Optics and Lasers in Engineering 88 (2017) 1–4

Contents lists available at ScienceDirect

Optics and Lasers in Engineering

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

Prediction of laser cutting heat affected zone by extreme learning

machine
Obrad Anicic a,n, Srđan Jović a, Hivzo Skrijelj a, Bogdan Nedić b
a
Faculty of Technical Sciences, Kosovska Mitrovica, University of Priština, Kneza Milosa 7, 38 220 Kosovska Mitrovica, Serbia
b
Faculty of Engineering, University of Kragujevac, Sestre Janic 6, 34 000, Kragujevac, Serbia

art ic l e i nf o a b s t r a c t

Article history: Heat affected zone (HAZ) of the laser cutting process may be developed based on combination of dif-
Received 17 June 2016 ferent factors. In this investigation the HAZ forecasting, based on the different laser cutting parameters,
Received in revised form was analyzed. The main goal was to predict the HAZ according to three inputs. The purpose of this
13 July 2016
research was to develop and apply the Extreme Learning Machine (ELM) to predict the HAZ. The ELM
Accepted 14 July 2016
results were compared with genetic programming (GP) and artificial neural network (ANN). The relia-
Available online 25 July 2016
bility of the computational models were accessed based on simulation results and by using several
Keywords: statistical indicators. Based upon simulation results, it was demonstrated that ELM can be utilized ef-
Extreme Learning Machine fectively in applications of HAZ forecasting.
Forecasting
& 2016 Elsevier Ltd. All rights reserved.
HAZ
Laser cutting

1. Introduction
Recently there is an increase in research in the field of laser
cutting process because laser machining offers an attractive al-
ternate among all nonconventional machining due to improved
end product quality, low cost, short processing time, smoothness
and precise cuts, narrow width and small heat affected zone (HAZ)
[1]. By controlling the different process parameters of laser, pre-
cision cuts can be made for different geometry and complex shape
[2]. Laser cutting is a non-contact, abrasion less technique elim-
inates tool wear, machine-tool deflections, vibrations and cutting
forces and can be used for almost all type of materials [3].
The interest of manufacturers using laser cutting is the opti-
mization of the productivity and the subsequent quality of com-
ponents made by the laser cutting process. Both aspects are gov-
erned by the selection of appropriate laser process parameters,
which are unique for each material and thickness. These para-
meters include laser power, cutting speed and assist gas parameter
[4]. These parameters are usually adjusted and tuned to provide
the quality of cut desired, but this consumes exhaustive amounts
of time and effort and still good quality cutting conditions may not
be found. Laser cut quality cannot be easily predicted. This is due
to the dynamic nature of the laser cutting process, and it is par-
ticularly clear when cutting ferrous alloys using oxygen as an as-
sisting gas. The oxidation reactions with iron and other alloying
n
Corresponding author.
E-mail address: anicicobrad@gmail.com (O. Anicic).
elements add another source of heat and material removal occurs
at two moving and often interacting fronts, the oxidation front and
the laser beam front. There are several investigations of the HAZ
during the laser cutting process.
Cai et al. [5] was aimed at introducing the laser induced ther-
mal-crack propagation (LITP) technology to solve the silicon-glass
double layer wafer dicing problems in the packaging procedure of
silicon-glass device packaged, investigating the feasibility of this
idea, and studying the crack propagation process of LITP cutting
double layer wafer. The analysis of experimental results and nu-
merical simulation results was shown that dual laser beam revi-
sion the separation path technology can effectively revise the de-
viation of the separation path in asymmetry linear cutting glass
with LITP [6]. In order to develop a detailed map of cracks and
understand the root cause of cracking, a comprehensive micro-
structural and numerical analysis was performed by Chamanfar
et al. [7]. Jebbari et al. [8] was shown that the absorbed power
melts a part of the material and preheats the part that would be
manufactured during the following interaction time. The numer-
ical predictions of the mechanical properties in the HAZ was
shown that there are large variations both in yield stress and work
hardening depending on the distance from the weld line [9].
McDaniels et al. [10] was found that the HAZ did not appear to
have an adverse effect on the high-cycle fatigue behavior. The
predicted temperature distribution is correlated with the HAZ and
a critical temperature range corresponding to the maximum depth
of the HAZ is identified using a combination of metallography,
hardness testing, and thermal modeling [11]. Parametric studies
showed that the depth and width of the heat affected zone

http://dx.doi.org/10.1016/j.optlaseng.2016.07.005
0143-8166/& 2016 Elsevier Ltd. All rights reserved.
2 O. Anicic et al. / Optics and Lasers in Engineering 88 (2017) 1–4

increased with an increase in the laser power and decreased with ● input 1 Laser power (Watt)
an increase of the laser spot size and the laser scan speed [12]. The ● input 2 Cutting speed (mm/min)
thermal model can be used to determine the laser parameters for a ● input 3 Gas pressure (Bar)
given cut geometry that will yield no residual heat affected zone in ● output Heat affected zone (HAZ) (mm)
the material after cutting.
Even though a number of new mathematical functions have
been proposed for modeling of the HAZ, in this investigation the
main aim was to overcome high nonlinearity by applying the soft 2.2. Extreme Learning Machine
computing method. Soft computing can be used as alternative to
analytical approach since soft computing offers advantages such as Huang et al. [19,20] developed Extreme Learning Machine
no required knowledge of internal system parameters, compact (ELM) as a novel learning algorithm for single hidden layer feed
solution for multi-variable problems. forward networks (SLFNs).
In this study laser cutting of Glass Fiber Reinforced Plastic This approach has some priority compared with conventional
Composites has been carried out to predict HAZ. Three input
neural networks including: 1) ELM is easy to use, and the method
parameters: laser power, cutting speed and gas pressure were
not only makes learning extremely fast but also produces good
taken under experimental investigation to find the effect and
generalization performance [19,20]; 2) in conventional neural
consequently models have been developed to predict the HAZ in
networks all the parameters of the networks such as learning rate,
context of cutting parameters.
In this investigation the main aim was to overcome high non- learning epochs and local minima are tuned iteratively by using
linearity since HAZ parameters are endogenous variables by ap- such learning algorithms; 3) ELM can be easily implemented and
plying the soft computing method. Recently, the Extreme Learning can obtain the smallest training error and the smallest norm of
Machine (ELM) has been introduced as a soft computing algorithm weights [19,20].
for single layer feed forward neural network (NN) [13,14]. It is In Fig. 1 the schematic topological structure of ELM network is
capable to solve problems caused by gradient descent based al- shown. For M arbitrary samples (xi, ti), in which xi ¼[xi1, xi2, …, xin]
T
gorithms like back propagation which applies in artificial neural ∈ Rn and ti ¼[ti1, ti2, …, tin]T∈ Rm, standard single hidden layer feed
networks (ANNs) and to decrease required time for training NN. It forward networks (SLFNs) with N hidden nodes and activation
has been proved that by utilizing the ELM, learning becomes very function g(x) are modeled as follows [19,20]:
fast and it produces good generalization performance [15]. It has M M
been widely utilized for the estimation of problems in many dif- ∑ βi gi (x j) = ∑ βi g (wi. xj + bi ) = oj , j = 1, ... , N
ferent fields of water resources [16–18]. i=1 i=1 (1)
In this investigation the main goal is to anticipate HAZ by using T
ELM approach. The primary objective is to analyze the HAZ fore- where wi ¼ [wi1, wi2, …, win] is the weight vector between input
casting based on laser power, cutting speed and gas pressure. and hidden nodes, βi ¼[ βi 1, βi 2, …, βi n]T is the weight vector be-
tween output and hidden nodes, and bi is the threshold of the ith
hidden node. That standard single hidden layer feed forward
2. Methodology networks with M hidden nodes with activation function g(x) as
follows:
2.1. Heat affected zone M
∑ βi g (wi. x j + bi ) = tj, j = 1, ... , N .
The experimental specimen was made of Glass Fiber Reinforced i=1 (2)
Plastic Composite in which E-glass chopped fibers were used as
The activation function can also be written as
reinforcement material and polyester resin as matrix material. The
size of the material was 300 mm
200 mm
5 mm. All experi-
ments were performed using a fiber laser cutting machine (Sa-
hajanand Laser Technology Limited, Gandhinagar) with CNC work
table having maximum output power 2000 W in continuous
mode. The laser beam was focused using a 125 mm focal length
lens, the nozzle diameter was 1.5 mm and the nozzle tip distance
was 1.5 mm. All parameters were precisely controlled as the ma-
chine is equipped with PLC controller. The machine was capable of
using nitrogen and oxygen as an assist gas at right angle and
coaxially to the laser beam. By trial experiments it was decided to
use nitrogen as inert gas.
For measurement of HAZ, straight cuts of about 50 mm in
length were made on sheet and they were measured using 3D
microscope MITUTOYO/QS-L2010ZB. Five readings were taken
throughout the length on both side of the cut and then average is
taken for each cut.
In experiments, number of fixed and independent parameters
were involved. The fixed parameters were focal length, nozzle
diameter, nozzle tip distance and material thickness and the va-
lues are as mention above whereas the variable parameters were
laser power, cutting speed and gas pressure. Input and output
parameters which are used in this investigation are listed as
bellow: Fig. 1. The topological structure of the extreme learning machine network used in
this study.
 O. Anicic et al. / Optics and Lasers in Engineering 88 (2017) 1–4 3

gi ( x j )=g (wi x j+bi ) (3)

where gi ( x j ) is an activation function, x j is an input sample, bi is a

threshold of ith hidden node an wi is an input weight.
These equations can be written as follows [19,20]:
Hβ=T (4)

where
⎡ g (w1. x1 + b1) ⋯ g (wM . x1 + bM ) ⎤
⎢ ⎥
H=⎢ ⋮ ⋯ ⋮ ⎥
⎢⎣ g (w1. x N + b1) ⋯ g (wM . x N + bM )⎥⎦
N *M

⎡ β T⎤ ⎡ tT ⎤ Fig. 2. Scatter plot of actual and ELM forecasted values of HAZ.

⎢ 1⎥ ⎢ 1⎥
β=⎢ ⋮⎥ and T = ⎢ ⋮ ⎥ .
⎢ T⎥ ⎢ tT ⎥
⎣ βM ⎦M ⎣ M ⎦N
*m *m
where H is the hidden layer output matrix on neural network. The
output weights can be constructed by finding least square solu-
tions mentioned equation which the result represents the fol-
lowing equation:
β = H †T (5)
†
where H is the Moore–Penrose generalized inverse of the hidden
layer output matrix H.

2.3. Models performance evaluation

Fig. 3. Scatter plot of actual and ANN forecasted values of HAZ.

Predictive performances of the proposed models were pre-

sented as the root means square error (RMSE) and coefficient of
determination (R2). These statistics are defined as follows:
n
∑i = 1 (Pi − Oi )2
RMSE = ,
n (6)

⎡ ∑n O − O ⋅ P − P ⎤2
⎣ i=1 i i (
i i ⎦ )( )
R2 = n n
( )
∑i = 1 Oi − Oi ⋅ ∑i = 1 Pi − Pi ( ) (7)

Here, Oi and Pi represent the forecast and experimental values,

respectively and n denotes the sum of test data.
Fig. 4. Scatter plot of actual and GP forecasted values of HAZ.

3. Results and discussion

follows that the prediction results are in very good agreement
with the measured values for the ELM method. This observation
3.1. Performance analysis
can be confirmed with acceptable value for the coefficient of de-
termination. The number of overestimated or underestimated
In this section, performance results of the ELM predictive
values is limited. Consequently, it is obvious that the predicted
model are reported for HAZ forecasting based on the parameters
values enjoy high level of precision.
in Table 1. Fig. 2 presents the accuracy of the developed ELM
predictive model for HAZ. Fig. 3 presents the accuracy of the de-
3.2. Performance comparison of ELM, ANN and GP
veloped ANN [21] predictive model for HAZ. Fig. 4 presents the
accuracy of the developed GP [22] predictive model for HAZ. The
In order to demonstrate the merits of the proposed ELM ap-
prediction accuracy of the ELM model is acceptable for this data. It
proach on a more definite and tangible basis, the ELM models'
can be seen that most of the points fall along the diagonal line. It
prediction accuracy was compared to the prediction accuracy of
the ANN [21] and GP [22] methods, which were used as bench-
Table 1
Comparative performance statistics of the ELM, ANN and GP models for HAZ
marks. The conventional statistical error indicators (RMSE and R2)
forecasting. were used for comparison. Table 1 summarizes the forecasting
accuracy results.
ELM ANN GP

2 2
4. Conclusion
RMSE R RMSE R RMSE R2
0.0046 0.9976 0.0169 0.9671 0.0342 0.8656
Forecasting of the HAZ is complex due to the many indicators
4 O. Anicic et al. / Optics and Lasers in Engineering 88 (2017) 1–4
and factors which influence the HAZ. Therefore in this study was
proposed a new approach to overcome the forecasting difficulties
of the HAZ by applying soft computing methodologies.
Forecasting of the future HAZ is complex based on different
machining parameters. An efficient learning model based upon
laser cutting parameters was developed to estimate the HAZ. Ac-
curacy level of predicted values was assessed in comparison to
ANN and GP results. The simulation results revealed that ELM
model is able to predict HAZ favorably so that it provides the most
accurate predictions. The ELM algorithm can generally be effec-
tively utilized in HAZ applications and particularly in the HAZ
estimations.
References
[1] Zhan Mei, Du Hongfei, Liu Jing, Ren Ning, Yang He, Jiang Haomin, Diao Keshan,
Chen Xinpin. A method for establishing the plastic constitutive relationship of
the weld bead and heat-affected zone of welded tubes based on the rule of
mixtures and a microhardness test. Mater Sci Eng A 2010;527:2864–74.
[2] Yilbas BS, Akhtar SS, Karatas C. Laser cutting of alumina tiles: heating and
stress analysis. J Manuf Process 2013;15:14–24.
[3] Kardas Omer Ozgur, Keles Omer, Akhtar Sohail, Yilbas Bekir Sami. Laser cut-
ting of rectangular geometry in 2024 aluminum alloy: thermal stress analysis.
Opt Laser Technol 2014;64:247–56.
[4] Huang Kuo-Cheng, Hsiao Wen-Tse, Hwang Chi-Hung, Lin Ru-Li, Yeh Jer-Liang
Andrew. The laser ablation model development of glass substrate cutting as-
sisted with the thermal fracture and ultrasonic mechanisms. Opt Lasers Eng
2015;67:31–5.
[5] Cai Yecheng, Yang Lijun, Zhang Hongzhi, Wang Yang. Laser cutting silicon-
glass double layer wafer with laser induced thermal-crack propagation. Opt
Lasers Eng 2016;82:173–85.
[6] Chunyang Zhao, Hongzhi Zhang, Lijun Yang, Yang Wang, Ye Ding. Dual laser
beam revising the separation path technology of laser induced thermal-crack
propagation for asymmetric linear cutting glass. Int J Mach Tools Manuf. 2016.
http://dx.doi.org/10.1016/j.ijmachtools.2016.04.005.
[7] Vlastimir Nikolić, Dalibor Petković, Lyubomir Lazov, Miloš Milovančević. Se-
lection of the most influential factors on the water-jet assisted underwater
laser process by adaptive neuro-fuzzy technique. Infrared Physics & Technol-
ogy 2016;77:45–50.
[8] Jebbari Neila, Jebari Mohamed Mondher, Saadallah Faycal, Tarrats-Saugnac
Annie, Bennaceur Raouf, Longuemard Jean Paul. Thermal affected zone ob-
tained in machining steel XC42 by high-power continuous CO2 laser. Opt Laser
Technol 2008;40:864–73.
[9] Dørum Cato, Lademo Odd-Geir, Myhr Ole Runar, Berstad Torodd, Hopperstad
Odd Sture. Finite element analysis of plastic failure in heat-affected zone of
welded aluminium connections. Comput Struct 2010;88:519–28.
[10] McDaniels RL, White SA, Liawa K, Chena L, McCayb MH, Liawa PK. Effects of a
laser surface processing induced heat-affected zone on the fatigue behavior of
AISI 4340 steel. Mater Sci Eng A 2008;485:500–7.
[11] Singh Ramesh, Alberts Matthew J, Melkote Shreyes N. Characterization and
prediction of the heat-affected zone in a laser-assisted mechanical micro-
machining process. Int J Mach Tools Manuf 2008;48:994–1004.
[12] Dalibor Petković, Vlastimir Nikolić, Miloš Milovančević, Lyubomir Lazov. Es-
timation of the most influential factors on the laser cutting process heat af-
fected zone (HAZ) by adaptive neuro-fuzzy technique. Infrared Physics &
Technology 2016;77:12–5.
[13] Huang GB, Zhu QY, Siew CK, Extreme Learning Machine: a new learning
scheme of feedforward neural networks. In: Proceedings of the international
joint conference on neural networks, Vol. 2, 2004, p. 985–990.
[14] Wang D, Huang GB, Protein sequence classification using Extreme Learning
Machine. In: Proceedings of the international joint conference on neural
networks, Vol. 3, 2005, p. 1406–1411.
[15] Huang GB, Zhu QY, Siew CK, Real-time learning capability of neural networks.
Technical Report ICIS/45/2003, School of Electrical and Electronic Engineering,
Nanyang Technological University, Singapore, 2003.
[16] Nian R, He B, Zheng B, Heeswijk MV, Yu Q, Miche Y, et al. Extreme learning
machine towards dynamic model hypothesis in fish ethology research. Neu-
rocomputing 2014;128:273–84.
[17] Abdullah SA, Malek AA, Abdullah NS, Kisi O, Yap KS. Extreme Learning Ma-
chines: A new approach for prediction of reference evapotranspiration. J Hy-
drol 2015;527:184–95.
[18] Mohammadi K, Shamshirband S, Motamedi S, Petkovic D, Hashim R, Gocic M.
Extreme learning machine based prediction of daily dew point temperature.
Comput Electron Agric 2015;117:214–25.
[19] Huang GB, Zhu QY, Siew CK. Extreme Learning Machine: theory and appli-
cations. Neurocomputing 2006;70:489–501.
[20] Huang GB, Chen L, Siew CK. Universal approximation using incremental
constructive feedforward networks with random hidden nodes. IEEE Trans
Neural Netw 2006;17:879–92.
[21] Liang N-Y, Huang G-B, Saratchandran P, Sundararajan N. A fast and accurate
online sequential learning algorithm for feedforward networks. IEEE Trans
Neural Netw 2006;17(6):1411–23.
[22] Koza J. Genetic programming: on the programming of computers by natural
selection.Cambridge, MA, USA: MIT Press; 1992.