7/29/2019 Kollmann(2010)

1/75

J O H A N N E S K E P L E R

U N I V E R S I T A T L I N ZN e t z w e r k f u r F o r s c h u n g , L e h r e u n d P r a x i s

Sensitivity Analysis: The Direct and Adjoint

Method

Masterarbeit

zur Erlangung des akademischen Grades

Diplomingenieur

in der Studienrichtung

Industriemathematik

Angefertigt am Institut fur Numerische Mathematik

Betreuung:

A. Univ.-Prof. Dipl.-Ing. Dr. Walter Zulehner

Eingereicht von:

Markus Kollmann

Linz, Februar 2010

Johannes Kepler UniversitatA-4040 Linz Altenbergerstrae 69 Internet: http://www.uni-linz.ac.at DVR 0093696

7/29/2019 Kollmann(2010)

2/75

Abstract

Shape optimization is widely used in practice. The typical problem is tofind the optimal shape which minimizes/maximizes a certain cost functionaland satisfies some given constraints. Usually shape optimization problems

are solved numerically, by some iterative method which also requires somegradient information.There are two approaches to provide such information: the direct approachand the adjoint approach.In this thesis the different approaches, for getting gradient information of thefunctional, will be presented and then the focus is to compare them. It willbe shown that the adjoint approach has a great advantage, but that therealso exist examples, where the adjoint technique does not work in the sensewe will introduce it.

Starting with the explanation of shape optimization, the direct and adjointapproach will be introduced. Then an introduction in shape derivatives isgiven and some necessary examples for them are shown.On the basis of a two-dimensional magnetostatic field problem, where themathematical model is derived using Maxwells equations, and is the varia-tional formulation of a boundary value problem, the process of getting gra-dient information will be shown.Finally, numerical results illustrate the two approaches.

i

7/29/2019 Kollmann(2010)

3/75

Zusammenfassung

Formoptimierung ist ein weit verbreitetes Gebiet der Optimierung. Das Pro-blem besteht darin, eine optimale Form (Geometrie) zu finden, die ein gegebe-nes Zielfunktional minimiert/maximiert und vorgegebene Nebenbedingungen

erfullt. Solche Probleme werden meistens mit iterativen Verfahren numerischgelost und diese Verfahren benotigen Gradienteninformation.Um diese zu bewerkstelligen gibt es zwei Methoden: die direkte Methode unddie adjungierte Methode.In dieser Arbeit werden die beiden Zugange, um Information uber den Gra-dienten des Funktionals zu bekommen, aufgezeigt und dann liegt der Fokusdarin, die Methoden zu vergleichen. Es wird sich herausstellen, dass der ad-jungierte Zugang einen groen Vorteil besitzt, es aber auch Beispiele existie-ren, wo er nicht anwendbar ist.

Zuerst folgt eine Erklarung der Idee der Formoptimierung und die direkte

und adjungierte Methode werden erlautert. Dann wird eine Einfuhrung indie Theorie der shape derivatives gegeben und einige wichtige Beispiele wer-den prasentiert.Auf Basis eines zweidimensionalen magnetostatischen Feldproblems, bei demdas mathematische Modell mithilfe der Maxwell Gleichungen hergeleitet wird(Variationsformulierung eines Randwertproblems), werden die Methoden ge-zeigt.Zum Schluss werden numerische Resultate die beiden Methoden veranschau-lichen.

ii

7/29/2019 Kollmann(2010)

4/75

Acknowledgement

First of all, I would like to express my thanks to my supervisor Prof. W.Zulehner, for giving me the chance to write this thesis, for his guidance andsuggestions on this work; especially, thanks for the providing me with the

knowledge on shape derivatives.

I also want to thank my colleagues, especially Stefan Muhlbock and LisaFrank for the valuable contributions.

Last but not least I want to thank my father for giving me the possibil-ity to study and my girlfriend Sylvia for the mental assistance.

This work has been carried out at the Institute of Computational Mathe-matics, JKU Linz.

iii

7/29/2019 Kollmann(2010)

5/75

Contents

1 Introduction 1

2 The Idea of Shape Optimization 3

2.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32.2 An Optimal Shape Design Problem . . . . . . . . . . . . . . . 32.3 The Direct and Adjoint Method in the Discrete Setting . . . . 4

2.3.1 The Direct Method . . . . . . . . . . . . . . . . . . . . 52.3.2 The Adjoint Method . . . . . . . . . . . . . . . . . . . 5

2.4 The Direct and Adjoint Method in the Continuous Setting . . 62.4.1 The Direct Method . . . . . . . . . . . . . . . . . . . . 72.4.2 The Adjoint Method . . . . . . . . . . . . . . . . . . . 7

3 Introduction to Shape Derivatives 9

3.1 The Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . 93.2 Shape Functionals . . . . . . . . . . . . . . . . . . . . . . . . . 10

3.2.1 Domain Integrals . . . . . . . . . . . . . . . . . . . . . 113.2.2 Boundary Integrals . . . . . . . . . . . . . . . . . . . . 14

4 Derivation of the Physical Problem 18

4.1 Physical Model . . . . . . . . . . . . . . . . . . . . . . . . . . 184.1.1 Maxwells Equations . . . . . . . . . . . . . . . . . . . 18

4.2 Mathematical Model . . . . . . . . . . . . . . . . . . . . . . . 204.2.1 Reduction to 2D . . . . . . . . . . . . . . . . . . . . . 204.2.2 Variational Formulation . . . . . . . . . . . . . . . . . 234.2.3 Existence and Uniqueness . . . . . . . . . . . . . . . . 25

4.3 The Cost Functional . . . . . . . . . . . . . . . . . . . . . . . 27

5 Shape Derivative Analysis for the Physical Problem 29

5.1 The Eulerian Derivative of the Cost Functional . . . . . . . . 295.2 The Boundary Value Problem for the Shape Derivative . . . . 31

5.2.1 The Direct Approach . . . . . . . . . . . . . . . . . . . 385.2.2 The Adjoint Approach . . . . . . . . . . . . . . . . . . 38

5.3 The Test Problem . . . . . . . . . . . . . . . . . . . . . . . . . 40

iv

7/29/2019 Kollmann(2010)

6/75

6 Discretization 45

6.1 Introduction to the Finite Element Method . . . . . . . . . . . 45

6.2 The Finite Element Method applied to the Test Problem . . . 486.3 The CG - Method . . . . . . . . . . . . . . . . . . . . . . . . . 50

7 Numerical Results 52

7.1 The Analytical Solutions . . . . . . . . . . . . . . . . . . . . . 527.2 Comparison of the Direct and Adjoint Method . . . . . . . . . 53

8 Conclusions and Outlook 64

8.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 648.2 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

Bibliography 66

Eidesstattliche Erklarung 68

Curriculum Vitae 69

v

7/29/2019 Kollmann(2010)

7/75

Chapter 1

Introduction

The thesis deals with the usage of the so-called shape derivatives in shapeoptimization. The whole thesis is mainly based on the books from Haslingerand Makinen [1], Sokolowski and Zolesio [2] and Delfour and Zolesio [3] (someparts can also be found in the book from Le Tallec and Laporte [4]).In this books one can find the mathematical approach for introducing shapederivatives and working with it. In this thesis the concentration is abouton how to calculate and work with shape derivatives and not about all theassumptions which have to be satisfied. It is assumed that everything issmooth enough, well defined and exists. So in this thesis it is shown howsomeone can work with shape derivatives.

The tasks of this thesis:

In this thesis we want to get an idea of the behaviour of two different ap-proaches (direct and adjoint method) for shape optimization by analysing atwo-dimensional problem. This happens in the following way:

Use the direct method for getting gradient information of some costfunctional.

Introduce the adjoint technique and use it for the task of providing

gradients.

Compare the results of these two methods and draw conclusions.

The organization of this thesis:

Chapter 1 is the introduction.

In chapter 2 the idea of shape optimization is introduced and the termshape optimization problem is defined. It will turn out that gradients

1

7/29/2019 Kollmann(2010)

8/75

have to be calculated. For this, two possible approaches are presented:the direct approach and the adjoint approach. The details of these

methods will be worked out for the discrete and the continuous case.

In chapter 3 some basic knowledge about shape derivatives is intro-duced, especially the definition of the change of the geometry of somebounded Lipschitz domain RN. Then the term Eulerian derivative(directional derivative of a functional) is defined and some examples forthis follow. So with this it will be shown how to compute the derivativewith respect to the change of the geometry of domain- and boundary-integrals.

In chapter 4 from the physical model for magnetostatics, i.e. Maxwells

equations, the mathematical model is derived, which is then the varia-tional formulation of a boundary value problem for a partial differentialequation. Then the cost functional, which describes the torque, is de-rived and will depend on the solution of the boundary value problem.

In chapter 5 a variational formulation for the shape derivative is derivedout of the given variational formulation of the boundary value problemfrom chapter 4 using the techniques introduced in chapter 3. Thenthe direct and adjoint approach are figured out for the (in chapter 4)derived problem.

In chapter 6 we give a short introduction to the finite element method(FEM). After applying it to the test problem, we shortly introducethe conjugate gradient method (CG method), with which we solve theresulting linear systems.

In chapter 7 the numerical results for the direct and adjoint methodare presented and compared.

In chapter 8 the results are summarized and conclusions about possiblecontinuations are drawn.

2

7/29/2019 Kollmann(2010)

9/75

Chapter 2

The Idea of Shape Optimization

In this chapter the whole idea and principle of shape optimization is in-troduced (as one can find it in [1] and [5]). This will lead us then to twoapproaches which are presented: the direct method and the adjoint method(see also [6]). There the so-called shape derivatives appear, which are thenformally introduced in chapter 3.

2.1 Motivation

The typical aim of a shape optimization problem is to modify the shape ofan object in such a way that the resulting object is optimal with respectto a certain criterion. This criterion is usually some sort of cost functionalwhich then has to be minimized or maximized. Since any minimizationproblem can be easily written as a maximization problem and vice versa,only maximization problems are considered.In most of the cases the functional itself is depending on a variable which isthe solution of a state problem and given through equations.

2.2 An Optimal Shape Design Problem

Definition 2.1 Let O be a family of admissible domains RN

, thedomain for the state problem and u() the solution of the state problem.Then the graph of the optimal shape design problem is defined as the set:

G := {(, u())| O}.

Now let I(, u()) with I : G R be the cost functional and setJ() = I(, u()). Then the optimal shape design problem can be writtenas:

Find O : J() J() for all O. (2.1)

3

7/29/2019 Kollmann(2010)

10/75

If the shape is given through a parametrization we have the following defini-tion:

Definition 2.2 Let P be a family of admissible parameters, P Rk, = (p) the domain for the state problem with p P and u(p) = u((p))the solution of the state problem. Then the graphof the optimal shape designproblem for parameterized shapes is defined as the set:

G := {(p, u(p))|p P}.

Now let again I(p, u(p)) with I : G R be the cost functional and setJ(p) = I(p, u(p)). Then the optimal shape design problem for parameterizedshapes can be written as:

Find p P : J(p) J(p) for all p P. (2.2)

The kind and number of design parameters is depending on the problem.

In the next sections the direct and adjoint methods will be introduced inthe discrete and continuous setting but the finite element method for dis-cretization will be discussed in detail in chapter 6. So for now only someresults from the finite element method are used which will be discussed in alater chapter.

2.3 The Direct and Adjoint Method in the

Discrete Setting

Assume that the state problem is already discretized and reads as follows:

Find u RNp : K(p)u(p) = f(p), (2.3)

where K is a sparse matrix.Additionally, consider the shape design problem (2.2) with constraints

cl(p, u(p)) 0, where l = 1,...,M and M is the number of constraints.

So in short the following problem has to be solved (with the settingcl(p) = cl(p, u(p))):

J(p) maxpRk

(2.4)

cl(p) 0.

4

7/29/2019 Kollmann(2010)

11/75

In optimization routines the gradients of J and c are needed and they lookas follows:

dJ

dpi=

I

pi+

I

u

u

pi, (2.5)

dcldpi

=clpi

+clu

u

pi,

where i = 1,...,k and l = 1,...,M.

The quantities u/pi are called sensitivities.

For calculating these gradients the special structure of the state equation

(Ku = f) is exploited and leads to two approaches: the direct method andthe adjoint method.

2.3.1 The Direct Method

Differentiation of the state equation with respect to pi gives:

(Ku)

pi=

f

pi

which is equivalent to

K

piu + K

u

pi=

f

pi

or, equivalently

Ku

pi=

f

pi

K

piu. (2.6)

The idea is now to solve (2.6) with respect to u/pi and use this solutionfor the calculation of the gradients of J and cl in (2.5).

Therefore, for each design parameter pi one solution of problem (2.6) isneeded.

2.3.2 The Adjoint Method

We have

u

pi= K1

fpi

K

piu

(2.7)

5

7/29/2019 Kollmann(2010)

12/75

Then for the calculation of the gradients of J and cl we obtain:

dJ

dpi =I

pi +I

u K1

fpi

K

pi u

=I

pi+ KT

I

u f

pi

K

piu

,

dcldpi

=clpi

+clu

K1 f

pi

K

piu

=clpi

+ KTclu

f

pi

K

piu

,

where i = 1,...,k and l = 1,...,M.

Now adjoint variables = (1,...,M+1)T are defined by

KT1 =I

u, (2.8)

KTl+1 =clu

.

The problem (2.8) is called the adjoint state problem.Plugging this into the formulas for calculating the gradients of J and cl leadsto

dJdpi

= Ipi

+ 1

fpi

Kpi

u

, (2.9)

dcldpi

=clpi

+ l+1 f

pi

K

piu

.

Compared to the direct approach, where we need one solution of the dif-ferentiated state problem (2.6) for each design parameter pi, here only onesolution of the adjoint state problem (2.8) is needed for the objective andeach restriction.

In the next section it is shown how to set up the approaches in the con-tinuous setting.

2.4 The Direct and Adjoint Method in the

Continuous Setting

Assume that the state problem is of the following form:

Find u V : a(u, v) = F, v , v V, (2.10)

6

7/29/2019 Kollmann(2010)

13/75

with an appropriate Hilbert space V, a bilinear form a(., .) and a linear formF, . = F(.). Or, in operator notation

Find u V : Au = F. (2.11)

Additionally, consider again problem (2.4) but for simplicity we assume thatthe functional and the constraints are not explicitly depending on the pa-rameter p, i.e. J(p) = I(u(p)) and cl(p) = cl(u(p)), and that there is onlyone design parameter:

J(p) maxpR

cl(p) 0.

Again the calculation of the gradients of the objective functional and theconstraints is needed. In chapter 3 we will see that these gradients are linearfunctionals of the following form:

dJ = f1, u , (2.12)

dcl = fl+1, u ,

where l = 1,...,M and u denotes the so-called shape derivative.

2.4.1 The Direct Method

By important techniques involving shape derivatives (see chapter 3) it ispossible to derive an equation for the unknown shape derivative u out of theequation for the state problem (2.10):

Find u V : a(u, v) = F, v , v V, (2.13)

where the bilinear form a(., .) is derived by differentiating the bilinear formin (2.10) and the linear form F, . is derived by differentiating the linearform in (2.10) (see chapter 3) . Again in operator notation

Find u

V : A

u

= F

. (2.14)

The idea is now to solve (2.13) and use this solution for the calculation ofthe gradients of J and cl in (2.12).

2.4.2 The Adjoint Method

The shape derivative u is determined as the solution of the following problem:

Find u V : Au, v = F, v , v V.

7

7/29/2019 Kollmann(2010)

14/75

Now the adjoint variables = (1,...,M+1)T are defined straight forward

by the following adjoint problem

Find (V)M+1 : (A)1, w = f1, w , w V, (2.15)

(A)l+1, w = fl+1, w , w V.

where (A) denotes the adjoint of A.

As u is an element in V we have

(A)1, u = 1, A

u = F, 1 ,

(A)l+1, u = l+1, A

u = F, l+1

and therefore

dJ = F, 1 , (2.16)

dcl = F, l+1 .

where l = 1,...,M.

Both approaches can be extended to problems with more than one designparameter. Then, in the direct approach, for each parameter we have tosolve one differentiated state problem for the shape derivative. In the ad-

joint approach only one solution of the adjoint state problem is needed forthe objective and each restriction.

8

7/29/2019 Kollmann(2010)

15/75

Chapter 3

Introduction to Shape

Derivatives

In this chapter the basic knowledge in shape derivatives is presented basedon the material derivative idea of continuum mechanics (see also [7] and [8]).This presentation will be formal, meaning that it is correct provided that alldata we need are sufficiently smooth.As seen in chapter 2, shape optimization deals with computations of deriva-tives of solutions to state problems as well as cost functionals with respectto shape variations. Here it will be shown, how to compute the derivative ofa functional (Eulerian derivative) using shape derivatives.

3.1 The Geometry

We want to study the geometric change of a bounded domain RN,N N, with Lipschitz boundary . is thought to be a collection of ma-terial particles changing their position in time. The space occupied by themat time t will determine a new configuration t. The change in the geom-etry of will be given by a process which deforms the initial configuration .

To formalize this mathematically, let

Tt : RN, t [0, )

be a family of transformations, which describes the motion of each materialparticle in the domain :

X x = Tt(X) x(t, X)

such that

T0() =

9

7/29/2019 Kollmann(2010)

16/75

i.e., T0 = identity.

The transformed geometry is given by:t = Tt()

i.e., t is the image of with respect to Tt.

Like in continuum mechanics we assume Tt to be a one-to-one transformationof onto t, such that

Tt (int) = intt, Tt() = t.

The point X may be thought of as the Lagrangian coordinates while x are

the Eulerian coordinates of a particle.Remark 3.1 In our purposes the parameter t will not be time but describesthe amount of change of the geometry.

Definition 3.2 The Eulerian velocity field V(t, x) at the point x(t) is givenby:

V(t, x) =x

t(t, T1t (x)).

By this definition one can see, that x(t, X) satisfies the following initial valueproblem

d

dtx(t, X) = V(t, x(t, X)) (3.1)

x(0, X) = X

and, vice versa, for each V(t, x) the family of transformations Tt(V)(X) canbe determined as solution of the problem (3.1).

3.2 Shape Functionals

In this section the so-called Eulerian derivative is introduced, calculated fordomain and boundary integrals and the term shape derivative is defined.

Definition 3.3 Let RN be a bounded Lipschitz domain and let J be afunctional with J(). Then the Eulerian derivative of the functional Jat in the direction of a vector field V is given by:

dJ(; V) =d

dtJ(t)

t=0

= limt0

1

t(J(t) J()) (3.2)

with t = Tt(V)().

10

7/29/2019 Kollmann(2010)

17/75

The Eulerian derivative is a directional derivative for the, in this sense oftencalled, shape functional J.

In the next subsections the Eulerian derivative of some important examplesof functionals is calculated. Let = .

3.2.1 Domain Integrals

Consider

J() =

F(x, y()(x), y()(x)) dx,

where F : R RN

R, (x,y,p) F(x,y,p).

Then

J(t) =

t

F(x, y(t)(x), y(t)(x)) dx.

Transforming the integral to an integral over (substitution rule) leads to

J(t) =

F(Tt(x), (y(t) Tt)(x), (DTTt (x)(y(t) Tt)(x)))det(DTt(x)) dx

with DTt = (Tt,i(V)

Xj(X))i,j=1,...,N.

With (3.2) we have

dJ(; V) =d

dtJ(t)

t=0

=d

dt

F(Tt(x), (y(t) Tt)(x), (DTTt (x)(y(t) Tt)(x)))(t)(x) dx

t=0

with (t)(x) := det(DTt(x)).

It can be easily shown that (0)(x) = 1.

Since the domain of integration does not depend on t anymore we can inter-change the order of differentiation and integration and get

dJ(; V)

=

d

dt

F(Tt(x), (y(t) Tt)(x), (DT

Tt (x)(y(t) Tt)(x)))(t)(x)

t=0

dx.

11

7/29/2019 Kollmann(2010)

18/75

Product rule and chain rule for differentiation leads to

dJ(; V) = F(x, y()(x), y()(x))

(0)(x) dx

+

xF(x, y()(x), y()(x)) V(0, x) dx

+

DyF(x, y()(x), y()(x))y(; V)(x) dx

+

DpF(x, y()(x), y()(x)) (y)(; V)(x) dx

+

DpF(x, y()(x), y()(x))d

dtDTTt (x)

t=0

y()(x) dx

with

V(0, x) :=dTt(x)

dt

t=0

, (3.3)

the material derivative

y(; V)(x) :=d

dt

(y(t) Tt)(x)

t=0

(3.4)

and where v w denotes the Euclidean inner product in RN. Dy (resp. Dp)

denotes the derivative with respect to y (resp. p).

It can be shown that (proof: see [2, page 76])

(0)(x) = divV(0, x) (3.5)

and

d

dt

DTTt (x)

t=0

= DVT(0, x). (3.6)

Under certain smoothness assumptions the material and spatial derivativescommute:

(y)(; V) = y(; V).

12

7/29/2019 Kollmann(2010)

19/75

With this we have

dJ(; V) = F(x, y()(x), y()(x))divV(0, x) dx

+

xF(x, y()(x), y()(x)) V(0, x) dx

+

DyF(x, y()(x), y()(x))y(; V)(x) dx

+

DpF(x, y()(x), y()(x)) y(; V)(x) dx

DpF(x, y()(x), y()(x))DVT(0, x)y()(x) dx.

Now we introduce the so called shape derivative y(; V):

Definition 3.4 The shape derivative of y() in the direction V is given by:

y(; V)(x) := y(; V)(x) y()(x) V(0, x). (3.7)

Using this definition we have

y(; V)(x) = y(; V)(x) +

y()(x) V(0, x)

= y(; V)(x) + DVT(0, x)y()(x) + (V(0, x) )y()(x).

Therefore,

y(; V)(x) DVT(0, x)y()(x) = y(; V)(x) + (V(0, x) )y()(x)

and

dJ(; V) =

F(x, y()(x), y()(x))divV(0, x) dx

+ xF(x, y()(x), y()(x)) V(0, x) dx+

DyF(x, y()(x), y()(x))y(; V)(x) dx

+

DyF(x, y()(x), y()(x))(V(0, x) )y()(x) dx

+

DpF(x, y()(x), y()(x)) y(; V)(x) dx

+

DpF(x, y()(x), y()(x)) (V(0, x) )y()(x) dx.

13

7/29/2019 Kollmann(2010)

20/75

Since

divF(x, y(x), p(x))V(x)= F(x, y(x), p(x))divV(x) + Dx

F(x, y(x), p(x))

V(x)

= F(x, y(x), p(x))divV(x) + xF(x, y(x), p(x)) V(x)

+ DyF(x, y(x), p(x))xy(x) V(x) + DpF(x, y(x), p(x))Dxp(x)V(x)

= F(x, y(x), p(x))divV(x) + xF(x, y(x), p(x)) V(x)

+ DyF(x, y(x), p(x))(V(x) )y(x) + DpF(x, y(x), p(x)) (V(x) )p(x),

we finally obtain

dJ(; V) =

DyF(x, y()(x), y()(x))y(; V)(x) dx

+

DpF(x, y()(x), y()(x)) y(; V)(x) dx

+

div

F(x, y()(x), y()(x))V(0, x)

dx,

or, by using Gauss theorem

dJ(; V) =

DyF(x, y()(x), y()(x))y(; V)(x) dx (3.8)

+

DpF(x, y()(x), y()(x)) y(; V)(x) dx

+

F(x, y()(x), y()(x))V(0, x) n(x) ds.

3.2.2 Boundary Integrals

Consider

J() =

F(x, y()(x), y()(x)) ds,

where F : R RN R, (x,y,p) F(x,y,p).

Then

J(t) =

t

F(x, y(t)(x), y(t)(x)) ds.

Transforming the boundary integral to an integral over (substitution rule)leads to

J(t) =

F(Tt(x), (y(t) Tt)(x), (DTTt (x)(y(t) Tt)(x)))(t)(x) ds

14

7/29/2019 Kollmann(2010)

21/75

with

(t)(x) = (t)(x)DT

T

t (x)n(x)l2

.We have (interchanging differentiation and integration)

d

dtJ(t)

t=0

=

d

dt

F(Tt(x), (y(t) Tt)(x), (DT

Tt (x)(y(t) Tt)(x)))(t)(x)

t=0

ds.

It can be shown that (proof: see [2, page 80])

(0)(x) = divV(0, x) DV(0, x)n(x) n(x). (3.9)

Definition 3.5 Let be a bounded Lipschitz domain with the boundary and V a vector field. Then the tangential divergenceis defined as

divV := divV DV n n.

With this we have

d

dtJ(t)

t=0

=

F(x, y()(x), y()(x))divV(0, x) ds

+

xF(x, y()(x), y()(x)) V(0, x) ds

+

DyF(x, y()(x), y()(x))y(; V)(x) ds

+

DyF(x, y()(x), y()(x))V(0, x) y()(x) ds

+

DpF(x, y()(x), y()(x)) y(; V)(x) ds

+

DpF(x, y()(x), y()(x)) (V(0, x) )y()(x) ds.

Definition 3.6 The tangential gradient of y in a point of is defined as:

y = y y

nn. (3.10)

Remark 3.7 For the tangential gradient of y the following holds:

y W d =

ydivW d (3.11)

W with W n = 0 on .

15

7/29/2019 Kollmann(2010)

22/75

Using (3.10) we get

d

dtJ(t)t=0

= F(x, y()(x), y()(x))div

V(0, x) ds

+

xF(x, y()(x), y()(x)) n(x)V(0, x) n(x) ds

+

F(x, y()(x), y()(x)) V(0, x) ds

+

DyF(x, y()(x), y()(x))y(; V)(x) ds

+

DyF(x, y()(x), y()(x))y

n()(x)V(0, x) n(x) ds

+

DyF(x, y()(x), y()(x))V(0, x) y()(x) ds

+

DpF(x, y()(x), y()(x)) y(; V)(x) ds

+

DpF(x, y()(x), y()(x)) (V(0, x) n(x))

n(y()(x)) ds

+

DpF(x, y()(x), y()(x)) (V(0, x) )y()(x) ds.

Since

div

F(x, y(x), p(x))V(x)

= (F(x, y(x), p(x))) V(x) + F(x, y(x), p(x))divV(x)

= F(x, y(x), p(x)) V(x) + DyF(x, y(x), p(x))V(x) y(x)

+ DpF(x, y(x), p(x)) (V(x) )p(x) + F(x, y(x), p(x))divV(x).

we get

d

dtJ(t)

t=0

=

xF(x, y()(x), y()(x)) n(x)V(0, x) n(x) ds

+

DyF(x, y()(x), y()(x))y(; V)(x) ds

+

DyF(x, y()(x), y()(x))y

n()(x)V(0, x) n(x) ds

+

DpF(x, y()(x), y()(x)) y(; V)(x) ds

+

DpF(x, y()(x), y()(x)) (V(0, x) n(x))

n(y()(x)) ds

+

div

F(x, y()(x), y()(x))V(0, x)

ds.

16

7/29/2019 Kollmann(2010)

23/75

Observe that (see [2, page 115])

div

y()(x)V(0, x)

ds =

y()(x)V(0, x) n(x) ds

with the mean curvature on the manifold

= divn(x).

So finally we get

d

dtJ(t)

t=0

=

xF(x, y()(x), y()(x)) n(x)V(0, x) n(x) ds (3.12)

+

DyF(x, y()(x), y()(x))y(; V)(x) ds

+

DyF(x, y()(x), y()(x))y

n()(x)V(0, x) n(x) ds

+

DpF(x, y()(x), y()(x)) y(; V)(x) ds

+

DpF(x, y()(x), y()(x)) (V(0, x) n(x))

n(y()(x)) ds

+

F(x, y()(x), y()(x))V(0, x) n(x) ds.

Example

Consider

J() =

y()(x) ds

with y() : R.Using (3.12) we get

d

dtJ(t)

t=0

= y(; V)(x)

+

y

n()(x)V(0, x) n(x) ds

+

y()(x)V(0, x) n(x) ds.

With the example above it is shown, how to compute the gradient of func-tionals (Eulerian derivative) which consist of integrals.In the next chapter the physical model will be discussed and the state equa-tion and the cost functional will be derived.

17

7/29/2019 Kollmann(2010)

24/75

Chapter 4

Derivation of the Physical

Problem

In this chapter the mathematical model is derived from the physical modelfor magnetostatics, i.e. Maxwells equations. The mathematical model is thevariational formulation of a boundary value problem for a partial differentialequation (PDE) (cf. [9]). The derivation is done as in [10] and [11].Then the cost functional, which describes the torque, is derived on the basisof [12]. This functional will depend on the solution of the boundary valueproblem.

4.1 Physical Model

4.1.1 Maxwells Equations

Electromagnetic phenomena can be described by Maxwells equations (4.1)- (4.4):

curlH = J +D

t(4.1)

curlE = B

t

(4.2)

divB = 0 (4.3)

divD = (4.4)

where the quantities involved are

18

7/29/2019 Kollmann(2010)

25/75

H magnetic field strengthE electric field strength

B magnetic flux densityD electric induction densityJ electric current density electric charge density.

The boldface letters denote three-dimensional vector fields. All these quan-tities depend on the position in space x = (x1, x2, x3) and on time t.

For the magnetic flux density B and the magnetic field strength H thereexists the following relation (constitutive laws):

B = or(H + H0) (4.5)

with

0 permeability of vacuum (0 := 4 107 V s

Am)

r relative permeability

and H0 given.By neglecting the facts of hysteresis in the sense that r can be representedas a function of |B| we get

B = or(|B|)(H + H0). (4.6)

where |.| denotes the Euclidean norm in R3.

Therefore, we have

(|B|)B = H + H0 (4.7)

with the so-called reluctivity:

(|B|) :=1

or(|B|) . (4.8)

Because of (4.3) a vector potential A can be found such that

B = curlA. (4.9)

A is unique up to a gradient field.

19

7/29/2019 Kollmann(2010)

26/75

Considering the low-frequency (quasi-static) case of electromagnetism, dis-placement currents are negligible in comparison with the impressed currents:Dt

|J|.If the fields H, B, J and consequently A are assumed to be independentof t, we obtain the following reduced set of equations, the magnetostaticformulation:

curlH = J

divB = 0 (4.10)

(|B|)B = H + H0.

By expressing B and H in terms of A we arrive at the magnetostatic vectorpotential formulation

curl[(|curlA|)curlA] = J + curlH0. (4.11)

The material influence appears in form of the reluctivity . We distinguishbetween the case of linear materials, where is constant, and the moregeneral case of nonlinear materials.We will consider the case of linear materials and the process of shape deriva-tives will only be worked out for such kind of materials.

4.2 Mathematical Model

4.2.1 Reduction to 2D

We now restrict the equations to a bounded domain R2. So we con-sider the magnetic field problem only in the x1-x2-plane. Also appropriateboundary conditions are needed: we will force the normal component of Bto vanish (cf. [11]).

Therefore, let R2 be a bounded domain with sufficiently smooth bound-ary = and let n = n(x1, x2) denote the outer unit normal vector on.We assume that the electric current density J is perpendicular to the mag-netic field strength H, which should lie in the x1-x2-plane, i.e.,

J =

00

J3(x1, x2)

, H =

H1(x1, x2)H2(x1, x2)

0

. (4.12)

20

7/29/2019 Kollmann(2010)

27/75

We additionally assume that

H0 = H01(x1, x2)H02(x1, x2)

0

and, therefore,

curlH0 =

00

H02x1

H01x2

. (4.13)

Because of (4.5) B has the following form:

B =

B1(x1, x2)B2(x1, x2)

0

.

On the other hand we have

B =

A3x2

A2x3A1x3

A3x1A2

x1 A1

x2

since B = curlA.

By comparison it follows that

A2x1

A1x2

= 0

and this leads to the following form for the vector potential A

A = 0

0A3(x1, x2)

(4.14)

and, therefore,

B = curlA =

A3x2

A3x10

. (4.15)

21

7/29/2019 Kollmann(2010)

28/75

In the following, A3 will be denoted by u viewed as a function from R2 R:

u := A3.

So we get

|curlA| = |u|,

curl[(|curlA|)curlA] =

00

div[(|u|)u]

.

where |.| denotes the Euclidean norm in R2 or R3.

As already mentioned we force the normal component of B to vanish onthe whole boundary . It can be shown that (cf. [11]) it is possible toreplace this boundary condition by the condition

n A = 0

which is equivalent to

n2A3n1A3

0

= 0

and, therefore,

u = 0.

Finally the problem is completely reduced to 2D and by summarizing wehave

div[(|u|)u] = J3 +H02x1

H01x2

in R2, (4.16)

u = 0 on .

or, equivalently

div[(|u|)u M] = J3 in R2, (4.17)

u = 0 on

with

M =

H02H01

.

22

7/29/2019 Kollmann(2010)

29/75

4.2.2 Variational Formulation

Equation (4.17) is the classical formulation of our problem.

We now consider m different materials in the domain such that betweentwo of them there is an interface. This means that the domain is split into

m non overlapping subdomains 1,...,m with =mi=1

i.

As already mentioned we only consider linear materials, i.e., in every subdo-main i it is assumed that the corresponding i is a positive constant.The interfaces are defined the following way:

i,j := i j , i, j = 1,...,m,i = j

where i and j are the two adjacent subdomains. Note that i,j = j,i.Let ni = ni(x1, x2) denote the outer unit normal vector to i on i,j andnj = nj(x1, x2) denote the outer unit normal vector to j on i,j.

Now let M(i) be the restriction of M to i and J(i)3 the restriction of J3

to i respectively. Then we can write the boundary value problem (4.17) inthe following form:

div(iu M(i) ) = J

(i)3 in i, i = 1,...,m (4.18)

u = 0 on .

One can show that from the three-dimensional physical interface conditions

H(i) ni + H(j) nj = 0 on i,j

the following interface conditions can be derived:iu M

(i)

ni +

ju M

(j)

nj = 0 on i,j. (4.19)

Multiplying the PDE from (4.18) by an arbitrary test function v with v = 0and integrating over the domain leads to

mi=1

i

div(iu M

(i) )

v dx =

mi=1

i

J(i)3 v dx.

23

7/29/2019 Kollmann(2010)

30/75

Integration by parts on the left hand side gives

mi=1

i

div(iu M

(i) )

v dx =

mi=1

i

(iu M(i) ) v dx

i

((iu M(i) ) n)v ds

mi=1

i1j=1j=i

i,j

((iu M(i) ) ni)v ds

+

i,j

((ju M(j) ) nj)v ds

By using the interface conditions (4.19) and the fact that v = 0

we get

mi=1

i

div(iu M

(i) )

v dx =

mi=1

i

(iu M(i) ) v dx.

Therefore, the linear variational formulation reads

Find u V0 = H10() := {v H

1() : v = 0 on := } : (4.20)

a(u, v) = F, v , v V0,

with

a(u, v) :=m

i=1

i

(iu M(i) ) v dx, (4.21)

F, v :=mi=1

i

J(i)3 v dx. (4.22)

The Sobolev space1 V0 is equipped with the following norm

||u||2H1() := ||u||2L2()

+ |u|2H1()

where

||u||L2() =

u

2

dx

1

2

|u|H1() =

|u|2 dx

12

.

The variational equation (4.20) can be written as an operator equation inthe dual space V0

Au = F (4.23)

1For a comprehensive introduction on Sobolev spaces see [13].

24

7/29/2019 Kollmann(2010)

31/75

with the linear operator A : V0 V0 defined by

Au,v = a(u, v). (4.24)

In order to ensure that the integrals appearing in (4.21) and (4.22) are well-defined, we need

J3 L2() and

M L2()2.

4.2.3 Existence and Uniqueness

Existence and uniqueness of the solution of linear variational formulations isguaranteed by the following theorem by Lax and Milgram (cf. [14]):

Theorem 4.1 (Lax-Milgram)Let V be a Hilbert space, F V and a(., .) : V V R be a bilinear formwith the following properties:

1) V-elliptic, i.e.,

1 = const > 0 : 1||v||2V a(v, v), v V,

2) V-bounded, i.e.,

2 = const > 0 : |a(w, v)| 2||w||V||v||V, v, w V.

Then there exists a unique solution u V of the variational problem

a(u, v) = F, v , v V, (4.25)

and we have the following two-side estimate

1

2 ||F||V

||u||V

1

1 ||F||V

. (4.26)

For problem (4.20) one can show that the assumptions of theorem 4.1 aresatisfied (the M - term is put to the right hand side) :

25

7/29/2019 Kollmann(2010)

32/75

F V0 :

| F, v | =mi=1

i

J(i)3 v dx +

i

M(i) v dx

=

J3v dx +

M v dx

J3v dx

+

M v dx

||J3||L2||v||L2 + ||M||L2

2|v|H1

||J3||L2 + ||M||L22

||v||H1

V0-boundedness:

|a(w, v)| =

mi=1

i

iw v dx

maxi=1,...,m imi=1

i

w v dx

= max

i=1,...,mi

w v dx

max

i=1,...,mi|w|H1|v|H1 max

i=1,...,mi||w||H1||v||H1

V0-ellipticity:

a(v, v) =

mi=1

i

iv v dx =

mi=1

ii

|v|2

dx

mini=1,...,m

i

mi=1

i

|v|2 dx = mini=1,...,m

i

|v|2 dx

= mini=1,...,m

i|v|2H1 .

By using Friedrichs inequality (4.27):

||u||2L2 c2F|u|

2H1, v H

10 () (4.27)

or, equivalently||u||2H1 (1 + c

2F)|u|

2H1

we get

a(v, v) min

i=1,...,mi

1 + c2F||u||2H1

Therefore existence and uniqueness of the solution of our problem is guaran-teed.

26

7/29/2019 Kollmann(2010)

33/75

4.3 The Cost Functional

The derivation of the functional is based on [12].

The force acting on a body in a magnetic field can be described using theMaxwell stress tensor as defined in [12, page 1]:

=

H21 12H

2 H1H2 H1H3

H2H1 H22

12H

2 H2H3

H3H1 H3H2 H23

12 H

2

with the magnetic field strength H, the magnetic permeability and

H =

H2

1 + H2

2 + H2

3 .

Now the force acting on the surface of a body inside a closed area A canbe calculated via the integral

F =

A

n ds

where n denotes the outer unit normal vector on the surface of A.

We are interested in the effect of the force acting on the rotor of an elec-

trical motor, therefore, we consider a cross section of A (x1-x2-plane), whichis a circular disk with radius r and boundary s.

For such a planar problem the normal stress n and the tangential stresst can be calculated as follows ([12, page 1-2])

nt

=

1

20

B2n B

2t

2BnBt

where Bn denotes the normal component, Bt the tangential component ofthe magnetic flux density and 0 is the permeability of vacuum.

For the calculation of the torque only the force in tangential direction isof importance and so the torque M is given by:

M = Ftr

=

s

rt ds =1

0

s

rBn(s)Bt(s) ds

where r =

x21 + x22.

27

7/29/2019 Kollmann(2010)

34/75

Because of

B = curlA =

u

x2 ux1

0

Bn, Bt can be calculated with n =1

x21 + x22

x1x2

0

:

Bn = B n =1

x21 + x22(

u

x2x1

u

x1x2)

Bt = B n = 1

x21 + x22

( ux2

x2 + ux1

x1),

where n is the vector perpendicular to n, i.e., n =1

x21 + x22

x2x1

0

.

So the functional (torque) becomes

M = J() =1

0 su()(x)TQ(x)u()(x) ds (4.28)

with

Q(x) =1

x21 + x22

x1x2

x21x2

2

2x21x2

2

2x1x2

. (4.29)

Remark 4.2 The functional describing the torque involves the matrix Q(x)as defined in (4.29) . It is possible to replace this special matrix by any othermatrix, e.g.: Q(x) = I, but the resulting functional will then not describethe torque anymore.

One can see that the functional depends on the solution of the boundaryvalue problem (BVP) (4.18).

As mentioned above, we assume that the domain for the state problem issplit into m non overlapping subdomains, which leads to the already dis-cussed interfaces. So s refers to the boundary of some circular subdomaininside.

This functional has to be maximized.

28

7/29/2019 Kollmann(2010)

35/75

Chapter 5

Shape Derivative Analysis for

the Physical Problem

For doing optimization, the Eulerian derivative of the functional (4.28) hasto be computed (like in chapter 3). Since it depends on the solution of theBVP, we already saw that the Eulerian derivative will then depend also onthe solution of the BVP and on its shape derivative. So we need to derive aBVP for the shape derivative. How this is done is shown in this chapter.

5.1 The Eulerian Derivative of the Cost Func-

tional

As discussed in the previous chapter the cost functional, which describes thetorque, has the following form:

J() =1

0

s

u()(x)TQ(x)u()(x) ds

with

Q(x) =1

x21 + x22

x1x2x21x2

2

2x21x2

2

2x1x2

and where u solves the following variational problem:

Find u V0 = H10 () : a(u, v) = F, v , v V0,

with

a(u, v) =mi=1

i

(iu M(i) ) v dx,

F, v =mi=1

i

J(i)3 v dx.

29

7/29/2019 Kollmann(2010)

36/75

Remark 5.1 The gradient of an H1- function is not well defined at bound-aries, so the functional would not be well defined in H1. Therefore we have

to assume that the solution u is more smooth, e.g. in H2

, then the functionalis well defined.

The functional can be written as

J() =1

0

s

F(x, u()(x)) ds

with

F(x, p) = pTQ(x)p.

Now the functional is of the form of the boundary integral in subsection 3.2.2.So we can use equality (3.12) for calculating the Eulerian derivative and thisyields

d

dtJ(t)

t=0

=

s

xF(x, u()(x)) n(x)V(0, x) n(x) ds

+

s

DpF(x, u()(x)) u(; V)(x) ds

+s DpF(x, u()(x)) (V(0, x) n(x))

n(u()(x)) ds

+

s

F(x, u()(x))V(0, x) n(x) ds.

We assume that the integration curve s is constant with respect to t. There-fore, V(0, x) n(x), which describes the velocity of particles on s in normaldirection due to variation of design parameters, equals zero.So finally we get

d

dtJ(t)

t=0

=1

0s

u(; V)(x)TQ(x)u()(x) ds (5.1)

+s

u()(x)TQ(x)u(; V)(x) ds

.

One can see that the derivative of the functional depends on the solution ofthe BVP (4.18) and its shape derivative.

For test purposes we consider a second functional of the following form:

J() =

s

u2()(x) ds (5.2)

30

7/29/2019 Kollmann(2010)

37/75

Again we can use equality (3.12) for calculating the Eulerian derivative andthis yields

d

dtJ(t)

t=0

= 2

s

u(; V)(x)u()(x) ds. (5.3)

5.2 The Boundary Value Problem for the Shape

Derivative

The variational formulation of the boundary value problem (4.18) reads

Find u V0 = H10 () : a(u, v) = F, v , v V0,

with

a(u, v) =mi=1

i

(iu M(i) ) v dx,

F, v =mi=1

i

J(i)3 v dx.

With the setting

yi() = yi,1()yi,2()

yi,3()

= u()v()M

(i) ()

we can write

a(u, v) =mi=1

i

Fi(yi()(x), yi()(x)) (5.4)

with

Fi(yi, pi) = (ipi,1 yi,3) pi,2. (5.5)

Now a(u, v) is of the form of the domain integral in subsection 3.2.1. So wecan use equality (3.8) for differentiating a(u, v) and this yields

31

7/29/2019 Kollmann(2010)

38/75

da(u, v) =

mi=1

i

Dyi,3Fi(yi()(x), yi()(x)) M(i) (; V)(x) dx

+

i

Dpi,1Fi(yi()(x), yi()(x)) u(; V)(x) dx

+

i

Dpi,2Fi(yi()(x), yi()(x)) v(; V)(x) dx

+

i

Fi(yi()(x), yi()(x))V(0, x) n(x) ds

+m

i=1

i1

j=1j=i

i,j

Fi(yi()(x), yi()(x))V(0, x) ni(x) ds

+

i,j

Fj(yj()(x), yj()(x))V(0, x) nj(x) ds

.

v is an arbitrary test function from H10 () and we take vt in the formvt = v T

1t H

10 (t). Then the material derivative v vanishes in and

therefore we have for the shape derivative v

v(; V)(x) = v()(x) V(0, x) in .

Continuing with da(u, v) we have

da(u, v) =mi=1

i

v()(x) M(i) (; V)(x) dx

+

i

iu(; V)(x) v()(x) dx

i

iu()(x) M

(i) ()(x)

(v()(x) V(0, x)) dx

+i iu()(x) M(i) ()(x) v()(x) (V(0, x) n(x)) ds

+mi=1

i1j=1j=i

i,j

iu()(x) M

(i) ()(x)

v()(x)

(V(0, x) ni(x)) ds

+

i,j

ju()(x) M

(j) ()(x)

v()(x)

(V(0, x) nj(x)) ds

.

By introducing the flux

i(u)()(x) = iu()(x) M(i) ()(x)

32

7/29/2019 Kollmann(2010)

39/75

we can further write

da(u, v) =

mi=1

i

v()(x) M(i) (; V)(x) dx

+

i

iu(; V)(x) v()(x) dx

i

i(u)()(x) (v()(x) V(0, x)) dx

+

i

(i(u)()(x) v()(x)) (V(0, x) n(x)) ds

+m

i=1

i1

j=1j=i

i,j

(i(u)()(x) v()(x)) (V(0, x) ni(x)) ds

+

i,j

(j(u)()(x) v()(x)) (V(0, x) nj(x)) ds

.

By doing integration by parts in the third i-integral we get

da(u, v) =mi=1

i

iu

(; V)(x) M(i) (; V)(x)

v()(x) dx

+i divi(u)()(x)(v()(x) V(0, x)) dx +

i Ii ds

+

mi=1

i1j=1j=i

i,j

Ii,j ds +

i,j

Ij,i ds

.

where

Ii := (i(u)()(x) v()(x)) (V(0, x) n(x))

(V(0, x) v()(x))(i(u)()(x) n(x)) ,

Ii,j := (i(u)()(x) v()(x)) (V(0, x) ni(x))

(V(0, x) v()(x))(i(u)()(x) ni(x)) ,

Ij,i := (j(u)()(x) v()(x)) (V(0, x) nj(x))

(V(0, x) v()(x))(j(u)()(x) nj(x)) .

Also the derivative with respect to t of the linear form can be computed usingequation (3.8):

33

7/29/2019 Kollmann(2010)

40/75

d F, v =

mi=1

i

J(i)3 (; V)(x)v()(x) dx +i

J(i)3 ()(x)v(; V)(x) dx

+

i

J(i)3 ()(x)v()(x)(V(0, x) n(x)) ds

+mi=1

i1j=1j=i

i,j

J(i)3 ()(x)v()(x)(V(0, x) ni(x)) ds

+

i,j

J(j)3 ()(x)v()(x)(V(0, x) nj(x)) ds

=mi=1

i

J(i)3 (; V)(x)v()(x) J

(i)3 ()(x)(v()(x) V(0, x))

dx

+

i

J(i)3 ()(x)v()(x)(V(0, x) n(x)) ds

+mi=1

i1j=1j=i

i,j

J(i)3 ()(x)v()(x)(V(0, x) ni(x)) ds

+i,j J(j)3 ()(x)v()(x)(V(0, x) nj(x)) ds.

This leads to the following variational problem for the shape derivativeu()(x):

34

7/29/2019 Kollmann(2010)

41/75

Find u

V0 = H

1

0 () :mi=1

i

iu

(; V)(x) M(i) (; V)(x)

v()(x) dx

+

i

divi(u)()(x)(v()(x) V(0, x)) dx +

i

Ii ds

+mi=1

i1j=1j=i

i,j

Ii,j ds +

i,j

Ij,i ds

=

mi=1

i

J

(i)

3 (; V)(x)v()(x) J

(i)

3 ()(x)(v()(x) V(0, x))

dx

+

i

J(i)3 ()(x)v()(x)(V(0, x) n(x)) ds

+mi=1

i1j=1j=i

i,j

J(i)3 ()(x)v()(x)(V(0, x) ni(x)) ds

+

i,j

J(j)3 ()(x)v()(x)(V(0, x) nj(x)) ds

, v V0.

We can further simplify this variational equation:

Since (cf. (4.18))

divi(u) = J(i)3 in i, i = 1,...,m

and v = 0

we obtain

m

i=1 i iu(; V)(x) M

(i) (; V)(x) v()(x) dx

=mi=1

i

J(i)3 (; V)(x)v()(x) dx

i

Ii ds

+mi=1

i1j=1j=i

i,j

J(i)3 ()(x)v()(x)(V(0, x) ni(x)) ds

+

i,j

J(j)3 ()(x)v()(x)(V(0, x) nj(x)) ds

i,j

Ii,j ds

i,j

Ij,i ds

.

35

7/29/2019 Kollmann(2010)

42/75

Now we further discuss the Ii - term:

i

Ii ds

=

i

(i(u)()(x) v()(x)) (V(0, x) n(x)) ds

i

(V(0, x) v()(x))(i(u)()(x) n(x)) ds,

and by using the definition of the tangential gradient (3.10)

i Ii ds=

i

(i(u)()(x) v()(x)) (V(0, x) n(x)) ds

i

(V(0, x) v()(x))(i(u)()(x) n(x)) ds

+

i

i(u)()(x)

v

nn(x)

(V(0, x) n(x)) ds

i

(V(0, x) v

nn(x))(i(u)()(x) n(x)) ds

=i (i(u)()(x) v()(x)) (V(0, x) n(x)) ds

i

(V(0, x) v()(x))(i(u)()(x) n(x)) ds

=

i

(V(0, x) n(x))i(u)()(x)

(i(u)()(x) n(x)) V(0, x)

v()(x) ds.

Additionally we can do integration by parts on (cf. (3.11)) and this yields

i

Ii ds

=

i

div

(V(0, x) n(x))i(u)()(x)

(i(u)()(x) n(x)) V(0, x)

v()(x) ds = 0.

The last equality is again due to v H10().

36

7/29/2019 Kollmann(2010)

43/75

Now we can do similar computations for the Ii,j - and Ij,i - term and get

i,j

Ii,j ds +i,j

Ij,i ds

=

i,j

(V(0, x) ni(x))i(u)()(x)

(i(u)()(x) ni(x)) V(0, x)

v()(x) ds

+

i,j

(V(0, x) nj(x))j(u)()(x)

(j(u)()(x) nj(x)) V(0, x)

v()(x) ds.

By using the interface conditions (4.19) this simplifies toi,j

Ii,j ds +

i,j

Ij,i ds

=

i,j

(V(0, x) ni(x))i(u)()(x) v()(x) ds

+

i,j

(V(0, x) nj(x))j(u)()(x) v()(x) ds

= i,j

(V(0, x) ni(x))iu()(x) M(i) ()(x) v()(x) ds+

i,j

(V(0, x) nj(x))

ju()(x) M(j) ()(x)

v()(x) ds.

Again we use the definition of the tangential gradient and obtaini,j

Ii,j ds +

i,j

Ij,i ds

=

i,j

(V(0, x) ni(x))

iu()(x) M(i) ()(x)

v()(x) ds

+i,j

(V(0, x) nj(x))

ju()(x) M(j) ()(x)

v()(x) ds.

Note that the term with the normal component of u vanishes.

Summarizing the variational problem for the shape derivative reads:

Find u V0 = H10 () : a

(u, v) = F, v , v V0, (5.6)

37

7/29/2019 Kollmann(2010)

44/75

with

a(u, v) :=

mi=1

i

iu(; V)(x) v()(x) dx,

F, v :=mi=1

i

J(i)3 (; V)(x)v()(x) + M

(i) (; V)(x) v()(x)

dx

+mi=1

i1j=1j=i

i,j

J(i)3 ()(x)v()(x)(V(0, x) ni(x)) ds

+

i,j

J(j)3 ()(x)v()(x)(V(0, x) nj(x)) ds

i,j

(V(0, x) ni(x))

iu()(x) M(i) ()(x)

v()(x) ds

i,j

(V(0, x) nj(x))

ju()(x) M(j) ()(x)

v()(x) ds

.

We discussed two types of functionals: the one which describes the torque(4.28) and the one which was introduced for test purposes (5.2). In thefollowing subsections it is shown, how the direct and adjoint method workfor this two functionals.

5.2.1 The Direct Approach

In chapter 2 we already saw how the direct method works in the continuoussetting:

First the solution of the variational problem (4.20) for u is computed. Thenthe solution of the variational problem (5.6) for the shape derivative u iscomputed. Now the Eulerian derivative of the two functionals (5.1) and(5.3) can be evaluated.

5.2.2 The Adjoint Approach

In chapter 2 we also saw how the adjoint method works in the continuoussetting:

First the solution of the variational problem (4.20) for u is computed. Thenwe have to set up one adjoint problem for each of the two functionals:

J() =1

0

s

u()(x)TQ(x)u()(x) ds:

38

7/29/2019 Kollmann(2010)

45/75

We define the adjoint variable as solution of the following problem(the bilinear form is symmetric):

Find V0 = H10 () : a(, w) = G, w , w V0,

with

a(, w) =mi=1

i

i()(x) w()(x) dx,

G, w :=1

0

s

w()(x)TQ(x)u()(x) ds

+s u()(x)TQ(x)w()(x) ds.

where G, w is the Eulerian derivative of the cost functional with winstead of u.

Remark 5.2 Since w is a test function in H1, the right hand side ofthe variational equation is not well defined in H1 and this means thatthe adjoint method does not make sense in H1 for such kind offunctionals.

Therefore, we cannot compare the direct and adjoint method when we

use this type of functional.

J() =

s

u2()(x) ds:

We define the adjoint variable as solution of the following problem(the bilinear form is symmetric):

Find V0 = H10 () : a

(, w) = G, w , w V0, (5.7)

with

a(, w) =m

i=1

i

i()(x) w()(x) dx,

G, w := 2

s

w()(x)u()(x) ds.

where G, w is now the Eulerian derivative of the second functionalwith w instead of u.

This variational problem is well defined in H1 and we can comparethe direct and adjoint method when we use this type of functional.

39

7/29/2019 Kollmann(2010)

46/75

Then we have to solve the adjoint problem (5.7) with respect to . Sinceu V0 and (5.7) holds for all functions in V0 we have for the Eulerian

derivatived

dtJ(t)

t=0

= G, u = a(, u) = F, .

For comparing the two different approaches, the following test problem wasconsidered:

5.3 The Test Problem

The domain is split into two subdomains 1 and 2 as follows:

Figure 5.1: The test problem

where n1(x) is the outer unit normal vector to 1 at the interface 1,2 andn1(x) is the outer unit normal vector to 2 at the interface 1,2. Furtherwe set 1 = 2 = 1 and the restrictions of the electric current density to

the subdomains J(1)3 = 0, J

(2)3 = 1 and, therefore, J

(1)3 = J

(2)3 = 0. We

additionally assume that the magnetic field H0 is 0 and, therefore, M = 0.So the boundary value problem (4.18) reads:

div(u) = J(i)3 in i, i = 1,...,m (5.8)

u = 0 on

and the corresponding variational formulation

Find u V0 = H10 () : a(u, v) = F, v , v V0, (5.9)

40

7/29/2019 Kollmann(2010)

47/75

with

a(u, v) :=

2i=1

i

u()(x) v()(x) dx,

F, v :=

2

v()(x) dx.

The analytical solution of this PDE can be computed and is given by:

u(x1, x2) =

14

(1 R2) + 12

R2 ln(R) if 0 r R14

(1 r2) + 12

R2 ln(r) if R < r 1(5.10)

with r =

x21 + x

22.

Next we define the deformation of the domain using the formalism intro-duced in chapter 3:

We assume to have only one parameter with respect to which shape vari-ations are considered: the radius of subdomain 1.Therefore, the deformation Tt is given by:

Tt :

x1x2

x1

x2 + tR

x1

x2 if 0 r R

x1

x2

+

t(1 r)r(1 R)

x1

x2

if R < r 1

(5.11)

with r =

x21 + x22.

The variational problem for the shape derivative now reads

Find u V0 : a(u, v) = F, v , v V0,

41

7/29/2019 Kollmann(2010)

48/75

with

a(u, v) :=

2i=1

i

u(; V)(x) v()(x) dx,

F, v :=

1,2

v()(x)(V(0, x) n1(x)) ds

1,2

(V(0, x) n1(x))u()(x) v()(x) ds

+

1,2

(V(0, x) n1(x))u()(x) v()(x) ds

= 1,2

v()(x)(V(0, x) n1(x)) ds.

By equation (3.3) we have

V(0, x) :=dTt(x)

dt

t=0

=

1R

x1

x2

if 0 r R

(1 r)

r(1 R) x1

x2 if R < r 1

with r =

x21 + x22.

Therefore,

V(0, x) n1(x) = 1 on 1,2

and the variational formulation for the shape derivative is simplified to

Find u

V0 : a

(u

, v) = F

, v , v V0, (5.12)

with

a(u, v) :=2i=1

i

u(; V)(x) v()(x) dx,

F, v :=

1,2

v()(x) ds.

42

7/29/2019 Kollmann(2010)

49/75

The analytical solution of this problem can be computed and is given by:

u(; V)(x1, x2) =R ln(R) if 0 r R

R ln(r) if R < r 1 (5.13)

with r =

x21 + x22.

Note that the shape derivative u equals the derivative ofu with respect to R,which is reasonable since the radius of subdomain 1 is the design parameter.

One can show that for the variational problem (5.12) the assumptions oftheorem 4.1 are satisfied by using trace theorems (cf. [13]) and thereforeuniqueness of the solution is guaranteed.

The cost functionals of consideration were already discussed and they hadthe following form

J() =1

0

s

u()(x)TQ(x)u()(x) ds

and

J() =

s

u2()(x) ds.

Remark 5.3 As already mentioned, the first functional is only well defined

if u is smooth enough, e.g. in H

2

. This is indeed true for this test problem.For the test problem we set 0 = 1 and, since taking the matrix Q(x) as in(4.29) would lead to trivial solutions, Q(x) = I. So the first functional issimplified to

J() =

s

u()(x) u()(x) ds (5.14)

and for its Eulerian derivative we have

d

dtJ(t)t=0 = 2s u

(; V)(x) u()(x) ds. (5.15)

The first functional will be evaluated at the boundary of some circle withradius re > R, the second functional will be evaluated at the interface 1,2.

The Direct Approach

First the solution of the variational problem (5.9) for u is computed. Thenthe solution of the variational problem (5.12) for the shape derivative u iscomputed. Now the Eulerian derivative of the functionals (5.15) and (5.3)can be evaluated.

43

7/29/2019 Kollmann(2010)

50/75

The Adjoint Approach for the Second Functional

First the solution of the variational problem (5.9) for u is computed. Thenwe define the adjoint variable as solution of the following problem:

Find V0 : a(, w) = G, w , w V0, (5.16)

with

a(, w) =2i=1

i

()(x) w()(x) dx,

G, w := 2 1,2

u()(x)w()(x) ds.

Since u V0 and (5.16) holds for all functions in V0 we have

d

dtJ(t)

t=0

= G, u = a(, u) = F, =

1,2

()(x) ds.

The analytical solution of this adjoint problem can be computed and is givenby:

(x1, x2) =2 R(14(1 R2) + 12R2 ln(R)) ln(R) if 0 r R

2 R(14

(1 R2) + 12

R2 ln(R)) ln(r) if R < r 1 (5.17)

with r =

x21 + x22.

One can show that for the variational problem (5.16) the assumptions oftheorem 4.1 are satisfied by using trace theorems and therefore uniquenessof the solution is guaranteed.

So far we discussed how gradients of functionals can be computed usingthe direct and adjoint method. We saw that the adjoint method is not well

defined for the functional, which describes the torque.Then we considered a particular test problem, applied the direct and adjointmethod there and this resulted in 3 variational problems: the problem for u,the problem for the shape derivative u and the adjoint problem for . Sofar, everything was formulated in the continuous setting. For solving thesevariational problems, we need to discretize them. This is done by using thefinite element method, which is introduced in the next chapter.

44

7/29/2019 Kollmann(2010)

51/75

Chapter 6

Discretization

The aim of this chapter is to describe the discretization by using the finiteelement method. First we give an introduction to the Galerkin-Ritz methodwith special ansatz functions and then we apply it to our linear problems(the original problem for u, the problem for the shape derivative u and theadjoint problem for ). The resulting systems of linear equations will besolved by the conjugate gradient (CG) method.

6.1 Introduction to the Finite Element Method

A comprehensive introduction to the finite element method (FEM) can befound in [15], [16] and [17].

As starting point we consider a linear variational problem

Find u V : a(u, v) = F, v , v V, (6.1)

in a Hilbert space V, whereby a(., .) and F, v should satisfy the assumptionsof the Lax-Milgram theorem (cf. Theorem 4.1), which ensures the existenceand uniqueness of a solution u of the problem (6.1).

The Galerkin-Ritz Method

Let us assume that we have a sequence of finite-dimensional subspacesVh V, and we associate the following discrete problems to the variationalproblem (6.1):

Find uh Vh V : a(uh, vh) = F, vh , vh Vh, (6.2)

where h indicates the discretization parameter and in fact we want to havethat for h 0 the solution of the discrete problem uh converges to the exact

45

7/29/2019 Kollmann(2010)

52/75

solution u. Since Vh V, the Lax-Milgram theorem guarantees the exis-tence and uniqueness of a solution uh of the discrete problem (6.2).

The next step is to choose appropriate basis functions for the finite-dimensionalspace Vh, such that Vh = span{p

(i)(x) : i = 1,...,Nh}, where p(i) denote the

basis functions and Nh the dimension of the space. Since uh Vh, we canuse the representation

uh(x) =

Nhi=1

u(i)h p(i)(x) (6.3)

with the coefficients u(i)h R. Therefore, we get the Galerkin system

Find uh RNh : Khuh = fh, (6.4)

with

uh :=

u(i)h

Nhi=1

,

Kh :=

a(p(j), p(i))Nhi,j=1

,

fh

:=

F, p(i)Nh

i=1.

The linear system (6.4) is equivalent to the discrete problem (6.2). Thisequivalence defines the so called Ritz isomorphism:

uh Vh uh RNh. (6.5)

Construction of Vh using the Courant Element

Now we want to construct finite element subspaces of certain Hilbert spacesand corresponding basis functions. First we introduce the notion of triangu-lation:

Definition 6.1 Let Rd

be an open, bounded set with sufficiently smoothboundary. Then Th := {r : r Rh} is called triangulation of if:

1) =

rRh

r

2) r s = 0.

Additionally we force the triangulation to be admissible, which means in 2D:

46

7/29/2019 Kollmann(2010)

53/75

Definition 6.2 A triangulation Th is admissible if:

r s =

0

common vertex

common edge.

In the Galerkin-Ritz FEM we have to face three main aspects:

A triangulation Th of the domain .

The finite element space Vh is constructed such that the functions arecontinuous and for each r Th the restriction of vh Vh is a polyno-mial.

The basis of Vh is constructed such that the basis functions have localsupport.

In 2D the widely used element is the so called Courant element. The do-main is subdivided into triangles, on which the functions vh should be linear.Therefore we have for the finite element space

Vh := {vh C()| vh|r P1 r Th}. (6.6)

The construction of this space is as follows:

Let x(j), j h denote all vertices (called nodes) of the triangulationand h the index set of the nodes.

We set

p(i)(x(j)) := ij, i, j h,

with

ij =

1 if i = j

0 otherwise.

This basis is called nodal basis, since the unknowns are the function values atthe nodes: u

(j)

h

= uh(x(j)). Therefore the basis function p(i) forms a pyramid

function with height 1 at the node x(i). Additionally this basis functionshave local support, therefore, since a(p(i), p(j)) = 0 if the supports are notoverlapping, the corresponding stiffness matrix Kh is sparse.

We can write

Vh :=

vh(x) =jh

v(j)p(j)(x)v(j) R. (6.7)

with Vh H1 for an admissible subdivision into triangles.

47

7/29/2019 Kollmann(2010)

54/75

The Reference Triangle

The stiffness matrix and the load vector (right hand side) are assembled bysumming over the elements r Th. The appearing integrals over the ele-ments can be evaluated efficiently using the concept of the reference triangle.For Courant triangles we define the reference triangle as :=

[1, 2]

[0, 1]21 + 2 1 with the following reference basis functions:

p(1)() := 1 1 2

p(2)() := 1

p(3)() := 2.

Now every r can be written as an affine linear transformation of and,consequently, the basis functions on every r can be written in terms ofthe reference basis functions. So the integrals can be easily transformedto the reference triangle using the three element nodes x(i,1), x(i,2), x(i,3). Forefficiently evaluating these integrals we used numerical integration (midpointrule).

6.2 The Finite Element Method applied to

the Test Problem

In our test problem we have to solve three variational problems:

1) The VP for u:

Find u V0 = H10 () : a(u, v) = F, v , v V0,

with

a(u, v) =2i=1

i

u()(x) v()(x) dx,

F, v =

2

v()(x) dx.

2) The VP for u:Find u V0 : a

(u, v) = F, v , v V0,

with

a(u, v) =2i=1

i

u(; V)(x) v()(x) dx,

F, v =

1,2

v()(x) ds.

48

7/29/2019 Kollmann(2010)

55/75

3) The VP for :

Find V0 : a(, w) = G, w , w V0,

with

a(, w) =2i=1

i

()(x) w()(x) dx,

G, w = 2

1,2

u()(x)w()(x) ds.

In all the problems we have the same Hilbert space H10 and the same bilinearform a(u, v). One can easily see that

V0h :=

vh(x) =jh

v(j)p(j)(x)v(j) R H10

where h is the set of all indices whose corresponding nodes do not lie on .

The construction of Kh starts with initializing all the entries with 0. Thenwe have

Kh(i,j) = a(p(j), p(i)) =

p(j) p(i)(x) dx =rBi,j

r

p(j) p(i)(x) dx

where Bi,j

:= Bi

Bj

and Bi

:= {r Rh

: x(i) r}.

Therefore we iterate over all elements r, compute the integrals using thereference triangle in each step and add this value to Kh(i,j).Because of the Ritz isomorphism the properties of the bilinear form carryover to the stiffness matrix and therefore the stiffness matrix in the threeproblems is symmetric and positive definite.

For calculating the load vectors, the same principle is used.

The implementation of the interface conditions works through iterating overall edges, which belong to the interface. As a reference element the line from

0 to 1 is taken. The rest is almost the same as above.

The implementation of the homogeneous Dirichlet boundary condition worksas follows:

set u(j)h = 0 for j h with h = h h

set Kh(i,j) = Kh(j,i) = ij

set f(j)h = 0 for j h.

49

7/29/2019 Kollmann(2010)

56/75

6.3 The CG - Method

If the matrix K in the linear systemKu = f, (6.8)

is symmetric and positive definite, problem (6.8) is equivalent to the followingminimization problem:

J(u) :=1

2(Ku,u) (f, u) min

u, (6.9)

because

J(u) = Ku f!

= 0,

2

J(u) = K = KT

> 0.J is called the energy functional.

The CG - algorithm reads as follows:

Initialization: choose initial guess u0

d0 = f Ku0

p0 = d0

Iteration: for n = 0, 1, 2,... until convergence do

un+1 = un + (n)pn with (n) =(dn, pn)

(Kpn, pn),

dn+1 = dn (n)Kpn,

pn+1 = dn+1 (n)pn with (n) =(dn+1, Kpn)

(Kpn, pn).

end.

For the CG - method we have the following convergence result:

Theorem 6.3 Let K Rnn be symmetric and positive definite.Then the CG - method is convergent and the following error estimate holds:

||un u||K 2qn

1 + q2n||u0 u||K 2q

n||u0 u||K, (6.10)

with

0 q =

(K) 1(K) + 1

< 1, (6.11)

where (K) is the condition number of the matrix K.

50

7/29/2019 Kollmann(2010)

57/75

Since the stiffness matrix in our three problems satisfies these assumptions,the CG - method can be used for solving them.

In the next chapter the numerical results are presented.

51

7/29/2019 Kollmann(2010)

58/75

Chapter 7

Numerical Results

In this chapter the numerical results for the direct and adjoint method appliedto the test problem from chapter 5 are presented. As already mentioned, forcomparing the two approaches we use the functional (5.2), since the adjointmethod is not well defined for the functional (5.14). But we also present theresults of the direct method used with the functional (5.14).

For the following results the radius of subdomain 1 was set to R = 0.4.

7.1 The Analytical Solutions

The analytical solutions of (5.9), (5.12) and (5.16) are:

u(x1, x2) =

14

(1 0.42) + 12

0.42 ln(0.4) 0.137 if 0 r R14

(1 r2) + 12

0.42 ln(r) if R < r 1,

u(; V)(x1, x2) =

0.4ln(0.4) 0.367 if 0 r R

0.4ln(r) if R < r 1

and

(x1, x2) =

2 0.4(1

4(1 0.42) + 1

20.42 ln(0.4)) ln(0.4) 0.1002 if 0 r R

2 0.4(14

(1 0.42) + 12

0.42 ln(0.4)) ln(r) if R < r 1

with r =

x21 + x22.

52

7/29/2019 Kollmann(2010)

59/75

Figure 7.1 shows u(x1, x2), u(; V)(x1, x2) and (x1, x2):

Figure 7.1: The analytical solutions u(x1, x2), u(; V)(x1, x2) and (x1, x2)

In the next section we present the results for the functional (5.2) with thedirect and adjoint method and the results for the functional (5.14) with thedirect method.

For the FEM the MATLAB mesh generator was used.

7.2 Comparison of the Direct and Adjoint

Method

First we consider the functional (5.2):

J() =

s

u2()(x) ds =

R=0.4

u2()(x) ds

=

1

4(1 0.42) +

1

20.42 ln(0.4)

22 0.4

0.046963 (7.1)

and its derivative

d

dtJ(t)

t=0

= 2

s

u(; V)(x)u()(x) ds = 2

R=0.4

u(; V)(x)u()(x) ds

= 2 0.4 ln(0.4)

1

4(1 0.42) +

1

20.42 ln(0.4)

2 0.4

0.25184. (7.2)

Therefore we can compare the numerical results with the exact solutions.

53

7/29/2019 Kollmann(2010)

60/75

For the FEM we used the following meshes:

mesh number number of unknowns (nodes) number of triangles1 152 2682 571 10723 2213 42884 8713 171525 34577 68608

The meshes were generated using the Refinement - option in Matlab.

In figure 7.2 one can see the generated meshes. Figures 7.3 - 7.7 show theFE solutions of the problems (5.9), (5.12) and (5.16) for u, u and .

54

7/29/2019 Kollmann(2010)

61/75

Figure 7.2: The used meshes

55

7/29/2019 Kollmann(2010)

62/75

Figure 7.3: The solutions u, u and with mesh 1

Figure 7.4: The solutions u, u

and with mesh 2

Figure 7.5: The solutions u, u and with mesh 3

56

7/29/2019 Kollmann(2010)

63/75

Figure 7.6: The solutions u, u and with mesh 4

Figure 7.7: The solutions u, u

and with mesh 5

With these FE solutions, we evaluated the functional and its derivative (us-ing numerical integration, in particular: midpoint rule) and got the followingresults:

mesh number functional derivative direct derivative adjoint1 0.0478194 -0.251122 -0.2511222 0.0471791 -0.251666 -0.2516663 0.0470173 -0.251796 -0.2517964 0.0469766 -0.251828 -0.251828

5 0.0469664 -0.251835 -0.251835

If we compare these results with the exact values (7.1) and (7.2) we immedi-ately see convergence of the discrete solution. We also see that the direct andthe adjoint approach deliver the same result for the derivative of the func-tional and this confirms the theory, which was done in the previous chapters.

For completeness we now consider functional (5.14) and provide gradient

57

7/29/2019 Kollmann(2010)

64/75

information using the direct method only (the adjoint one is not well definedfor this type of functional).

The functional (5.14) looks as follows:

J() =

s

u()(x) u()(x) ds

with the derivative:

d

dtJ(t)

t=0

= 2

s

u(; V)(x) u()(x) ds.

As already mentioned, this functional is evaluated at the boundary of somecircle with radius re > 0.4. We take re = 0.5. Therefore,

J() =

re=0.5

u()(x) u()(x) ds

=

re=0.5

0.42x120.52

x12

0.42x220.52

x22

0.42x120.52

x12

0.42x220.52

x22

ds

=

0.082

0.52 0.08 +

0.52

4

2 0.5

0.0254469 (7.3)

and

d

dtJ(t)

t=0

= 2

re=0.5

u(; V)(x) u()(x) ds

= 2

re=0.5

0.4x10.520.4x20.52

0.42x120.52

x12

0.42x220.52

x22

ds

= 2

0.43

2 0.52 0.2

2 0.5

0.4523893. (7.4)

For computing this functional and the derivative using the FE solutions, wehave to guarantee that there exist nodes at the radius re = 0.5. Thereforewe used different meshes than before:

mesh number number of unknowns (nodes) number of triangles1 248 4602 955 18403 3749 73604 14857 294405 59153 117760

58

7/29/2019 Kollmann(2010)

65/75

In figure 7.8 one can see the generated meshes.

Figure 7.8: The used meshes

For evaluating the functional and the derivative, we need to calculate gradi-ents out of the FE solutions. Since we used P1-functions for the FEM, thegradient is constant on each finite element, and therefore will not be contin-uous across the edge of two neighbouring elements. This means that we havethe following possibilities:

59

7/29/2019 Kollmann(2010)

66/75

use the information about the gradient in elements which touch thecircle with radius re = 0.5 inside

use the information about the gradient in elements which touch thecircle with radius re = 0.5 outside

take a mean value of the inside and outside gradient information.

How this is meant is illustrated in figure 7.9

Figure 7.9: The touching elements

The blue triangles are the elements which touch the circle with radiusre = 0.5 inside and the red ones are the elements which the touch this circleoutside. For the calculation of the gradient of the FE solution according tothe first possibility, only the blue triangles are used. For the calculation ofthe gradient of the FE solution according to the second possibility, only thered triangles are used. For the calculation of the gradient of the FE solutionaccording to the third possibility, the gradient is calculated in the red andblue triangles and then a mean value of the gradient of the neighbouring redand blue triangles is taken.

60

7/29/2019 Kollmann(2010)

67/75

For the different possibilities we got the following results:

gradient inside

mesh number functional derivative1 0.00849013 -0.2901042 0.0156795 -0.3755723 0.0202496 -0.4152434 0.0227719 -0.4341465 0.0240905 -0.443351

gradient outside

mesh number functional derivative1 0.048333 -0.5525962 0.036016 -0.5075363 0.0304912 -0.4811644 0.0279072 -0.4670685 0.0266615 -0.459801

mean value of gradient inside and outside

mesh number functional derivative1 0.0242782 -0.442411

2 0.0247937 -0.4468633 0.0251064 -0.4495264 0.0252736 -0.4509365 0.0253595 -0.451658

If we compare these results with the exact values (7.3) and (7.4) we see, thatthe solutions in all cases converge. But we also see that the one with themean values is the fastest converging solution.

For interest, consider the adjoint problem with this type of functional:

Find V0 = H10 () : a(, w) = G, w , w V0,

with

a(, w) =2i=1

i

()(x) w()(x) dx,

G, w = 2

s

u()(x) w()(x) ds.

61

7/29/2019 Kollmann(2010)

68/75

As already mentioned, w is a test function from H1, but the gradient of aH1 - function is not well defined at boundaries and, therefore, the variational

problem for the adjoint variable is not well defined in H1

.

Nevertheless, one could ignore this fact and try to discretize and implementthis variational problem using the FEM with Courant elements, like we didfor the previous (well defined) problems. But one has to think about how toimplement it. Again we have the three possibilities like above:

use the information about the gradient of u in elements which touchthe circle with radius re = 0.5 inside

use the information about the gradient of u in elements which touch

the circle with radius re = 0.5 outside

take a mean value of the inside and outside gradient information.

The following tables show the results for the different possibilities comparedwith the results from the direct approach:

gradient inside

mesh number derivative direct derivative adjoint1 -0.290104 -0.631029

2 -0.375572 -1.674323 -0.415243 -3.748374 -0.434146 -7.885315 -0.443351 -16.1521

gradient outside

mesh number derivative direct derivative adjoint1 -0.552596 -1.553892 -0.507536 -2.79243 -0.481164 -5.24236

4 -0.467068 -10.13095 -0.459801 -19.903

62

7/29/2019 Kollmann(2010)

69/75

mean value of gradient inside and outside

mesh number derivative direct derivative adjoint1 -0.442411 -0.2643512 -0.446863 -0.5361843 -0.449526 -1.079024 -0.450936 -2.163815 -0.451658 -4.33281

One immediately sees, that the solutions of the adjoint approach in all casesdo not converge and, therefore, the results are, as expected, useless.

This can be seen as a warning example in the sense that before starting

discretizing it is very important that one first makes sure that the formu-lated problems are well defined.

63

7/29/2019 Kollmann(2010)

70/75

Chapter 8

Conclusions and Outlook

8.1 Conclusion

In this thesis the entire mathematical treatment of using the shape derivativeapproach in shape optimization is embraced. We introduced two differentmethods for providing gradient information of functionals, which consist ofdomain and/or boundary integrals and depend on the solution of some stateproblem: the direct and the adjoint technique. We have seen that the adjointapproach has a great advantage compared to the direct one: the adjoint stateequation has to be solved only once.We considered a magnetostatic problem, where we have derived a mathemat-

ical formulation from the governing physical equations (Maxwell). Then weapplied the two approaches and this resulted in three linear variational prob-lems. Solvability of the problems has been proven under the assumptions ofthe Lax-Milgram theorem.

We have discretized the continuous problems with the finite element method.In order to solve the resulting systems of equations, where the involved ma-trices are symmetric and positive definite, we used the conjugate gradientmethod.

Numerical results have shown that indeed the direct and the adjoint ap-proach lead to the same results for the test problem. But we also have seenthat there exist examples where the adjoint technique is not useful in thesense of providing suitable results.

8.2 Outlook

This work can be continued in the following directions:

64

7/29/2019 Kollmann(2010)

71/75

The nonlinear problem:

When deriving the mathematical formulation we assumed the involved

materials to be linear and therefore the variational problem got linear.But in most of the problems in physics there is a nonlinear part in-volved. Therefore the adjoint method has to be set up in the nonlinearcase.

The adjoint method for general functionals:

We have shown an example of a functional where the adjoint problem,which was derived in the usual way (straight forward), is not well de-fined and therefore the adjoint approach is inapplicable in the sense weintroduced it. For this one has to think about how to reformulate the

adjoint problem.

65

7/29/2019 Kollmann(2010)

72/75

Bibliography

[1] J.HASLINGER and R.A.E.MAKINEN. Introduction to Shape Optimiza-tion: Theory, Approximation and Computation. SIAM, 2003.

[2] JAN SOKOLOWSKI and JEAN-P. ZOLESIO. Introduction to ShapeOptimization: Shape sensitivity analysis. Springer Verlag Berlin, 1992.

[3] JEAN-P. ZOLESIO and M.C. DELFOUR. Shapes and Geometries:Analysis, Differential Calculus and Optimization. SIAM, 2001.

[4] P.LE TALLEC and E.LAPORTE. Numerical Methods in Sensitivity andShape Optimization. BIRKHAUSER, 2002.

[5] W.MUHLHUBER. Spezialforschungsbereich F013: Various Aspects ofShape Optimization. Johannes Kepler University Linz, Presentation.

[6] E.J.HAUG K.K.CHOI and V.KOMKOV. Mathematics in Science andEngineering, Volume 177: Design Sensitivity Analysis of Structural Sys-

tems. ACADEMIC PRESS, INC, 1986.

[7] H.ENGL. Lecture Notes for the Course Mathematical Methods in Con-tinuum Mechanics. Johannes Kepler University Linz, Winter Semester2008/09.

[8] W.ZULEHNER. Lecture Notes for the Course Numerical Methodsin Continuum Mechanics. Johannes Kepler University Linz, SummerSemester 2009.

[9] B.HEISE. Mehrgitter-Newton-Verfahren zur Berechnung nichtlinearermagnetischer Felder. Technische Universitat Chemnitz, 1991.

[10] S.MUHLBOCK and A.LECHNER. Numerical solution of nonlinear sta-tionary magnetic field problems. Johannes Kepler University Linz, Bach-elor Thesis, 2008.

[11] C.PECHSTEIN. Multigrid-Newton-Methods for Nonlinear Magneto-static Problems. Johannes Kepler University Linz, Master Thesis, 2004.

66

7/29/2019 Kollmann(2010)

73/75

[12] W.GRUBER W.AMRHEIN S.SILBER H.GRABNER andM.REISINGER. Theoretical analysis of force and torque calcula-

tion in magnetic bearing systems with circular airgap. Institute ofElectrical Drives and Power Electronics JKU Linz.

[13] R.ADAMS. Sobolev Spaces. Academic Press, New York, 1975.

[14] U.LANGER. Lecture Notes for the Course Numerical Methods for Par-tial Differential Equations. Johannes Kepler University Linz, WinterSemester 2007/08.

[15] P.G. CIARLET. The Finite Element Method for Elliptic Problems. Clas-sics in Applied Mathematics, Vol. 40, SIAM, Philadelphia, 2002.

[16] U.LANGER and M.JUNG. Methode der finiten Elemente fur Inge-nieure. Teubner-Verlag, Stuttgart, Leipzig, Wiesbaden, 2001.

[17] W.ZULEHNER. Lecture Notes for the Course Numerical Methods forElliptic Problems. Johannes Kepler University Linz, Summer Semester2008.

67

7/29/2019 Kollmann(2010)

74/75

Eidesstattliche Erklarung

Ich, Markus Kollmann, erklare an Eides statt, dass ich die vorliegende Mas-terarbeit selbststandig und ohne fremde Hilfe verfasst, andere als die angegebe-nen Quellen und Hilfsmittel nicht benutzt bzw. die wortlich oder sinngema

entnommenen Stellen als solche kenntlich gemacht habe.

Leonding, Februar 2010 Markus Kollmann

68

7/29/2019 Kollmann(2010)

75/75

Curriculum Vitae

Name: Markus Kollmann

Nationality: Austria

Date of Birth: 1st of October, 1984

Place of Birth: Linz, Austria

Education:

1991 - 1995 Elementary SchoolKeferfeld, Linz

1995 - 1999 Secondary Comprehensive SchoolLandwiedstrae, Linz

1999 - 2004 PolytechnicLeonding

2005 - 2008 Studies of Technical Mathematics (Bachelor)Johannes Kepler University, Linz

2008 - 2010 Studies of Industrial Mathematics (Master)Johannes Kepler University, Linz

Special Activities:

August 2009 23rd ECMI Modelling Week, Wroclaw, Poland