Artificial neural network model & hidden layers in multilayer artificial neural networks
Analysis of Durability of High Performance Concrete Using Artificial Neural Networks
-
Upload
baneeishaquek -
Category
Documents
-
view
74 -
download
2
description
Transcript of Analysis of Durability of High Performance Concrete Using Artificial Neural Networks
Construction and Building Materials 23 (2009) 910–917
Contents lists available at ScienceDirect
Construction and Building Materials
journal homepage: www.elsevier .com/locate /conbui ldmat
Analysis of durability of high performance concrete using artificial neural networks
R. Parichatprecha, P. Nimityongskul *
School of Engineering and Technology, Asian Institute of Technology, P.O. Box 4, Pathumthani 12120, Thailand
a r t i c l e i n f o
Article history:Received 1 March 2007Received in revised form 14 April 2008Accepted 29 April 2008Available online 24 June 2008
Keywords:Artificial neural networksHigh performance concreteDurabilityChloride ions permeability
0950-0618/$ - see front matter � 2008 Elsevier Ltd. Adoi:10.1016/j.conbuildmat.2008.04.015
* Corresponding author. Tel.: +66 25246586; fax: +E-mail address: [email protected] (P. Nimityongskul
a b s t r a c t
This study aims to determine the influence of the content of water and cement, water–binder ratio, andthe replacement of fly ash and silica fume on the durability of high performance concrete (HPC) by usingartificial neural networks (ANNs). To achieve this, an ANNs model is developed to predict the durability ofhigh performance concrete which is expressed in terms of chloride ions permeability in accordance withASTM C1202-97 or AASHTO T277. The model is developed, trained and tested by using 86 data sets fromexperiments as well as previous researches. To verify the model, regression equations are carried out andcompared with the trained neural network. The results indicate that the developed model is reliable andaccurate. Based on the simulating durability model built using trained neural networks, the optimumcement content for designing HPC in terms of durability is in the range of 450–500 kg/m3. The results alsorevealed that the durability of concrete expressed in terms of total charge passed over a 6-h period can besignificantly improved by using at least 20% fly ash to replace cement. Furthermore, it can be concludedthat increasing silica fume results in reducing the chloride ions penetrability to a higher degree than flyash. This study also illustrates how ANNs can be used to beneficially predict durability in terms of chlo-ride ions permeability across a wide range of mix proportion parameters of HPC.
� 2008 Elsevier Ltd. All rights reserved.
1. Introduction
The use of concrete possessing both high strength and durabil-ity, hereinafter called high performance concrete (HPC), has beenincreasing all over the world. The factors which justify its popular-ity are high workability, high strength, and high durability for var-ious structural purposes. Although the definitions of HPC arevaried, the essence of HPC emphasizes three main characteristics.Apart from the three basic ingredients, i.e. cement, aggregatesand water in conventional concrete, active mineral additives likefly ash, silica fume, and superplasticizer have been incorporatedto make highly workable, high-strength and durable concrete [1–3]. HPC mix design seems to be complicated because HPC includesmore ingredients. In addition, maintaining a low water–binder ra-tio with adequate workability makes the design process more com-plicated. Traditionally, expert civil engineers can produce HPC mixproportions by using empirical results from previous research plustheir experience to achieve the required performance [4]. However,available results are often of limited value because some types ofingredients and some properties have not been investigated. Now-adays, concrete can be made with about 10 different components.The number of properties to be adjusted has also increased, soempirical methods are no longer sufficient in concrete mix design[5]. Compressive strength, workability, and durability of concrete
ll rights reserved.
66 25246059.).
are fundamentals of concrete properties. Most research has fo-cused on properties of normal concrete. Many mathematical mod-els have been suggested to describe the relationship betweencomponents and materials behavior. The traditional approach usedin modeling HPC properties starts with an assumed form of analyt-ical equation and is followed by regression analysis employingexperimental data to determine unknown coefficients in the equa-tions. Unfortunately, rational and easy-to-use equations are not yetavailable in design codes to accurately predict the properties ofHPC. Furthermore, with the aforementioned models, the evalua-tion of the effect of each parameter on the properties of concreteis almost impossible [6–8].
Durability is a fundamental property of concrete based on itsimpermeability, and it can be explicated by electrical conductivityin accordance with ASTM C1202-97 and AASHTO T277 [9,10]. If the6-h period charge passed of concrete is lower than 1000 C, the con-crete is said to possess very high impermeability and durability.The permeability of concrete depends on the pore structure of con-crete, while the electrical conductivity of concrete is determined byboth pore structure and the chemistry of the pore solution. Manyresearchers have found that the microstructure of concrete canbe improved and charge passed can be decreased by adding sup-plementary cementing materials such as fly ash, silica fume, andblast furnace slag. Since high performance concrete is a highly het-erogeneous material, the modeling of its behavior is a very difficulttask. However, there are no guidelines or specifications for durabil-ity of high performance concrete [11–15,20].
Nomenclature
OPC ordinary Portland cementF fly ashSF silica fumeW waterSP superplasticizerCA coarse aggregateFA fine aggregate
W/B water–binder ratioMSE mean square errorR2 absolute fraction of varianceMAPE mean absolute percentage errortj the target value of jth patternoj the output value of jth patternp the number of patterns
R. Parichatprecha, P. Nimityongskul / Construction and Building Materials 23 (2009) 910–917 911
In recent years, there has been interest in a class of computingdevices known as artificial neural networks (ANNs) that operate ina manner analogous to biological nervous systems. The neural net-work modeling approach is simpler and more direct than tradi-tional statistical methods, particularly when modeling nonlinearmultivariate interrelationships [16]. Very recently, many research-ers have applied neural networks to predict various properties ofconcrete. Most of the research emphasizes two basic properties,namely compressive strength and workability of HPC. Unfortu-nately, up to the present there has been no research on modelingthe durability of HPC using neural networks.
The aim of this study is to construct an ANNs model to investi-gate the influence of mix proportion parameters on the resistanceof chloride ion penetrability of high performance concrete. For thispurpose, data for developing the neural network model are col-lected from the experiments and previous research. The design ofthe experimental program is based on the relevant parameters,namely W/B, cement content, fly ash content, and silica fume con-tent. The data used in the ANNs model are arranged in a format ofeight input parameters which include the content of ordinary Port-land cement (OPC), fly ash (F), silica fume (SF), water (W), superp-lasticizer (SP), coarse aggregate (CA), fine aggregate (FA) andwater–binder ratio (W/B). To verify the model, linear and non-lin-ear regression equations are carried out and compared with theproposed neural network model. The influence of mix proportionparameters on the durability of HPC is also investigated by usingthe proposed ANNs model.
2. Artificial neural networks
Artificial neural networks are computing systems that simulatethe biological neural systems of human brain. They are based on asimplified modeling of the brain’s biological functions exhibiting
Fig. 1. Single processing element of ANNs.
the ability to learn, think, remember, reason, and solve problems.Conceptually, a neural networks model consists of a set of compu-tational units and a set of one-way data connection joining units orweights as shown in Fig. 1. Units that receive no input from othersare called input nodes, while those with no outgoing links arecalled output nodes. All other intermediate units are called hiddennodes.
The multi-layered model has several layers, and each layer con-sists of numerous neurons which are connected with each other. Inthis model, information is sent from input layer to output in onedirection, and learning is preceded so as to minimize the differencebetween the output of the model and the target output.
ANNs can solve challenging problems of interest to computerscientists and engineers such as pattern classification, categoriza-tion, function approximation, prediction and forecasting, optimiza-tion, content-addressable memory, and control robotics.Rumellhart et al. [17] developed a method called error back-prop-agation, or more simply back-propagation, for learning associa-tions between input and output patterns using more than thetwo layers of Rosenblat’s original perceptron. Back-propagation isa supervised learning technique that compares the responses ofthe output units to the desired response, and readjusts the weightsin the network so that the next time when the same input is pre-sented to the network, the network’s response will be closer tothe desired response. Traditionally, the learning process is usedto determine proper interconnection weights, and the network istrained to make proper associations between the inputs and theircorresponding outputs [5,7,18]. Errors that arise during the learn-ing process can be expressed in terms of mean square error(MSE) and are calculated using Eq. (1).
MSE ¼ 1p
� ��X
j
ðtj � ojÞ2; ð1Þ
In addition, the absolute fraction of variance (R2) and mean absolutepercentage error (MAPE) and are calculated using Eqs. (2) and (3),respectively.
R2 ¼ 1�P
jðtj � ojÞ2PjðojÞ2
!; ð2Þ
MAPE ¼ 1p
Xj
oj � tj
oj
�������� � 100
� �; ð3Þ
where tj is the target value of jth pattern, oj is the output value of jthpattern, and p is the number of patterns.
3. Experimental program and data collection
The first step in developing the network is to obtain good andreliable training and testing examples. To obtain the data for devel-oping the neural network models, a database of high strength anddurable concrete is produced by collecting the data sets fromexperiments by Parichatprecha et al. [19] combined with data setsfrom previous researches [11–15,20]. The influence of using
Table 1Ranges of components of data sets for chloride ions permeability prediction
Component Data set of high strength and durable concrete
Min (kg/m3) Max (kg/m3) Avg (kg/m3)
OPC 135 611 387F 0 275 66SF 0 110 11W/B 0.21 0.60 0.36Water 120 220 165SP 0 17.3 5.7CA 895 1167 1056FA 536 914 693
912 R. Parichatprecha, P. Nimityongskul / Construction and Building Materials 23 (2009) 910–917
different pozzolanic materials, cement content, and water-to-bin-der (W/B) ratios on the durability of concrete was experimentallyinvestigated by measuring the charge passed of concrete in accor-dance with ASTM C1202-97. The workability of concrete expressedin terms of slump was kept constant by varying the dosage ofsuperplasticizer based on poly-carboxylic ether (PCE). Two typesof pozzolanic material were used, namely pulverized fly ash anda combination of pulverized fly ash and condensed silica fume.The cementitous materials were varied from 400–550 kg/m3 withW/B ranging from 0.3 to 0.4. Control specimens without pozzolanicmaterials of concrete were also cast and tested for comparison.ASTM C1202-97 Rapid Chloride Permeability Test (RCPT) was usedin this experimental program for hardened concrete. The completeRCPT apparatus is illustrated in Fig. 2. This test method covers thedetermination of the electrical conductance of concrete to providea rapid indication of its resistance to the penetration of chlorideions. After 28 days’ curing, cylindrical specimens of 100 mm diam-eter and 200 mm length were cut to 50 mm thick on each end.These specimens were saturated in water for 18 ± 2 h until fullysaturated and then allowed to surface dry in air for at least 1 h.Next the specimens were placed on suitable silicon and completecoating of all surfaces was ensured. One side of the cell contained3.0% NaCl solution and the other 0.3 M NaOH solution. The current(ampere-seconds) was recorded at 30-min intervals during a test-ing period of 6 h. Based on the charge that passed through the sam-ple, a qualitative rating was made of the concrete’s permeability, asshown in Table 2 in accordance with ASTM C1202-97. A total of 30mixes were made and the specimens were tested for their chargepassed over a duration of 6 h.
Furthermore, to expand the prediction range of the model builtwith the experimental data, 56 concrete mixtures and their test re-sults were culled from previous researches [11–15,20]. The 28-daycompressive strength of all data is in the range of 30–120 MPa. Ofthese, the ANNs model is developed, trained and tested by using atotal of 86 data sets. Table 1 illustrates the general details of theconcrete evaluation in this study. The data used in ANNs modelare arranged in a format of eight input parameters which includeOPC, F, SF, W, SP, CA, FA, and W/B ratio. To test the reliability andaccuracy of the models, 20% of the 86 data sets were randomly se-lected as test sets, while the remaining 70 samples were used totrain the network. The output of the model is the total chargepassed in accordance with ASTM C1202 or AASHTO T277. The in-put and output of a typical neural network is in the range of 0–1.
Fig. 2. Testing apparatus for rapid ch
The use of the higher number is not desirable as the networksare generally simulated on a computer and this can create float-ing-point overflow problems [5]. Therefore, setting the input andoutput in the range of 0–1 is essential to normalize their valuesto suit the network’s functioning. In this study, x/xmax normaliza-tion technique was applied for transforming the input and outputvalues remaining in the range of 0–1.
4. Neural networks for modeling durability of HPC
The electrical conductivity of concrete is determined by bothpore structure and the chemistry of the pore solution, which aredependent on the dosage of cement, water, SP, fine aggregate,coarse aggregate and type and dosage of pozzolanic materials.The ANNs model developed in this study has eight neurons inthe input layer, one hidden layer, and an output layer as shownin Fig. 3. The selection of the number of nodes in the hidden layeris the most challenging part in the total network development pro-cess. Unfortunately, there are no fixed guidelines available for thispurpose and hence this has to be done by the trial-and-errormethod.
In this study, the neural networks were developed and per-formed under MATLAB programming. The learning algorithm usedin the study was gradient descent with adaptive learning rateback-propagation, a network training function that updates weightand bias values according to gradient descent with adaptive learn-ing rate [21]. The error incurred during the learning process wasexpressed in terms of mean-squared-error (MSE).
loride permeability test (RCPT).
Fig. 3. Architecture of neural network for predicting durability of HPC.
Fig. 4. Selection of number of neurons in hidden layer for training sets with variouslearning rate.
Fig. 5. Performance of training set of total charge passed prediction with ANNsmodel comparing with regression techniques.
Fig. 6. Performance of testing set of total charge passed prediction with ANNsmodel comparing with regression techniques.
Table 2Chloride ion penetrability based on charge passed (ASTM C1202-97)
Charge passed (Coulombs) Chloride ion penetrability
>4000 High2000–4000 Moderate1000–2000 Low100–1000 Very low<100 Negligible
R. Parichatprecha, P. Nimityongskul / Construction and Building Materials 23 (2009) 910–917 913
After a number of trials as shows in Fig. 4, the best networkarchitecture and parameters that minimize the MSE error of train-ing data were selected as follows:
� 8 input units;� 1 hidden layer;� 25 hidden units;� 1 output unit;� Activation function = sigmoidal function;� Learning rate = 0.1;� Learning cycles = 10,000.
5. Results and discussion
5.1. ANNs model analysis
The performance of the ANNs model for predicting the totalcharge passed of training and testing sets is illustrated in Figs. 5and 6, respectively. The results indicate that the proposed ANNsmodel is successful in learning the relationship between the differ-ent input and the output parameters. Fig. 5 illustrates that the
ANNs model is capable of generalizing between inputs and the out-put variables with high accuracy predictions. The statistical param-eters of the training and testing sets are shown in Table 2. All of thestatistical values in Table 3 demonstrate that the proposed ANNsmodel is suitable and can help predict the total charge passed closeto the experimental values.
To compare with statistical techniques, the seven input param-eters, namely C, F, SF, W, SP, CA, and FA, are remodeled with linearand nonlinear regression techniques. In this research, SPSS version
914 R. Parichatprecha, P. Nimityongskul / Construction and Building Materials 23 (2009) 910–917
10 was applied to determine the best fit of linear and nonlinearregression. For the observation of the performance of the classicalstatistical method, linear regression was employed to characterizemapping among seven input parameters and total charge passed ofconcrete. The result of multiple regression analysis is given as
Q ¼ �170:51� 4:85 � C � 11:02 � F � 46:14 � SFþ 34:42 �W
þ 33:03 � SP� 1:82 � CAþ 1:17 � FA ð4Þwhere Q is the total charge passed of concrete, measured incoulombs.
In addition, the second trial for the characterization of mappingamong input parameters was made by using nonlinear multipleregression analysis. Various different nonlinear equations weretried and the best equation was determined by considering relatedR2 and scatter plots between measured and calculated results. Thedetail of the best nonlinear regression equation is given as
Q ¼ �125;525:05� 47:53 � C � 72:26 � F � 106:35 � SF� 69:45 �W � 242:28 � SP þ 203:06 � CAþ 88:27 � S
þ 0:05 � C2 þ :11 � F2 þ 1:39 � SF2 þ 0:32 �W2 þ 16 � SP2
� :08 � CA2 � :03 � FA2 þ 0:13 � C � F þ 0:30 � C � SF� 0:04 � CA � FA ð5Þ
Table 3 shows the statistical parameters of ANNs and regressionmodels. It can be seen that the ANNs model gives a higher degreeof accuracy than the regression techniques. Furthermore, Figs. 5and 6 verify that the ANNs model is the most accurate of the threecompared paradigms for estimating the chloride ions permeabilityof high performance concrete.
5.2. Influence of relevant materials on chloride ions penetrability
The ANNs model was trained and tested on cases reflecting awide range of concrete mix proportions, validated on independenttest data, and compared with the models from regression tech-
Fig. 7. Influence of cement content on total char
Table 3Statistical parameters of neural networks and regression models
Statistical parameters ANNs model Line
Training set Testing set Train
MSE 0.00082 0.00301 0.00MAPE(%) 8.64 13.88 52.8R2 0.9782 0.9741 0.91
niques. The results show that the ANNs model is the most accurateamong the three compared paradigms for estimating the totalcharge passed of HPC. Although there are eight input parametersin the model, it is more meaningful to investigate the influenceof water and cement contents, water–binder ratio (W/B), fly ash-binder ratio (F/B), and silica fume-binder ratio (SF/B) the on dura-bility of HPC. The binder is a cementitous material, that is, cementplus fly ash and silica fume. The range of each variable is shown asfollows:
� The cement content was varied from 300 to 600 kg/m3.� The water content was varied from 110 to 200 kg/m3.� The water to binder ratio was varied from 0.25 to 0.5.� The fly ash-binder ratio (F/B) was varied from 0% to 50%, and
binder content was kept constant at 450 kg/m3.� The silica fume content was varied from 0% to 15%, and binder
content was kept constant at 450 kg/m3.
All other components or ratios were kept constant: SP contentswere kept constant at 1% of binder; the fine aggregate to coarseaggregate ratio by weight was kept at 0.67; entrapped air was keptat 1.00%; and the volume of concrete was 1.000 m3.
5.2.1. Influence of cement and water contentFig. 7 illustrates the variations in total charge passed with
increasing cement content at various levels of water content whichis produced by using the trained neural networks developed in thisstudy, and the response surface is shown in Fig. 8. For concretehaving cement content of 300–450 kg/m3, the higher the cementcontent, the lower the total charge passed, and for concrete havingcement content of 450–600 kg/m3, an increase in cement contentresults in a slight decrease of the total charge passed. As shownin Figs. 7 and 8, it was also found that the optimum cement contentfor design of HPC in terms of chloride penetration resistance ran-ged from 450 to 500 kg/m3.
ge passed at various levels of water content.
ar regression model Non-linear regression model
ing set Testing set Training set Testing set
89 0.0144 0.0039 0.04768 33.02 34.75 12.5941 0.3260 0.9647 0.8968
Fig. 8. Response surface of durability with the variations of W/B ratio and cement content.
Fig. 9. Influence of percent replacement of fly ash on relative charge passed.
R. Parichatprecha, P. Nimityongskul / Construction and Building Materials 23 (2009) 910–917 915
According to ASTM C1202, for water content of concrete greaterthan 150 kg/m3, the chloride ion penetrability of concrete contain-ing any level of cement content can be classified in the level of low
Fig. 10. Response surface of durability with the va
to high, and for water content of concrete lower than 150 kg/m3, itcan be classified in the level of negligible to low. Furthermore, it isof interest to note that the charge passed of concrete was found todecrease with decreasing water content.
5.2.2. Influence of percent replacement of fly ash and water–binderratio
Fig. 9 shows the variations in the total charge passed for HPCwhen increasing fly ash at different levels of W/B, produced byusing the trained neural network developed in this study. In addi-tion, response surface is shown in Fig. 10. The relative chargepassed means the percentage of total charge passed of concretecontaining fly ash to total charge passed of concrete without flyash. At a low level of W/B ratio (W/B = 0.25–0.4), the relative chargepassed significantly decreases when the replacement of fly ash isgreater than 20%, and the chloride ions penetrability of concretecontaining any level of fly ash replacement can be classified atthe level of negligible to low. However, at a higher level of W/B ra-tio (W/B = 0.45–0.5), the reduction in relative charge passed is pro-portional to the increase in percent replacement of fly ash, and thechloride ions penetrability of concrete containing any level of flyash replacement can be classified in the level of low to high.
riations of W/B ratio and fly ash-binder ratio.
916 R. Parichatprecha, P. Nimityongskul / Construction and Building Materials 23 (2009) 910–917
5.2.3. Influence of percent replacement of silica fume and water–binder ratio
Fig. 11 shows the variations in the total charge passed for HPCwith increasing silica fume at different levels of W/B, producedby using the trained neural networks developed in this study. Inaddition, response surface is shown in Fig. 12. The relative chargepassed means the percentage of total charge passed of concretecontaining silica fume to total charge passed of concrete withoutsilica fume. At a low level of W/B ratio (W/B = 0.25–0.35), the chlo-ride ions penetrability of concrete containing any replacement ofsilica fume can be classified at the level of negligible to very low(0–1000 C). However, at a higher level of W/B ratio (W/B = 0.4-0.5), the chloride ions penetrability of concrete containing anyreplacement of silica fume can be classified at the level of verylow to high (500–5000 C). For any level of water–binder ratio,when replacing cement with at least 5% of silica fume, the chlorideions penetrability of concrete can be classified at the level of neg-ligible to very low. Furthermore, it can be summarized that an in-crease in silica fume content results in significantly reducing thechloride ions penetrability to a higher degree when compared withthe fly ash results. It can be pointed out that silica fume is a veryfine particle and has higher chemical reactivity compared to ce-ment and fly ash.
Fig. 11. Influence of percent replacement of silica fume on relative charge passed.
Fig. 12. Response surface of durability with the v
6. Conclusions
The following conclusions were drawn from this investigation:
� The mean absolute percentage error (MAPE) of prediction of testresults was found to be 13.88 % and the absolute fraction of var-iance (R2) was 0.9741. The results indicate that the models arereliable, accurate, and illustrate how ANNs can be used to effi-ciently predict the chloride ions permeability across a widerange of ingredients of HPC.
� Based on the simulated total charge passed model built usingtrained neural networks, the optimum cement content fordesign of HPC in terms of total charge passed ranges from 450to 500 kg/m3.
� The chloride penetration resistance of concrete with any level ofwater–binder ratio can be significantly improved by using atleast 20% fly ash to replace cement.
� At any level of water–binder ratio, when replacing cement withat least 5% of silica fume, the chloride ions penetrability of con-crete can be classified at the level of negligible to very low.
� Increasing silica fume results in reducing the chloride ions pen-etrability to a higher degree than pulverized fly ash. It can bepointed out that silica fume is a very fine particle and has ahigher chemical reactivity compared to cement and fly ash.
� Although the capability of the proposed network is limited tothe data located within the available range of training data inthe database, the available range of the system could be easilyexpanded by retraining the neural networks with additionaldata from trial mixes.
References
[1] Neville AM. Properties of concrete. 4 ed. Essex: Longman Group Limited; 1995.[2] Metha PK, Aı̈tcin PC. Principles underlying the production of high-performance
concrete. Cem Concr Aggr 1990;12(2):70–8.[3] Alves MF, Cremonini RA, Dal Molin DCC. A comparison of mix proportioning
methods for high-strength concrete. Cem Concr Comp 2004;26(6):613–21.[4] Zain MFM, Islam MN, Basri IH. An expert system for mix design of high
performance concrete. Adv Eng Software 2005;36:325–77.[5] Yeh IC. Design of high-performance concrete mixture using neural networks
and nonlinear programming. J Comput Civil Eng. 1999;13(1):36–42.[6] Akkurt S, Ozdemir S, Tayfur G, Akyol B. The use of GA-ANNs in the modeling of
compressive strength of cement mortar. Cem Concr Res 2003;33:973–9.
ariations of W/B and silica fume-binder ratio.
R. Parichatprecha, P. Nimityongskul / Construction and Building Materials 23 (2009) 910–917 917
[7] Öztas� A, Pala M, Özbay E, Kanca E, Çagar N, Bhatti MA. Predicting thecompressive strength and slump of high strength concrete using neuralnetworks. Constr Build Mater 2006;21(2):384–94.
[8] Pala M, Özbay E, Öztas� A, Yuce MI. Appraisal of long term effects of fly ash andsilica fume on compressive strength of concrete by neural networks. ConstrBuild Mater 2007;20(9):769–75.
[9] ASTM C1202. Standard test method for electrical indication of concrete’sability to resist chloride ion penetration. In: Annual book of ASTM standards,vol. 04.02; 1997.
[10] AASHTO T277. Rapid determination of the chloride permeability of concrete,American Association of states highway and transportation officials, Standardspecifications –Part II Tests, Washington (DC); 1990.
[11] Wee TH, Suryavanshi AK, Tin SS. Influence of aggregate fraction in the mix onthe reliability of the rapid chloride permeability test. Cem Concr Res1999;21:59–72.
[12] Feng N, Feng X, Hao T, Xing F. Effect of ultrafine mineral powder on chargepassed of the concrete. Cem Concr Res 2002;32:623–7.
[13] Shi Caijun. Effect of mixing proportions of concrete on its electricalconductivity and the rapid chloride permeability test (ASTM C1202 orAASHTO 227) results. Cem Concr Res 2004;34:537–45.
[14] Zhao TJ, Zhou ZH, Zhu JQ, Feng NQ. An alternating test method for concretepermeability. Cem Concr Res 1998;28:7–12.
[15] Sirivivatnanon V, Cao HT. Binder dependency of durability properties of HPC.Canadian international symposium of HPC and reactive powder concrete,Canada; 1998. p. 227–40.
[16] Yeh IC. Analysis of strength of concrete using design of experiments and neuralnetworks. J Matr Civil Eng. 2006;18(4):598–604.
[17] Rumellhert D, Hinto G, Williams R. Learning internal representations by errorpropagation. Cambridge, Mass: MIT Press; 1986.
[18] Kasperkiewicz J, Racz J, Dubrawski A. HPC strength prediction using artificialneural networks. J Comput Civil Eng. 1995;9(4):279–84.
[19] Parichatprecha R, Subeidi BP, Nimityongskul P. Influence of pozzolanicmaterials and cement content on the charge passed of high strength anddurable concrete. In; Proceeding of the 8th of the ferrocement and thinreinforced cement composites, Thailand; 2006. p. 535–44.
[20] Chindaprasirt P, Chotithanorm C, Cao HT, Sirivivatnanon V. Influence of fly ashfineness on chloride penetration of concrete. Constr Build Mater2007;21(2):356–61.
[21] Hagen Martin T, Demuth Howard B, Beale Mark. Neural networkdesign. Boston: PWS Publishing, Co.; 1995.