You are on page 1of 6

Materials and Design 30 (2009) 764769

Contents lists available at ScienceDirect

Materials and Design


journal homepage: www.elsevier.com/locate/matdes

Articial neural network approach to predict the mechanical properties of CuSnPbZnNi cast alloys
Mehmet Sirac Ozerdem a,*, Sedat Kolukisa b
a b

Faculty of Engineering and Architecture, Department of Electrical and Electronics Engineering, Dicle University, Diyarbakir 21680, Turkey Faculty of Engineering and Architecture, Department of Mechanical Engineering, Dicle University, Diyarbakir 21680, Turkey

a r t i c l e

i n f o

a b s t r a c t
In this study, an articial neural network approach is employed to predict the mechanical properties of CuSnPbZnNi cast alloys. In articial neural network (ANN), multi layer perceptron (MLP) architecture with back-propagation algorithm is utilized. In Articial Neural Network training module, CuSn PbZnNi (wt%) contents were employed as input while yield strength, tensile strength and elongation were employed as outputs. ANN system was trained using the prepared training set (also known as learning set). After training process, the test data were used to check system accuracy. As a result of the study neural network was found successful for the prediction of yield strength, tensile strength and elongation of CuSnPbZnNi alloys. 2008 Elsevier Ltd. All rights reserved.

Article history: Received 9 February 2007 Accepted 7 May 2008 Available online 17 May 2008 Keywords: Articial neural network Prediction of mechanical properties CuSnPbZnNi cast alloys

1. Introduction The mechanical properties of cast alloys mainly depend on the weight fractions of alloying elements, applied heat treatments, microstructures, morphologies of the various phases constituting. Copper is alloyed with other elements because pure copper is extremely difcult to cast as well as being prone to surface cracking, porosity problems and to the formation of internal cavities. The casting characteristics and mechanical properties of copper can be improved by the addition of small amounts of elements including beryllium, silicon, nickel, tin, zinc, chromium, and silver [1]. Copper alloys are poured into many types of casting such as sand, shell, investment, permanent mold, chemical sand, centrifugal, and die. The copper-base casting alloy family can be subdivided into three groups according to solidication (freezing) range. In this study group III Cu cast alloys were employed which have a freezing range well over 110 C, even up to 170 C. Alloy type, chemical compositions and UNS codes were presented in Table 1 [2]. Effects and relationships of alloying elements on mechanical properties have always been studied for each alloying element. Repeated tests are employed to observe the effect of alloying elements on the materials behavior, since these parameters are essential for dimensional approaches. Recently with the developments in articial intelligence; researchers have a great deal of attention to the solution of non-

linear problems in physical and mechanical properties of metal alloys [3]. In this study, features of multi layer perceptron architecture with back-propagation learning algorithm [4] were employed to predict the mechanical properties, namely, tensile strength, yield strength, elongation of CuSnPbZnNi (group III) cast alloys. 2. Articial neural network (ANN) model An ANN is a mathematical model consisting of a number of highly interconnected processing elements organized into layers, the geometry and functionality of which have been likened to that of the human brain. ANNs learn by experience, generalize from previous experiences to new ones, and can make decisions [5]. Neural elements of a human brain have a computing speed of a few milliseconds, whereas the computing speed of electronic circuits is on the order of microseconds. The ANNs are parallel process elements which has characteristic such as (a) ANN is a mathematical model of a biological neuron, (b) ANN has various process elements which are related to each other, (c) ANN keeps knowledge with connection weights. Neural network models provide an alternative approach to implementing enhancement techniques. A simple process element of the ANN is given in Fig. 1. The network has one input layer, one hidden layer and one output layer. The input layer consists of all the input factors. Information from the input layer is then processed in the course of one

* Corresponding author. Tel.: +90 412 24 88 403; fax: +90 412 24 88 405. E-mail addresses: sozerdem@dicle.edu.tr (M.S. Ozerdem), kolukisa@dicle.edu.tr (S. Kolukisa). 0261-3069/$ - see front matter 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.matdes.2008.05.019

M.S. Ozerdem, S. Kolukisa / Materials and Design 30 (2009) 764769 Table 1 Nominal chemical composition of group III copper cast alloys Alloy type UNS No Leaded red brass C83450 C83600 C83800 C84400 C84800 C90300 C90500 C90700 C91100 C91300 C92200 C92300 C92600 C92700 C92900 C93200 C93400 C93500 C93700 C93800 C94300 Composition (%) Cu 88 85 83 81 76 88 88 89 84 81 86 87 87 88 84 83 84 85 80 78 70 Sn 2.5 5 4 3 2.5 8 10 11 16 19 6 8 10 10 10 7 8 5 10 7 5 Pb 2 5 6 7 6.5 1.5 1 1 2 2.5 7 8 9 10 15 25 Zn 6.5 5 7 9 15 4 2 4.5 4 2 3 1 Ni 1 3.5

765

Leaded semired brass Tin bronze

Leaded tin bronze

High-leaded tin bronze

Fig. 2. The structure of three layered neural network proposed in present study.

weights are adjusted to minimize the error between the network output and actual value. Among the various kinds of ANN approaches that exit, the multi layer perceptron (MLP) architecture with back-propagation learning algorithm has become the most popular in engineering applications [9,10]. Back-propagation algorithm, which is common in literature, has been used to update the forward path parameters in ANNs. This method is based on minimization of the quadratic cost function by tuning the network parameters. The mean square error is considered as a measurement criterion for a training set. Parameters which minimize this cost function are determined. The averaged square error is given by Eqs. (1) and (2)
Fig. 1. The mathematical model of neuron.

ej n dj n yj n
hidden layer, following output vector is computed in the nal (output) layer. A schematic description of the layers is given in Fig. 2. In developing an ANN model, the available data set is divided into two sets, one to be used for training of the network, an the remaining is to be used to verify the generalization capability of the network [68]. Inputoutput pairs are presented to the network and
P 1 XX 2 e n en 2P j2C n1 j

1 2

In these equations e, n, d, y, P and C indicate error signal at the output, iteration number, desired output, generated output by network, total number of patterns contained in the training set and

Fig. 3. General architecture of the learning and predicting system.

766

M.S. Ozerdem, S. Kolukisa / Materials and Design 30 (2009) 764769

g indicates learning-rate parameters and has different values in different problems. The adjustment of synaptic weight coefcients between input layer and hidden layer are given by Eq. (5)
dj n u
0 m X i0

! wji nyi n

X
k

dk nwkj n

In case of the network does not converge, the formula, which is used to determine weight coefcients including a momentum parameter been generalized [11] and has been given in Eq. (6)

Dwji n aDwji n 1 gdj nyi n

The general architecture of learning and predicting the mechanical properties system is given in Fig. 3.

3. Collecting the experimental data and working platform Chemical compositions, mechanical properties were collected from related standards and ASM Hand Books [2]. MATLAB platform was used to train and test the ANN. In the training, increased number of neurons (714) in the hidden layer has been used in order to dene the output accurately. After training the network successfully, it has been tested by using the known data. Statistical methods were used to compare the results produced by the network. Errors occurring at the learning and testing stages are called the root-mean squared (RMS), absolute fraction of variance (R2), and mean percentage error (MPE) values. They are dened as follows, respectively !1=2 X jt j oj j2 7 RMS 1=p ! 2 j t j oj R2 1 P 2 j oj  X tj oj 1 100 MPE tj p j P
j

Fig. 4. Performance changing of neural network in training stage.

Table 2 MLP architecture and training parameters The number of layers The number of neuron on the layers The initial weights and biases Activation functions for hidden and output layers Training parameters Learning rule Adaptive learning-rate for hidden layer Adaptive learning-rate for output layer Number of iteration Momentum constant Duration of learning time Acceptable mean-squared error 3 Input: 5, Hidden: 10, Output: 3 Randomly between 1 and +1 Log-sigmoid Back-propagation From 0.9 to 0.887583 From 0.7 to 0.690342 13,893 0.95 2 min 27 s 0.001

8 9

number of neurons at output layer respectively. The adjustment of synaptic weights between hidden layer and output layer is given by Eqs. (3) and (4)

4. Evaluation of results and discussions The aim of employing an ANN model is to predict the mechanical properties, namely yield strength, tensile strength and elongation of group III CuSnPbZnNi cast alloys. The network has ve input parameters: CuSnPbZnNi (wt%), contents and three

Dwji n gdj nyi n


dj n ej nu0
m X i0

3 4

wji nyi n

Fig. 5. (a) Comparison (b) scatter plot, for the training stage of yield strength.

M.S. Ozerdem, S. Kolukisa / Materials and Design 30 (2009) 764769

767

Fig. 6. (a) Comparison, (b) scatter plot, for the training stage of tensile strength.

output parameters: yield strength, tensile strength, and elongation. So, the architecture of ANN becomes 5-10-3, 5 corresponding to the input values, 10 to the number of hidden layer neurons and 3 to the outputs. The experimental data set includes 21 patterns, of which 13 patterns were used for training the network and 8 patterns were selected randomly to test the performance of the trained network. All the input and output values were normalized between 0.1 and 0.9 by using linear scaling. The log-sigmoid transfer function was used in the hidden and output layer. During the training period, the averaged square error decreased with increasing number of iteration. After 13,893 training cycles, signicant effect on error reduction has not been traced. The performance changing of ANN in training stage is given in Fig. 4. With the increasing number of reliable input data, improves the integrity of the training session and target outputs. The MLP architecture and training parameters used in learning stage for ANN structure are presented in Table 2. A comparison of the measured

reveal that the well-trained network model has great accuracy in predicting yield strength, tensile strength and elongation. The decision as to the number of neurons used in the hidden layer usually depends on the arithmetical mean of the number of inputs and outputs. In this application 714 hidden layers were employed to test. The algorithm with 10 hidden layer neurons is suggested to be used in present application. Weights between input layer and hidden layer, produced for the present outputs, namely, yield strength, tensile strength and elongation were presented in Table 3. The activation function used in this study is given as follows

Mi

1 1 eEi

10

where Ei is the weighted sum of the input depending on percentage of Cu, Sn, Pb, Zn and Ni. The constants for the calculation of Mi were taken from Table 3. The outputs are calculated as follows

1 1 e1:26M1 0:11M2 5:54M4 2:63M5 2:10M6 1:24M7 3:35M8 1:75M9 2:86M10 4:28 1 Yield strength 1 e0:4M1 2:76M2 8:3M3 0:54M4 3:28M5 1:08M6 1:06M7 6:11M8 0:22M9 2:08M10 2:6 1 Elongation 1 e1:42M1 2:01M25:1M3 6:5M4 1:57M5 1:3M6 6:34M7 4:14M8 2M9 1:3M10 2:42 Tensile strength

and predicted calculated yield strength, tensile strength and elongation at training stage, are presented in Figs. 5a, 6a, 7a, respectively. From these comparison charts, it can be clearly seen that the ANN is properly trained and showed a consistency among the properties. A comparison of the measured and predicted yield strength, tensile strength and elongation, at testing stage, are presented in Figs. 8a, 9a, 10 a, respectively. Comparison of measured and predicted mechanical properties at testing stage indicated that there is a high correlation between them. The results of verifying by the ANN are shown in Figs. 5b, 6b, 7b, 8b, 9b, 10b, as scatterplots. It can be seen from the scatter diagrams that the slope and intercept of the regression equations for the outputs are signicantly near to 1 and 0, respectively. The scatterplots

Statistical values of yield strength, tensile strength and elongation of Cu, Sn, Pb, Zn and Ni alloys were presented in Table 4. As a result of statistical values namely, root-mean squared (RMS), absolute fraction of variance (R2) and mean percentage error (MPE), are within acceptable ranges which meets the integrity of the ANN learning and testing stages. 5. Conclusion The main quality indicator of a neural network is its generalization ability, its ability to predict accurately the output of unseen test data. In this study, we have beneted from these features of ANN. It has been demonstrated that MLP architecture with back-

768

M.S. Ozerdem, S. Kolukisa / Materials and Design 30 (2009) 764769

Fig. 7. (a) Comparison, (b) scatter plot, for the training stage of elongation.

Fig. 8. (a) Comparison, (b) scatter plot, for the testing stage of yield strength.

Fig. 9. (a) Comparison, (b) scatter plot, for the testing stage of tensile strength.

propagation learning algorithm can be used as a tool for predicting of mechanical properties namely, yield strength, tensile strength, and elongation of group III CuSnPbZnNi cast alloys. The chemical composition (wt%) of alloying elements was employed as input parameters of the ANN. The ANN architecture, 5-10-3, was employed and the algorithm with 10 hidden layer neurons is sug-

gested to be used in present application since minimum mean square error was obtained. Regarding to the scatter plots (Figs. 5b, 6b, 7b, 8b, 9b and 10b) at training and testing stages, yield strength, tensile strength, and elongation of group III CuSnPbZnNi cast alloys were predicted with a high success rate since the slopes of the functions are very

M.S. Ozerdem, S. Kolukisa / Materials and Design 30 (2009) 764769

769

Fig. 10. (a) Comparison, (b) scatter plot, for the testing stage of elongation.

Table 3 Weights between input layer and hidden layer i Ei = G1Cu + G2Sn + G3Pb + G4Zn + G5Ni + G6 G1 1 2 3 4 5 6 7 8 9 10 0.9508 2.1212 7.3432 7.8754 2.1098 3.3508 0.13847 4.4011 1.7534 1.1278 G2 1.5753 4.0658 10.929 5.0688 5.3311 1.4709 2.4829 5.2269 2.5048 2.2662 G3 0.21861 2.187 4.3224 2.1528 2.1736 1.8413 9.4855 6.0877 1.7547 2.7564 G4 1.1119 0.57483 2.3968 0.77378 4.6547 0.23087 1.9803 2.792 1.1469 1.1492 G5 0.32043 0.79812 0.90819 1.8429 2.7018 0.7163 0.56159 0.71459 1.0523 0.34662 G6 1.0741 2.085 4.6221 8.8576 0.45383 0.17508 2.1388 3.5737 0.37387 0.71906

Table 4 Statistical values of Tensile strength, Yield strength and Elongation RMS train Tensile strength Yield strength Elongation 4.36353 3.78465 1.90119 R2 train 0.999094 0.999829 0.996512 MPE train 7.58977 2.05653 24.5353 RMS test 5.91402 5.9123 3.55947 R2 test 0.998518 0.999631 0.987208 MPE test 9.32542 1.90538 48.8848

close to 1. With the increasing number of input patterns, the success rate of capturing those patterns was being improved gradually. This method (ANN) could be also employed in predicting other physical properties of metal alloys if reliable process parameters and test results are given as ANN input and output data, respectively [1215]. Acknowledgments The author would like to thank to Dicle University Research Committee since a part of this study is supported through Grant DUAPK 03-MF-86. Special thanks to TUBITAK (The Scientic and Technological Research Council of Turkey) for their unfailing support to the researchers. References
[1] Derek, Teyler E. Introduction to copper and copper alloys, ASM hand book, vol. 2, 10th ed. p. 21640. [2] Schmidt Robert F, Schmidt Donald G, Sahoo Mahi. Copper and Copper Alloys, ASM Hand Book, vol. 15, 9th ed. p.77185. [3] Tsao CC. Prediction of ank wear of different coated drills for JIS SUS 304 stainless steel using neural network. J Mater Process Technol 2002;123: 35460.

[4] Haykin S. Neural networks, a comprehensive foundation. New York: McMillian College Publishing Company; 1994. p. 198203. [5] Kohonen T. Content Addressable Memories. NewYork: Springer-Verlag; 1980. p. 16872. [6] Haykin S. Neural networks, a comprehensive foundation. Upper Saddle RiverRiver, NJ: Prentice Hall; 1999. p.15773. [7] Rojas R. Neural network, a systematic introduction. Germany: Springer; 1996. p. 6371. [8] Bernhaut R, Pfalzgraf J. On mathematical modeling of networks and implementation aspects. AISC-Calculemus 2002, LNAI 2385, Berlin: SpringerVerlag; 2002. p.16880. [9] Okuyucu H, Kurt A, Arcaklioglu E. Articial neural network application to the friction stir welding of aluminium plates. J Mater Des 2007;28(1):7884. [10] Altinkok N, Koker R. Neural network approach to prediction of bending strength and hardening behaviour of particulate reinforced (AlSiMg)aluminium matrix composites. J Mater Des 2004;25:595602. [11] Zurada JM. Introduction to articial neural networks. West Publishing Co; 1992. p. 42326. [12] Vasudevan M, Rao BPC, Venkatraman B, Jayakumar T, Raj B. Articial neural network modelling for evaluating austenitic stainless steel and Zircaloy-2 welds. J Mater Process Technol 2005;169:396400. [13] Ohdar RK, Pasha S. Prediction of the process parameters of metal powder preform forging using articial neural network (ANN). J Mater Process Technol 2003;132:22734. [14] Miaoquan L, Xuemei L, Aiming X. Prediction of the mechanical properties of forged TC11 titanium alloy by ANN. J Mater Process Technol 2002;121:14. [15] Ozerdem MS, Kolukisa S. Articial neural network approach to predict mechanical properties of hot rolled, nonresulfurized, AISI 10xx series carbon steel bars. J Mater Process Technol 2008;199:4379.

You might also like