Identification of Water/Cement Ratio of Cement Pastes ... - Springer Link

KSCE Journal of Civil Engineering (2013) 17(4):763-768 DOI 10.1007/s12205-013-0156-9

Structural Engineering

www.springer.com/12205

Identification of Water/Cement Ratio of Cement Pastes, Basing on the Microstructure Image Analysis Data and using Artificial Neural Network Ali Ugur Ozturk* and Okan Onal** Received May 26, 2011/Revised January 16, 2012/Accepted July 29, 2012

···································································································································································································································

Abstract Artificial Neural Network (ANN) analysis has been established to forecast the Water/Cement (w/c) ratio values of cement pastes by using image analysis techniques in the scope of this study. W/c ratio values have reasonably great effects on the performance of cement based structural members. The service life or ultimate performances such as strength and durability characteristics are strongly affected by w/c ratios of cementitious materials. In this study, the relationship between microstructural phases such as unhydrated cement part, hydration products, capillary porosity, and w/c ratios predicted by ANN analysis, has been established. The predicted values are compared with estimated values obtained by proposed method in the literature. The study indicated that, using a contemporary data analysis technique, which is capable of searching nonlinear relationships more thoroughly, would result in more realistic prediction of the w/c ratios compared to the proposed method. Keywords: paste, backscattered electron imaging, microstructure; compressive strength, artificial neural network ··································································································································································································································· 1. Introduction

Cement is the most widely used structural material in construction technology. Cementitious materials such as concrete are basically a mixture of aggregates and paste. The paste, comprised of cement and water, binds aggregates into a rocklike mass as the paste hardens because of the chemical reaction of the cement and water (Kosmatka et al., 2002). Furthermore the knowledge of mix design properties has a remarkable role on the performance in service life. Proportions of cement, water and aggregates in concrete consist traditional concrete mixture design values (ACI Committee 318, 2002; ACI 211.4R, 1993; ACI, 1996). The amounts of cement, water, fine and coarse aggregates are determined respectively. Proportions of these basic ingredients are the most important parameters in service life performance in the view of material concept. The knowledge of the performance of cementitious materials is a great issue to realize the usage limits of cement based structural systems. Properties such as strength and durability (i.e., abrasion resistance, freeze-thawing resistance and alkalisilica reaction potential) are most important factors affecting this performance in service life of a structural member. These two factors depend on some characteristics of cement-based structural members such as Water/Cement ratio (w/c) and porosity. Especially w/c ratio has significant effects on most properties of cementitious materials (Neville, 1995).

Thus, it is difficult to ascertain the w/c ratio after the cementitious materials is hardened; w/c ratio indicates the quality of members produced by cementitious materials. To determine w/c ratio of cement paste is easy by using new methods depending on microstructural studies (Wong and Buenfeld, 2009). Microstructural characteristics of cementitious materials can be investigated more easily by using new computer hardware and software technologies. Microstructural characterizations are required for detailed inner structure analysis and allow modifications of macro properties. Image analysis of micrographs of cementitious materials bring benefits for quantifying microstructural characteristics of cement pastes such as porosity, pore structure and phases including undifferentiated hydration products and anhydrous cement content (Scrivener et al., 1987; Zhao and Darwin, 1992; Kjeilsen et al., 1991; Lange et al., 1994). Undifferentiated hydration product indicates the rest of the hydration products except phases such as calcium hydroxide, anhydrous cement, pores and aggregates without discrimination. So, in order to determine w/c ratios of cement pastes, it is necessary to implement current experimental technologies and analyze their results. Recently, artificial neural networks and fuzzy logic approaches, a subfield of intelligent systems, are being widely used to solve a wide variety of problems in civil engineering applications (Ince, 2004; Adhikary and Mutsuyoshi, 2006; Kewalramani and Gupta, 2006; Alshihri et al., 2009; Onal

*Assistant Professor, Dept. of Civil Engineering, Faculty of Engineering, Celal Bayar University, 45140 Manisa, Turkey (Corresponding Author, E-mail: [email protected]) **Assistant Professor, Dept. of Civil Engineering, Faculty of Engineering, Dokuz Eylul University, 35160 Izmir, Turkey (E-mail: [email protected]) − 763 −

Ali Ugur Ozturk and Okan Onal

and Ozturk, 2010). The importance and applications of this technique increased significantly during last years (Lefik et al., 2009). Nowadays, ANNs are successfully used in computeraided management, system identification, modeling of different physical dynamic processes depending on several fuzzy variables. Now, it was possible to quantify the effect of age, w/c, coarse aggregate volume percentage, and curing temperature as main parameters in the evolution of concrete elastic modulus. It was concluded that with no need of complex modeling procedures, early-age concrete elastic modulus can be properly modeled by using an ANN technique based on original mixture proportions (Venkiteela et al., 2010). In the scope of this study, w/c ratios of cement pastes have been predicted by ANN analysis using amounts of microstructural phases such as unhydrated cement (uc), undifferentiated hydration product (hp) and capillary pores (cp) achieved by digital image processing techniques. Cement pastes were prepared with w/c ratios; 0.3, 0.4 and 0.5. Flat sections of cement pastes covered with polyester film providing smooth and flat surfaces, were investigated by using Scanning Electron Microscope (SEM) in backscattered electron mode (BE) at different ages (1, 2, 7, 28 and 90 days) to determine the amounts of microstructural phases. The predicted results of ANN model were compared to estimated w/c ratio values calculated by proposed model by Wong and Buenfeld (2009). It has been observed that, the performance of the ANN analysis is superior to the proposed model by Wong and Buenfeld (2009) based on the calculated errors for the w/c ratios.

2. Experimental Part In order to predict the w/c ratio of cement pastes using ANN, microstructural investigation under SEM in BE mode has been performed. Cement pastes were prepared with three different w/c ratios (0.3, 0.4, 0.5). Specific gravity is 3.15 for CEM I 42,5 R type cement. The mineral composition of cement is 69.58% C3S, 9.2% C2S, 6.74% C3A and 10.74% C4AF. Cement pastes were cast into plastic containers with dimensions 2,5 cm in width and 10 cm in length, and compacted in two equal layers. After 1, 2, 7, 28 and 90 days of curing at 20° to 22° C ambient laboratory temperature, hydration processes of the mixes were terminated by submerging specimens into isopropyl alcohol for five days. Before microstructure studies, the specimens were impregnated by a polyester resin, and then the surface of each specimen was polished. Each specimen was sanded by 600 and 1200 grid sandpapers. Sanding should follow sequence of 9, 3, 1, and 0.25 µm diamond paste for 120 s (Stutzman and Clifton 1999). Microstructural investigations require qualified expert practice. Microstructural phases of cementitious materials should be segmented with accuracy and consistency. During imaging under SEM, brightness and contrast setting were done, and the histogram of each micrograph was checked to be centered and stretched as to cover the whole dynamic range of the current grey scale (Ozturk and Baradan, 2008). After the suitable settings for

Fig. 1. Micrographs of Cement Pastes: (a) With w/c 0.5 by 500X Magnification at 1 Day, (b) With w/c 0.5 by 500X Magnification at 2 Day, (c) With w/c 0.5 by 500X Magnification at 7 Day, (d) With w/c 0.5 by 500X Magnification at 28 Day, (e) With w/c 0.5 by 500X Magnification at 90 Day

brightness and contrast were found, all 300 images (5 images for each sample) were taken under these arrangements to avoid errors occurred by imaging process (Fig. 1). Accurate analyses lead to meaningful quantitative data that can be used for comparative studies and to characterize a relationship between microstructure and mechanical behavior. The phases in a polished section can be segmented by their grey level thresholds in the micrograph (Fig. 2). The grey level values of phases compose separate peaks in the grey level histogram with their heights proportional to the relative fractions of each phase. To define the amounts of area ratio of each phase (Fig. 3), segmentation analysis was conducted for each micrograph. Phases were separated into four groups, these are pores, undifferentiated hydrates, unhydrated cement, calcium hydroxide (Ca(OH)2). The volume fractions of microstructural phases (unhydrated cement, undifferentiated hydration products and capillary pores) were determined by image analysis on backscattered electron

− 764 −

KSCE Journal of Civil Engineering

Identification of Water/Cement Ratio of Cement Pastes, Basing on the Microstructure Image Analysis Data and using Artificial Neural Network

Fig. 2. Grey Value Histogram of Cement Mortar at 28 days Fig. 3. Microstructural Phases of a Cement Mortar

images taken by SEM at 500x magnification (Wong and Buenfeld, 2009). Grey value histograms were determined for each micrograph and all amounts of phases were taken into account to image based proposed method in the literature (Wong et al., 2006; Scrivener et al., 1987). The volumetric ratio of hydration products of to the reacted cement, named dV was determined 2.06 in this study calculated by the proposed method by Wong and Buenfeld (2009). Estimated w/c ratio values and amounts of volume fractions required microstructural phases are given below in Table 1, where w/c and Ew/c are actual and estimated (by proposed method by Wong and Buenfeld, 2009) values, and where Vah, Vhp and Vp indicate anhydrated part, hydration product and capillary porosity, respectively. The values were obtained for each investigation day (1, 2, 7, 28 and 90 days) of samples with different w/c ratios.

3. Artificial Neural Network Analysis In order to obtain a more realistic prediction of the w/c ratios compared to the model proposed by Wong and Buenfeld, a contemporary data analysis technique, which is capable of searching nonlinear relationships more thoroughly, have been employed. ANNs are non-linear statistical data modeling tools, which may be used to model complex relationships between inputs and outputs. This technique allows investigation of the relationship between several visual features of cement paste and actual w/c ratio by simulating the structure of biological neural networks. ANN consists of an interconnected group of artificial neurons and processes information using a connectionist approach to computation. ANN is an adaptive system that changes its

Table 1. Volumetric Fractions of Phases and w/c Ratios of Cement Pastes Day 1 2 7 28 90 1 2 7 28 90 1 2 7 28 90 1 2 7 28 90

w/c 0,5 0,5 0,5 0,5 0,5 0,5 0,5 0,5 0,5 0,5 0,5 0,5 0,5 0,5 0,5 0,5 0,5 0,5 0,5 0,5

Ew/c 0,48 0,51 0,49 0,49 0,48 0,49 0,48 0,49 0,48 0,49 0,47 0,49 0,48 0,47 0,49 0,51 0,49 0,48 0,48 0,49

Vol. 17, No. 4 / May 2013

Vah 19,68 14,79 11,26 8,64 5,46 18,66 15,64 10,27 8,57 3,43 20,97 16,86 11,07 9,55 4,12 17,25 13,45 11,36 8,60 4,00

Vhp 45,30 47,90 53,40 57,40 60,02 44,20 47,59 51,36 57,10 61,45 46,45 48,24 52,85 58,60 62,15 43,80 49,05 54,40 56,84 61,15

Vp 38,40 35,20 28,70 25,60 20,30 37,90 32,87 26,75 24,32 18,60 39,20 36,05 27,16 24,90 19,80 38,05 31,07 27,90 24,30 19,40

w/c 0,4 0,4 0,4 0,4 0,4 0,4 0,4 0,4 0,4 0,4 0,4 0,4 0,4 0,4 0,4 0,4 0,4 0,4 0,4 0,4

Ew/c 0,38 0,4 0,41 0,39 0,39 0,39 0,41 0,38 0,38 0,39 0,4 0,39 0,39 0,38 0,37 0,41 0,4 0,38 0,39 0,4

Vah 26,12 20,62 14,23 12,59 10,06 26,11 20,91 19,01 17,48 10,74 28,18 26,70 18,59 16,62 12,07 25,88 20,80 17,54 14,19 10,61

− 765 −

Vhp 47,20 49,25 54,60 58,90 63,20 46,15 48,76 50,72 55,10 60,70 46,90 47,98 53,24 59,20 62,45 45,80 50,24 55,18 57,60 62,15

Vp 33,20 29,60 23,40 19,20 16,45 34,62 31,29 25,03 23,47 17,10 38,74 35,46 26,05 22,68 16,20 37,29 29,90 23,55 21,04 18,30

w/c 0,3 0,3 0,3 0,3 0,3 0,3 0,3 0,3 0,3 0,3 0,3 0,3 0,3 0,3 0,3 0,3 0,3 0,3 0,3 0,3

Ew/c 0,31 0,32 0,31 0,29 0,32 0,31 0,32 0,3 0,29 0,32 0,31 0,31 0,3 0,31 0,32 0,3 0,29 0,3 0,31 0,32

Vah 31,07 25,11 22,87 20,57 15,42 33,89 30,54 27,81 26,60 18,03 40,06 37,32 35,84 28,59 22,06 41,69 35,38 28,22 23,27 18,47

Vhp 44,16 46,10 49,70 53,90 55,12 43,60 47,92 51,38 54,19 61,24 44,23 48,10 55,29 60,12 62,71 42,19 48,50 53,56 56,12 60,94

Vp 27,50 23,15 19,36 14,05 13,24 30,23 28,46 22,37 19,43 15,60 36,10 33,28 29,56 24,30 19,53 35,85 27,76 22,61 19,41 16,05


structure based on information that flows through the network during the learning phase (Onal and Ozturk, 2010). Feed forward back propagation technique is the preferred method in function approximation problems (Werbos, 1994). Therefore, the prediction of the actual w/c ratio values of the cement paste specimens has been established using the algorithm of this neural network approach. The back-propagation idea was first developed by David Rumelhart, Geoffrey Hinton and R.J. Williams in 1986. The Multilayer perceptron neural network architecture typically incorporates the back propagation technique that uses the gradient descent algorithm. Back-propagation means the computation of derivatives of the sum-of-squares error function with respect to the weight estimates with the gradient descent method used to determine a smaller error as the number of iterations increases (Matignon, 2005). In this study, the calculations and ANN model were established for samples with three different w/c ratios for five different examination ages. At this point, the initial w/c ratio can be obtained by using different microstructural properties formed at different ages. In fact we have three different sets with three different w/c ratios, and so ANN model always determines one initial w/c ratio for each set. The actual w/c ratios and the features derived from the digital image processing operations (Vah, Vhp and Vp) have been used as the data set in the neural network analysis. For the evaluation of the ANN model, a widely used cross validation technique, called as holdout method has been selected. The holdout method partitions the data into two subsets called a training set and a test set or holdout set. The network fits a function using the training set only. Then the trained network is asked to predict the output values for the data in the testing set. These output values have never been evaluated by the network before. The errors between the actual and predicted values are used to evaluate the model in this methodology. It is common to designate the 2/3 of the data as the training set, and the remaining 1/3 part as the test set (Selver et al., 2008). However, the half of the data set (i.e., belonging to 30 test specimens) has been employed as the training set in the supervised learning process in recent study, in order to have enough data allocated to the testing data set. The odd numbered data rows have been used as the training set in the setup of the network structure. The remaining data have been employed as the test set. Since randomized data selection has not been used, each individual runs of the neural network can be evaluated in order to judge the success of the neural network. However, since the initial weight factors are selected in a randomized manner, each individual run of the training algorithm resulted with slightly varying outputs. Since the input data set consists of input and target values ranging between different upper and lower limits, a scaling approach has been applied to the data. This procedure normalizes the inputs and targets so that they will have zero mean and unity standard deviation, respectively. The transformation vectors contain the mean and standard deviations of the original inputs and targets. After the network has been trained, these vectors

should be used to transform any future inputs that are applied to the network. Namely, the output of the network is trained to produce outputs with zero mean and unity standard deviation. In order to convert these outputs back in to the same units that were used for the original targets, same vectors have to be used in the transformation process. These vectors effectively become a part of the network, just like the network weights and biases. This preprocessing approach maps the data in a new form, more suitable to train the network. The success of the neural network analysis has been increased by performing principle component analysis, which captures the variance in a data set in terms of principle components. The principal component analysis has been conducted using Matlab’s Statistics Toolbox functions. The principal component analysis approach searches an orthogonal set of vectors that maximize the variance of the projected data. Principal component analysis is a linear projection of the data into another frame of reference with minimum error in order to extract relevant information from high dimensional data sets (Onal and Ozturk, 2010). The preprocessed data set has been used in the gradient descent with adaptive learning rate back propagation algorithm offered by the neural network toolbox of the MatLab Technical Computing Language. In practice, networks with several hidden layers distribute the invariance task throughout the net. Naturally, the weight initialization, learning rate, data preprocessing arguments apply to these networks too. It has been found empirically that networks with multiple hidden layers are more prone to getting caught in undesirable local minima. In the absence of a problemspecific reason for multiple hidden layers, then, it is simplest to proceed using just a single hidden layer (Duda et al., 2000). The selection of the number of the neurons in the hidden layer has been determined by a trial and error procedure. Thus, the topology of the network has been organized to consist of one hidden layer with ten neurons, having tangent sigmoid function, and one output layer, having linear function as in the general function approximation applications (Fig. 4). The performance improvement of the network during training has been monitored using the performance tool of the MatLab Technical Programming Language. After several runs of the training algorithm, it has been observed that best training

− 766 −

Fig. 4. ANN Architecture KSCE Journal of Civil Engineering

Identification of Water/Cement Ratio of Cement Pastes, Basing on the Microstructure Image Analysis Data and using Artificial Neural Network

performance (i.e., minimum value of the mean squared error) has been achieved at around 250 epochs. Therefore, the training has been limited with 250 epochs. The number of epoch was predefined as 250, in order to avoid over-learning, because, a decrease in errors over the training data does not mean a decrease over novel data and in fact may lead to less generalization. Overlearning results in an over fitted function, which has good approximation for the training set, but poor approximation for the test set. The test data set have been normalized using the pre-calculated mean and standard deviation of the training set after the completion of the network training process. The normalized test data set have been preprocessed by applying the same principal component transformation that was previously computed for the training data. Thus, the test data set has been converted to the same units that the network training has been processed. The preprocessed test data has been employed in the simulation of the ANN model. However, in order to evaluate the success of the ANN model, the output of the network must be converted to the same quantity, as the raw data before the preprocessing of the data set. Therefore, the same transformation matrices used in the preprocessing of the train and test data (i.e., principle component

Table 2. Comparison of the Mean % Errors by Wong & Buenfeld and ANN Analysis

0.3 w/c Ratio 0.4 w/c Ratio 0.5 w/c Ratio

Mean % Error by Wong & Buenfeld 3.10 3.59 4.44

Mean % Error by ANN Analysis 1.83 2.91 1.64

analysis and then normalization), have been used in the post processing operations (Fig. 5). In order to interpret the results of the ANN analysis, the training set has been also simulated by using the established network and outputs have been marked in the Fig. 6. The performance of the network has been evaluated calculating the mean % errors between the actual and predicted w/c ratio values for 0.3, 0.4 and 0.5 (Table 2). It should be noted that, relatively low % errors for the predictions of the w/c ratios according to the Wong and Buenfeld model have been achieved. The results of the ANN analysis showed that a good agreement between microstructural characteristics and actual w/c ratio of cement pastes can be establish by using feed forward backpropagation algorithm.

4. Conclusions

Fig. 5. Schematic Presentation of the Data Analysis Process

Fig. 6. Relationship between Predicted and Actual Values of w/c Ratio of Cement Paste Vol. 17, No. 4 / May 2013

Cementitious materials are well known and the most used materials in the production of structural members. Thus cement is the worldwide used construction material; the usage limits and the performance of structural members comprised with cement are not well known. Therefore the knowledge of mix design properties has a remarkable role on the performance in service life. Sometimes, in order to realize and determine this performance, side engineers need to get the knowledge of mix design and properties such as mineral additives amount, w/c ratio or presence chemical admixtures of concrete members before casting. Water to cement ratio is one of these properties to realize and produce an idea about the ultimate performance for a structural member. Currently, engineers and scientists have opportunities to perform some new experimental techniques such as microstructural analysis using optical or scanning electron microscopy in order to realize these ultimate performances. Microstructural phases such as undifferentiated hydrated part and pores formed during hydration process can easily be investigated. Developments of such phases and their effects on the performance (strength and durability characteristics) in the service life can be researched. In the scope of this study, the actual w/c ratios affecting on the performance of cementitious materials were tried to be predicted by using ANN values. The predicted values are also being compared with estimated values by proposed method using volumetric fractions of microstructural phases of cement pastes in the literature. The ANN analysis indicates a great potential on

− 767 −


forecast of w/c ratios of cement pastes. Relatively low % errors for the predictions of the w/c ratios according to the Wong and Buenfeld model have been achieved. Future studies using new sample sets would be beneficial for the validation of the proposed w/c ratio prediction methodology. It has to be mentioned that, the methodology, proposed in this study, may produce inappropriate results under conditions such as carbonation, alkali-silica reaction especially when the microstructure has some different phases with different chemical formations. However studies on the cementitious materials, subjected to these aggressive environmental conditions, require suitable micrographs which reflect different microstructures compared to the reference samples used in current work. Furthermore, a good image analysis procedure must be performed to get quantitative parameters of these microstructural phases. By implementing these proficiencies, the proposed ANN methodology may overcome different aggressive environmental conditions. Also, the ANN analysis has the ability to adapt itself to these new experimental sets, which is a major advantage over Wong and Buenfeld method.

References ACI Committee 318 (2002). Building code requirements for structural concrete, (ACI 3118-02) and Commentary (ACI 318-02), ACI, Detroit. ACI Standard 211.1 (1996). Standard practice for selecting proportions for normal, heavyweight, and mass concrete, ACI Manual of Concrete Practice, Part 1. ACI 211.4R (1993). Guide for selecting properties for high strength concrete with Portland cement and fly ash, ACI, Detroit. Adhikary, B. B. and Mutsuyoshi, H. (2006). “Prediction of shear strength of steel fiber RC beams using neural networks.” Construct. Build Mater., Vol. 20, No. 9, pp. 801-811. Alshihri, M. M., Azmy, M., and El-Bisy, M. S. (2009). “Neural networks for predicting compressive strength of structural light weight concrete.” Construct. Build Mater., Vol. 23, No. 6, pp. 22142219. Duda, R. O., Hart, P. E., and Stork, D. G. (2000). Pattern classification (2nd ed.), John Wiley & Sons. Ince, R. (2004). “Prediction of fracture parameters of concrete by artificial neural networks.” Eng. Fract. Mech., Vol. 71, No. 15, pp. 2143-59. Kewalramani, A. M. and Gupta, R. (2006). “Concrete compressive strength prediction using ultrasonic pulse velocity through artificial neural networks.” Auto. Construct., Vol. 15, No. 15, pp. 374-379. Kjeilsen, K. O., Detwiler, R. J., and Gjurv, O. E. (1991). “Backscattered electron image analysis of cement paste specimens; Specimen preparation and analytical methods.” Cem. Conc. Res., Vol. 21, No.

2, pp. 388-390. Kosmatka, S. H., Kerkhoff, B., and Panarese, W. C. (2002). Design and control of concrete mixtures, Editors: Steven H. Kosmatka, Beatrix Kerkhoff, and Willim C. Panarese, 14th ed., Portland Cement Association. Lange, D. A., Jennings, H. M., and Shah, S. P. (1994). “Image analysis techniques for characterization of pore structure of cement based materials.” Cem. Conc. Res., Vol. 24, No. 5, pp. 841-853. Lefik, M., Boso, D. P., and Schrefler, B. A. (2009). “Artificial neural networks in numerical modelling of composites.” Comput. Methods. Appl. Mech. Eng., Vol. 198, Nos. 21-26, pp. 1785-1804. Matignon, R. (2005). Neural network modeling using sas enterprise miner, pp. 90. Neville, A. M. (1995). Properties of concrete, Fourth Edition Addison Wesley Longman Ltd. Onal, O. and Ozturk, A. U. (2010). “Artificial neural network application on microstructure-compressive strength relationship of cement mortar.” Advances in Engineering Software, Vol. 41, No. 2, pp. 165169. Ozturk, A. U. and Baradan, B. (2008). Porosity and mechanical properties of cement mortars based on Microstructural investigation/ Computational Materials, 978-1-60456-896-7, Nova Science Publishers, New York. Scrivener, K. L., Patel, H. H., Pratt, P. L., and Parrott, L. J. (1987). “Analysis of phases in cement paste using backscattered electron images. In: L.J. Struble, P.W. Brown, editors. Microstructural development during hydration of cement.” Mater. Res. Soc. Symp. Proc., Vol. 85, pp. 67-76. Selver, M., Ardali, E., Önal, O., and Akay, O. (2008). “Predicting uniaxial compressive strengths of brecciated rock specimens using neural networks and different learning models.” 23rd International Symposium on Computer and Information Sciences, I·STANBUL, October. Stutzman, P. E. and Clifton, J. R. (1999). “Sample preparation for scanning electron microscopy.” In: Proceedings of 21st International Conference - Cement Microscopy, Las Vegas, pp. 10-22. Venkiteela, G., Gregori, A., Sun, Z., and Shah, S. P., (2010). “Artificial neural network modeling of the early-age dynamic Young's modulus of normal concrete.” ACI Materials Journal, Vol. 107, No. 3, pp. 282-290. Werbos, P. J. (1994). The roots of backpropagation: From ordered derivatives to neural networks and political forecasting, John Wiley and Sons, New York, Chichester, Brisbane, Toronto and Singapore. Wong, H. S. and Buenfeld, N. R. (2009). “Determining the watercement ratio, cement content, water content and degree of hydration of hardened cement paste: Method develop and validation on paste samples.” Cem. Concr. Res., Vol. 39, No. 10, pp. 957-965. Wong, H. S., Head, M. K., and Buenfeld, N. R. (2006). “Pore segmentation of cement-based materials from backscattered electron images.” Cem. Concr. Res., Vol. 36, No. 6, pp. 1083-1090. Zhao, H. and Darwin, D. (1992). “Quantitative backscattered electron analysis for cement paste.” Cem. Concr. Res., Vol. 22, No. 4, pp. 695-706.

− 768 −

KSCE Journal of Civil Engineering