Design of medium carbon steels by computational intelligence techniques N.S. Reddy a,⇑ , J. Krishnaiah b , Hur Bo Young a , Jae Sang Lee c a School of Materials Science and Engineering, Engineering Research Institute, Gyeongsang National University, Jinju 660-701, Republic of Korea b Research and Development, Bharat Heavy Electricals Limited, Tiruchirappalli, India c Graduate Institute for Ferrous Technology, Pohang University of Science and Technology, Pohang 790-784, Republic of Korea article info Article history: Received 25 August 2014 Received in revised form 21 November 2014 Accepted 22 January 2015 Keywords: Neural networks Genetic algorithms Index of relative importance Medium carbon steels Desired properties Optimization abstract Steel design with the targeted properties is a challenging task due to the involvement of many variables and their complex interactions. Artificial neural networks (ANN) recognized for representing the complex relationships and genetic algorithms (GA) are successful for optimization of many real world problems. ANN has been used to identify the relative importance of variables those control the mechanical proper- ties of medium carbon steels. We propose the combination of ANN and GA to optimize composition and heat treatment parameters for the desired mechanical properties. The trained ANN model was used as a fitness function and also as a predictive model. The predicted properties were realistic and higher for the model suggested with the optimum combination of composition and heat treatment variables. The pro- posed framework is expected to be useful in reducing the experiments required for designing new steels. Ó 2015 Elsevier B.V. All rights reserved. 1. Introduction The design of alloy steels with the desired properties is a chal- lenging task as it involves a multi-objective optimization of various problems [1,2]. For example, it is hard to design a material that com- bines high strength and ductility, which are the two most key mechanical characteristics of metals [3]. There are excellent facili- ties to measure the composition, microstructure, and properties. However, it is difficult to predict reliable mechanical properties with the comprehensive description of the chemical composition, pro- cessing parameters and structure of steels. The microstructure in steels determine the mechanical properties and the typical metallurgical approach for the prediction of properties through structure-properties relationships follows the hierarchy of composi- tion and processing conditions ? microstructure ? mechanical properties. Developing models for the analysis of complex multicomponent steel properties is a difficult task as a quantitative treatment is nec- essary. Physical models are not capable of predicting the mechan- ical properties of steels. The conventional linear regression is not sufficient to describe the relationships; hence the biologically inspired artificial neural networks (ANNs) have been identified as appropriate tools. ANN models are well-known for function approximation and feature extraction of the highly complex non- linear relationships from the data [4,5]. In the present approach, both the compositions and heat treatment variables which deter- mine the microstructure are related directly to the mechanical properties. Hence, it is appropriate to attempt these techniques to enable the quantitative expression and understanding of the complicated nonlinear problem [6–9]. ANN model is a combination of a mathematical function and associated weights between the inputs, hidden units and outputs. Experimental data are presented to the network in the form of input and output parameters, and optimized nonlinear relationship is found by minimizing mean square error. The error adjustment step takes place once the data is presented to the input layer and forward propagation is finished. Every processing element in the output layer estimates an output and compared with the actual output specified in the data set. An error value for every unit is cal- culated based on the difference. The weights of these units are adjusted based on the error for all of the interconnections estab- lished with the output layer. After this, the subsequent sets of weights are adjusted for the interconnections coming into the hid- den layer located just beneath the output layer. Upon each presen- tation, the weights are adjusted to decrease the error between the network’s output and the actual output. This process is continued http://dx.doi.org/10.1016/j.commatsci.2015.01.031 0927-0256/Ó 2015 Elsevier B.V. All rights reserved. ⇑ Correspoding author at: School of Materials Science and Engineering, Depart- ment of Metallic & Materials Engineering, Gyeongsang National University, 900 Gazwa-dong, Jinju 660-701, Gyeongnam, Republic of Korea. Tel.: +82 55 772 1669 (O), +82 70 8745 0793 (R), mobile: +82 10 8999 0793; fax: +82 55 772 1670. E-mail addresses: nsreddy@gnu.ac.kr, dr.subba@me.com (N.S. Reddy). Computational Materials Science 101 (2015) 120–126 Contents lists available at ScienceDirect Computational Materials Science journal homepage: www.elsevier.com/locate/commatsci