Determination of cutting parameters for silicon wafer with a Diamond Wire Saw using an artificial neural network Erhan Kayabasi a,⇑ , Savas Ozturk b,c , Erdal Celik d,e,f , Huseyin Kurt a a Faculty of Engineering, Department of Mechanical Engineering, Karabuk University, Karabuk, Turkey b Faculty of Engineering and Architecture, Department of Material Science and Engineering, Izmir Katip Çelebi University, Izmir, Turkey c The Graduate School of Natural and Applied Sciences, Dokuz Eylul University, Izmir, Turkey d Faculty of Engineering, Department of Metallurgical and Materials Engineering, Dokuz Eylul University, Izmir, Turkey e Center for Production and Application of Electronic Materials (EMUM), Dokuz Eylul University, Izmir, Turkey f The Graduate School of Natural and Applied Sciences, Department of Nanoscience & Nanoengineering, Dokuz Eylul University, Izmir, Turkey article info Article history: Received 5 December 2016 Received in revised form 30 March 2017 Accepted 11 April 2017 Keywords: Si wafer Artificial neural network Cutting parameters Surface roughness abstract An Artificial Neural Network (ANN) simulation was utilized to predict surface roughness values (R a ) for a Silicon (Si) ingot cutting operation with a Diamond Wire Saw (DWS) cutting machine. Experiments were done on a DWS cutting machine to obtain data for training, testing and validation of the ANN. The DWS cutting operation had three parameters affecting surface quality: spool speed, z axis speed and oil ratio in a coolant slurry. Other parameters such as wire tension, wire thickness, and work piece diameter were assumed as constant. The DWS cutting machine performed 28 cutting operations with different values of the selected three parameters and new cutting parameters were derived for different cutting conditions to achieve the best surface quality by using the ANN. Wafers 400 mm thick were cut from a n-type single crystalline Si ingot in a STX 1202 DWS cutting machine. R a values were measured three times from different regions of the wafers. In ANN simulation 70% of R a values were used as training, 15% of R a values were used as validation and 15% of R a values were used to test data in ANN. The ANN simulation results validated training output data with success above 99%. Consequently, the R a values corresponding to the cutting parameters, and also proper cutting parameters for specific R a values were determined for DWS cutting using the ANN. Ó 2017 Elsevier Ltd. All rights reserved. 1. Introduction In solar cell production, growing and cutting ingot into wafers (wafering) comprise 28% of the total cost distribution of solar mod- ule production (Anspach et al., 2014; Ranjan et al., 2011; Schwinde et al., 2015). In recent decades, cutting single crystalline and poly- crystalline Si ingot with DWS became a conventional method due to its higher production capability, low material consumption, pre- cise thickness determination and low R a values (Zhuang et al., 2016; Bidiville et al., 2015; Sun et al., 2004; Pei et al., 2004; Yu et al., 2012). Despite these advantages, however, there are some disadvantages such as high duration of cutting time, expensive and high quantity coolant slurries, corrugated surface shape for- mation, and diamond wire and silicon wafer breakage due to non-optimized cutting parameters. All of these issues increase the total production cost of a solar cell (Schwinde et al., 2015; Bidiville et al., 2015; Yu et al., 2012). Moreover, the surface quality of wafers obtained after a cutting process directly affects the dura- tion, energy and material consumption of the lapping operation (Schwinde et al., 2015). Thus, DWS parameters must be deter- mined precisely to optimize material and energy consumption, minimize R a values and control total process duration (Wu et al., 2014). An Artificial Neural Network (ANN) is a simple and cost- effective method to derive new parameters and predict results for all science branches. It is mostly used to discover complex rela- tions between input and output data that may not be recognized by theoretical expressions (Çetinel et al., 2006; Boutorh and Guessoum, 2016; Za ˘voianu et al., 2013; Oliveira et al., 2015). In ANN, some data are used for training the network, which provides network weights to achieve desired results (Boutorh and Guessoum, 2016; Ahmadizar et al., 2015). A network consists of three layers such as an input layer, hidden layer and output layer. The neurons are connected to each other by trained weights. Back propagation (BP) is the most common method for training net- works to minimize errors (Çetinel et al., 2006). A feedforward ANN is also another method used for training ANN. Feedforward http://dx.doi.org/10.1016/j.solener.2017.04.022 0038-092X/Ó 2017 Elsevier Ltd. All rights reserved. ⇑ Corresponding author. E-mail address: erhankayabasi@karabuk.edu.tr (E. Kayabasi). Solar Energy 149 (2017) 285–293 Contents lists available at ScienceDirect Solar Energy journal homepage: www.elsevier.com/locate/solener