFD approach with respect to the mathematical methods
are:
1. In essence, the CFD process requires expert intervention by an ex-
pert researcher at every stage (mesh generation, solver settings and
post processing).
2. The license cost of the grid generator, solver and post processor.
3. The running cost especially for unsteady simulations which need also
parallel processing.
4. CFD results always need (i) validation with experimental results (ii)
perform the same simulation on different grids to be sure that the
obtained results are grid independent.
2.5 Discrepancy in the previous studies
In this section, the effect of the geometrical parameters on the performance
in terms of the two indicators; the pressure drop and the separation effi-
ciency (cut-off diameter) will be discussed briefly based on the available
literatures. More details will be presented in the subsequent chapters.
2.5.1 The cone tip diameter
Very little information is available on the effects of changing the cone bot-
tom (tip) diameter, which determines the cone shape if other cyclone di-
mensions are fixed [184]. Regarding this effect, discrepancies and uncer-
tainties exist in the literature. Bryant et al. [17] observed that if the vor-
tex touched the cone wall, particle re-entrainment occurred and efficiency
decreased, so collection efficiency will be lower for cyclones with a small
18
2.5. Discrepancy in the previous studies
cone opening (cone tip diameter). While according to Xiang et al. [184], a
cone is not an essential part for cyclone operation, although it serves the
practical purpose of delivering collected particles to the central discharge
point. However, Zhu and Lee [200] stated that the cone provides greater
tangential velocities near the bottom for removing smaller particles.
2.5.2 The dust outlet geometry
Conventional cyclones always have a dustbin attached to the cone to collect
the separated solid particles. When a gas flow stream enters the dustbin
(closed at bottom), some of the flow will return into the cone and distribute
some of the separated particles. This phenomena called “re-entrainment”
and it will affect the separation efficiency of the cyclone [138]. There has
been little work concerning the dust outlet geometries [e.g., 40, 47, 78,
123].
Regarding this influence, discrepancies and uncertainties exist in the lit-
erature. Xiang and Lee [186] reported that the dustbin connected to the
cyclone should be incorporated in the flow domain as it affects the results
obtained. On the other hand, numerous studies were performed with-
out dustbin [e.g., 159, 178] with good matching with experimental results.
Obermair et al. [123] performed cyclone tests with five different dust outlet
geometries to find the influence of the dust outlet geometry on the sepa-
ration process. They showed that separation efficiency can be improved
significantly by changing the dust outlet geometry, and they reported that
further research is needed to clarify precise effects of dust outlet geometry.
The effect of a dipleg (a vertical tube between the cyclone and the dustbin)
was posed and investigated by several researchers [e.g., 78, 92].
2.5.3 The inlet dimensions
The effects of the inlet dimensions on the cyclone performance (pressure
drop and cut-off diameter) have been reported in many articles. Casal
and Martinez-Benet [21] proposed the following empirical formula for the
dimensionless pressure drop (Euler number),
Eu = 11.3
(ab
Dx
)2
+ 2.33 (2.1)
implying proportionality with the square of the inlet area. On the other
hand, Ramachandran et al. [139] proposed,
Eu = 20
a b
D2x
SD
HD
hD
Bc
D
1/3
(2.2)
19
Chapter 2. Literature Review
i.e., a linear relation with the inlet area. Iozia and Leith [84, 85] presented
a correlation to estimate the cut-off diameter x50 and found proportionality
to (ab)0.61. The importance of the inlet dimensions becomes clear after the
study of the natural length (or vortex length) by several researchers, e.g.,
Alexander [1]. The cyclone has two spiral motions, outer and inner. In the
reverse flow cyclone, the outer vortex weakens and changes its direction
at a certain axial distance Ln from the vortex finder [29]. This distance
is usually called the turning length, natural length or vortex length of the
cyclone. The inlet area is one of the relevant parameters influencing the
natural length. Alexander [1] found that Ln decreased proportionally to
the inlet area but the opposite trend has been also reported [29].
Numerous studies have been performed for the effect of geometrical pa-
rameters on the flow pattern and performance [e.g., 15, 62, 102, 140, 184]
while the effect of cyclone inlet dimensions remained largely unexplored.
The articles investigating the effect of cyclone geometry report only briefly
on the effect of inlet section dimensions without sufficient details about the
effects on the flow pattern and velocity profiles. The new trend is to study
the multi-inlet cyclone [e.g., 103, 187, 195].
2.5.4 The vortex finder dimensions
The vortex finder size is an especially important dimension, which sig-
nificantly affects the cyclone performance as its size plays a critical role
in defining the flow field inside the cyclone, including the pattern of the
outer and inner spiral flows. Saltzman and Hochstrasser [151] studied
the design and performance of miniature cyclones for repairable aerosol
sampling, each with a different combination of three cyclone cone lengths
and three gas outlet diameters. Iozia and Leith [84] optimized the cyclone
design parameters, including the gas outlet diameter, to improve the cy-
clone performance using their optimization program. Kim and Lee [95]
described how the ratio of the diameters of cyclone body D and the vortex
finder Dx affected the collection efficiency and pressure drop of cyclones,
and proposed an energy-effective cyclone design. Moore and Mcfarland
[111] also tested cyclones, with six different vortex finders, and concluded
that the variation in the gas outlet diameter under the constraint of a
constant cyclone Reynolds number produced a change in the aerodynamic
particle cut-off diameter. Recently, Hoekstra [74] investigated the effect of
gas outlet diameter on the velocity profile using 2-D axisymmetric simu-
lations. Lim et al. [102] examined experimentally the effect of the vortex
finder shape on the collection efficiency at different flow rates but without
any explanation on its effect of the flow field pattern and velocity profiles.
Raoufi et al. [140] duplicated numerically the same study of Lim et al.
20
2.5. Discrepancy in the previous studies
[102] with limited details about the effect of the gas outlet diameter on the
flow field pattern and velocity profile.
2.5.5 The cyclone heights
Limited literatures are available for the effect of cyclone height. Zhu and
Lee [200] have conducted detailed experiments on cyclones of different
height and found that, the cyclone height can influence considerably the
separation efficiency of the cyclones. However, they did not provide any
information about the flow pattern or even explanation for the efficiency
results. Hoffmann et al. [76] investigated the effect of the cyclone length
on the separation efficiency and the pressure drop experimentally and the-
oretically. The cyclone performance improves with increasing length up
to 5.5 cyclone diameters beyond this length the separation efficiency was
dramatically reduced. However, they did not present any contour plot or
velocity profile to assist the explanation for the effect of cyclone height
on performance. Recently, Xiang and Lee [186] have repeated the same
study of Zhu and Lee [200] for the effect of cyclone height computationally
via steady three-dimensional simulation using Reynolds stress turbulence
model (RSM). They found that the tangential velocity decreases with in-
creasing cyclone height, which is responsible for the lower separation effi-
ciency observed in long cyclones. The explanation of this behavior was not
adequate. Moreover, no particle tracking study was presented.
2.5.6 Previous optimization studies
Due to the wide range of industrial applications of the cyclone separator,
it was a matter of study for decades. However, the optimization studies on
it is quite limited in literature. Moreover, many of these studies are not
coherent studies. Ravi et al. [141] carried out a multi-objective optimiza-
tion of a set of N identical reverse-flow cyclone separators in parallel by
using the non-dominated sorting genetic algorithm (NSGA). Two objective
functions were used: the maximization of the overall collection efficiency
and the minimization of the pressure drop. Non-dominated Pareto optimal
solutions were obtained for an industrial problem in which 165 m3/s of
air was treated. In addition, optimal values of several decision variables,
such as the number of cyclones and eight geometrical parameters of the cy-
clone, are obtained. Their study shows that the barrel diameter, the vortex
finder diameter, and the number of cyclones used in parallel, are the impor-
tant decision variables influencing the optimal solutions. Moreover, their
study illustrates the applicability of NSGA in solving multi-objective opti-
mization problems involving gas-solid separations. The main drawbacks of
21
Chapter 2. Literature Review
their study are: (1) They used the model of Shepherd and Lapple [157] for
predicting the dimensionless pressure drop (Euler number). In the Shep-
herd and Lapple model, the Euler number depends on only three factors
(Eu = 16ab/D2x) and they used it to optimize the seven geometrical parame-
ters. (2) The barrel diameter, number of parallel cyclones and the gas veloc-
ity have been included into the optimization design space. Consequently, it
is not devoted to the geometrical ratio. (3) They used many side constraints
on the geometrical values (0.4 ≤ a/D ≤ S/D, 0.15 ≤ b/D ≤ (1 − Dx/D)/2if 0.5 ≤ Dx/D ≤ 0.6) these constraints prevent searching for the global op-
timization geometrical ratios for the seven geometrical parameters. (4) No
table for the non-dominated Pareto front points are presented from which
the designer can select certain geometrical ratio set (optimal solution).
Swamee et al. [172] investigated the optimum values of the number of
cyclones to be used in parallel, the diameter of cyclone barrel D and exit
pipe Dx, when a specified flow rate of gas is to be separated from solid
particles, and the cut diameter is already specified. They used Stairmand
model for calculation of pressure drop and Gerrard and Liddle formula
for the cut-off diameter [172] which is not a widely used model. Instead
of handling two objective functions, they blended the two objective into a
single objective problem which is not the suitable method to considering
two conflicting objectives (the pressure drop and cut-off diameter).
Safikhani et al. [148] performed a multi-objective optimization of cyclone
separators. First, they simulated many cyclones to obtain the pressure
drop and the cut-off diameter and used artificial neural network approach
to obtain the objective function values. Finally, a multi-objective genetic
algorithms are used for Pareto based optimization of cyclone separators
considering two conflicting objectives. However, the design variables are
only four (instead of seven): the barrel height, the cone height, the vortex
finder diameter and length. So they ignored the effect of inlet dimensions,
which has been acknowledged by other researchers as significant geometri-
cal parameters for the cyclone flow field and performance (cf. Elsayed and
Lacor [50, 52–54]). Moreover, they did not explain why they selected these
particular parameters. Furthermore, they applied four side constraints on
the four tested variables, which prevent searching for the global optimiza-
tion.
Pishbin and Moghiman [130] applied genetic algorithm for optimum cy-
clone design. They studied the seven geometrical parameters. The data
used for optimization was obtained from 2-D axisymmetric simulations.
However, the flow in the cyclone separator is 3-D unsteady. Instead of
using multi-objective genetic algorithm (e.g., non-dominated sorting ge-
netic algorithm II (NSGA-II) [34]) they used the weighted-sum genetic al-
gorithm. In this technique, a weighting factor is assigned for each objective
22
2.6. Summary and research plan
function based on the user preference. The main shortages of the Pishbin
and Moghiman [130] study are: (1) How to select the weighting factor,
in scientific and engineering problems, it is a non-trivial task to find the
one solution of interest to the decision maker [26]. The decision maker’s
weight (no matter how defined) could be greater than necessary as more
acceptable solutions are missed. Optimizing mostly profit could lead to
poor quality or reliability, not a good compromise [26]. The weighted-sum
genetic algorithm usually does not find all Pareto front points of inter-
est. But this approach is a simple approach for handling multi-objective
optimization problem. Another simple but better result can be obtained
using the desirability function approach [54, 126]. (2) No table for the non-
dominated Pareto front points is presented from which the designer can
select a certain geometrical ratio.
Safikhani et al. [149] carried out a multi-objective optimization using the
genetic algorithm technique to obtain the best vortex finder dimension (di-
ameter and length) and shape (convergent and divergent). Four design
variables have been investigated; the vortex finder diameter, angle, upper-
part length and lower-part length of the vortex finder. They applied neural
networks to obtain a meta-model for the pressure drop and collection effi-
ciency from CFD dataset. The main shortages of the Safikhani et al. [149]
study are: (1) They used dimensional values instead of dimensionless, and
applied side constraints, which prevent the optimization procedure from
obtaining global optimization. (2) The selection of only the vortex finder
dimension as the design variables and neglecting the interaction with the
vortex finder diameter with the other dimensions, especially the inlet di-
mensions [52, 53].
2.6 Summary and research plan
After studying the existing literature on cyclones, the following conclusion
can be drawn:
• The separation mechanism inside cyclone separators is not well un-
derstood yet, and needs more investigations.
• Nearly all published articles have no systematic and complete study
for the effect of geometrical parameters on the flow field and perfor-
mance. In more details: the geometry parameters are not given as
dimensionless numbers; the effect of a certain parameter is obtained
with no knowledge about the effect of others or possible interactions;
the results of different articles are sometimes in contradiction.
• In some cases, there are a lot of results about the effect of a certain
parameter on the performance but on different dimensions (and not
23
Chapter 2. Literature Review
on the same dimensionless ratio) and also sometimes at different op-
erating conditions
• Some parameters have less interest compared with others like the
effect of vortex finder shape and number of inlet sections.
• The particle tracking in almost all computational investigations are
not well reported (or missing), and need more investigations
• Some articles give only the flow pattern results without a study of
the effect of any parameter on the performance, or with limited dis-
cussion.
From the remarks above, it is clear that more investigation is still needed
about the flow patterns inside cyclone separators and how the geometri-
cal parameters affect the performance. This guided us to construct the
following research plan:
1. Undertake a systematic study for the effect of seven geometrical pa-
rameters to obtain the most significant factors on the cyclone perfor-
mance. This step can be subdivided into:
(a) Application of design of experiment to obtain a table of runs, the
performance parameters will be estimated using the most robust
mathematical models.
(b) Application of the response surface methodology to determine
the most significant factors and any possible interactions.
2. Once, the most significant factors are selected, CFD simulations for
each particular parameter will be performed to investigate in detail
the effect of these parameters on the flow field pattern and perfor-
mance (pressure drop and cut-off diameter).
3. Optimization of the cyclone geometry, first single-objective optimiza-
tion for minimum pressure drop and then for best performance (multi-
objective optimization for minimum pressure drop and minimum cut-
off diameter). The required data for optimization can be obtained us-
ing mathematical models, artificial neural networks data (based on
experimental measurements available in literature) or CFD simula-
tions.
24
Chapter 3
Governing Equations
3.1 Turbulence
Hinze [73] described turbulence as follows: “Turbulence fluid motion is an
irregular condition of flow in which various quantities show a random vari-
ation with time and space coordinates so that statistically distinct average
values can be discerned”. Wilcox [183] explains that turbulence consists
of a continuous spectrum of scales that vary from smallest to largest over
several orders of magnitude. The idea of a series turbulent eddies is often
used. He also stated that: “A turbulent eddy can be thought of as a lo-
cal swirling motion whose characteristic dimension is the local turbulence
scale.” These eddies overlap in space and the large ones carry the smaller
ones. The conversion of energy in a turbulent flow follows a cascading pro-
cess where the kinetic energy is transferred from the larger eddies to the
smaller ones.
The energy cascade
The concept of the energy cascade has been introduced in 1922 by Richard-
son [132]. The idea is that the kinetic energy enters the turbulence at the
largest scales of motion. This energy is then transferred by inviscid pro-
cesses to smaller and smaller scales until the smallest scales. The energy
is dissipated by viscous processes [35], Fig. 3.1. In 1941, Kolmogorov iden-
tified the smallest scales of turbulence to be those that now bear his name
[35, 132]. In this concept, the turbulence can be considered to be composed
of eddies of different sizes that overlap in space. The largest eddies are
characterized by the length scale l which is comparable to the flow scale
25
Chapter 3. Governing Equations
Figure 3.1: Schematic representation of the energy cascade [32, 35]. η is the Kol-
mogorov length scale, lDI is the length scale dividing the dissipation and the iner-
tial subrange, lEI is the length scale dividing the energy-containing range and the
inertial subrange, l0 is the length scale and L is the characteristic length [35].
(the large scales are of the order of the flow geometry), and a characteristic
velocity u which is on the order of the turbulence intensity.
If l and u are the length and velocity scales of the largest eddy, the time
scale is derived as,
τ =l
u(3.1)
The large energy containing eddies give away their kinetic energy to slightly
smaller-scale eddies with which the large scales interact. The process of
kinetic energy transfer continues in a similar fashion until the smallest
scale eddies are reached, where the frictional forces become so large that
the kinetic energy is converted into internal energy. This process of energy
transfer and dissipation is referred to as the energy cascade process. The
scales at which the dissipation (ε) takes place are the smallest scales, and
are also referred to as the Kolmogorov scales. They can be estimated from
the large-scale properties as follows,
ε =u2
τ=
u3
l(3.2)
26
3.1. Turbulence
Figure 3.2: Energy spectrum for a turbulent flow [183]
Since the processes of dissipation in the smallest scales are due to viscous
forces, the properties of the smallest eddies can be estimated using the flow
kinematic viscosity (ν) and the dissipation (ε) itself. The length, velocity
and time scales are given by:
lη =
(ν3
ε
)1/4
(3.3)
uη = (νε)1/4
(3.4)
τη =(νε
)1/2(3.5)
The turbulent length scale l is related to the wave number κ as κ = 2π/l.The energy spectrum E(κ) for a turbulent flow is as shown in Fig. 3.2.
From dimensional analysis, the Kolmogorov -5/3 law characterizes the in-
ertial subrange which is given by,
E(κ) = Cκε2/3κ−5/3 (3.6)
Cκ is the Kolmogorov constant.
27
Chapter 3. Governing Equations
3.2 The governing equations for the gas phase
All fluid motions (laminar or turbulent) are governed by a set of dynami-
cal equations namely the continuity, momentum and the energy equation
(Navier-Stokes equations),
∂ρ
∂t+
∂
∂xi(ρui) = 0 (3.7)
∂
∂t(ρui) +
∂
∂xj(ρuiuj) = − ∂p
∂xi+
∂τij∂xj
(3.8)
∂
∂t(ρE) +
∂
∂xi(ρHui) =
∂
∂xi(τjiuj − qi) (3.9)
ui(~x, t) represents the i-th component of the fluid velocity at a point in
space ~x and time t.
p(~x, t) is the static pressure.
τij(~x, t) are the viscous stresses.
ρ(~x, t) is the fluid density.
E and H are the total energy and total enthalpy per unit mass.
qi in Eq. 3.9 is the heat flux which is proportional to the temperature
gradient.
qi = −κ∂T
∂xi(3.10)
where κ is the thermal conductivity.
The Mach numbers associated with air flow in cyclone separators are very
nominal, which allows the flow to be treated as incompressible. Further-
more, the air behaves as a Newtonian fluid, in which case the viscous
stresses are related to the incompressible fluid motion using a property
of fluid, viscosity [86].
τij = 2µ
(sij −
1
3skkδij
)(3.11)
sij is the instantaneous strain rate tensor given by,
28
3.2. The governing equations for the gas phase
sij =1
2
(∂ui
∂xj+
∂uj
∂xi
)(3.12)
For incompressible flows, Eqs. 3.7 and 3.8 are simplified to the following
form,
∂uj
∂xj= 0 (3.13)
∂ui
∂t+ uj
∂ui
∂xj= −1
ρ
∂p
∂xi+ ν
∂2ui
∂xj∂xj(3.14)
In this thesis, the temperature effects are ignored and hence Eq. 3.9 is
uncoupled from the continuity and momentum equations.
The four main numerical procedures for solving the Navier-Stokes equa-
tions are the direct numerical simulation (DNS), the large eddy simula-
tion (LES), the detached eddy simulation (DES) and the reynolds aver-
aged navier stokes (RANS) approach. The most accurate approach is DNS
where the whole range of spatial and temporal scales of turbulence are
resolved. Since all the spatial scales, from the smallest dissipative Kol-
mogorov scales (lη) up to the energy containing integral length scales (l),are needed to be resolved by the computational mesh, the number of points
required in one direction is of the order,
N =l
lη(3.15)
The number of points required for a resolved DNS in three dimensions can
be estimated as,
N =
(l
lη
)3
∼(ul
ν
)9/4
= Re9/4 (3.16)
The number of grid points required for fully resolved DNS is enormously
large, especially for high Reynolds number flows, and hence DNS is re-
stricted to relatively low Reynolds number flows. DNS is generally used
as a research tool for analyzing the mechanics of turbulence, such as tur-
bulence production, energy cascade, energy dissipation, noise production,
drag reduction, etc [86].
3.2.1 Reynolds averaged Navier Stokes (RANS)
When the flow is turbulent, it is convenient to analyze the flow in two
parts, a mean (time-averaged) component and a fluctuating component
29
Chapter 3. Governing Equations
[86],
Ui = U i + u′
i
P = P + p′
Tij = T ij + τ′
ij
Overline is a shorthand for the time average and in case of RANS, Ui ≡ Ui
and u′
i=0. The above technique of decomposing is referred to as Reynolds
Decomposition. Inserting this decomposition into the instantaneous equa-
tions and time averaging results in the Reynolds averaged Navier-Stokes
equations (RANS).
∂U j
∂xj= 0 (3.17)
∂U i
∂t+ U j
∂U i
∂xj= −1
ρ
∂P
∂xi+ ν
∂2U i
∂xj∂xj− ∂
∂xj
(u
′
iu′
j
)(3.18)
u′
iu′
j in the last term of Eq. 3.18 represents the correlation between fluctu-
ating velocities and is called the Reynolds stress tensor. All the effects of
turbulent fluid motion on the mean flow are lumped into this single term
by the process of averaging [86]. This will enable great savings in terms
of computational requirements. On the other hand, the process of aver-
aging generates six new unknown variables. Now, in total there are ten
unknowns (3-velocity, 1-pressure, 6-Reynolds stresses) and only four equa-
tions (1-continuity, 3 components of momentum equation). Hence, we need
six equations to close this problem. This is referred to as the Closure prob-
lem. Based on the way we close the Reynolds stress tensor, there are two
main categories, namely the eddy viscosity models and the Reynolds stress
model.
The Reynolds stress tensor resulting from time averaging of the Navier-
Stokes equations is closed by replacing it with an eddy viscosity multiplied
by velocity gradients. This is referred to as the Boussinesq assumption.
u′
iu′
j = −νt
(∂U i
∂xj+
∂U j
∂xi
)(3.19)
where νt is the turbulent (eddy) kinematic viscosity. In order to make Eq.
3.19 valid upon contraction because of Eq. 3.17, it should be rewritten as,
30
3.2. The governing equations for the gas phase
u′
iu′
j = −νt
(∂U i
∂xj+
∂U j
∂xi
)+
2
3ρδijk (3.20)
where δij is the Kronecker delta, δij = 1 if i = j and δij = 0 if i 6= j. k is the
turbulent kinetic energy given by,
k =1
2u
′
iu′
i (3.21)
The eddy viscosity is treated as a scalar quantity and is determined using
a turbulent velocity scale v and a length scale l, based on the dimensional
analysis.
νt ∼ vl (3.22)
There are different types of eddy viscosity models (EVM) based on the way
we close the eddy viscosity. Algebraic or zero equation EVM’s normally
use a geometric relation to compute the eddy viscosity. In one equation
EVM’s, one turbulence quantity is solved and a second turbulent quantity
is obtained from algebraic expression. These two quantities are used to
describe the eddy viscosity. In two equation EVM models the two turbulent
quantities are solved to describe the eddy viscosity. The interested reader
can refer to [24, 177, 183] for more details.
In the Reynolds stress models (RSM), an equation is solved for each Reynolds
stress component as well as one length scale determining equation. How-
ever, RSM’s are computationally much more demanding when compared
to EVM’s.
3.2.2 Reynolds stress model (RSM)
RSM is regarded as the most appropriate RANS turbulence model for cy-
clone flows [178]. Equation 3.18 can be written as [147],
∂Ui
∂t+ Uj
∂Ui
∂xj= −1
ρ
∂P
∂xi+ ν
∂2Ui
∂xj∂xj− ∂
∂xjRij (3.23)
where Rij = u′iu
′j is the Reynolds stress tensor. The RSM turbulence model
provides differential transport equations for evaluation of the turbulence
31
Chapter 3. Governing Equations
stress components (Eq. 3.24).
∂
∂tRij + Uk
∂
∂xkRij =
∂
∂xk
(νtσk
∂
∂xkRij
)−[Rik
∂Uj
∂xk+Rjk
∂Ui
∂xk
]
− C1ε
K
[Rij −
2
3δijK
]− C2
[Pij −
2
3δijP
]− 2
3δijε
(3.24)
where the turbulence production terms Pij are defined as [147]:
Pij = −[Rik
∂Uj
∂xk+Rjk
∂Ui
∂xk
], P =
1
2Pij (3.25)
With P being the fluctuating kinetic energy production. νt is the turbulent
(eddy) kinematic viscosity; and σk = 1, C1 = 1.8, C2 = 0.6 are empirical
constants The transport equation for the turbulence dissipation rate, ε, is
given as [98]:
∂ε
∂t+ Uj
∂ε
∂xj=
∂
∂xj
[(ν +
νtσε
) ∂ε
∂xj
]− Cε1 ε
KRij
∂Ui
∂xj− Cε2 ε
2
K(3.26)
In Eq. (3.26), K =1
2u′iu
′i is the fluctuating kinetic energy, and ε is the
turbulence dissipation rate. The values of constants are σε = 1.3, Cε1 =1.44 and Cε2 = 1.92.
3.2.3 Large eddy simulation (LES)
The large eddy simulation technique was developed based on an implica-
tion from Kolmogorov’s theory of self-similarity that the large eddies of the
flow are dependent on the geometry while the smaller scales are more uni-
versal [86]. Hence, the big three-dimensional eddies which are dictated by
the geometry and boundary conditions of the flow involved are directly cal-
culated (resolved) whereas the small eddies which tend to be more isotropic
are modeled.
Large eddy simulation (LES) treats the large eddies more exactly than
the small ones. In LES, the large scales in space are calculated directly as
illustrated in Fig. 3.3. The LES technique is based on a separation between
large and small scales. A grid size first has to be determined. Those scales
that are of a characteristic size greater than the grid size are called large or
resolved scales, and others are called small or subgrid scales. The subgrid
scales are included by way of a model called the subgrid model [154]. An
32
3.2. The governing equations for the gas phase
elaborate explanation on LES can be found in several text books such as
[132, 150, 183].
Figure 3.3: Decomposition of the energy spectrum [154]
3.2.3.1 LES equations
In case of RANS, the instantaneous continuity and momentum equations
(Eq. 3.7 and 3.8) are time averaged to obtain a steady form of the averaged
equations (Eq. 3.17 and 3.18). In case of LES, instead of time-averaging,
the instantaneous time-dependent equations are filtered. Filtering is a
method that separates the resolvable scales from the subgrid scales. Fil-
tering can be performed in either wave number space or the physical space.
The filter cut-off should lie somewhere in the inertial range of the spectrum
(Fig. 3.2).
In finite volume methods, box filters are always used because the finite
volume discretization itself implicitly provides the filtering operation. One
of the earliest volume average box filters was given by Deardorff [33].
φ(X, t) =1
∆3
∫ x−0.5∆x
x−0.5∆x
∫ y−0.5∆y
y−0.5∆y
∫ z−0.5∆z
z−0.5∆z
φ(ξ, t)dξdηdζ (3.27)
φ = φ+ φs (3.28)
In the above equation, φ denotes the resolvable scale filtered variable and
φs denotes the sub-grid scale fluctuation. ∆ is the filter width given by
∆ = (∆x∆y∆z)1/3.
Leonard [100] defined a generalized filter as a convolution integral which
is given by,
φ(X, t) =
∫ ∫ ∫G(X − ξ; ∆)φ(ξ, t) d3ξ (3.29)
33
Chapter 3. Governing Equations
G is the filter function that determines the scale of resolved eddies. The
filter function is normalized by requiring that,
∫ ∫ ∫G(X − ξ; ∆) d3ξ = 1 (3.30)
The filter function in terms of the volume average box filter (Eq. 3.27) can
be written as,
G(X − ξ; ∆) =
1/∆3, |x− ξ| < ∆x/2
0, otherwise
Finally, the decomposition of the flow into a filtered part and a sub-grid
part looks like,
Ui = Ui + usi
P = P + ps
Tij = Tij + τsij
The hat operator . in the above equations represents the filtering oper-
ation as opposed to the time-averaging in case of RANS. Moreover, con-
trary to RANS, where the average of fluctuations is zero, in LES,Ui 6= Ui
[150] and us 6= 0. Further details on the filtering methods can be found in
[32, 97, 132, 183].
Inserting the above decomposition into the instantaneous equations re-
sults in the following filtered Navier-Stokes equations,
∂Uj
∂xj= 0 (3.31)
[∂Ui
∂t+ Uj
∂Ui
∂xj
]= −1
ρ
∂P
∂xi+ ν
∂2Ui
∂xj∂xj−
∂τsgsij
∂xj(3.32)
where τsgsij are the sub-grid scale stresses.
3.2.3.2 SGS modeling
From the energy cascade, explained in the beginning of this chapter, it
is apparent that the energy transfer occurs from the bigger scales to the
34
3.2. The governing equations for the gas phase
smaller scales. Hence, the main purpose of an SGS model is to represent
the energy sink [86]. The representation of the energy cascade is an aver-
age process. However, locally and instantaneously the transfer of energy
can be much larger or much smaller than the average. Moreover, there is
also the phenomenon of energy backscatter in the opposite direction [129].
Ideally speaking, SGS models should actually account for all these phe-
nomena. However, if the grid scale is much finer than the dominant scales
of the flow, even a crude SGS model will result in good predictions of the
behavior of the dominant scales [86].
The sub-grid scale stresses τsgsij in Eq. 3.32 are given by,
τsgsij = UiUj − UiUj (3.33)
By using the definition of filtering as given by Eq. 3.28 we can further work
out τsgsij as,
τsgsij = −(UiUj −
[
(Ui + usi )(Uj + us
j)
])(3.34)
τsgsij = −usiu
sj︸ ︷︷ ︸
Reynolds
+(−Uius
j −Ujus
i )︸ ︷︷ ︸Cross−term
+ UiUj − UiUj︸ ︷︷ ︸
Leonard
(3.35)
Leonard [100] shows that the Leonard stresses can significantly drain en-
ergy from the resolvable scales and they can be directly computed. On the
other hand, Wilcox [183] mentions that Leonard stresses are of the same
order of magnitude as the truncation error when a finite-difference scheme
of second-order accuracy is used, and thus it is implicitly represented. The
cross-term stresses are dispersive in nature and largely account for the
backscatter effects. Modeling them with a purely dissipative model such as
Smagorinsky would be in conflict because of its dispersive nature [97]. In
many applications, it is assumed that the Leonard and cross-term stresses
can be neglected, and only the Reynolds stresses remain to be modeled. It
is the same case in the present work [86]. The interested reader is referred
to Sagaut [150] for the detailed review of various SGS models available in
literature.
Smagorinsky model
One of the simplest SGS model is the Smagorinsky model [161]. The un-
known subgrid-scale stresses are modeled employing the Boussinesq as-
sumption as in the case of RANS. The subgrid-scale stress are related to
35
Chapter 3. Governing Equations
the eddy viscosity as follows,
τij −1
3τkkδij = −νt
(∂Ui
∂xj+
∂Uj
∂xi
)(3.36)
The eddy-viscosity is modeled as,
νt = L2s
√2SijSij (3.37)
where Ls is the length-scale for the sub-grid scale and is given by CsV1/3,
where V is the computational cell volume. It is interesting to note that the
length scale is now the filter width rather than the distance to the closest
wall as in RANS. Cs is a constant which is taken to be 0.17. The only
disadvantage of the Smagorinsky model is the constant Cs, which is not
really a constant, but is flow dependent. It is found to vary between 0.065
[110] and 0.3 [89]. In the dynamic version, which was first proposed by
Germano et al. [60], Cs is dynamically computed based on the information
provided by the resolved scales of motion. The specification of Ls as CsV1/3
is not justifiable in the viscous wall region as it incorrectly leads to a non-
zero turbulent shear-stress at the wall. In order to rectify this, Moin and
Kim [110] use a Van Driest damping function to specify the length scale
as,
Ls = CsV1/3
[1− exp
(y+
A+
)](3.38)
where y+ = uτd/ν is the non-dimensional distance from wall, uτ is the wall
shear stress velocity, d is the distance to the nearest wall and A=25 is the
Van Driest constant.
The above-described SGS model is a standard version as defined in Smagorin-
sky [161]. The LES simulations in the present thesis are performed em-
ploying the Fluent flow solver. The Smagorinsky model implemented in
Fluent deviates slightly from the standard version in the following ways
[86],
• The length-scale for the sub-grid scale is computed as min(κd,CsV1/3).
κ is the von Karman constant (typically a value of 0.41 is used), d is
the distance to the closest wall. κd is indeed one of the first mixing
length models in the literature to handle the turbulent viscosity and
was proposed by Prandtl [133]. Van Driest damping is basically an
improved version of Prandtl’s mixing length model. Both the Prandtl
and the Van Driest model are algebraic and from the zero-equation
models category.
36
3.3. Discrete phase modeling
• The constant Cs in Fluent is taken to be 0.1 instead of 0.17 as was
originally proposed. The value of 0.17 for Cs was originally derived
for homogeneous isotropic turbulence in the inertial subrange. How-
ever, this value was found to cause excessive damping of large-scale
fluctuations in transitional flows near solid boundaries, and has to be
reduced in such regions [59]. A Cs value of around 0.1 has been found
to yield the best results for a wide range of flows, and is the default
value in Fluent.
Dynamic Smagorinsky-Lilly model
Germano et al. [60] and subsequently Lilly [101] conceived a procedure in
which the Smagorinsky model constant Cs is dynamically computed based
on the information provided by the resolved scales of motion [59]. The
dynamic procedure thus obviates the need for users to specify the model
constant Cs in advance. The Smagorinsky model constant is dynamically
computed instead of given as an input to the solver, but clipped to zero or
0.23 if the calculated model constant is outside this range to avoid numer-
ical instabilities [59]. The second advantage of the dynamic Smagorinsky-
Lilly model over the Smagorinsky model is the treatment near the wall.
In the dynamic Smagorinsky-Lilly model, a damping function for the eddy
viscosity near the wall is not required, since the model constant goes to
zero in the laminar region just near the wall [55, 113].
3.3 Discrete phase modeling
3.3.1 Governing equations for the particles
Based partly on the physical properties of dust particles and partly on
the mathematical modeling effort required, there are certain reasonable
assumptions made to describe the particles transport in a fluid medium
[86]. The major simplifying assumptions are as follows,
• The particles are assumed to be spherical.
• The ratio of particle to fluid density is very large: The density of the
dust particle is much higher when compared to the fluid medium
which is air.
• Drag force is the dominant force: This is a direct result of the previous
assumption. Since the density of the particles are much higher than
the density of the fluid medium, several forces such as the lift force,
Basset force and buoyancy force can be readily discarded as they have
negligible effect on the particles transport [86].
37
Chapter 3. Governing Equations
102
100
10-2
10-4
10-7 10-5 10-110-3
One-Way
Coupling
Two-Way
Coupling
Four-Way
Coupling
Negligable
effect on
turbulence
Particles
enhance
production
Particles
enhance
dissipation
Fluid Particles Fluid Particles Fluid Particles Particles
Φp
τp /τe
Dilute suspension Dense suspension
Figure 3.4: Map for particle-turbulence modulation [44]. φp is the ratio of particles
volume to the volume occupied by particles and fluids. τp is the particle response
time, τp = ρpd2/(18µ), where ρp is the particle density, d is the particle diameter, ρ
is the fluid viscosity. τe is the turnover time of large eddy (time scale= l/u) [44].
• One-way coupling: The phenomenon of mutual mass, momentum
and energy transfer between the phases is termed as coupling. El-
ghobashi [44] proposed a map of regimes of interactions between
particles and fluid turbulence as shown in Fig. 3.4. For values of
dispersed-phase volume fraction less than 10−6, particles have negli-
gible effects on turbulence and this is termed as one-way coupling.
The volume fraction of dust particles we are dealing with in the
present thesis is much less than 10−6 and hence one-way coupling
is assumed. In the second regime which lies between 10−6 − 10−3,
the existence of particles can augment the turbulence if the ratio of
the particle response time to the turnover time of a large eddy is
greater than unity, or can attenuate turbulence if the ratio is less
than unity. This interaction is called two-way coupling. In the third
regime where the volume fractions are greater than 10−3, in addi-
tion to two-way coupling between particles and turbulence, particle
collisions take place and hence this regime is termed as four-way cou-
pling.
38
3.3. Discrete phase modeling
Incorporating all the above assumptions, the Lagrangian equations gov-
erning the particle motion can be written as [86]:
dxp
dt= up (3.39)
dup
dt= Fd (u− up) + gx
(ρp − ρ)
ρp(3.40)
xp is the particle position, gx is the gravitational force, ρ and ρp are the
density of the fluid and the particle respectively.
Generally, the particle moves with a different velocity than the fluid at any
given point. The difference in fluid velocity (u) and the particle velocity
(up), termed as the slip velocity (u − up), leads to an unbalanced pressure
distribution as well as viscous stresses on the particle surface which yields
a resulting force called drag force. In Eq. 3.40, the term Fd (u− up) is the
drag force per unit particle mass. Fd is given by [86]:
Fd =1
τp
CdRep24
(3.41)
where τp is the particle relaxation time given by,
τp =ρpd
2p
18µ(3.42)
Laws of drag coefficient
The drag coefficient Cd is a function of particle Reynolds number (Rep).
Various experimentally based empirical correlations for the drag coeffi-
cient based on Rep are available in the literature. The Reynolds number of
the particle is defined as:
Rep = ρ dp|u− up|
µ(3.43)
In Fluent, the drag coefficient for spherical particles is calculated by using
the correlations developed by Morsi and Alexander [112]. It is given by,
Cd = a1 +a2Rep
+a3Re2p
(3.44)
where a1, a2 and a3 are constants that apply to smooth spherical particles
in a stipulated range of Rep as given in Table 3.1.
39
Chapter 3. Governing Equations
Table 3.1: Drag coefficient parameters used for Morsi and Alexander [112]
Rep a1 a2 a3< 0.1 0 24 0
0.1 < 1.0 3.69 22.73 0.0903
1 < 10.0 1.222 29.1667 -3.8889
10.0 < 100.0 0.6167 46.5 -116.67
100.0 < 1000.0 0.3644 98.33 -2778
1000.0 < 5000.0 0.357 148.62 -4.75
5000.0 < 10000.0 0.46 -490.546 57.87
10000.0 < 50000.0 0.5191 -1662.5 5.4167
3.3.2 Modeling the particle phase
Coming to the fundamental mathematical modeling of two-phase flow, the
two most widely used approaches are the Eulerian continuum approach
and the Lagrangian trajectory approach.
Eulerian continuum approach
In an Eulerian approach, the particles are treated as a second fluid which
behaves like a continuum and the equations are developed for average
properties of the particles. For example, the particle velocity is the average
velocity over an averaging volume. This approach is most suitable when
one requires a macroscopic field description of dispersed phase properties
such as pressure, mass flux, concentration, velocity and temperature. Eu-
lerian approach is more suitable for simulating large-scale particle flow
processes. However, this approach requires sophisticated modeling in or-
der to describe the key effects and phenomena found in industrial pro-
cesses [31, 86].
Lagrangian trajectory approach
A Lagrangian approach is useful when the particle phase is so diluted that
the description of particle behavior by continuum models is not feasible.
The motion of a particle is expressed by ordinary differential equations
in Lagrangian coordinates and are directly integrated to obtain individ-
ual tracks of particles [86]. To solve the Lagrangian-equation for a par-
ticular moving particle, the dynamic behavior of the gas phase (generally
obtained by an Eulerian approach) and other particles surrounding this
moving particle should be pre-determined. Since the particle velocity and
the corresponding particle trajectory are calculated for each particle, this
approach is more suitable to obtain the discrete nature of motion of par-
ticles. However, to obtain statistical averages with reasonable accuracy, a
40
3.3. Discrete phase modeling
large number of particles will have to be tracked. An advantage of using
the Lagrangian approach is the ability to vary easily the physical prop-
erties associated with individual particles such as diameter, density, etc.
Moreover, local physical phenomena related to the particle flow behavior
can be easily probed. Hence, the Lagrangian models can also be used for
validation, testing and development of continuum models [31].
The Lagrangian approach is classified into two types namely, determinis-
tic trajectory methods and Stochastic trajectory methods based on the effect
of turbulence. In a deterministic method, all the turbulent transport pro-
cesses of the particle phase are neglected where as the stochastic method
takes into account the effect of fluid turbulence on the particle motion by
considering instantaneous fluid velocity in the formulation of the equation
of particle motion. In the present thesis, the dust particles are modeled
with a stochastic Lagrangian approach [86].
3.3.3 Stochastic trajectory approach
One of the most frequently used models is the eddy interaction model
(EIM) first introduced by Hutchinson et al. [83] and further developed
by Gosman and Ioannides [67].
The instantaneous motion of particles governed by Equations 3.39 and
3.40 can be written in a general form as given below [86].
dx
dt= up (3.45)
dup
dt=
1
τp(u− up) + g (3.46)
The instantaneous fluid velocity u in the above equation is represented as
the sum of the mean and fluctuating velocity,
u = U + u′
(3.47)
Assuming isotropic turbulence, we have,
u′2 = v′2 = w′2 =2
3k (3.48)
where k is the turbulent kinetic energy. Furthermore, it is assumed that
the local velocity fluctuations of the fluid phase obey a Gaussian probabil-
ity density distribution. Most stochastic models in practical use are de-
41
Chapter 3. Governing Equations
rived from the formulation of Gosman and Ioannides [67], which is given
by,
u′
=
√2
3k ∗ ζ (3.49)
where ζ is a random number drawn from a normal probability distribu-
tion with zero mean and unit standard deviation. The minimal random
number generator of Park and Miller with Bays-Durham shuffle [135] is
implemented [86]. The random number generator returns a uniform ran-
dom derivative with zero mean and unit standard deviation.
The chosen fluctuation is referred to a turbulent eddy whose size (length
scale) and life-time (time scale) is known. Sommerfeld et al. [164] proposed
the following relations for eddy parameters,
te = ctk
ε(3.50)
le = te
√2
3k (3.51)
where ct was taken to be 0.3.
Figure 3.5 shows a 2-D schematic representation of an eddy inside a rect-
angular domain. At any given particle position (xp, yp), the eddy param-
eters are first evaluated based on the local fluid kinetic energy and dissi-
pation rate. The particle position (xp, yp) is assumed to be located at the
center of this hypothetical eddy. It is accepted that each eddy has its own
fluctuation u′
, which remains constant until the particle leaves this eddy.
The particle leaving an eddy is based on a certain interaction time of the
particle with the eddy. Once this interaction time is reached while time
integration of particle equations, the particle is assumed to have left the
present eddy. Now, based on the new position of the particle, new eddy pa-
rameters are calculated and a new fluctuation u′
is assigned to this eddy.
This procedure may be repeated for as many interaction times as required
for the particle to traverse the required distance. If a statistically signif-
icant number of particles are tracked in this way, the ensemble averaged
behavior should represent the turbulent dispersion induced by the prevail-
ing fluid field [67]. The interaction time is the minimum of two time scales,
one being a typical turbulent eddy lifetime and the other the crossing-time
of the particle in the eddy [67].
tint = min(te, tc) (3.52)
42
3.3. Discrete phase modeling
Figure 3.5: 2-D illustration of a particle within an eddy [86]
The crossing-time is defined as,
tc = −τp ln
[1−
(le
τp|u− up|
)](3.53)
where τp is the particle relaxation time, le the eddy length scale and |u−up|the magnitude of slip velocity. In circumstances where le/(τp|u − up|) > 1,
Eq. 3.53 has no solution. This can be interpreted as the particle trapped
by an eddy, in which case tint = te [67].
The mentioned eddy interaction model is needed only for RANS simula-
tion to take into account the effect of turbulence on the particle. In LES
simulations, the effect of the resolved velocity fluctuations on the parti-
cles is accounted for and there is no need for an eddy interaction model
like in RANS [86]. In this thesis, the effect of the subgrid scale velocity
fluctuations on the particle dispersion is assumed negligible and hence not
modeled. Figure 3.6 represents a flow chart demonstrating the steps in-
volved in tracking one injected particle.
43
Chapter 3. Governing Equations
Start with the location of one particle injected from the inlet surface at a certain point
Compute the surrounding cells of the current control volume where the particle lies
Interpolate the flow variables at the particle position
Determine the eddy parameters (and fluctuating velocity using interpolated flow variables and the
random number generator in case of RANS)
Compute the forces acting on the particle
Determine the integration time-step based on the cell size & eddy parameters
Perform a time-step integration to obtain updated particle position and velocity
Check the distance between particle position & the nearest boundary cell (wall or outlet)
Particle is stuck on wall (trapped) or reached outlet (escaped)
Is the particle still in the current cell?
Is the current cell a boundary cell?
No
No
Is the distance to wall/outlet <= particle radius?
No
Figure 3.6: Flow chart demonstrating the steps involved in tracking one injected
particle [86]
44
Chapter 4
Sensitivity Analysis of
Geometrical Parameters
4.1 Sensitivity analysis
The cyclone separator performance and the flow field are affected mainly
by the cyclone geometry where there are seven geometrical parameters,
namely, the inlet section height a and width b, the vortex finder diameter
Dx and length S, the barrel height h, the cyclone total height Ht and the
cone-tip diameter Bc. all of these parameters are always expressed as a
ratio of cyclone diameter D, as shown in Fig. 4.1 and Table 4.1.
In this study, only the effect of geometry was taken into account. Neverthe-
less, what about the effect of flow rate on the performance. Overcamp and
Scarlett [125, p 369, Fig. 6] studied the effect of changing Reynolds num-
ber on the cut-off diameter (Stokes number) and found that for Reynolds
number values beyond 1E4, the effect of increasing the Reynolds number
on the cut-off diameter is very limited. Furthermore, Karagoz and Avci
[90, p 863, Fig. 7] studied the effect of increasing the Reynolds number
on the pressure drop and found that beyond Reynolds number of 2E4 any
increase in the Reynolds number has nearly no effect on the pressure drop.
The values of the Reynolds number for all tested cases in this thesis have
Reynolds number higher than 2E4. Consequently, the effect of flow rate
Table 4.1: The Stairmand high-efficiency design
a/D b/D Dx/D Ht/D h/D S/D Bc/D0.5 0.2 0.5 4.0 1.5 0.5 0.375
45
Chapter 4. Sensitivity Analysis of Geometrical Parameters
Figure 4.1: The cyclone separator dimensions
can be safely neglected.
4.1.1 Response surface methodology (RSM)
The usual method of optimizing any experimental set-up is to adjust one
parameter at a time, keeping all others constant, until the optimum work-
ing conditions are found. Adjusting one parameter at a time is necessar-
ily time consuming, and may not reveal all interactions between the pa-
rameters. In order to fully describe the response and interactions of any
complex system a multivariate parametric study must be conducted [30].
Since there are seven geometrical parameters to be investigated, the best
technique is to perform this study using the response surface methodology
(RSM).
RSM is a powerful statistical analysis technique which is well suited to
model complex multivariate processes, in applications where a response
is influenced by several variables, and the objective is to optimize this re-
sponse. Box and Wilson first introduced the theory of RSM in 1951 [13].
RSM today is the most commonly used method of process optimization. Us-
ing RSM one may model and predict the effect of individual experimental
parameters on a defined response output, as well as locating any inter-
actions between the experimental parameters which otherwise may have
been overlooked. RSM has been employed extensively in the field of engi-
neering and manufacturing, where many parameters are involved in the
process [70, 106, 118, 167–170].
46
4.1. Sensitivity analysis
In order to conduct a RSM analysis, one must first design the experiment,
identify the experimental parameters to adjust, and define the process re-
sponse to be optimized. Once the experiment has been conducted and the
recorded data tabulated, the RSM analysis software models the data and
attempts to fit a second-order polynomial to this data [30]. The generalized
second-order polynomial model used in the response surface analysis was
as follows:
Y = β0 +
7∑
i=1
βiXi +
7∑
i=1
βiiX2i +
∑∑
i<j
βijXiXj (4.1)
where β0, βi, βii, and βij are the regression coefficients for intercept, linear,
quadratic and interaction terms, respectively. Xi and Xj are the indepen-
dent variables, and Y is the response variable (Euler number).
4.1.2 Design of experiment (DOE)
The statistical analysis is performed through three main steps. Firstly,
construct a table of runs with combinations of values of the independent
variables via the commercial statistical software Statgraphics centurion
XV by giving the minimum and maximum values of the seven geometrical
factors under investigation as input. Secondly, perform the runs by esti-
mating the pressure drop (Euler number) using the MM model (cf. Sec. A.3,
page 256). Thirdly, fill in the values of pressure drop in the Statgraphics
worksheet and obtain the response surface equation with main effect plot,
interaction plots, Pareto chart and response surface plots beside the opti-
mum settings for the new cyclone design.
Table 4.2 depicts the parameters ranges selected for the seven geometrical
parameters. The study was planned using Box–Behnken design, with 64
combinations. A significant level of P < 0.05 (95% confidence) was used in
all tests. Analysis of variance (ANOVA) was followed by an F-test of the
individual factors and interactions.
Table 4.2: The values of the independent variables
Variables minimum center maximum
Inlet height, a/D =X1 0.4 0.55 0.7
Inlet width, b/D =X2 0.14 0.27 0.4
Vortex finder diameter, Dx/D =X3 0.2 0.475 0.75
Total cyclone height, Ht/D =X4 3.0 5.0 7.0
Cylinder height, h/D =X5 1.0 1.5 2.0
Vortex finder length, S/D =X6 0.4 1.2 2.0
cone-tip diameter, Bc/D =X7 0.2 0.3 0.4
47
Chapter 4. Sensitivity Analysis of Geometrical Parameters
Fitting the model
The analysis of variance (ANOVA) shows that the resultant quadratic poly-
nomial models adequately represented the input data with the coefficient
of multiple determination R2 being 0.92848. This indicates that the ob-
tained quadratic polynomial model was adequate to describe the influence
of the independent variables studied [189]. Analysis of variance (ANOVA)
was used to evaluate the significance of the coefficients of the quadratic
polynomial models (see Table 4.3). For any of the terms in the models, a
large F-value (small P-value) would indicate a more significant effect on
the respective response variables.
Based on the ANOVA results presented in Table 4.3, the variable with the
largest effect on the pressure drop (Euler number) was the linear term of
vortex finder diameter, followed by the linear term of inlet width and inlet
height (P < 0.05); the other four linear terms (barrel height, vortex finder
length, cyclone total height and cone-tip diameter) did not show a signif-
icant effect (P > 0.05). The quadratic term of vortex finder diameter also
had a significant effect (P < 0.05) on the pressure drop; however, the effect
of the other six quadratic terms was insignificant (P > 0.05). Furthermore,
the interaction between the inlet dimensions and vortex finder diameters
(P < 0.05) also had a significant effect on the pressure drop, while the effect
of the remaining terms was insignificant (P > 0.05).
4.1.3 Analysis of response surfaces
For visualization of the calculated factor, main effects plot, Pareto chart
and response surface plots were drawn. The slope of the main effect curve
is proportional to the size of the effect, and the direction of the curve spec-
ifies a positive or negative influence of the effect [61](Fig.4.2(a)). Based on
the main effect plot, the most significant factor on the Euler number are
(1) the vortex finder diameter, with a second-order curve with a wide range
of inverse relation and a narrow range of direct relation, (2) direct relation
with inlet dimensions, (3) inverse relation with cyclone total height and
insignificant effects for the other factors.
Pareto charts were used to summarize graphically and display the rela-
tive importance of each parameter with respect to the Euler number. The
Pareto chart shows all the linear and second-order effects of the param-
eters within the model and estimates the significance of each with re-
spect to maximizing the Euler number response. A Pareto chart displays
a frequency histogram with the length of each bar proportional to each
estimated standardized effect [30]. The vertical line on the Pareto chart
judges, whether each effect is statistically significant within the generated
48
4.1. Sensitivity analysis
response surface model; bars that extend beyond this line represent ef-
fects that are statistically significant at a 95% confidence level. Based on
the Pareto chart (Fig. 4.2(b)) and ANOVA table (Table 4.3) there are four
significant parameters (six terms in the ANOVA table ) at a 95% confidence
level: the negative linear vortex finder diameter; the linear inlet width; the
linear total cyclone height; a second-order vortex finder diameter; negative
interaction between vortex finder diameter and inlet dimensions. These
are the major terms in a polynomial fit to the data. Therefore, the pareto
chart is a perfect supplementation to the main effects plot.
To visualize the effect of the independent variables on the dependent ones,
surface response of the quadratic polynomial models were generated by
varying two of the independent variables within the experimental range
while holding the other factors at their central values [189]. Thus, Fig. 4.2(c)
was generated by varying the inlet height and the inlet width while hold-
ing the other five factors fixed at their central value. The trend of the curve
is linear, with more significant effect for inlet width, with no interaction
between the inlet height and width. The response surface plots given by
Figs. 4.2(d), 4.2(e) and 4.2(f) show that there are interactions between both
inlet width and inlet height with the vortex finder diameter. The effect of
cyclone total height is less significant with respect to the vortex finder di-
ameter, but its effect is higher than that of the vortex finder length, the
barrel height and the cone-tip diameter.
4.1.4 Conclusions
Mathematical modeling (the Muschelknautz method of modeling (MM))
has been used to understand the effect of the cyclone geometrical parame-
ters on the cyclone performance. The most significant geometrical param-
eters are:
1. the vortex finder diameter
2. the inlet section width
3. the inlet section height
4. the cyclone total height.
The effect of both the barrel height and the vortex finder length on the cy-
clone separator performance are small in comparison with these most sig-
nificant geometrical parameters. There are strong interactions between
the effects of inlet dimensions and the vortex finder diameter on the cy-
clone performance.
This study confirms the insignificant effect of the cone-tip diameter on the
cyclone performance. However, the discrepancy exists in literature for this
49
Chapter 4. Sensitivity Analysis of Geometrical Parameters
Table 4.3: Analysis of variance and the regression coefficients∗
Variable Regression coefficient F-Ratio P-Value
β0 -43.0742
Linearβ1 178.176 8.11 0.0075β2 372.26 19.79 0.0001β3 -161.452 232.04 0.0000β4 -1.55344 0.6 0.446β5 8.5875 0 0.9691β6 -7.23112 0.1 0.757β7 19.5663 0 0.9537
quadraticβ11 1.08238 0 0.9931β22 -12.2111 0 0.9446β33 403.419 107.8 0.0000β44 -0.223597 0.09 0.7641β55 -2.67108 0.05 0.8223β66 1.81257 0.15 0.6994β77 -62.1739 0.04 0.8364
Interactionβ12 91.0488 0.22 0.6427β13 -355.892 14.75 0.0005β14 0.459314 0 0.9726β15 -3.27883 0 0.9514β16 2.19997 0 0.9465β17 26.2787 0.01 0.9191β23 -720.685 42.42 0.0000β24 1.03571 0 0.9467β25 -2.53478 0 0.9675β26 4.2616 0.01 0.9112β27 -5.28466 0 0.9862β34 5.2034 0.51 0.4799β35 2.77536 0.01 0.9249β36 0.985086 0 0.9568β37 32.579 0.05 0.8221β45 -0.0452174 0 0.9911β46 0.345301 0.02 0.8902β47 -1.5016 0.01 0.9404β56 -0.422227 0 0.9667β57 3.82354 0 0.9622β67 -6.40134 0.02 0.8945
R2 0.92848
∗ Bold numbers indicate significant factors as identified by the analysis of variance (ANOVA) at the 95%confidence level.
50
4.1. Sensitivity analysis
(a) Main effect plot (b) Pareto chart
(c) X1 versus X2 (d) X1 versus X3
(e) X2 versus X3 (f) X3 versus X4
Figure 4.2: Analysis of design of experiment (cf. Table 4.2)
51
Chapter 4. Sensitivity Analysis of Geometrical Parameters
geometrical factor motivates us to study this individual parameter in more
details using large eddy simulation methodology as given in Sec. 4.2. Be-
fore proceed to apply CFD technique in the optimization of the cyclone ge-
ometry and to study the effect of each significant factor factor, a question
appears. Should the dust outlet geometry be included in the simulation
domain? The answer of this question is given in Sec. 4.3.
4.2 The cone-tip diameter
Until now, a considerable number of investigations has been performed ei-
ther on small sampling cyclones or larger industrial cyclone separators, for
example; Buttner [19], Iozia and Leith [84], Kim and Lee [95], Zhu et al.
[199], Elsayed and Lacor [50, 52, 53] and Safikhani et al. [147]. In these
studies, almost all of the cyclone dimensions listed in Table 4.4, were var-
ied and the changes in cyclone performance characteristics brought about
by these variations were studied. However, very little information is avail-
able on the effect of changing the cone bottom (tip) diameter (which deter-
mines the cone shape if other cyclone dimensions are fixed [184]) on the
flow pattern and performance. Regarding this effect, discrepancies and
uncertainties exist in the literature. Bryant et al. [17] observed that if the
vortex touched the cone wall, particle re-entrainment occurred and effi-
ciency decreased, so the collection efficiency will be lower for cyclones with
a small cone opening (cone-tip diameter). However, according to Stern et
al. [171] (cited in Xiang et al. [184]), a cone is not an essential part for
cyclone operation, whereas it serves the practical purpose of delivering
collected particles to the central discharge point. However, Zhu and Lee
[200] stated that the cone provides greater tangential velocities near the
bottom for removing smaller particles. Furthermore, the sensitivity anal-
ysis presented in Sec. 4.1 indicates the insignificant effect of the cone-tip
diameter on the cyclone performance.
However, the understanding and knowledge of the flow field inside a cy-
clone has been developed rapidly over the last few years, the exact mecha-
nisms of removing particles are still not fully understood. Therefore, most
existing cyclone theories are based on simplified models or depend upon
empirical correlations [23]. Xiang et al. [184] carried out experiments with
cyclones of different cone dimensions and evaluated a few models, namely
Barth [9], Leith and Licht [99] and Iozia and Leith [85]. All these models
could simulate correctly the trend of Xiang’s experimental data. However,
the quantitative agreement was not satisfactory. CFD has a great poten-
tial to predict the flow field characteristics and particle trajectories as well
as the pressure drop inside the cyclone [68]. Chuah et al.[23] carried out a
numerical investigation on the same cyclone dimensions used by Xiang et
52
4.2. The cone-tip diameter
al. [184] with the commercial finite volume code Fluent. Using different
turbulence models they proved that Fluent with Reynolds stress turbu-
lence model (RSM) predicts well the cyclone collection efficiency and the
pressure drop. The CFD simulation results from Chuah et al. [23] agree
well with Xiang’s experimental results in that cyclones with a smaller cone
diameter result in a slightly higher collection efficiency compared to cy-
clones with a bigger cone-tip diameter (only if the cone-tip diameter is not
smaller than the gas exit tube diameter). Moreover, the change in the
pressure drop will not be significant when the cone size is varied. Both
Xiang and Chuah did not give any results about the effect of the cone-tip
diameter on the flow field inside the cyclone separator, except some plots
for axial and tangential velocity profiles at two stations in the flow field for
Chuah et al. [23]. Xiang and Lee [185] computationally investigated the
effect of the cone-tip diameter on the flow field using the Reynolds stress
turbulence model. They did not present any contour plots for either the
static pressure, tangential and axial velocity. However, the comparisons
between the tangential and axial velocity profiles at different sections in-
dicating no valuable difference between the three cyclones [185, Fig. 8, p.
216 and Fig. 9, p. 217 ], they mentioned that the cone-tip diameter has
a significant effect on the flow field. No particle tracking study has been
performed in the study of Xiang and Lee [185].
Currently a better understanding of the flow field inside cyclone separators
is an important concern, especially with the application of large eddy sim-
ulation (LES). The present study was undertaken in an effort to carry out
a numerical study on the effect of the cone-tip diameter on the flow field
and the cyclone performance using LES available in Fluent commercial
finite volume solver.
Table 4.4: The geometrical dimensions of the three cyclones§
DimensionLength Dimension ratio(mm) (dimension/D)
Body diameter, D 31 1Gas outlet diameter, Dx 15.5 0.5Inlet height, a 12.5 0.4Inlet width, b 5 0.16Cyclone height, Ht 77 2.5Cylinder height, h 31 1Gas outlet duct length, S 15.5 0.5
Cone-tip diameter, Bc
Cyclone I 19.4 0.625Cyclone II 15.5 0.5Cyclone III 11.6 0.375
§ The outlet section is above the cylindrical barrel surface by Le = 0.5D. The inlet section located at adistance Li = 0.75D from the cyclone center (cf. Fig. 4.3).
53
Chapter 4. Sensitivity Analysis of Geometrical Parameters
!
"
#
"
Sa
h
D
b
H t
D x
L i
L e
B c
Figure 4.3: Schematic diagram for the cyclone geometry and coordinate definition
4.2.1 Numerical simulation
4.2.1.1 Configuration of the three cyclones
The cyclones used in this study had a reversed flow tangential inlet. The
geometry and dimensions are shown in Fig. 8.1 and Table 4.4. Three cy-
clones with different cone-tip diameters are used viz., Bc/D= 0.625, 0.5
and 0.375. The three cyclones are identical to those used by both Xiang
et al. [184, 186] and Chuah et al. [23]. Four plotting sections are used to
investigate the effect of the cone-tip diameter Bc on the velocity profiles as
given by Table 4.5.
4.2.1.2 Selection of the turbulence model (RANS versus LES)
For the turbulent flow in cyclones, the key to the success of CFD lies with
the accurate description of the turbulent behavior of the flow [68]. To
model the swirling turbulent flow in a cyclone separator, there are dif-
ferent turbulence models available in Fluent. These range from the stan-
dard k − ε model to the more complicated Reynolds stress model (RSM)
Table 4.5: The position of different plotting sections
Section S1 S2 S3 S4
z (mm) 5 15 30 50z/D 0.16 0.48 0.97 1.61
54
4.2. The cone-tip diameter
and large eddy simulation (LES) methodology as an alternative for RANS
models. The standard k−ε, RNG k−ε and Realizable k−ε models were not
optimized for strongly swirling flows found in cyclones [23]. The Reynolds
stress turbulence model (RSM) requires the solution of transport equations
for each of the Reynolds stress components and yields an accurate predic-
tion on swirl flow pattern, axial velocity, tangential velocity and pressure
drop on cyclone simulation [159].
Large eddy simulation (LES) has been widely accepted as a promising nu-
merical tool for solving the large-scale unsteady behavior of complex tur-
bulent flows. Encouraging results have been reported in recent literature
and demonstrate the ability of LES to capture the swirling flow instability
and the energy containing coherent motion of such highly swirling flows
[41]. LES methodology has been used in many articles to study the highly
swirling flow in cyclone separators, [e.g., 38, 39, 48, 155, 156, 159, 191].
It will be used in this study to reveal the effect of changing the cone-tip
diameter on the turbulent flow in the cyclone separator. The simulation
will start with the Reynolds stress turbulence model for flow initialization,
then the large eddy simulation methodology will be applied.
4.2.1.3 Solver settings
Selection of the discretization schemes
The choice of the discretization schemes has a tremendous influence on the
simulation results and the Fluent solver offers many different schemes
for pressure-velocity coupling, pressure, momentum, kinetic energy, rate
of kinetic energy dissipation discretization [59]. Both Kaya and Karagoz
[91] and Shukla et al. [158] investigated the performance of different dis-
cretization schemes in the steady and unsteady simulation of cyclone sep-
arators. The schemes used in this study are given in the following para-
graphs together with an explanation of the reasons behind their selection.
Kaya and Karagoz [91] reported the advantages of SIMPLEC (semi im-
plicit method for pressure linked equations consistent) scheme for pressure-
velocity coupling in terms of convergence. For the pressure discretization
they stated that only the PRESTO (pressure staggering option) pressure
interpolation scheme only can predict precisely the mean velocity profiles
static pressure distribution and the pressure drop in the cyclone separator
with good agreement with the experimental values. This scheme is also
recommended by the Fluent manual [59] for highly swirling flows.
For momentum, the QUICK (quadratic upwind differencing) scheme has
been recommended by both Kaya and Karagoz [91] and Fluent manual
[59] for the flow in cyclone separators. For the discretization of kinetic
55
Chapter 4. Sensitivity Analysis of Geometrical Parameters
energy and its dissipation rate equation, the second-order upwind scheme
has been used [158]. The first-order upwind scheme has been used for
the discretization of the Reynolds stress equations [91, 158]. For the LES
simulations the bounded central difference scheme is the default and the
recommended convection scheme by the Fluent manual [59].
The time step
The unsteady simulation started with the Reynolds stress model for ini-
tialization with a time step of 1E-4s using implicit coupled solution al-
gorithm. For the LES simulation, the time step is 1E-5 s. The selected
time step results in an average inlet Courant number of 2.3 for the three
cyclones. But as the solver is a segregated implicit solver, there is no sta-
bility criterion that needs to be met in determining the time step (and
consequently the Courant number) [59, 137]. However, to model transient
phenomena properly the Fluent manual [59] suggested using a time step
of at least one order of magnitude smaller than the smallest time constant
in the system. In the cyclone separator studies, the average residence time
(cyclone volume/ gas volume flow rate) is widely used to estimate the time
step [23, 43, 52]. In this study, the cyclone volumes (calculated by the Flu-
ent solver) are 5.045387E-5 m3, 4.738813E-5 m3 and 4.468306E-5 m3 for
cyclones I-III respectively. For a flow rate of 30E-3 m3/s, the corresponding
average residence time values are 0.00168s, 0.00158s and 0.001489s for
cyclones I-III respectively, i.e., the used time step is just a small fraction
of the average residence time. This confirms that the used time step can
reveal the transient phenomena properly. However, the interest in this
study lies in the simulation of averaged scalars and vectors (the average
velocities and pressures in order to estimate the cyclone performance). To
resolve the high-frequency phenomena in the time domain may be a time
step smaller by two orders of magnitude (tiny fraction) than the average
residence time would be required which is not the case in this study. Fur-
thermore, to verify that the choice for the time step was appropriate after
the calculation is complete; Fluent manual [59] suggests to check the max-
imum value of Courant number at the most sensitive transient regions of
the domain (in this study, it is the central region) should not exceed a value
of 20-40 [59]. For the three cyclones the maximum values of Courant num-
ber are 9.78, 12.1, 16.92 for cyclones I-III respectively. This verifies again
that the choice of the time step was proper.
Convergence criteria
With regard to the convergence criteria, two aspects should be considered.
Firstly, the scaled residuals should be below 1E-5 (The default convergence
56
4.2. The cone-tip diameter
criterion of Fluent is that scaled residuals of all equations fall below 1E-
3). Secondly, some representative quantities such as velocity and pressure
should be monitored until they are constant [137]. Although the present
simulations were converged at about (t=1.5-1.6 s), they were only termi-
nated at t=2s to get more accurate time averaged values. After achieving
constant tangential velocity with time at a certain point in the middle of
the cyclone domain, Fluent begins the data sampling for time statistics for
the whole domain (the velocity components and the static pressure) each
time step for sufficiently long time (t=1.5s until t=2.0s). From this step,
the following time averaged values are available; the mean and the root
mean square values of the static pressure p, the velocity magnitude v, the
x-velocity vx, the y-velocity vy and the z-velocity vz. The mean z-velocity vz
is identical to the mean axial velocity vaxial. The mean tangential velocity
vθ can be obtained using Fluent custom field function [cf. 59] according to
Eq. (4.2).
vθ = vx cos θ + vy sin θ (4.2)
where vx is the time averaged x-velocity, vy is the time averaged y-velocity
and θ is the angular coordinate.
Boundary conditions and other settings
Velocity inlet boundary condition is applied at inlet, outflow at gas outlet
and wall (no-slip) boundary condition at all other boundaries. The air in-
let velocity Uin equals 8 m/s, corresponding to air inlet volume flow rate
Qin=30 l/s, air density 1.0 kg/m3 and dynamic viscosity of 2.11E-5 Pa s,
leading to a Reynolds number of 1.18E4 based on the cyclone diameter and
the area averaged inlet velocity. The turbulence intensity I equals 5% and
the turbulence characteristic length equals 0.07 times the inlet width [74].
At the cyclone inlet, the Reynolds stress specific method in Fluent solver is
the Reynolds stress components. The diagonal components of the Reynolds
stress tensor (normal stresses) are assigned to 2kin/3, kin = 32
(I U2
in
),
where kin is the kinetic energy at the inlet [59, 81]. The shear stresses
(non-diagonal components) at the inlet are set to zero. To take into ac-
count the stochastic component of the turbulent flow at the inlet for the
LES simulation, artificial perturbations have been generated using the
spectral synthesizer method available in the Fluent solver [59, 82, 162],
where the fluctuation velocity components are computed by synthesizing
a divergence-free velocity-vector field from the summation of 100 Fourier
harmonics [59]. The fluctuations are added to the mean inlet velocity.
The reason for introducing these artificial perturbations instead of select-
ing the no-perturbation option in the Fluent solver, is that the unpertur-
57
Chapter 4. Sensitivity Analysis of Geometrical Parameters
bated flat turbulent profiles at inlet generates unrealistic turbulent eddies
[59]. For the near-wall treatment, the enhanced wall function [59] has
been used in the RSM simulation. For subgrid scale model, the dynamic
Smagorinsky-Lilly model [59, 60, 101] has been used. The Smagorinsky
model constant is dynamically computed instead of given as an input to
the solver, but clipped to zero or 0.23 if the calculated model constant is
outside this range to avoid numerical instabilities [59]. The second advan-
tage of the dynamic Smagorinsky-Lilly model over the Smagorinsky model
is the treatment near the wall. In the dynamic Smagorinsky-Lilly model, a
damping function for the eddy viscosity near the wall is not required, since
the model constant goes to zero in the laminar region just near the wall
[113].
The boundary condition at the gas outlet is the outflow boundary condition
provided by Fluent [59], where all transport variables have a zero normal
gradient. This boundary condition is valid for fully developed flow, and
this is the reason why the vortex finder is extended 1/2 cyclone diameter
above the top of the cyclone in the present study to allow the exit flow
to be a fully developed flow [185]. The effect of this distance on the flow
field has been the subject of different investigations [e.g., 153, 178], where
this distance was varied between zero and four cyclone diameters. Wang
et al. [178] investigated the effect of outflow length on the flow field and
velocity profile and suggested to have the gas outlet boundary condition at
a distance longer than the cyclone radius.
The grid independency study
The grid independence study has been performed for the three tested cy-
clones. Three levels of grid for each cyclone have been tested, to be sure
that the obtained results are grid independent. The hexahedral computa-
tional grids were generated using GAMBIT grid generator and the simula-
tions were performed using Fluent 6.3.26 commercial finite volume solver
on a 8 nodes CPU Opteron 64 Linux cluster.
The computational results of the three grid types are presented in Ta-
ble 4.6. As seen the maximum difference between the results obtained
from the fine and medium meshes is 1% for the calculation of the cut-off
diameter and the Euler number which is in the range of experimental error
[52, 137]. It has been observed that even medium grids provide a sufficient
grid independency. However, for excluding any uncertainty, computations
have been performed using the fine grid, where the total number of grid
points was not that critical with respect to the computation overhead [10].
Consequently, the used grid produces grid independent results (the author
only checked the mean values, so for future studies with unsteady phe-
58
4.2. The cone-tip diameter
Table 4.6: The details of the grid independence study for cyclones I-IIIa
Cyclone I II III
N Eu x50 N Eu x50 N Eu x50
Coarse 632153 2.48 1.396 513021 2.36 1.335 513991 2.869 1.24Medium 861077 2.405 1.355 863852 2.28 1.257 712576 2.712 1.21Fine 1021616 2.39 1.35 1025778 2.27 1.25 1027982 2.687 1.2
% differenceb 3.76 3.4 3.96 6.8 6.77 3.33
% differencec 0.63 0.37 0.44 0.56 0.93 0.83
a N is the number of hexahedral cells, Eu is the Euler number (dimensionless pressure drop = pressuredrop / average kinetic energy at inlet) and x50 is the cut-off diameter; the particle diameter that willproduce 50% collection efficiency (cf. Sec. 3.3).b The percentage absolute difference between the coarse and fine grid values for Euler number andcut-off diameter.c The percentage absolute difference between the medium and fine grid values for Euler number andcut-off diameter.
nomena like vortex-core precession (cf. Ref. [38]), the effect of the grid on
the Strouhal number associated with the simulated vortex-core precession
[38] should be included in the grid independency study probably requiring
finer grids, but this is not part of the present study). Moreover, to evaluate
accurately the numerical uncertainties in the computational results (espe-
cially because of the large difference between the results obtained on the
coarse and the fine mesh which is about 7%), the concept of grid conver-
gence index (GCI) was adopted using three grid levels per cyclone.
Grid convergence index (GCI)
Roache [143–145] suggested a quantitative measure for the grid conver-
gence; the grid convergence index (GCI). The GCI can be computed using
two levels of grid; however, three levels are recommended in order to es-
timate accurately the order of convergence and check that the solution is
within the asymptotic range of convergence [160]. For a consistent numer-
ical analysis the discretized equations will approach the solution of the
actual equations as the grid resolution approaches zero [160]. The appro-
priate level of grid resolution is a significant issue in numerical investiga-
tions. It is a function of many variables including the flow condition, type
of analysis, geometry and many other variables.
The GCI is based upon a grid refinement error estimator derived from
the theory of the generalized Richardson extrapolation [160]. The GCI
is a measure of how far the computed value is away from the value of
the asymptotic numerical value. Consequently, it indicates how much the
solution would change with a further refinement of the grid. A small value
of GCI indicates that the computation is within the asymptotic range.
59
Chapter 4. Sensitivity Analysis of Geometrical Parameters
The GCI on the fine grid is defined as:
GCIfine =
Fs|ε|(rp − 1)
(4.3)
where Fs is a factor of safety. Fs = 3 for comparison of two grids and 1.25
for comparison over three grids or more.
For the coarse grid:
GCIcoarse =
Fs|ε|rp(rp − 1)
(4.4)
ε is a relative error measure of the key variable f between the coarse and
fine solutions,
ε =f2 − f1
f1(4.5)
where f2 is the coarse-grid numerical solution obtained with grid spacing
h2. f1 is the fine-grid numerical solution obtained with grid spacing h1. ris the grid refinement ratio (r = h2/h1 > 1). For complicated geometries
r is replaced by the ratio of the number of control volumes in the fine and
coarse mesh [107] which is the case in this study,
r12 =
(N1
N2
) 1D
(4.6)
where D = 2 and 3 for two-dimensional and three-dimensional geometries
respectively [143, pp. 410]. N1 and N2 are the number of control volumes
in the fine and coarse mesh respectively.
p is the order of the discretization method. p equals two if the second order
discretization is used for all terms in space [107] (However, Slater [160]
stated that if all discretization in space was of second-order, p will be less
than 2. The difference is due to grid stretching, grid quality, non linearity
in the solution, presence of shocks, turbulence modeling and perhaps other
factors). For the grid refinement study, three meshes have been used with
N1, N2 and N3 cells for the fine, medium and coarse three-dimensional
mesh respectively.
r12 =(
N1
N2
) 13
, r23 =(
N2
N3
) 13
, e12 = f2 − f1, e23 = f3 − f2, where ei,i+1 =
fi+1 − fi is the difference in the key variable f resulting from the use of
different grids. If r12 = r23 then,
p = ln
(e23e12
)/ln(r) (4.7)
60
4.2. The cone-tip diameter
If r12 6= r23 which is the case in this study, Roache [145] proposed to solve
Eq. 4.8
e23(rp23 − 1)
= rp12
[e12
(rp12 − 1)
](4.8)
Equation 4.8 is transcendental in p. Using the iterative technique with
relaxation factor introduced in Roache [144, 145]
p = ω ρ+ (1− ω)ln(β)
ln(r12)(4.9)
where β =(rp12−1)e23(rp23−1)e12
, ω = 0.5 and ρ is the previous iteration of p. The
author suggest to use ρ = ln(e23e12
)/ln(r12) as a first guess. The iteration
will stop if |p−ρp | < 1E − 5.
Now one can calculate, ε12 = f2−f1f1
, ε23 = f3−f2f2
, GCIfine12 = 1.25|ε12|
(rp12−1)and
GCIfine23 = 1.25|ε23|
(rp23−1). GCI
fine12 should be smaller than GCI
fine23
To check if the solution is in the asymptotic range, α ≈ 1 (cf. Eq. 4.10)
α =rp12GCI
fine12
GCIfine23
(4.10)
The Richardson extrapolation can be used to obtain the value of f when
the grid spacing h vanishes (h → 0) [2, 143].
fexact = f1 + (f1 − f2)/ (rp12 − 1) (4.11)
Table 4.7 presents the grid convergency calculations using GCI method
and three grid levels for cyclones I -III. The following conclusions have
been obtained from the GCI analysis:
• The results are in the asymptotic range for the three cyclones, be-
cause the obtained values for α are close to unity.
• The ratio R is less than unity this means monotonic convergence [2]
(Ali et al. [2] classified the possible convergence conditions into three
groups, namely (1) monotonic convergence; 0 < R < 1 (2) oscillatory
convergence; R < 0 (3) divergence; R > 1.) .
• There is a reduction in the GCI value for the successive grid refine-
ments (GCIfine12 < GCI
fine23 ) for the two variables (Eu and x50). This in-
dicates that the dependency of the numerical results on the cell size
has been reduced. Moreover, a grid independent solution has been
achieved. Further refinement of the grid will not give much change
61
Chapter 4. Sensitivity Analysis of Geometrical Parameters
in the simulation results. For the two variables (Eu and x50), the ex-
trapolated value is only slightly lower than the finest grid solution.
Therefore, the solution has converged with the refinement from the
coarser grid to the finer grid [2]. Figure 4.4 presents a qualitative
proof that the obtained results are in the asymptotic range, i.e., the
obtained results are mesh independent.
Table 4.7: Grid convergency calculations using GCI method and three grid levels
for cyclones I - III
i Ni fi ri,i+1 ei,i+1 εi,i+1 GCIfinei,i+1% Ra αb
I
Eu
0c 2.37501 1021616 2.3900
1.0586 0.0150 0.0063 0.78492 861077 2.4050 0.2013 1.0063
1.1085 0.0750 0.0312 1.55973 632153 2.4800
x50
0 1.34711 1021616 1.3500
1.0586 0.0050 0.0037 0.26492 861077 1.3550 0.1224 1.0037
1.1085 0.0410 0.0303 0.72523 632153 1.3960
II
Eu
0 2.25231 1025778 2.2700
1.0589 0.0100 0.0044 0.97542 863852 2.2800 0.1256 1.0044
1.1897 0.0800 0.0351 1.51943 513021 2.3600
x50
0 1.24131 1025778 1.2500
1.0589 0.0070 0.0056 0.87402 863852 1.2570 0.0902 1.0056
1.1897 0.0780 0.0621 1.56523 513021 1.3350
III
Eu
0 2.68351 1027982 2.6870
1.1299 0.0250 0.0093 0.16192 712576 2.7120 0.1607 1.0093
1.1150 0.1570 0.0579 1.31273 513991 2.8690
x50
0 1.19621 1027982 1.2000
1.1299 0.0100 0.0083 0.39132 712576 1.2100 0.3361 1.0083
1.1150 0.0300 0.0248 1.42113 513991 1.2400
a R=ε12/ε23.b α=
(
rp12 GCI12)
/GCI23.c The value at zero grid space (h → 0). i=1, 2 and 3 denote the calculations at the fine, medium andcoarse mesh respectively.
62
4.2. The cone-tip diameter
N -1
Eul
ernu
mbe
r
Cut
-off
diam
eter
0 5E-07 1E-06 1.5E-062
2.2
2.4
2.6
2.8
3
1.15
1.2
1.25
1.3
1.35
1.4
1.45
1.5Cyclone I (Eu)Cyclone I (X50)Cyclone II (Eu)Cyclone II (X50)Cyclone III (Eu)Cyclone III (X50)
(h --> 0)
Figure 4.4: Qualitative representation of the grid independency study. The Euler
number and the cut-off diameter for each cyclone at the three grid levels. N−1 is
the reciprocal of the number of cells, h → 0 means the value at zero grid size (cf.
Table 4.7). To obtain a smooth curve; the spline curve fitting has been applied in
Tecplot post-processing software.
4.2.2 Results and discussion
Validation of results
The obtained numerical results are compared with the LDA velocity mea-
surements of Hoekstra [74] measured using laser doppler anemometry
(LDA) system. Figure 4.5 shows the comparisons between the LES simula-
tion and the measured axial and tangential velocity profiles at axial station
Z=94.25 cm from the cyclone bottom [74]. The LES simulation predicts
a similar trend as observed experimentally. The non exact matching be-
tween experimental and LES simulation has been reported in some other
literatures [e.g., 38]. Considering the complexity of the turbulent swirling
flow in the cyclones, the agreement between the simulations and measure-
ments is considered to be quite acceptable. Another comparison between
the current LES results and the Reynolds stress turbulence model (RSM)
results of Xiang and Lee [186] for cyclone III is given in Fig. 4.6 which in-
dicates LES can also depict the main flow features of cyclonic flow as the
Reynolds stress turbulence model can do.
63
Chapter 4. Sensitivity Analysis of Geometrical Parameters
Radial position / Cyclone radius
Tan
gent
ialv
eloc
ity/I
nlet
velo
city
-0.5 0 0.50
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
LDASimulation
Radial position / Cyclone radiusA
xial
velo
city
/Inl
etve
loci
ty-0.5 0 0.5
-0.4
-0.2
0
0.2
0.4
0.6
LDASimulation
Figure 4.5: Comparison of the time averaged axial and tangential velocity between
LDA measurements [74] and the LES simulations at section Z=94.25 cm from the
cyclone bottom (cf. Hoekstra [74] for more details about the used cyclone geometry
in this validation).
4.2.3 The flow pattern in the three cyclones
The pressure field
Figure 4.7 shows the time-averaged static pressure contours plots. In the
three cyclones, the static pressure decreases radially from the wall to the
center. A negative pressure zone appears in the forced vortex region (cen-
tral region) due to high swirling velocity. The pressure gradient is largest
along the radial direction, while the gradient in the axial direction is very
limited. The cyclonic flow is not symmetrical as is clear from the shape of
the low-pressure zone at the cyclone center (twisted cylinder). The static
pressure contour plots for the three cyclones are almost the same.
The velocity field
Based on the contours plots of the time averaged tangential velocity (Fig. 4.7)
and the radial profiles at sections S1, S2, S3 and S4 shown in Fig. 4.8 (cf.
Table 4.5), the following comments can be drawn. The maximum tangen-
tial velocity equals around 1.25 times the average inlet velocity and occurs
in the annulus cylindrical part. The tangential velocity distribution for
the three cyclones are nearly identical at the corresponding sections. The
tangential velocity profile at any section is composed from two regions, in-
ner and outer. In the inner region the flow rotates approximately like a
solid body rotation (forced vortex), where the tangential velocity increases
with radius. After reaching its peak the tangential velocity decreases with
radius in the outer part of the profile (free vortex). This profile is so-called
64
4.2. The cone-tip diameter
Dimensionless distance
Dim
en
sio
nle
ss
tan
ge
nti
al
ve
locit
y
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 10
0.25
0.5
0.75
1
1.25Present studyXiang and Lee (2005)
Dimensionless distance
Dim
en
sio
nle
ss
axia
lve
locit
y
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1-0.5
-0.25
0
0.25
0.5Present studyXiang and Lee (2005)
Figure 4.6: Comparison of the time averaged tangential and axial velocity between
RSM results of Xiang and Lee [186] and the current LES results at Z/D=1.29 for
Cyclone III (Bc/D = 0.375). Note: the dimensionless radial distance = the distance
/ the cyclone radius, the dimensionless velocity = the velocity / inlet velocity.
Rankine type vortex which include a quasi-forced vortex in the central re-
gion and a quasi free vortex in the outer region.
The radial profiles given in Fig. 4.8 represent the time averaged tangential
velocity in the lower part of the cyclone. The tangential velocity distribu-
tions at the bottom sections show good axis-symmetrical distribution. The
tangential velocity profiles for the three cyclones are identical in the inner
region, where the maximum tangential velocity nearly equals the inlet ve-
locity and occurs at a position 0.25 - 0.45 of the cyclone radius as given in
Table 4.8.
The axial velocity contours (Fig. 4.7) indicate the existence of two flow
streams. Downward flow directed to the cyclone bottom (negative axial
velocity), and upward flow directed to the vortex finder exit. The axial ve-
locity plots for the three cyclones are nearly identical to the corresponding
sections in the conical part. The axial velocity equals zero at the walls and
maximum close to the position of maximum tangential velocity. The axial
velocity profiles shown in Fig. 4.8 exhibit a severe asymmetrical feature.
Table 4.8: Comparison between the maximum tangential velocity value and its
position at different sections.
Section S1 S2 S3 S4
Cyclone I II III I II III I II III I II III
vθmax/ v∗in 1.08 0.97 1.02 1.02 0.95 1.05 0.96 0.89 1.04 0.94 0.88 1.035
x/R† 0.3 0.28 0.26 0.3 0.33 0.33 0.32 0.37 0.41 0.35 0.43 0.425
∗ The ratio between the maximum tangential velocity and the area average inlet velocity.† The dimensionless distance between the centerline and the point of maximum velocity, R is the cycloneradius.
65
Chapter 4. Sensitivity Analysis of Geometrical Parameters
Comparison of the velocity profiles in the three cyclones
However, from the previous discussions it is clear that, the effect of the
cone-tip diameter on the flow field in the conical section is insignificant, in
comparison with other geometrical parameters such as the vortex finder
diameter. Nevertheless, in this section a comparison between the axial and
tangential velocity profiles at four sections (Table 4.5) will be analyzed as
presented in Fig. 4.9.
The tangential velocity profiles in the forced vortex region are nearly iden-
tical in the three cyclones at each sections. The tangential velocity in the
free vortex region increases as the cone-tip diameter is reduced. The tan-
gential velocity profiles for the three cyclones are almost the same.
The axial velocity profile has the shape of an inverted W for all cyclones.
The highest axial velocity occurs at 0.25 - 0.5 of the cyclone radius down
the vortex finder until the cyclone bottom, and between 0.25 and 0.5 of the
cyclone radius in the annulus and through the vortex finder. No consid-
erable difference exists in the axial velocity profiles for the three cyclones.
Since the axial velocity profiles is almost the same for the three cyclones,
the average residence time of particles is nearly the same. Furthermore,
the position of the highest axial velocity moves inward in the conical part
as the cone-tip diameter reduced.
From the previous analysis, the region of downward flow is nearly the
same, for the three cyclones, while the tangential velocity slightly increases
as the cone-tip diameter reduced, so the particles will experience a higher
tangential velocity for cyclone III than in other cyclones for the same time
(as the region of downward axial velocity is nearly equal). This results in
a slightly higher collection efficiency. This is consistent with the measured
results reported by Xiang et al. [184] and simulation by Xiang and Lee
[186] and Chuah et al. [23]. The change of the cone-tip diameter affects
the flow field in the cyclone separator but this change is so limited, i.e.,
the reduction of cone-tip diameter enhances the collection efficiency but
with a small percentage, as the flow field pattern is so closed for the three
cyclones.
The DPM results
In order to calculate the effect of the cone-tip diameter on the cut-off diam-
eter, 104 particles were injected from the inlet surface with particle veloc-
ity equals the gas inlet velocity. The particle density is 860 kg/m3 and the
maximum number of time steps for each injection was 9E5 steps. The DPM
analysis results for the three cyclones are shown in Table 4.9 and Fig. 4.10.
It is found that the cut-off diameter decreases slightly with decreasing the
66
4.2. The cone-tip diameter
Figure 4.7: The contours plots for the time averaged flow variables at Y=0 . From
top to bottom: the static pressure N/m2, the tangential velocity and axial velocity
m/s. From left to right cyclone I, II and III respectively.
67
Chapter 4. Sensitivity Analysis of Geometrical Parameters
Dimensionless distance
Dim
en
sio
nle
ss
tan
ge
nti
al
ve
locit
y
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 10
0.25
0.5
0.75
1
1.25S1S2S3S4
Dimensionless distance
Dim
en
sio
nle
ss
axia
lve
locit
y-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1
-0.5
-0.25
0
0.25
0.5S1S2S3S4
Dimensionless distance
Dim
en
sio
nle
ss
tan
ge
nti
al
ve
locit
y
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 10
0.25
0.5
0.75
1
1.25S1S2S3S4
Dimensionless distance
Dim
en
sio
nle
ss
axia
lve
locit
y
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1-0.5
-0.25
0
0.25
0.5S1S2S3S4
Dimensionless distance
Dim
en
sio
nle
ss
tan
ge
nti
al
ve
locit
y
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 10
0.25
0.5
0.75
1
1.25S1S2S3S4
Dimensionless distance
Dim
en
sio
nle
ss
axia
lve
locit
y
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1-0.5
-0.25
0
0.25
0.5S1S2S3S4
Figure 4.8: The radial profile for the time averaged tangential and axial velocity
at different sections on the X-Z plane (Y=0) for each cyclone. From top to bottom:
Cyclone I, II and III respectively.
68
4.2. The cone-tip diameter
Dimensionless distance
Dim
en
sio
nle
ss
tan
ge
nti
al
ve
locit
y
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 10
0.25
0.5
0.75
1
1.25Cyclone ICyclone IICyclone III
Dimensionless distance
Dim
en
sio
nle
ss
axia
lve
locit
y
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1-0.5
-0.25
0
0.25
0.5Cyclone ICyclone IICyclone III
Dimensionless distance
Dim
en
sio
nle
ss
tan
ge
nti
al
ve
locit
y
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 10
0.25
0.5
0.75
1
1.25Cyclone ICyclone IICyclone III
Dimensionless distance
Dim
en
sio
nle
ss
axia
lve
locit
y
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1-0.5
-0.25
0
0.25
0.5Cyclone ICyclone IICyclone III
Dimensionless distance
Dim
en
sio
nle
ss
tan
ge
nti
al
ve
locit
y
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 10
0.25
0.5
0.75
1
1.25Cyclone ICyclone IICyclone III
Dimensionless distance
Dim
en
sio
nle
ss
axia
lve
locit
y
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1-0.5
-0.25
0
0.25
0.5Cyclone ICyclone IICyclone III
Dimensionless distance
Dim
en
sio
nle
ss
tan
ge
nti
al
ve
locit
y
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 10
0.25
0.5
0.75
1
1.25Cyclone ICyclone IICyclone III
Dimensionless distance
Dim
en
sio
nle
ss
axia
lve
locit
y
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1-0.25
0
0.25
0.5Cyclone ICyclone IICyclone III
Figure 4.9: Comparison between the radial profile for the time averaged tangential
and axial velocity at different sections on the X-Z plane (Y=0). From top to bottom:
section S4 - S1 respectively.
69
Chapter 4. Sensitivity Analysis of Geometrical Parameters
cone-tip diameter while the pressure drop is increasing slightly. Conse-
quently, the effect of the cone-tip diameter on the cyclone performance is
insignificant.
The trend of changing the cut-off diameter with the cone-tip diameter
given by Chuah et al. [23] (Qin = 60l/min) supports the conclusion of
the slightly decrease of the cut-off diameter by decreasing the cone-tip di-
ameter (insignificant effect), Table 4.9.
4.2.4 Comparison with mathematical models
Table 4.10 presents a comparison between the Euler number (dimension-
less pressure drop) and the cut-off diameter obtained from CFD, experi-
mental investigation [184] and seven mathematical models, viz. (the Barth
model [9], the Muschelknautz method of modeling (MM) [116, 174], the
Stairmand model [165], the Casal and Martnez-Benet model [21], the Shep-
herd and Lapple model [157], the Iozia and Leith model [85] and the Ritema
model [142] (cf. Sec. 2.2)
The Euler numbers obtained from the models of Shepherd and Lapple,
Casal and Martnez-Benet are constant, because these models do not in-
clude the effect of the cone-tip diameter Bc in their formulas. The three
other models (Barth, MM and Stairmand) indicate less effect on both the
Euler number and the cut-off diameter by changing the cone-tip diameter.
The models of Iozia and Leith in addition to that of Rietma indicate no
change in the cut-off diameter with changing the cone-tip diameter. The
results of mathematical models and the experimental investigation sup-
port the CFD results that the cone-tip diameter has an insignificant effect
on the cyclone separator performance.
Table 4.9: The cut-off diameter and pressure drop for the three cyclones
Cyclone I II III
Bc/D 0.625 0.5 0.375
Cut-off diameter [µm] 1.35 1.25 1.2
Cut-off diameter [µm] (Chuah et al. [23]∗) 1.65 1.45 1.1
Pressure drop [N/m2] 76.5 72.7 86
Euler number Eu 2.39 2.27 2.687
∗ Qin = 60 l/min.
70
4.2. The cone-tip diameter
Bc /D
Eul
ernu
mbe
r[-]
Cut
-off
diam
eter
[mic
ron]
0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.70
1
2
3
4
5
1
1.2
1.4
1.6
1.8
Euler number [-] (Exp.)
Euler number [-] (LES)
Cut-off diameter [micron] (LES)
Figure 4.10: The effect of cone-tip diameter on the pressure drop (Euler number)
and the cut-off diameter (with spline curve fitting to get a smooth curve).
4.2.5 Conclusions
Large eddy simulation has been used to study the effect of the cone-tip
diameter on the cyclone flow field and performance. Three cyclones with
different values of Bc/D viz. 0.625, 0.5 and 0.375 (at constant vortex finder
diameter Dx/D = 0.5) have been investigated. The following conclusion
can be drawn.
• The cone-tip diameter has an insignificant effect on the flow pattern
and performance.
• As the cone-tip diameter decreases, the maximum tangential velocity
increases slightly, while its position is almost the same.
Table 4.10: The cyclone performance parameters using CFD, Experimental [184]
and different seven mathematical models
Cyclone CFD Barth MM Stairmand∗ Sphered∗ Casal∗ Iozia† Ritema† Exp.§
I 2.39 6.94 4.88 6.68 4.1 4.07 - - 2.8
Euler number Eu [-] II 2.27 7.40 4.95 6.69 4.1 4.07 - - 2.8
III 2.687 7.43 4.95 6.69 4.1 4.07 - - 3.25
I 1.35 1.22 2.13 - - - 1.44 1.4 3.01
Cut-off diameter x50 [µm] II 1.25 1.22 2.089 - - - 1.44 1.4 2.60
III 1.2 1.28 2.089 - - - 1.44 1.4 2.36
∗ The mathematical model used for estimation of the pressure drop only.† The mathematical model used for estimation of the cut-off diameter only.§ Different particle density.
71
Chapter 4. Sensitivity Analysis of Geometrical Parameters
• The flow pattern and performance parameters of the three cyclones
are almost the same.
• Decreasing the cone-tip diameter increases the pressure drop slightly.
The reverse trend is obtained for the cut-off diameter.
• Seven mathematical models used for estimation of the effect of cone-
tip diameter on the cyclone performance, and all of them support the
CFD results for the insignificance of varying the cone-tip diameter on
the cyclone performance.
• However, the main finding of the current study is the insignificant
effect of the cone-tip diameter on the cyclone performance, in com-
parison to the other geometrical parameters like the vortex finder
diameter Dx or the inlet dimensions [cf. 50, 52], the cone is an essen-
tial part for cyclone operation. If the cone would be removed from the
cyclone, it will be only cylindrical part and that one will have another
geometry (not the conventional cyclone). Moreover, the particles are
collected over the cone wall and then moved to the cyclone bottom.
Consequently, a very low collection efficiency will be expected for a
cyclone separator without a cone.
4.3 The dust outlet geometry
Conventional cyclones always have a dustbin attached to the cone to collect
the separated solid particles. When a gas flow stream enters the dustbin
(closed at bottom), some of the flow will return the cone and carry some of
the separated particles. This phenomena is called “re-entrainment” and it
will affect the separation efficiency of the cyclone [138].
However, many works have been carried out to investigate the influence of
different geometrical parameters such as cyclone length, inlet and outlet
pipe geometries on the performance of cyclones [15, 62, 102, 140, 184],
there has been little work concerning the dust outlet geometries [40, 47,
78, 123].
Regarding this influence, discrepancies and uncertainties exist in the liter-
ature. Xiang and Lee [186] reported that the dustbin connected to the cy-
clone should be incorporated in the flow domain as it affects the results. On
the other hand, numerous studies were performed without dustbin [e.g.,
159, 178] with good matching with experimental results.
Obermair et al. [123] performed cyclone tests with five different dust outlet
geometries to find the influence of the dust outlet geometry on the sepa-
ration process. They showed that separation efficiency can be improved
significantly by changing the dust outlet geometry, and they reported that
further research is needed to clarify precise effects of dust outlet geometry.
72
4.3. The dust outlet geometry
The effect of a dipleg (a vertical tube between the cyclone and the dustbin)
was posed and investigated by several researchers [78, 92].
The previous studies on the effect of dust outlet geometry on the cyclonic
flow can be classified into the following categories:
1. Comparison between cyclones with dustbin and that with dustbin
plus dipleg [78, 123, 138]. Obermair et al. [123] investigated exper-
imentally the effect of different dust outlet geometries on the flow
pattern. Nevertheless, the given flow pattern was limited to the cy-
clone bottom. The effects of the dust outlet geometry on the velocity
profiles are not given.
2. Comparison between the cyclone with and without dipleg (prolonged
cyclones). Kaya and Karagoz [92] numerically investigated the flow
characteristics and particle collection efficiencies of conventional (with-
out dipleg) and prolonged cyclones.
3. Comparison between cyclone with and without dustbin [47]. Elsayed
and Lacor [47] numerically investigated two cyclones with and with-
out dustbin. They reported a negligible effect of the dustbin on the
performance. They reported further research is needed with the cy-
clone with dipleg and dust hopper (dustbin plus dipleg).
All studies above do not contain enough information about the effect of
the dust outlet geometry on the flow pattern, velocity profiles and sepa-
ration performance. Furthermore, there is no clear comparison between
the four cases: separation space only (cylinder on cone), cyclone with cylin-
drical shaped dustbin, cyclone with vertical tube (dipleg) and cyclone with
dustbin plus dipleg. The present numerical investigation aims at exam-
ining the influence of the dust outlet geometry on the flow pattern and
the cyclone performance using the Reynolds stress turbulence model. This
serves two purposes: (1) to demonstrate whether simulation of the separa-
tion space only without including the dustbin or the dipleg is sufficient to
estimate the cyclone performance. (2) to give information for the designer
about the effect of the configuration under the separation space (the dust-
bin or the dipleg) on the flow pattern and performance.
4.3.1 Numerical simulation
4.3.1.1 Configurations of the four tested cyclones
The numerical simulations were performed on four cyclones. Cyclone I has
only the cylinder on cone shape (separation space only), cyclone II has a
dustbin, cyclone III has a dipleg while cyclone IV has a dust hopper (dust-
bin plus dipleg). Figure 4.11 and Table 4.11 give the cyclones dimensions.
73
Chapter 4. Sensitivity Analysis of Geometrical Parameters
Table 4.12 gives more details for the used cyclones, including the number
of cells, cyclone volume and the flow residence time for each cyclone. Nine
sections are used to plot the velocity profiles as shown in Table 4.13.
(a) I (b) II (c) III (d) IV
(e) The surface mesh
Figure 4.11: Schematic diagrams and surface meshes for the four tested cyclone
separators
4.3.1.2 Solver settings
Based on the study of Kaya and Karagoz [91] for the best selection of nu-
merical schemes to be used with RSM model, the following discretization
schemes have been used. The PRESTO scheme has been used for the pres-
74
4.3. The dust outlet geometry
Table 4.11: The geometrical dimensions of the four tested cyclones∗
Dimension Length Dimension ratio
(m) (Dimension/D)
Body diameter, D 0.205 1
Inlet height, a 0.105 0.5
Inlet width, b 0.041 0.2
Gas outlet diameter, Dx 0.105 0.5
Gas outlet duct length, S 0.105 0.5
Cone-tip diameter, Bc 0.076875 0.375
Cylinder height, h 0.3075 1.5
Cyclone height, Ht 0.82 4
∗ The outlet section is above the cyclone surface by Le = 0.5D. The inlet section located at a distanceLi = 0.75D from the cyclone center, the height of the dustbin and the dipleg, LD = 2D.
sure interpolation, the SIMPLEC algorithm for pressure velocity coupling,
the QUICK scheme for momentum equations, the second-order upwind
for the turbulent kinetic energy and the first-order upwind discretization
scheme for the Reynolds stresses [52], cf. Sec. 4.2.1.3 for more details.
Chuah et al. [23] stated that the time step should be selected as a tiny
fraction of the residence time tres. From Table 4.12, the value of tres varies
between 0.237 and 0.399 s. Therefore a time step of 1E-4 is an accept-
able value for the current simulations [52]. The simulations have been
performed using FLUENT 6.3.26.
4.3.1.3 CFD grid and boundary conditions
The mesh sensitivity study has been performed for the four tested cyclones
with three levels for each cyclone, to be sure that the obtained results
are grid independent. For example, three different meshes with respec-
tively 130596, 260230 and 478980 cells have been used for cyclone I. The
computational results on the three grids are presented in Table 4.14. As
the maximum difference between the results is less than 5%, so the grid
template 130596 produces the grid independent results [52, 146]. It has
been observed that even 130596 grid provides a sufficient grid indepen-
Table 4.12: The details of the four tested cyclones
Cyclone I II III IV
Number of cellsa 260230 593125 322286 441062
Cyclone volume x102 [m3] 1.99 3.35 2.18 2.76
tres [s] b 0.237 0.399 0.259 0.328
a The total number of hexahedral cells after the mesh sensitivity study
b The average residence time, tres = V/Qin where V is the cyclone volume and Qin is the gas flow rate.
75
Chapter 4. Sensitivity Analysis of Geometrical Parameters
Table 4.13: The position of different plotting sections∗
Section S1 S2 S3 S4 S5 S6 S7 S8 S9
z`/D 2.75 2.5 2.25 2 1.75 1.5 1.25 1.0 0.75∗ z` is measured from the top of the inlet section (cf. Fig. 4.11(a)).
dency. However, for excluding any uncertainty, computations have been
performed using the 260230 cells grid, where the total number of grid
points was not that critical with respect to the computation overhead [10].
Figure 4.11(e) shows the surface mesh of the four cyclones. The hexahedral
computational grids were generated using the GAMBIT grid generator.
The boundary condition at the inlet section is the velocity inlet. An out-
flow boundary condition is used at the outlet. The no-slip (wall) boundary
condition is used at the other boundaries [52]. The air volume flow rate
Qin=0.08405 m3/s for all cyclones, air density 1.225 kg/m3 and dynamic
viscosity of 178.940E-6 Pa s. The turbulent intensity equals 5% and char-
acteristic length equals 0.07 times the inlet width [74].
DPM settings
A discrete phase modeling (DPM) study has been performed by injecting
104 particles from the inlet surface with a particle density of 860 kg/m3
and with a particle size ranging from 0.025 until 5 µm at a velocity equals
to the gas velocity.
4.3.2 Results
Validation of the numerical model
In order to validate the obtained results, it is necessary to compare the pre-
diction with experimental data. The comparison performed with the mea-
surements of Hoekstra [74] of the Stairmand cyclone using Laser doppler
anomemetry (LDA). The present simulation are compared with the mea-
sured axial and tangential velocity profiles at an axial station located at
94.25 cm from the cyclone bottom (Dx/D = 0.5) as shown in Fig. 4.12 (cf.
Table 4.14: The details of the grid independency study for cyclone I
Total number of cellsStatic pressure drop Cut-off diameter
N/m2 µm
130596 955.51 1.48260230 960.25 1.5478980 961.12 1.51% difference∗ 0.584 1.987
∗ The percentage difference between the coarsest and finest grid
76
4.3. The dust outlet geometry
Radial position /cyclone radius
Tan
gint
ialv
eloc
ity/i
nlet
velo
city
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1
0
0.25
0.5
0.75
1
1.25
1.5
1.75
2
LDARSM
Radial position /cyclone radius
Axi
alve
loci
ty/i
nlet
velo
city
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1
-0.25
0
0.25
0.5
LDARSM
Figure 4.12: Comparison of the time averaged tangential and axial velocity be-
tween the LDA measurements, Hoekstra [74] and the current Reynolds stress tur-
bulence model (RSM) results at 94.25 cm from the cyclone bottom.
Hoekstra [74] for more details about the used cyclone in this validation).
The RSM simulation matches the experimental velocity profile with un-
derestimation of the maximum tangential velocity, and overestimation of
the axial velocity at the central region. Considering the complexity of the
turbulent swirling flow in the cyclones, the agreement between the simu-
lations and measurements is considered to be quite acceptable.
The cyclone pressure drop is calculated as the pressure difference between
the inlet and the average pressure across the vortex finder exit [74]. The
experimental pressure drop of the cyclone can be calculated by the differ-
ence between the static pressures at the inlet and outlet [137]. A com-
parison of the pressure drop, the cut-off diameter (at particle density of
2740 kg/m3) obtained from the experimental data [74], CFD prediction is
shown in Table 4.15. Table 4.15 indicates a very small deviations from
the experimental values in both the calculated pressure drop and cut-off
diameter. As the errors are less than 4%, so it is in the same magnitude
as the experimental error [137]. The above comparison results show that
the numerical model employed in this study can be used to analyze the gas
flow field and performance of the cyclone separator.
Table 4.15: Validation of the computational pressure drop and cut-off diameter
Static pressure drop [N/m2] Cut-off diameter [micron]
Experimental [74] 300 1
CFD 309 0.965
% error 3 3.5
77
Chapter 4. Sensitivity Analysis of Geometrical Parameters
The dominant velocity component of the gas flow in cyclones is the tan-
gential velocity, which results in the centrifugal force for particle separa-
tion [186]. The axial velocity is responsible more than the gravity for the
transport of particles to the collection devices [29, 104]. These velocity
components will be discussed in details in order to investigate the effect
of the dust outlet geometry on the flow properties. Moreover, the pressure
distribution in the swirling flow in these four cyclone separators will be
discussed in details.
The axial variation of the flow properties
Figures 4.13 and 4.14 present the radial profiles of the time-averaged tan-
gential and axial velocity and static pressure at nine axial stations. As
expected, the tangential velocity profiles exhibit the so-called Rankine vor-
tex, which consists of two parts, an outer free vortex and an inner solid
rotation in the center (Fig. 4.13). The tangential velocity distribution in
the inner region is rather similar in different sections for the same cy-
clone. In the outer region, due to the sharp drop in velocity magnitude in
the near wall region, the distribution is different but the maximum tan-
gential velocity is similar in all sections. Generally, the tangential velocity
distribution varies only slightly with axial positions for the same cyclone,
which is also reported in other articles [e.g., 66, 127, 163, 186]. This means
that, if the tangential velocity increases at one section of the cyclone, it will
increase at all other sections. The axial velocity profiles at nine different
stations are shown in Fig. 4.13. Two types of axial velocity profiles are
observed. Cyclone I, III and IV show an inverted W profile. Only cyclone
II (with dustbin) has an inverted V profile. The reason for the two differ-
ent axial velocity profiles can be explained by the change in the flow field
pattern caused by the dustbin.
The radial profiles of the time averaged static pressure are given in Fig. 4.14
for the four cyclones. Like for the tangential velocity, the axial variations
of static pressure are very small for the same cyclone. An exception is the
cyclone I which shows some variations in the central part. Figure 4.15
compares the static pressure profiles for the four cyclones at sections S7-
S9 (located at the cylindrical part of the cyclone, which is the most effective
part of the cyclone in the separation process, the location of the highest area
average tangential velocity). The plots at sections S7-S9 are also represen-
tative for the other sections, as the axial variations in the flow variables
are small. From the comparison between the radial profiles of the four
cyclones, the minimum pressure at the cyclone center is almost the same
for all cyclones. The static pressure radial profiles of cyclones I, III and IV
78
4.3
.T
he
du
stou
tlet
geom
etry
Radial position/ Cyclone radius
Ta
ng
en
tia
lve
locit
y(m
/s)
-1 -0.5 0 0.5 10
5
10
15
20
25
30
S1S2S3S4S5S6S7S8S9
Radial position/ Cyclone radius
Ta
ng
en
tia
lve
locit
y(m
/s)
-1 -0.5 0 0.5 10
5
10
15
20
25
30
S1S2S3S4S5S6S7S8S9
Radial position/ Cyclone radius
Ta
ng
en
tia
lve
locit
y(m
/s)
-1 -0.5 0 0.5 10
5
10
15
20
25
30
S1S2S3S4S5S6S7S8S9
Radial position/ Cyclone radius
Ta
ng
en
tia
lve
locit
y(m
/s)
-1 -0.5 0 0.5 10
5
10
15
20
25
30
S1S2S3S4S5S6S7S8S9
Radial position/ Cyclone radius
Axia
lve
locit
y(m
/s)
-1 -0.5 0 0.5 1-5
0
5
10
15
20S1S2S3S4S5S6S7S8S9
Radial position/ Cyclone radius
Axia
lve
locit
y(m
/s)
-1 -0.5 0 0.5 1-5
0
5
10
15
20S1S2S3S4S5S6S7S8S9
Radial position/ Cyclone radius
Axia
lve
locit
y(m
/s)
-1 -0.5 0 0.5 1-5
0
5
10
15
20S1S2S3S4S5S6S7S8S9
Radial position/ Cyclone radius
Axia
lve
locit
y(m
/s)
-1 -0.5 0 0.5 1-5
0
5
10
15
20S1S2S3S4S5S6S7S8S9
Figure 4.13: The radial profile for the time averaged tangential and axial velocity at different sections.
79
Chapter 4. Sensitivity Analysis of Geometrical Parameters
almost coincide. Cyclone II (with dustbin) depicts fewer gradients in the
radial direction. It has the lowest maximum pressure value. Referring to
Fig. 4.15, there is no difference between the radial pressure profiles at the
three sections for the same cyclone.
Radial position/ Cyclone radius
Sta
tic
pre
ssu
re(N
/m2)
-1 -0.5 0 0.5 1
-200
0
200
400
600
800
1000
1200
1400
S1S2S3S4S5S6S7S8S9
(a) Cyclone I
Radial position/ Cyclone radius
Sta
tic
pre
ssu
re(N
/m2)
-1 -0.5 0 0.5 1
0
200
400
600
800
1000
1200
S1S2S3S4S5S6S7S8S9
(b) Cyclone II (dustbin)
Radial position/ Cyclone radius
Sta
tic
pre
ssu
re(N
/m2)
-1 -0.5 0 0.5 1
-200
0
200
400
600
800
1000
1200
1400
S1S2S3S4S5S6S7S8S9
(c) Cyclone III (dipleg)
Radial position/ Cyclone radius
Sta
tic
pre
ssu
re(N
/m2)
-1 -0.5 0 0.5 1
0
200
400
600
800
1000
1200
1400
S1S2S3S4S5S6S7S8S9
(d) Cyclone IV (dustbin plus dipleg)
Figure 4.14: The radial profiles for the time-averaged static pressure at different
sections for the four cyclones.
The velocity profiles
Figure 4.16 compares the tangential and axial velocity profiles at sections
S7-S9. The variation of the velocity profiles (both the axial and the tan-
80
4.3. The dust outlet geometry
gential) from cyclone to cyclone is mainly located at the central region for
both the tangential and axial velocity. Since the effect of the dust outlet ge-
ometry on the tangential velocity (centrifugal force) is minor, it is expected
that the collection efficiency (cut-off diameter) of the four cyclones will be
comparable.
Radial position/ Cyclone radius
Sta
tic
pre
ssu
re(N
/m2)
-1 -0.5 0 0.5 10
200
400
600
800
1000
1200
1400
1600With dust binWithout dust binWith diplegWith dipleg & dust bin
(a) The four cyclones at section S7
Radial position/ Cyclone radius
Sta
tic
pre
ssu
re(N
/m2)
-1 -0.5 0 0.5 10
200
400
600
800
1000
1200
1400
1600With dust binWithout dust binWith diplegWith dipleg & dust bin
(b) The four cyclones at section S8
Radial position/ Cyclone radius
Sta
tic
pre
ssu
re(N
/m2)
-1 -0.5 0 0.5 10
200
400
600
800
1000
1200
1400
1600With dust binWithout dust binWith diplegWith dipleg & dust bin
(c) The four cyclones at section S9
Figure 4.15: The radial profile for the time-averaged static pressure at different
sections for the four cyclones.
The axial velocity in the inner vortex is either reported as an inverted V
81
Chapter 4. Sensitivity Analysis of Geometrical Parameters
or inverted W-shaped profile,i.e., with a maximum (V-shaped) or a dip (W-
shaped) at the symmetry axis [29]. Hoekstra et al. [75] stated that the
shape of the axial velocity profile is affected by the cyclone geometry. They
referred the dip in the inverted W profile to the loss of swirl in the vortex
finder (the friction force of the vortex finder wall attenuates the swirling
flow), which results in an adverse pressure gradient at the centerline [78].
Hence, fluid with fewer swirls is drawn back from the exit pipe into the cy-
clone. This core flow prevails throughout the entire separation space of the
cyclone in spite of the attenuation of swirl in the conical part of the cyclone
[78]. This explains the reason behind the inverted W-shaped profile ex-
hibited by cyclones I (cylinder on cone), cyclone III (dipleg) and cyclone IV
(dustbin plus dipleg). However, why cyclone II (with dustbin) exhibit the
inverted V-shaped profile? In cyclone II, the dustbin has an equal diame-
ter to the cyclone barrel, which means sudden expansion to the downward
flow and sudden contraction to the upward flow (the flow inside the dustbin
also has two streams due to the absence of a vortex stabilizer [75] which
can prevent the re-entrainment of the collected particles). The upward gas
flow (directed from the end of dustbin) has a higher kinetic energy espe-
cially at the cone-tip diameter (sudden contraction) which can overcome
the adverse pressure gradient at the centerline (caused by the swirl at-
tenuation in the vortex finder) and results in the inverted V axial velocity
profile. One more question may appear now, why cyclones III (dipleg) and
IV (dustbin plus dipleg) did not exhibit the inverted V-shaped profile? The
reason can be referred to the diameter of the dustbin directly connected to
the cyclone. For cyclones III and IV there is no change in the flow area at
the connection, consequently no flow acceleration happens. The effect of
dustbin dimensions (diameter, height) still need more investigations. The
author believe, if a cone is inserted at the entrance of the cone-tip inside
the dustbin of cyclone II, the axial velocity will become inverted W-shaped.
However, the dimensions and location of this cone still need more investi-
gations (cf. Obermair et al. [123] for more details).
The flow pattern
Regarding the effect of neglecting the dust outlet geometry (dustbin or
dipleg or dustbin plus dipleg) in the simulating domain on the flow field
pattern, the following comments can be drawn (Fig. 4.17).
1. From the comparison between the static pressure contour plots of
each cyclone and that of cyclone I, the highest value of static pres-
sure is obtained in case of cyclone III (dipleg). The lowest value is
given by cyclone II (dustbin). The highest value of the static pressure
82
4.3. The dust outlet geometry
Radial position/ Cyclone radius
Ta
ng
en
tia
lve
locit
y(m
/s)
-1 -0.5 0 0.5 10
5
10
15
20
25
30
35With dust binWithout dust binWith diplegWith dipleg & dust bin
Radial position/ Cyclone radius
Axia
lve
locit
y(m
/s)
-1 -0.5 0 0.5 1-10
-5
0
5
10
15
20
25With dust binWithout dust binWith diplegWith dipleg & dust bin
Radial position/ Cyclone radius
Ta
ng
en
tia
lve
locit
y(m
/s)
-1 -0.5 0 0.5 10
5
10
15
20
25
30
35With dust binWithout dust binWith diplegWith dipleg & dust bin
Radial position/ Cyclone radius
Axia
lve
locit
y(m
/s)
-1 -0.5 0 0.5 1-10
-5
0
5
10
15
20
25With dust binWithout dust binWith diplegWith dipleg & dust bin
Radial position/ Cyclone radius
Ta
ng
en
tia
lve
locit
y(m
/s)
-1 -0.5 0 0.5 10
5
10
15
20
25
30
35With dust binWithout dust binWith diplegWith dipleg & dust bin
Radial position/ Cyclone radius
Axia
lve
locit
y(m
/s)
-1 -0.5 0 0.5 1-10
-5
0
5
10
15
20
25With dust binWithout dust binWith diplegWith dipleg & dust bin
Figure 4.16: The radial profile for the time averaged tangential and axial velocity
at different sections on the X-Z plane (Y=0) at the inlet region (sections S7, S8 and
S9). From top to bottom: section S7-S9.
83
Chapter 4. Sensitivity Analysis of Geometrical Parameters
in cyclone IV (dustbin plus dipleg) is in between that for cyclones III
and II. This indicates a slight underestimation of the pressure drop
by neglecting the effect of dipleg or dustbin plus dipleg (cyclone III, IV
versus cyclone I). Furthermore, a slight overestimation of the pres-
sure drop is observed by neglecting the effect of the dustbin (cyclone
II versus cyclone I).
2. The tangential velocity pattern is very similar in the four cyclones
(Rankine profile). The highest value is almost the same for the four
cyclones. Consequently, the collection efficiency (cut-off diameter) of
the four cyclones will be almost the same. Moreover, the contour plots
for the tangential velocity of the four cyclones are quite similar in the
main separation space (cylinder and cone).
3. The axial velocity patterns for cyclones I, III and IV have the shape
of an inverted W profile while that of cyclone II has an inverted V
profile, indicating different flow behavior. Close to the cone bottom
there are different flow patterns as a consequence of different dust
outlet geometry.
Figure 4.18 shows a qualitative view of the complex flow in the four tested
cyclones with the streamtraces plots of the time averaged velocities colored
by the time-average axial velocity. The swirling, downward flow at the
outer region of the cyclone is clearly visible. Near the bottom of the cyclone,
it is diverted into an upward flow near the cyclone center. In cyclone II
(with dustbin) and IV (with dustbin and dipleg), the flow behavior in the
dustbin is quite different because of the different length of the dustbin.
The performance
Figure 4.19 presents the grade efficiency curves (GECs) for the four tested
cyclones. As expected, the frictional efficiencies of all the cyclones are seen
to increase with the increase in particle size. The shapes of the grade
collection efficiency curves of all models have a so-called S shape [195].
It is clear from Fig. 4.19 that the effect of the cyclone dustbin or dipleg,
on the cut-off diameter (particle diameter of 50% collection efficiency) is
small. Neglecting the effect of dust outlet geometry in the cyclone simu-
lation slightly overestimates the cut-off diameter (when compare it with
cyclone I). A deeper look to the GECs indicates the variation of the collec-
tion efficiency for particles diameters larger than 1.5 µm, with a higher
efficiency for cyclone II and the lower efficiency for cyclone I. This behavior
is due to the increase in the separation space and the change in the max-
imum tangential velocity. For particles with diameters less than 0.8 µm,
84
4.3. The dust outlet geometry
Figure 4.17: The contour plots for the time averaged flow variables at sections Y=0
and throughout the inlet section. From left to right: the static pressure (N/m2), the
tangential and the axial velocity (m/s).
85
Chapter 4. Sensitivity Analysis of Geometrical Parameters
Figure 4.18: The streamtraces plots for the time averaged flow variables, colored
by the average axial velocity (m/s).
86
4.3. The dust outlet geometry
the collection efficiency tends to zero except in case of cyclone IV which
depict a non zero collection efficiency.
In order to estimate the effect of dust outlet geometry on the performance
parameters, the pressure drop and cut-off diameter have been calculated.
Table 4.16 shows the effect of dust outlet geometry absence from the simu-
lation domain on the pressure drop and cut-off diameter. From the compar-
ison between the estimated pressure drop of cyclone I and the other three
cyclones, the following comments are obtained. A slight underestimation
is obtained when omitting the dipleg or dustbin plus dipleg from the sim-
ulation space. A slight overestimation result from omitting the dustbin.
The pressure drop values given in Table 4.16 support the results obtained
from the analysis of the static pressure contour plots.
Table 4.16: The effect of dust outlet geometry on the cyclone performance
Cyclone I II (dustbin) III (dipleg) IV (dustbin plus dipleg)
Pressure drop (N/m2) 960 890 1017 1008
Cut-off diameter (µm) 1.5 1.0 1.25 1.2
Particle diameter [micron]
Col
lect
ion
effic
ienc
y
1 2 3 4 5
0.2
0.4
0.6
0.8
1Without dustbinWith dustbinWith diplegWith dustbin plus dipleg
Figure 4.19: The grade efficiency curves for the four cyclones.
4.3.3 Conclusions
Four cyclones have been simulated using the Reynolds stress model (RSM),
to study the effect of the dust outlet geometry on the cyclone separator
performance, flow pattern and velocity profiles. The following conclusions
have been obtained.
87
Chapter 4. Sensitivity Analysis of Geometrical Parameters
• The maximum tangential velocity in the four cyclones is very similar.
• No radial acceleration occurs in the cyclone space (the maximum tan-
gential velocity is nearly constant throughout the cyclone).
• The cyclone without dustbin slightly overestimates both the pressure
drop and cut-off diameter. So the simulations with and without dust-
bin will produce nearly the same performance parameters.
• The cyclone without dipleg slightly underestimates the pressure drop
and overestimates the cut-off diameter. Consequently, the simula-
tions with and without dipleg will produce nearly the same perfor-
mance parameters.
• The axial velocity patterns obtained by the four cyclones are differ-
ent.
• If the main target of the CFD investigation is the performance pa-
rameters, one can safely simulate only the main separation space
(cylinder on cone). However, if the aim is to investigate the flow field
pattern, the dust outlet geometry should be included in the simula-
tion domain.
4.4 Closure
From the previous sections, it becomes clear that the most significant fac-
tors are four (the vortex finder diameter, the inlet width, the inlet height
and the total cyclone height (cone height)). Consequently, any minor mod-
ification in these particular factors will result in a considerable change in
the cyclone performance. The effect of the cone-tip diameter on the cyclone
performance is insignificant. Since the main target of this thesis is to op-
timize the cyclone performance, it is accepted to exclude the dust hopper
(dustbin or dipleg) from the simulation domain and use the obtained CFD
data to obtain the optimized cyclone design.
88
Chapter 5
The Vortex Finder Dimensions
5.1 Introduction
The vortex finder size is an especially important dimension, which sig-
nificantly affects the cyclone performance as its size plays a critical role
in defining the flow field inside the cyclone, including the pattern of the
outer and inner spiral flows. Numerous studies have been performed for
the effect of geometrical parameters on the flow pattern and performance
[15, 52, 62, 102, 140, 184] whereas only limited number of studies have
been devoted to the effect of the vortex finder dimensions.
Iozia and Leith [84] optimized the cyclone design parameters, including
the vortex finder diameter, to improve the cyclone performance using an
optimization code. Kim and Lee [95] described how the ratio of the di-
ameters of cyclone body, and the vortex finder diameter Dx/D affected the
collection efficiency and pressure drop of cyclones, and proposed an energy-
effective cyclone design. Moore and Mcfarland [111] also tested cyclones,
with six different vortex finders, and concluded that the variation in the
gas outlet diameter under the constraint of a constant cyclone Reynolds
number produced a change in the aerodynamic particle cut-off diameter.
Bakari and Hamdullahpur [7] investigated experimentally the effects of
both the inlet gas velocity, the cyclone inlet width, the vortex finder length
and the vortex finder diameter on the cyclone performance. They reported
that the vortex finder length and the vortex finder diameter have a strong
effect on the cyclone performance parameters. Moreover, the results in-
dicated that the vortex finder length has a direct effect on the cyclone
performance. The longer vortex finder minimizes the short circuiting of
incoming gases, preventing the dusty gas from flowing directly from the
inlet to the outlet. Zhu and Lee [200] carried out a set of experimental
89
Chapter 5. The Vortex Finder Dimensions
(a) The cyclone geometry (b) The surface mesh for cyclone S1 (D5)
Figure 5.1: Schematic diagram for the cyclone separator
investigations on the particle collection efficiency of small cyclones oper-
ating at high-flow rates. Special emphasis was given to the effects of the
barrel height and the vortex finder length on the particle collection effi-
ciency. The length ratios of the barrel height h and vortex finder length Sto the cyclone body diameter D were varied from 0.75 to 4.5 and from 0.5 to
1.5, respectively. Pressure drop decreased substantially either as the bar-
rel height h became longer or as the vortex finder length S became shorter.
It was also found that the difference between the cyclone barrel height hand the vortex finder length S affects the particle collection characteristics
significantly. The optimum performance will be obtained if (h−S)/D = 1.0.
The performance of a cyclone, with different vortex finders, was evaluated
by Lim et al. [102] to examine the effect of the vortex finder shape on
the characteristics of the collection efficiency. Four cylinder-shaped and
six cone-shaped vortex finders were designed and employed to compare
the collection efficiencies of the cyclone, at flow rates of 30 and 50 l/min.
The cylinder-shaped-vortex finders had different diameters and the cone-
shaped vortex finders had different cone lengths. The result indicates that
two cone-shaped vortex finders, with different diameters, had the collec-
tion efficiencies between those of the cylinder-shaped vortex finders with
the same diameter, and that a smaller pressure drop per flow rate unit
could be achieved for the cone-shaped design, but the cone length did not
affect the collection efficiency and pressure drop of the cyclone. Raoufi et
al. [140] duplicated numerically the same study of Lim et al. [102] with
limited details about the effect of the gas outlet diameter on the flow field
90
5.1. Introduction
pattern and velocity profile.
You-hai et al. [188] simulated the three-dimensional gas-phase flow field
in the cyclone separator with different vortex finder diameters. The results
show that when the diameter of the vortex finder decreases, the downward
flow decreases and the tangential velocities of the whole cyclone separa-
tor increase, but at the cost of the pressure drop increase. Horvath et al.
[79] refereed the reason behind the two classes of axial velocity in cyclone
separator (class V and class W) to the influence of the vortex finder di-
ameter. The pressure drop and collection efficiency of a swirl tube with
different vortex finder geometries were studied numerically by Jian and
You-hai [87]. The gas flow fields were simulated by the Reynolds stress
model (RSM) and the stochastic tracking approach in discrete phase model
(DPM). The results indicate that the decrease of the vortex finder diameter
leads to higher tangential velocity, which helps to improve the separation
efficiency. The back flow can be observed in converging coned-shaped vor-
tex finder, meanwhile the diverging coned-shape vortex finder can make
the flow move smoothly to reduce pressure drop.
Ficici et al. [57] performed an experimental study using three cylinder-
shaped vortex finders with diameters of 80, 120 and 160 mm. They inves-
tigated the effects of gas inlet velocity, the vortex finder diameter Dx and
length S on the cyclone performance. They reported a linear relationship
between the length of the vortex finder and the pressure loss. Khalkhalia
and Safikhania [94] performed a multi-objective optimization of a cyclone
vortex finder shape using CFD simulations data set. Two meta-models
based on the evolved group method of data handling (GMDH) type neural
networks are used as fitness functions for Pareto-based optimization.
In summary, all previous studies reported the significant effect of the vor-
tex finder dimensions on the cyclone performance and flow pattern. Nev-
ertheless, the previous studies are not coherent, and did not present suf-
ficient details about the effect of these two geometrical parameters on the
pressure drop and the cut-off diameter. Moreover, detailed studies about
the effect of the vortex finder dimensions on the velocity profiles are scarce
in the literature.
The present study is intended to computationally investigate the effect of
increasing the vortex finder diameter Dx and length S on the pressure drop
and cut-off diameter and to obtain more details about the flow field pattern
and velocity profiles using the large eddy simulation (LES) methodology.
91
Chapter 5. The Vortex Finder Dimensions
5.2 Numerical settings
5.2.1 Configuration of the tested cyclones
The numerical simulations were performed on five cyclones with different
vortex finder diameters (at constant S) and five cyclones with different
vortex finder length (at constant Dx). Figure 5.1 and Table 5.1 give the
cyclones geometrical dimensions.
5.2.2 Solver settings
The simulations started with unsteady simulation using the Reynolds stress
turbulence model with a time step of 1E-4 for initialization of the flow field.
Afterwards, the turbulence model switched to the large eddy simulation
with a time step of 1E-5s using implicit coupled solution algorithm. The
selected time step results in an average inlet Courant number of 0.0288
for the tested cyclones. The cyclones volumes’ and the corresponding res-
idence times for the tested cyclones are given in Table 5.2. The minimum
value of residence time is 0.0812s i.e., the used time step is just a small
fraction of the average residence time. This confirms that the used time
step can reveal the transient phenomena properly. For the tested cyclones
the maximum values of the courant number are less than 0.2. This verifies
again that the choice of the time step was proper (cf. Sec. 4.2.1.3, page 56
for more details).
Table 5.1: The geometrical dimensions of the tested cyclonesa
Dimension Cycloneb dimension/D (h − S)/D Dx/SInlet height, a 0.375Inlet width, b 0.2625Barrel height, h 1.5c
Total cyclone height, Ht 4.0Cone tip-diameter, Bc 0.375
Vortex finder diameter, Dx S/D = 0.5
D1 0.30 1.0 0.6D2 0.35 1.0 0.7D3 0.40 1.0 0.8D4 0.45 1.0 0.9D5 0.50 1.0 1.0
Vortex finder length, S Dx/D = 0.5
S1 0.5 1.0 1.0S2 0.625 0.875 0.8S3 0.875 0.625 0.625S4 1.0 0.5 0.5S5 1.0 1.0c 1.0
a Body diameter, D = 31 mm. The outlet section is above the cylindrical barrel surface by Le = 0.5D.The inlet section located at a distance Li = D from the cyclone center, cf. Fig. 5.1(a).
b Cyclone D5 and S1 are identical.
c cyclone S6 has different barrel height h/D = 2.0 .
92
5.2. Numerical settings
Table 5.2: The details of the tested cyclones
Cyclone D1 D2 D3 D4 D5 =S1 S2 S3 S4 S5
Number of cellsa 601482 620088 657841 666336 714029 795321 798143 798256 888612
Cyclone volume x 105 [m3] 6.765 6.802 6.845 6.894 6.949 6.948677 6.947 6.946 8.116
tresb [s] 0.0812 0.0816 0.0821 0.0827 0.0834 0.0834 0.0834 0.0834 0.0974
a The total number of hexahedral cells after the grid independency study, cf. Sec. 5.2.4
b The residence time tres = V/Qin where V is the cyclone volume and Qin is the gas flow rate (50 l/min).
5.2.3 Boundary conditions
Velocity inlet boundary condition is applied at inlet, outflow at the gas
outlet and wall (no-slip) boundary condition at all other boundaries. The
air inlet volume flow rate Qin=50 l/min, air density 1.0 kg/m3 and dynamic
viscosity of 2.11E-5 Pa s. The turbulence intensity I equals 5% and the
turbulence characteristic length equals 0.07 times the inlet width [52, 55,
74]. All other settings are identical to that given in Sec. 4.2.1.3.
5.2.4 Grid independency study
The grid independence study has been performed for the tested cyclones.
Three levels of grid for each cyclone have been tested, to be sure that
the obtained results are grid independent. The hexahedral computational
grids were generated using GAMBIT grid generator and the simulations
were performed using Fluent 6.3.26 commercial finite volume solver on
an eight nodes CPU Opteron 64 Linux cluster. To evaluate accurately the
numerical uncertainties in the computational results, the concept of grid
convergence index (GCI) was adopted using three grid levels per cyclone.
Grid convergence index (GCI)
The grid convergence index (GCI) proposed by Roache [143–145] was em-
ployed to test the grid independence of the simulations. The GCI are com-
Table 5.3: The details of the grid independence study for cyclone S1 (D5)
Number of cells Eua Cut-off diameter
714029 3.375 1.9851174029 3.18 1.4361793459 3.145 1.318
% differenceb -6.8148 -33.602% differencec -1.1006 -8.2173
a Euler number is the dimensionless pressure drop Eu = ∆P/( 12ρV 2
in) where ∆P is the static pressure
drop, ρ is the gas density, Vin is the gas inlet velocity. b The percentage difference between the coarsestand finest grid. c The percentage difference between the fine and finest grid.
93
Chapter 5. The Vortex Finder Dimensions
puted using three levels of grids in order to estimate accurately the order of
convergence and check that the solution is within the asymptotic range of
convergence [160]. The GCI is based upon a grid refinement error estima-
tor derived from the theory of the generalized Richardson extrapolation
[55, 160]. The GCI is a measure of how far the computed value is away
from the value of the asymptotic numerical value. Consequently, it indi-
cates how much the solution would change with a further refinement of
the grid. A small value of GCI indicates that the computation is within the
asymptotic range (cf. Sec. 4.2.1.3 for more details).
Table 5.4 presents the grid convergency calculations using GCI method for
three grid levels for cyclone S1 as an example for the tested cyclones. The
following conclusions have been obtained from the GCI analysis [55]:
• The results are in the asymptotic range, because the obtained values
for α are close to unity.
• The ratio R is less than unity this means monotonic convergence [2].
• There is a reduction in the GCI value for the successive grid refine-
ments (GCIfine12 < GCI
fine23 ) for the two variables (Eu and x50). This in-
dicates that the dependency of the numerical results on the cell size
has been reduced and a grid independent solution has been achieved.
Further refinement of the grid will not give much change in the sim-
ulation results. For the two variables (Eu and x50), the extrapolated
value is only slightly lower than the finest grid solution. Therefore,
the solution has converged with the refinement from the coarser grid
to the finer grid [2]. Figure 5.2 presents a qualitative proof that the
obtained results are in the asymptotic range.
Table 5.4: Grid convergency calculations using GCI method and three grid levels
for cyclone S1
i Ni fi ri,i+1 ei,i+1 εi,i+1 GCIfinei,i+1
% Ra αb
Eu
0c 3.1336
1 1793459 3.1450
1.1517 0.0350 0.0111 0.45202 1174029 3.1800 0.1815 1.0111
1.1803 0.1950 0.0613 1.8228
3 714029 3.3750
x50
0 1.2703
1 1793459 1.3180
1.1517 0.1180 0.0895 4.5223
2 1174029 1.4360 0.2342 1.08951.1803 0.5490 0.3823 14.4222
3 714029 1.9850
a R=ε12/ε23 .
b α=(
rp12
GCI12
)
/GCI23 .
c The value at zero grid space (h → 0). i=1, 2 and 3 denote the calculations at the fine, medium and coarse mesh respectively.
94
5.3. Results and discussions
N -1
Eul
ernu
mbe
r
Cut
-off
diam
eter
0 5E-07 1E-063
3.1
3.2
3.3
3.4
1.2
1.4
1.6
1.8
2
2.2Euler number
Cut-off diameter
(h --> 0)(h --> 0)(h --> 0)(h --> 0)(h --> 0)(h --> 0)(h --> 0)(h --> 0)(h --> 0)(h --> 0)(h --> 0)(h --> 0)(h --> 0)(h --> 0)(h -- > 0)(h --> 0)
Figure 5.2: Qualitative representation of the grid independency study. The Euler
number and the cut-off diameter for cyclone S1 at the three grid levels. N−1 is
the reciprocal of the number of cells, h → 0 means the value at zero grid size (cf.
Table 5.4). To obtain a smooth curve; the spline curve fitting has been applied in
Tecplot post-processing software.
Table 5.5: The position of different plotting sections
Section S1 S2 S3 S4 S5 S6 S7 S8 S9
z`/Da 2.75 2.5 2.25 2 1.75 1.5 1.25 1.0 0.75
a z`measured from the inlet section top (cf. Fig. 5.1(a)).
5.3 Results and discussions
The flow velocity can be decomposed into three components. The tangen-
tial and the axial velocity components are the major velocity components
in comparison with the radial velocity component. Xiang and Lee [186]
stated that the tangential velocity is the dominant gas velocity in gas cy-
clones, which results in the centrifugal force for particle separation. The
axial component is responsible for the two flow streams (downward and
upward).
5.3.1 The axial variation
Figures 5.3 and 5.4 present the radial profiles of the time-averaged tan-
gential and axial velocity and static pressure at nine axial stations (cf. Ta-
ble 5.5). As expected, the tangential velocity profiles exhibit the so-called
Rankine vortex, which consists of two parts, an outer free vortex and an
inner solid rotation in the center. The tangential velocity distribution in
the inner region is rather similar at different sections for the same cyclone.
95
Chapter 5. The Vortex Finder Dimensions
In the outer region, due to the sharp drop in the velocity magnitude in the
near wall region, the distribution is different but the maximum tangen-
tial velocity is similar at all sections for the same cyclone (S1-S5). The
maximum tangential velocity increases with decreasing the vortex finder
diameter.
The cyclones S1 to S5 show the inverted W profile but cyclones D1 to D3
show the inverted V profile. The radial profiles of the time averaged static
pressure are given in Fig. 5.3. Like for the tangential velocity, the ax-
ial variations of static pressure are very small for the same cyclone. The
maximum value of the static pressure decreases when the vortex finder
diameter is increased for cyclones D1 to D5. However, the maximum value
of the static pressure slightly increases when the vortex finder length is
decreased for cyclones S1 to S4. Cyclone S5 (Dx/D = 0.5, S/D = 0.5,
h/D = 2.0) differs than cyclone S4 in only the barrel height (h/D = 1.5).
From Fig. 5.4, a slight difference in the maximum tangential velocity be-
tween the two cyclones, which predict close values for the collection effi-
ciency. Moreover, the axial velocity profiles for the two cyclones are close
but less central dip in the axial velocity is exhibited in cyclone S5. The dis-
tinct difference between cyclones S4 and S5 is that in the maximum static
pressure, where cyclone S5 indicates a reduction in the maximum static
pressure. This indicates that the effect of changing the barrel height is
more significant than that of the vortex finder length. The sharp changes
in the radial profiles appear in Fig. 5.4 are due to crossing the vortex finder.
5.3.2 The flow pattern
Figure 5.5 shows the contour plots of the time-averaged static pressure,
tangential and axial velocity for cyclones D1-D5. The time-averaged static
pressure decreases radially from the wall to the center. A negative pres-
sure zone appears in the forced vortex region (central region) due to high
swirling velocity. The pressure gradient is largest along the radial direc-
tion, whereas the gradient in axial direction is very limited. The cyclonic
flow is not symmetrical as is clear from the shape of the low-pressure zone
at the cyclone center (twisted cylinder). Two vortical motions are exist one
moving down (outer vortex) and the other moving up (inner vortex). The
highest value of the static pressure decreases with increasing the vortex
finder diameter. Consequently, a smaller pressure drop can be expected
when increasing the vortex finder diameter. The tangential velocity pat-
tern is very similar in all cyclones (Rankine profile). The highest value
decreases with increasing the vortex finder diameter. Consequently, a bet-
ter collection efficiency can be expected when decreasing the vortex finder
96
5.3
.R
esu
ltsan
dd
iscussio
ns
D1 D2 D3 D4 D5
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-4
0
4
8
12
16
S1S2S3S4S5S6S7S8S9
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015
0
4
8
12
16
S1S2S3S4S5S6S7S8S9
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-4
0
4
8
12
16
S1S2S3S4S5S6S7S8S9
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.0150
2
4
6
8
10
12
14
S1S2S3S4S5S6S7S8S9
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.0150
2
4
6
8
10
12
S1S2S3S4S5S6S7S8S9
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-5
0
5
10
15S1S2S3S4S5S6S7S8S9
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-4
0
4
8
12S1S2S3S4S5S6S7S8S9
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-4
-2
0
2
4
6
8S1S2S3S4S5S6S7S8S9
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-4
-2
0
2
4
6S1S2S3S4S5S6S7S8S9
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-5
-2.5
0
2.5
5
7.5
10S1S2S3S4S5S6S7S8S9
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.0150
100
200
300
400
500
S1S2S3S4S5S6S7S8S9
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015
0
100
200
300
400
S1S2S3S4S5S6S7S8S9
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015
0
80
160
240
320
S1S2S3S4S5S6S7S8S9
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015
-40
0
40
80
120
160
200
240
280
S1S2S3S4S5S6S7S8S9
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-100
-50
0
50
100
150
200
S1S2S3S4S5S6S7S8S9
Figure 5.3: The radial profile for the time-averaged tangential and axial velocity and static pressure at different sections for
cyclones D1 - D5.
97
Ch
ap
ter
5.
Th
eV
orte
xF
ind
er
Dim
en
sion
s
S1 S2 S3 S4 S5
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.0150
2
4
6
8
10
12
S1S2S3S4S5S6S7S8S9
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.0150
2
4
6
8
10
12
S1S2S3S4S5S6S7S8S9
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.0150
2
4
6
8
10
12
S1S2S3S4S5S6S7S8S9
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.0150
2
4
6
8
10
12
S1S2S3S4S5S6S7S8S9
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.0150
2
4
6
8
10
12
S1S2S3S4S5S6S7S8S9
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-5
-2.5
0
2.5
5
7.5
10S1S2S3S4S5S6S7S8S9
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-5
0
5
10S1S2S3S4S5S6S7S8S9
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-5
0
5
10S1S2S3S4S5S6S7S8S9
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015
-4
-2
0
2
4
6
8
10S1S2S3S4S5S6S7S8S9
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015
-4
-2
0
2
4
6
8
10S1S2S3S4S5S6S7S8S9
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-100
-50
0
50
100
150
200
S1S2S3S4S5S6S7S8S9
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-100
-50
0
50
100
150
200
S1S2S3S4S5S6S7S8S9
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-100
-50
0
50
100
150
200
S1S2S3S4S5S6S7S8S9
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-100
-50
0
50
100
150
200
S1S2S3S4S5S6S7S8S9
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-100
-50
0
50
100
150
200
S1S2S3S4S5S6S7S8S9
Figure 5.4: The radial profile for the time-averaged tangential and axial velocity and static pressure at different sections for
cyclones S1 - S5.
98
5.3. Results and discussions
diameter.
Figure 5.6 shows the contour plots of the time-averaged static pressure,
tangential and axial velocity for cyclones S1-S5. The highest value of the
static pressure slightly decreases with increasing the vortex finder length.
The tangential velocity pattern is very similar in all cyclones (Rankine
profile). The highest value is very closed for cyclones S1 to S4 but there is
a small difference between cyclones S4 and S5. The axial velocity patterns
for the five cyclones (S1 to S5) have the shape of an inverted W profile.
Cyclone S5 exhibit a less central dip in the axial velocity.
5.3.3 The radial variation
The tangential and axial velocity profiles at section S6 (as a representative
for the other sections, because the axial variations in the flow variables are
small) for the tested cyclones are compared in Fig. 5.7. From the compar-
ison between the radial profiles of the five D cyclones, the minimum pres-
sure at the cyclone center is almost the same for cyclones (D3 - D5). The
slope of the static pressure radial profile becomes flatter with increasing
the vortex finder diameter. Contrarily, the variation of the static pressure
for cyclones S1 to S5 is very small.
Decreasing the cyclone vortex finder diameter, increases the maximum
tangential velocity. The maximum tangential velocity approaches asymp-
totically 1.589 times the inlet velocity when decreasing the vortex finder
diameter. The effect of increasing the vortex finder length on the maximum
tangential velocity is limited and small reduction in the maximum tan-
gential velocity by increasing the barrel height for the same vortex finder
dimensions (cyclone S5).
The variation of axial velocity with changing the vortex finder length is
limited close to the wall, especially in the outer part. In the central region,
the change in axial velocity profile is more pronounced with the dip in
axial velocity decreasing with increasing the vortex finder length. Cyclone
S5 exhibit the smallest dip among cyclones S1 to S5.
The variation of axial velocity in the five D cyclones is limited close to the
wall with a big change in the central part. The axial velocity profile grad-
ually changes from the inverted W to be inverted V with decreasing vortex
finder diameter. Again, limited effect on the axial velocity profile (inverted
W) for the vortex finder length is recorded. Moreover, increasing the vortex
finder length has only small influence on the central part. Furthermore,
The influence of the barrel height is more pronounced than that of the vor-
tex finder length. Generally, the effect of changing the vortex finder length
99
Ch
ap
ter
5.
Th
eV
orte
xF
ind
er
Dim
en
sion
s
D1 D2 D3 D4 D5
Figure 5.5: The contour plots for the time averaged flow variables at sections Y=0 and throughout the inlet section for
cyclones D1 - D5. From top to bottom: the static pressure (N/m2), the tangential velocity (m/s) and the axial velocity (m/s).
From left to right cyclone D1 - D5.
100
5.3
.R
esu
ltsan
dd
iscussio
ns
S1 S2 S3 S4 S5
Figure 5.6: The contour plots for the time averaged flow variables at sections Y=0 and throughout the inlet section for
cyclones S1 - S5. From top to bottom: the static pressure (N/m2), the tangential velocity (m/s) and the axial velocity (m/s).
From left to right cyclone S1 - S5.
101
Chapter 5. The Vortex Finder Dimensions
on the axial velocity profile is very small in comparison with the vortex
finder diameter.
The shape of the axial velocity profile is affected by the cyclone geometry
[75]. Hoekstra et al. [75] referred the dip in the inverted W profile to the
loss of swirl in the vortex finder (the friction force of the vortex finder wall
attenuates the swirling flow), which results in an adverse pressure gradi-
ent at the centerline [78]. Hence, fluid with fewer swirls is drawn back
from the exit pipe into the cyclone. This core flow prevails throughout the
entire separation space of the cyclone in spite of the attenuation of swirl
in the conical part of the cyclone [78]. This explains the reason behind the
inverted W-shaped profile exhibited by cyclones S1 to S5. However, why
decreasing the vortex finder diameter gradually change the axial velocity
profile from inverted W to exhibit the inverted V-shaped profile? The an-
swer is, when the vortex finder diameter decreases, the swirl in the cyclone
increases (as is clear from the tangential velocity profiles for cyclones D1
to D5). Consequently, the flow can overcome the adverse pressure gradient
and exhibit the inverted V profile for cyclone D1.
5.3.4 The cyclone performance
In order to estimate the effect of the vortex finder dimensions on the per-
formance parameters, the Euler number (the dimensionless pressure drop)
have been calculated. A discrete phase modeling (DPM) study has been
performed by injecting 104 particles from the inlet surface with a particle
density of 860 kg/m3 and with a particle size ranging from 0.025 until 5
micron.
Figure 5.8 shows a sharp decrease of the Euler number (dimensionless
pressure drop) with increasing the vortex finder diameter Dx and a small
increase with increasing the vortex finder length S. This behavior can
be explained as follows. The pressure drop in the cyclone is composed of
three main contributions [52]: (1) the pressure drop at the inlet section.
(2) the pressure drop in the cyclone body due to swirling motion and due
to wall friction, this contribution may increase with increasing the cyclone
barrel height (cyclone S5) or with increasing the vortex finder length (cy-
clones S1 - S4) as the wall friction will increase due to friction with a larger
wall surface, or decreases as the vortex strength will decrease because the
maximum tangential velocity decreases. (3) the main contribution to the
cyclone pressure drop is the energy loss in the vortex finder, which mainly
depends on the maximum tangential velocity in the cyclone. As is clear
from Fig. 5.7 the maximum tangential velocity decreases with increasing
the vortex finder diameter. As the inlet section is the same in the all tested
102
5.3. Results and discussions
Radial position (m)
Sta
ticpr
essu
re(N
/m2)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015
0
100
200
300
400
500
Dx/D=0.30Dx/D=0.35Dx/D=0.40Dx/D=0.45Dx/D=0.50
Radial position (m)
Sta
ticpr
essu
re(N
/m2)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-100
-50
0
50
100
150
200
S/D=0.5
S/D=0.625
S/D=0.875
S/D=1.0 (h/D=1.5)
S/D=1.0 (h/D=2.0)
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-2
0
2
4
6
8
10
12
14
Dx/D=0.30Dx/D=0.35Dx/D=0.40Dx/D=0.45Dx/D=0.50
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-2
0
2
4
6
8
10
12
S/D=0.5
S/D=0.625
S/D=0.875
S/D=1.0 (h/D=1.5)
S/D=1.0 (h/D=2.0)
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-2
0
2
4
6
8Dx/D=0.30Dx/D=0.35Dx/D=0.40Dx/D=0.45Dx/D=0.50
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-5
-2.5
0
2.5
5
S/D=0.5
S/D=0.625
S/D=0.875
S/D=1.0 (h/D=1.5)
S/D=1.0 (h/D=2.0)
Figure 5.7: Comparison between the radial profiles for the time averaged static
pressure, tangential and axial velocity at section S6.
103
Chapter 5. The Vortex Finder Dimensions
cyclones, the pressure drop in the inlet section does not vary with increas-
ing the vortex finder dimensions. The sharp decrease of the Euler number
between the D cyclones with increasing the Dx is mainly due to the de-
crease in the pressure drop as a result of the decrease in the maximum
tangential velocity. The Euler number in cyclones (S1 - S4) increases as
the vortex finder length increases. Since, the maximum tangential veloc-
ity in the four cyclones is almost the same. Consequently, this trend is due
to the small increase in the energy loss at the wall surface with increasing
S. Cyclone S exhibits a sudden (small) drop in the Euler number. There
are two competing contributions: increase of the pressure drop due to fric-
tion (the barrel height in cyclone S5 is larger than that for cyclone S4)
and decrease of the pressure drop due to the small reduction in the vortex
strength (maximum tangential velocity, cf. Fig. 5.7).
Dx
S
Eul
ernu
mbe
r
Stk
50x
10
3
0.3 0.35 0.4 0.45 0.5
0.6 0.8 1
3
4
5
6
7
8
9
10
11
12
0
0.5
1
1.5
2
2.5
Euler number (Dx)
Stk50 x 103 (Dx)
Euler number (S)
Stk50 x 103 (S)
Figure 5.8: The variation of the Euler number and the Stokes number with the
vortex finder dimensions.
The trend of increasing the Stokes number with increasing the vortex
finder diameter is quite reasonable, as the centrifugal force affecting par-
ticles attenuates when the swirl intensity (maximum tangential velocity)
decreases (Fig. 5.8). The Stokes number slightly increases as the vortex
finder length is increased (cyclone S1 - S4). The insignificant change of
Stokes number is quite reasonable because of very limited changes in the
flow pattern (cf. Fig. 5.7). The small increase in the Stokes number can be
explained with the aid of the contours plots given in Fig. 5.6. The zone of
peak axial velocity increases with increasing S. Consequently, the possi-
bility of carrying bigger particles to escape with the upward flow slightly
increased. Cyclone S5 depicts a sudden reduction in the Stokes number
104
5.3. Results and discussions
values in comparison with cyclone S4. This drop is due to separation space
increase in cyclone S5, and the possibility of particles to be captured in-
creases. Although, the vortex strength decreased with a small amount due
to the reduction in the maximum tangential velocity.
Figure 5.9 presents the variation of the cyclone performance parameters
with the ratio Dx/S. It is clear that the effect of changing Dx/S depends
on the variables (Dx or S). Consequently, there are two curves per perfor-
mance parameter (Euler number and Stokes number). The common point
in Fig. 5.9 is cyclone D5 (=S1).
Dx /S
Eul
ernu
mbe
r
Stk
50x
103
0.6 0.8 13
4
5
6
7
8
9
10
11
12
0.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6Euler number (Dx)
Stk50 x 103 (Dx)
Euler number (S)
Stk50 x 103 (S)
Figure 5.9: The variation of the Euler number and the Stokes number with the
ratio of DxS
(cyclone S5 is excluded).
The variation of the performance parameters with the ratio of (h−S)/D is
presented in Fig. 5.10 for cyclones S1 - S5. Increasing (h− S)/D decreases
both the Euler number and Stokes number. However, the values of the
performance parameters depend on the value of the barrel height as is
clear from the values for cyclone S5.
In order to obtain the Euler number-Stokes number relationship, Fig. 5.11
has been drawn. It indicates a general relationship (trend) between the
two dimensionless numbers irrespective to the geometrical parameters
values. Two second-order polynomials have been proposed by Elsayed and
Lacor [54, 56], Eqs. 5.1 and 5.2.
Equation 5.2 presents a good matching for the performance parameters
105
Chapter 5. The Vortex Finder Dimensions
(h-S)/ D
Eul
ernu
mbe
r
Stk
50x
103
0.5 0.625 0.75 0.875 13
3.2
3.4
3.6
3.8
4
2
2.1
2.2
2.3
2.4Euler number (S1-S4)
Stk50 x 103 (S1-S4)
Euler number (S5)
Stk50 x 103 (S5)
Figure 5.10: The variation of the Euler number and the Stokes number with the
ratio of h−SD
.
for cyclones D1 -D5. For smaller values of Euler number, there is un-
derestimation of the Stokes number if the Eq. 5.2 is applied. The differ-
ence between the two correlations can be referred to two reasons. Firstly,
Eq. 5.1 is based on the CFD simulations data for both Euler number and
Stokes number, whereas Elsayed and Lacor correlation [56] (Eq. 5.2) was
obtained from experimental data for Euler number values and Iozia and
Leith model for the Stokes number values. Secondly, Eq. 5.1 is limited
to only four geometrical parameters. The other three factors are fixed,
h = 1.5, S = 0.5 and Bc = 0.375. Consequently, Eq. 5.1 is not suitable
to fit the Euler number - Stokes number relationship for cyclones S1 - S5
because the values of S is away from the range of applicability.
Stk50 = 100.3533(log10(Eu))2−1.1645log10(Eu)−2.3198 (5.1)
Stk50 = 100.3016(log10(Eu))2−0.9479log10(Eu)−2.5154 (5.2)
5.4 Conclusions
Nine cyclones of different vortex finder dimensions (diameter and length)
have been simulated using the large eddy simulation (LES) methodology,
106
5.4. Conclusions
Euler number
Sto
kes
num
berx
10
3
3 4 5 6 7 8 9
1
1.5
2
2.5
3
Simulation
Correlation 1
Correlation 2
Figure 5.11: The variation of the Stokes number with the Euler number for cy-
clones D1–D5.
to study the effect of the vortex finder dimensions on the performance and
flow pattern. The following conclusions have been obtained.
• The maximum tangential velocity in the cyclone decreases with in-
creasing the vortex finder diameter. A negligible change is noticed
with increasing the vortex finder length.
• Very limited axial variations in the flow variables are reported with
changing the vortex finder dimensions for the same cyclone.
• Increasing the vortex finder length, makes a small change in both
the static pressure, axial and tangential velocity profiles. However,
decreasing the vortex finder diameter gradually changes the axial
velocity profile from the inverted W to the inverted V profile.
• Decreasing the cyclone vortex finder diameter, increases the maxi-
mum tangential velocity. The maximum tangential velocity approaches
asymptotically 1.589 times the inlet velocity when decreasing the vor-
tex finder diameter.
• The Euler number (dimensionless pressure drop) decreases with in-
creasing the vortex finder diameter Dx. Increasing the vortex finder
length S slightly increases the Euler number.
• The Stokes number increases with increasing the vortex finder di-
ameter, because the centrifugal force affecting particles attenuates
when the swirl intensity (maximum tangential velocity) decreases.
The Stokes number slightly increases as the vortex finder length is
107
Chapter 5. The Vortex Finder Dimensions
increased (cyclone S1 - S4). The insignificant change of Stokes num-
ber is quite reasonable because of very limited changes in the flow
pattern.
• The effect of changing Dx/S on the performance parameters depends
on the variables (Dx or S).
• Increasing (h − S)/D decreases both the Euler number and Stokes
number. However, the values of the performance parameters depend
on the value of the barrel height.
108
Chapter 6
The Inlet Dimensions
6.1 Introduction
The effects of cyclone inlet section dimensions on the cyclone performance
(pressure drop and cut-off diameter) have been reported in many articles.
Casal and Martinez-Benet [21] proposed the following empirical formula
for the dimensionless pressure drop (Euler number),
Eu = 11.3
(aD
bD
Dx
D
)2
+ 2.33 (6.1)
implying proportionality with the square of the inlet area. Ramachandran
et al. [139] on the other hand proposed,
Eu = 20
a b
D2x
SD
HD
hD
Bc
D
1/3
(6.2)
i.e. a linear relation with the inlet area. Iozia and Leith [84, 85] presented
a correlation to estimate the cut-off diameter d50 and found proportional-
ity to (a b)0.61. The importance of inlet dimensions becomes clearer after
the study of natural length (or vortex length) by several researchers, e.g.,
Alexander [1]. The cyclone has two spiral motions, outer and inner. In the
reverse flow cyclone, the outer vortex weakens and changes its direction
at a certain axial distance Ln from the vortex finder [29]. This distance
is usually called the turning length, natural length or vortex length of the
cyclone. The inlet area is one of the relevant parameters influencing the
natural length. Alexander [1] found that Ln decreased proportionally to
the inlet area (Ai = a b) but the opposite trend has been also reported [29].
109
Chapter 6. The Inlet Dimensions
Numerous studies have been performed for the effect of geometrical pa-
rameters on the flow pattern and performance [e.g., 15, 62, 102, 140, 184]
while the effect of cyclone inlet dimensions remained largely unexplored.
The articles investigating the effect of cyclone geometry report only briefly
on the effect of inlet section dimensions on the cyclone performance with-
out sufficient details about their effects on the flow pattern and velocity
profiles. A new trend is the use of multi-inlet cyclone [e.g., 103, 187, 195].
The effects of cyclone inlet on the flow field and performance of cyclone
separators have been numerically investigated by Zhao et al. [198]. They
compared the performance of two types of cyclones with the conventional
single inlet and spiral double inlets using the Reynolds stress turbulence
model. The results show that the new type cyclone separator with spiral
double inlet can improve the symmetry of gas flow pattern and enhance
the particle separation efficiency. While their finding is for double inlets
cyclone, it supports the importance of the effect of the inlet section dimen-
sions on the performance of cyclone separator. The significant effects of the
cyclone inlet dimensions on the cyclone performance have been acknowl-
edged in many articles [e.g., 5]. For two inlets cyclone separators, Zhao
[193] reported the possibility of increasing the cyclone efficiency without
significantly increasing the pressure drop by improving the inlet geometry
of the cyclone. The effect of inlet section angle has been tested by many re-
searchers. Qian and Zhang [137] computationally investigated the effect of
the inlet section angle. The pressure drop of the cyclone decreases to a 30%
lower value than that for conventional cyclone, if the inlet section angle θbecomes 45 . However, Qian and Wu [136] reported only 15% reduction in
the pressure drop for θ = 45 .
In summary, all articles mentioned above did not study the effect of the
inlet height or width dimensions on the performance and flow pattern but
they studied the effect of the inlet configurations (inclined instead of tan-
gential), or the effect of the number of inlets (single or double) or the shape
of the inlet section (rectangular duct or nozzle). The present study is in-
tended to computationally investigate the effect of increasing the cyclone
inlet width and height on the pressure drop and cut-off diameter and ob-
taining more details about the flow field pattern and velocity profiles. The
study will be done using RANS and the Reynolds stress turbulence model
(RSM) will be used as a closure turbulence model.
110
6.2. Numerical settings
6.2 Numerical settings
6.2.1 Configuration of the five cyclones
The numerical simulations were performed on five cyclones with different
inlet dimensions. Figure 8.1 and Table 6.1 give the cyclones dimensions.
Table 6.2 gives more details for the used cyclones including the number of
cells, geometric swirl number, cyclone volume, flow residence time and the
inlet velocity for each cyclone.
In swirling flow, the swirl number usually characterizes the degree of swirl.
In cyclone separators, the swirling flow is characterized by the geometric
swirl number. The geometric swirl number Sg is a measure for the ratio of
tangential to axial momentum [75, 154], defined by [74],
Sg =π Dx D
4 Ain(6.3)
where Dx is the vortex finder diameter, D is the cyclone body diameter, and
Ain is the inlet cross-sectional area. For industrial cyclones, the geomet-
rical swirl number usually varies between 1 and 5 [74]. Table 6.2 shows
that, the tested cyclones cover this range. Implying that the obtained re-
sults can be applied to the industrial cyclones.
6.2.2 Boundary conditions
A velocity inlet boundary condition is used at the cyclone inlet, i.e., a veloc-
ity normal to the inlet is specified. An outflow boundary condition is used
at the outlet. The no-slip boundary condition is used at the other bound-
aries. The air volume flow rate Qin=50 l/min for all cyclones, air density
1.0 kg/m3 and dynamic viscosity 2.11E-5 Pa s. The turbulent intensity
equals 5% and the characteristic length equals 0.07 times the inlet width
[74].
6.2.3 Selection of the time step
The average residence time in the cyclone is determined from the cyclone
dimensions and gas flow rate [43]. The residence time tres = V/Qin where
V is the cyclone volume and Qin is the gas flow rate. This value is used
to select the time step. The time step for the unsteady simulation should
be a tiny fraction of the average residence time [23]. The tres ≈ 0.08s(for all tested cyclones) as shown in Table 6.2. So a time step of 5E-4 is
111
Chapter 6. The Inlet Dimensions
Table 6.1: The geometrical dimensions of the tested cyclones§
Dimension Cyclone∗ dimension/D b/a
Gas outlet diameter, Dx 0.5Vortex finder insertion length, S 0.5Cone tip-diameter, Bc 0.375Cylinder height, h 1.5Cyclone height, Ht 4.0
Inlet height, a b/D=0.2625A1 0.25 1.05A2 0.375 0.7A3 0.50 0.525
Inlet Width, b a/D=0.375B1 0.15 0.4B2 0.2625 0.7B3 0.375 1.0
§ Body diameter, D = 31 mm. The outlet section is above the cylindrical barrel surface by Le = 0.5D.The inlet section located at a distance Li = D from the cyclone center.
∗ Cyclones A2 and B2 are identical.
an acceptable value for the current simulation for accurate results and
achieve scaled residuals less than 1e-5 for all variables. The selected time
step results in an average inlet Courant number of 29.25, 28.88, 21.67,
30.40, 21.45 for cyclones A1, A2, A3, B1 and B3 respectively. However,
as the solver is a segregated implicit solver, there is no limitations on the
Courant number for stability.
6.2.4 CFD grid
Figure 6.1(b) shows the surface grid of cyclone A2 used in this study. The
hexahedral computational grids were generated using GAMBIT grid gen-
erator.
A grid independency study has been performed for the five tested cyclones.
Three grid levels for each cyclone have been tested, to be sure that the
obtained results are grid independent. For example, for cyclone A2 (B2)
three levels of meshes with respectively 490164, 714029 and 1174029 cells
have been used. The computational results are presented in Table 6.3. As
seen, the maximum difference between the results is less than 5%, so the
grid with 490164 cells can be considered as adequate [146]. It has been ob-
Table 6.2: The details of the five tested cyclones
Cyclone A1 A2=B2 A3 B1 B3
Number of cells∗ 705088 714029 820362 706370 816714
Sg 5.984 3.989 2.992 6.981 2.793
Cyclone volume x105 [m3] 6.878 6.95 7.017 6.875 7.012
tres [s] 0.0825 0.0834 0.0842 0.0825 0.0841
Inlet velocity [m/s] 13.214 8.809 6.607 15.416 6.166
∗ The total number of hexahedral cells after the grid independence study
112
6.3. Results
!
"
#
"
Sa
h
D
b
Ht
Dx
L i
Le
Bc
(a) The cyclone geometry (b) The surface mesh for cyclone A2 (B2)
Figure 6.1: Schematic diagram and surface mesh for the cyclone separator
served that even 490164 grid provides a sufficient grid independency. How-
ever, for excluding any uncertainty, computations have been performed us-
ing the 714029 cells grid, where the total number of grid points was not
that critical with respect to the computation overhead [10].
Table 6.3: The details of the grid independence study for cyclone A2
Number of cells Static pressure drop [N/m2] Cut-off diameter [µm]
490164 95.99 0.99714029 100.922 1.01174029 98.68 1.02% difference∗ 2.73 2.94
∗ The percentage difference between the coarsest and finest grid
6.3 Results
6.3.1 The axial variation of the flow properties
The tangential velocity is the dominant component of the gas flow in cy-
clones, which results in the centrifugal force for particle separation [186].
113
Chapter 6. The Inlet Dimensions
Moreover, the development of axial velocity profile in axial direction will
be analyzed for the five cyclones. Nine sections are used to plot the velocity
profiles as shown in Table 6.4.
Figures 6.2, 6.3 and 6.4 present the radial profiles of the time-averaged
static pressure, tangential and axial velocity at 9 axial stations. As ex-
pected, the tangential velocity profiles exhibit the so-called Rankine vor-
tex, which consists of two parts, an outer free vortex and an inner solid
rotation in the center (Fig. 6.3). The tangential velocity distribution in the
inner region is rather similar at different sections for the same cyclone. In
the outer region, due to the sharp drop in velocity magnitude in the near
wall region, the distribution is different and the change in the value of
maximum tangential velocity is rather limited. Generally, the tangential
velocity distribution varies only slightly with axial positions for the same
cyclone, which is also reported in other articles [e.g., 66, 127, 163, 186].
This means that, if the tangential velocity increases at one section of the
cyclone, it will increase at all other sections. The same conclusion can be
drawn from the radial profile of static pressure with higher values of pres-
sure drop expected for cyclone A1 in comparison with cyclones A2 and A3.
Cyclone B1 also depicts higher values of pressure in comparison with cy-
clones B2 and B3 (Fig. 6.2). The axial velocity profile has the shape of an
inverted W for all cyclones except B3 with the shape of an inverted V, as a
result of the change in the flow field pattern due to very wide inlet section.
Table 6.4: The position of different plotting sections†
Section S1 S2 S3 S4 S5 S6 S7 S8 S9
z`/D 2.75 2.5 2.25 2 1.75 1.5 1.25 1.0 0.75
† z`measured from the inlet section top
To evaluate the effect of increasing the cyclone inlet width on the tangen-
tial and axial velocity profiles, the tangential and axial velocity profiles at
section S9 (close to the inlet section) for the three cyclones (both for A cy-
clones and B cyclones) are compared in Fig. 6.5. As is clear from Fig. 6.5
the variation of axial velocity close to the wall is limited when changing
the inlet width or height. The axial velocity profiles for the three cyclones
are very similar except at the central region. The most important is the ef-
fect of cyclone inlet height or width on the tangential velocity (proportional
to the centrifugal force, which is the main force in the separation process).
Increasing the cyclone inlet width or height decreases the maximum tan-
gential velocity. Cyclone A1 and B1 have the maximum tangential velocity
in comparison with other cyclones. This means that decreasing the cyclone
inlet dimension will enhance the collection efficiency.
114
6.3. Results
6.3.2 The flow pattern
Figure 6.6 shows the contour plots of the time-averaged static pressure,
tangential and axial velocity for cyclones A1, A2 and A3. It is observed
that, the highest value of the static pressure decreases with increasing the
inlet height. The tangential velocity pattern is very similar in all cyclones
(Rankine profile). The highest value decreases with increasing the inlet
height, so that a better collection efficiency can be expected when decreas-
ing the inlet height. The axial velocity patterns for the three cyclones have
the shape of an inverted W profile.
Figure 6.7 shows the contour plots of the time-averaged static pressure,
tangential and axial velocity for cyclones B1, B2 and B3. It can be seen
that, the non-symmetry of the flow increases with increasing the inlet
width. The highest value of the static pressure decreases with increas-
ing the inlet width. The tangential velocity pattern is very similar for all
cyclones (Rankine profile). The highest value decreases with increasing
the inlet width, so that a better collection efficiency can be expected when
decreasing the inlet width. The axial velocity patterns for cyclones B1 and
B2 have the shape of an inverted W profile while that of cyclone B3 has an
inverted V profile. In cyclone B3, the inlet width (b/D=0.375) is wider than
the gap between the cyclone barrel and the vortex finder wall. As a result,
some part of the incoming flow will impact the vortex finder and hence will
not experience any swirling motion around the vortex finder. This results
in different axial velocity profile in cyclone B3 in comparison with the other
two cyclones. Furthermore, this will cause excessive stresses on the vortex
finder, vibrations and noise.
From the inspection of Fig. 6.7 (also Fig. 8.2 page 152, Fig. 8.15 page 181 &
Fig. 8.24 page 199), two observations can be made: the flow is asymmetric
near the cone bottom and there is a large variation in the time-averaged
axial velocity in this region. These observations are also found in the liter-
ature; for example, Gronald and Derksen [69, Fig. 5], Kaya and Karagoz
[92, Fig. 8] and Wang et al. [178, 179, Fig. 6]. However, no explana-
tion for this flow pattern is given. The author thinks this flow pattern can
be explained as follows: (1) The cyclone has only one inlet; consequently,
the flow should be asymmetric. However, this asymmetry should diminish
far away from the inlet section. (2) Another effect comes from the cyclone
bottom, which is closed in these simulations (no flow from the cyclone bot-
tom). As a consequence of this boundary condition, the descending flow
reverses its direction at the cone bottom resulting in a large asymmetry
115
Chapter 6. The Inlet Dimensions
A1 B1
Radial position [-]
Sta
ticpr
essu
re[N
/m2 ]
-1 -0.5 0 0.5 1-50
0
50
100
150
200
250
300
S1S2S3S4S5S6S7S8S9
Radial position [-]
Sta
ticpr
essu
re[N
/m2 ]
-1 -0.5 0 0.5 1-50
0
50
100
150
200
250
300
350
S1S2S3S4S5S6S7S8S9
A2 B2
Radial position [-]
Sta
ticpr
essu
re[N
/m2 ]
-1 -0.5 0 0.5 1-50
0
50
100
150
200
S1S2S3S4S5S6S7S8S9
Radial position [-]
Sta
ticpr
essu
re[N
/m2 ]
-1 -0.5 0 0.5 1-50
0
50
100
150
200
S1S2S3S4S5S6S7S8S9
A3 B3
Radial position [-]
Sta
ticpr
essu
re[N
/m2 ]
-1 -0.5 0 0.5 1-50
0
50
100
150
200
S1S2S3S4S5S6S7S8S9
Radial position [-]
Sta
ticpr
essu
re[N
/m2 ]
-1 -0.5 0 0.5 1-25
0
25
50
75
100
125
S1S2S3S4S5S6S7S8S9
Figure 6.2: The radial profile for the time-averaged static pressure at different
sections. Note: A2=B2.116
6.3. Results
A1 B1
A2 B2
A3 B3
Figure 6.3: The radial profile for the time-averaged tangential velocity at different
sections. Note: A2=B2.117
Chapter 6. The Inlet Dimensions
A1 B1
A2 B2
A3 B3
Figure 6.4: The radial profile for the time-averaged axial velocity at different sec-
tions. Note: A2=B2.118
6.3. Results
Radial position [-]
Sta
ticpr
essu
re[N
/m2 ]
-1 -0.5 0 0.5 1-50
0
50
100
150
200
250
300a/D=0.25a/D=0.375a/D=0.5
Radial position [-]
Sta
ticpr
essu
re[N
/m2 ]
-1 -0.5 0 0.5 1-50
0
50
100
150
200
250
300
350b/D=0.15b/D=0.2625b/D=0.375
Figure 6.5: Comparison between the radial profiles for the time averaged static
pressure, tangential and axial velocity at section S9.
119
Chapter 6. The Inlet Dimensions
at this region. The author believes that the extent of asymmetry is not
just a function of the cyclone geometry and boundary condition at the cone
bottom, but also the operating condition like the gas flow rate. This asym-
metry causes the obtained axial velocity variations (spot-like distribution
close to the cone bottom). (3) Another reason can be the boundary condi-
tion at the gas outlet (vortex finder exit). A fully developed flow (outflow
boundary condition) is assumed. Actually, there is no guarantee that the
flow is fully developed. May be the flow is still developing throughout a
long distance after the vortex finder exit. However, Wang et al. [178] re-
ported that the effect of the gas outlet length on the simulated velocity
becomes insignificant after a distance of 1/2 the cyclone diameter.
In spite of the above-mentioned interpretations, the following parameters
need to be investigated to accurately explain the reason of this flow behav-
ior: (i) The effect of including the dustbin in these simulations. (ii) The ef-
fect of the gas outlet tube length. (iii) The effect of the operating condition;
e.g., the gas Reynolds number (it includes the effect of inlet velocity, gas
density, and viscosity), the operating pressure, and temperature. More-
over, it may be worthwhile to perform these investigations experimentally
(using the particle image velocimetry (PIV) technique or any other flow
visualizations technique) to check also if these phenomena (observations)
come only from the numerical solution (schemes, settings, boundary con-
ditions, etc.).
6.3.3 The cyclone performance
In order to estimate the effect of cyclone inlet dimensions on the perfor-
mance parameters, the pressure drops have been calculated and presented
in Fig. 6.8. A discrete phase modeling (DPM) study has been performed by
injecting 104 particles from the inlet surface with a particle density of 860
kg/m3 and with a particle size ranging from 0.025 until 5 micron.
6.3.3.1 The effect of the inlet height
Figure 6.8 shows a rapid decrease in the pressure drop when increasing
the inlet height for 0.25 ≤ a/D ≤ 0.4 and a smaller decrease for a/D ≥ 0.4.
This behavior can be explained as follows. The pressure drop in the cyclone
is composed of three main contributions: (1) the pressure drop at the inlet
section (decreased by increasing the inlet dimensions). (2) the pressure
drop in the cyclone body due to swirling motion and due to wall friction,
this contribution decreases with increasing the cyclone inlet height (as the
vortex strength will decrease). (3) the main contribution to the cyclone
120
6.3. Results
A1 A2 A3
Th
est
ati
cp
ress
ure
N/m
2T
he
tan
gen
tia
lvelo
city
m/s
Th
ea
xia
lvelo
city
m/s
Figure 6.6: The contour plots for the time averaged flow variables at sections Y=0
and throughout the inlet section. From top to bottom : the static pressure N/m2,
the tangential velocity m/s and the axial velocity m/s. From left to right cyclone A1
through cyclone A3.
121
Chapter 6. The Inlet Dimensions
B1 B2 B3
Th
est
ati
cp
ress
ure
N/m
2T
he
tan
gen
tia
lvelo
city
m/s
Th
ea
xia
lvelo
city
m/s
Figure 6.7: The contour plots for the time averaged flow variables at sections Y=0
and throughout the inlet section. From top to bottom : the static pressure N/m2,
the tangential velocity m/s and the axial velocity m/s. From left to right cyclone B1
through cyclone B3.
122
6.3. Results
pressure drop is the energy loss in the exit tube, which mainly depends on
the maximum tangential velocity in the cyclone. As is clear from Fig. 6.5
the maximum tangential velocity decreases with increasing cyclone inlet
height. Consequently, in general the total cyclone pressure drop will de-
crease with increasing the cyclone inlet height. Figure 6.8 also depicts
the pressure drop for the three cyclones using four different mathemati-
cal models; Muschelknautz method of modeling (MM) [29, 175], Casal and
Martinez-Benet [21], Shepherd and Lapple [157], Stairmand [166] indicat-
ing the same overall trend obtained using RSM simulations for the pres-
sure drop with better matching between the MM model and the current
RSM simulation.
The effect of the cyclone inlet height on the cut-off diameter (particle di-
ameter of 50% collection efficiency) is shown in Fig. 6.8. The general trend
is an increase of the cut-off diameter with increasing cyclone inlet height,
due to weakness of the vortex. Figure 6.8 also depicts the cut-off diameter
for the three cyclones using two mathematical models; Iozia and Leith [85]
and Rietema [142] indicating the same overall trend obtained using RSM
simulations for the cut-off diameter with exact matching between the cur-
rent RSM results and Rietema model.
6.3.3.2 The effect of the inlet width
Figure 6.8 shows a rapid decrease in the pressure drop when increasing the
inlet width for 0.15 ≤ b/D ≤ 0.27 and a smaller decrease for b/D ≥ 0.27.
Generally, both the cyclone pressure drop and the cut-off diameter de-
crease with increasing the cyclone inlet width. The used mathematical
models again indicate the same overall trend obtained using RSM simula-
tions for cut-off diameter and pressure drop.
Figure 6.8 indicates the need of applying a multi-objective optimization
procedure to get the optimum value for the inlet dimensions. Both the col-
lection efficiency (cut-off diameter) and the pressure drop in cyclone sep-
arator are important objective functions to be optimized simultaneously
[146]. The effects of changing the cyclone inlet dimensions on pressure
drop and collection efficiency are opposite. Increasing the inlet width will
save more driving power but leads to reduced collection efficiency (larger
cut-off diameter). From the graph, the optimum value will be close to b/D= 0.25 with large reduction in pressure drop (energy losses) and small in-
crease in cut-off diameter. Elsayed and Lacor [50] estimated the optimum
value for inlet width (b/D) equals 0.236.
123
Chapter 6. The Inlet Dimensions
6.3.3.3 The inlet height versus the inlet width
Figure 6.9 shows that, the impact of changing either the inlet height or
width on the pressure drop is almost the same. However, the effect of
changing the inlet width on the cut-off diameter is more significant in
comparison with that of the inlet height. When plotting the pressure drop
against the ratio of inlet width to inlet height, it becomes clear the opti-
mum range of b/a is from 0.5 until 0.7.
6.4 Conclusions
Five cyclones of different inlet width and height have been simulated, us-
ing the Reynolds stress model (RSM), to study the effect of cyclone inlet
dimensions on the cyclone separator performance and flow pattern. The
following conclusions have been obtained.
• The maximum tangential velocity in the cyclone decreases with in-
creasing both the cyclone inlet width and height.
• No acceleration occurs in the cyclone space (the maximum tangential
velocity nearly constant throughout the cyclone). The axial variation
of both the static pressure and axial velocity is very limited.
• Increasing the cyclone inlet width or height decreases the pressure
drop at the cost of increasing the cut-off diameter. So an optimiza-
tion procedure is needed to estimate the optimum value of inlet di-
mensions.
• Wider inlet cyclones (b/D > gap between the cyclone barrel and the
vortex finder are not preferred.
• The effect of changing the inlet width on the cut-off diameter is more
significant in comparison with that of the inlet height.
• The optimum ratio of the inlet width to the inlet height b/a is from
0.5 to 0.7.
124
6.4. Conclusions
a/D
Pre
ssu
red
rop
[N/m
2]
Cu
t-o
ffd
iam
ete
r[m
icro
n]
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.550
100
200
300
400
500
600
700
800
900
1000
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2RSMMMCasalShepherdStairmandRSMIoziaRietma
Pressure drop
Cut-off diameter
b/D
Pre
ssu
red
rop
[N/m
2]
Cu
t-o
ffd
iam
ete
r[m
icro
n]
0.1 0.15 0.2 0.25 0.3 0.35 0.40
100
200
300
400
500
600
700
800
900
1000
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2RSMMMCasalShepherdStairmandRSMIoziaRietma
Pressure drop
Cut-off diameter
Figure 6.8: The effect of inlet dimensions on the pressure drop and cut-off diameter
using CFD simulations and different mathematical models.
125
Chapter 6. The Inlet Dimensions
Inlet height a/D
Inlet width b/DP
ressu
red
rop
[N/m
2]
Cu
t-o
ffd
iam
ete
r[m
icro
n]
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55
0.15 0.2 0.25 0.3 0.35 0.4
0
100
200
300
400
500
600
700
800
900
1000
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2Effect of inlet heightEffect of inlet width
Pressure drop
Cut-off diameter
b/a
Pre
ssu
red
rop
[N/m
2]
Cu
t-o
ffd
iam
ete
r[m
icro
n]
0.2 0.4 0.6 0.8 1 1.20
100
200
300
400
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6Pressure drop [N/m 2]
Cut-off diameter [micron]
Figure 6.9: Comparison between the effect of inlet height and width on the pres-
sure drop and cut-off diameter using CFD simulations.
126
Chapter 7
The Cyclone Height
7.1 Introduction
Numerous studies have been performed for the effect of geometrical pa-
rameters on the flow pattern and performance [15, 52, 62, 102, 140, 184]
whereas only limited number of studies have been devoted to the effect
of the cyclone height. Zhu and Lee [200] have conducted detailed exper-
iments on cyclones of different height and found that the cyclone height
can influence considerably the separation efficiency of the cyclones. How-
ever, they did not provide any information about the flow pattern nor an
explanation for the efficiency results. Hoffmann et al. [76] investigated the
effect of the cyclone height on the separation efficiency and the pressure
drop experimentally and theoretically. They found improvement in cyclone
performance with increasing the total height Ht up to 5.5 times cyclone
diameters beyond this length the separation efficiency was dramatically
reduced. However, they did not present any contour plot or velocity profile
to support the explanation for the effect of the cyclone height on the perfor-
mance. Recently, Xiang and Lee [186] have repeated the study of Zhu and
Lee [200] for the effect of cyclone height via steady three-dimensional sim-
ulations using the Reynolds stress turbulence model (RSM). They found
that the tangential velocity decreases with increasing the cyclone height,
which is responsible for the lower separation efficiency observed in long
cyclones. The reason for this behavior however, was not analyzed in detail.
Moreover, no particle tracking study was presented.
The present study is intended to computationally investigate the effect of
increasing cyclone (barrel and cone) height on the pressure drop and cut-
off diameter and to obtain more details about the flow field pattern and
velocity profiles. A RANS approach with the Reynolds stress turbulence
127
Chapter 7. The Cyclone Height
(a) The cyclone geometry (b) The surface mesh for cyclone
C1 (B2)
Figure 7.1: Schematic diagram for the cyclone separator
model (RSM) is used.
7.2 Numerical settings
7.2.1 Configuration of the tested cyclones
The numerical simulations were performed on four cyclones with different
barrel heights (at constant cone height) and three cyclones with different
cone heights (at constant barrel height). Figure 7.1 and Table 7.1 give the
cyclones dimensions.
7.2.2 Boundary conditions
Table 7.2 shows the details of the boundary conditions. The air volume
flow rate Qin=50 L/min for all cyclones, air density 1.0 kg/m3 and dynamic
viscosity of 2.11E-5 Pa s. The turbulent intensity equals 5% and charac-
teristic length equals 0.07 times the inlet width [52, 74].
128
7.2. Numerical settings
Table 7.1: The geometrical dimensions of the tested cyclonesa
Dimension Cyclone b dimension/D hc/h Htc
Inlet height, a 0.375Inlet width, b 0.2625Gas outlet diameter, Dx 0.5Vortex finder insertion length, S 0.5Cone tip-diameter, Bc 0.375
Cone height, hc h/D = 1.5C1 2.5 1.666 4.0C2 3.5 2.333 5.0C3 4.5 3.0 6.0
Barrel height, h hc/D = 2.5
B1 1.0 2.5 3.5B2 1.5 1.666 4.0B3 2.0 1.25 4.5B4 2.5 1.0 5.5
a Body diameter, D = 31 mm. The outlet section is above the cylindrical barrel surface by Le = 0.5D.The inlet section located at a distance Li = D from the cyclone center.
b The cone height is hc = Ht − h, where Ht is the total cyclone height. Cyclone C1 and B2 areidentical.
c Cyclone C2 and cyclone B4 are equal in Ht/D but they are different in both h/D and hc/D.
Grid independency study
A grid independency study has been performed for the tested cyclones.
Three different grids have been tested for each cyclone, to be sure that the
obtained results are mesh independent. For example, meshes with respec-
tively 490164, 714029 and 1174029 cells have been used for cyclone B2.
The computational results on the three grids are presented in Table 7.4.
As it can be seen the maximum difference between the results is less than
5%, so the results on the 490164 cells grid can already be considered as suf-
ficiently accurate. However, for excluding any uncertainty, computations
have been performed using the 714029 cells grid, where the total number
of grid points was not that critical with respect to the computation over-
head [10, 52]. Figure 7.1(b) shows the surface grid of cyclone C1(B2) used
in this study. Table 7.3 gives the total number of cells used for each cyclone
after the grid independency study. The hexahedral grids were generated
using the GAMBIT grid generator.
Table 7.2: The boundary conditions
Boundary Inlet Outlet Cone tip Other surfaces
Condition Velocity inlet Outflow Wall (no-slip) Wall (no-slip)
129
Chapter 7. The Cyclone Height
Table 7.3: The details of the tested cyclonesa
Cyclone B1 B2 (C1) B3 B4 C2 C3
Number of cells 688170 714029 712183 786865 770556 820362
a The total number of hexahedral cells after the grid independency study
7.3 Results
The flow velocity can be decomposed into three components. The tangen-
tial and the axial velocity components are the major velocity components
in comparison with the radial velocity component. Xiang and Lee [186]
stated that the tangential velocity is the dominant gas velocity in gas cy-
clones, which results in the centrifugal force for particle separation. The
axial component is responsible for the two flow streams (downward and
upward).
7.3.1 The axial variation
Figures 7.2 - 7.4 present the radial profiles of the time-averaged tangential
and axial velocity and static pressure at nine axial stations (cf. Table 7.5).
As expected, the tangential velocity profiles exhibit the so-called Rankine
vortex, which consists of two parts, an outer free vortex and an inner solid
rotation in the center (Fig. 7.2). The tangential velocity distribution in the
inner region is rather similar at different sections for the same cyclone.
In the outer region, due to the sharp drop in velocity magnitude in the
near wall region, the distribution is different but the maximum tangential
velocity is similar at all sections for the same cyclone. The axial velocity
profiles at nine different stations are shown in Fig. 7.3. All cyclones show
the inverted W profile but the central dip decreases with increasing the
Table 7.4: The details of the grid independence study for cyclone C1 (B2)
Number of cells Eua Stk50x103b
490164 3.475 1.188714029 3.654 1.21174029 3.573 1.224% differencec 2.74 2.94
a Euler number is the dimensionless pressure drop Eu = ∆P/( 12ρV 2
in) where ∆P is the static pressuredrop, ρ is the gas density, Vin is the gas inlet velocity.
b The Stokes number based on the cut-off diameter; Stk50 = ρpx250Vin/(18µD) [37]. It is the ratio
between the particle relaxation time; ρpx250/(18µ) and the gas flow integral time scale; D/Vin where ρp
is the particle density =860 kg/m3, µ is the gas viscosity.
c The percentage difference between the coarsest and finest grid
130
7.3. Results
Table 7.5: The position of different plotting sections
Section S1 S2 S3 S4 S5 S6 S7 S8 S9
z`/Da 2.75 2.5 2.25 2 1.75 1.5 1.25 1.0 0.75
a z`measured from the inlet section top (cf. Fig. 7.1(a)).
barrel (or cone) height (Except at the cyclone bottom for cyclones C2 and
C3, where the axial velocity shows an inverted V profile). The radial pro-
files of the time averaged static pressure are given in Fig. 7.4. Like for
the tangential velocity, the axial variations of the static pressure are very
small for the same cyclone. The variations become negligible with increas-
ing barrel (or cone) height. Furthermore, the maximum value of the static
pressure decreases when the barrel height is increased.
7.3.2 The radial variation
The tangential and axial velocity profiles at section S6 (as a representative
for the other sections, because the axial variations in the flow variables are
small) for the six cyclones are compared in Fig. 7.5. The variation of axial
velocity with changing barrel height is limited close to the wall, especially
in the cylindrical part. In the central region, the change in axial velocity
profile is more pronounced with the dip in the axial velocity decreasing
with increasing the barrel height. This is the result of the flatting of the
pressure distribution results in a smaller pressure force. This may explain
also the change of the axial velocity from cyclones B1 to B4. Increasing the
cyclone barrel height decreases the maximum tangential velocity. Cyclone
B1 has the maximum tangential velocity in comparison with the other cy-
clones. The effect of increasing the barrel height on the maximum tan-
gential velocity is limited. The variation of axial velocity in the three C
cyclones is limited close to the wall with changing the cone height. The
axial velocity profiles are very similar except at the central region due to
change in the axial velocity profile. Increasing the cyclone cone height de-
creases the maximum tangential velocity. Cyclone C1 has the maximum
tangential velocity in comparison with the two other cyclones (C2 and C3).
From the comparison between the radial profiles of the four B cyclones,
the minimum pressure at the cyclone center is almost the same for all cy-
clones (B1 - B4). The slope of the static pressure radial profile becomes
flatter with increasing the barrel height. The minimum pressure at the
cyclone center is almost the same for all C cyclones. The static pressure
radial profiles of cyclones C2 and C3 are very close. Increasing the cyclone
131
Chapter 7. The Cyclone Height
C1 (B2) B1
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-2
0
2
4
6
8
10
12
S1S2S3S4S5S6S7S8S9
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.0150
2
4
6
8
10
12
S1S2S3S4S5S6S7S8S9
C2 B3
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-2
0
2
4
6
8
10
12
S1S2S3S4S5S6S7S8S9
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.0150
2
4
6
8
10
12
S1S2S3S4S5S6S7S8S9
C3 B4
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-2
0
2
4
6
8
10
12
S1S2S3S4S5S6S7S8S9
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.0150
2
4
6
8
10
12
S1S2S3S4S5S6S7S8S9
Figure 7.2: The radial profile for the time-averaged tangential velocity at different
sections. Note: C1=B2.
132
7.3. Results
C1 (B2) B1
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-4
-2
0
2
4
6
8S1S2S3S4S5S6S7S8S9
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-4
-2
0
2
4
6
8S1S2S3S4S5S6S7S8S9
C2 B3
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-4
-2
0
2
4
6
8S1S2S3S4S5S6S7S8S9
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-4
-2
0
2
4
6
8S1S2S3S4S5S6S7S8S9
C3 B4
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-4
-2
0
2
4
6
8S1S2S3S4S5S6S7S8S9
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-4
-2
0
2
4
6
8S1S2S3S4S5S6S7S8S9
Figure 7.3: The radial profile for the time-averaged axial velocity at different sec-
tions. Note: C1=B2.
133
Chapter 7. The Cyclone Height
C1 (B2) B1
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-50
0
50
100
150
200
S1S2S3S4S5S6S7S8S9
Radial position (m)
Sta
ticpr
essu
re(N
/m2)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-100
-50
0
50
100
150
200
250
300
S1S2S3S4S5S6S7S8S9
C2 B3
Radial position (m)
Sta
ticpr
essu
re(N
/m2)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-50
0
50
100
150
200
S1S2S3S4S5S6S7S8S9
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-100
-50
0
50
100
150
200
250
300
S1S2S3S4S5S6S7S8S9
C3 B4
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-50
0
50
100
150
S1S2S3S4S5S6S7S8S9
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-100
-50
0
50
100
150
200
250
300
S1S2S3S4S5S6S7S8S9
Figure 7.4: The radial profile for the time-averaged static pressure at different
sections. Note: C1=B2.
134
7.3. Results
height (either barrel or cone) decreases the pressure drop, the maximum
tangential velocity (vortex strength) and the dip in the axial velocity pro-
file. The effect of increasing the cone height on the axial velocity profile is
predominant with respect to the barrel height.
The swirling motion of the gas generates a strong radial pressure gradient,
the pressure being low in the centre of the vortex and high at the periph-
ery. As the strongly swirling gas enters the confines of the vortex finder
on its way out of the cyclone, the swirl is attenuated through friction with
the wall. This means that further up the vortex finder the pressure in the
centre is higher than at the exit of the separation space: a reverse pres-
sure gradient is present [78] as is clear from Fig.7.5. This drives an axial
flow with dip in the centre of the vortex finder (inverted W profile); this
core flow prevails throughout the entire separation space of the cyclone in
spite of the attenuation of swirl in the conical part of the cyclone. With
increasing the cone height the pressure distribution becomes flatter conse-
quently the pressure force causes the dip in the axial velocity at the center
line becomes less and less. That may explain also the change of the axial
velocity from cyclones C1 to C3, Figs. 7.5 and 7.6.
7.3.3 The flow pattern
Figure 7.6 shows the contour plots of the time-averaged static pressure,
tangential and axial velocity for cyclones C1-C3. The time-averaged static
pressure decreases radially from the wall to the center. A negative pres-
sure zone appears in the forced vortex region (central region) due to high
swirling velocity. The pressure gradient is largest along the radial direc-
tion, whereas the gradient in axial direction is very limited. The cyclonic
flow is not symmetrical as is clear from the shape of the low-pressure
zone at the cyclone center (twisted cylinder). Two vortical motions are
exist one moving down (outer vortex) and the other moving up (inner vor-
tex). The highest value of the static pressure decreases with increasing
the cone height. The tangential velocity pattern is very similar in all cy-
clones (Rankine profile). The highest value decreases with increasing the
cone height but the differences between cyclones C2 and C3 are small, so
that a better collection efficiency can be expected when decreasing the cone
height.
Figure 7.7 shows the contour plots of the time-averaged static pressure,
tangential and axial velocity for cyclones B1-B4. The highest value of the
static pressure decreases with increasing the barrel height. The tangen-
tial velocity pattern is very similar in all cyclones (Rankine profile). The
highest value decreases with increasing the barrel height, but the differ-
135
Chapter 7. The Cyclone Height
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-50
0
50
100
150
200
250hc/D=2.5hc/D=3.5hc/D=4.5
Radial position (m)S
tatic
pres
sure
(N/m
2 )-0.015 -0.01 -0.005 0 0.005 0.01 0.015
-100
-50
0
50
100
150
200
250
300
h/D=1.0
h/D=1.5
h/D=2.0
h/D=2.5
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-2
0
2
4
6
8
10
12
14hc/D=2.5hc/D=3.5hc/D=4.5
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.0150
2
4
6
8
10
12
h/D=1.0
h/D=1.5
h/D=2.0
h/D=2.5
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-4
-2
0
2
4
6
8hc/D=2.5hc/D=3.5hc/D=4.5
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-4
-2
0
2
4
6
8h/D=1.0
h/D=1.5
h/D=2.0
h/D=2.5
Figure 7.5: Comparison between the radial profiles for the time averaged static
pressure, tangential and axial velocity at section S6.
136
7.3. Results
ences between the four cyclones are small. The axial velocity patterns for
the four cyclones have the shape of an inverted W profile.
7.3.4 The performance
In order to estimate the effect of cyclone height on the performance pa-
rameters, the Euler number (the dimensionless pressure drop) have been
calculated. A discrete phase modeling (DPM) study has been performed by
injecting 104 particles from the inlet surface with a particle density of 860
kg/m3 and with a particle size ranging from 0.025 until 5 micron.
Figure 7.8 and Table 7.6 show a sharp decrease of the Euler number with
increasing the barrel height until h/D > 1.8 (Ht/D > 4.3) and a gradual
decrease beyond. This behavior can be explained as follows. The pres-
sure drop in the cyclone is composed of three main contributions [52]: (1)
the pressure drop at the inlet section. (2) the pressure drop in the cy-
clone body due to swirling motion and due to wall friction, this contribu-
tion may increase with increasing the cyclone height as the wall friction
will increase due to friction with a larger wall surface, or decreases as
the vortex strength will decrease because the maximum tangential veloc-
ity decreases. (3) the main contribution to the cyclone pressure drop is
the energy loss in the exit tube, which mainly depends on the maximum
tangential velocity in the cyclone. As is clear from Fig. 7.5 the maximum
tangential velocity decreases with increasing cyclone barrel height. As the
inlet section is the same in all cyclones, the pressure drop in the inlet sec-
tion does not vary with increasing barrel height. The sharp decrease of
the Euler number between cyclones B1 and B2 is due to the decrease in
the maximum tangential velocity. There are two competing contributions:
increase of the pressure drop due to friction and decrease of the pressure
drop due to the reduction in the vortex strength. At the beginning, the wall
friction effect is small in comparison with the effect of vortex strength. For
longer cyclones, this effect becomes larger (but still less than that of vortex
strength decay). This explains the small variation of the Euler number
with the barrel height for h/D > 1.8, which is clear from Fig. 7.5 where the
maximum tangential velocity of cyclones B3 and B4 are very close.
The behavior of the Stokes number curve as a function of barrel height
is quite reasonable with increasing barrel height (separation space), the
possibility of particles to be captured increases due to the increased cy-
clone space. However, the vortex strength decreased with a small amount
due to the reduction of the maximum tangential velocity, the main contri-
bution here is the collecting surface. The Stokes number curve becomes
137
Chapter 7. The Cyclone Height
Figure 7.6: The contour plots for the time averaged flow variables at sections Y=0
and throughout the inlet section. From left to right : the static pressure (N/m2), the
tangential velocity (m/s) and the axial velocity (m/s). From top to bottom cyclone
C1-C3.
138
7.3
.R
esu
lts
Figure 7.7: The contour plots for the time averaged flow variables at sections Y=0 and throughout the inlet section. From
top to bottom: the static pressure (N/m2), the tangential velocity (m/s) and the axial velocity (m/s). From left to right cyclone
B1 - B4.
139
Chapter 7. The Cyclone Height
nearly flat between cyclones B3 and B4 due to the small changes in both
the axial and the tangential velocity profile between the two cyclones. Both
Ramachandran [139] and Iozia [84] models agree with the CFD results in
the trend of decreasing both the Euler number and Stokes number with
increasing barrel height, but differ in slope and values.
Ht /D
Eul
ernu
mbe
r
Stk
50x
103
3.5 4 4.5 53
3.5
4
4.5
5
5.5
6
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4Euler number (CFD)Euler number (Ramachandran model)Stk50 x 103 (CFD)Stk50 x 103 (Iozia model)
Figure 7.8: The effect of barrel height on the Euler number and the Stokes number.
Table 7.6: The Euler numbers and Stokes numbers for cyclones B1-B4
Cyclone B1 B2 B3 B4
h/D 1.0 1.5 2.0 2.5
hc/h 2.5 1.666 1.25 1.0
Ht/D 3.5 4 4.5 5.0
Eu 4.39 3.654 3.33 3.09
Eu (Ramachandran model [139]) 5.71 4.77 4.17 3.73
Stk50 x 103 1.32 1.2 1.01 0.95
Stk50 x 103 (Iozia model [84]) 1.94 1.82 1.72 1.64
Figure 7.9 and Table 7.7 show a sharp decrease of both the Euler number
and the Stokes number with increasing the cone height until hc/D=3.3 and
a gradual decrease when 3.3 > hc/h > 4.0. This behavior can be explained
as follows. As the inlet section is the same in all cyclones, the pressure
drop in the inlet section may not vary with increasing the cone height.
140
7.3. Results
The sharp decrease in the Euler number between cyclones C1 and C2 is
due to the huge decrease in the pressure drop in the cyclone body due
to the drop in the maximum tangential velocity (the decay of the vortex
strength). There are two competing contributions, increase of the pressure
drop due to friction and decrease of the pressure drop due to the reduction
in the vortex strength. At the beginning, the decay in the pressure drop
due to the decay of the vortex strength overrides the effect of increasing
the pressure drop due to wall friction for longer cyclones. This explains
the small variation of the Euler number with the cone height for hc > 4.0,
which is clear from Fig. 7.5 where the maximum tangential velocities of
cyclones C2 and C3 are very close.
The trend of decreasing Stokes number with increasing cone height is
quite reasonable, as more separation space exists, and the possibility of
particles to be captured increases. Although, the vortex strength decreased
with a small amount - due to the reduction in the maximum tangential
velocity- we estimate that the main contribution to the collection efficiency
comes from the increased collecting surface with increasing the cone height.
The reason of nearly constant Stokes number after Ht/D =5.5 is the change
in the axial velocity profile. Figure 7.5 shows a higher kinetic energy of the
flow at the cyclone bottom for cyclone C3, that will enhance re-entrainment
of some of the captured particles to escape with the upward flow. Conse-
quently low collection efficiency and higher cut-off diameter (Stokes num-
ber). Because, the differences between the axial and tangential velocity
profile between cyclones C2 and C3 are limited.
Table 7.7: The Euler numbers and Stokes numbers for cyclones C1-C3
Cyclone C1 C2 C3
hc/h 1.666 2.333 3.0
hc/D 2.5 3.5 4.5
Ht/D 4 5 6
Eu 3.654 2.749 2.584
Stk50 x 103 1.2 0.465 0.315
7.3.5 The cone height versus the barrel height
• Increasing the cyclone total height (either by increasing the cone or
the barrel height) will decrease the maximum tangential velocity.
• The effect of cone height on the axial velocity profile is much larger
than that of the barrel height.
• Both the Euler and the Stokes numbers decrease with increasing the
141
Chapter 7. The Cyclone Height
hc /D
Ht /D
hc /hE
uler
num
ber
Stk
50x
103
2 2.5 3 3.5 4 4.5 5
3.5 4 4.5 5 5.5 6 6.5
1.5 2 2.5 3 3.5
1
1.5
2
2.5
3
3.5
4
0
0.5
1
1.5
2
2.5
3
3.5
4Euler number
Stk50 x 103
Figure 7.9: The effect of cone height on the pressure drop (Euler number) and
cut-off diameter (Stokes number).
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-4
-2
0
2
4
6
8hc /D=2.5hc /D=3.5hc /D=4.5
Figure 7.10: Comparison between the radial profiles for the time averaged static
pressure, tangential and axial velocity at section S9.
142
7.3. Results
total height either by increasing the barrel or cone height. The effect
of changing the cone height is more important than that of changing
the barrel height (cf. Table 7.6, 7.7 and Fig. 7.11).
• The effect of increasing the ratio of cone to barrel height hc/h on the
cyclone performance depends on the dependent variable. If hc/h in-
creases at constant cone height, the Euler number increases linearly
with decreasing the barrel height. If hc/h increases at constant barrel
height, the Euler number decreases with increasing the cone height.
• The effect of changing hc/h at constant cone height has a negligible
effect on the Stokes number.
• Increasing hc/h at constant barrel height decreases the Stokes num-
ber. This effect becomes negligible for hc/h > 2.75 (Fig. 7.11). This be-
havior can be explained by inspecting the variation of the time aver-
aged static pressure, tangential and axial velocity profiles with cone
and barrel height, Fig. 7.12. As is clear from Fig. 7.12 the differences
between the maximum tangential velocity in cyclones C2 and C3 are
negligible. Furthermore, the effect of the cone height on the flow field
is more significant than that of the barrel height.
143
Chapter 7. The Cyclone Height
Ht /D
Eul
ernu
mbe
r
Stk
50x
10
3
3 3.5 4 4.5 5 5.5 6 6.50
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6Euler number (Barrel height)Euler number (Cone height)Stk50 x 103 (Barrel height)Stk50 x 103 (Cone height)
hc /h
Eul
ernu
mbe
r
Stk
50
x1
03
0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.250
1
2
3
4
5
6
0
1
2
3
4
5
6Euler number (Barrel height)Euler number (Cone height)Stk50 x 103 (Barrel height)Stk50 x 103 (Cone height)
Figure 7.11: The variation of the Euler number and the Stokes number with the
barrel and cone height.
144
7.3. Results
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-100
-50
0
50
100
150
200
250
300
h/D=1.0
h/D=1.5
h/D=2.0
h/D=2.5
hc/D=3.5
hc/D=4.5
Radial position (m)
Sta
ticpr
essu
re(N
/m2 )
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-100
-50
0
50
100
150
200
250
300
h/D=1.0
h/D=1.5
h/D=2.0
h/D=2.5
hc/D=3.5
hc/D=4.5
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-2
0
2
4
6
8
10
12
h/D=1.0
h/D=1.5
h/D=2.0
h/D=2.5
hc/D=3.5
hc/D=4.5
Radial position (m)
Tan
gent
ialv
eloc
ity(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015-2
0
2
4
6
8
10
12
h/D=1.0
h/D=1.5
h/D=2.0
h/D=2.5
hc/D=3.5
hc/D=4.5
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015
-4
-2
0
2
4
6
8h/D=1.0
h/D=1.5
h/D=2.0
h/D=2.5
hc/D=3.5
hc/D=4.5
Radial position (m)
Axi
alve
loci
ty(m
/s)
-0.015 -0.01 -0.005 0 0.005 0.01 0.015
-4
-2
0
2
4
6
8h/D=1.0
h/D=1.5
h/D=2.0
h/D=2.5
hc/D=3.5
hc/D=4.5
Figure 7.12: The radial profile for the time averaged static pressure, tangential and
axial velocity at three different sections for the six cyclones. From top to bottom
: static pressure, tangential and axial velocity. From left to right: S6 - S9. Note:
h/D = 1.5 also represents hc/D = 2.5
145
Chapter 7. The Cyclone Height
7.4 Conclusions
Six cyclones of different barrel and cone height have been simulated using
the Reynolds stress model (RSM), to study the effect of cyclone height on
the performance and flow pattern. The following conclusions have been
obtained.
• The maximum tangential velocity in the cyclone decreases with in-
creasing the cyclone (barrel or cone) height.
• No acceleration occurs in the cyclone space (the maximum tangential
velocity nearly constant throughout the same cyclone).
• Increasing the barrel height, makes a small change in the axial ve-
locity.
• Increasing the cyclone barrel height decreases the pressure drop and
the cut-off diameter. The changes in the performance beyond h/D =1.8 are small.
• Increasing the cone height makes a considerable change in the axial
velocity.
• Both the pressure drop and the cut-off diameter decrease with in-
creasing the cyclone cone height. The performance improvement stops
after hc/D = 4.0 (Ht/D = 5.5).
• The effect of changing the barrel height is less significant on the per-
formance and the flow pattern in comparison with the effect of the
cone height.
146
Chapter 8
Optimization
8.1 Introduction
Mathematical optimization refers to the selection of a best element from
some set of available alternatives. In the simplest case, this means solving
problems in which one seeks to minimize or maximize a real function by
systematically choosing the values of real or integer variables from within
an allowed set. Generally, the use of the word optimization implies the
best result under the circumstances [176].
In 1951, Stairmand [166] presented one of the most popular design guide-
lines for the high-efficiency cyclone separators [147]. Stairmand presented
the geometrical ratios for the seven geometrical parameters as: a/D = 0.5,
b/D = 0.2, Dx/D = 0.5, Ht/D = 4.0, h/D = 1.5, S/D = 0.5 and Bc/D =0.375. These values have been obtained based on the Stairmand model
[165] which suffers from many shortages [50].
In the Stairmand model [165], the velocity distribution has been obtained
from a moment-of-momentum balance, estimating the pressure drop as
entrance and exit losses combined with the loss of static pressure in the
swirl. The main drawbacks of the Stairmand model are: (1) neglecting the
entrance loss by assuming no change of the inlet velocity occurs at the inlet
area; (2) assuming a constant friction factor; (3) the effect of the particle
mass loading on the pressure drop is not included. All these drawbacks
are overcome in the Muschelknautz method of modeling (MM) [77] intro-
duced by Muschelknautz and Trefz [116, 117]. The main benefit of MM
over other models is its ability to take the following effects into account: a)
wall roughness due to both the physical roughness of the materials of con-
struction and to the presence of collected solids. b) the effect of the mass
loading and the Reynolds number on cyclone performance. c) the change
147
Chapter 8. Optimization
of flow velocity throughout the cyclone [77]. Consequently, the optimiza-
tion procedure using a data obtained from MM model via response sur-
face methodology will result in better results than the simple Stairmand
model. An alternative approach is using the available experimental data
in literature using the artificial neural networks (ANNs) approach for the
optimization process. The third source of data is the CFD simulations to
train the ANNs.
A general unconstrained optimization problem can be expressed as [182]:
min fi(x), i = 1, 2, . . . , Isubject to
x = [x1, x2, . . . , xn]T ,x ∈ X
(8.1)
where x = [x1, x2, . . . , xn]T is the vector of design variables (cyclone dimen-
sions) with total number of n, X denotes the design space. The fi(x) stands
for objective functions with total number of I. I should be larger than or
at least equal to 1, which suggests the multi-objective or single objective
optimal problem. In Eq. 8.1, f1(x) is the Euler number and f2(x) is the
cut-off diameter.
The design variables are seven geometrical parameters, namely, Dx/D,
a/D, b/D, Ht/D, h/D, S/D and Bc/D. The side constraints are: 0.3 ≤Dx/D ≤ 0.65, 0.2 ≤ a/D ≤ 0.65, 0.15 ≤ b/D ≤ 0.3, 3 ≤ Ht/D ≤ 6, 0.75 ≤h/D ≤ 2, 0.4 ≤ S/D ≤ 0.75 and 0.1 ≤ Bc/D ≤ 0.45.
There are numerous optimization techniques that can be used for the ge-
ometry optimization in the cyclone separator. However, only two tech-
niques have been used in this thesis (Nelder-Mead technique and the ge-
netic algorithms (GA)) for robustness, and availability in the used soft-
ware. In cases of multi-objective optimization studies, two approaches
have been applied, the desirability function and the non-sorted dominated
genetic algorithm (NSGA-II) techniques have been selected due to robust-
ness, and availability. More details about the used optimization techniques
are given in appendix B. A detailed literature review for the previous opti-
mization studies is given in Sec. 2.5.6, page 21.
The present study is an attempt to obtain new optimized cyclone separator
designs based on the MM model, experimental data sets and CFD simu-
lations data sets and to investigate the effect of each cyclone geometrical
parameter on the cyclone performance using response surface methodol-
ogy (RSM) and CFD simulations. Table 8.1 summaries the conducted opti-
mization studies.
148
8.2. Single-objective using MM model
Table 8.1: Summary of optimization studies
No.Design Objective Source Meta-model Optimization
variables functions of data technique
1 7 1 MM model RSM Nelder-Mead2 7 1 Experimental data (Exp.) ANN Nelder-Mead3 7 2 Exp. and Iozia and Leith model ANN GA and NSGA-II
4 4 2 CFD simulations data ANNDesirability function,
GA and NSGA-II
8.2 Optimization of the Euler number using
MM model and Nelder-Mead technique
The present study aims to obtain a new optimized cyclone separator for
minimum pressure drop based on the MM model and to investigate the
effect of each cyclone geometrical parameter on the cyclone performance
using CFD simulations. This section is an extension to the sensitivity
analysis study presented in Sec. 4.1, page 45.
!
"
#
"
Sa
h
D
b
Ht
Dx
L i
Le
Bc
Figure 8.1: Schematic diagram for Stairmand cyclone separator
Table 8.2 gives the optimum values for cyclone geometrical parameters
for minimum pressure drop estimated by MM using the downhill simplex
optimization technique available in Statgraphics XV software.
149
Chapter 8. Optimization
Table 8.2: The geometrical parameters for minimum pressure drop using MM
model
Factor Low High Optimum
a/D 0.5 0.75 0.618b/D 0.14 0.4 0.236Dx/D 0.2 0.75 0.622Ht/D 3 7 4.236h 1.0 2.0 1.618S 0.4 2 0.620Bc 0.2 0.4 0.382
8.2.1 CFD comparison between the two designs
Numerical settings
The air volume flow rate Qin=0.08 m3/s for the two cyclones (inlet velocity
for Stairmand design is 19 m/s and 13.1 m/s for the new design), air density
1.0 kg/m3 and dynamic viscosity of 2.11E-5 Pa s. The turbulent intensity
equals 5% and characteristic length equals 0.07 times the inlet width [75].
Velocity inlet boundary condition is applied at inlet, outflow at the gas
outlet and wall boundary condition at all other boundaries.
The finite volume method has been used to discretize the partial differen-
tial equations of the model using the SIMPLEC (Semi-Implicit Method for
Pressure-Linked Equations-Consistent) method for pressure velocity cou-
pling and QUICK scheme to interpolate the variables on the surface of
the control volume. The implicit coupled solution algorithm was selected.
The unsteady Reynolds stress turbulence model (RSM) was used in this
study with a time step of 0.0001 s. The residence time (cyclone volume/gas
volume flow rate) of the two cyclones are close (≈ 0.25 s).
The grid refinement study shows that a total number of about 134759 hex-
ahedral cells for Stairmand cyclone and 154746 hexahedral cells for the
new design are sufficient to obtain a grid-independent solution, and fur-
ther mesh refinement yields only small, insignificant changes in the nu-
merical solution. These simulations were performed on an eight nodes
CPU Opteron 64 Linux cluster using Fluent commercial software. The
geometrical values are given in Table 8.3 for the two cyclones (cf. Fig. 8.1).
Table 8.3: The values of geometrical parameters for the two designs (D=0.205 m)
Cyclone a/D b/D Dx/D Ht/D h/D S/D Bc/D Li/D Le/DStairmand design 0.5 0.2 0.5 4 1.5 0.5 0.36 1.0 0.618
New design 0.618 0.236 0.622 4.236 1.618 0.620 0.382 1.0 1.618
150
8.2. Single-objective using MM model
Table 8.4: The position of different sectionsa
Section S1 S2 S3 S4 S5 S6 S7
z’/Db 2.75 2.5 2.25 2.0 1.75 1.5 0.25
a Sections S1–S5 are located in the conical section, section S6 at the cylindrical part and S7 locatedthrough the inlet section.b z’ measured from the inlet section top
Results and discussion
The pressure field
Figure 8.2 shows the contour plot at Y=0 and at section S7 (at the middle of
inlet section, Table 8.4). In the two cyclones, the time-averaged static pres-
sure decreases radially from the wall to center. A negative pressure zone
appears in the forced vortex region (central region) due to high swirling
velocity. The pressure gradient is largest along the radial direction, while
the gradient in the axial direction is very limited. The cyclonic flow is not
symmetrical as is clear from the shape of the low pressure zone at the
cyclone center (twisted cylinder). However, the two cyclones have almost
the same flow pattern, but the highest pressure of the Stairmand design
is nearly twice that of the new design, implying that the new design has a
lower pressure drop.
The pressure distributions presented in Figs. 8.3 and 8.4 of the two cy-
clones at sections S1–S6 depict the two parts pressure profile (for Rankine
vortex). Once again, the highest static pressure for Stairmand design is
more than twice that of the new design at all sections while the central
value is almost the same for the two cyclones irrespective to the section
location. This indicates that, the new design has a lower pressure drop
with respect to the Stairmand design.
The velocity field
Based on the contour plots of the time–averaged tangential velocity, Fig. 8.2,
and the radial profiles at sections S1–S6 shown in Figs. 8.3 and 8.4, the
following comments can be drawn. The tangential velocity profile at any
section is composed of two regions, an inner and an outer one. In the
inner region, the flow rotates approximately like a solid body (forced vor-
tex), where the tangential velocity increases with radius. After reaching
its peak the velocity decreases with radius in the outer part of the profile
(free vortex). This profile is a so-called Rankine type vortex as mentioned
151
Chapter 8. Optimization
Figure 8.2: The contour plots for the time averaged flow variables at sections Y=0
and S7. From top to bottom: static pressure [N/m2], tangential velocity [m/s] and
axial velocity [m/s]. From left to right Stairmand design and new design respec-
tively.
152
8.2. Single-objective using MM model
Table 8.5: The performance parameters for the two cyclones
Method Eu [-] ∆p [N/m2] x50 [µm]
Stairmand designMM 5.79 1045 1.54CFD 6.592 1190 1.0
New designMM 5.24 450 1.77CFD 5.672 487 1.6
before, including a quasi-forced vortex in the central region and a quasi-
free vortex in the outer region. The maximum tangential velocity may
reach twice the average inlet velocity and occurs in the annular cylindrical
part. The tangential velocity distributions for the two cyclones are nearly
identical in pattern and values (dimensionless), with the highest velocity
occurring at 1/4 of the cyclone radius for both cyclones. This implies a
nearly equal collection efficiency for both cyclones, as the centrifugal force
is the main driving force for particle collection in the cyclone separator.
The axial velocity profiles for the two cyclones are also very close, exhibit-
ing a M letter shape (also known as inverted W axial velocity profile in
some other literatures (cf. Horvath et al. [79])). Part of the flow in the
central region moves downward in the two cyclones. This phenomena has
been shown in the axial velocity pattern in other published articles [e.g.,
79, 159].
The DPM results
In order to calculate the cut-off diameters of the two cyclones, 104 particles
were injected from the inlet surface with a velocity equals the inlet gas ve-
locity. The particle density ρp is 860 kg/m3. The grade efficiency curves for
the two designs are plotted in Fig. 8.5. The DPM analysis results and the
pressure drops for the two cyclones are depicted in Table 8.5. An accept-
able agreement between the CFD results and the MM mathematical model
has been obtained. While the difference between the two cyclone cut-off di-
ameters is small, the saving in pressure drop is considerable (nearly half
the value of Stairmand cyclone).
Based on the flow pattern analysis and the DPM results. One can con-
clude that the cyclone collection efficiency for the two cyclones should be
very close, with the advantage of low pressure drop in the new design.
The authors want to emphasis that only small changes in the geometrical
dimensions of the two designs led to this improvement in the performance.
153
Chapter 8. Optimization
8.2.2 Conclusions
Both mathematical modeling (the Muschelknautz method of modeling (MM))
and CFD investigation have been used to understand the effect of the cy-
clone geometrical parameters on the cyclone performance and a new opti-
mized cyclone geometrical ratios based on MM model has been obtained.
The new cyclone design is very close to the Stairmand high efficiency de-
sign in the geometrical parameter ratio, but superior in low pressure drop
at nearly the same cut-off diameter. The new cyclone design results in
nearly one-half the pressure drop obtained by the old Stairmand design at
the same volume flow rate.
154
8.2. Single-objective using MM model
Figure 8.3: The radial profile for the time averaged tangential and axial velocity
at different sections on the X-Z plane (Y=0) at sections S1–S3 . From top to bot-
tom: section S1–S3. From left to right: time-averaged static pressure, tangential
velocity and axial velocity respectively.
155
Chapter 8. Optimization
Figure 8.4: The radial profile for the time–averaged tangential and axial velocity
at different sections on the X-Z plane (Y=0) at sections S4–S6 . From top to bot-
tom: section S4–S6. From left to right: time-averaged static pressure, tangential
velocity and axial velocity respectively.
156
8.2. Single-objective using MM model
Particle diameter [micron]
Col
lect
ion
effic
ienc
y[-]
10-1 100 101
0.25
0.5
0.75
1Stairmand designNew design
Figure 8.5: The grade efficiency curves for the two designs
157
Chapter 8. Optimization
8.3 Optimization of the Euler number using
RBFNN and Nelder-Mead technique
Recently, Artificial neural networks (ANNs) have been widely applied in
the fields of modeling, prediction, fault detection and process control. In
the field of performance evaluation for cyclone separators, unfortunately,
ANNs have not been paid enough attention on their algorithmic advan-
tages [197]. There are few articles about the application of neural network
in the field of cyclone separator.
There are six main objectives of this study. (1) Application of the artificial
neural network to model the pressure drop using experimental dataset.
(2) Investigation of the effect of the seven geometrical parameters on the
pressure drop based on the trained RBFNN. (3) A detailed comparison be-
tween the experimental pressure drop values and the estimated values
obtained from different mathematical models. In order to recommend the
best mathematical model for future use. (4) Application of the response
surface methodology to study the effect of each geometrical parameter on
the pressure drop and test the interaction between these parameters using
the trained RBFNN. (5) Obtaining the optimum design (geometrical ratios)
for minimum pressure drop. (6) CFD study of the new cyclone separator
and compare its performance and flow pattern with the Stairmand design.
8.3.1 Radial basis function neural networks (RBFNN)
Why RBFNN
Radial basis function neural networks (RBFNNs) are powerful and inter-
esting networks due to their rapid training, generality and simplicity [16].
Girosi and Poggio [65] and Hartman and Keeler [72] proved that RBFNNs
are universal approximators and can approximate any continuous func-
tion with arbitrary accuracy. Training of these networks is very fast, and
they are very good at interpolation [190]. Niros and Tsekouras [120] stated
that radial basis function neural networks (RBFNN) have certain advan-
tages over other types of neural networks including better approximation
capabilities, simple network structure, and faster learning.
Zhao and Su [197] tested three different types of artificial neural network
to model the pressure drop in cyclone separators, viz. the back propa-
gation neural network (BPNN), the radial basis function neural network
(RBFNN) and the generalized regression neural network (GRNN). They
stated that compared with the BPNN and GRNN, the RBFNN provides
superior prediction performance criteria, better capability of approxima-
tion and high robustness.
158
8.3. Single-objective using RBFNN
In this study, the radial basis function neural network (RBFNN) is em-
ployed to model the pressure drop in the cyclone separator.
The structure of RBFNN
The radial basis function neural network (RBFNN) is a kind of 3-layered
forward network with multi inputs and multi outputs. The first layer has
(m) inputs, while the second is a hidden layer with (L) units, and the third
layer has (n) outputs, Fig. 8.6. The transformation function from the in-
puts to the hidden units, varied radial basis functions (RBF) on different
occasions, is nonlinear, whereas the mapping of the hidden units to the
outputs is linear [180].
Theoretically, RBFNN has the ability of approaching nonlinear mapping
arbitrarily, fr : Rm → Rn, as defined by:
y = fr(x) = W0 +
L∑
i=1
Wiφ (‖x− ci‖) (8.2)
where x ∈ Rm is the input vector, y ∈ Rn is the output vector, ci ∈ Rm
(i = 1, 2, · · ·L) is the center vector, ‖ · ‖ is the Euclidian norm, Wi ∈ Rm
(i = 1, 2, · · ·L) is the weight vector, W0 ∈ Rm is the bias vector, and φ(·) is
the Gaussian function, a non negative and nonlinear function with radial
symmetry and attenuation versus center, in the form of:
φ(v) = exp(−v2/2σ2) (8.3)
where σ is a width constant (spread factor ) [180].
In order to comprehensively compare the model performance, the evalua-
tion parameters, normalized mean squared error E2 and correlation coef-
ficient R are employed as follows [197]:
E2 =1
nΣn
i=1(yNi − yNi)2
(8.4)
where n is the number of test cases, y is the actual variable, y is the
RBFNN output variable and Ni is the neurons number of input layer in
the RBFNN.
R =Σn
i=1(yNi − yN )(yNi − yN
)√Σn
i=1(yNi − yN )2Σn
i=1
(yNi − yN
)2 (8.5)
159
Chapter 8. Optimization
1
∑
∑
∑
x1
x2
xm
W0
Wi
y1
y2
yn
Input layer Hidden layer Output layer
Figure 8.6: Schematic diagram for the radial basis function neural network
Modeling procedures
The ANN modeling procedure can be divided into the following steps: se-
lecting the variables, dividing the sample, optimizing parameters, training
and testing simulation and evaluating performance [197]. In this study, all
calculations are carried out on a MacBook pro laptop with the hardware
configurations: processor, Intel Core 2 Duo (2.4 GHz); memory, 4.0 GB
(DDR3-1067 2G 2); hard drive, 320 GB (7200 rpm); with Mac OS X 10.5
system.
The Euler number
The pressure drop across a cyclone separator essentially depends on the
dimensions and operating conditions. Generally, it is proportional to the
average dynamic pressure at the inlet and is often defined as [197]
∆P = Eu
(12ρgV
2in
)(8.6)
where Eu is Euler number (the dimensionless pressure drop also called
the pressure drop coefficient [197]). The Euler number is a complex non-
linear function of the cyclone geometrical dimensions and is not affected
by operating conditions in the high Reynolds number ( Re > 5E4) [50, 77].
The Euler number will be constant for any cyclone configuration regard-
less of size as long as the dimension ratios remain the same, although the
pressure drop varies with different operating conditions (due to the effect
160
8.3. Single-objective using RBFNN
of ρg and Vin). Therefore, pressure drop can be established by determin-
ing experimentally or theoretically for a particular cyclone design and also
be modified by the semi-empirical correlations to take the effect of solid
loading [197].
In order to determine the Euler number more accurately, all eight dimen-
sions of the cyclone are selected to establish the ANN models because they
have the effect on the Euler number to different extent [139, 197]. Usu-
ally, these dimensions can be characterized by the barrel diameter D and
expressed as seven dimensionless geometric ratios [197]:
Eu = f
(Dx
D,a
D,b
D,S
D,Ht
D,h
D,Bc
D
)(8.7)
According to Eq. 8.7, seven independent dimensionless geometrical vari-
ables and one dependent variable (the Euler number of the cyclone) are se-
lected as respectively the input and output parameters in the ANN model,
as presented in Table 8.6. For simplicity, the division of each factor by the
barrel diameter D will be dropped.
A dataset of 98 samples obtained from the measurements of pressure drop
for different cyclone designs available in the literature [42, 139, 197] is
used in the present investigation to evaluate the prediction performance
of the ANN models. Table 8.7 presents more details about the used dataset
including the minimum, mean, maximum and range of the seven dimen-
sionless geometrical ratios. Due to the large difference in the order of mag-
nitude of the value (cf. Table 8.7), the available dataset is transformed into
-1 to 1 interval using the Matlab intrinsic function; mapminmax in order
to avoid solution divergence [197]. The ANN calculations have been per-
formed using the neural network toolbox available from Matlab commer-
cial software 2010a.
Descriptive statistical parameters of the input dataset
Table 8.8 shows Pearson product moment correlations between each pair of
variables. These correlation coefficients range between -1 and +1 and mea-
sure the strength of the linear relationship between the variables. More-
over, shown in parentheses is the P-value which tests the statistical signif-
Table 8.6: The input and output variables for ANN model
Input parameters Output diameter
Variables X1 X2 X3 X4 X5 X6 X7 ySpecification Dx a b S Ht h Bc Eu
161
Chapter 8. Optimization
Table 8.7: Descriptive statistical parameters for the training dataset
Variable Dx a b S Ht h Bc
Minimum 0.25 0.113 0.067 0.39 1.158 0.501 0.14Mean 0.429 0.630 0.211 0.891 3.283 1.189 0.342Maximum 0.667 1.0 0.4 3.052 10.97 3.5 1.0Range 0.417 0.887 0.333 2.662 9.812 2.999 0.86
icance of the estimated correlations. P-values below 0.05 indicate statisti-
cally significant non-zero correlations at the 95.0% confidence level. The
following pairs of variables have P-values below 0.05: (1) Dx with a and Bc;
(2) a with b, Ht and Bc; (3) b with Ht and Bc;(4) S with Ht, h and Bc; (5) Ht
with h and Bc; (6) h with Bc. These conclusions can be obtained also from
Fig. 8.7. From this analysis, however there are some correlations between
the input variables, but this dataset is still reliable. The same dataset has
been used successfully by other researchers [e.g., 194, 197]. Furthermore,
it is the only available experimental data set in the literature.
K-fold cross validation
For the calibration of the RBFNN, the spread factor σ plays an important
role in the regression model. To obtain the optimum value for σ a multi-
step search technique is used. In the multi-step search technique, the
seeking for the optimum value is performed in two steps [197]. The first
step is a coarse search to determine the best range of values. The second
search is a fine search in the best range. To avoid overfitting, the K-fold
cross validation has been employed [8, 80, 197]. The original sample of
data is randomly portioned into K subsamples. A single subsample is used
for the validation (testing) and the other remaining K-1 subsamples are
used for the training. The process of training and testing is then repeated
for each of the K possible choices of the subset omitted from the training.
The average performance on the K omitted subsets is then our estimate
of the generalization performance. This procedure has the advantage that
is allows us to use a high proportion of the available data (a fraction 1 -
1/K) for training, while making use of all the data points in estimating the
Table 8.8: Correlations between each pair of input variables∗
Dx a b S Ht h BcDx -0.377 (0.000) -0.121 (0.234) 0.092 ( 0.367) 0.189 (0.061) 0.096 (0.346) 0.199 (0.049)a -0.377 (0.000) 0.442 (0.000) 0.148 (0.145) 0.288 (0.004) -0.052 (0.610) 0.270 (0.007)
b - 0.121 (0.234) 0.442 (0.000) 0.170 (0.093) 0.239 (0.017) 0.194 (0.054) 0.243 (0.015)
S 0.092 (0.367) 0.148 (0.145) 0.170 (0.093) 0.378 (0.000) 0.685 (0.000) 0.526 (0.000)
Ht 0.189 (0.061) 0.288 (0.004) 0.239 (0.017) 0.378 (0.000) 0.393 (0.000) 0.555 (0.000)
h 0.096 (0.346) -0.052 (0.610) 0.194 (0.054) 0.685 (0.000) 0.393 (0.000) 0.470 (0.000)Bc 0.199 (0.049) 0.270 (0.007) 0.243 (0.015) 0.526 (0.000) 0.555 (0.000) 0.470 (0.000)∗ The bold value shown in parentheses is the P-value.
162
8.3. Single-objective using RBFNN
Figure 8.7: Qualitative representation of the correlations between each pair of
input variables. For uncorrelated pair of variables, the data will be well distributed
and no linear correlation can be obtained e.g., Dx and b (first row, third column).
The high correlation between Ht and h is clear (fifth row, sixth column), where one
can fit easily fit a straight line.
generalization error. The disadvantage is that we need to train the net-
work K times. Typically K = 10 is considered reasonable and most widely
used [18]. For the radial basis function neural networks (newrb in Matlab
2010a), the learning process is a must to obtain the weights (the width of
the radial basis function units). In order to obtain the optimum value of
the spread factor σ, the multi-step search technique with 10-fold cross val-
idation from the interval of (0 - 1) with the performance goal of 1E-5, the
maximum number of neurons in the hidden layer equals the training sam-
ple size of 98 neurons, the number of neurons to add between displays is
10. Based on the mentioned settings, the optimum value of σ equals 0.191
(E2=1.1321E-06), this value is different than that obtained by Zhao and
Su [197] (σ equals 0.32 (E2=5.84E-04)) the reason can be referred to the
lower goal used in the current study (the goal used in Zhao and Su [197]
was 1E-4, and all other settings are identical).
8.3.1.1 Fitting the ANN
Table 8.9 presents more details about the validation of the used RBFNNs.
Both the average, minimum, maximum and range of the input (Euler num-
ber) and the predicted Euler number are given. It is clear from Table 8.9
that the ANNs preserved the descriptive statistical parameters of the in-
163
Chapter 8. Optimization
Table 8.9: Validation of the used RBFNN∗
Experimental MM Stairmand Ramachandran Shepherd
x y x y x y x y x y
Average 23.268 23.268 15.150 15.150 20.206 20.206 22.543 22.543 17.774 17.774
Minimum 2.3 1.745 1.34 1.164 2.88 2.892 1.85 1.793 0.957 1.197Maximum 155.3 155.985 138.0 137.235 132.0 133.521 153.0 150.854 92.2 92.543
Range 153.0 154.24 136.66 136.071 129.12 130.629 151.15 149.061 91.243 91.346
Correlation Coefficient, R 0.999 0.996 0.999 0.999 0.999
Mean squared error, E2 1.311E-4 1.212E-4 9.185E-5 1.442E-4 5.411E-5
Intercept 0.017 0.012 0.011 0.020 0.005Slope 0.999 0.999 0.999 0.999 1.0
∗ x is the input to the RBFNN and y is the predicted value. Both x and y represent the Euler number.The values of R, E2, intercept and slope are that for the testing stage.
put data. The correlation coefficient between the input and the output and
the mean squared error are given for each RBFNN. The intercept and the
slope of the adjusted line between the input and the predicted value of the
ANN are also given.
The configured RBFNN predictions versus experimental data and four
other models for cyclone Euler number are shown in Fig. 8.8. According
to Fig. 8.8, it can be seen that the ANN models are able to attain the
high training accuracy. The training mean square errors for the exper-
imental values and the four mathematical models (MM, Stairmand, Ra-
machandran and Shepherd and Lapple), have the values 1.311E-4, 1.212E-
4, 9.185E-5, 1.442E-4 and 5.411E-5 respectively (Fig. 8.8). This indicates
that, compared with traditional models of curve fitting, the models based
on artificial intelligence algorithm have a superior capability of nonlinear
fitting. Especially, the RBFNN has its unique and optimal approximation
characteristics in learning process [197].
Figure 8.8 illustrates the agreement between the ANN input and output.
The obtained relation is a typical linear relation with a coefficient of cor-
relation close to 1 (R > 0.999). The agreement between the input and
output of the ANN is also clear from the value of the mean squared error
E2 (< 1.5E − 4). That means, the trained neural network predicts very
well the Euler number values and can be used in cyclone design and per-
formance estimation. Table 8.9 and Fig. 8.8 present different performance
indicators as a validation of the proposed model for experimental values.
8.3.2 Evaluation of different mathematical models
In order to evaluate the performance of the four tested mathematical mod-
els in comparison with the experimental values, the percentage residual
164
8.3. Single-objective using RBFNN
Input value
Pre
dict
edva
lue
50 100 150
50
100
150 Data pointLinear fit
(a) Experimental values
Input value
Pre
dict
edva
lue
50 100 150
50
100
150 Data pointLinear fit
(b) MM model
Input value
Pre
dict
edva
lue
50 100 150
50
100
150 Data pointLinear fit
(c) Stairmand
Input value
Pre
dict
edva
lue
50 100 150
50
100
150 Data pointLinear fit
(d) Ramachandran model
Input value
Pre
dict
edva
lue
20 40 60 80 100
20
40
60
80
100Data pointLinear fit
(e) Shepherd and Lapple model
Figure 8.8: Linear regression of the RBFNN.
165
Chapter 8. Optimization
Run number
Res
idua
lerr
or%
0 20 40 60 80 100-200
-100
0
100
200
300
400MMStairmandRamachandranShepherd
Figure 8.9: Percentage residual error for the four tested models based on 98
dataset.
error (Eq. 8.8) for each model has been plotted in Fig. 8.9.
% error =Model value - Experimental value
Experimental value∗ 100 (8.8)
Figure 8.9 depicts that MM model underestimate the pressure drop by
around 50%. Also the percentage errors for other models are between ±50%. The peaks in error are almost the same for all models. The residual
error of MM model is almost the lowest. The reasons of these peaks may be
due to the high values of geometrical swirl number for some of the avail-
able dataset (cf. Ramachandran et al. [139]), where Sg = πDx D/(4a b). Sg
varies between 1 and 5 for industrial cyclones [74]. The Sg values in the 98
dataset used in this study varies from 2.18 to 92.67. As the swirl number is
a measure for the ratio of tangential to axial momentum [74, 75, 154], the
high values of Sg may cause violation of the simplified assumptions used
in the models. Figure 8.9 seems to indicate that the MM model among the
other mathematical models is best suited for estimation of the pressure
drop.
The effect of geometrical parameters on the Euler number
The effects of the geometrical parameters on the Euler number are de-
picted in Fig. 8.10. To study the effect of each parameter, the tested RBFNN
model has been used by varying one parameter at a time from its minimum
to maximum values of the available 98 dataset, while the other parameters
are kept constant at their mean values (cf. Table 8.7). Figure 8.10(a) indi-
cates the significant effect of the vortex finder diameter Dx and the vortex
finder length S, the inlet width b and the total height Ht. Less effect is due
to the cylinder height h (for h > 2.5) and the inlet height a (for a > 0.55).
166
8.3. Single-objective using RBFNN
The effect of the geometrical parameters on the Euler number obtained
from the MM model (Fig. 8.10(b)) is very close to that obtained from the
analysis of the experimental dataset but with underestimation of the Eu-
ler number. This supports the use of the MM model in the Euler number
estimation for cyclone separators [29, 50, 77]. The situation for the Stair-
mand model is not the same as that for the experimental data or MM
model. Here, the effect of many geometrical parameters attenuated. The
effect of the inlet width becomes insignificant. That is not realistic, as for
high values of inlet width, a considerable part of the incoming flow will im-
pact directly the vortex finder which increases the entrance loss and conse-
quently, the total pressure drop in the cyclone separator. The Ramachan-
dran model was initially constructed by curve fitting based on the used 98
dataset (cf., Ramachandran et al. [139] for more details). So it depicts a
better agreement with the experimental values of the Euler number than
that of Stairmand model. The Ramachandran model predicts almost the
same Euler number variation with the inlet width. In the Shepherd and
Lapple model only the inlet dimensions (aandb) and vortex finder diameter
Dx affect the pressure drop (Eu = 16ab/D2x) as is clear from Fig. 8.10(e).
Figure 8.11 compares the effect of each individual geometrical parameter
using predictions with the ANN based on respectively the experimental
data and the four tested models. The following conclusions can be drawn
from the analysis of Fig. 8.11:
Dx : All the models (except Shepherd and Lapple) show the same variation
of the Euler number with increasing Dx. For Dx > 0.5 there is an ex-
act matching between the results of MM model and the experimental
values.
a : Both the Ramachandran and the Shepherd and Lapple models predict
a (nearly) linear relation between the inlet height and the Euler num-
ber. The effect of changing a predicted by the Stairmand model is in
accordance with the experimental values for a > 0.8.
b : For small values of b (b < 0.15), the MM model results are in good agree-
ment with the experimental values. Also the trends of all models in
this range are matching the trend of the experimental values. Be-
yond this range, Both the Ramachandran and the Stairmand models
agree well with the experimental values trend.
S : The trend of the MM model results is similar to that of the experi-
mental values, although there is a shift in the values of the Euler
number. The Shepherd and Lapple model does not present any effect
of changing S on the Euler number (for Shepherd and Lapple model,
Eu = 16ab/D2x), while the MM model always underestimates the ef-
167
Chapter 8. Optimization
Dx, a, b, B c
h, S
Ht
Eul
ernu
mbe
r
0.2 0.4 0.6 0.8 1
0.5 1 1.5 2 2.5 3 3.5
2 4 6 8 10
0
10
20
30
40
Dx
abSHt
hBc
(a) Experimental values
Dx, a, b, B c
h, S
Ht
Eul
ernu
mbe
r
0.2 0.4 0.6 0.8 1
0.5 1 1.5 2 2.5 3 3.5
2 4 6 8 10
0
10
20
30
40
Dx
abSHt
hBc
(b) MM model
Dx, a, b, B c
h, S
Ht
Eul
ernu
mbe
r
0.2 0.4 0.6 0.8 1
0.5 1 1.5 2 2.5 3 3.5
2 4 6 8 10
0
10
20
30
40
Dx
abSHt
hBc
(c) Stairmand model
Figure 8.10: The effect of geometrical parameters on the Euler number using the
trained neural networks based on experimental dataset and four different mathe-
matical models.
168
8.3. Single-objective using RBFNN
Dx, a, b, B c
h, S
Ht
Eul
ernu
mbe
r
0.2 0.4 0.6 0.8 1
0.5 1 1.5 2 2.5 3 3.5
2 4 6 8 10
0
10
20
30
40 Dx
abSHt
hBc
(d) Ramachandran model
Dx, a, b, B c
h, S
Ht
Eul
ernu
mbe
r
0.2 0.4 0.6 0.8 1
0.5 1 1.5 2 2.5 3 3.5
2 4 6 8 10
0
10
20
30
40 Dx
abSHt
hBc
(e) Shepherd and Lapple model
Figure 8.10: (continued) The effect of geometrical parameters on the Euler number
using the trained neural networks based on experimental dataset and four differ-
ent mathematical models.
fect of S. Both the Stairmand and the Ramachandran models may
over/underestimate its effect on Euler number.
Ht : Both the Stairmand and the Ramachandran models do not show a
significant effect of Ht on the Euler number. The trends of both MM
and Ramachandran model are almost the same as that for the exper-
imental values.
h : For higher values of h all models (except the Shepherd and Lapple
model) give nearly the same value of Euler number.
Bc : Nearly, all the models (except the Shepherd and Lapple model) show
the same trend in the changing the Euler number with the cone tip
diameter.
8.3.3 Design of experiment (DOE)
Table 8.10 depicts the parameters ranges selected for the seven geometri-
cal parameters. The study was planned using Box-Behnken design, with
64 combinations. A significant level of P < 0.05 (95% confidence) was used
in all tests. Analysis of variance (ANOVA) was followed by an F-test of the
169
Chapter 8. Optimization
Table 8.10: The values of the independent variables used in the design of experi-
ment
Variables minimum center maximum
Vortex finder diameter, Dx=X1 0.2 0.475 0.75
Inlet height, a=X2 0.4 0.55 0.7
Inlet width, b=X3 0.14 0.27 0.4
Vortex finder length, S=X4 0.4 1.2 2.0
Total cyclone height, Ht=X5 3.0 5.0 7.0
Cylinder height, h=X6 1.0 1.5 2.0
Cone tip diameter, Bc=X7 0.2 0.3 0.4
individual factors and interactions.
Fitting the model
Analysis of variance (ANOVA) showed that the resultant quadratic poly-
nomial models adequately represented the experimental data with the co-
efficient of multiple determination R2 being 0.965843. This indicates that
the quadratic polynomial model obtained was adequate to describe the in-
fluence of the independent variables studied [189]. Analysis of variance
(ANOVA) was used to evaluate the significance of the coefficients of the
quadratic polynomial models (see Table 8.12). For any of the terms in the
models, a large F-value (small P-value) would indicate a more significant
effect on the respective response variables.
Based on the ANOVA results presented in Table 8.12, the variable with
the largest effect on the pressure drop (Euler number) was the linear term
of vortex finder diameter, followed by the linear term of inlet width and
vortex finder length (P < 0.05); the other four linear terms (inlet height,
barrel height, cyclone total height and cone tip diameter) did not show a
significant effect (P > 0.05). The quadratic term of vortex finder diameter,
vortex finder length and cyclone total height also had a significant effect (P
< 0.05) on the pressure drop; however, the effect of the other four quadratic
terms was insignificant (P > 0.05). Furthermore, the interaction between
the inlet dimensions and vortex finder diameters (P < 0.05) also had a
significant effect on the pressure drop, while the effect of the remaining
terms was insignificant (P > 0.05).
Analysis of response surfaces
For visualization of the calculated factor, main effects plot, Pareto chart
and response surface plots were drawn. The slope of the main effect curve
170
8.3. Single-objective using RBFNN
Table 8.11: The geometrical dimensions and Euler number for the used cyclones
X1 X2 X3 X4 X5 X6 X7 Y1
Exp. No. a/D b/D Dx/D Ht/D h/D S/D Bc/D Euler Number
1 0.7 0.27 0.75 5 1 1.2 0.3 4.16
2 0.55 0.27 0.475 7 2 2 0.3 6.93
3 0.55 0.27 0.475 3 2 0.4 0.3 9.30
4 0.55 0.27 0.2 7 1.5 1.2 0.2 65.705 0.7 0.4 0.475 7 1.5 1.2 0.3 11.90
6 0.55 0.14 0.475 5 1 1.2 0.4 4.34
7 0.55 0.4 0.475 5 1 1.2 0.4 9.19
8 0.55 0.27 0.475 5 1.5 1.2 0.3 7.94
9 0.55 0.27 0.475 3 1 2 0.3 8.87
10 0.4 0.4 0.475 7 1.5 1.2 0.3 5.8811 0.55 0.27 0.475 5 1.5 1.2 0.3 7.94
12 0.55 0.4 0.75 5 1.5 0.4 0.3 3.24
13 0.55 0.14 0.2 5 1.5 0.4 0.3 24.50
14 0.7 0.4 0.475 3 1.5 1.2 0.3 13.40
15 0.7 0.27 0.2 5 1 1.2 0.3 105.0016 0.7 0.27 0.475 5 1.5 0.4 0.4 10.30
17 0.4 0.4 0.475 3 1.5 1.2 0.3 7.05
18 0.7 0.27 0.75 5 2 1.2 0.3 4.07
19 0.55 0.4 0.475 5 2 1.2 0.2 9.19
20 0.55 0.27 0.475 7 1 0.4 0.3 7.5021 0.55 0.27 0.2 3 1.5 1.2 0.2 77.50
22 0.55 0.14 0.75 5 1.5 0.4 0.3 2.42
23 0.55 0.4 0.2 5 1.5 0.4 0.3 130.00
24 0.55 0.27 0.475 5 1.5 1.2 0.3 7.94
25 0.4 0.14 0.475 7 1.5 1.2 0.3 3.17
26 0.4 0.27 0.2 5 1 1.2 0.3 41.9027 0.4 0.27 0.2 5 2 1.2 0.3 40.70
28 0.55 0.27 0.475 5 1.5 1.2 0.3 7.94
29 0.55 0.27 0.475 5 1.5 1.2 0.3 7.94
30 0.55 0.27 0.75 7 1.5 1.2 0.2 3.35
31 0.55 0.4 0.475 5 1 1.2 0.2 9.4532 0.55 0.4 0.2 5 1.5 2 0.3 128.00
33 0.55 0.14 0.2 5 1.5 2 0.3 23.80
34 0.55 0.27 0.75 3 1.5 1.2 0.4 3.82
35 0.55 0.14 0.475 5 2 1.2 0.4 4.22
36 0.55 0.27 0.475 5 1.5 1.2 0.3 7.9437 0.55 0.14 0.475 5 2 1.2 0.2 4.33
38 0.55 0.27 0.75 3 1.5 1.2 0.2 3.86
39 0.4 0.27 0.75 5 1 1.2 0.3 2.99
40 0.55 0.27 0.475 5 1.5 1.2 0.3 7.94
41 0.55 0.27 0.475 7 1 2 0.3 7.09
42 0.7 0.14 0.475 3 1.5 1.2 0.3 6.0343 0.55 0.14 0.75 5 1.5 2 0.3 2.23
44 0.7 0.27 0.2 5 2 1.2 0.3 103.00
45 0.55 0.4 0.75 5 1.5 2 0.3 3.05
46 0.55 0.27 0.2 7 1.5 1.2 0.4 64.10
47 0.4 0.27 0.75 5 2 1.2 0.3 2.9248 0.4 0.27 0.475 5 1.5 2 0.4 5.70
49 0.7 0.27 0.475 5 1.5 2 0.4 9.64
50 0.4 0.27 0.475 5 1.5 0.4 0.2 6.34
51 0.4 0.27 0.475 5 1.5 0.4 0.4 6.12
52 0.55 0.27 0.475 3 2 2 0.3 8.5153 0.7 0.14 0.475 7 1.5 1.2 0.3 4.70
54 0.4 0.27 0.475 5 1.5 2 0.2 5.88
55 0.55 0.14 0.475 5 1 1.2 0.2 4.51
56 0.7 0.27 0.475 5 1.5 0.4 0.2 10.60
57 0.55 0.27 0.2 3 1.5 1.2 0.4 76.30
58 0.4 0.14 0.475 3 1.5 1.2 0.3 4.0659 0.55 0.27 0.75 7 1.5 1.2 0.4 3.25
60 0.55 0.27 0.475 3 1 0.4 0.3 9.77
61 0.7 0.27 0.475 5 1.5 2 0.2 9.91
62 0.55 0.27 0.475 5 1.5 1.2 0.3 7.94
63 0.55 0.4 0.475 5 2 1.2 0.4 9.0164 0.55 0.27 0.475 7 2 0.4 0.3 7.30
New design∗ 0.618 0.236 0.618 4.236 1.618 0.618 0.3819 5.24
Stairmand design† 0.5 0.2 0.5 4 1.5 0.5 0.36 5.79
∗ The new design based on MM model and downhill simplex optimization scheme (Euler number =5.24, cut-off diameter =1.77µm ).†The standard Stairmand high efficiency cyclone design, (Euler number =5.79, cut-off diameter =1.54µm based on MM model).
171
Chapter 8. Optimization
Table 8.12: Analysis of variance of the regression coefficients of the fitted quadratic
equationa
Variable Regression coefficient F-Ratio P-Value
β0 138.604
Linearβ1 -485.694 96.89 0.0000β2 20.2354 0.06 0.8152β3 325.517 59.28 0.0000β4 -10.7497 6.87 0.0140β5 -8.7097 0.06 0.8161β6 -19.1361 0.57 0.4575β7 -167.261 0.60 0.4464
Quadraticβ11 515.706 489.19 0.0000β22 113.392 2.09 0.1590β33 -13.0018 0.02 0.9017β44 10.9817 15.89 0.0004β55 1.04634 5.63 0.0247β66 6.06181 0.74 0.3974β77 232.346 1.74 0.1983
Interactionβ12 -158.258 7.61 0.0101β13 -345.242 27.22 0.0000β14 -4.21438 0.15 0.6981β15 8.42261 3.83 0.0602β16 31.8848 3.43 0.0744β17 95.7341 1.24 0.2752β23 -143.344 1.40 0.2473β24 20.5571 1.09 0.3060β25 -3.23107 0.17 0.6851β26 -16.3805 0.27 0.6076β27 -6.62067 0.00 0.9668β34 -26.3282 1.34 0.2569β35 0.0205865 0.00 0.9982β36 32.7526 0.81 0.3758β37 -28.156 0.02 0.8782β45 -0.62389 0.18 0.6763β46 -2.06662 0.12 0.7294β47 -18.6467 0.40 0.5334β56 -2.60425 1.21 0.2803β57 2.35394 0.04 0.8437β67 -5.32565 0.01 0.9112
R2 0.965843
a Bold numbers indicate significant factors as identified by the analysis of variance (ANOVA) at the 95%confidence level.
172
8.3. Single-objective using RBFNN
is proportional to the size of the effect and the direction of the curve spec-
ifies a positive or negative influence of the effect [50, 61] (Fig.8.12(a)).
Based on the main effect plot, the most significant factors on the Euler
number are: (1) the vortex finder diameter Dx, with a second–order curve
with a wide range of inverse relation and a narrow range of direct relation,
(2) the inlet width b linearly related to the Euler number, (3) the cyclone
total height Ht, (4) the vortex finder length S, whereas the other factors
have an insignificant effect.
Pareto charts were used to graphically summarize and display the rela-
tive importance of each parameter with respect to the Euler number [50].
The Pareto chart shows all the linear and second-order effects of the pa-
rameters within the model and estimates the significance of each with re-
spect to maximizing the Euler number response. A Pareto chart displays
a frequency histogram with the length of each bar proportional to each
estimated standardized effect [30]. The vertical line on the Pareto charts
judges whether each effect is statistically significant within the generated
response surface model; bars that extend beyond this line represent effects
that are statistically significant at a 95% confidence level. Based on the
Pareto chart (Fig. 8.12(b)) and ANOVA table (Table 8.12) there are five sig-
nificant parameters (eight terms in the ANOVA table; Table 8.12) at a 95%
confidence level: the vortex finder diameter Dx, the inlet width b, the total
cyclone height Ht, the vortex finder length S and the inlet height a (due to
interaction with Dx). Therefore, the Pareto chart is a perfect supplement
to the main effects plot.
To visualize the effect of the independent variables on the dependent ones,
surface response of the quadratic polynomial models were generated by
varying two of the independent variables within the experimental range
while holding the other factors at their central values [189] as shown in
Fig. 8.13. Thus, Fig. 8.13(b) was generated by varying the inlet height aand the inlet width b while keeping the other five factors constant. The
trend of the curve is linear, with a more significant effect for the inlet
width b, and a weak interaction between the inlet height a and width b.The response surface plots of Figs. 8.13(a), 8.13(c) and 8.13(d) show that
there are strong interactions between the vortex finder diameter Dx and
respectively the inlet height a, the cyclone total height Ht and the vortex
finder length S. The effect of vortex finder length S is less significant with
respect to the vortex finder diameter Dx, but its effect is higher than that
of the barrel height h and the cone tip diameter Bc (cf., Fig. 8.12).
173
Chapter 8. Optimization
8.3.3.1 Optimization (Nelder-Mead method)
In this optimization problem, the objective function is the Euler number
(f1(x) in Eq. 8.1). Table 8.13 gives the optimum values for cyclone geomet-
rical parameters for minimum pressure drop. It is clear from Table 8.13
that the new optimized design is very close to the Stairmand design in
many geometrical parameters, whereas the new ratios will lead to the min-
imum pressure drop. To understand the effect of this small change in the
geometrical ratios on the flow field pattern and performance, a CFD study
for the two designs is needed.
Table 8.13: The geometrical parameters for minimum pressure drop using RBFNN
Factor Low High Stairmand design Optimum
Dx 0.2 0.75 0.5 0.487
a 0.5 0.75 0.5 0.629
b 0.14 0.4 0.2 0.203
S 0.4 2.0 0.5 0.733
Ht 3.0 7.0 4.0 4.852
h 1.0 2.0 1.5 1.633
Bc 0.2 0.4 0.375 0.383
174
8.3. Single-objective using RBFNN
Dx
Eul
ernu
mbe
r
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.70
10
20
30
40
50Exp.MMStairmandRamchandranShepherd
(a) Dx
a
Eul
ernu
mbe
r
0 0.2 0.4 0.6 0.8 10
5
10
15
20
25Exp.MMStairmandRamchandranShepherd
(b) a
b
Eul
ernu
mbe
r
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
5
10
15
20
25
30Exp.MMStairmandRamchandranShepherd
(c) b
S
Eul
ernu
mbe
r
0.5 1 1.5 2 2.5 30
5
10
15
20
25
30
35
40
45
50
55
60
65Exp.MMStairmandRamchandranShepherd
(d) S
Figure 8.11: Comparison between the effect of each geometrical parameters on the
Euler number using different models and experimental values based on RBFNN.
175
Chapter 8. Optimization
Ht
Eul
ernu
mbe
r
2 4 6 8 10 120
10
20
30
Exp.MMStairmandRamchandranShepherd
(e) Ht
h
Eul
ernu
mbe
r
0 0.5 1 1.5 2 2.5 3 3.5 40
5
10
15
20Exp.MMStairmandRamchandranShepherd
(f) h
Bc
Eul
ernu
mbe
r
0.2 0.4 0.6 0.8 10
10
20
30
40
Exp.MMStairmandRamchandranShepherd
(g) Bc
Figure 8.11: (continued) Comparison between the effect of each geometrical param-
eters on the Euler number using different models and experimental values based
on RBFNN.
176
8.3. Single-objective using RBFNN
(a) Main effects plot
(b) Pareto chart
Figure 8.12: Analysis of design of experiment
177
Chapter 8. Optimization
(a) Dx versus a
(b) a versus b
(c) Dx versus Ht
(d) Dx versus S
Figure 8.13: The response surface plots.178
8.3. Single-objective using RBFNN
8.3.4 CFD Comparison between the two designs
Numerical settings
The air volume flow rate Qin=0.08 m3/s for the two cyclones (inlet velocity
for Stairmand design is 19 m/s and 14.9 m/s for the new design), air density
1.0 kg/m3 and dynamic viscosity of 2.11E-5 Pa s. the turbulent intensity
equals 5% and characteristic length equals 0.07 times the inlet width [75].
Velocity inlet boundary condition is applied at inlet, outflow at gas outlet
and wall boundary condition at all other boundaries. The Reynolds stress
turbulence model has been used to reveal the turbulence characteristics in
the cyclone separators.
The finite volume method has been used to discretize the partial differen-
tial equations of the model using the SIMPLEC (Semi-Implicit Method for
Pressure-Linked Equations-Consistent) method for pressure velocity cou-
pling and QUICK scheme to interpolate the variables on the surface of the
control volume. The implicit coupled solution algorithm was selected. The
unsteady Reynolds stress turbulence model (RSM) was used in this study
with a time step of 0.0001 s.
The grid refinement study using different levels of grid shows that a total
number of 134759 hexahedral cells for the Stairmand cyclone and 381709
hexahedral cells for the new design are sufficient to obtain a grid-independent
solution, and further mesh refinement yields only small, insignificant changes
in the numerical solution. The hexahedral meshes have been obtained us-
ing the GAMBIT commercial software. These simulations were performed
on an 8 nodes CPU Opteron 64 Linux cluster using Fluent 6.3.26 commer-
cial software. The geometrical values for the two cyclones (cf. Fig. 8.1) are
given in Table 8.14. The surface mesh for Stairmand cyclone is given in
Fig. 8.14(a).
Table 8.14: The values of geometrical parameters for the two designs (D=0.205 m)∗
Cyclone a/D b/D Dx/D Ht/D h/D S/D Bc/DStairmand design 0.5 0.2 0.5 4 1.5 0.5 0.375
New design 0.628 0.203 0.487 4.852 1.633 0.732 0.382
∗ The outlet section is above the cyclone surface by Le = 0.618D. The inlet section located at adistance Li = D from the cyclone center.
179
Chapter 8. Optimization
(a) The Stairmand design (b) The new design
Figure 8.14: The surface meshes for the two designs
Results and discussion
The pressure field
Figure 8.15 shows the contour plot at Y=0. In the two cyclones, the time-
averaged static pressure decreases radially from the wall to center. A neg-
ative pressure zone appears in the forced vortex region (central region)
due to high swirling velocity. The pressure gradient is largest along the
radial direction, while the gradient in the axial direction is very limited.
The cyclonic flow is not symmetrical as is clear from the shape of the low
pressure zone at the cyclone center (twisted cylinder). However, the two
cyclones have almost the same flow pattern, but the highest pressure of
the Stairmand design is nearly 1.5 times that of the new design, implying
that the new design has a lower pressure drop.
The pressure distribution presented in Figs. 8.16 and 8.17 of the two cy-
clones at sections S1–S6 depict the two parts pressure profile (for Rankine
vortex). Again, the highest static pressure for Stairmand design is more
than 1.5 times that of the new design at all sections while the central value
is almost the same for the two cyclones irrespective of the section location.
This indicates that, the new design has a lower pressure drop with respect
to the Stairmand design.
180
8.3. Single-objective using RBFNN
Figure 8.15: The contour plots for the time averaged flow variables at sections
Y=0 (cf. Fig.8.1). From top to bottom: Stairmand design and the new design re-
spectively. From left to right: the static pressure (N/m2), the tangential and axial
velocity (m/s). Note: both cyclones have the same barrel diameter and air volume
flow rate.
181
Chapter 8. Optimization
Distance from center (m)
Sta
ticpr
essu
re(N
/m2)
-0.1 -0.05 0 0.05 0.1
0
500
1000
1500
2000
2500New designStairmand design
Distance from center (m)
Tan
gent
ialv
eloc
ity/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1
0
0.5
1
1.5
2
2.5New designStairmand design
Distance from center (m)
Axi
alve
loci
ty/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1
-0.2
0
0.2
0.4
0.6
0.8
1New designStairmand design
Distance from center (m)
Sta
ticpr
essu
re(N
/m2)
-0.1 -0.05 0 0.05 0.1
0
500
1000
1500
2000
2500New designStairmand design
Distance from center (m)
Tan
gent
ialv
eloc
ity/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1
0
0.5
1
1.5
2
2.5New designStairmand design
Distance from center (m)
Axi
alve
loci
ty/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1-0.4
-0.2
0
0.2
0.4
0.6
0.8
1New designStairmand design
Distance from center (m)
Sta
ticpr
essu
re(N
/m2)
-0.1 -0.05 0 0.05 0.1
0
500
1000
1500
2000
2500New designStairmand design
Distance from center (m)
Tan
gent
ialv
eloc
ity/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1
0
0.5
1
1.5
2
2.5New designStairmand design
Distance from center (m)
Axi
alve
loci
ty/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1-0.4
-0.2
0
0.2
0.4
0.6
0.8
1New designStairmand design
Figure 8.16: The radial profile for the time averaged tangential and axial velocity
at different sections on the X-Z plane (Y=0) at sections S1–S3 . From top to bottom:
section S1–S3. From left to right: time-averaged static pressure, tangential and
axial velocity respectively.
182
8.3. Single-objective using RBFNN
Distance from center (m)
Sta
ticpr
essu
re(N
/m2)
-0.1 -0.05 0 0.05 0.1
0
500
1000
1500
2000
2500New designStairmand design
Distance from center (m)
Tan
gent
ialv
eloc
ity/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1
0
0.5
1
1.5
2
2.5New designStairmand design
Distance from center (m)
Axi
alve
loci
ty/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1-0.4
-0.2
0
0.2
0.4
0.6
0.8
1New designStairmand design
Distance from center (m)
Sta
ticpr
essu
re(N
/m2)
-0.1 -0.05 0 0.05 0.10
500
1000
1500
2000
2500New designStairmand design
Distance from center (m)
Tan
gent
ialv
eloc
ity/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1
0
0.5
1
1.5
2
2.5New designStairmand design
Distance from center (m)
Axi
alve
loci
ty/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1-0.4
-0.2
0
0.2
0.4
0.6
0.8
1New designStairmand design
Distance from center (m)
Sta
ticpr
essu
re(N
/m2)
-0.1 -0.05 0 0.05 0.10
500
1000
1500
2000
2500New designStairmand design
Distance from center (m)
Tan
gent
ialv
eloc
ity/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1
0
0.5
1
1.5
2
2.5New designStairmand design
Distance from center (m)
Axi
alve
loci
ty/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1
-0.2
0
0.2
0.4
0.6
0.8
1New designStairmand design
Figure 8.17: The radial profile for the time–averaged tangential and axial velocity
at different sections on the X-Z plane (Y=0) at sections S4–S6 . From top to bottom:
section S4–S6. From left to right: time-averaged static pressure, tangential and
axial velocity respectively.
183
Chapter 8. Optimization
Table 8.15: The position of different sectionsa
Section S1 S2 S3 S4 S5 S6
z`/D b 2.75 2.5 2.25 2.0 1.75 1.5
a Sections S1–S5 are located in the conical section, section S6 at the cylindrical part.b z` is measured from the top of the inlet section (cf. Fig. 8.1).
The velocity field
Based on the contour plots of the time–averaged tangential velocity, Fig. 8.15,
and the radial profiles at sections S1–S6 shown in Figs. 8.16 and 8.17, the
following conclusions can be drawn. The tangential velocity profile at any
section is composed of two regions, an inner and an outer one. In the
inner region, the flow rotates approximately like a solid body (forced vor-
tex), where the tangential velocity increases with radius. After reaching
its peak the velocity decreases with radius in the outer part of the profile
(free vortex). This profile is a so-called Rankine type vortex as mentioned
before, including a quasi-forced vortex in the central region and a quasi-
free vortex in the outer region. The maximum tangential velocity may
reach twice the average inlet velocity and occurs in the annular cylindrical
part. The tangential velocity distribution for the two cyclones are approx-
imately nearly identical in pattern and values (dimensionless), with the
highest velocity occurring at 1/4 of the cyclone radius for both cyclones.
This implies a nearly equal collection efficiency for both cyclones, as the
centrifugal force is the main driving force for particle collection in the cy-
clone separator. The axial velocity profiles for the two cyclones are also
very close, exhibiting the inverted W axial velocity profile [79]. Part of the
flow in the central region moves downward in the two cyclones. This phe-
nomena has been shown in the axial velocity pattern in other published
articles [e.g., 79, 159].
The DPM results
In order to calculate the cut-off diameters of the two cyclones, 104 particles
were injected from the inlet surface with a velocity equals the air inlet ve-
locity. The particle density ρp is 860 kg/m3 and the maximum number of
time steps for each injection was 200000 steps. The DPM analysis results
and the pressure drops for the two cyclones are depicted in Table 8.16. Al-
though, the difference between the two cyclone cut-off diameters is small,
the saving in the pressure drop is considerable (nearly 25% the value of
Stairmand cyclone).
Based on the flow pattern analysis and the DPM results, one can conclude
that the cyclone collection efficiency for the two cyclones is very close, with
184
8.3. Single-objective using RBFNN
Table 8.16: The pressure drop and the cut-off diameter for the two cyclones
∆p (N/m2) x50 (µm)
Stairmand design 1190 1.0New design 864 0.8
the advantage of low pressure drop in the new design. The authors want
to emphasis that only small changes in the geometrical dimensions of the
two designs lead to this improvement in the performance.
8.3.5 Conclusions
In order to predict accurately the complexly non linear relationships be-
tween pressure drop and geometrical dimensions, a radial basis neural
network (RBFNN) is developed and employed to model the pressure drop
for cyclone separators. The neural network has been trained and tested by
the experimental data available in literatures. The following conclusions
can be drawn from analysis of the obtained results:
• The result demonstrates that artificial neural networks can offer an
alternative and powerful approach to model the cyclone pressure drop.
• Four mathematical models (Muschelknautz method “MM”, Stairmand,
Ramachandran and Shepherd and Lapple) have been tested against
the experimental values. The residual error of MM model is the low-
est. Also, one can multiply the calculated value using MM by 1.5 to
get the experimental value (as a rough approximation)
• The analysis indicates the significant effect of the vortex finder diam-
eter Dx and the vortex finder length S, the inlet width b and the total
height Ht.
• The response surface methodology has been used to fit a second order
polynomial to the RBFNN.
• The second-order polynomial has been used to get a new optimized
cyclone for minimum pressure drop using Nelder-Mead technique.
• A comparison between the new design and the standard Stairmand
design has been performed using CFD simulation with the Reynolds
stress model to get a clear vision of the flow field pattern in the new
design.
• CFD results show that, the new cyclone design is very close to the
Stairmand high efficiency design in the geometrical parameter ratio,
and superior for low pressure drop at nearly the same cut-off diame-
ter.
• The new cyclone design results in nearly 75% of the pressure drop
obtained by the old Stairmand design at the same volume flow rate.
185
Chapter 8. Optimization
8.4 Single and multi-objective optimization us-
ing RBFNN and GA
Both the pressure drop and the cut-off diameter in a cyclone separator
can be decreased or increased by varying the cyclone dimensions. For an
accurate optimal design of a cyclone, it is quite necessary to use a reli-
able model for its performance parameters. Optimization of gas cyclone is,
indeed, a multi-objective optimization problem rather than a single objec-
tive optimization problem that has been considered so far in the literature
[53, 146]. Both the pressure drop and the collection efficiency in gas cy-
clones are important objective functions to be optimized simultaneously
in such a real-world complex multi-objective optimization problem [54].
These objective functions are either obtained from experiments, empirical
models or computed using very timely and high-cost computational fluid
dynamic (CFD) approaches. Modeling and optimization of the parameters
are investigated in the present study, by using radial basis function arti-
ficial neural networks and multi-objective genetic algorithm optimization
technique in order to maximize the collection efficiency (minimize the cut-
off diameter) and minimize the pressure drop.
The optimization studies conducted in the previous sections used the Nelder-
Mead technique [119] which suffer from two drawbacks, (1) the final solu-
tion may be affected by the start point, (2) the obtained optimum may be
a local minimum. To avoid this situation, the application of evolutionary
method like the genetic algorithm must.
Study objectives
There are four objectives of this study. (1) Investigation of the effect of the
seven geometrical parameters on the cyclone separator performance (the
pressure drop and cut-off diameter) based on the experimental data for
the pressure drop and the most robust mathematical models for the cut-off
diameter. (2) Study the possible interaction between the seven geometri-
cal parameters affecting the cyclone performance using response surface
methodology. (3) Multi-objective optimization to obtain new geometrical
ratios for optimum performance (minimum pressure drop and minimum
cut-off diameter). (4) Obtaining the optimum design (geometrical ratios)
of the cyclone separator for minimum pressure drop using the genetic al-
gorithm optimization technique, followed by a comparison of the numeri-
cal simulations of the optimal design and the Stairmand design using the
Reynolds stress turbulence model.
186
8.4. Multi-objective optimization using GA
8.4.1 Artificial neural network (ANN) approach
In this study, the radial basis function neural network (RBFNN) has been
used to model the effect of cyclone dimensions on the Euler number (pres-
sure drop). The details of this step have been explained in Sec. 8.3.1,
page 160.
The cut-off diameter (Stokes number)
The source of the training data for the cut-off diameter has been obtained
from the application of Iozia and Leith model [85]. This model has been
approved as an acceptable approach for calculating the cyclone cut-off di-
ameter [84, 85]. The cut-off diameter x50 is the particle diameter which
produces 50% collection efficiency.
A question may appear here, why the authors employed the cut-off diame-
ter instead of the collection efficiency, like in the study of Ravi et al. [141].
Firstly, for low mass loading cyclone separator, the cut-off diameter can
replace the collection efficiency, since one can fit the grade efficiency curve
using the cut-off diameter via some correlations, cf. Hoffmann and Stein
[77, page 91] for more details. Moreover, many models can predict well the
cut-off diameter but exhibit different grade efficiency curves [77, page 97].
Secondly, the cut-off diameter (instead of the collection efficiency) has been
used as an objective function in many recent publications [e.g., 50, 148].
Moreover, the selection of the cut-off diameter or the collection efficiency
for low mass loading cyclones can be considered as a researcher choice.
The Iozia and Leith model [84] is similar to the model of Barth [9] both
based on the equilibrium-orbit theory (Force balance). The cut-off diameter
can be calculated as [84]:
x50 =(9 µ C) /
(π HCS ρp V 2
θmax
)1/2(8.9)
where µ is the gas viscosity, Q is the gas volume flow rate, ρp is the particle
density. HCS is the core height (height of the control surface, Fig. 8.18),
Vθmax is the maximum tangential velocity, that occur at the edge of the
control surface CS. In this model, the value of core diameter dc, and the
tangential velocity at the core edge; Vθmax are calculated from regression
of experimental data using Eq. A.34.
Vθmax = 6.1Vin
(a b/D2
)−0.61(Dx/D)
−0.74(Ht/D)
−0.33
(8.10)
dc = 0.52D(a b/D2
)−0.25(Dx/D)
1.53(8.11)
187
Chapter 8. Optimization
Figure 8.18: The control surface (core edge) used in the Iozia and Leith model [85].
where Vin is the area-average inlet velocity. The following expression ob-
tained from trigonometry relations and can be used to calculate HCS .
HCS =(R− Dx
2 )(Ht − h)
R− (Bc2 )
+ (h− S) if Bc > Dx
= (Ht − S) if Bc ≤ Dx (8.12)
It is clear from this model that the most important geometrical parameters
that affect the cyclone cut-off diameter are the vortex finder diameter,the
ratio of inlet area to exit area, cyclone height.
Based on this model, the cut-off diameter is a function of the inlet gas ve-
locity (i.e., a function of both gas volume flow rate, Barrel diameter, inlet
section height and width), gas viscosity and particle density. For this par-
ticular study presented in Fig.8.20. The following values have been used:
Barrel diameter D =0.1 m, air flow rate =0.8333 l/s, air viscosity 1.0E-5 Pa
s and particle density 860 kg/m3. This means that the obtained results will
be valid for this particular case (used for just demonstration). However,
the authors believe the variation of the cut-off diameter due to variations
of cyclone geometrical dimensions is superior to the effect of these operat-
ing parameters, which is quite difficult to cover their range of operating
conditions.
The cut-off diameter x50 for cyclone separator is always given in units
of µm. Another way to represent x50 is using a dimensionless number;
Stokes number. The Stokes number based on the cut-off diameter; Stk50 =ρpx
250Vin/(18µD) [37]. It is the ratio between the particle relaxation time;
ρpx250/(18µ) and the gas flow integral time scale; D/Vin.
188
8.4. Multi-objective optimization using GA
Table 8.17: Validation of the used RBFNN to model the Euler number∗
x y
Average 23.268 23.268Minimum 2.3 1.745
Maximum 155.3 155.985
Range 153.0 154.24
Correlation Coefficient, R 0.999
Mean squared error, E2 1.311E-4Intercept 0.017
Slope 0.999∗ x is the input to the RBFNN and y is the predicted value.
Both x and y represent the Euler number.
The values of R, E2, intercept and slope are that for the testing stage.
Fitting the ANNs
Tables 8.17 and 8.18 present more details about the validation of the used
RBFNNs. Both the average, standard deviation, minimum, maximum and
range of the input and the predicted values are given. It is clear from
Tables 8.17 and 8.18 that the ANNs preserved the descriptive statistical
parameters of the input data. The correlation coefficient between the input
and the output and the mean squared error are given for each RBFNN.
The intercept and the slope of the adjusted line between the input and the
predicted value of the ANN are also given.
The configured RBFNN predictions versus experimental data for the Euler
number and the Iozia and Leith mathematical model for the Euler number
and cut-off diameter are shown in Fig. 8.19. According to Fig. 8.19, it can
be seen that the ANN models are able to attain the high training accu-
racy. The training mean square errors for the experimental values and the
Iozia and Leith model have the values 1.311E-4 and 3.258E-4 respectively
(Fig. 8.19). This indicates that, compared with traditional models of curve
fitting, the models based on artificial intelligence algorithm have a supe-
rior capability of nonlinear fitting. Especially, the RBFNN has its unique
and optimal approximation characteristics in learning process [197].
Figure 8.19 illustrates the agreement between the ANNs input and out-
put. The obtained relation is a typical linear relation with a coefficient of
correlation close to 1 (R > 0.999). The agreement between the input and
output of the ANN is also clear from the value of the mean squared error
E2. That means, the trained neural networks predict very well both the
Euler number and cut-off diameter values and can be used in cyclone de-
sign and performance estimation. Tables 8.17, 8.18 and Fig. 8.19 present
different performance indicators as a validation of the proposed model for
experimental values.
189
Chapter 8. Optimization
Table 8.18: Validation of the used RBFNN to model the cut-off diameter∗
x y
Average 8.219 8.219Minimum 3.64 3.716
Maximum 15.3 15.405
Range 11.66 11.689
Correlation Coefficient, R 0.99915
Mean squared error, E2 3.258E-4Intercept 0.014
Slope 0.999∗ x is the input to the RBFNN and y is the predicted value.
Both x and y represent the cut-off diameter.
The values of R, E2 , intercept and slope are that for the testing stage.
Input value (x)
Pre
dict
edva
lue
(y)
40 80 120
40
80
120
Data pointLinear fit
y=0.999 x + 0.0167
R=0.99964
E2= 1.311E-4
(a) The Euler number
Input value (x)
Pre
dict
edva
lue
(y)
4 6 8 10 12 14
4
6
8
10
12
14
Data pointLinear fit
y=0.998 x + 0.014
R=0.99915
E2= 3.258E-4
(b) The cut-off diameter
Figure 8.19: Linear regression of the RBFNNs for the Euler number and the cut-off
diameter.
8.4.1.1 The effect of geometrical parameters on the cut-off diam-
eter based on RBFANN
The effect of the geometrical parameters on the Euler number based on
the trained RBFNN has been presented in Sec. 8.3.2, page 166. The study
acknowledge the significant effect of the vortex finder diameter Dx and
the vortex finder length S, the inlet width b and the total height Ht. Less
effect is due to the cylinder height h (for h > 2.5) and the inlet height a (for
a > 0.55).
The effects of the geometrical parameters on the cut-off diameter are de-
picted in Fig. 8.20. To study the effect of each parameter, the tested RBFNN
model has been used by varying one parameter at a time from its minimum
to maximum values of the available 98 dataset, while the other parameters
190
8.4. Multi-objective optimization using GA
Dx, a, b, B c
h, S
Ht
X50
[mic
ron]
0.2 0.4 0.6 0.8 1
0.5 1 1.5 2 2.5 3 3.5
2 4 6 8 10
0
2
4
6
8
10
12
14
Dx
abSHt
hBc
Figure 8.20: The effect of geometrical parameters on the cut-off diameter based on
the Iozia and Leith model [85]. Note: The plotted curves are obtained for a test
case with the following settings, Barrel diameter=0.1 m, air flow rate=0.8333l/s,
air viscosity=1.0E-5 Pa s, particle density=860 kg/m3.
are kept constant at their mean values (cf. Table 8.7). Figure 8.20 indicates
the significant effect of the vortex finder diameter Dx and the vortex finder
length S, the inlet width b, the inlet height a and the total height Ht. Less
effect is due to the cylinder height h and the cone tip diameter Bc. More
analysis is given in Table 8.19.
The significant geometrical parameters on the cut-off diameter
(Stokes number) using the response surface methodology (RSM)
Table 8.20 represents the parameters ranges selected for the seven geo-
metrical parameters. The study was planned using Box–Behnken design,
with 64 combinations. A significant level of P < 0.05 (95% confidence) was
used in all tests. Analysis of variance (ANOVA) was followed by an F-test
of the individual factors and interactions [53].
191
Chapter 8. Optimization
Table 8.19: The variation of the cut-off diameter with cyclone dimensions using the
RBFNN model (cf., Fig. 8.20)
Factor Analysis
Dx
The vortex finder diameter has the most significant effect on
the cut-off diameter x50 (the highest slope in Fig. 8.20). The
slope is very high until Dx = 0.5 and any further increase in
Dx produces a small change in x50. In general, increasing Dx
increases x50 (decreasing the collection efficiency), this is one
of the main reasons of the trade-off between the Euler number
and the cut-off diameter objectives. This makes the
optimization of cyclone geometry a multi-objective procedure.
b
The variation of x50 with the inlet width is similar in trend
and significance to that for Dx but here the slope changes at
b = 0.25.
S and a
The effect of the vortex finder length and the inlet section
height on the cut-off diameter is almost paralleled up to
S = 1.5 and a = 0.6 afterwards they lose their significance and
become nearly constant.
h
Increasing the barrel height slightly decreases the cut-off
diameter with nearly linear relation. This trend has been
reported by other researchers using CFD simulations, e.g.,
Elsayed and Lacor [51].
Ht∗
The effect of the cyclone total height is basically due to two
effects the cone height and barrel height. The curve can be
subdivided into four main regions. Sharp decrease in x50 up to
Ht = 3.25, no valuable difference between 3.25 and 5.25, sharp
increase between 5.25 and 8, and insignificant effect beyond 8.
Bc
The effect of the cone-tip diameter on the cut-off diameter is
quite small. First, increasing the cone-tip diameter slightly
decreases the cut-off diameter up to Bc = 0.55 and any further
increment increases the cut-off diameter. This trend has been
reported by other researchers, e.g., Elsayed and Lacor [46, 55].
∗ Due to the interaction between the geometrical parameters, especially between Ht with S and Dx (cf.,Fig. 8.22 and Table 8.21), the obtained conclusions may not be applicable generally and the application ofresponse surface methodology to analysis the effect of each particular parameter must.
192
8.4. Multi-objective optimization using GA
Table 8.20: The values of the independent variables used in the design of experi-
ment
Variables minimum center maximum
Inlet height, a=X1 0.4 0.55 0.7Inlet width, b=X2 0.14 0.27 0.4Cone tip diameter, Bc=X3 0.2 0.3 0.4Vortex finder diameter, Dx=X4 0.2 0.475 0.75Barrel height, h=X5 1.0 1.5 2.0Total cyclone height, Ht=X6 3.0 5.0 7.0Vortex finder length, S=X7 0.4 1.2 2.0
Analysis of variance (ANOVA)
Analysis of variance (ANOVA) showed that the resultant quadratic polyno-
mial models adequately represented the experimental data with the coef-
ficient of multiple determination R2 being 0.984099 (cf., Table 8.21). This
indicates that the quadratic polynomial model obtained was adequate to
describe the influence of the independent variables studied [189]. Analy-
sis of variance (ANOVA) was used to evaluate the significance of the coef-
ficients of the quadratic polynomial models (see Table 8.21). For any of the
terms in the models, a large F-value (small P-value) would indicate a more
significant effect on the respective response variables [50, 53].
Based on the ANOVA results presented in Table 8.21, the variable with the
largest effect on the Stokes number (cut-off diameter) was the linear term
of vortex finder diameter, followed by the linear term of the cyclone total
height, the vortex finder length and the inlet width (P < 0.05); the other
three linear terms (inlet height, barrel height, and cone tip diameter) did
not show a significant effect (P > 0.05). The quadratic term of cyclone
total height, vortex finder diameter and vortex finder length also had a
significant effect on the pressure drop; however, the effect of the other four
quadratic terms was insignificant. Furthermore, the interaction between
Dx with (Ht, S, b) and between S with (h, Ht) also had a significant ef-
fect on the Stokes number, while the effect of the remaining terms was
insignificant.
Analysis of response surfaces
For visualization of the calculated factor, main effects plot, Pareto chart
and response surface plots were drawn. The slope of the main effect curve
is proportional to the size of the effect and the direction of the curve speci-
fies a positive or negative influence of the effect [50, 61], Fig.8.21(a). Based
on the main effect plot, the most significant factors on the Stokes number
are: (1) the vortex finder diameter Dx, with a second–order curve of di-
rect relation. (2) the cyclone total height Ht inversely related to the Stokes
number. (3) the vortex finder length S with direct relationship. (4) the inlet
193
Chapter 8. Optimization
dimensions width b and height a inversely related to the Stokes number.
Whereas the other factors have an insignificant effect. The main effect plot
supports the analysis given in Table 8.19, except for Ht where the strong
interaction between the cyclone total height and the vortex finder length
affected the trend given in Fig. 8.20.
Pareto charts were used to summarize graphically and display the relative
importance of each parameter with respect to the Stokes number [50]. The
Pareto chart shows all the linear and second-order effects of the parame-
ters within the model and estimates the significance of each with respect
to maximizing the Stokes number response. A Pareto chart displays a
frequency histogram with the length of each bar proportional to each es-
timated standardized effect [30]. The vertical line on the Pareto charts
judges whether each effect is statistically significant within the generated
response surface model; bars that extend beyond this line represent ef-
fects that are statistically significant at a 95% confidence level. Based on
the Pareto chart (Fig. 8.21(b)) and ANOVA table (Table 8.21) there are five
significant parameters at a 95% confidence level: the vortex finder diame-
ter Dx, the total cyclone height Ht, the vortex finder length S and the inlet
dimensions a and b. Therefore, the Pareto chart is a perfect supplement to
the main effect plot.
To visualize the effect of the independent variables on the dependent ones,
surface response of the quadratic polynomial models were generated by
varying two of the independent variables within the experimental range
while holding the other factors at their central values (cf., Table 8.20) [189]
as shown in Fig. 8.22. Thus, Fig. 8.22(a) was generated by varying the
total height Ht and the vortex finder length S while keeping the other
five factors constant. The response surface plots presented in Fig. 8.22
illustrate the strong interactions between Ht with (S and Dx) and Dx with
(S and b).
8.4.2 Single objective optimization using GA
The genetic algorithm optimization technique has been applied to obtain
the geometrical ratios for minimum pressure drop (Euler number). The
objective function is the Euler number (using the trained radial basis func-
tion neural network presented in Sec. 8.3.1.1, page 163). The design vari-
ables are the seven geometrical dimensions of the cyclone separator.
194
8.4. Multi-objective optimization using GA
(a) Main effects plot
(b) Pareto chart. A=a, B=b, C=Bc, D=Dx, E=h, F=Ht, G=S,
FG=Ht*S, etc.
Figure 8.21: Analysis of design of experiment for the Stokes number
195
Chapter 8. Optimization
(a) Ht versus S
(b) Dx versus Ht
(c) Dx versus S
(d) b versus Dx
Figure 8.22: The response surface plots for the Stokes number. Note: the stokes
number values are multiplied by 1000.
196
8.4. Multi-objective optimization using GA
GA settings
Table 8.22 presents the settings used to obtain the optimum design for
minimum pressure using global optimization Matlab toolbox (Matlab 2010a
commercial package). The evolution of the cost function for the best indi-
vidual is given in Fig. 8.23. After 70 iterations (14400 function counts), the
Matlab code stops generating a new population since the average change
in the fitness value becomes less than 1E-6. The total calculation time for
this optimization problem was 102 seconds. Table 8.23 gives the optimum
values for cyclone geometrical parameters for minimum pressure drop es-
timated by the artificial neural network using the genetic algorithm op-
timization technique. It is clear from Table 8.23 that the new optimized
design is very close to the Stairmand design in many geometrical param-
eters, whereas the new ratios will result in minimum pressure drop. To
understand the effect of this small change in the geometrical ratios on the
flow field pattern and performance, a CFD study for the two designs is
needed [53].
Generation
Fitn
ess
valu
e
10 20 30 40 50 60 704
4.5
5
5.5
6
6.5
7
7.5
8
Figure 8.23: Evolution of the cost function for the best individual
8.4.2.1 Comparison between the two cyclone designs using CFD
Numerical settings
The air volume flow rate Qin=0.08 m3/s for the two cyclones (inlet velocity
for Stairmand design is 19 m/s and 16 m/s for the new design), air density
197
Chapter 8. Optimization
1.0 kg/m3 and dynamic viscosity of 2.11E-5 Pa s. the turbulent intensity
equals 5% and characteristic length equals 0.07 times the inlet width [75].
Velocity inlet boundary condition is applied at inlet, outflow at gas outlet
and wall boundary condition at all other boundaries [53].
The finite volume method has been used to discretize the partial differen-
tial equations of the model using the SIMPLEC (Semi-Implicit Method for
Pressure-Linked Equations-Consistent) method for pressure velocity cou-
pling and QUICK scheme to interpolate the variables on the surface of the
control volume. The implicit coupled solution algorithm was selected. The
unsteady Reynolds stress turbulence model (RSM) was used in this study
with a time step of 0.0001 s.
The grid refinement study using different levels of grid shows that a total
number of 134759 hexahedral cells for the Stairmand cyclone and 378963
hexahedral cells for the new design are sufficient to obtain a grid indepen-
dent solution, and further mesh refinement yields insignificant changes in
the numerical solution. The hexahedral meshes have been obtained using
the GAMBIT commercial software. These simulations were performed on
an 8 nodes CPU Opteron 64 Linux cluster using Fluent 6.3.26 commer-
cial software. The geometrical values for the two cyclones are given in
Table 8.24.
Results and discussion
The pressure field
Figure 8.24 shows the contour plot at Y=0. In the two cyclones, the time-
averaged static pressure decreases radially from the wall to center. A neg-
ative pressure zone appears in the forced vortex region (central region)
due to high swirling velocity. The pressure gradient is largest along the
radial direction, while the gradient in the axial direction is very limited.
The cyclonic flow is not symmetrical as is clear from the shape of the low-
pressure zone at the cyclone center (twisted cylinder). However, the two
cyclones have almost the same flow pattern, but the highest pressure of
the Stairmand design is nearly 1.5 times that of the new design, implying
that the new design has a lower pressure drop.
The pressure distributions presented in Figs. 8.25 and 8.26 of the two cy-
clones at sections S1–S6 depict the two parts pressure profile (for Rankine
vortex). Again, the highest static pressure for Stairmand design is more
than 1.5 times that of the new design at all sections while the central value
is almost the same for the two cyclones irrespective of the section location.
198
8.4. Multi-objective optimization using GA
Figure 8.24: The contour plots for the time averaged flow variables at sections Y=0.
From top to bottom: Stairmand design and the new design respectively. From left
to right: the static pressure (N/m2), the tangential and axial velocity (m/s). Note:
both cyclones have the same barrel diameter and air volume flow rate.
199
Chapter 8. Optimization
Distance from center (m)
Sta
ticpr
essu
re(N
/m2 )
-0.1 -0.05 0 0.05 0.10
500
1000
1500
2000
2500New designStairmand design
Distance from center (m)
Tan
gent
ialv
eloc
ity/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1
0
0.5
1
1.5
2
2.5New designStairmand design
Distance from center (m)
Axi
alve
loci
ty/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1-0.5
0
0.5
1New designStairmand design
Distance from center (m)
Sta
ticpr
essu
re(N
/m2 )
-0.1 -0.05 0 0.05 0.10
500
1000
1500
2000
2500New designStairmand design
Distance from center (m)
Tan
gent
ialv
eloc
ity/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1
0
0.5
1
1.5
2
2.5New designStairmand design
Distance from center (m)
Axi
alve
loci
ty/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1-0.5
0
0.5
1New designStairmand design
Distance from center (m)
Sta
ticpr
essu
re(N
/m2 )
-0.1 -0.05 0 0.05 0.10
500
1000
1500
2000
2500New designStairmand design
Distance from center (m)
Tan
gent
ialv
eloc
ity/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1
0
0.5
1
1.5
2
2.5New designStairmand design
Distance from center (m)
Axi
alve
loci
ty/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1-0.5
0
0.5
1New designStairmand design
Figure 8.25: The radial profile for the time averaged tangential and axial velocity
at different sections on the X-Z plane (Y=0) at sections S1–S3 (cf., Table 8.25). From
top to bottom: section S1–S3. From left to right: time-averaged static pressure,
tangential and axial velocity respectively.
This indicates that, the new design has a lower pressure drop with respect
to the Stairmand design.
The velocity field
Based on the contour plots of the time–averaged tangential velocity pre-
sented in Fig. 8.24, and the radial profiles at sections S1–S6 shown in
Figs. 8.25 and 8.26, the following conclusions can be drawn. The tangen-
tial velocity profile at any section is composed of two regions, an inner and
an outer one. In the inner region, the flow rotates approximately like a
solid body (forced vortex), where the tangential velocity increases with ra-
dius. After reaching its peak the velocity decreases with radius in the outer
200
8.4. Multi-objective optimization using GA
Distance from center (m)
Sta
ticpr
essu
re(N
/m2 )
-0.1 -0.05 0 0.05 0.10
500
1000
1500
2000
2500New designStairmand design
Distance from center (m)
Tan
gent
ialv
eloc
ity/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1
0
0.5
1
1.5
2
2.5New designStairmand design
Distance from center (m)
Axi
alve
loci
ty/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1-0.5
0
0.5
1New designStairmand design
Distance from center (m)
Sta
ticpr
essu
re(N
/m2 )
-0.1 -0.05 0 0.05 0.10
500
1000
1500
2000
2500New designStairmand design
Distance from center (m)
Tan
gent
ialv
eloc
ity/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1
0
0.5
1
1.5
2
2.5New designStairmand design
Distance from center (m)
Axi
alve
loci
ty/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1-0.5
0
0.5
1New designStairmand design
Distance from center (m)
Sta
ticpr
essu
re(N
/m2 )
-0.1 -0.05 0 0.05 0.10
500
1000
1500
2000
2500New designStairmand design
Distance from center (m)
Tan
gent
ialv
eloc
ity/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1
0
0.5
1
1.5
2
2.5New designStairmand design
Distance from center (m)
Axi
alve
loci
ty/I
nlet
velo
city
-0.1 -0.05 0 0.05 0.1-0.5
0
0.5
1New designStairmand design
Figure 8.26: The radial profile for the time–averaged tangential and axial velocity
at different sections on the X-Z plane (Y=0) at sections S4–S6 (cf., Table 8.25). From
top to bottom: section S4–S6. From left to right: time-averaged static pressure,
tangential and axial velocity respectively.
part of the profile (free vortex). This profile is a so-called Rankine type vor-
tex. The maximum tangential velocity may reach twice the average inlet
velocity and occurs in the annular cylindrical part. The tangential veloc-
ity distributions for the two cyclones are approximately nearly identical in
pattern and values (dimensionless), with the highest velocity occurring at
1/4 of the cyclone radius for both cyclones. This implies a nearly equal col-
lection efficiency for both cyclones, as the centrifugal force is the main driv-
ing force for particle collection in the cyclone separator. The axial velocity
profiles for the two cyclones are also very close, exhibiting the inverted W
axial velocity profile.
201
Chapter 8. Optimization
Particle diameter [micron]
Col
lect
ion
effic
ienc
y[-]
10-1 100 101
0.25
0.5
0.75
1Stairmand designNew design
Figure 8.27: The grade efficiency curves for the two designs
The DPM results
In order to calculate the cut-off diameters of the two cyclones, 104 particles
were injected from the inlet surface with a velocity equals the air inlet ve-
locity and particle density ρp = 860kg/m3. The grade efficiency curves for
the two designs are plotted in Fig. 8.27. The DPM analysis results and the
pressure drops for the two cyclones are depicted in Table 8.26. Although,
the difference between the two cyclone cut-off diameters is small, the sav-
ing in pressure drop is considerable (nearly 32.5% the value of Stairmand
cyclone).
Based on the flow pattern analysis and the DPM results, one can conclude
that the cyclone collection efficiency for the two cyclones is very close, with
the advantage of low pressure drop in the new design. The authors want
to emphasis that only small changes in the geometrical dimensions of the
two designs lead to this improvement in the performance.
8.4.3 Optimal cyclone design for best performance
NSGA-II settings
Table 8.27 presents the genetic operators and parameters for multi-objective
optimization. The Euler number values have been obtained from the arti-
ficial neural network trained by experimental values. The Stokes number
values are obtained from Iozia and Leith model [85]. In order to investi-
gate the effect of different geometrical and operational parameters on the
Pareto front, sixteen test cases with different barrel diameter, gas flow rate
and particle density have been tested, cf. Table 8.28. The sixteen test cases
covers: 1) Two barrel diameters, 31 mm and 205 mm. 2) Four levels of air
flow rates, 50, 60, 70 and 80 l/min. 3) Five values of particle density, 860,
202
8.4. Multi-objective optimization using GA
1000, 1500, 1750 and 2000 kg/m3.
Pareto front
The Pareto front (non dominated points) for the sixteen test cases are pre-
sented in Fig. 8.28(a). Figure 8.28(a) clearly demonstrate tradeoffs in ob-
jective functions Euler number and Stokes number from which an appro-
priate design can be compromisingly chosen by the designer [148]. All the
optimum design points in the Pareto front are non-dominated and could be
chosen by a designer as optimum cyclone separator [148]. The correspond-
ing geometrical ratios of the Pareto front shown in Fig. 8.28(a) are given
in Table 8.29 for test case 1 and Table 8.30 for test case 9. Three points
A, B and C are indicated in Figs. 8.28(c) and 8.28(c) and Tables 8.29 and
8.30. Point A indicates the point of minimum Euler number and maximum
Stokes number. Point B indicates the point of maximum Euler number and
minimum Stokes number. Point C indicates an optimal point for the multi-
objective optimization problem.
In order to obtain the Euler number- Stokes number relationship, Fig. 8.28(b)
has been drawn. It indicates a general relationship (trend) between the
two dimensionless numbers irrespective to the barrel diameter, gas flow
rate, particle density. A second-order polynomial has been fitted between
the logarithms of Euler number and Stokes number, Eq. 8.13. The ob-
tained correlation can fit the data with a coefficient of correlation R2 =0.98643 as shown in Fig. 8.28(b).
Stk50 = 100.3016(log10(Eu))2−0.9479log10(Eu)−2.5154 (8.13)
Bubble plots for Pareto front
For visual inspection of the effect of the seven geometrical parameters on
the two conflicting performance parameters, the bubble plots on Pareto
front points have been drawn for each geometrical parameter. However,
only figures for test case 1 (Fig. 8.29) and 9 (Fig. 8.30) are presented, but
all other cases depict the same results (trend).
Figure 8.29 indicates that: a) Decreasing the vortex finder diameter Dx de-
creases the Stokes number and increases the Euler number, Fig. 8.29(b).
b) Generally speaking, increasing the inlet height a increases the Euler
number and decreases the Stokes number. c) A similar trend is exhibited
203
Chapter 8. Optimization
by the inlet width b but due to interaction with other geometrical and op-
erational variables, one could see a range of bubble sizes in the region of
best performance (lower values for both the Euler and Stokes numbers).
d) The higher values of total cyclone height Ht will produce less Stokes
number, intermediate values could produce less Euler number, smaller-
intermediate values could produce the optimum performance due to inter-
action with other variables. e) Short barrels will produce better collection
efficiency (low Stokes number) and higher Euler numbers. Intermediate
values results in low Euler number values. Long barrels can produce the
best performance. f) Short vortex finder may produce higher values of Eu-
ler numbers or higher values of Stokes number due to strong interaction
with other variables. Long vortex finder can produce the optimum perfor-
mance. g) Generally speaking, the variation of the cone-tip diameter Bc
has no effect on the performance parameter. The above comments is re-
stricted to the range of each geometrical variables located on the Pareto
front and not for the whole range of values (cf., Fig. 8.29 for the range of
each geometrical parameters).
8.4.4 Conclusions
To predict the complex non-linear relationships between the performance
parameters and the geometrical dimensions, two radial basis neural net-
works (RBFNNs) are developed and employed to model the Euler number
and Stokes number for cyclone separators. The neural networks have been
trained and tested by the experimental data available in literature for Eu-
ler number (pressure drop) and Iozia and Leith model [85] for the Stokes
number (cut-off diameter). The effects of the seven geometrical parame-
ters on the Stokes number have been investigated using the trained ANN.
To declare any interaction between the geometrical parameters affecting
the Stokes number, the response surface methodology has been applied.
The trained ANN has been used as an objective function to obtain the
cyclone geometrical ratios for minimum Euler number using the genetic
algorithms optimization technique. A CFD comparison between the new
optimal design and the Stairmand design using the Reynolds stress turbu-
lence model has been performed. A multi-objective optimization technique
using NSGA-II technique has been applied to determine the Pareto front
for the best performance cyclone separator.
The following conclusions can be drawn from analysis of the obtained re-
sults:
204
8.4. Multi-objective optimization using GA
• The result demonstrates that artificial neural networks can offer an
alternative and powerful approach to model the cyclone performance.
• The analysis indicates the significant effect of the vortex finder di-
ameter Dx and the vortex finder length S, the inlet width b, the inlet
height a and the total height Ht on the cyclone performance.
• The response surface methodology has been used to fit a second-order
polynomial to the RBFNN for the cut-off diameter. The analysis of
variance of the cut-off diameter indicates a strong interaction be-
tween Dx with (Ht, S, b) and between S with (h, Ht).
• The trained RBFNN has been used to get a new optimized cyclone for
minimum pressure drop (Euler number) using the genetic algorithm
optimization technique.
• A comparison between the new design and the standard Stairmand
design has been performed using CFD simulation with the Reynolds
stress turbulence model to get a clear vision of the flow field pattern
and performance in the new design.
• CFD results shows that, the new cyclone design are very close to the
Stairmand high efficiency design in the geometrical parameter ratio,
and superior in low pressure drop at nearly the same cut-off diameter.
• The new cyclone design results in nearly 68% of the pressure drop
obtained by the old Stairmand design at the same volume flow rate.
This confirms that the obtained design using the genetic algorithm is
better than that obtained using Nelder-Mead technique which results
in 75% of the Stairmand pressure drop [53].
• The two trained RBFNNs have been used in a multi-objective opti-
mization process using NSGA-II technique. Sixteen test cases with
different barrel diameter, gas flow rate and particle density have
been tested and plotted. The Pareto fronts for the 16 test cases are
very close. A second-order polynomial has been fitted between the
logarithms of Euler number and Stokes number to obtain a general
formula, Stk50 = 100.3016(log10(Eu))2−0.9479 log10(Eu)−2.5154 with a coeffi-
cient of correlation R2 = 0.98643. This formula can be used to obtain
the Stokes number if the Euler number is known.
205
Chapter 8. Optimization
Table 8.21: Analysis of variance and the regression coefficients of the fitted
quadratic equation for the Stokes numbera
Source Regression coefficient F-Ratio P-Value
β0 -0.0470554Linearβ1 0.942933 6.65 0.0154β2 -1.3178 24.12 0.0000β3 2.10188 0 1β4 8.3493 1145.05 0.0000β5 -0.843633 0 1β6 -0.527695 243.98 0.0000β7 1.46453 137.75 0.0000Quadraticβ11 -1.17696 0.21 0.6489β22 1.44904 0.18 0.6737β33 -3.50314 0.37 0.5476β44 2.33086 9.38 0.0048β55 0.281211 1.49 0.232β66 0 39.06 0.0000β77 0.243505 7.33 0.0114Interactionβ12 0.726218 0.03 0.8558β13 0 0 1β14 -2.09219 1.25 0.2732β15 0 0 1β16 0.121646 0.22 0.6401β17 -0.187708 0.09 0.7726β23 0 0 1β24 -4.84457 5.03 0.0330β25 0 0 1β26 0.262495 0.78 0.3842β27 -0.466053 0.39 0.5352β34 0 0 1β35 0 0 1β36 0 0 1β37 0 0 1β45 0 0 1β46 -0.869946 38.41 0.0000β47 1.44156 16.87 0.0003β56 0 0 1β57 0 39.06 0.0000β67 -0.370128 58.83 0.0000
R2 0.984099
a Bold numbers indicate significant factors as identified by the analysis of variance (ANOVA) at the 95%confidence level.
Table 8.22: Genetic operators and parameters for single objective optimization
Population type: Double vectorInitial range: [0.2 0.1 0.1 0.3 2.0 0.65 0.05 ; 0.8 0.8 0.6 0.8 8.0 2.5 0.75]Fitness scaling: RankSelection operation: Tournament (tournament size equals 4)Elite count: 2Crossover fraction: 0.8Crossover operation: Intermediate crossover with the default value of 1.0Mutation operation: The constraint dependent defaultMaximum number of generations (iterations): 1400Population size: 200
206
8.4. Multi-objective optimization using GA
Table 8.23: The optimized cyclone separator design for minimum pressure drop
Factor Low High Stairmand design Optimum design
Dx 0.2 0.75 0.5 0.549a 0.5 0.75 0.5 0.595b 0.14 0.4 0.2 0.201S 0.4 2.0 0.5 0.595Ht 3.0 7.0 4.0 4.549h 1.0 2.0 1.5 1.411Bc 0.2 0.4 0.375 0.275
Table 8.24: The geometrical parameters for the two designs (D=0.205 m)∗
Cyclone a/D b/D Dx/D Ht/D h/D S/D Bc/DStairmand design 0.5 0.2 0.5 4 1.5 0.5 0.375New design 0.595 0.201 0.549 4.549 1.411 0.595 0.275
∗ The outlet section is above the cyclone surface by Le = 0.618D. The inlet section located at adistance Li = D from the cyclone center.
Table 8.25: The position of different sectionsa
Section S1 S2 S3 S4 S5 S6
z`/D b 2.75 2.5 2.25 2.0 1.75 1.5
a Sections S1–S5 are located in the conical section, section S6 at the cylindrical part.b z` is measured from the top of the inlet section.
Table 8.26: The performance parameters for the two cyclones
Design Method ∆p (N/m2) Euler number x50 (µm) Stokes number x 103
CFD 803 6.338 0.804 0.114
New design ANN 584.4 4.613 2.938 1.815
Ramachandran model [139] 877.98 5.523
CFD 1190 6.567 1.0 0.209
Stairmand design ANN 1015.8 5.606 3.314 1.931Ramachandran model [139] 699.66 4.846
Table 8.27: Genetic operators and parameters for multi-objective optimization
Population type: Double vectorPopulation size: 105 (i.e., 15* number of variables which is the default in Matlab)Initial range: [0.2 0.1 0.1 0.3 2.0 0.65 0.05 ; 0.8 0.8 0.6 0.8 8.0 2.5 0.75]Selection operation: tournament (tournament size equals 2)Crossover fraction: 0.8Crowding distance fraction 0.35Crossover operation: Intermediate crossover with the default value of 1.0Number of generations (iterations): 1400 (i.e., 200* number of variables which is the default in Matlab)
207
Chapter 8. Optimization
Table 8.28: The diameters, air flow rates and the particle densities for the sixteen
test cases
Case D [mm] Q [l/min] ρp [kg/m3]
1 205 50 860
2 205 60 860
3 205 70 860
4 205 80 860
5 205 50 1000
6 205 50 1500
7 205 50 1750
8 205 50 2000
9 31 50 860
10 31 60 860
11 31 70 860
12 31 80 860
13 31 50 1000
14 31 50 1500
15 31 50 1750
16 31 50 2000
208
8.4. Multi-objective optimization using GA
Table 8.29: The seven geometrical parameters and the obtained Euler number and
Stokes number for the nondominated points (Pareto-front) for test case 1 (cf. Table
8.28)
point Dx a b S Ht h Bc Euler number Stokes number x 103
1 0.306 0.659 0.385 0.410 6.957 1.779 0.387 23.843 0.552
2B 0.306 0.688 0.398 0.404 6.993 1.779 0.317 27.322 0.540
3 0.618 0.229 0.213 0.411 6.774 1.885 0.495 1.026 2.785
4 0.326 0.331 0.240 0.451 6.668 1.885 0.444 6.485 0.798
5C 0.360 0.295 0.253 0.443 6.664 1.910 0.459 4.892 0.937
6 0.670 0.230 0.213 0.410 6.777 1.859 0.496 0.879 3.478
7 0.585 0.226 0.217 0.419 6.763 1.901 0.492 1.159 2.089
8 0.306 0.685 0.299 0.401 6.985 1.789 0.425 18.474 0.576
9 0.306 0.666 0.361 0.407 6.965 1.782 0.398 22.300 0.55810 0.305 0.516 0.271 0.439 6.840 1.868 0.428 12.977 0.633
11 0.303 0.286 0.318 0.449 6.611 1.930 0.462 8.404 0.699
12 0.476 0.240 0.217 0.434 6.743 1.901 0.471 1.915 1.524
13 0.312 0.622 0.277 0.419 6.943 1.909 0.429 14.816 0.619
14 0.308 0.686 0.379 0.427 6.974 1.781 0.459 23.145 0.55415 0.592 0.229 0.213 0.413 6.831 1.895 0.494 1.117 2.491
16 0.320 0.570 0.273 0.423 6.898 1.909 0.434 12.767 0.658
17 0.598 0.229 0.213 0.412 6.770 1.893 0.494 1.095 2.565
18 0.309 0.441 0.266 0.450 6.769 1.881 0.435 10.629 0.674
19 0.430 0.307 0.231 0.438 6.735 1.898 0.465 3.211 1.22920 0.306 0.667 0.389 0.417 6.950 1.779 0.367 24.941 0.550
21 0.306 0.392 0.277 0.441 6.708 1.882 0.443 9.963 0.676
22 0.326 0.331 0.209 0.451 6.668 1.900 0.471 5.515 0.823
23 0.308 0.686 0.348 0.427 6.974 1.807 0.459 21.132 0.564
24 0.680 0.229 0.213 0.409 6.787 1.857 0.495 0.849 3.640
25 0.307 0.653 0.296 0.421 6.972 1.816 0.420 17.606 0.58826 0.307 0.368 0.338 0.439 6.708 1.895 0.430 11.399 0.659
27 0.514 0.235 0.217 0.429 6.709 1.901 0.479 1.596 1.718
28 0.310 0.596 0.275 0.424 6.918 1.899 0.429 14.387 0.622
29 0.559 0.246 0.221 0.453 6.785 1.898 0.455 1.486 1.915
30 0.399 0.308 0.229 0.438 6.707 1.891 0.460 3.732 1.10231 0.516 0.244 0.241 0.422 6.728 1.904 0.486 1.797 1.674
32 0.413 0.256 0.280 0.444 6.667 1.927 0.472 3.510 1.162
33 0.306 0.507 0.352 0.437 6.811 1.848 0.460 16.110 0.602
34 0.307 0.644 0.333 0.423 6.960 1.807 0.401 19.929 0.575
35 0.646 0.229 0.219 0.409 6.797 1.863 0.497 0.963 3.113
36A 0.692 0.228 0.213 0.408 6.819 1.855 0.498 0.815 3.805
37 0.558 0.216 0.230 0.439 6.802 1.922 0.486 1.311 1.945
Minimum 0.303 0.216 0.209 0.401 6.611 1.779 0.317 0.815 0.540
Maximum 0.692 0.688 0.398 0.453 6.993 1.930 0.498 27.322 3.805
A indicates the point of minimum Euler number and maximum Stokes number. B indicates the point ofmaximum Euler number and minimum Stokes number. C indicates an optimal point for themulti-objective optimization problem. (cf. Fig. 8.28(c)).
209
Chapter 8. Optimization
Table 8.30: The seven geometrical parameters and the obtained Euler number and
Stokes number for the nondominated points (Pareto-front) for test case 9 (cf. Table
8.28)
point Dx a b S Ht h Bc Euler number Stokes number x 103
1A 0.686 0.236 0.236 0.478 6.910 1.915 0.473 1.008 3.743
2 0.308 0.655 0.390 0.423 6.902 1.995 0.471 21.688 0.5583 0.308 0.692 0.390 0.423 6.902 1.995 0.472 22.898 0.551
4 0.585 0.235 0.235 0.418 6.925 1.727 0.471 1.355 2.436
5 0.314 0.330 0.236 0.431 6.837 1.864 0.470 6.568 0.750
6 0.309 0.486 0.335 0.426 6.869 1.931 0.471 13.947 0.621
7 0.578 0.237 0.235 0.421 6.684 1.975 0.471 1.356 2.0088 0.309 0.272 0.347 0.424 6.897 1.967 0.472 8.025 0.697
9 0.310 0.538 0.369 0.429 6.882 1.984 0.467 16.799 0.597
10 0.473 0.238 0.235 0.478 6.923 1.971 0.471 2.105 1.481
11 0.308 0.375 0.382 0.423 6.901 1.989 0.471 12.168 0.634
12 0.432 0.238 0.236 0.445 6.785 1.947 0.472 2.494 1.30013 0.321 0.327 0.236 0.431 6.839 1.862 0.470 6.223 0.777
14 0.308 0.606 0.383 0.424 6.896 1.990 0.471 19.736 0.571
15 0.308 0.643 0.381 0.424 6.895 1.986 0.472 20.775 0.565
16 0.407 0.255 0.240 0.470 6.899 1.907 0.472 3.119 1.165
17 0.309 0.333 0.353 0.424 6.882 1.958 0.471 10.021 0.665
18 0.452 0.236 0.236 0.448 6.793 1.945 0.472 2.255 1.39419 0.678 0.236 0.236 0.473 6.893 1.919 0.473 1.029 3.615
20 0.625 0.237 0.236 0.449 6.800 1.944 0.472 1.187 2.891
21 0.309 0.596 0.379 0.425 6.894 1.988 0.471 19.039 0.578
22 0.313 0.325 0.262 0.430 6.847 1.884 0.471 7.197 0.730
23 0.664 0.237 0.236 0.476 6.911 1.914 0.473 1.076 3.40424 0.309 0.365 0.345 0.425 6.876 1.946 0.471 10.740 0.656
25 0.608 0.235 0.235 0.438 6.921 1.797 0.472 1.261 2.691
26B 0.308 0.692 0.390 0.423 6.902 1.995 0.472 22.898 0.551
27 0.309 0.436 0.381 0.424 6.899 1.989 0.464 14.120 0.616
28 0.308 0.427 0.360 0.424 6.886 1.962 0.471 13.112 0.626
29C 0.330 0.290 0.252 0.465 6.890 1.955 0.471 5.639 0.818
30 0.658 0.245 0.239 0.477 6.910 1.924 0.473 1.146 3.290
31 0.308 0.362 0.365 0.424 6.883 1.668 0.467 11.924 0.64832 0.520 0.251 0.244 0.470 6.906 1.684 0.472 1.996 1.671
33 0.631 0.236 0.236 0.449 6.796 1.945 0.472 1.164 2.964
34 0.549 0.242 0.240 0.430 6.909 1.783 0.471 1.613 1.823
35 0.316 0.270 0.341 0.425 6.891 1.966 0.472 7.482 0.726
36 0.686 0.236 0.236 0.478 6.910 1.915 0.473 1.008 3.74337 0.380 0.239 0.237 0.477 6.890 1.922 0.473 3.327 1.071
Minimum 0.308 0.235 0.235 0.418 6.684 1.668 0.464 1.008 0.551Maximum 0.686 0.692 0.390 0.478 6.925 1.995 0.473 22.898 3.743
A indicates the point of minimum Euler number and maximum Stokes number. B indicates the point ofmaximum Euler number and minimum Stokes number. C indicates an optimal point for themulti-objective optimization problem. (cf. Fig. 8.28(d)).
210
8.4. Multi-objective optimization using GA
Euler number
Sto
kes
num
berx
10
3
0 5 10 15 20 25
0.5
1
1.5
2
2.5
3
3.5
4
Case 1Case 2Case 3Case 4Case 5Case 6Case 7Case 8Case 9Case 10Case 11Case 12Case 13Case 14Case 15Case 16
(a) Pareto fronts for 16 test cases,
linear scale
Euler number
Sto
kes
num
berx
10
3
5 10 15 20 25 30350.5
1
1.5
2
2.5
3
3.5
44.5
55.5
66.5
7
Pareto pointsStk50=10 0.3016[LOG10(Eu)] 2-0.9479 LOG10(Eu)-2.5154
(b) Pareto fronts for 16 test cases, log
scale
Euler number
Sto
kes
num
berx
103
5 10 15 20 25
0.5
1
1.5
2
2.5
3
3.5
4A
BC
(c) Pareto front for test case 1
Euler number
Sto
kes
num
berx
103
5 10 15 20 25
0.5
1
1.5
2
2.5
3
3.5
4A
BC
(d) Pareto chart for test case 9
Figure 8.28: Pareto fronts for different test cases.
211
Ch
ap
ter
8.
Op
timiz
atio
n
(a) Dx, range: 0.303 - 0.692 (b) a, range: 0.216 - 0.688 (c) b, range: 0.209 - 0.398
(d) Ht, range: 6.611 - 6.993 (e) h, range: 1.779 - 1.930 (f) S, range: 0.401 - 0.453
Figure 8.29: Bubble plots for different geometrical parameters for test case 1 (cf. Fig. 8.28(c) and Table 8.29).
212
8.4
.M
ulti-o
bje
ctive
op
timiz
atio
nu
sing
GA
(a) Dx, range: 0.308 - 0.686 (b) a, range: 0.235 - 0.692 (c) b, range: 0.235 - 0.390
(d) Ht, range: 6.684 - 6.925 (e) h, range: 1.668 - 1.995 (f) S, range: 0.418 - 0.478
Figure 8.30: Bubble plots for different geometrical parameters for test case 9 (cf. Fig. 8.28(d) and Table 8.30).
213
Chapter 8. Optimization
Table 8.31: The seven geometrical parameters and the obtained Euler number and
Stokes number for the three points, A, B and C for the 16 test cases (cf. Table 8.28)
Point Dx a b S Ht h Bc Euler number Stokes number x 103
A1 0.692 0.228 0.213 0.408 6.819 1.855 0.498 0.815 3.805
B1 0.306 0.688 0.398 0.404 6.993 1.779 0.317 27.322 0.540
C1 0.360 0.295 0.253 0.443 6.664 1.910 0.459 4.892 0.937
A2 0.700 0.204 0.227 0.446 6.422 1.811 0.470 0.820 4.260
B2 0.301 0.697 0.399 0.561 6.989 1.840 0.449 28.222 0.538
C2 0.415 0.414 0.287 0.449 6.803 1.776 0.443 6.052 1.037A3 0.676 0.209 0.204 0.431 4.378 1.980 0.496 0.868 4.108
B3 0.304 0.624 0.398 0.581 6.983 1.116 0.344 32.296 0.561
C3 0.363 0.278 0.230 0.459 6.404 1.354 0.445 4.791 1.000
A4 0.671 0.205 0.212 0.427 6.524 1.824 0.467 0.819 3.796
B2 0.303 0.622 0.399 0.467 6.767 1.975 0.479 22.621 0.556C4 0.347 0.237 0.271 0.582 6.769 1.913 0.476 4.873 0.931
A5 0.669 0.224 0.204 0.537 6.946 1.895 0.479 0.900 3.654
B5 0.301 0.672 0.387 0.416 6.851 1.905 0.463 23.579 0.538
C5 0.343 0.228 0.282 0.475 6.925 1.822 0.487 4.647 0.893
A6 0.697 0.207 0.207 0.400 6.734 1.974 0.465 0.706 4.142B6 0.301 0.671 0.400 0.500 6.982 1.228 0.477 29.367 0.538
C6 0.380 0.241 0.303 0.486 6.715 1.845 0.427 4.557 1.026
A7 0.660 0.214 0.205 0.432 6.962 1.934 0.495 0.810 3.391
B7 0.301 0.690 0.322 0.459 6.988 1.918 0.450 20.739 0.557
C7 0.401 0.350 0.258 0.451 6.928 1.920 0.454 4.735 1.039
A8 0.660 0.214 0.205 0.432 6.962 1.934 0.495 0.810 3.391B8 0.301 0.690 0.322 0.459 6.988 1.918 0.450 20.739 0.557
C8 0.333 0.278 0.263 0.457 6.869 1.940 0.450 5.582 0.832
A9 0.686 0.236 0.236 0.478 6.910 1.915 0.473 1.008 3.743
B9 0.308 0.692 0.390 0.423 6.902 1.995 0.472 22.898 0.551
C9 0.330 0.290 0.252 0.465 6.890 1.955 0.471 5.639 0.818A10 0.699 0.211 0.205 0.409 6.499 1.937 0.444 0.741 4.328
B10 0.302 0.636 0.399 0.424 6.854 1.614 0.488 23.886 0.543
C10 0.375 0.298 0.250 0.465 6.726 1.875 0.445 4.619 0.997
A11 0.687 0.212 0.232 0.568 6.702 1.900 0.465 0.955 4.035
B11 0.300 0.670 0.399 0.481 6.988 1.693 0.484 25.971 0.534C11 0.367 0.259 0.282 0.461 6.688 1.857 0.424 4.801 0.972
A12 0.694 0.202 0.202 0.512 6.999 1.794 0.499 0.736 4.071
B12 0.301 0.686 0.400 0.431 6.840 1.191 0.439 29.948 0.533
C12 0.392 0.368 0.281 0.455 6.874 1.731 0.472 5.822 0.980
A13 0.700 0.235 0.201 0.602 6.752 1.810 0.481 0.901 4.286
B13 0.300 0.574 0.365 0.408 6.743 1.907 0.454 19.174 0.565C13 0.348 0.308 0.239 0.426 6.577 1.925 0.477 5.047 0.895
A14 0.692 0.202 0.203 0.464 6.963 1.901 0.463 0.722 4.161
B14 0.305 0.695 0.398 0.434 6.881 1.785 0.498 24.604 0.543
C14 0.398 0.389 0.272 0.440 6.899 1.849 0.467 5.614 0.992
A15 0.676 0.237 0.205 0.437 6.779 1.879 0.469 0.890 3.717B15 0.301 0.688 0.367 0.562 6.975 1.704 0.480 25.765 0.549
C15 0.332 0.224 0.278 0.450 6.645 1.799 0.456 4.912 0.869
A16 0.684 0.217 0.202 0.424 6.725 1.962 0.380 0.823 4.299
B16 0.314 0.687 0.388 0.400 6.943 1.792 0.495 21.743 0.565
C16 0.348 0.292 0.245 0.457 6.850 1.887 0.477 4.969 0.891
A indicates the point of minimum Euler number and maximum Stokes number. B indicates the point ofmaximum Euler number and minimum Stokes number. C indicates an optimal point for themulti-objective optimization problem, cf. Fig. 8.28(c).
214
8.5. Multi-objective optimization using CFD data
8.5 Single and multi-objective optimization us-
ing CFD data
8.5.1 Design variables and approaches
The sensitivity analysis presented in Sec. 4.1 reported that the most sig-
nificant factors affecting the cyclone performance are, the vortex finder di-
ameter Dx, the inlet width b, the inlet height a and the total cyclone height
Ht. The effects of the barrel height h, the vortex finder length S and the
cone-tip diameter Bc are insignificant. Therefore, only four significant ge-
ometrical factors have been used in this study. The three other factors are
fixed based on the Stairmand high efficiency design, i.e. Bc/D = 0.375,
h/D = 1.5 and S/D = 0.5. The selection of the values for both h and S are
also based on the conclusion of Zhu and Lee [200] (when both the pressure
drop and the particle collection efficiency are considered, a cyclone which
has (h-S)/D of 1.0 would be an optimum design).
Optimization of a gas cyclone is, indeed, a multi-objective optimization
problem rather than a single objective optimization problem that has been
considered so far in the literature [146]. Both the pressure drop and the
cut-off diameter in gas cyclones are important objective functions to be op-
timized simultaneously in a multi-objective optimization problem. These
objective functions values are either obtained from experiments, empirical
models or computed using computational fluid dynamic (CFD) approaches.
CFD modeling and optimization of the parameters are investigated in the
present study and multi-objective Nedler-Mead optimization algorithms
are used in order to maximize the collection efficiency (minimize the cut-off
diameter) and minimize the Euler number (dimensionless pressure drop).
The desirability function approach has been used to handle the two objec-
tive function.
The application of the Nelder-Mead technique suffers from one big dis-
advantage. The optimal point may be local minimum because the tech-
nique depends on the starting point. This was the motivation to use also
the genetic algorithm technique. Furthermore, the radial basis function
neural networks can attain high accuracy as curve fitting approach than
the response surface [53]. Consequently, the application of the radial ba-
sis function neural networks to model the relationship between the per-
formance parameters and the four geometrical parameters will result in
more accurate results. The optimization studies given in the previous sec-
tions were applied for the seven geometrical parameters but the current
study is focusing only on the most significant four geometrical parameters.
The above-mentioned studies used meta-models using mathematical mod-
els and experimental measurements but this study is based only on CFD
215
Chapter 8. Optimization
simulations performed on sampling cyclone with a diameter of 31E-3 m.
Study objectives
There are four objectives of this study. (1) Investigation of the effect of the
four geometrical parameters on the cyclone separator performance based
on CFD simulations results. (2) Study the possible interaction between
the four geometrical parameters using the response surface methodology.
(3) Application of the multi-objective optimization technique to obtain new
geometrical ratios for minimum pressure drop and minimum cut-off diam-
eter, followed by a comparison of the numerical simulations for the new
design and the Stairmand design using the Reynolds stress turbulence
model. (4) Application of the genetic algorithm technique instead of the
desirability function approach using the two trained radial basis function
neural networks as fitness functions.
Study outline
This study is performed in six stages. The first stage is the application
of the response surface methodology (Box-Behnken design) to design an
experiment to study the effect of four geometrical parameters (the inlet
height and width, the vortex finder diameter, and the cyclone total height)
on the cyclone performance using Statgraphics statistical software. Sec-
ondly, the obtained 27 test cases (designs) have been computationally sim-
ulated using the Reynolds stress turbulence model and discrete phase
modeling with the Fluent solver. In the third stage, the Euler numbers
and the cut-off diameters obtained are used to fit a second order poly-
nomial (response surface) for each response (the Euler number and the
cut-off diameter). The obtained polynomials have been used to study the
variation of the two responses with the four geometrical parameters. Fur-
thermore, the obtained polynomials have been used to obtain new cyclone
geometrical ratios using the multi-objective optimization between the two
conflicting objectives (the Euler number and the cut-off diameter) using
the desirability function approach. The fourth stage is a computational
investigation: a numerical comparison between the new design and the
Stairmand design has been performed. Fifthly, replace the response sur-
face methodology with the artificial neural networks approach and study
the effect of each geometrical parameter on the cyclone separator perfor-
mance. The last stage is an optimization study using the genetic algorithm
instead of desirability function.
216
8.5. Multi-objective optimization using CFD data
Design of experiment
In order to model a complex multivariate process where the responses are
influenced by several variables, the response surface statistical technique
seems the best approach [50]. The steps are as follows: (1) construct the
design of experiment by identifying the four tested geometrical parameters
(minimum and maximum values) and also to decide upon the dependent
variables (the Euler number and the cut-off diameter). Statgraphics com-
mercial statistical software gives 27 runs to be performed (cf. Table 8.32)
using the Box-Behnken design of experiment method [118]. (2) Once the
runs have been conducted (using CFD simulations) and recorded data in-
serted in the table; Statgraphics software fits a second order polynomial to
this data (one response surface per dependent variable) [30]. The second-
order polynomial (response surface) has the form [50]:
Yk = β0 +4∑
i=1
βiXi +4∑
i=1
βiiX2i +
∑∑
i<j
βijXiXj (8.14)
where β0, βi, βii, and βij are the regression coefficients for intercept, linear,
quadratic and interaction terms, respectively. Xi and Xj are the indepen-
dent variables, and Yk is k the response variable (k = 1 for the Euler num-
ber and k = 2 for the cut-off diameter). (3) The third step, is the analysis
of the response surface plot, main effect plots, Pareto chart and interaction
plots.
Analysis of response surface plots
Analysis of variance (ANOVA) showed that the resultant quadratic poly-
nomial models adequately represent the used data with the coefficient of
determination R2, being 0.94284 and 0.973468 for the Euler number and
cut-off diameter respectively. This indicates that the quadratic polynomial
models obtained were adequate to describe the influence of the indepen-
dent variables studied [189]. Analysis of variance (ANOVA) was used to
evaluate the significance of the coefficients of the quadratic polynomial
models (cf. Tables 8.33 and 8.34). For any of the terms in the models, a
large F-value (small P-value) would indicate a more significant effect on
the respective response variables.
Based on the ANOVA results presented in Table 8.33, the variable with
the largest effect on the pressure drop (Euler number) was the linear term
of vortex finder diameter (β3), the linear terms of inlet height and width
(P < 0.05) (β1 and β2 respectively), whereas the linear term of cyclone
total height did not show a significant effect (P > 0.05). The quadratic
217
Chapter 8. Optimization
Table 8.32: The Box-Behnken’s design matrix and the responses of the Euler num-
ber and cut-off diametera
Run X1 X2 X3 X4 Euler number Cut-off diameterb
No. a/D b/D Dx/D Ht/D (-) (µm)
1 0.250 0.2625 0.50 3 3.500 1.546
2 0.500 0.2625 0.75 4 2.827 2.541
3 0.375 0.3750 0.25 4 27.257 1.1584 0.375 0.2625 0.50 4 3.475 1.683
5 0.375 0.2625 0.75 3 2.333 2.444
6 0.250 0.2625 0.50 5 2.952 1.364
7 0.375 0.1500 0.50 4 2.726 1.353
8 0.375 0.1500 0.50 5 2.530 1.2849 0.375 0.2625 0.25 5 17.712 0.956
10 0.500 0.3750 0.50 4 9.086 2.163
11 0.250 0.2625 0.75 4 1.413 1.939
12 0.375 0.1500 0.50 3 3.000 1.455
13 0.375 0.1500 0.75 4 1.211 1.826
14 0.375 0.3750 0.50 3 7.500 2.08115 0.500 0.2625 0.50 5 5.904 1.787
16 0.375 0.3750 0.50 5 6.326 1.836
17 0.250 0.2625 0.25 4 12.720 0.860
18 0.375 0.2625 0.25 3 21.000 1.084
19 0.500 0.2625 0.25 4 25.440 1.12720 0.250 0.3750 0.50 4 4.543 1.651
21 0.375 0.2625 0.50 4 4.770 1.683
22 0.375 0.3750 0.75 4 3.029 2.610
23 0.500 0.1500 0.50 4 3.634 1.513
24 0.250 0.1500 0.50 4 1.817 1.15525 0.500 0.2625 0.50 3 7.000 2.025
26 0.375 0.2625 0.50 4 4.770 1.683
27 0.375 0.2625 0.75 5 1.968 2.156
a The values of h/D = 1.5, S/D = 0.5 and Bc/D = 0.375 are identical to that of Stairmand highefficiency design, so the variation in the total cyclone height is due to the variations of the cone height.
b The value of cut-off diameter depends on the gas velocity and density, particle density. In this study,D = 31mm, gas volume flow rate Qin = 50 l/min, gas density ρ = 1.0, gas viscosity µ = 2.11E − 5 Pa sand particle density ρp = 860 kg/m3.
term of vortex finder diameter also had a significant effect on the pressure
drop; however, the effect of the other three quadratic terms was insignifi-
cant. Furthermore, the interaction between the inlet dimensions and vor-
tex finder diameter also had a significant effect on the pressure drop (β13
and β23), whereas the effect of the remaining terms was insignificant. Ta-
ble 8.34 confirms the significant effect of all linear terms of the inlet di-
mensions, vortex finder diameter and total cyclone height on the cut-off
diameter. Moreover, the interaction between the inlet width and the vor-
tex finder diameter also had a significant effect on the cut-off diameter,
whereas the effect of the remaining terms was insignificant.
For the visualization of the results of the analysis, main effects plot, Pareto
chart and response surface plots were drawn. The slope of the main effect
curve is proportional to the size of the effect, and the direction of the curve
specifies a positive or negative influence of the effect [50, 61] (Fig.8.31(a)).
Based on the main effect plot and the Pareto chart shown in Fig. 8.31,
the most significant factors on the Euler number are: (1) the vortex finder
diameter Dx, with a non-linear relation with a wide range where it is in-
versely proportional to Euler number and a narrow range with a direct
218
8.5. Multi-objective optimization using CFD data
Table 8.33: Analysis of variance and regression coefficients for the Euler number∗
Variable Regression coefficient F-Ratio P-Value
β0 1.88508Linearβ1 41.3522 14.21 0.0012β2 134.36 33.78 0.0000β3 -49.1776 209.03 0.0000β4 -4.14082 0.94 0.3430Quadraticβ11 16.9991 0.1 0.7509β22 -44.8511 0.5 0.4878β33 87.9024 44.97 0.0000β44 0.368311 0.2 0.6602Interactionβ12 48.4622 0.41 0.5306β13 -90.448 7.01 0.0155β14 -1.096 0.02 0.8992β23 -185.69 28.5 0.0000β24 -1.56444 0.03 0.8708β34 2.923 0.47 0.5017
R2 0.94284
∗ Bold numbers indicate significant factors as identified by the analysis of variance (ANOVA) at the 95%confidence level.
proportional relation. (2) the inlet width b and the inlet height a almost
linearly related to the Euler number, (3) the cyclone total height Ht has
an insignificant effect. Pareto charts were used to summarize graphically
and display the relative importance of each parameter with respect to the
Euler number [50]. The Pareto chart shows all the linear and second-
order effects of the parameters within the model and estimates the signif-
icance of each with respect to maximizing the Euler number response. A
Pareto chart displays a frequency histogram with the length of each bar
proportional to each estimated standardized effect [30]. The vertical line
on the Pareto charts judges, whether each effect is statistically significant
within the generated response surface model; bars that extend beyond this
line represent effects that are statistically significant at a 95% confidence
level. Based on the Pareto chart (Fig. 8.31(b)) there are six significant pa-
rameters at a 95% confidence level: the vortex finder diameter Dx; the
inlet width b; the inlet height a and the combinations a Dx, b Dx and D2x.
Therefore, the Pareto chart is a perfect supplement to the main effects
plot. To visualize the effect of the independent variables on the dependent
ones,response surfaces of the quadratic polynomial models were generated
by varying two of the independent variables within the experimental range
whereas holding the other factors at their central values [189] as shown in
Fig. 8.31. Thus, Fig. 8.31(e) was generated by varying the inlet height aand the inlet width b whereas keeping the other factors fixed at their cen-
tral values. The trend of the curve is linear, with a more significant effect
for the inlet width b, and a weak interaction between the inlet height a
219
Chapter 8. Optimization
Table 8.34: Analysis of variance and regression coefficients for the cut-off
diameter∗
Variable Regression coefficient F-Ratio P-Value
β0 0.365539Linearβ1 1.84206 67.21 0.0000β2 -1.23737 89.76 0.0000β3 0.494019 524.79 0.0000β4 -0.00682157 15.14 0.0009Quadraticβ11 -2.25164 0.82 0.3763β22 -0.708116 0.06 0.8152β33 -0.307782 0.25 0.6237β44 0.00841817 0.05 0.8308Interactionβ12 2.73778 0.59 0.4531β13 2.68 2.77 0.1116β14 -0.112 0.08 0.7837β23 7.07071 18.61 0.0003β24 -0.164444 0.14 0.7170β34 -0.16 0.63 0.4360
R2 0.973468
∗ Bold numbers indicate significant factors as identified by the analysis of variance (ANOVA) at the 95%confidence level.
and width b. The response surface plots of Figs. 8.31(c), 8.31(d) and 8.31(f)
show that there are strong interactions between the vortex finder diameter
Dx and the inlet dimensions a and b.
From the analysis of the design of experiment for the cut-off diameter, the
effect of variation of the vortex finder diameter Dx on the cut-off diameter
is opposite to that on the Euler number (cf. Fig. 8.32(a)). The Pareto chart
given in Fig. 8.32(b) indicates the significance effect of the cyclone total
height Ht on the cut-off diameter, but its effect is minor in comparison with
the three other factors. The significant interaction exists only between the
vortex finder diameter and the inlet dimensions.
8.5.2 Multi-objectives optimization using the desirabil-ity function
The desirability function approach
From the previous analysis, it is observed that the optimal values for the
geometrical parameters that minimize the pressure drop are different from
the values that minimize the cut-off diameter (cf. Figs. 8.31 and 8.32).
As a result, a multi-objective optimization procedure is needed. The uti-
220
8.5. Multi-objective optimization using CFD data
(a) Main effects plot
(b) Pareto chart
Figure 8.31: Analysis of design of experiment for the Euler number
lization of desirability function proposed by Harrington [71] is the most
popular and strongly suggested method for multiple response optimiza-
tion problems [20] to convert the problem into single-objective. The Stat-
graphics statistical package uses this desirability function approach for
optimization of multiple response problems. The desirability function is
first defined for each response. The desirability function d(y) expresses the
desirability of a response value equal to y on a scale of 0 (if the response
value is in an unacceptable range) to 1 (for the optimum value), for mini-
mization of response variable. The procedure will then find the settings of
the experimental factors that maximize a combined desirability function,
which is a function that expresses the desirability of a solution involving
m, where m here equals 2 (one for the Euler number and the other for the
221
Chapter 8. Optimization
(c) Dx versus a
(d) Dx versus b
(e) a versus b
(f) Dx versus Ht
Figure 8.31: (continued) Analysis of design of experiment for the Euler number.
Note: In Figs 8.31(c) - 8.31(f), for each plot of two independent variable, all other
variables are hold at their central values.
222
8.5. Multi-objective optimization using CFD data
(a) Main effects plot
(b) Pareto chart
Figure 8.32: Analysis of design of experiment for the cut-off diameter
cut-off diameter), responses through the function of the form [131],
D =dI11 dI22 . . . dImm
1/(∑m
j=1 Ij)(8.15)
where dj is the calculated desirability of the jth response and Ij is an im-
pact coefficient that ranges between 1 and 5 [131]. It represents the im-
portance relative of each response over the other responses [88]. Where
the default value is 3. In this study, more importance is given to the Eu-
ler number (I1 = 5 and I2 = 3). When a response is to be minimized, the
223
Chapter 8. Optimization
(c) Dx versus a
(d) Dx versus b
(e) a versus b
(f) Dx versus Ht
Figure 8.32: (continued) Analysis of design of experiment for the cut-off diameter.
Note: In Figs 8.32(c) - 8.32(f), for each plot of two independent variable, all other
variables are hold at their central values.
224
8.5. Multi-objective optimization using CFD data
Table 8.35: The optimized cyclone separator design for best performance using the
desirability function
Factor Minimum Center Maximum Stairmand design Optimum
a 0.25 0.375 0.5 0.5 0.256b 0.15 0.2625 0.375 0.2 0.151Dx 0.25 0.5 0.75 0.5 0.415Ht 3 4 5 4.0 4.56
desirability of a predicted response equal to yj is defined as,
dj =
1 yj < lowj(yj−highj
lowj−highj
)lowj ≤ yj ≤ highj
0 yj > highj
(8.16)
where lowj and highj are the minimum and maximum values of jth re-
sponse. In this study, the minimum and maximum values obtained in the
data sheet have been used (cf. Table 8.32). The desirability plots are given
in Fig. 8.33. For more information about statistical model used for mul-
tiple response optimization refer to Ref. [131]. Table 8.35 presents the
optimum values of geometry parameters that minimize the values of the
Euler number and the cut-off diameter, which gives optimum desirability
D = 0.83 to minimize the Euler number and cut-off diameter. It is clear
from Table 8.35 that the new optimized design differs from the Stairmand
design in many geometrical parameters, whereas the new ratios will re-
sult in minimum pressure drop and minimum cut-off diameter. The ratio
of inlet width to height b/a = 0.589 lays in the optimal cyclone lies in the
recommended range of b/a from 0.5 to 0.7 proposed by Elsayed and Lacor
[52]. To understand the effect of these changes in the geometrical ratios on
the flow field pattern and performance, a CFD study for the two designs is
needed.
8.5.2.1 CFD comparison between the Stairmand and optimal de-
sign
The Fluent solver has many turbulence models available for simulating
turbulent flow. It is generally recognized that only the Reynolds stress
model (RSM) and large eddy simulation (LES) can capture the main fea-
tures of the highly complicated swirling flow in cyclone separators [11, 23,
50, 52, 62–64, 74, 91, 159, 186, 198]. The Reynolds stress turbulence model
has been used in this study to reveal the turbulent flow in the two cyclone
separators. For the detailed governing equation for both the Reynolds av-
eraged Navier-Stokes equation (RANS) and the discrete phase modeling
225
Chapter 8. Optimization
(a) Dx versus a
(b) Dx versus b
(c) a versus b
(d) Dx versus Ht
Figure 8.33: The desirability plots. Note: for each plot of two independent variable,
all other variables are hold at their optimal values.
226
8.5. Multi-objective optimization using CFD data
(DPM) the reader can refer to Elsayed and Lacor [52]. The geometrical
values are given in Table 8.36 for the two cyclones (cf. Fig. 8.34).
Numerical settings
The air volume flow rate Qin=50 l/min for the two cyclones, air density is
1.0 kg/m3 and dynamic viscosity 2.11E-5 Pa s. The turbulent intensity
equals 5% and characteristic length equals 0.07 times the inlet width [75].
A velocity inlet boundary condition is applied at inlet, outflow at gas outlet
and wall boundary conditions at all other boundaries.
The finite volume method has been used to discretize the partial differen-
tial equations of the model using the SIMPLEC (Semi-Implicit Method for
Pressure-Linked Equations-Consistent) method for pressure velocity cou-
pling and QUICK scheme to interpolate the variables on the surface of the
control volume. The implicit coupled solution algorithm was selected. The
unsteady Reynolds stress turbulence model (RSM) was used in this study
with a time step of 0.0001s.
(a) The cyclone geometry (b) The surface mesh for the new design
Figure 8.34: The cyclone geometry and the surface mesh for the new design.
Table 8.36: The values of geometrical parameters for the two designs (D=31E-3 m)
Cyclone a/D b/D Dx/D Ht/D h/D S/D Bc/D Li/D Le/DStairmand design 0.5 0.2 0.5 4 1.5 0.5 0.375 1.0 0.5
New design 0.256 0.151 0.415 4.56 1.5 0.5 0.375 1.0 0.5
227
Chapter 8. Optimization
Grid independency study
The grid independence study has been performed for the tested cyclones.
Three levels of grid for each cyclone have been tested, to be sure that
the obtained results are grid independent. The hexahedral computational
grids were generated using Gambit grid generator and the simulations
were performed using Fluent 6.3.26 commercial finite volume solver on a
8 nodes CPU Opteron 64 Linux cluster. To evaluate accuratelyto estimate
accurately the numerical uncertainties in the computational results, the
concept of grid convergence index (GCI) was adopted using three grid lev-
els per cyclone.
Grid convergence index (GCI)
Table 8.37 presents the grid convergency calculations using GCI method
for three grid levels for each cyclone. The following conclusions have been
obtained from the GCI analysis [55]:
• The results are in the asymptotic range, because the obtained values
for α are close to unity.
• The ratio R is less than unity this means monotonic convergence [2].
• There is a reduction in the GCI value for the successive grid refine-
ments
(GCIfine12 < GCI
fine23 ) for the two variables (Eu and x50). This indicates
that the dependency of the numerical results on the cell size has been
reduced. Moreover, a grid independent solution has been achieved.
Further refinement of the grid will not give much change in the sim-
ulation results. For the two variables (Eu and x50), the extrapolated
value is only slightly lower than the finest grid solution. Therefore,
the solution has converged with the refinement from the coarser grid
to the finer grid [2]. Figure 8.35 presents a qualitative proof that the
obtained results are in the asymptotic range.
• The value of εi,i+1 represent the relative change in each value from
coarse to medium and from medium to fine mesh. For example, ε1,2 =0.0256 for the Euler number in the new design means the percentage
change in the Euler number when the mesh becomes 986748 cells
instead of 717353 cells equals 2.56%. This means no need to use
the fine mesh and the usage of the medium mesh of 717353 cells is
sufficient. Another example, ε1,2 = 0.0052 for the cut-off diameter
in the Stairmand design means the percentage change in the cut-
off diameter when the mesh becomes 848783 cells instead of 622253
228
8.5. Multi-objective optimization using CFD data
cells equals 0.52%. This means no need to use the fine mesh and the
usage of the medium mesh of 622253 cells is sufficient.
In summary, the grid refinement study shows that a total number of about
622253 hexahedral cells for Stairmand cyclone and 717353 hexahedral
cells for the new design are sufficient to obtain a grid-independent solu-
tion, and further mesh refinement yields only small, insignificant changes
in the numerical solution.
Table 8.37: Grid convergency calculations using GCI method using three grid levels
for each cyclone.
i Ni fi ri,i+1 ei,i+1 εi,i+1 GCIi,i+1% Ra αb
New
desi
gn
Eu
ler
nu
mber 0c 2.5195
1 986748 2.5380
1.1121 0.0650 0.0256 0.9112
2 717353 2.6030 0.0331 1.0256
1.2553 2.0140 0.7737 4.00993 362679 4.6170
Cu
t-off
dia
mete
r 0 0.6621
1 986748 0.6710
1.1121 0.0190 0.0283 1.6647
2 717353 0.6900 0.0669 1.02831.2553 0.2920 0.4232 5.0609
3 362679 0.9820
Sta
irm
an
dd
esi
gn
Eu
ler
nu
mber 0 5.4310
1 848783 5.4860
1.1090 0.1200 0.0219 1.2521
2 622253 5.6060 0.0999 1.0219
1.2044 1.2270 0.2189 3.90103 356181 6.8330
Cu
t-off
dia
mete
r 0 1.7187
1 848783 1.7210
1.1090 0.0090 0.0052 0.1678
2 622253 1.7300 0.0489 1.00521.2044 0.1850 0.1069 0.8173
3 356181 1.9150
a R=ε12/ε23 .
b α=(
rp12
GCI12
)
/GCI23 .
c The value at zero grid space (h → 0). i=1, 2 and 3 denote the calculations at the fine, medium and coarse mesh respectively.
Results and discussion
Flow field pattern
Figure 8.36 shows the contour plot at Y=0 and throughout the inlet sec-
tion. In the two cyclones, the time-averaged dimensionless static pressure
(Euler number) decreases radially from the wall to center. A negative pres-
sure zone appears in the forced vortex region (central region) due to high
swirling velocity. The pressure gradient is largest along the radial direc-
tion, whereas the gradient in the axial direction is very limited. The cy-
clonic flow is not symmetrical as is clear from the shape of the low-pressure
zone at the cyclone center (twisted cylinder). The flow asymmetry is more
pronounced in the new cyclone. However, the two cyclones have almost the
229
Chapter 8. Optimization
N -1
Eul
ernu
mbe
r
Cut
-off
diam
eter
0 1E-06 2E-062
3
4
5
6
7
1
1.5
New design (Eu)
New design (X50)
Stairmand design (Eu)
Stairmand design (X50)
(h --> 0)(h --> 0)(h --> 0)(h --> 0)(h --> 0)(h --> 0)
Figure 8.35: Qualitative representation of the grid independency study. The Euler
number and the cut-off diameter for the two cyclones using the three grid levels.
N−1 is the reciprocal of the number of cells, h → 0 means the value at zero grid
size (cf. Table 8.37). To obtain a smooth curve; the spline curve fitting has been
applied in Tecplot post-processing software.
same flow pattern, but the Euler number of the Stairmand design is nearly
twice that of the new design.
The dimensionless static pressure distribution presented in Fig. 8.37 for
the two cyclones indicates that the highest dimensionless static pressure
for the Stairmand design is more than twice that of the new design at all
sections whereas the central value is almost the same for the two cyclones.
This indicates that, the new design has a lower dimensionless pressure
drop than the Stairmand design. However, these results are obtained at
different inlet velocity for the two cyclones (to have the same air flow rate).
The same Euler number values would be obtained if the two cyclones work
at the same inlet velocity because the Euler number is not a function of
flow velocity if the Reynolds number is higher than 2E4 [50].
The tangential velocity profile is composed of two regions. In the inner re-
gion, the flow rotates approximately like a solid body (forced vortex), where
the tangential velocity increases with radius. After reaching its peak the
velocity decreases with radius in the outer part of the profile (free vortex).
The tangential velocity distributions for the two cyclones are nearly iden-
tical in pattern (Rankine profile). The inner part of the tangential velocity
distribution of the two cyclones is very similar. The outer part for the new
design is flatter in comparison with the Stairmand cyclone. This implies
that there is more space in the optimal cyclone where the particles are
subjected to high centrifugal force. Whereas, the maximum dimension-
less tangential velocity for Stairmand cyclone is higher than that for the
230
8.5. Multi-objective optimization using CFD data
new design, the cyclone performance is not only affected by the maximum
tangential velocity but also with the separation space (the new design is
longer than the Stairmand design)
The axial velocity profiles for the two cyclones are different in values and
shape. Stairmand cyclone exhibit the inverted W axial velocity profile. The
new design exhibit the inverted W axial velocity profile away from the inlet
section and the inverted V elsewhere. This is due to the very high swirl
exist at the inlet section in case of the new design.
Performance parameters
To calculate the cut-off diameters of the two cyclones, 104 particles were
injected from the inlet surface with a particles density ρp = 860kg/m3 and
the maximum number of time steps for each injection was 2E9 steps. The
DPM analysis results and the Euler number for the two cyclones are given
in Table 8.38. Table 8.38 introduces also a comparison between the CFD
results and four different mathematical models viz., the Ramachandran
model [139], the Muschelknautz method of modeling (MM) [50, 116, 174],
the Iozia model [85], the Ritema model [142](cf. Hoffmann and Stein[77]
for more details about these mathematical models). However, no-good
matching between the two approaches (CFD and mathematical models)
is obtained, they agree in the trend of superior performance of the new
design. The Euler number and cut-off diameter for the new design is ap-
proximately half that of Stairmand design.
8.5.3 Artificial neural network (ANN) approach
Artificial neural networks (ANNs) have become an attractive approach for
modeling highly complicated and nonlinear system [53, 180, 197]. In this
study, the radial basis function neural network (RBFNN) has been used to
model the effect of cyclone dimensions on both the pressure drop and the
cut-off diameter. For more details about the radial basis function neural
networks, the reader can refer Sec. 8.3.1, page 158. Two RBFNNs have
been trained using the 27 data set obtained from the CFD simulations for
both the Euler number (the dimensionless pressure drop) and the cut-off
diameter.
In this study, the performance parameters are assumed as functions of
only the four geometrical parameters, whereas all other parameters kept
constant, Eq. 8.17.
231
Chapter 8. Optimization
New design Stairmand design
Figure 8.36: The contour plots for the time averaged flow variables at sections Y=0
and throughout the inlet section. From top to bottom: the dimensionless static
pressure (divided by the dynamic pressure at inlet), the dimensionless tangential
velocity, axial velocity. Note: both cyclones have the same air volume flow rate.
232
8.5. Multi-objective optimization using CFD data
Table 8.38: The cyclone performance parameters using CFD simulations and four
mathematical models
Parameter Cyclone Statgraphics CFD Ramachandran MM Iozia Ritema
Euler numberStairmand design 5.673 5.606 4.85 5.33 - -New design 2.221 2.603 2.71 2.99 - -
Cut-off diameter, µmStairmand design 1.706 1.73 - - 1.69 1.55New design 0.865 0.69 - - 0.95 0.71
Radial position / Cyclone radius
Sta
ticpr
essu
re/D
ynam
icpr
essu
reat
inle
t
-1 -0.5 0 0.5 1-0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5New designStairmand design
Radial position / Cyclone radius
Tan
gent
ialv
eloc
ity/I
nlet
velo
city
-1 -0.5 0 0.5 10
0.2
0.4
0.6
0.8
1
1.2 New designStairmand design
Figure 8.37: The radial profiles of the time averaged static pressure and tangential
velocity at z’/D=1.5.
Eu = f1
(a
D,b
D,Dx
D,Ht
D
)
x50 = f2
(a
D,b
D,Dx
D,Ht
D
)(8.17)
Due to the large difference in the order of magnitude of the value (cf. Ta-
ble 8.35), the available dataset is transformed into -1 to 1 interval using
the Matlab intrinsic function; mapminmax in order to avoid solution diver-
gence [53, 197]. The RBFNN calculations have been performed using the
neural network toolbox available from Matlab commercial software 2010a.
The cut-off diameter x50 for cyclone separator is always given in units of
µm. Another way to represent x50 is using a dimensionless number; Stokes
number Stk50 = ρpx250Vin/(18µD) [37]. It is the ratio between the particle
relaxation time; ρpx250/(18µ) and the gas flow integral time scale; D/Vin.
Fitting the RBFNNs
The configured RBFNNs predictions versus the CFD data for the Euler
number and cut-off diameter are shown in Fig. 8.38. It can be seen that
233
Chapter 8. Optimization
the RBFNN models are able to attain the high training accuracy. The
training mean square errors are zeros (i.e., identical matching between
the input and output, the reason behind that may refer to the consistency
between the used data in the simulation using the design of experiment,
which is not the case for the study of Elsayed and Lacor [53] using ex-
perimental data set), Fig. 8.38. This indicates that, in comparison with
traditional models of curve fitting, the models based on an artificial intelli-
gence algorithm have a superior capability of nonlinear fitting. Especially,
the RBFNN has its unique and optimal approximation characteristics in
learning process [53, 197].
Figure 8.38 illustrates the agreement between the RBFNNs input and out-
put. The obtained relation is a typical linear relation with a coefficient of
correlation close to 1 (R > 0.999). The agreement between the input and
output of the RBFNN is also clear from the value of the mean squared er-
ror E2. Consequently, the trained neural networks predict very well both
the Euler number and cut-off diameter values and can be used in cyclone
design and performance estimation. Figure 8.38 present different perfor-
mance indicators as a validation of the proposed models for the trained
data.
Input value (x)
Pre
dict
edva
lue
(y)
5 10 15 20 25
2
4
6
8
10
12
14
16
18
20
22
24
26 Data pointLinear fit
y=0.999 x -2.2E-15
R=0.99999
E2= 0.0
(a) The Euler number
Input value (x)
Pre
dict
edva
lue
(y)
1 1.5 2 2.5
1
1.5
2
2.5 Data pointLinear fit
y=0.999 x +4.03E-16
R=0.99999
E2= 0.0
(b) The cut-off diameter
Figure 8.38: Linear regression of the RBFNNs for the Euler number and the cut-off
diameter.
234
8.5. Multi-objective optimization using CFD data
Dx, a, b
HtE
uler
num
ber
0.2 0.3 0.4 0.5 0.6 0.7
3 3.2 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5
5
10
15
Dx
abHt
(a) Euler number
Dx, a, b
Ht
X5
0[m
icro
n]
0.2 0.4 0.6
3 3.2 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5
0.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
Dx
abHt
(b) Cut-off diameter
Figure 8.39: The effect of geometrical parameters on the cyclone performance.
The effect of the four geometrical parameters on the cyclone per-
formance based on RBFNNs
The effects of the geometrical parameters on both the Euler number and
the cut-off diameter are depicted in Figs. 8.39(a) and 8.39(b). To study the
effect of each parameter, the tested RBFNNs models have been used by
varying one parameter at a time from its minimum to maximum values
of the available CFD dataset, whereas the other parameters are kept con-
stant at their mean values (cf. Table 8.35). Figures 8.39(a) and 8.39(b)
indicate the significant effect of the vortex finder diameter Dx, the inlet
width b, the inlet height a. Less effect is assigned to the total cyclone
height Ht. More analysis is given in Tables 8.39 and 8.40.
8.5.4 Optimization Using Genetic Algorithms
Optimal cyclone design for minimum pressure drop
The genetic algorithm optimization technique has been applied to obtain
the geometrical ratios for minimum pressure drop (Euler number). The ob-
jective function is the Euler number (using the trained radial basis func-
tion neural network). The design variables are four geometrical dimen-
sions, the inlet height a, the inlet width b, the vortex finder diameter Dx
and the total cyclone height Ht. These four variables are the most signifi-
cant factors which affect the cyclone performance [50].
235
Chapter 8. Optimization
Table 8.39: The variation of the Euler number with cyclone dimensions using the
RBFNN model (cf., Fig. 8.39(a))
Factor Analysis
Dx
The most significant effect is that of the vortex finder diameter Dx
with inverse relationship when increasing Dx up to ≈ 0.65 after
which the relation becomes direct, Fig. 8.39(a). This can be explained
as follows: Although the pressure loss in the vortex finder decreases
with increasing the vortex finder diameter like the case of viscous
flow in a pipe, the Euler number in the cyclone body instead will
increase due to the decrease of the flow area just after the flow
entrance from the inlet region (the annular space between the barrel
and the vortex finder). This analysis indicates the large contribution
of the pressure loss in the vortex finder to the total Euler number (the
pressure loss at the entrance, the pressure loss in the cyclone body,
and the pressure loss in the vortex finder [50]).
a and b
The inlet height a and width b is almost linearly related to the Euler
number. But, why the relation is direct (at the same flow rate
increasing the inlet dimensions decreases the inlet velocity.
Consequently,the loss in the vortex finder will decrease)? The reason
is the increase in the pressure drop at the inlet section, due to
deviation of the inlet flow when it mixes with the swirling flow.
Ht
The effect of changing the total cyclone height Ht on the Euler
number is very small.
Table 8.41 presents the settings used to obtain the optimum design for
minimum pressure using global optimization Matlab toolbox in Matlab
2010a commercial package. Table 8.42 gives the optimum values for cy-
clone geometrical parameters for minimum pressure drop estimated by the
artificial neural network using the genetic algorithm optimization tech-
nique. It is clear from Table 8.42 that the new optimized design is very
close to the Stairmand design in many geometrical parameters, whereas
the new ratios will result in the minimum pressure drop.
Optimal cyclone design for best performance using NSGA-II
In case of cyclone separator geometry optimization for minimum Euler
number and minimum cut-off diameter, the objectives are conflicting with
each other. There is no best solution for which all objectives are optimal
simultaneously [181]. The increase of one objective will lead to the de-
236
8.5. Multi-objective optimization using CFD data
Table 8.40: The variation of the cut-off diameter with cyclone dimensions using the
RBFNN model (cf., Fig. 8.39(b))
Factor Analysis
Dx
The vortex finder diameter has the most significant effect on the
cut-off diameter x50 (the highest slope in Fig. 8.39(b)). The slope is
very high for Dx = 0.4 − 0.55 and any further increase or decrease in
Dx beyond the above range produces a small change in x50.
Increasing the vortex finder diameter decreases the swirling intensity
in the cyclone (i.e. Reduction in the centrifugal force). Consequently,
low collection efficiency (higher x50) is obtained. In the meantime, the
increase of Dx decreases the pressure drop. This is one of the main
reasons of the trade-off between the Euler number and the cut-off
diameter objectives. This makes the optimization of cyclone geometry
a multi-objective procedure.
a and b
The variation of x50 with the inlet width and height are similar in
trend and significance to that for Dx but here the slope is high in the
range of a = 0.3− 0.4 b = 0.2− 0.275.
Ht
The effect of the cyclone total height on the cut-off diameter is due to
the effect of the cone height as the barrel height is fixed in this study.
The slope is very small, with a general trend of inverse relation
(increasing the separation space, enhances the collection efficiency).
crease of the other objective. Then, there should be a set of solutions, the
so-called Pareto optimal set or Pareto front, in which one solution cannot
be dominated by any other member of this set [56].
Recently, a number of multi-objective genetic algorithms (MOGAs) based
on the Pareto optimal concept have been proposed. The well known non-
dominated sorting genetic algorithm II (NSGA-II) proposed by Deb et al.
[34] is one of the most widely used MOGAs since it provides excellent re-
sults as compared with other multi-objective genetic algorithms proposed
[25].
Table 8.43 presents the genetic operators and parameters for multi-objective
optimization. The Euler number and the Stokes number values have been
obtained from the artificial neural network trained by the CFD data set.
The Pareto front (non dominated points) is presented in Fig. 8.40(a) and
Table 8.44. Figure 8.40(a) clearly demonstrate tradeoffs in objective func-
tions (Euler number and Stokes number). All the optimum design points
in the Pareto front are non-dominated and could be chosen by a designer
as optimum cyclone separator [148]. This set of designs makes the Pareto
237
Chapter 8. Optimization
Table 8.41: Genetic operators and parameters for single objective optimization
Population type Double vectorInitial range [0.25 0.15 0.25 3 ; 0.5 0.375 0.75 5 ] for a, b, Dx and Ht respectivelyFitness scaling RankSelection operation: Tournament (tournament size equals 4)Elite count 2Crossover fraction 0.8Crossover operation Intermediate crossover with the default value of 1.0Mutation operation The constraint dependent defaultMaximum number of generations: 800Population size 200
front approach more preferred than the desirability function approach which
gives only one design point. Three points A, B and C are indicated in
Fig. 8.40(a). Point A indicates the point of minimum Euler number (max-
imum Stokes number). Point B indicates the point of maximum Euler
number (minimum Stokes number). Point C indicates an optimal point for
the multi-objective optimization problem.
In order to obtain the Euler number-Stokes number relationship, Fig. 8.40(b)
has been drawn. It indicates a general relationship (trend) between the
two dimensionless numbers irrespective to the four geometrical parame-
ters values. A second-order polynomial has been fitted between the loga-
rithms of Euler number and Stokes number, Eq. 8.18. The obtained corre-
lation can fit the data with a coefficient of determination R2 = 0.99613 as
shown in Fig. 8.40(b). Elsayed and Lacor [56] presented another correla-
tion between the Euler number and the Stokes number, Eq. 8.19. Equa-
tion 8.19 presents a good matching for only high values of Euler numbers.
For smaller values of Euler number, there is underestimation of the Stokes
number. The difference between the two correlations can be referred to two
reasons. Firstly, the new correlation is based on the CFD simulations data
for both Euler number and Stokes number, whereas Elsayed and Lacor
Table 8.42: The optimized cyclone separator design for minimum pressure drop
using genetic algorithm
Factor Low High Stairmand design∗ Optimum design†
a 0.25 0.5 0.5 0.492
b 0.15 0.375 0.2 0.158
Dx 0.25 0.75 0.5 0.617
Ht 3.0 5.0 4.0 4.535
Euler number 5.606 2.369
Cut-off diameter 1.706 1.704
∗ The values for the Euler number and the cut-off diameter for the Stairmand design have been obtainedfrom CFD simulations, cf. Table 8.38.
† The values for the Euler number and the cut-off diameter for the new optimal design have beenobtained from the trained RBFNN.
238
8.5. Multi-objective optimization using CFD data
Table 8.43: Genetic operators and parameters for multi-objective optimization
Population type Double vectorPopulation size 60 (i.e., 15* number of variables which is the default in Matlab)Initial range [0.25 0.15 0.25 3 ; 0.5 0.375 0.75 5 ] for a, b, Dx and Ht respectivelySelection operation tournament (tournament size equals 2)Crossover fraction 0.8Crowding distance fraction 0.35Crossover operation Intermediate crossover with the default value of 1.0Maximum number of generations 800 (i.e., 200* number of variables which is the default in Matlab)
correlation was obtained from experimental data for Euler number values
and Iozia and Leith model for the Stokes number values. Secondly, the
new correlation is limited to only four geometrical parameters. The other
three factors are fixed, h = 1.5, S = 0.5 and Bc = 0.375. Consequently, the
new correlation (Eq. 8.18) is valid only for these values.
Stk50 = 100.3533(log10(Eu))2−1.1645log10(Eu)−2.3198 (8.18)
Stk50 = 100.3016(log10(Eu))2−0.9479log10(Eu)−2.5154 (8.19)
8.5.5 Conclusions
CFD simulations data have been used to understand the effect of four
geometrical parameters on the cyclone performance and to optimize the
cyclone geometry. Two meta-models have been used viz., the response
surface and the radial basis function neural network approaches. Two
optimization techniques have been applied, the desirability function with
Nelder-Mead technique and the non-sorted dominated genetic algorithm
NSGA-II.
• The response surface methodology has been used to fit two second-
order polynomials to the Euler number and cut-off diameter obtained
from CFD simulations. The analysis of variance of the Euler number
indicates a strong interaction between Dx with (a, b) and between Dx
with b only for the cut-off diameter.
• The bi-objective functions have been converted to single-objective func-
tion using the desirability function approach. A new optimal de-
sign has been obtained using the Nelder-Mead technique available in
Statgraphics commercial software. The ratio of inlet width to height
b/a = 0.589 lays in the optimal cyclone lies in the recommended range
of b/a from 0.5 to 0.7 proposed by Elsayed and Lacor [52]. The new de-
sign and the Stairmand design have been computationally compared
239
Chapter 8. Optimization
Euler number
Sto
kes
num
berx
103
5 10 15 20 25 30 35
1
2
3
4
5
6
7 A
BC
(a) Pareto front in linear scale
Euler number
Sto
kes
num
berx
10
3
5 10 15 20 25 30
2
4
6
8
10
Non-dominated pointsElsayed & Lacor correlationCurve-fitting
Stk50=10 0.3533 (LOG10(Eu))^2- 1.1645 (LOG10(Eu) -2.31 98
R2=0.99613
(b) Pareto front in log scale. Note: The correlation
(Eq. 8.19) given by Elsayed and Lacor [45] are based on
seven geometrical parameters.
Figure 8.40: Pareto front plots obtained from the variation of four geometrical
ratios.
240
8.5. Multi-objective optimization using CFD data
Table 8.44: The seven geometrical parameters and the obtained Euler number and
Stokes number for the nondominated points (Pareto-front).
point a b Dx Ht h S Bc Euler number Stokes number x 103
1 0.500 0.375 0.250 4.994 1.5 0.5 0.375 33.750 0.511
2 0.391 0.296 0.250 4.979 1.5 0.5 0.375 20.806 0.569
3 0.374 0.270 0.252 4.961 1.5 0.5 0.375 17.883 0.595
4A 0.250 0.150 0.750 4.899 1.5 0.5 0.375 0.756 6.958
5 0.252 0.152 0.362 4.893 1.5 0.5 0.375 3.326 1.262
6 0.250 0.150 0.732 4.846 1.5 0.5 0.375 0.799 6.4467 0.320 0.214 0.300 4.929 1.5 0.5 0.375 8.617 0.840
8 0.363 0.268 0.300 4.945 1.5 0.5 0.375 12.160 0.778
9 0.259 0.159 0.472 4.960 1.5 0.5 0.375 2.086 2.080
10 0.391 0.284 0.252 4.966 1.5 0.5 0.375 19.737 0.581
11 0.419 0.327 0.256 4.983 1.5 0.5 0.375 23.514 0.56812 0.251 0.151 0.711 4.853 1.5 0.5 0.375 0.854 5.903
13 0.455 0.365 0.251 4.976 1.5 0.5 0.375 29.693 0.528
14 0.251 0.151 0.695 4.858 1.5 0.5 0.375 0.895 5.504
15 0.298 0.189 0.441 4.910 1.5 0.5 0.375 3.277 1.554
16 0.376 0.270 0.262 4.960 1.5 0.5 0.375 16.657 0.63017 0.472 0.358 0.251 4.986 1.5 0.5 0.375 30.324 0.525
18B 0.500 0.375 0.250 4.994 1.5 0.5 0.375 33.750 0.51119 0.255 0.155 0.608 4.810 1.5 0.5 0.375 1.214 3.820
20 0.252 0.152 0.497 4.879 1.5 0.5 0.375 1.759 2.380
21C 0.286 0.183 0.346 4.907 1.5 0.5 0.375 4.954 1.103
22 0.351 0.233 0.295 4.949 1.5 0.5 0.375 10.568 0.789
23 0.251 0.151 0.638 4.897 1.5 0.5 0.375 1.050 4.354
24 0.265 0.154 0.460 4.881 1.5 0.5 0.375 2.193 1.780
25 0.493 0.321 0.250 4.991 1.5 0.5 0.375 28.441 0.531
26 0.312 0.206 0.320 4.869 1.5 0.5 0.375 7.136 0.94127 0.352 0.234 0.295 4.950 1.5 0.5 0.375 10.728 0.784
28 0.486 0.365 0.250 4.992 1.5 0.5 0.375 31.833 0.518
29 0.252 0.152 0.730 4.840 1.5 0.5 0.375 0.817 6.387
30 0.252 0.153 0.676 4.861 1.5 0.5 0.375 0.957 5.071
31 0.251 0.151 0.650 4.895 1.5 0.5 0.375 1.013 4.56532 0.426 0.333 0.252 4.977 1.5 0.5 0.375 25.126 0.551
33 0.261 0.150 0.747 4.838 1.5 0.5 0.375 0.798 6.822
34 0.272 0.161 0.553 4.880 1.5 0.5 0.375 1.630 2.945
35 0.253 0.154 0.553 4.880 1.5 0.5 0.375 1.451 3.022
36 0.251 0.151 0.711 4.853 1.5 0.5 0.375 0.854 5.90337 0.266 0.284 0.252 4.966 1.5 0.5 0.375 13.428 0.633
Minimum 0.25 0.150 0.250 4.810 1.5 0.5 0.375 0.7555 0.511Maximum 0.50 0.375 0.750 4.995 1.5 0.5 0.375 33.7497 6.958
A indicates the point of minimum Euler number and maximum Stokes number. B indicates the point of maximum Euler number and minimum
Stokes number. C indicates an optimal point for the multi-objective optimization problem. (cf. Fig. 8.40(a)).
to get a clear vision for the differences between the flow field pattern
and performance in the two designs. The CFD simulations results
and four mathematical models confirmed the better performance of
the new design in comparison with the Stairmand design.
• The result demonstrates that artificial neural networks can offer an
alternative and powerful approach to model the cyclone performance
better than the response surface methodology. The used RBFNN pre-
sented zero mean squared error and almost unity coefficient of deter-
mination.
• The analysis indicates the significant effect of the vortex finder diam-
eter Dx and the inlet dimensions a and b on the cyclone performance.
Moreover, the range of high influence is given for each geometrical
parameters using the trained RBFNNs.
• The trained RBFNN for the Euler number has been used to get a
new optimized cyclone for minimum pressure drop (Euler number)
241
Chapter 8. Optimization
using the genetic algorithm optimization technique. The new cyclone
design is very close to the Stairmand high efficiency design in the
geometrical parameter ratio, and superior in low pressure drop at
nearly the same cut-off diameter. But, the optimal design obtained
from the desirability function results in a better collection efficiency
(smaller cut-off diameter) as is clear from Table 8.38 because there
the cut-off diameter is included in the desirability function, whereas,
the obtained result are for single objective (Euler number).
• The two trained RBFNNs have been used in a multi-objective opti-
mization process using NSGA-II technique. The Pareto front is pre-
sented for the designer with a wide choice for selection.
• A second-order polynomial has been fitted between the logarithms
of Euler number and Stokes number to obtain a general formula,
Stk50 = 100.3533(log10(Eu))2−1.1645 log10(Eu)−2.3198 with a coefficient of de-
termination R2 = 0.98643. This formula can be used to obtain the
Stokes number if the Euler number is known at h = 1.5, S = 0.5 and
Bc = 0.375.
242
Chapter 9
Conclusions and Future
Directions
9.1 Conclusions
9.1.1 The most significant geometrical factors
The geometrical parameters in cyclone separators affect significantly the
flow field and performance parameters. There are seven geometrical pa-
rameters, which can be classified into four classes, namely, the inlet di-
mensions (height and width), the vortex finder dimensions (diameter and
length), the cyclone height (cone and barrel) and the cone-tip diameter.
These dimensions do not have the same influence on the cyclone perfor-
mance. Consequently, the first step was to identify the most significant
parameters and the possible interaction between them.
The Muschelknautz method of modeling (MM)) has been used to under-
stand the effect of the cyclone geometrical parameters on the cyclone per-
formance. The most significant geometrical parameters are:
1. The vortex finder diameter
2. The inlet section width
3. The inlet section height
4. The cyclone total height (cone height).
The effect of both the barrel height and the vortex finder length on the cy-
clone separator performance are small in comparison with these most sig-
nificant geometrical parameters. There are strong interactions between
243
Chapter 9. Conclusions and Future Directions
the effects of inlet dimensions and the vortex finder diameter on the cy-
clone performance.
Large eddy simulation methodology has been used to study the effect of the
cone-tip diameter on the cyclone flow field and performance. The analysis
of results indicates the insignificant effect of the cone-tip diameter on the
flow pattern and performance.
Four cyclones have been simulated using the Reynolds stress model (RSM),
to study the effect of the dust outlet geometry on the cyclone separator
performance, flow pattern and velocity profiles. The results approved that
if the main target of the CFD investigation is the performance parameters,
one can safely simulate only the main separation space (cylinder on cone).
However, if the aim is to investigate the flow field pattern, the dust outlet
geometry should be included in the simulation domain.
9.1.2 The impact of geometry
The vortex finder dimensions: The maximum tangential velocity in the
cyclone decreases with increasing the vortex finder diameter. A neg-
ligible change is noticed with increasing the vortex finder length. In-
creasing the vortex finder length, makes a small change in both the
static pressure, axial and tangential velocity profiles. However, de-
creasing the vortex finder diameter gradually changes the axial ve-
locity profile from the inverted W to the inverted V class. Decreasing
the cyclone vortex finder diameter, increases the maximum tangen-
tial velocity. The Euler number (dimensionless pressure drop) de-
creases with increasing the vortex finder diameter. Increasing the
vortex finder length slightly increases the Euler number. The Stokes
number increases with increasing the vortex finder diameter, because
the centrifugal force affecting particles attenuates when the swirl in-
tensity (maximum tangential velocity) decreases. The Stokes number
slightly increases as the vortex finder length is increased.
The inlet dimensions: The maximum tangential velocity in the cyclone
decreases with increasing both the cyclone inlet width and height. In-
creasing the cyclone inlet width or height decreases the pressure drop
at the cost of increasing the cut-off diameter. The effect of changing
the inlet width on the cut-off diameter is more significant in compari-
son with that of the inlet height. The optimum ratio of the inlet width
to the inlet height is from 0.5 to 0.7.
The cyclone heights: The maximum tangential velocity in the cyclone
decreases with increasing the cyclone (barrel or cone) height. In-
creasing the barrel height, makes a small change in the axial veloc-
244
9.1. Conclusions
ity. Increasing the cyclone barrel height decreases the pressure drop
and the cut-off diameter. The changes in the performance beyond
h/D = 1.8 are small. Increasing the cone height makes a consider-
able change in the axial velocity. Both the pressure drop and the cut-
off diameter decrease with increasing the cyclone cone height. The
performance improvement stops after hc/D = 4.0 (Ht/D = 5.5). The
effect of changing the barrel height is less significant on the perfor-
mance and the flow pattern in comparison with the effect of the cone
height.
9.1.3 Optimization
Several new optimized cyclone geometrical ratios have been obtained and
presented in this thesis. All the new geometrical ratios result in better
performance than the Stairmand design. For example, The new cyclone
design ratios obtained using MM model and Nelder-Mead technique for
minimum pressure drop, are very close to the Stairmand high efficiency
design in the geometrical parameter ratios, and superior for low pressure
drop at nearly the same cut-off diameter. The new cyclone design results
in nearly one-half the pressure drop obtained by the Stairmand design at
the same volume flow rate.
Artificial neural networks
In order to accurately predict the complexly non linear relationships be-
tween pressure drop and geometrical dimensions, a radial basis neural
network (RBFNN) is developed and employed to model the pressure drop
for cyclone separators. The neural network has been trained and tested by
the experimental data available in literatures.
• The result demonstrates that artificial neural networks can offer an
alternative and powerful approach to model the cyclone performance.
• Four mathematical models (Muschelknautz method “MM”, Stairmand,
Ramachandran and Shepherd and Lapple) have been tested against
the experimental values. The residual error of MM model is the low-
est. Moreover, one can multiply the calculated value using MM by
1.5 to get the experimental value (as a rough approximation).
• The response surface methodology has been used to fit a second-order
polynomial to the RBFNN.
• The second-order polynomial has been used to get a new optimized
cyclone for minimum pressure drop using the Nelder-Mead technique.
245
Chapter 9. Conclusions and Future Directions
• A comparison between the new design and the standard Stairmand
design has been performed using CFD simulation to obtain a clear
vision of the flow field pattern in the new design.
• The new cyclone design results in nearly 75% of the pressure drop
obtained by the Stairmand high efficiency design at the same volume
flow rate.
9.1.4 Multi-objective optimization
Two multi-objective optimization approaches have been applied in this
thesis, namely the desirability function and the non-dominated sorted ge-
netic algorithm (NSGA-II). Moreover, two sources of data have been used,
namely: (1) The experimental data for the pressure drop (Euler number)
and the Iozia and Leith model for the cut-off diameter (Stokes number) for
the optimization of the seven geometrical parameters. (2) CFD simulations
for the optimization of only four geometrical parameters.
Seven geometrical parameters
Two radial basis neural networks (RBFNNs) are developed and employed
to model the Euler number and the Stokes number for cyclone separators.
The neural networks have been trained and tested by the experimental
data available in literature for Euler number (pressure drop) and the Iozia
and Leith model [85] for the Stokes number (cut-off diameter). The effects
of the seven geometrical parameters on the Stokes number have been in-
vestigated using the trained ANN. To declare any interaction between the
geometrical parameters affecting the Stokes number, the response surface
methodology has been applied. The trained ANN has been used as an ob-
jective function to obtain the cyclone geometrical ratios for minimum Euler
number using the genetic algorithms optimization technique. A CFD com-
parison between the new optimal design and the Stairmand design has
been performed. A multi-objective optimization technique using NSGA-
II technique has been applied to determine the Pareto front for the best
performance cyclone separator.
• The trained RBFNN has been used to get a new optimized cyclone for
minimum pressure drop (Euler number) using the genetic algorithm
optimization technique.
• The new cyclone design results in nearly 68% of the pressure drop
obtained by the Stairmand design at the same volume flow rate. This
246
9.1. Conclusions
confirms that the obtained design using the genetic algorithm is bet-
ter than that obtained using Nelder-Mead technique which results in
75% of the Stairmand pressure drop [53].
• The two trained RBFNNs have been used in a multi-objective opti-
mization process using NSGA-II technique. Sixteen test cases with
different barrel diameter, gas flow rate and particle density have
been tested. The Pareto fronts for these test cases are very close.
A second-order polynomial has been fitted between the logarithms
of the Euler number and the Stokes number to obtain a general for-
mula, Stk50 = 100.3016(log10(Eu))2−0.9479 log10(Eu)−2.5154. This formula
can be used to obtain the Stokes number if the Euler number is
known.
Four geometrical parameters
CFD simulations data have been used to understand the effect of four
geometrical parameters on the cyclone performance and to optimize the
cyclone geometry. Two meta-models have been used viz., the response
surface and the radial basis function neural network approaches. Two
optimization techniques have been applied, the desirability function with
Nelder-Mead technique and the non-sorted dominated genetic algorithm
NSGA-II.
• The response surface methodology has been used to fit two second-
order polynomials to the Euler number and the cut-off diameter ob-
tained from CFD simulations. The analysis of variance of the Euler
number indicates a strong interaction between Dx with (a, b) and be-
tween Dx with b only for the cut-off diameter.
• The bi-objective functions have been converted to single-objective func-
tion using the desirability function approach. A new optimal design
has been obtained using the Nelder-Mead technique. The ratio of
inlet width to height b/a = 0.589 i.e., in the optimal cyclone lies in
the recommended range of b/a from 0.5 to 0.7 proposed by Elsayed
and Lacor [52]. The new design and the Stairmand design have been
computationally compared to get a clear vision for the differences be-
tween the flow field pattern and performance in the two designs. The
CFD simulations results confirmed the superior performance of the
new design in comparison with the Stairmand design.
• The result demonstrates once more that artificial neural networks
can offer an alternative and powerful approach to model the cyclone
performance better than the response surface methodology. The used
247
Chapter 9. Conclusions and Future Directions
RBFNN presented zero mean squared error and almost unity coeffi-
cient of determination.
• The analysis indicates the significant effect of the vortex finder diam-
eter Dx and the inlet dimensions a and b on the cyclone performance.
Moreover, the range of high influence is given for each geometrical
parameters using the trained RBFNNs.
• The trained RBFNN for the Euler number has been used to get a
new optimized cyclone for minimum pressure drop (Euler number)
using the genetic algorithm optimization technique. The new cyclone
design is very close to the Stairmand high efficiency design in the
geometrical parameter ratio, and superior in low pressure drop at
nearly the same cut-off diameter. But, the optimal design obtained
from the desirability function results in a better collection efficiency
(smaller cut-off diameter) because there the cut-off diameter is in-
cluded in the desirability function, whereas, the obtained result are
for single objective (Euler number).
• The two trained RBFNNs have been used in a multi-objective opti-
mization process using NSGA-II technique. The Pareto front is pre-
sented for the designer with a wide choice for selection.
• A second-order polynomial has been fitted between the logarithms
of the Euler number and the Stokes number to obtain a general for-
mula, Stk50 = 100.3533(log10(Eu))2−1.1645 log10(Eu)−2.3198. This formula
can be used to obtain the Stokes number if the Euler number is
known at h = 1.5, S = 0.5 and Bc = 0.375.
9.2 Future Directions
As a recommendation of future work, the same study is to be performed but
at different flow rates and different particle densities. Furthermore, the
effect of the cyclone dimensions on natural vortex length and precessing
vortex core needs more investigation using large eddy simulation method-
ology.
The present study on the dust outlet geometry is to be enlarged to include
the following. (1) The effect of dustbin and dipleg shape, length, diameter,
and interior details (like the inner cone at the dustbin inlet) and optimiza-
tion of these dimensions. (2) The effect of air and dust flow rates, and
particle properties. (3) perform the same study but using large eddy simu-
lation (LES) to check the effect of dust outlet geometry on the vortex core
precession.
248
9.2. Future Directions
The following issues still need more investigation. (1) test more mathe-
matical models against the experimental value. (2) comparison between
the support vector machine approach and neural networks approach for
cyclone separator performance estimation. (3) create a neural network
model to design the cyclone separator and estimate its performance pa-
rameters. (4) generate performance curves for each geometrical and oper-
ating parameters that affect the cyclone performance to help the designer
in predicting the change of the performance due to change in the cyclone
loading and operating conditions.
Furthermore, many parameters can be taken into consideration in the
optimization process, e.g., erosion rate for harsh environments, surface
roughness, interaction between particles using two-phase flow modeling.
Another idea is to perform an geometry optimization (aeroacoustic) study
to reduce the cyclone-noise especially for domestic applications.
249
Chapter 9. Conclusions and Future Directions
250
Appendix A
Mathematical models
Since the first application of aerocyclones in 1886 [3], theories for the es-
timation of both particle collection efficiency and pressure drop of cyclone
have been developed by many contributors using different methods with
various simplifying assumptions. During the past 50 years, interest in
particle collection and pressure theories has steadily increased [196]. The
most widely used mathematical models for the cut-off diameter and pres-
sure drop estimation are:
• Barth model [9]
• The Muschelknautz method of modeling (MM) [29, 77, 114–116, 174,
175]
• Stairmand model [165]
• Shepherd and Lapple model [157]
• Casal and Martinez-Bent model [21]
• Ramachandran model [139]
• Iozia and Leith model [84]
• Rietema model [142]
A.1 General assumptions
Some simplifying assumptions are common to all these models. They can
be considered as offering a good compromise between accurate prediction
and simplification of the equations [3]. They are:
• The particles are spherical.
• The particle motion is not influenced by the presence of neighboring
particles.
• The radial velocity of the gas equals zero.
251
APPENDIX A. Mathematical models
!"#$%&'(
)*+,#*&(
(-.#/"%'()0(
Figure A.1: The control surface concept in the equilibrium-orbit model.
• The radial force on the particle is given by Stokes’ law.
A.2 Barth model
In 1956, Barth [9] proposed a simple model based on force balance (clas-
sified as one of the equilibrium-orbit models [77]). This model enables to
obtain the cut-off diameter and the pressure drop values. Barth proposed
calculating the wall velocity (the tangential velocity near the wall of the
cyclone surface) and the tangential velocity at the control surface CS (
Fig. A.1) through two steps:
1. calculate the wall velocity vθw (the velocity outside CS) from the inlet
velocity vin.
2. use its value to calculate the tangential velocity, vθCS at CS.
In a cyclone with a slot type of rectangular inlet, the inlet jet is compressed
against the wall, resulting in a decrease in the area available for the in-
coming flow and an increase in the velocity. Barth accounted for this by
introducing α, which is defined as the ratio of the moment-of-momentum
of the gas in the inlet and the gas flowing along the wall.
α =vinRin
vθwR(A.1)
252
A.2. Barth model
Figure A.2: Inlet flow pattern for tangential inlet cyclone [77]
where Rin is the radial position of the center of the inlet (Fig. A.2), For a
slot inlet Rin = R− b2 , where b is the inlet width and R = D/2 is the cyclone
radius.
Hoffmann and Stein [77] gave an algebraic relations for α, the simplest of
which is
α = 1− 0.4
(b
R
)0.5
(A.2)
So vθw can be calculated as vθw = vinRin
αR , as the inlet velocity is given or
calculated from vin = Qa b where Q is the gas volume flow rate, a is the inlet
height.
To get from vθw to the tangential velocity at CS vθCS , Barth gave the fol-
lowing relation
vθCS =vθw(
RRx
)
1 + HCSRπfvθwQ
(A.3)
Introducing Eq. (A.1), one obtains
vθCS =πRinRxvx
a bα+HCSπfRin(A.4)
Where Rx is the vortex finder radius Rx = Dx
2 , vx is the mean axial velocity
in the vortex finder vx = QπR2
x, and HCS is the height of the control surface
extending from the bottom of the vortex finder to the cyclone bottom.
253
APPENDIX A. Mathematical models
For the friction factor f , the following relation can be used for hydraulically
smooth cyclone surface according to Hoffmann and Stein [77],
f = 0.05(1 + 3√co) (A.5)
Where co is the mass ratio of dust feeding the cyclone to the gas flow rate
(dimensionless).
Estimation of the pressure drop
Barth subdivided the pressure drop into three contributions:
1. the inlet losses (Barth assumed that this loss could be effectively
avoided by good design).
2. the losses in the cyclone body.
3. the losses in the vortex finder.
The pressure drop in the cyclone body can be estimated from,
∆Pbody =1
2ρv2x
(Dx
D
) 1(
vxvθCS
− H−S0.5Dx
f)2 −
(vθCs
vx
)2
(A.6)
where f is the friction factor (calculated from Eq. A.5). This model ac-
counts for the effect of solid loading upon pressure loss via the total friction
factor f .
The pressure drop in the vortex finder can be estimated using a semi- em-
pirical approach as,
∆Px =1
2ρv2x
((vθCS
vx
)2
+K
(vθCS
vx
) 43
)(A.7)
where K is the vortex finder entrance factor (K = 3.41 for rounded edge
and K = 4.4 for sharp edge)
The total pressure drop ∆P = ∆Pbody + ∆Px can be made dimensionless
using the average inlet velocity vin = Qab leading to the so called Euler
number Eu
Eu =∆P
0.5ρv2in(A.8)
254
A.2. Barth model
Estimation of the cut-off diameter
As mentioned above, the Barth’s model [9] is based on an equilibrium-orbit
model [77]. This model considers the imaginary cylindrical surface CS that
is formed by continuing the vortex finder wall to the bottom of the cyclone
Fig. A.1. It is based on the force balance of a particle that is rotating in CSat radius Rx = Dx
2 . The outwardly directed centrifugal force is balanced
against the inward drag caused by the gas flowing through surface CS into
the inner part of the vortex.
Fcentrifugal = ma = mpv2θCS
Rx= ρp
(π6x3) v2θCS
Rx(A.9)
where the subscript p is for particle properties and x is the particle diame-
ter. Since, for a creep flow, Rep ≪ 1 where Rep =ρgxvrp
µg, the drag coefficient
for sphere CD = 24Rep
. Hence, the Stokesian drag force can be estimated as:
Fdrag = 3πµgvrCS x (A.10)
where vrCS is the uniform radial gas velocity in the surface of CS given by:
vrCS =Q
πDxHCS(A.11)
The following expression obtained from trigonometry relations can be used
to calculate HCS
HCS =(R−Rx)(H − h)
R− (Bc2 )
+ (h− S) if Bc > Dx
= (H − S) if Bc ≤ Dx (A.12)
Large particles are therefore centrifuged out to the cyclone wall (because
the centrifugal force is larger than the drag force. Whereas, small particles
are dragged in and escape out through the vortex finder. The particle size
for which the two forces balance the particles that orbit in equilibrium in
CS is taken as the cyclone’s cut-off diameter x50; it is the particle size that
stands a 50 - 50 chance of being captured. This particle size is called of
fundamental importance and is a measure of the intrinsic separation ca-
pability of the cyclone. Here, all the gas velocity components are assumed
constant over CS for the computation of the equilibrium-orbit size. Equat-
ing the forces, in Eqs. A.9 and A.10 gives the cut-off diameter x50 as:
x50 =
√9vrCS µg Dx
ρp v2θCS
(A.13)
255
APPENDIX A. Mathematical models
A.3 The Muschelknautz method of modeling
(MM)
Hoffmann and Stein [77] stated that the most practical method for mod-
eling cyclone separators at the present time is the Muschelknautz method
(MM) [29, 77, 114–116, 174, 175]. The roots of the Muschelknautz method
(MM) extend back to an early work performed by Barth [9] as it is based
on the equilibrium orbit model [77].
The pressure loss in cyclone
According to the MM model, the pressure loss across a cyclone occurs,
primarily, as a result of friction with the walls and irreversible losses
within the vortex core, the latter often dominating the overall pressure
loss, ∆p = ∆pbody+∆px. In a dimensionless form, it is defined as the Euler
number.
Eu =1
12ρv
2in
[∆pbody +∆px] (A.14)
The wall loss, or the loss in the cyclone body is given by,
∆pbody = fAR
(0.9 Q)
ρ
2(vθwvθCS)
1.5(A.15)
where ρ is the gas density, Q is the gas volume flow rate, AR is the total
inside area of the cyclone contributing to frictional drag. It encompasses
the inside area of the roof, the barrel cylinder, the cone, and the external
surface of the vortex finder.
AR = Aroof +Abarrel +Acone +Avortex finder
= π[R2 −R2
x + 2Rh+ (R+Rb)√(Ht − h)2 + (R −Rb)2
]+2πRxS
(A.16)
where R = D/2, Rx = Dx/2, Rb = Bc/2.
As in the Barth’s model [9, 77], the tangential gas velocity in the entire
space between the wall and the vortex finder can be significantly higher
than the inlet velocity due to constriction of the inlet jet (see Fig. A.2). For
tangential inlet, the inlet jet is compressed against the wall, resulting in
a decrease in the area available for the incoming flow, and an increase in
the velocity. Barth accounts for this by introducing α (see Eq. A.1).
256
A.3. The Muschelknautz method of modeling (MM)
Muschelknautz computes the entrance constriction coefficient α for a con-
ventional slot-type inlet from the following empirical formula, Eq.A.17
[77].
α =1
β
1−
√√√√1 + 4
[(β
2
)2
−(β
2
)]√1− (1− β2)(2β − β2)
1 + c0
(A.17)
where β = b/R, c0 is the ratio of the mass of the incoming solids to the
mass of the incoming gas in the stream feeding the cyclone.
Knowing α, along with vin, Rin and R, one can compute the wall veloc-
ity, vθw (velocity in the vicinity of the wall), vθw = (vinRin)/(αR). Now
compute the geometric mean radius, Rm =√RxR which is needed in the
computation of a wall axial velocity vzw .
vzw =0.9Q
π(R2 −R2m)
(A.18)
Trefz and Muschelknautz [174] found that, approximately 10% of the in-
coming gas “short-circuits” the cyclone and flows radially inwards in a spi-
ral like manner along the roof and down the outside of the vortex finder.
As a consequence, approximately 90% of the incoming flow Q directly par-
ticipates in the flow along the walls and in the formulation of the inner
vortex [77].
To calculate the friction factor f , Muschelknautz and Trefz [77] defined the
cyclone body Reynolds number ReR as:
ReR =RinRmvzwρ
Htµ(A.19)
with ρ and µ are the gas phase density and absolute viscosity, respectively.
Ht is the cyclone total height. The friction factor of the clean gas fair =
f
(ksR,ReR
)where
ksR
is the wall relative roughness of the cyclone wall
(ks = 0.046 [mm] for commercial steel pipe,ksR
6< 6E − 4 for non-negative
logarithm in Eq. A.22). The gas friction factor can be expressed as the sum
of two components, one for smooth wall, fsm, plus an added contribution
due to wall roughness fr.
fair = fsm + fr (A.20)
fsm = 0.323Re−0.623R (A.21)
257
APPENDIX A. Mathematical models
fr =
log
1.6
ksR
− 0.0005999
2.38
−21 +
2.25E5
Re2R
(ksR
− 0.000599
)0.213
(A.22)
The total frictional drag f within a cyclone consists of two components in
the MM, that due to drag on the (pure) gas phase fair and that due to an
additional drag imposed by the moving strand of solids, which is present
at the walls. The total friction factor f becomes,
f = fair + 0.25
(R
Rx
)−0.625√
η c0 Frx ρ
ρstr(A.23)
The second term in this equation is the frictional contribution due to the
solids, where η is the overall efficiency, that is the fraction of incoming
solids collected by the cyclone (Hoffmann and Stein [77] suggested to as-
sume a value of 0.9 to 0.99). Frx is the Froude number (Frx = vx√2 Rx g
)
where vx is the average axial velocity through the vortex finder vx (vx =Q
π R2x
), ρstr term represents the bulk density of the dust or strand layer at
the walls and can be taken as 0.4ρbulk where ρbulk is the bulk density of the
solid [77].
The tangential velocity of the gas at the inner core radius Rcs (see Fig. A.3)
is given by,
vθcs = vθw
(RRx
)
[1 +
f AR
√
RRx
2 Q
] (A.24)
The second contribution to pressure drop is the loss in the core and in the
vortex finder and is given by,
∆px =
[2 +
(vθcsvx
)2
+ 3
(vθcsvx
)4/3]1
2ρv2x (A.25)
Cut-off diameter
A very fundamental characteristic of any lightly loaded cyclone is its cut-off
diameter x50 produced by the spin of the inner vortex. This is the practical
diameter that has a 50% probability of capture. The cut-off diameter is
258
A.4. Stairmand model for pressure drop
Q
0.1Q
Rc
Rcs
Vz
Vx
Vөw
Vөcs
Figure A.3: Geometric parameters and velocities used in MM model [77]
analogous to the screen openings of an ordinary sieve or screen [77]. In
lightly loading cyclones, x50 exercises a controlling influencing on the cy-
clone’s separation performance. It is the parameter that determines the
horizontal position of the cyclone grade-efficiency curve (fraction collected
versus particle size). For low mass loading, the cut-off diameter can be
estimated in MM using Eq. A.26 [77].
x50 =
√9 µ (0.9 Q)
π (ρp − ρ)v2θcs(Ht − S)(A.26)
A.4 Stairmand model for pressure drop
Stairmand [165] estimated the pressure drop as entrance and exit losses
combined with the static pressure loss in the swirl.
Eu = 1 + 2q2(2 (D − b)
Dx− 1
)+ 2
(4 a b
πD2x
)2
(A.27)
259
APPENDIX A. Mathematical models
where q is given by,
q =−(
Dx
2(D−b)
)0.5+(
Dx
2(D−b) +4ARGa b
)0.5(2ARGa b
) (A.28)
where AR is the total wall area of the cyclone body, including the inner
walls of the lid, the cylindrical and the conical sections and the outer wall
of the vortex finder, given by Eq. A.29. G(= f/2 where f is the friction factor)is a wall friction factor, which Stairmand set equal to 0.005.
AR =π(D2 −D2
x
)
4+πDh+πDxS+
π (D +Bc)
2
((H − h)
2+
(D −Bc
2
)2)0.5
(A.29)
A.5 Purely empirical models for pressure drop
1. Sphered and Lapple model [157]
Eu =16 a b
D2x
(A.30)
2. Casal and Martinez-Bent model [21]
Eu = 3.33 + 11.3
(a b
D2x
)2
(A.31)
3. Ramachandran model
The Ramachandran et al. [139] model was developed through a sta-
tistical analysis of pressure drop data for ninety-eight cyclone de-
signs. The model is shown to perform better than the pressure drop
models of Shepherd and Lapple [157], and Barth [9] in comparison
with experimental results.
Eu = 20
a b
D2x
SD
HD
hD
Bc
D
1/3
(A.32)
260
A.6. Iozia and Leith model for the cut-off diameter
!!
"!!
!!
!!!
!!!
!!!
!
!!!
!!!
!!!
!!!!!!!!
!!!!
!!!!
Figure A.4: The control surface used in Iozia and Leith model [85].
A.6 Iozia and Leith model for the cut-off di-
ameter
The Iozia and Leith model [84] is similar to the model of Barth [9] as it is
also based on the equilibrium-orbit theory (Force balance). Iozia and Leith
[84] gave the following expression for the cut-off diameter,
x50 =(9 µ Q) /
(π HCS ρp V 2
θmax
)1/2(A.33)
where HCS is the core height (height of the control surface of Barth’s
model) Vθmax is the maximum tangential velocity, that occurs at the edge
of the control surface CS, Fig. A.4. In this model however the value of
the core diameter dc and the tangential velocity at the core edge Vθmax are
calculated from regression of experimental data using the following equa-
tions.
Vθmax = 6.1vin
(ab/D2
)−0.61(Dx/D)
−0.74(Ht/D)
−0.33
(A.34)
dc = 0.52D(ab/D2
)−0.25(Dx/D)1.53 (A.35)
It is clear from this model that the most important geometry parameters
that affect the cyclone collection efficiency are the vortex finder diameter,
the ratio of inlet area to exit area and the cyclone height.
261
APPENDIX A. Mathematical models
A.7 Rietema model for cut-off diameter
The Rietema model [142] estimates the cut-off diameter x50 using Eq. A.36.
x50 =
√µ ρ Q
(ρp − ρ) Ht ∆P(A.36)
This model relates, the separation cut-off diameter x50 to the pressure
drop. Hence, the pressure drop needs to be predicted to use the model.
A good pressure drop model for this purpose is that of Shepherd and Lap-
ple. The interested reader can refer to Hoffmann and Stein [77] for more
details.
262
Appendix B
Optimization Techniques
Two optimization techniques have been used throughout this thesis, namely
the Nelder-Mead technique and the genetic algorithms. In the following
sections, the details of these techniques will be explained.
B.1 Nelder-Mead
The Nelder-Mead method, also known as downhill simplex method is a
commonly used nonlinear optimization technique, The technique was pro-
posed by Nelder and Mead [119] and is a technique for minimizing an
objective function in a many-dimensional space. It requires only function
evaluations, and no calculation of derivatives [134]. The downhill simplex
nonlinear optimization technique has been used by many researchers [e.g.,
4, 12]. According to Bernon et al. [12] Powell’s algorithm and the down-
hill simplex one are ones of the most used minimization algorithms; the
downhill-simplex algorithm became the most performant. Further more,
the Statgraphics XV package has been used for design of experiment and
optimization, with the only available optimization technique is the Nelder-
Mead technique. In this study, the target is to obtain the global optimum
values. Consequently, no linear constrains applied.
The idea is to employ a moving simplex in the design space to surround
the optimal point and then shrink the simplex until its dimensions reach a
specified error tolerance [96]. In n-dimensional space, a simplex is a figure
of n+1 vertices connected by straight lines and bounded by polygonal faces.
If n = 2, a simplex is a triangle; if n = 3, it is a tetrahedron.
For two variables, the simplex is a triangle and the method is a pattern
search that compares function values at the three vertices of a triangle.
The worst vertex where f(x, y) is largest, is rejected and replaced with a
263
APPENDIX B. Optimization Techniques
new vertex. A new triangle is formed, and the search is continued. The pro-
cess generates a sequence of triangles (which might have different shapes),
for which the function values at the vertices get smaller and smaller. The
size of the triangles is reduced, and the coordinates of the minimum point
are found [108]. The algorithm is stated using the term simplex (a general-
ized triangle in N dimensions) and will find the minimum of a function of
N variables. It is effective and computationally compact. In the following
paragraphs, a brief explanation is given for this technique. The interested
reader can refer to Mathews and Fink [108].
!
R
G
W
Md
d
Reflection
B
GW
M C2
Contraction
R
M
C1
B
GW
M
Shrinkage
S
B
GW
Md
d
Expansion
d
E
R
B
GW
Original Simplex
Figure B.1: Basic operations in the downhill simplex method for two dimensions
space [108]
The initial triangle BGW
Let f(x, y) be the function that is to be minimized. To start, we are given
three vertices of a triangle: V k = (xk, yk), k = 1, 2, 3. The function f(x, y) is
then evaluated at each of the three points zk = f(xk, yk) for k = 1, 2, 3. The
subscripts are reordered so that z1 ≤ z2 ≤ z3. We use the notation [108],
B = (x1, y1), G = (x2, y2), and W = (x3, y3) (B.1)
to help remember that B is the best vertex, G is good (next to best), and W
is the worst vertex.
264
B.1. Nelder-Mead
Midpoint of the good side
The construction process uses the midpoint of the line segment joining B
and G. It is found by averaging the coordinates:
M =B +G
2=
(x1 + x2
2,y1 + y2
2
)(B.2)
Reflection using the point R
The function decreases as we move along the side of the triangle from W
to B, and it decreases as we move along the side from W to G. Hence it is
feasible that f(x, y) takes on smaller values at points that lie away from
W on the opposite side of the line between B and G. We choose a test point
R that is obtained by ‘Reflecting’ the triangle through the side BG. To
determine R, we first find the midpoint M of the side BG. Then draw
the line segment from W to M and call its length d. This last segment is
extended a distance d through M to locate the point R (see Fig. B.1). The
vector formula for R is [108]
R = M + (M −W ) = 2M −W (B.3)
Expansion using the point E
If the function value at R is smaller than the function value at W, then
we have moved in the correct direction toward the minimum. Perhaps the
minimum is just a bit further than the point R. So we extend the line
segment through M and R to the point E. This forms an expanded triangle
BGE. The point E is found by moving an additional distance d along the
line joining M and R (see Fig. B.1). If the function value at E is less than
the function value at R, then we have found a better vertex than R. The
vector formula for E is [108]
E = E + (R−M ) = 2R−M (B.4)
Contraction using the point C
If the function values at R and W are the same, another point must be
tested. Perhaps the function is smaller at M, but we cannot replace W
with M because we must have a triangle (for 2-D space). Consider the two
midpoints C1 and C2 of the line segments WM and MR, respectively
(see Fig. B.1). The point with the smaller function value is called C, and
the new triangle is BGC.
265
APPENDIX B. Optimization Techniques
Shrink toward B
If the function value at C is not less than the value at W, the point G and
W must be shrunk towards B (see Fig. B.1). The point G is replaced with
M, and W is replaced with S, which is the midpoint of the line segment
joining B with W [108].
Logical decisions for each step
A computationally efficient algorithm should perform function evaluation
only is needed. In each step, a new vertex is found, which replace W. As
soon as it is found, further investigation is not needed, and the iteration
step is completed. The logical details for two-dimensional cases are ex-
plained in Fig. B.2.
B.2 Genetic algorithms (GA)
The genetic algorithm is an optimization technique for solving both con-
strained and unconstrained optimization problems that is based on natu-
ral selection, the process that drives biological evolution [109]. The genetic
algorithm repeatedly modifies a population of individual solutions. At each
step, the genetic algorithm selects individuals at random from the current
population to be parents and uses them to produce the children for the
next generation. Over successive generations, the population evolves to-
ward an optimal solution. The genetic algorithm can be used to solve a
variety of optimization problems that are not well suited for standard opti-
mization algorithms, including problems in which the objective function is
discontinuous, no differentiable, stochastic, or highly nonlinear [109]. The
genetic algorithm uses three main types of rules at each step to create the
next generation from the current population [109]:
• Selection rules select the individuals, called parents, that contribute
to the population at the next generation.
• Crossover rules combine two parents to form children for the next
generation.
• Mutation rules apply random changes to individual parents to form
children.
The genetic algorithm differs from a classical, derivative-based, optimiza-
tion algorithm in two main ways, as summarized in Table B.1.
266
B.2
.G
en
etic
alg
orith
ms
(GA
)
If f (R) < f (G) es No
Case (i)either reflect or extend
Case (ii)either contract or shrink
If f (B) < f (R)Yes No
Replace W with R Compute E and f (E)
If f (E) < f (B)Yes No
Replace W with E Replace W with R
If f (R) < f (W)Yes No
Replace W with R
Compute C-(W+M)/2or C=(M+R)/2and f (C)
Compute E and f (E)
If f (C) < f (W)Yes No
Replace W with C Compute S and f (S)
Replace W with SReplace G with M
Start next cycle ...
Start next cycle ...
Start next cycle ... Start next cycle ...
Start next cycle ...
Start next cycle ...
Start a new cycle
Figure B.2: The logical decisions for the Nelder-Mead algorithm
267
APPENDIX B. Optimization Techniques
Table B.1: Comparison between the classical algorithms and the genetic algorithm
Classical Algorithm Genetic Algorithm
• It generates a single point at each itera-tion.
• The sequence of points approaches anoptimal solution.
• It selects the next point in the sequenceby a deterministic computation.
• In most cases, it find a relative (local)optimum that is closest to the startingpoint.
• It generates a population of points ateach iteration.
• The best point in the population ap-proaches an optimal solution.
• It selects the next population by compu-tation, which uses random number gen-erators.
• It can find the global optimum solutionwith a high probability.
B.2.1 Description of the genetic algorithm process
Genetic algorithms use a population of configurations, called individual, to
evolve over a number of generations. Each individual is represented by its
genetic material, called a chromosome and every variable in each individ-
ual is termed as the gene. For optimization purpose, the chromosome is
described by the design variables.
The process starts with an initial population of n individuals. The perfor-
mance of each individual is then evaluated in regard to the fitness function
and the handling of constraints (if some are considered) [28]. A selection
is done in the population to identify valuable parents. Higher is the per-
formance of an individual, higher is its probability to become parent [28].
Two parents are match randomly to exchange their genetic materials to
form the offspring for the next generation. This exchange process is called
crossover. If this process does not happen, the parents are directly trans-
ferred to the next generation meaning the cloning of these individuals. Af-
ter the crossover operator and before forming the next generation, all the
individuals are forced to undergo a mutation process [28]. The evolution
procedure is repeated until the population converges to a certain level or
simply if the maximum number of generation is reached, Fig. B.3.
B.2.2 Genetic operators
Population
Population type specifies the type of the input to the fitness function [109].
There are two options in Matlab, double vector and Bit string (binary chro-
mosome). The double vector option has been used in this study. A double
vector chromosome is simply a row vector of double values. These values
can be thought of as genes. Thus, 7 genes/chromosome means a double
268
B.2. Genetic algorithms (GA)
Start
Create initial random population
Evaluate fitness for each individual
Store best individuals (chromosomes)
Creating the mating pool (selection of parents)
Create the next generation by applying crossover & mutation
(new generation: offspring)
Optimal solution found?Stooping criteria met?
Yes
Stop
Evaluate fitness for each individual
No
Figure B.3: Flow chart for the genetic algorithms process
vector of 7 elements [27]. The population size n specifies how many indi-
viduals there are in each generation.
Fitness scaling
The scaling function converts raw fitness scores returned by the fitness
function to values in a range that is suitable for the selection function.
Scaling function specifies the function that performs the scaling. One can
choose from the following functions [109]: Rank, proportional, Top or Shift
linear. The rank option has been used in this study. Rank scales the raw
scores based on the rank of each individual, rather than its score. The
rank of an individual is its position in the sorted scores. The rank of the
fittest individual is 1, the next fittest is 2, and so on. Rank fitness scaling
removes the effect of the spread of the raw scores.
269
APPENDIX B. Optimization Techniques
Selection
Different types of selections are implanted in the optimization toolbox in
Matlab 2010a [109], but only the tournament selection is used in this study
(it is the only available selection type for multi-objective optimization in
the Matlab toolbox). The tournament selection randomly identifies some
competitors from the population to compete against each other. The one
with the highest performance win a parent status. The tournament selec-
tion permits to control the selective pressure put on the population. The
population diversity is adjusted by modification of the competitor num-
ber. Greater competitor numbers in the tournament increase the chances
to focus the search over the best individuals meaning a greater selective
pressure. On the other hand, with only two competitors, the possibility of
becoming a parent remains open to a larger band of the population mean-
ing a lesser selective pressure [28]. In other words, the tournament size
can take values between 1 and n (population size). Larger values give
more chances to the best samples to be selected and to create offsprings.
It favors a rapid, although perhaps premature, convergence to a local op-
timum. Very small values result in a more random selection of parents
[173]. The default value of tournament size of four has been used.
Reproduction
Reproduction options determine how the genetic algorithm creates chil-
dren at each new generation. Elite count specifies the number of individ-
uals who are guaranteed to survive to the next generation. the Matlab
manual [109] suggested to use Elite count to be two (a positive integer
less than or equal to population size). Crossover fraction specifies the frac-
tion of the next generation that crossover produces. Mutation produces
the remaining individuals in the next generation. Matlab manual [109]
suggested to use a crossover fraction of 0.8 (a fraction between 0 and 1).
Crossover
Crossover combines two individuals, or parents, to form a new individual,
or child (offspring), for the next generation [109]. the Matlab optimiza-
tion toolbox offers six functions: Scattered, single point, two point, inter-
mediate, heuristic and arithmetic. The intermediate function has been
used in this study. Intermediate (also, called weighted crossover [28]) cre-
ates children by a random weighted average of the parents. Intermedi-
ate crossover is controlled by a single parameter ratio: child1 = parent1+
rand*Ratio*(parent2 - parent1). If the ratio is in the range [0,1], the chil-
dren produced are within the hypercube defined by the parents locations
270
B.2. Genetic algorithms (GA)
at the opposite vertices. If Ratio is in a larger range, say 1.1, children can
be generated outside the hypercube. Ratio can be a scalar or a vector of
length equals the number of variables. If the ratio is a scalar, all the chil-
dren lie on the line between the parents. If the ratio is a vector, children
can be any point within the hypercube.
Mutation
Mutation functions make small random changes in the individuals in the
population, which provide genetic diversity and enable the genetic algo-
rithm to search for a broader space. To specify the function that performs
the mutation in the mutation function field, one can choose from the follow-
ing functions: constraint dependent default, gaussian, uniform and adap-
tive feasible. The constraint dependent default chooses: Gaussian if there
are no constraints or adaptive feasible otherwise. 1) Gaussian adds a ran-
dom number to each vector entry of an individual. This random number
is taken from a gaussian distribution centered on zero. The standard de-
viation of this distribution can be controlled with two parameters. The
scale parameter (default value of 1.0) determines the standard deviation
at the first generation. The shrink parameter controls how the standard
deviation shrinks as generations go by. If the shrink parameter is 0, the
standard deviation is constant. If the Shrink parameter is 1 (the default
value), the standard deviation shrinks to 0 linearly as the last generation
is reached. 2) Uniform is a two-step process. First, the algorithm selects
a fraction of the vector entries of an individual for mutation, where each
entry has the same probability as the mutation rate of being mutated. In
the second step, the algorithm replaces each selected entry by a random
number selected uniformly from the range for that entry. 3) Adaptive fea-
sible randomly generates directions that are adaptive with respect to the
last successful or unsuccessful generation. A step length is chosen along
each direction so that linear constraints and bounds are satisfied. The
constraint dependent default has been used in this study.
Stopping criteria
The stopping criteria determines what causes the algorithm to terminate.
Matlab optimization toolbox has the following stopping criteria, the default
values are given in parentheses.
• Generations specific the maximum number of iterations the genetic
algorithm performs (100).
• Time limit specifies the maximum time in seconds the genetic algo-
rithm runs before stopping (∞).
271
APPENDIX B. Optimization Techniques
• Fitness limit - If the best fitness value is less than or equal to the
value of fitness limit, the algorithm stops (−∞).
• Stall generations - If the weighted average change in the fitness func-
tion value over stall generations is less than Function tolerance, the
algorithm stops (50).
• Stall time limit - If there is no improvement in the best fitness value
for an interval of time in seconds specified by Stall time limit, the
algorithm stops (∞).
• Function tolerance - If the cumulative change in the fitness func-
tion value over stall generations is less than Function tolerance, the
algorithm stops (1E-6).
B.3 Multi-objective optimization
In case of cyclone separator geometry optimization for minimum Euler
number and minimum cut-off diameter, the objectives are conflicting with
each other. There is no best solution for which all objectives are optimal si-
multaneously [181]. The increase of one objective will lead to the decrease
of the other objective. Then, there should be a set of solutions, the so-called
Pareto optimal set or Pareto front, in which one solution cannot be domi-
nated by any other member of this set. The definition of domination is as
given in Wang et al. [181]
For minimal problem, a solution a ∈ X dominates a solution
b ∈ X(a ≻ b) if and only if it is superior or equal in all objectives
and at least superior in one objective. This can be expressed as
follows [181]:
a ≻ b, if
∀i ∈ 1, 2, . . . ,m : fi(a) ≤ fi(b)∧∃j ∈ 1, 2, . . . ,m : fj(a) < fj(b)
where m is the number of objective functions.
Each solution of Pareto optimal set is called a non-inferior solution, which
is corresponding to one point on the Pareto front. A general example of the
Pareto front with two objectives is illustrated in Fig. B.4. In this example,
the Pareto front is composed of six points, which are indifferent to each
other (denoted by filled circles • in Fig. B.4). While, points with hollow
circles are not belonging to the Pareto front, since they are dominated by
the Pareto front points.
272
B.3. Multi-objective optimization
Figure B.4: Pareto front for two objective functions.
Nondominated sorting genetic algorithm (NSGA-II)
Recently, a number of multi-objective genetic algorithms based on the Pareto
optimal concept has been proposed. The well known nondominated sort-
ing genetic algorithm (NSGA-II) proposed by Deb et al. [34] is one of the
most widely used multi-objective genetic algorithm since it provides ex-
cellent results as compared with other multi-objective genetic algorithms
proposed [25]. A brief description of processes of NSGA-II is presented
below, Fig. B.5.
Initially, a parent population of size n is generated randomly. All individ-
uals in this population are sorted into different front levels based on the
domination of pair comparison. Each front level is assigned a fitness (or
a rank) which equals its non-domination level. Level 1 is the top level in
which the individual is dominated by none of the other individuals; level 2
is the secondary level in which the individual is dominated by some indi-
viduals only in level 1, and so on. In the same front level, the location of the
finite number of solutions is expected to be distributed uniformly. In other
words, a large diversity of the individuals can prevent the results from
sticking into a local optimum. Therefore, another feature, called crowding
distance, is adopted to evaluate the local aggregation of individuals. The
273
APPENDIX B. Optimization Techniques
definition of crowding distance is [34, 181]:
Cj =
Nobj∑
i=1
F j+1i − F j−1
i
(F ji )max − (F j
i )min
(B.5)
where Cj is the crowding distance of point j on the Pareto front, (F ji ) is the
value of the fitness function i at point j (i = 1, 2 for two objective functions).
For boundary points, the crowding distance is set to the maximum value
of the system in order to ensure that these points can survive to the next
generation. In the same rank level, individuals who have larger crowding
distance also have more opportunities to be selected [182].
Simulated binary crossover
In NSGA-II, a simulated binary crossover operator is used, which simu-
lates the working principle of the single point crossover operator on binary
strings [182]. Let x1,ti and x2,t
i denote two parent individuals for ith selec-
tion in generation t. The procedure of computing the children individuals
x1,t+1i and x2,t+1
i from parent individuals x1,ti and x2,t
i is described below. A
spread factor βi is defined as the ratio of the absolute difference in children
values to that of parent values:
βi =
∣∣∣∣∣x2,t+1i − x1,t+1
i
x2,ti − x1,t
i
∣∣∣∣∣ (B.6)
Firstly, a random number ui ∈ [0, 1] is generated, whereafter, from a speci-
fied probability distribution function, the ordinate βqi is found so that the
area under the probability curve from 0 to βqi is equal to the chosen ran-
dom number ui. The probability distribution used to create a child indi-
vidual is derived to have a similar search power as that in a single-point
crossover in binary-coded GAs and is given as follows [182]:
P (βi) =
0.5(η + 1)βiη, if βi ≤ 1
0.5(η + 1)/βiη+2, otherwise
(B.7)
where, η is the distribution index which can be any nonnegative real num-
ber. A lager value of η gives a higher probability for creating near parent
individuals and a small value of η allows distant individuals to be selected
as children individuals. Using Eq.B.7, the βqi can be calculated as follows
[182]:
βqi =
(2ui)1
η+1 , if ui ≤ 0.5(
12(1−ui)
) 1η+1
, otherwise(B.8)
274
B.3. Multi-objective optimization
Thereafter, the children individuals can be obtained by [182]:
x1,t+1i = 0.5
[(1 + βqi)x
1,ti + (1− βqi)x
2,ti
]
x2,t+1i = 0.5
[(1 + βqi)x
1,ti + (1− βqi)x
2,ti
] (B.9)
Note that two children individuals are symmetric about the parent indi-
viduals. This is deliberately used to avoid any bias towards any particular
parent individual in a single crossover operation.
Mutation Operator
Let xk be the component of an individual xi, which is going to be mu-
tated. xuk and xl
k stand for the maximum and minimum value of this com-
ponent in all individuals, respectively. The mutated individual yk can be
calculated as follows [182]:
yk = xk + δq(xuk − xl
k
)(B.10)
Here δq is a mutation parameter, which stands for:
δq =
(2r + (1− 2r)(1− δ1)
ηm+1)1/(ηm+1) − 1, if r ≤ 0.5
1−(2r(1 − r) + 2(r − 0.5)(1− δ2)
ηm+1)1/(ηm+1)
otherwise
(B.11)
where ηm is the mutation index which is set to 20 (1- crossover fraction) in
general, and r is a random number. The intermediate variables δ1 and δ2stand for:
δ1 =xk − xl
k
xuk − xl
k
(B.12)
δ2 =xuk − xk
xuk − xl
k
(B.13)
275
APPENDIX B. Optimization Techniques
Figure B.5: Flow chart for the NSGA-II process
276
List of Publications
Journal Articles
Published
1. K. Elsayed, C. Lacor. Modeling and Pareto Optimization of Gas Cy-clone Separator Performance Using RBF Type Artificial Neural Net-works and Genetic Algorithms. Powder Technology. In Press, Ac-cepted Manuscript, Available online 17 October 2011doi: http://dx.doi.org/10.1016/j.powtec.2011.10.015
2. K. Elsayed, C. Lacor. Numerical and Empirical Modeling of the FlowField and Performance in Cyclones of Different Cone-Tip Diameters.Computers & Fluids, Vol. 51, No. 1, pp. 48 - 59, 2011.doi: http://dx.doi.org/10.1016/j.compfluid.2011.07.010
3. K. Elsayed, C. Lacor. Modeling, Analysis and Optimization of Air-cyclones Using Artificial Neural Network, Response Surface Method-ology and CFD Simulation Approaches. Powder Technology, Vol.212, No. 1, pp. 115 - 133, 2011.doi: http://dx.doi.org/10.1016/j.powtec.2011.05.002
4. K. Elsayed, C. Lacor. The Effect of Cyclone Inlet Dimensions onthe Flow Pattern and Performance. Applied Mathematical Mod-elling, Vol. 35, No. 4, pp. 1952 - 1968, 2011.doi: http://dx.doi.org/10.1016/j.apm.2010.11.007
5. K. Elsayed, C. Lacor. Optimization of the Cyclone Separator Geom-etry for Minimum Pressure Drop Using Mathematical Models andCFD Simulations.Chemical Engineering Science, Vol. 65, No. 22,pp. 6048 - 6058, 2010.doi: http://dx.doi.org/10.1016/j.ces.2010.08.042 (listed inthe Top 25 Hottest Articles Chemical Engineering > Chemical Engi-neering Science, October to December 2010).
Accepted with revision
1. K. Elsayed, C. Lacor. The Effect of Cyclone Vortex Finder Dimen-sions on the Flow Pattern and Performance using LES.Computers & Fluids, 2011.
277
With reviewers
1. K. Elsayed, C. Lacor. CFD Modeling and Multi-Objective Optimiza-tion of Cyclone Geometry Using Desirability Function, Artificial Neu-ral Networks and Genetic Algorithms.Applied Mathematical Modelling. Submitted, 2011.
2. K. Elsayed, C. Lacor. A CFD Study of the Effect of the Dust OutletGeometry on the Performance and Hydrodynamics of Gas Cyclones.Computers & Fluids. Submitted, 2011.
3. K. Elsayed, C. Lacor. Numerical Study of the Effect of Cyclone Cone& Barrel Height On The Flow Pattern And Performance.Applied Mathematical Modelling. Submitted, 2011.
International conference proceedings
1. K. Elsayed, C. Lacor. Single and Multi-Objective Optimization of theCyclone Separator Geometry Using Artificial Neural Network andGenetic Algorithm, EUROGEN 2011, Evolutionary and Determinis-tic Methods for Design, Optimization and Control with Applicationsto Industrial and Societal Problems, Capua, Italy, 14-16 September2011.
2. K. Elsayed, C. Lacor. The Effect of Dust Outlet Shape on the FlowPattern and Performance of Cyclone Separators, 10th InternationalSymposium on Experimental and Computational Aerothermodynam-ics of Internal Flows (ISAIF10), Brussel, Belgium, 4-7 July 2011.
3. K. Elsayed, C. Lacor. The Effect of the Cyclone Separator ConeHeight on the Performance Using Artificial Neural Network Modeland CFD Simulations, 10th International Symposium on Experimen-tal and Computational Aerothermodynamics of Internal Flows (ISAIF10),Brussel, Belgium, 4-7 July 2011.
4. K. Elsayed, C. Lacor. A CFD Study of the Effect of Cyclone BarrelHeight on Its Performance Parameters, 8th International Conferenceon Computational Fluid Dynamics in the Oil & Gas, Metallurgicaland Process Industries, Trondheim, Norway, 21-23 June 2011.
5. K. Elsayed, C. Lacor. Numerical Investigations of the Effect of Dif-ferent Dust Outlet Designs on the Cyclone Performance and FlowPattern, 14th International Conference on Aerospace Sciences andAviation Technology (ASAT-14), Cairo, Egypt, 24-26 May 2011.
6. K. Elsayed, C. Lacor. Optimization of the Cyclone Separator Geom-etry for Minimum Pressure Drop Based on Artificial Neural NetworkModel And CFD Simulation, ECCOMAS thematic conference, CFD &Optimization, Antalya, Turkey, 23-25 May 2011.
278
7. K. Elsayed, C. Lacor. Multi-Objective Optimization of Gas CycloneBased On CFD Simulation, ECCOMAS thematic conference, CFD &Optimization, Antalya, Turkey, 23-25 May 2011.
8. K. Elsayed, C. Lacor. The effect of cyclone height on the flow patternand performance using LES, Tenth International Congress of FluidDynamics (ICFD10), ASME, Egypt, ICFD10-EG-3003, Ain Soukhna,Red Sea, Egypt, 16-19 December 2010.
9. K. Elsayed, C. Lacor. The effect of cyclone inlet width on the flowpattern and performance, Tenth International Congress of Fluid Dy-namics (ICFD10), ASME, Egypt, ICFD10-EG-3085, Ain Soukhna,Red Sea, Egypt, 16-19 December 2010.
10. K. Elsayed, C. Lacor. Numerical study on the effect of cyclone in-let height on the flow pattern and performance, Tenth InternationalCongress of Fluid Dynamics (ICFD10), ASME, Egypt, ICFD10-EG-3068, Ain Soukhna, Red Sea, Egypt, 16-19 December 2010.
11. K. Elsayed, C. Lacor. The effect of cyclone dustbin on the flow pat-tern and performance, Tenth International Congress of Fluid Dynam-ics (ICFD10), ASME, Egypt, ICFD10-EG-3092, Ain Soukhna, RedSea, Egypt, 16-19 December 2010.
12. K. Elsayed, C. Lacor. Application of Response Surface Methodologyfor Modeling and Optimization of the Cyclone Separator for Mini-mum Pressure Drop, Fifth European Conference on ComputationalFluid Dynamics (ECCOMAS CFD10), Lisbon, Portugal, 14-17 June2010.
13. K. Elsayed, C. Lacor. The Effect of Vortex Finder Diameter on Cy-clone Separator Performance and Flow Field, Fifth European Confer-ence on Computational Fluid Dynamics (ECCOMAS CFD10), Lisbon,Portugal, 14-17 June 2010.
14. V. Agnihotri, K. Elsayed, C. Lacor, S. Verbanc. Numerical Studyof Particle Deposition in the Human Upper Airways With Emphasison Hot Spot Formation and Comparison Of LES and RANS Models,Fifth European Conference on Computational Fluid Dynamics (EC-COMAS CFD10), Lisbon, Portugal, 14-17 June 2010.
15. K. Elsayed, C. Lacor. Modeling of the Gas and Particle Flow in theCyclone Separator Using LES, RANS and Mathematical Models, 14th
International Conference on Applied Mechanics and Mechanical En-gineering (AMME-14), Military Technical College, Cairo, Egypt, 25-27 May 2010.
16. K. Elsayed, C. Lacor. Optimization of the Cyclone Separator Geom-etry Based On CFD Simulation, ERCOFTAC day, Vrije UniversiteitBrussel, Brussels, Belgium, 3rd December 2009.
279
17. K. Elsayed, C. Lacor. A CFD Study of the Effects of Cone Dimen-sions on the Flow Field of Cyclone Separators Using LES, 13th In-ternational Conference on Aerospace Sciences & Aviation Technology(ASAT-13), Military Technical College, Cairo, Egypt, 26-28 May 2009.
18. K. Elsayed, C. Lacor. Investigation of the Geometrical ParametersEffects on the Performance and the Flow-Field of Cyclone Separa-tors Using Mathematical Models and Large Eddy Simulation, 13th
International Conference on Aerospace Sciences & Aviation Technol-ogy (ASAT-13), Military Technical College, Cairo, Egypt, 26-28 May2009.
National Conference proceedings
1. K. Elsayed, C. Lacor. Study Of the Effects of Geometrical Parame-
ters on the Performance of Cyclone Separators, 8th National Congress
on Theoretical and Applied Mechanics (NCTAM2009), Vrije Univer-
siteit Brussel, Brussels, Belgium, 28-29 May 2009.
2. K. Elsayed, C. Lacor. Effects of Geometry Parameters on the Flow
Field of Cyclone Separator, Poster day, the GRAduate School in ME-
CHanics (GRASMECH 2008), Royal Military Academy, Brussels, Bel-
gium, 3rd October 2008.
Technical reports
1. K. Elsayed, S. Jayaraju, C. Lacor. Wood Pellet Transport in the
Biomass Burner of the Rodenhuize Power Plant, Research report, La-
borelec Company, Belgium, April 2009.
2. K. Elsayed, C. Lacor. The State of the Art for Flow in Cyclone Sepa-
rator, Internal report, Vrije Universiteit Brussel, Brussels, Belgium,
April 2008.
280
Bibliography
[1] R. M. Alexander. Fundamentals of cyclone design and operation. InProceedings of the Australian Institute of Mineral and Metallurgy,number 152, pages 203–228, 1949.
[2] M. S. M. Ali, C. J. Doolan, and V. Wheatley. Grid convergence studyfor a two-dimensional simulation of flow around a square cylinderat a low reynolds number. In Seventh International Conference onCFD in the Minerals and Process Industries, CSIRO, Melbourne,Australia, 9-11 December 2009.
[3] S. Altmeyer, V. Mathieu, S. Jullemier, P. Contal, N. Midoux, S. Rode,and J. P. Leclerc. Comparison of different models of cyclone predic-tion performance for various operating conditions using a generalsoftware. Chemical Engineering and Processing, 43:511–522, 2004.
[4] N. Amoura, H. Kebir, S. Rechak, and J.M. Roelandt. Axisymmetricand two-dimensional crack identification using boundary elementsand coupled quasi-random downhill simplex algorithms. Engineer-ing Analysis with Boundary Elements, 34(6):611–618, 2010.
[5] A. Avci and I. Karagoz. Theoretical investigation of pressure lossesin cyclone separators. International Communications in Heat andMass Transfer, 28(1):107–117, 1 2001.
[6] Mehdi Azadi, Mohsen Azadi, and Ali Mohebbi. A CFD study of theeffect of cyclone size on its performance parameters. Journal of Haz-ardous Materials, 182(1-3):835 – 841, 2010.
[7] M. I. Bakari and F. Hamdullahpur. Optimization of the overall per-formance of cyclone separators. In 2nd Trabzon International En-ergy and Environment Symposium, 26-29 July, Turkey, 1998.
[8] M. Balcilar, A.S. Dalkilic, and S. Wongwises. Artificial neural net-work techniques for the determination of condensation heat transfercharacteristics during downward annular flow of R134a inside a ver-tical smooth tube. International Communications in Heat and MassTransfer, 38(1):75 – 84, 2011.
281
[9] W. Barth. Design and layout of the cyclone separator on the ba-sis of new investigations. Brennstow-Waerme-Kraft (BWK), 8(4):1–9,1956.
[10] A. C. Benim, K. Ozkan, M. Cagan, and D. Gunes. Computationalinvestigation of turbulent jet impinging onto rotating disk. In-ternational Journal of Numerical Methods for Heat & Fluid Flow,17(3):284 – 301, 2007.
[11] S. Bernardo, M. Mori, A.P. Peres, and R.P. Dionisio. 3-D computa-tional fluid dynamics for gas and gas-particle flows in a cyclone withdifferent inlet section angles. Powder Technology, 162(3):190 – 200,2006.
[12] J. L. Bernon, V. Boudousq, J. F. Rohmer, M. Fourcade, M. Zanca,M. Rossi, and D. Mariano-Goulart. A comparative study of pow-ell’s and downhill simplex algorithms for a fast multimodal surfacematching in brain imaging. Computerized Medical Imaging andGraphics, 25(4):287–297, 2001.
[13] G. E. P. Box and K. B. Wilson. On the experimental attainment ofoptimum conditions. Journal of the Royal Statistical Society, 13:1–45, 1951.
[14] F. Boysan, W.H. Ayer, and J. A. Swithenbank. Fundamentalmathematical-modelling approach to cyclone design. Transaction ofInstitute Chemical Engineers, 60:222–230, 1982.
[15] F. Boysan, B. C. R. Ewan, J. Swithenbank, and W. H. Ayers. Exper-imental and theoretical studies of cyclone separator aerodynamics.IChemE Symp Series, 69:305 – 320, 1983.
[16] D. S. Broomhead and D. Lowe. Multivariable functional interpola-tion and adaptive networks. Complex Systems, 2:321–355, 1988.
[17] H. S. Bryant, R.W. Silverman, and F. A. Zenz. How dust in gas affectscyclone pressure drop. Hydrocarbon Processing, 62(4):87–90, 1983.
[18] John A. Bullinaria. Introduction to neural networks, 2004.http://www.cs.bham.ac.uk/ ˜ jxb/NN/l10.pdf .
[19] H. Buttner. Size separation of particles from aerosol samples usingimpactors and cyclones. Particle & Particle Systems Characteriza-tion, 5(2):87–93, 2004.
[20] W. M. Carlyle, D. C. Montgomery, and G. C. Runger. Optimizationproblems and methods in quality control and improvement. Journalof Quality Technology, 32(1):1–17, 2000.
[21] J. Casal and J. M. Martinez-Benet. A better way to calculate cyclonepressure drop. Chemical Engineering, 90(2):99–100, 1983.
282
[22] J. Chen and M. Shi. A universal model to calculate cyclone pressuredrop. Powder Technology, 171(3):184 – 191, 2007.
[23] T. G. Chuah, J. Gimbun, and T. S.Y. Choong. A CFD study of theeffect of cone dimensions on sampling aerocyclones performance andhydrodynamics. Powder Technology, 162:126 – 132, 2006.
[24] T. J. Chung. Computational Fluid Dynamics. Cambridge UniversityPress, 1st edition, 2002.
[25] C. A. C. Coello. Twenty years of evolutionary multi-objective opti-mization: a historical view of the field. IEEE Computational Intelli-gence Magazine, 1:28–36, 2006.
[26] C. A. C. Coello, G. B. Lamont, and D. A. Van Veldhuizen. Evolution-ary Algorithms for Solving Multi-Objective Problems. Springer, 2ndedition, 2007.
[27] D. Cooley. Using optimtool GA in Matlab, 2011.digital.cs.usu.edu/ ˜ cooley/CS6665_10_Optimtool_GA.pptx .
[28] G. Corriveau, R. Guilbault, and A. Tahan. Genetic algorithms andfinite element coupling for mechanical optimization. Advances inEngineering Software, 41(3):422 – 426, 2010.
[29] C. Cortes and A. Gil. Modeling the gas and particle flow insidecyclone separators. Progress in Energy and Combustion Science,33(5):409 – 452, 2007.
[30] J. S. Cowpe, J. S. Astin, R. D. Pilkington, and A. E. Hill. Applicationof response surface methodology to laser-induced breakdown spec-troscopy: influences of hardware configuration. Spectrochimica ActaPart B, 62:1335 – 1342, 2007.
[31] J. S. Curtis and B. V. Wachem. Modeling particle-laden flows: Aresearch outlook. Wiley InterScience, 50-11, 2004.
[32] L. Davidson. Lecture notes. Dept. of Thermo and Fluid Dynamics,Chalmers University of Technology, Goteborg, Sweden, 2000.
[33] J. W. Deardorff. The use of subgrid transport equations in a three-dimensional model of atmospheric turbulence. Journal of Fluids En-gineering, 95:429–438, 1973.
[34] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan. A fast and eli-tist multiobjective genetic algorithm: NSGA-II. IEEE TransactionEvolitionery Computation, 6:182–197, 2002.
[35] W. Deconinck. Design and application of discrete explicit filters forlarge eddy simulation of compressible turbulent flows. Master’s the-sis, University of Toronto, 2008.
283
[36] J. J. Derksen. Separation performance predictions of a stairmandhigh-efficiency cyclone. AIChE Journal, 49(6):1359–1371, 2003.
[37] J. J. Derksen, S. Sundaresan, and H. E. A. van den Akker. Simula-tion of mass-loading effects in gas–solid cyclone separators. PowderTechnology, 163:59–68, 2006.
[38] J. J. Derksen and H. E. A. Van den Akker. Simulation of vortex coreprecession in a reverse-flow cyclone. AIChE Journal, 46(7):1317 –1331, 2000.
[39] J. J. Derksen, H. E. A. van den Akker, and S. Sundaresan. Two-way coupled large-eddy simulations of the gas-solid flow in cycloneseparators. AIChE Journal, 54(4):872–885, April 2008.
[40] D. B. Dias, M. Mori, and W. P. Martignoni. Boundary condition ef-fects in CFD cyclone simulations. In 8th World Congress of ChemicalEngineering (WCCE8), Montreal, August 2009.
[41] K. K. J. R. Dinesh and M. P. Kirkpatrick. Study of jet precession,recirculation and vortex breakdown in turbulent swirling jets usingLES. Computers & Fluids, 38:1232–1242, 2009.
[42] J. Dirgo. Relationship between cyclone dimensions and performance.PhD thesis, Harvard University, USA, 1988.
[43] J. Dirgo and D. Leith. Cyclone collection efficiency: Comparison ofexperimental results with theoretical predictions. Aerosol Scienceand Technology, 4:401–415, 1985.
[44] S. E. Elghobashi. On predicting particle-laden turbulent flows. Ap-plied Scientific Research, 52:309–329, 1994.
[45] K. Elsayed and C. Lacor. Modeling and pareto op-timization of gas cyclone separator performance usingRBF type artificial neural networks and genetic algo-rithms. Powder Technology, Accepted 9 October, 2011.http://dx.doi.org/10.1016/j.powtec.2011.10.015 .
[46] K. Elsayed and C. Lacor. Investigation of the geometrical param-eters effects on the performance and the flow-field of cyclone sep-arators using mathematical models and large eddy simulation. In13th Aerospace Sciences & Aviation Technology (ASAT-13), MilitaryTechnical College, Cairo, Egypt, 26-28 May 2009.
[47] K. Elsayed and C. Lacor. The effect of cyclone dustbin on the flowpattern and performance. In Tenth International Congress of FluidDynamics (ICFD10), ASME, Egypt, ICFD10-EG-3092, 2010, AinSoukhna, Red Sea, Egypt, 16-19 December 2010.
284
[48] K. Elsayed and C. Lacor. The effect of cyclone height on the flow pat-tern and performance using LES. In Tenth International Congressof Fluid Dynamics (ICFD10), ASME, Egypt, ICFD10-EG-3003, AinSoukhna, Red Sea, Egypt, 16-19 December 2010.
[49] K. Elsayed and C. Lacor. The effect of vortex finder diameter on cy-clone separator performance and flow field. In Fifth European Con-ference on Computational Fluid Dynamics (ECCOMAS CFD 2010),Lisbon, Portugal, 14-17 June 2010.
[50] K. Elsayed and C. Lacor. Optimization of the cyclone separator ge-ometry for minimum pressure drop using mathematical models andCFD simulations. Chemical Engineering Science, 65(22):6048–6058,2010.
[51] K. Elsayed and C. Lacor. A CFD study of the effect of cyclone barrelheight on its performance parameters. In 8th International Confer-ence on Computational Fluid Dynamics in the Oil & Gas, Metallur-gical and Process Industries, Trondheim, Norway, 21-23 June 2011.
[52] K. Elsayed and C. Lacor. The effect of cyclone inlet dimensions onthe flow pattern and performance. Applied Mathematical Modelling,35(4):1952–1968, 2011.
[53] K. Elsayed and C. Lacor. Modeling, analysis and optimizationof aircyclones using artificial neural network, response surfacemethodology and CFD simulation approaches. Powder Technology,212(1):115–133, 2011.
[54] K. Elsayed and C. Lacor. Multi-objective optimization of gas cyclonebased on CFD simulation. In ECCOMAS thematic conference, CFD& Optimization, Antalya, Turkey, 23-25 May 2011.
[55] K. Elsayed and C. Lacor. Numerical modeling of the flow field andperformance in cyclones of different cone-tip diameters. Computers& Fluids, 51(1):48–59, 2011.
[56] K. Elsayed and C. Lacor. Single and multi-objective optimizationof the cyclone separator geometry using artificial neural networkand genetic algorithm. In EUROGEN 2011, Evolutionary and De-terministic Methods for Design, Optimization and Control with Ap-plications to Industrial and Societal Problems, Capua, Italy, 14-16September, 2011.
[57] Ferit Ficici, Vedat Ari, and Murat Kapsiz. The effects of vortex finderon the pressure drop in cyclone separators. International Journal ofthe Physical Sciences, 5(6):804–813, 2010.
[58] M. W. First. Cyclone dust collector design. ASME Annual GeneralMeeting, Paper No. 49A127 (1949).
285
[59] Fluent. FLUENT 6.3 user’s guide. Fluent Incorporated, Lebanon,2006.
[60] M. Germano, U. Piomelli, P. Moin, and W. H. Cabot. A dynamicsubgrid-scale eddy viscosity model. Physics of Fluids A, 3(7):1760–1765, 1991.
[61] M. Gfrerer and E. Lankmayr. Screening, optimization and valida-tion of microwave-assisted extraction for the determination of per-sistent organochlorine pesticides. Analytica Chimica Acta, 533:203–211, 2005.
[62] J. Gimbun, T. G. Chuah, T. S. Y. Choong, and A. Fakhru’l-Razi. ACFD study on the prediction of cyclone collection efficiency. Inter-national Journal for Computational Methods in Engineering Scienceand Mechanics, 6(3):161 – 168, 2005.
[63] J. Gimbun, T.G. Chuah, T.S. Choong, and Y.A. Fakhru’l-Razi. Predic-tion of the effects of cone tip diameter on the cyclone performance.Aerosol Science and Technology, 36:1056–1065, 2005.
[64] J. Gimbun, T.G. Chuah, A. Fakhru’l-Razi, and T. S. Y. Choong. Theinfluence of temperature and inlet velocity on cyclone pressure drop:a CFD study. Chemical Engineering & Processing, 44(1):7–12, 2005.
[65] F. Girosi, T. Poggio, and A. Caprile. Extensions of a theory of net-works for approximations and learning: outliers and negative ex-amples. In NIPS-3: Proceedings of the 1990 conference on Advancesin neural information processing systems 3, pages 750–756, Denver,Colorado, United States, 1990.
[66] A. Gorton-Hulgerth, J. Woisetschlager, Graham Wigley, andG. Staudinger. Investigation of the flow field in the upper part ofa cyclone with laser and phase doppler anemometry. Particle & Par-ticle Systems Characterization, 17:21–27, 2000.
[67] A. D. Gosman and E. Ioannides. Aspects of computer simulation ofliquid-fuelled combustor. AIAA Journal, 81-0323, 1981.
[68] W. D. Griffiths and F. Boysan. Computational fluid dynamics (CFD)and empirical modelling of the performance of a number of cyclonesamplers. Journal of Aerosol Science, 27(2):281–304, 1996.
[69] G. Gronald and J. J. Derksen. Simulating turbulent swirling flow ina gas cyclone: A comparison of various modeling approaches. PowderTechnology, 205(1-3):160 – 171, 2011.
[70] P. O. Haaland. Experimental design in biotechnology. Marcel Dekker,New York, 1989.
286
[71] E. C. Jr. Harrington. The desirability function. Industrial QualityControl, 21:494–498, 1965.
[72] Eric Hartman, James D. Keeler, and Jacek M. Kowalski. Layeredneural networks with gaussian hidden units as universal approxi-mations. Neural Computation, 2(2):210–215, 1990.
[73] J. O. Hinze. Turbulence. McGraw-Hill, New York, 2nd ed. edition,1975.
[74] A. J. Hoekstra. Gas flow field and collection efficiency of cyclone sep-arators. PhD thesis, Technical University Delft, 2000.
[75] A. J. Hoekstra, J. J. Derksen, and H. E. A. Van Den Akker. Anexperimental and numerical study of turbulent swirling flow in gascyclones. Chemical Engineering Science, 54:2055–2065, 1999.
[76] A. C. Hoffmann, M. de Groot, W. Peng, H. W. A. Dries, and J. Kater.Advantages and risks in increasing cyclone separator length. AIChEJournal, 47(11):2452–2460, 2001.
[77] A. C. Hoffmann and L. E. Stein. Gas cyclones and swirl tubes: Prin-ciple, Design and Operation. Springer, 2nd edition, 2008.
[78] A.C. Hoffmann, M. De Groot, and A. Hospers. The effect of thedust collection system on the flow pattern and separation efficiencyof a gas cyclone. The Canadian Journal of Chemical Engineering,74:464–470, August 1996.
[79] A. Horvath, C. Jordan, and M. Harasek. Influence of vortex-finderdiameter on axial gas flow in simple cyclone. Chemical Product andProcess Modeling, 3(1), 2008.
[80] Y. H. Hu and J. Hwang, editors. Handbook of Neural Network SignalProcessing. CRC Press, Washington, D.C., 2002.
[81] S. Huang, Q. S. Li, and S. Xu. Numerical evaluation of wind effectson a tall steel building by CFD. Journal of Constructional Steel Re-search, 63(5):612 – 627, 2007.
[82] S. H. Huang, Q. S. Li, and J. R. Wu. A general inflow turbulencegenerator for large eddy simulation. Journal of Wind Engineeringand Industrial Aerodynamics, 98(10-11):600 – 617, 2010.
[83] P. Hutchinson, G. F. Hewitt, and A. E. Dukler. Deposition of liquidor solid dispersions from turbulent gas streams: a stochastic model.Chemical Engineering Science, 26:419–439, 1971.
[84] D. L. Iozia and D. Leith. Effect of cyclone dimensions on gas flowpattern and collection efficiency. Aerosol Science and Technology,10(3):491–500, 1989.
287
[85] D. L. Iozia and D. Leith. The logistic function and cyclone fractionalefficiency. Aerosol Science and Technology, 12(3):598–606, 1990.
[86] S. T. Jayaraju. Study of the air flow and aerosol transport in thehuman upper airway using LES and DES methodology. PhD thesis,Vrije Universiteit Brussel, 2009.
[87] Zhang Jian and Jin You-hai. CFD analysis on collection performanceof swirl tube with different vortex finder geometries. Journal of Sys-tem Simulation, 21(4):1214–17, 2009.
[88] Viviana P. Jofre, Mariela V. Assof, Martn L. Fanzone, Hctor C.Goicoechea, Luis D. Martnez, and Mara F. Silva. Optimization ofultrasound assisted-emulsification-dispersive liquid-liquid microex-traction by experimental design methodologies for the determina-tion of sulfur compounds in wines by gas chromatography-massspectrometry. Analytica Chimica Acta, 683(1):126 – 135, 2010.
[89] W. P. Jones and M. Wille. Large-eddy simulation of a plane jet in across-flow. International Journal of Heat and Fluid Flow, 17-3:296–306, 1995.
[90] I. Karagoz and A. Avci. Modelling of the pressure drop in tangentialinlet cyclone separators. Aerosol Science and Technology, 39(9):857–865, 2005.
[91] F. Kaya and I. Karagoz. Performance analysis of numerical schemesin highly swirling turbulent flows in cyclones. Current Science,94(10):1273 – 1278, 2008.
[92] F. Kaya and I. Karagoz. Numerical investigation of performancecharacteristics of a cyclone prolonged with a dipleg. Chemical Engi-neering Journal, 151:39–45, 2009.
[93] L. C. Kenny and R. A. Gussman. A direct approach to the designof cyclones for aerosol-monitoring applications. Journal of AerosolScience, 31(12):1407–1420, 2000.
[94] Abolfazl Khalkhalia and Hamed Safikhania. Pareto based multi-objective optimization of a cyclone vortex finder using CFD,GMDH type neural networks and genetic algorithms. Engi-neering Optimization, , Available online: 05 Jul 2011, 2011.http://dx.doi.org/10.1080/0305215X.2011.564619 .
[95] J. C. Kim and K. W. Lee. Experimental study of particle collectionby small cyclones. Aerosol Science and Technology, 12:1003–1015,1990.
[96] J. Kiusalaas. Numerical Methods in Engineering With MATLAB.Cambridge University Press, Cambridge, UK, 2nd edition, 2010.
288
[97] C. Lacor. Lecture notes. Dept. Fluid Mechanics, Vrije UniversiteitBrussel, Brussels, Belgium, 2007.
[98] B. E. Launder, G. J. Reece, and W. Rodi. Progress in the developmentof a Reynolds-stress turbulence closure. Journal of Fluid Mechanics,68(03):537–566, 1975.
[99] D. Leith and W. Licht. The collection efficiency of cyclone type par-ticle collectors: a new theoretical approach. A.I.Ch.E. SymposiumSeries, 68:196–206, 1972.
[100] A. Leonard. Energy cascade in LES of turbulent fluid flows. Ad-vances in Geophysics, 18:237–248, 1974.
[101] D. K. Lilly. A proposed modification of the germano subgrid-scaleclosure method. Physics of Fluids A, 4:633–635, 1992.
[102] K. S. Lim, H. S. Kim, and K. W. Lee. Characteristics of the collectionefficiency for a cyclone with different vortex finder shapes. Journalof Aerosol Science, 35(6):743–754, 2004.
[103] K. S. Lim, S. B. Kwon, and K. W. Lee. Characteristics of the collectionefficiency for a double inlet cyclone with clean air. Journal of AerosolScience, 34(8):1085–1095, 2003.
[104] A. J. Linden. Investigations into cyclone dust collectors. Proceedingsof the Institution of Mechanical Engineers – Part B – EngineeringManufacture (Professional Engineering Publishing), 160:233–251,1949.
[105] Z. Liu, Y. Zheng, L. Jia, J. Jiao, and Q. Zhang. Stereoscopic PIVstudies on the swirling flow structure in a gas cyclone. ChemicalEngineering Science, 61:4252 – 4261, 2006.
[106] C. Liyana-Pathirana and F. Shahidi. Optimization of extraction ofphenolic compounds from wheat using response surface methodol-ogy. Food Chemistry, 93:47–56, 2005.
[107] P. Worth Longest and Samir Vinchurkar. Effects of mesh style andgrid convergence on particle deposition in bifurcating airway mod-els with comparisons to experimental data. Medical Engineering &Physics, 29(3):350 – 366, 2007.
[108] J. H. Mathews and K. D. Fink. Numerical Methods Using MATLAB.Prentice Hall, 3rd edition, 1999.
[109] Matlab. Global Optimization Toolbox User’s Guide, Matlab R2010a.The MathWorks, Inc., 2010.
[110] P. Moin and J. Kim. Numerical investigation of turbulent channelflow. Journal of Fluid Mechanics, 155:441, 1982.
289
[111] M. E. Moore and A. R. McFarland. Performance modeling of single-inlet aerosol sampling cyclones. Environmental Science and Tech-nology, 27(9):1842–1848, September 1993.
[112] S. A. Morsi and A. J. Alexander. An investigation of particle tra-jectories in two-phase flow systems. Journal of Fluid Mechanics,55(02):193–208, 1972.
[113] S. Murakami. Overview of turbulence models applied in CWE-1997.Journal of Wind Engineering and Industrial Aerodynamics, 74-76:1– 24, 1998.
[114] E. Muschelknautz. Die berechnung von zyklonabscheidern fur gas.Chemie Ingenieur Technik, 44:63–71, 1972.
[115] E. Muschelknautz and W. Kambrock. Aerodynamische beiwerte deszyclon abscheiders aufgrund neuer und verbesserter messungen.Chem-Ing-Tech, 42(5):247–55, Chemie Ingenieur Technik 1970.
[116] E. Muschelknautz and M. Trefz. Design and calculation of higherand design and calculation of higher and highest loaded gas cy-clones. in ‘Proceedings of Second World Congress on Particle Tech-nology’ Kyoto, Japan, pages 52–71, October 1990.
[117] E. Muschelknautz and M. Trefz. VDI-Warmeatlas. Druckverlustund Abscheidegrad in Zyklonen, 6 Auflage:Lj1 – Lj9, 1991.
[118] Raymond H. Myers, Douglas C. Montgomery, and Christine M.Anderson-Cook. Response Surface Methodology: Process and Prod-uct Optimization Using Designed Experiments. John Wiley & Sons,Inc., 3rd edition, 2009.
[119] J. A. Nelder and R. Mead. A simplex method for function minimiza-tion. The Computer Journal (1965), 7(4).
[120] Antonios Niros and George Tsekouras. A fuzzy clustering algorithmto estimate the parameters of radial basis functions neural networksand its application to system modeling. In John Darzentas, GeorgeVouros, Spyros Vosinakis, and Argyris Arnellos, editors, ArtificialIntelligence: Theories, Models and Applications, volume 5138 of Lec-ture Notes in Computer Science, pages 194–204. Springer Berlin /Heidelberg, 2008.
[121] Stefan Obermair, Christoph Gutschi, Jakob Woisetschlager, andGernot Staudinger. Flow pattern and agglomeration in the dustoutlet of a gas cyclone investigated by phase doppler anemometry.Powder Technology, 156:34–42, 2005.
[122] Stefan Obermair and Gernot Staudinger. The dust outlet of a gas cy-clone and its effects on separation efficiency. Chemical EngineeringTechnology, WILEY-VCH Verlag GmbH, 24(12):1259–1263, 2001.
290
[123] Stefan Obermair, Jakob Woisetschlager, and Gernot Staudinger. In-vestigation of the flow pattern in different dust outlet geometriesof a gas cyclone by laser doppler anemometry. Powder Technology,138:239–251, 2003.
[124] T. O’Doherty, A. J. Griffiths, N. Syred, P. J. Bowen, and W. Fick. Ex-perimental analysis of rotating instabilities in swirling and cyclonicflows. Developments in Chemical Engineering and Mineral Process-ing, 7:245–267, 1999.
[125] T. J. Overcamp and S. E. Scarlett. Effect of Reynolds number onthe Stokes number of cyclones. Aerosol Science and Technology,19(3):362–370, 1993.
[126] S. H. R. Pasandideh and S. T. A. Niaki. Multi-response simula-tion optimization using genetic algorithm within desirability func-tion framework. Applied Mathematics and Computation, 175(1):366– 382, 2006.
[127] W. Peng, A. C. Hoffmann, P. J. A. J. Boot, A. Udding, H.W.A. Dries,A. Ekker, and J. Kater. Flow pattern in reverse-flow centrifugal sep-arators. Powder Technology, 127:212 – 222, 2002.
[128] R. H. Perry and D. W. Green, editors. Perry’s Chemical Engineers’Handbook. McGraw-Hill, 1997.
[129] U. Piomelli, Y. Yu, and R. J. Adrain. Sub-grid scale energy transferand near-wall turbulence structure. Physics of Fluids, 8:215 –224,1996.
[130] S. I. Pishbin and M. Moghiman. Optimization of cyclone separatorsusing genetic algorithm. International Review of Chemical Engineer-ing (I.RE.CH.E.), 2(6):683–690, 2010.
[131] N. W. Polhemus. How to perform an optimiza-tion experiment using statgraphics centurion, 2005.http://www.sigmaplus.fr/telecharg/StatG/howto9.pdf .
[132] S. B. Pope. Turbulent flows. Cambridge university press, New York,2000.
[133] L. Prandtl. Bericht uber die entstehung der turbulenz. Zeitschrift
f’ur Angewandte Mathematik und Mechanik, 5:136–139, 1925.
[134] W. H. Press, B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling. Nu-merical Recipes in FORTRAN 77: The Art of Scientific Computing.Cambridge University Press, Cambridge, UK, 1992.
[135] William H. Press, Brian P. Flannery, Saul A. Teukolsky, andWilliam T. Vetterling. Numerical Recipes in C The Art of ScientificComputing. Cambridge University Press, Cambridge, UK, 1992.
291
[136] F. Qian and Y. Wu. Effects of the inlet section angle on the separa-tion performance of a cyclone. Chemical Engineering Research andDesign, 87(12):1567–1572, 2009.
[137] F. Qian and M. Zhang. Effects of the inlet section angle on the flowfield of a cyclone. Chemical Engineering & Technology, 30(11):1521–4125, 2007.
[138] Fuping Qian, Jiguang Zhang, and Mingyao Zhang. Effects of theprolonged vertical tube on the separation performance of a cyclone.Journal of Hazardous Materials, B(136):822–829, 2006.
[139] G. Ramachandran, D. Leith, J. Dirgo, and H. Feldman. Cyclone opti-mization based on a new empirical model for pressure drop. AerosolScience and Technology, 15:135–148, 1991.
[140] A. Raoufi, M. Shams, M. Farzaneh, and R. Ebrahimi. Numericalsimulation and optimization of fluid flow in cyclone vortex finder.Chemical Engineering and Processing, 47:128–137, 2008.
[141] G. Ravi, Santosh K. Gupta, and M. B. Ray. Multiobjective optimiza-tion of cyclone separators using genetic algorithm. Ind. Eng. Chem.Res., 39:4272–4286, 2000.
[142] K. Rietema. Het mechanisme van de afscheiding van fijnverdeeldestoffen in cyclonen (in dutch). De Ingenieur, 71(39):ch59–ch65, 1959.
[143] P. J. Roache. Perspective: A method for uniform reporting of gridrefinement studies. Journal of Fluids Engineering, 116(3):405–413,1994.
[144] Patrick J. Roache. Verification of codes and calculations. AIAA Jour-nal, 36(5):696–702, 1998.
[145] PJ Roache. Quantification of uncertainty in computational fluid dy-namics. Annual Review of Fluid Mechanics, 29:123–160, 1997.
[146] H. Safikhani, M. A. Akhavan-Behabadi, N. Nariman-Zadeh, andM. J. Mahmood Abadi. Modeling and multi-objective optimizationof square cyclones using CFD and neural networks. Chemical Engi-neering Research and Design, 89(3):301 – 309, 2011.
[147] H. Safikhani, M.A. Akhavan-Behabadi, M. Shams, and M. H.Rahimyan. Numerical simulation of flow field in three types of stan-dard cyclone separators. Advanced Powder Technology, 21(4):435–442, 2010.
[148] H. Safikhani, A. Hajiloo, M.A. Ranjbar, and N. Nariman-Zadeh.Modeling and multi-objective optimization of cyclone separators us-ing CFD and genetic algorithms. Computers & Chemical Engineer-ing, 35(6):1064–1071, 2011.
292
[149] H. Safikhani, S.A. Nourbakhsh, and N. Nariman-zadeh. Modelingand multi-objective optimization of cyclone vortex finder using CFDand neural networks. In 2nd International Conference on Engineer-ing Optimization, September 6-9, 2010, Lisbon, Portugal, pages 1–9,2010.
[150] P. Sagaut. Large-Eddy Simulation for Incompressible Flows. Scien-tific Computation, Springer, 2002.
[151] B.E. Saltzman and J.M. Hochstrasser. Design and performance ofminiature cyclone for respirable aerosol sampling. EnvironmentalScience & Technology, 17(7):418–424, July 1983.
[152] S. Schmidt, H. M. Blackburn, M. Rudman, and I. Sutalo. Simula-tion of turbulent flow in a cyclonic separator. In Third internationalconference on CFD in the minerals and process industries, CSIRO,Melbourne, Australia, 10-12 December 2003.
[153] St. Schmidt, H. M. Blackburn, and M. Rudman. Impact of outletboundary conditions on the flow properties within cyclone. In 15th
Australasian Fluid Mechanics Conference, The University of Sydney,Sydney, Australia, 2004.
[154] H. Shalaby. On the Potential of Large Eddy Simulation to SimulateCyclone Separators. PhD thesis, Chemnitz University of Technology,Chemnitz, Germany, 2007.
[155] H. Shalaby, K. Pachler, K. Wozniak, and G. Wozniak. Compara-tive study of the continuous phase flow in a cyclone separator usingdifferent turbulence models. International Journal for NumericalMethods in Fluids, 48(11):1175–1197, 2005.
[156] H. Shalaby, K. Wozniak, and G. Wozniak. Particle-laden flow simula-tion in a cyclone separator. WILEY-VCH Verlag GmbH & Co. KGaA,Weinheim, 6(1):547–548, 2006.
[157] C. B. Shepherd and C. E. Lapple. Flow pattern and pressure dropin cyclone dust collectors cyclone without intel vane. Industrial &Engineering Chemistry, 32(9):1246–1248, September 1940.
[158] Sujeet Kumar Shukla, Prashant Shukla, and Pradyumna Ghosh.Evaluation of numerical schemes using different simulation meth-ods for the continuous phase modeling of cyclone separators. Ad-vanced Powder Technology, 22(2):209–219, 2010.
[159] M. D. Slack, R. O. Prasad, A. Bakker, and F. Boysan. Advances incyclone modeling using unstructured grids. Trans IChemE., 78 PartA, (2000), 2000.
[160] John W. Slater. Examining spatial (grid) convergence, 2008.http://www.grc.nasa.gov/WWW/wind/valid/tutorial/spa tconv.html .
293
[161] J. Smagorinsky. General circulation experiments with the primitiveequations: The basic experiment. Monthly weather review, 91(3):99–164, 1963.
[162] A. Smirnov, S. Shi, and I. Celik. Random flow generation techniquefor large eddy simulations and particle-dynamics modeling. Journalof Fluids Engineering, 123(2):359–371, 2001.
[163] G. Solero and A. Coghe. Experimental fluid dynamic characteriza-tion of a cyclone chamber. Experimental Thermal and Fluid Science,27:87–96, 2002.
[164] M. Sommerfeld, A. Ando, and D. Wennerberg. Swirling, particle-laden flows through a pipe expansion. Journal of Fluids Engineer-ing, 114:648–656, 1992.
[165] C. J. Stairmand. Pressure drops in cyclone separators. Industrialand Engineering Chemistry, 16(B):409–411, 1949.
[166] C. J. Stairmand. The design and performance of cyclone separators.Industrial and Engineering Chemistry, 29:356–383, 1951.
[167] G. Steenackers, R. Versluys, M. Runacres, and P. Guillaume.Reliability-based design optimization of computation-intensive mod-els making use of response surface models. Quality and ReliabilityEngineering International, 27(4):555–568, 2011.
[168] Gunther Steenackers. Finite Element Model Updating and Opti-mization of Mechanical Systems Making Use of Regressive Tech-niques. PhD thesis, Vrije Universiteit Brussel, 2008.
[169] Gunther Steenackers and Patrick Guillaume. Bias-specified robustdesign optimization: A generalized mean squared error approach.Computers & Industrial Engineering, 54(2):259 – 268, 2008.
[170] Gunther Steenackers, Patrick Guillaume, and Steve Vanlanduit. Ro-bust optimization of an airplane component taking into account theuncertainty of the design parameters. Quality and Reliability Engi-neering International, 25(3):255–282, 2009.
[171] A.C. Stern, K.J. Caplan, and P.D. Bush. Cyclone dust collectors. NewYork: American Petroleum Institute, 1955.
[172] P. K. Swamee, N. Aggarwal, and K. Bhobhiya. Optimum design ofcyclone separator. American Institute of Chemical Engineers (AIChE), 55(9):2279–2283, 2009.
[173] Dominique Thevenin and Gabor Janiga. Optimization and compu-tational fluid dynamics. Springer Verlag, Berlin, 2008.
294
[174] M. Trefz. Die vershiedenen abscheidevorgange im hoher un hochbeladenen gaszyklon unter besonderer berucksichtigung der sekun-darstromung. Forschritt-Berichte VDI; VDI-Verlag GmbH: Dussel-dorf, Germany (1992), 295.
[175] M. Trefz and E. Muschelknautz. Extended cyclone theory for gasflows with high solids concentrations. Chemie Ingenieur Technik,16:153–160, 1993.
[176] P. Venkataraman. Applied optimization with MATLAB program-ming. John Wiley and Sons, 2nd edition, 2009.
[177] H.K. Versteeg and W. Malalasekera. Introduction to ComputationalFluid Dynamics, the Finite Volume Method. Longman Scientific &Technical, 1995.
[178] B. Wang, D.L. Xu, K.W. Chu, and A.B. Yu. Numerical study of gas-solid flow in a cyclone separator. Applied Mathematical Modelling,30:1326–1342, 2006.
[179] B. Wang, D.L. Xu, G. X. Xiao, K. W. Chu, and A. B. Yu. Numericalstudy of gas-solid flow in a cyclone separator. In Third InternationalConference on CFD in the Minerals and Process Industries, pages371–376, CSIRO, Melbourne, Australia, December 2003.
[180] Lei Wang, Cheng Shao, Hai Wang, and Hong Wu. Radial basis func-tion neural networks-based modeling of the membrane separationprocess: Hydrogen recovery from refinery gases. Journal of NaturalGas Chemistry, 15:230–234, 2006.
[181] X.D. Wang, C. Hirsch, Sh. Kang, and C. Lacor. Multi-objective opti-mization of turbomachinery using improved NSGA-II and approxi-mation model. Computer Methods in Applied Mechanics and Engi-neering, 200(9-12):883 – 895, 2011.
[182] Xiaodong Wang. CFD Simulation of Complex Flows in Turboma-chinery and Robust Optimization of Blade Design. PhD thesis, VrijeUniversiteit Brussel, 2010.
[183] D. C. Wilcox. Turbulence Modeling for CFD. DCW Industries, Inc,2nd ed. edition, 1994.
[184] R. Xiang, S. H. Park, and K. W. Lee. Effects of cone dimension on cy-clone performance. Journal of Aerosol Science, 32(4):549–561, 2001.
[185] R. B. Xiang and K. W. Lee. Numerical simulation of flow patternsin cyclones of different cone dimensions. Part. Part. Syst. Charact.,22(3):212–218, 2005.
295
[186] R. B. Xiang and K. W. Lee. Numerical study of flow field in cyclonesof different height. Chemical Engineering and Processing, 44:877–883, 2005.
[187] H. Yoshida, Y. Inada, K. Fukui, and T. Yamamoto. Improvement ofgas-cyclone performance by use of local fluid flow control method.Powder Technology, 193:6–14, 2009.
[188] Jin You-hai, Ji Guang-qin, Cao Qing-yun, and Wang Jian-jun. Nu-merical simulation of gas-phase flow field in vortex finder of cycloneseparators. Journal of China University of Petroleum, 32(6):109–12,2008.
[189] Y. Yuan, Y. Gao, L. Mao, and J. Zhao. Optimisation of conditions forthe preparation of β-carotene nanoemulsions using response surfacemethodology. Food Chemistry, 107:1300–1306, 2008.
[190] Aishe Zhang and Ling Zhang. RBF neural networks for the pre-diction of building interference effects. Computers & Structures,82(27):2333–2339, 2004.
[191] B. Zhang and S. Hui. Numerical simulation and PIV study of theturbulent flow in a cyclonic separator. In International Conferenceon Power Engineering, Hangzhou, China, October 2007.
[192] B. Zhao. A theoretical approach to pressure drop across cyclone sep-arators. Chemical Engineering Technology, 27:1105–1108, 2004.
[193] B. Zhao. Experimental investigation of flow patterns in cycloneswith conventional and symmetrical inlet geometries. Chemical En-gineering & Technology, 28(9):969 – 972, 2005.
[194] B. Zhao. Modeling pressure drop coefficient for cyclone separatorsa support vector machine approach. Chemical Engineering Science,64:4131–4136, 2009.
[195] B. Zhao, H. Shen, and Y. Kang. Development of a symmetrical spiralinlet to improve cyclone separator performance. Powder Technology,145:47–50, 2004.
[196] B. Zhao and Y. Su. Particle collection theory for cyclone separators:summary and comparison. Particle & Particle Systems Characteri-zation, 23:484–488, 2006.
[197] B. Zhao and Y. Su. Artificial neural network-based modeling of pres-sure drop coefficient for cyclone separators. chemical engineeringresearch and design, 88:606–613, 2010.
[198] B. Zhao, Y. Su, and J. Zhang. Simulation of gas flow pattern andseparation efficiency in cyclone with conventional single and spiraldouble inlet configuration. Chemical Engineering Research and De-sign, 84:1158–1165, 2006.
296
[199] Y. Zhu, M. C. Kim, K. W. Lee, Y. O. Park, and M. R. Kuhlman. De-sign and performance evaluation of a novel double cyclone. AerosolScience and Technology, 34(8):373–380, 2000.
[200] Y. Zhu and K. W. Lee. Experimental study on small cyclones oper-ating at high flow rates. Journal of Aerosol Science, 30:1303–1315,1999.
297