Solar Energy 149 (2017) 285–293

Contents lists available at ScienceDirect

Solar Energy
journal homepage: www.elsevier.com/locate/solener

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

a r t i c l e i n f o a b s t r a c t

Article history: An Artificial Neural Network (ANN) simulation was utilized to predict surface roughness values (Ra) for a
Received 5 December 2016 Silicon (Si) ingot cutting operation with a Diamond Wire Saw (DWS) cutting machine. Experiments were
Received in revised form 30 March 2017 done on a DWS cutting machine to obtain data for training, testing and validation of the ANN. The DWS
Accepted 11 April 2017
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
Keywords:
of the selected three parameters and new cutting parameters were derived for different cutting
Si wafer
Artificial neural network
conditions to achieve the best surface quality by using the ANN. Wafers 400 mm thick were cut from a
Cutting parameters n-type single crystalline Si ingot in a STX 1202 DWS cutting machine. Ra values were measured three
Surface roughness times from different regions of the wafers. In ANN simulation 70% of Ra values were used as training,
15% of Ra values were used as validation and 15% of Ra values were used to test data in ANN. The ANN
simulation results validated training output data with success above 99%. Consequently, the Ra values
corresponding to the cutting parameters, and also proper cutting parameters for specific Ra 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 Ra 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
⇑ Corresponding author.
E-mail address: erhankayabasi@karabuk.edu.tr (E. Kayabasi).
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 Ra 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; Zăvoianu 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.
 286 E. Kayabasi et al. / Solar Energy 149 (2017) 285–293

ANNs learn much like a human brain, but they don’t need a specific conjugate gradient algorithm (SCGA). Consequently, a Ra corre-
user for a defined problem algorithm. An additional feature is an sponding to several cutting parameters and several optimum cut-
inherent generalization ability (Boutorh and Guessoum, 2016; ting parameters for specific Ra values were predicted. Results
Benardos and Vosniakos, 2007; Ahmed and Hasegawa, 2013; Düg obtained from the simulation were represented graphically and
˘ enci et al., 2015; Yaghini et al., 2013). Because feedforward ANNs evaluated in terms of Ra, cutting duration, material and energy con-
are static networks, they do not have any feedback elements and sumption. The best cutting parameters were determined to mini-
do not include any delays. Thus, results are computed directly by mize the surface Ra, cutting duration, and material and energy
feedforward connections between input data (Boutorh and consumption for the next steps.
Guessoum, 2016). Together, these capabilities make ANN the best
choice for complex non-linear relations in many engineering appli-
2. Experimental and method
cations due to its precise and correct outputs (Düğenci et al., 2015;
Ozturk et al., 2016).
In a DWS process, z axis speed, spool speed and coolant oil ratio
Obtaining a surface with a low Ra in the cutting operation is sig-
are major factors on surface Ra, which are adjustable in a STX 1202
nificant in many engineering applications. There are several stud-
DWS cutting machine. As such, these three parameters were
ies for finalizing a surface with a desired Ra by using parameters
selected for experimental study. A simple general factorial design
derived with artificial intelligence. For instance, Zain et al. (2010)
was adopted for experimentation. Cutting operations were per-
predicted surface Ra in a machining process using ANN. They col-
formed in a STX 1202 DWS cutting machine with a capacity of a
lected 24 instances of data from samples machined with uncoated,
maximum 5 m/s spool speed and 10 mm/min z axis speed. Due
TiAlN coated, and SNTR coated cutting equipment. Their method
to many wire breakages at high cutting speeds and low spool
utilized a feedforward back propagation algorithm with Traingdx
speeds, cutting parameters were selected between 2.5 to 4.5 m/s
for training, Learngdx for learning, MSE for the performance func-
and 0.5 to 1 mm/min for spool speed and the z axis speed, respec-
tion, and Logsis for a transfer function. Several ANN models were
tively with a pure oil coolant. Process parameters and their factor
utilized to determine the best ANN structure. A 3-1-1 network
levels are summarized in Table 1. A Mitutoyo SJ 210 instrument
structure machined with SNTR coated cutting equipment gave the
was used in surface Ra measurements. To understand the effects
best Ra values. Another Ra study with ANN was performed by
of cutting parameters on Ra, material and energy consumption,
Mia and Dhar for hard turning using a coolant (Mia and Dhar,
and cutting duration per one wafer, Ra values were recorded for a
2016). For input data, they used material feed rate, cutting speed,
specific spool speed, z axis speed and coolant oil ratio. This was
material hardness and cutting conditions including dry and high
done because extended operating durations always increases
pressure coolant jet environments. ANN was trained with
material and energy consumption. Spool speed and z axis speed
Levenberg-Marquardt (Trainlm), Bayesian regularization (BR) and
should be adjusted properly to cutting duration for the lowest Ra.
scaled conjugate gradient (SCG). Root mean square error (RMSE)
Hence, high Ra values extend process durations for lapping because
was used to calculate the performance function. The effects of
of deep DWS damage on a Si wafer, as shown in Fig. 1.
parameters on Ra values were determined in a developed model
Moreover, some DWS damage remained after a chemical tex-
with a regression coefficient (R) higher than 0.997. Kumar et al.
turization process because of non-optimized cutting parameters.
(2014), studied Ra determination of titanium alloys during electric
A SEM image of a sample wafer surface subjected to our study is
discharge machining using a hybrid Taguchi-ANN approach and a
given in Fig. 2. There are many deep DWS prints between upright
feedforward back propagation algorithm. A Levenberg-Marquardt
pyramidal shapes formed in a micro texturization process. This
(Trainlm) algorithm was used for the training network. Tansig
deep damage decreases total efficiency and creates weak regions
and Purelin were preferred for the transfer function in the input
that can result in wafer breakages, thus impacting subsequent
and output layers. Tiryaki et al. (2014) used ANN to model Ra for
steps in solar cell production.
a wood machining process with the aim of low operation duration
Measurements were repeated three times from different points
and energy consumption. They selected wood species, number of
on the wafers and average Ra values were considered for ANN sim-
saws, cutting depth, wood zone, and abrasive grain size as input
ulation. The operating technique for the DWS cutting system is
data for their ANN. Çaydasß and Hasçalık (2008) studied ANN and
given schematically in Fig. 3. The main driver provides the spool
a regression analysis method to predict Ra values in an abrasive
speed for a diamond wire in the right and left direction. All system
waterjet machining process. Traverse speed, water jet pressure,
moves through the z direction to cut the Si ingot into wafers. The Si
standoff distance, abrasive grit size and abrasive flow rate param-
ingot was tightly fastened on a table to prevent vibrations to the
eters were used as input data to predict Ra. Feedforward back prop-
diamond wire and Si ingot. Dogit cutting oil was used as coolant
agation ANN was used consisting of 13 input neurons, 22 hidden
slurry, which is soluble in water. Wafers were washed in distilled
neurons and one output neuron. Karayel (2009) used ANN to pre-
water, acetone and a 5% HF solution for the cleaning step.
dict and control Ra in computer numerically controlled (CNC) turn-
ing machine. There were three input data: cutting depth, cutting
speed, material feed speed. Despite the considerable research in Table 1
Cutting settings used in experiments.
this area, there is no study with any type of artificial intelligence
approach predicting Ra values that correspond to determined Spool speed (m/s) z Axis speed (mm/min) Coolant oil [%]
parameters for a Si ingot cutting process in a DWS cutting machine. 2.5 0.5 100
In this study, we determined new cutting parameters that pre- 3.0 0.75
dicted new Ra values corresponding to these parameters for a DWS 3.5 1.0
4.0
process with ANN for the first time. Firstly, 28 wafers were cut
4.5
with two different coolant options in a STX 1202 DWS machine
2.5 1.0 30
and Ra values were measured. Secondly, data collected from mea-
3.0 2.0
surements were rearranged and used to train a feedforward ANN 3.5
with back propagation. Thirdly, new operating conditions were 3.75
determined and simulated in a trained ANN to predict Ra values 4.0
for new operating conditions. This study utilized a feedforward 4.5
5.0
back propagation multilayer ANN and trained with a scaled
 E. Kayabasi et al. / Solar Energy 149 (2017) 285–293
Fig. 1. Deep damage of DWS on Si wafer.
Ra values measured under conditions in Table 1 were inserted
as input and output data to the ANN. The study input variables
were spool speed, z axis speed and coolant oil ratio in slurry; and
the output variable was Ra as shown in Fig. 4. A feedforward back
propagation algorithm was used in the ANN, and is considered a
gradient-decent method. Furthermore, it shifts threshold and
weight values automatically while training ANN to provide a min-
imum difference between target and output values (Kumar et al.,
2014; Beale et al., 2016). The ANN was trained by a Levenberg-
Marquardt (Trainlm) algorithm with a MATLAB ANN tool since
the Trainlm is the fastest function for training a network. Addition-
ally, trainlm works better on function fitting (nonlinear regression)
problems than pattern recognition problems (Beale et al., 2016).
The Trainlm function also renews weights and biases with a
Levenberg-Marquardt optimization (Kumar et al., 2014). Transfer
function Tansig is much preferred once the speed of the process
is significant, but the structure of ANN is not necessary (Kumar
et al., 2014). Mean square error (MSE) as a performance function
provided evaluation and comparison for minimum errors.
Consequently, a feedforward backpropagation method, Trainlm
training function, Learngdm adaptation learning function, mean
square error (MSE) performance function, and Tansig transfer func-
tion were selected for the ANN. The number of neurons was
selected as 10 for the hidden layer. Maximum epochs were chosen
at 1000 and the error goal was written as 0. The learning rate for
ANN was 0.01. Initial weights and biases were derived automati-
cally by the MATLAB ANN tool (Beale et al., 2016). The architecture
of the ANN in our study is given in Fig. 5. Relative error between
Fig. 2. Deep damage of DWS on textured Si wafer.
287
Fig. 3. Schematic of STX 1202 machine cutting system.
measurements and ANN simulation results were calculated with
Eq. (1) (Kumar et al., 2014).
 
Experimental Results  ANN Predicted Results
% Relativ e Error ¼
Experimental Results
 100 ð1Þ
Spool speed, z axis speed and coolant oil ratio were inserted as
Input data and Ra values were inserted as Output data into the ANN
in part 1. With these values, ANN was trained several times in
order to reach the most successful trained ANN by examining
training, test and validation graphs in Part 2. Lastly, accuracy con-
trol of ANN results was performed in Part 3. Two different simula-
tions were performed with a trained ANN. The first was performed
to predict Ra values versus spool speed, and the second was per-
formed to predict Ra values versus z axis speed. A flowchart of
the whole process is given in Fig. 6.
During the experimental studies, the diamond wire saws had
the same properties and all wafers were 400 mm thick and cut from
the same 100 mm diameter n-type single crystalline Si ingot. The
duration of one wafer cutting operation can be calculated by divid-
ing the ingot diameter by the z axis speed. To increase the speed of
the operation, the z axis speed should be increased directly up to a
288 E. Kayabasi et al. / Solar Energy 149 (2017) 285–293
Fig. 4. Structure of ANN.
limited level. If z axis speed augments more than spool speed com-
pensation, this will cause breakage of the diamond wire, breakage
of the wafer, and formation of a corrugated surface shape. In addi-
tion, when the diamond wire breaks, the connecting wire is also
broken, resulting in a complicated replacement process and exces-
sive wasted time. Moreover, after replacing the connection of wire
into the system, the first cutting process is used to obtain an initial
flat surface for cutting, and not for a new wafer. Furthermore, the
total wire length decreases because the broken diamond wire can-
not be spliced. In experimental studies, all these factors were con-
sidered with an aim of obtaining minimum material consumption
and a flat surface with the lowest Ra.
3. Results and discussion
This study performed 84 measurements to determine 28 aver-
age Ra values that were inserted into an ANN simulation from dif-
ferent regions of wafer surfaces. After several trainings of ANN, a
success level of more than 99% was achieved as given in Fig. 8.
Training, test and validation slopes from ANN simulation proves
that there is a strong correlation between spool speed, z axis speed,
coolant oil/water ratio and the Ra values in terms of obtaining
excellent accuracy as seen in the regression curves in Fig. 7.
Fig. 5. A multi-layer tansig-purelin network with two input neurons and one hidden layer with 10 neurons and one output layer.
Fig. 6. Process flow chart of ANN.
All studied parameters, duration per one cutting process, aver-
age measured Ra values, ANN results for these average measured
values and relative error between measured and simulated values
are given in Table 2. Table 2 shows the similarity between mea-
sured and simulated Ra values. On the line, some results were
obtained at almost identical values with measured values. More
than half of relative errors are observed below 1%. Minimum rela-
tive error was observed in Experiment 2 with a value of 0.009%.
Maximum relative error was observed in Experiment 22 with a
value of 13.167%.
Figs. 8 and 9 present some information about Ra values of the
cutting process that was performed with a slurry consisting of pure
coolant oil. As shown in Fig. 8, Ra versus spool speed can be
observed for each z axis speed from 0.5 mm/min to 1 mm/min of
the first simulation mentioned before Fig. 8. Generally, it is under-
stood from Fig. 9 that higher z axis speeds increase Ra values. In
 E. Kayabasi et al. / Solar Energy 149 (2017) 285–293 289

Fig. 7. Training, test and validation results of ANN.

Fig. 8. ANN results representing difference of Ra versus spool speed for the wafers cut with pure coolant oil.

detail, low z axis speeds, for example 0.5 mm/min and 0.6 mm/
min, generally result in lower Ra values between the interval of
3–4.5 m/s in spool speed. In contrast, for spool speeds below
3 m/s, Ra values suddenly increasing up to 2.81 mm and 3.16 mm,
respectively, for z axis speeds of 0.5 mm/min and 0.6 mm/min.
Fig. 8 indicates that there is no advantage for the z axis speeds of
0.5 mm/min and 0.6 mm/s at higher values of a spool speed above
3 m/s. A 3 m/s spool speed provides the best Ra value when z axis
speeds are 0.5 mm/min and 0.6 mm/min. In addition, even though
the Ra values are lowest for 0.5 mm/min and 0.6 mm/min, the
290 E. Kayabasi et al. / Solar Energy 149 (2017) 285–293

Table 2
Cutting parameters, average surface Ra values and ANN results for the experiments.

Exp. no. Cutting parameters Measured Ra [mm] ANN results for Ra [mm] Relative error [%]
Spool speed [m/s] Z axis speed [mm/min] Coolant oil [%] Duration of DWS [min]
1 2.5 0.5 1 200 3.269 3.165 3.177
2 2.5 0.75 1 134 2.036 2.036 0.009
3 2.5 1 1 100 2.080 2.074 0.298
4 2.5 1 0.3 100 1.745 1.745 0.020
5 2.5 2 0.3 50 1.431 1.430 0.051
6 3 0.5 1 200 1.141 1.142 0.075
7 3 0.75 1 134 1.432 1.428 0.271
8 3 1 1 100 2.219 2.178 1.863
9 3 1 0.3 100 1.569 1.542 1.738
10 3 2 0.3 50 1.424 1.427 0.245
11 3.5 0.5 1 200 1.157 1.088 5.942
12 3.5 0.75 1 134 1.499 1.538 2.587
13 3.5 1 1 100 2.176 2.209 1.499
14 3.5 1 0.3 100 1.377 1.376 0.078
15 3.5 2 0.3 50 1.406 1.410 0.282
16 3.75 2 0.3 50 1.416 1.395 1.473
17 4 0.5 1 200 1.087 1.088 0.137
18 4 0.75 1 134 1.786 1.714 4.008
19 4 1 1 100 2.303 2.299 0.154
20 4 1 0.3 100 1.321 1.322 0.100
21 4 2 0.3 50 1.358 1.376 1.346
22 4.5 0.5 1 200 1.256 1.091 13.167
23 4.5 0.75 1 134 1.839 1.881 2.283
24 4.5 1 1 100 2.906 2.905 0.036
25 4.5 1 0.3 100 1.304 1.298 0.454
26 4.5 2 0.3 50 1.340 1.336 0.332
27 5 1 0.3 100 1.254 1.256 0.133
28 5 2 0.3 50 1.259 1.258 0.060

Fig. 9. ANN results representing difference of Ra versus z axis speed for the wafers cut with pure coolant oil.

cutting duration takes more than three hours for one wafer as seen
in Table 2. This creates other problems such as material and energy
consumption and low production level. In order to shorten the
duration of the cutting operation higher z axis speeds were inves-
tigated. A 2.5 mm/s spool speed gives the lowest Ra values when z
axis speeds was selected between 0.9 and 1 mm/min. In addition, a
z axis speed 0.8 mm/min gives lower Ra values when the spool
speed was 3 m/s. In other words, there is no advantage at higher
spool speeds of more than 2.5 m/s for z axis speeds at 0.9 mm/
min and 1 mm/min.
Fig. 9 denotes output values of ANN in terms of Ra versus z axis
speed, for each spool speed from 2.5 m/s to 4.5 m/s. It is clear from
Fig. 9 that for z axis speeds between 0.5 mm/min and 0.6 mm/min,
Ra values were lower for all spool speeds except for 2.5 m/s and
2.75 m/s. Ra values then suddenly increase with spool speeds
below 2.75 m/s and for z axis speeds between 0.5 and 0.7 mm/
min. In comparison, Ra values increasing after 0.6 mm/min for all
spool speeds as shown in Fig. 9.
Comparing Figs. 8 and 9, they almost demonstrate the same
results for Ra values and draw the same roadmap for optimum
parameters. Considering Figs. 8 and 9 together, the lowest Ra can
be obtained with spool speeds between 3.5 and 4.5 m/s and with
between 0.50 and 0.6 mm/min z axis speeds, but at a maximum
duration. To decrease the duration with minor losses of Ra values,
Figs. 8 and 9 are in agreement for selecting a spool speed of 3.5 m/s
and z axis speed of 0.60 mm/min. A further advantage in duration
 E. Kayabasi et al. / Solar Energy 149 (2017) 285–293 291

Fig. 10. ANN results representing difference of Ra versus spool speed for the wafers cut with coolant oil/distilled water ratio 0.3.

Fig. 11. ANN results representing difference of Ra versus z axis speed for the wafers cut with coolant oil/distilled water ratio 0.3.

can be obtained at a 3.25 m/s spool speed and 1 mm/min z axis
speed, but with a loss in surface quality.
Results for the second ANN simulation are given in Figs. 10 and
11. Ra values were considerably lower than the experiments with
pure coolant oil. When cutting parameters were selected maxi-
mum (spool speed = 5 m/s, z axis speed = 2 mm/min), the Ra value
was at a minimum (1.258 mm). Fig. 10 indicates that when the z
axis speed is 1 mm/min, the Ra value becomes minimum at a
5 m/s spool speed. However, Ra values increase gradually with an
increase in z axis spool speeds, and a 2 mm/min z axis speed at a
5 m/s spool speed is more advantageous due to minor Ra loss, com-
pared to a 1 mm/min z axis speed and 5 m/s spool speed. Hence,
with a minor loss of Ra values, the cutting operation ended at only
50 min per wafer and the loss in Ra value can be compensated with
the short duration in the lapping step.
Fig. 11 shows that a 5 m/s spool speed in all z axis speeds gives
minimum Ra values, except at a 1 mm/min z axis speed. Ra values
gradually increase with a decrease of spool speeds for all z axis
speeds. However, when the z axis speed was 1 mm/min, there
was no advantage in selecting a spool speed of 5 m/s versus select-
ing 4.75 m/s, as the same values were obtained at both spool
speeds. Furthermore, the two operations ended with the same
duration and selecting a higher spool speed results in high energy
consumption.
When Figs. 10 and 11 are evaluated together, the slopes always
validate each other. It’s clearly understood that there is a signifi-
cant advantage in using distilled water and a coolant oil mixture
at a 7:3 volumetric ratio. Using distilled water and an oil mixture
coolant permits utilization of higher capacities of the DWS
machine, resulting in the lowest Ra values. Figs. 10 and 11 also
reflect proper parameters for special cutting conditions of cutting
studies. Fig. 12 represents the surface view of a wafer cutting oper-
ation under cutting conditions at a 5 m/s spool speed, 1.5 mm/min
z axis speed and 3:10 oil and distilled water mixture. Surface qual-
ity improved visibly compared to the SEM view given in Fig. 1.
In DWS cutting, it’s not sufficient to only obtain a low surface Ra,
but it is also important to have cutting done in a short period of
time to minimize costs with various cutting parameters. Longer
292 E. Kayabasi et al. / Solar Energy 149 (2017) 285–293
Fig. 12. SEM image of a wafer cut with optimized parameters.
cutting durations will result in a low Ra, but also results in the use
of additional material and energy consumption that increases total
cost. Meanwhile, short cutting durations utilize less material and
energy consumption, but result in a longer lapping duration in
the next step. For these reasons, the total duration of DWS cutting
and lapping operations should be considered together to optimize
overall duration, materials, and energy consumption.
4. Conclusion
To satisfy current requirements for manufacturing Si wafers, a
cutting process with the above-mentioned specifications was iden-
tified and the effects of cutting parameters on Ra were investigated.
Most researchers have used a Taguchi orthogonal design and cen-
tral composite design in order to reduce experimentation. How-
ever, these models are based on only a few experimental results
(Çaydasß and Hasçalık, 2008; Asiltürk and Akkusß, 2011; Allen
et al., 2008; Mansour and Abdalla, 2002). In comparison, this paper
adopted a general full factorial design. This investigation con-
ducted 28 wafer cutting experiments and Ra values for all wafers
were measured from three different regions.
In this study, cutting parameters were determined for several
cutting conditions from 2.5 to 4.5 m/s and 0.5 to 1 mm/min for
spool speed and z axis speed respectively with a pure oil coolant,
and from 2.5 to 5 m/s and 1 to 2 mm/min for spool speed and
the z axis speed respectively with diluted coolant oil, using a feed-
forward back propagation ANN. Firstly, several cutting processes
were performed from a n-type single crystalline Si ingot of
100 mm diameter. All cutting parameters and Ra values were rear-
ranged in the form of tables to insert into an ANN. After training,
testing and validation of the ANN, new cutting processes were per-
formed to observe how the simulation results fit with newly mea-
sured Ra values, which provided results for various cutting
parameters. These results permit researchers to select the best
spool speed, z axis speed and coolant oil ratio for DWS operation
while considering low Ra, low material and low energy consump-
tion. In addition, optimized parameters also prevent the formation
of corrugated shapes and breakage of the diamond wire and Si
wafer. This study provides a roadmap for researchers studying
the cutting of Si ingot in DWS and facilitates the selection of opti-
mal parameters for future studies.
Conflict of interest
The authors declare that they have no conflict of interest.
Acknowledgements
This experimental research was supported by the Scientific
Research Projects Coordination Unit of Karabük University (Project
Number KBU-BAP-14/1-DR-003), Turkey and Electronic Materials
Production and Application Center (EMUM) at Dokuz Eylul Univer-
sity, Turkey.
References
Ahmadizar, F. et al., 2015. Artificial neural network development by means of a
novel combination of grammatical evolution and genetic algorithm. Eng. Appl.
Artif. Intell. 39, 1–13.
Ahmed, Y.A., Hasegawa, K., 2013. Automatic ship berthing using artificial neural
network trained by consistent teaching data using nonlinear programming
method. Eng. Appl. Artif. Intell. 26 (10), 2287–2304.
Allen, R.D. et al., 2008. Surface roughness prediction model for CNC machining of
polypropylene. Proc. Instit. Mech. Eng., Part B: J. Eng. Manuf. 222 (2), 137–157.
Anspach, O., Hurka, B., Sunder, K., 2014. Structured wire: From single wire
experiments to multi-crystalline silicon wafer mass production. Sol. Energy
Mater. Sol. Cells 131, 58–63.
_ Akkusß, H., 2011. Determining the effect of cutting parameters on surface
Asiltürk, I.,
roughness in hard turning using the Taguchi method. Measurement.
Beale, H.M., Hagan, M.T., Demuth, H.B., 2016. Neural Network Toolbox User’s Guide.
The MathWorks, Inc., USA, p. 3-2/3-30.
Benardos, P.G., Vosniakos, G.C., 2007. Optimizing feedforward artificial neural
network architecture. Eng. Appl. Artif. Intell. 20 (3), 365–382.
Bidiville, A. et al., 2015. Wire-sawing processes: parametrical study and modeling.
Sol. Energy Mater. Sol. Cells 132, 392–402.
Boutorh, A., Guessoum, A., 2016. Complex diseases SNP selection and classification
by hybrid Association Rule Mining and Artificial Neural Network—based
Evolutionary Algorithms. Eng. Appl. Artif. Intell. 51, 58–70.
Çaydasß, U., Hasçalık, A., 2008. A study on surface roughness in abrasive waterjet
machining process using artificial neural networks and regression analysis
method. J. Mater. Process. Technol. 202 (1–3), 574–582.
Çetinel, H. et al., 2006. Artificial neural network-based prediction technique for
wear loss quantities in Mo coatings. Wear 261 (10), 1064–1068.
Düğenci, M. et al., 2015. Creep modelling of polypropylenes using artificial neural
networks trained with Bee algorithms. Eng. Appl. Artif. Intell. 45, 71–79.
Karayel, D., 2009. Prediction and control of surface roughness in CNC lathe using
artificial neural network. J. Mater. Process. Technol. 209 (7), 3125–3137.
Kumar, S. et al., 2014. A hybrid Taguchi-artificial neural network approach to
predict surface roughness during electric discharge machining of titanium
alloys. J. Mech. Sci. Technol. 28 (7), 2831–2844.
Mansour, A., Abdalla, H., 2002. Surface roughness model for end milling: a semi-free
cutting carbon casehardening steel (EN32) in dry condition. J. Mater. Process.
Technol. 2002 (124), 183–191.
Mia, M., Dhar, N.R., 2016. Prediction of surface roughness in hard turning under
high pressure coolant using Artificial Neural Network. Measurement 92, 464–
474.
Oliveira, G.G., Pedrollo, O.C., Castro, N.M.R., 2015. Simplifying artificial neural
network models of river basin behaviour by an automated procedure for input
variable selection. Eng. Appl. Artif. Intell. 40, 47–61.
Ozturk, T. et al., 2016. Extracting the dielectric constant of materials using ABC
based ANNs and NRW algorithms. J. Electromag. Waves Appl. 30 (13), 1785–
1799.
Pei, Z.J. et al., 2004. An experimental investigation into soft-pad grinding of wire-
sawn silicon wafers. Int. J. Mach. Tools Manuf. 44 (2–3), 299–306.
Ranjan, S. et al., 2011. Silicon solar cell production. Comput. Chem. Eng. 35 (8),
1439–1453.
Schwinde, S., Berg, M., Kunert, M., 2015. New potential for reduction of kerf loss and
wire consumption in multi-wire sawing. Sol. Energy Mater. Sol. Cells 136, 44–47.
Sun, X. et al., 2004. Waviness removal in grinding of wire-sawn silicon wafers: 3D
finite element analysis with designed experiments. Int. J. Mach. Tools Manuf. 44
(1), 11–19.
Tiryaki, S., Malkoçoğlu, A., Özsßahin, Sß., 2014. Using artificial neural networks for
modeling surface roughness of wood in machining process. Constr. Build. Mater.
66, 329–335.
Wu, H., Yang, C., Melkote, S.N., 2014. Effect of reciprocating wire slurry sawing on
surface quality and mechanical strength of as-cut solar silicon wafers. Precision
Eng. 38 (1), 121–126.
 E. Kayabasi et al. / Solar Energy 149 (2017) 285–293 293

Yaghini, M., Khoshraftar, M.M., Fallahi, M., 2013. A hybrid algorithm for artificial
neural network training. Eng. Appl. Artif. Intell. 26 (1), 293–301.
Yu, X. et al., 2012. Thin Czochralski silicon solar cells based on diamond wire sawing
technology. Sol. Energy Mater. Sol. Cells 98, 337–342.
Zain, A.M., Haron, H., Sharif, S., 2010. Prediction of surface roughness in the end
milling machining using Artificial Neural Network. Expert Syst. Appl. 37 (2),
1755–1768.
Zăvoianu, A.-C. et al., 2013. Hybridization of multi-objective evolutionary
algorithms and artificial neural networks for optimizing the performance of
electrical drives. Eng. Appl. Artif. Intell. 26 (8), 1781–1794.
Zhuang, Y.F. et al., 2016. Versatile strategies for improving the performance
of diamond wire sawn mc-Si solar cells. Sol. Energy Mater. Sol. Cells 153,
18–24.