Applied Thermal Engineering 129 (2018) 893–906

Contents lists available at ScienceDirect

Applied Thermal Engineering

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

Research Paper

Robust optimization for reducing welding-induced angular distortion in

fiber laser keyhole welding under process parameter uncertainty
Qi Zhou a,b,c,⇑, Yan Wang c, Seung-Kyum Choi c, Longchao Cao b, Zhongmei Gao b
a
School of Aerospace Engineering, Huazhong University of Science & Technology, 430074 Wuhan, PR China
b
The State Key Laboratory of Digital Manufacturing Equipment and Technology, School of Mechanical Science and Engineering, Huazhong University of Science & Technology,
430074 Wuhan, PR China
c
George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USA

h i g h l i g h t s

A robust optimization framework for reducing welding-induced distortion is proposed.

A three-dimensional thermo-mechanical finite element model (FEM) is developed.

The FEM is validated by laser welding experiments.

The process parameter uncertainty is quantified.

The reliability of the obtained robust optimum is verified by Monte Carlo method.

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

Article history: Welding-induced angular distortion is a typical out-of-plane distortion, which brings negative effects on
Received 12 June 2017 the joints’ quality. Therefore, the selection of appropriate process parameters to minimize or control
Revised 10 October 2017 welding-induced distortion under uncertainty has become of critical importance. In this paper, a robust
Accepted 14 October 2017
process parameter optimization framework is proposed to reduce welding-induced distortion in fiber
Available online 16 October 2017
laser keyhole welding under parameter uncertainty. Firstly, a three-dimensional thermal-mechanical
finite element model (FEM) for simulating the welding-induced distortion is developed and validated
Keywords:
by laser welding experiment. Secondly, a Gaussian process (GP) model is constructed to build the rela-
Welding-induced distortion
Thermal-mechanical analysis
tionship between the input process parameters and output responses. Finally, uncertainty quantification
Finite element model of both process parameter uncertainty and GP model uncertainty is derived. The obtained uncertainty
Gaussian process model quantification formulas are used in the robust optimization problem to minimize welding-induced dis-
tortion. The effectiveness and reliability of the obtained robust optimum are verified by the Monte
Carlo method.
Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction mechanical strength and dimensional accuracy. Fig. 1 shows a

Laser welding (LW), as an advanced manufacturing technology,
has been widely used in the aerospace, shipbuilding, energy, and
automotive industries [1–3]. It has advantages over other joining
technologies such as a high degree of automation, enhanced joint
strength, high energy density, and a narrow heat-affected zone
[4–7]. However, one of the major issues in laser welding is
welding-induced distortion during the LW process, which signifi-
cantly affects the quality of the welding joints, especially in
⇑ Corresponding author at: School of Aerospace Engineering, Huazhong Univer-
sity of Science & Technology, 430074 Wuhan, PR China.
E-mail addresses: qizhouhust@gmail.com, qi.zhou@me.gatech.edu (Q. Zhou).
schematic diagram of the laser keyhole welding process. Generally,
the welding distortions are mainly caused by the changes of the
temperature field of the welding pool during the welding process.
The most severe one is the angular distortion, as illustrated in Fig. 2
(a). Therefore, additional post-weld treatments are usually
required to correct the welding-induced distortions toward a
satisfactory level of joint strength and dimensional accuracy [8].
Post-weld treatments are always time-consuming and costly,
therefore only practical in the most critical applications. For indus-
trial applications where the financial budget consideration is criti-
cal, the best practice is to actively control the welding-induced
distortions during the welding process by selecting the optimum
welding process parameters.

https://doi.org/10.1016/j.applthermaleng.2017.10.081
1359-4311/Ó 2017 Elsevier Ltd. All rights reserved.
894 Q. Zhou et al. / Applied Thermal Engineering 129 (2018) 893–906

Nomenclature

Abbreviations MLE maximum likelihood estimate

BPNN back propagation neural network OLHS optimal Latin hypercube sampling
FCD faces centered design PI prediction intervals
FEM finite element method PSR polynomial surface regression
GA genetic algorithm P/W the ratio of weld penetration to the width
GP Gaussian process RMAE relative maximum absolute error
HSSD Hammersley sequence sampling design RMSE root mean square error
LFP laser focal position SQP sequential quadratic program
LP laser power UD uniform design
LW laser welding WS welding speed
MC Monte Carlo

at certain sample points. Metamodels are constructed to approxi-

mate the relationship between input process parameters and out-
n put welding-induced distortion. Metamodels are used in numerical
ctio Weld bead
ng dire optimization algorithms to predict the best parameter values. For
ldi
Laser beam We example, Narwadkar and Bhosle [17] applied the Taguchi method
to generate a three-level three-factor sampling plan. Then, the
3mm

Weld pool
mm
100
Keyhole
Stainless steel 316L
80mm
Fig. 1. Schematic plot of laser welding process.
Several welding process parameters, such as laser power, weld-
ing speed, laser focal position, contribute to the distortion [9,10].
Better control of the most influential process parameters will be
most effective in eliminating the distortion. Most research on con-
trolling welding-induced distortions has been done by conducting
physical experiments [11–13]. In those efforts, a large number of
experiments were conducted in a predefined process parameter
domain to obtain the welding-induced distortions. Based on the
experimental data, the process parameter values that can meet
the final weld joint requirement will be chosen. This resource-
consuming trial-and-error method often leads to sub-optimal solu-
tions [8,14,15]. To improve the efficiency, the design of experiment
and optimization approaches have been proposed [9,16,17]. In
these approaches, laser welding experiments are only conducted
optimal input parameters were obtained. Although the design of
experiment approaches help obtain the optimum process parame-
ters, the required laser welding experiments for metamodeling are
still time-consuming and costly because of the highly nonlinear
and non-smooth relationships between input and output
[8,10,18,19].
With the fast advancement of computer’s capability and speed,
computational simulation methods, e.g. finite element method
(FEM), has made it possible to replace the physical experiments
by simulating the thermo-mechanical behavior of structures
during laser welding process. Several researchers have studied
the problem of distortion through FEM in welding processes
[18,20–25]. For example, Deng and Murakawa [18] developed a
large deformation and thermo-elastic-plastic FEM for simulating
welding distortion in a low-carbon steel butt-welded joint with a
thickness of 1 mm. Wang et al. [24] employed an elastic FEM to
simulate welding distortion in the fabrication of a cantilever beam
component of a jack-up drilling rig; Manurung et al. [25] used a
linear thermal elastic FEM to analyze the welding-induced distor-
tion in combined butt and T-joints with 9-mm low carbon steel. It
was concluded from these studies that FEM is a very powerful and
reliable tool for modeling and analyzing welding processes. Never-
theless, limited research work has been conducted in integrating
FEM in process optimization. Song et al. [26] successfully per-
formed a welding residual stress minimization problem by adopt-
ing the Broyden–Fletcher–Goldfarb–Shanno (BFGS) line search

P /W = Weld Penetration / Weld Width

Wead Width
Penetration
Weld

Weld bead
Angular distortion

(a) Illustration of angular distortion (b) Illustration of P/W

Fig. 2. The illustration of angular distortion and P/W.
 Q. Zhou et al. / Applied Thermal Engineering 129 (2018) 893–906 895

Start

Numerical Welding Simulation Gaussian Process Modeling Robust Process Parameters

Geometry modeling Design of experiment by optimal
Optimization
Latin hypercube sampling Establish robust process parameters
Thermal-mechanical analysis optimization formulation
Conduct numerical welding
• Heat source model
• Material properties dependent
simulations to obtain angular Uncertainty quantification of
temperature distortion at sample points process parameters and GP model
• Initial conditions and boundary uncertainties
conditions
Obtain the P/W at sample points

Build Gaussian prosess model Implement sequence quadratic

program for optimum searching
Verify FE model by conducting Valid the accuracy of the
laser welding experiment Gaussian process model Does stopping No
Ye criteria met?
No Does simulation Yes No Does accuracy
sYes Yes
result reliable? criteria met?
Output the process parameters

End

Fig. 3. Flowchart of the proposed approach.

method; Islam et al. [8] developed a process parameter optimiza-
tion framework for a lap joint fillet weld by integrating FEM with
polynomial surface regression (PSR) and genetic algorithm (GA).
Rong et al. [9] built a back-propagation neural network (BPNN)
based on FEM results to predict the angular distortion of the
welded structures in butt joint without gap and then used the pre-
diction to reduce the weld distortion. These methods demon-
strated that it is possible to obtain optimal process parameters
with considerably less cost than physical experiments.
However, there are some shortcomings in these approaches for
FEM based process optimization. PSR is a local approximation
method and is only suitable for local optimization problems.
Machine learning techniques (e.g., BPNN) require a large quantity
of training sampling points to ensure their prediction accuracy
[27–29]. More importantly, the above-mentioned studies do not
assess the uncertainty associated with the prediction of welding-
induced distortion and optimization of process parameters is based
on the assumption of deterministic parameters. As a consequence,
there is no variability in the simulation outputs. In reality, uncer-
tainty is involved in the process parameters of laser welding. The
uncertainty may cause distortion variation and result in defective
weld joints.
To overcome the above two drawbacks related to efficiency and
uncertainty, in this work, a robust process parameters optimization
framework under uncertainty for reducing welding-induced distor-
tion in laser keyhole welding is developed. The developed frame-
work is composed of three main parts: numerical welding
simulation, Gaussian process (GP) modeling, and robust process
parameter optimization. Specifically, a three-dimensional
thermal-mechanical FEM for welding-induced distortion is devel-
oped and validated by laser welding experiments. Then, a GP model,
which can provide a confidence prediction interval for an unob-
served point and exhibits more robust prediction performance than
other metamodels with small sample sizes, is introduced for data fit-
ting. Uncertainties associated with both GP model and process
parameters are quantified. The uncertainty information is included
in the robust optimization problem. Monte Carlo (MC) method is
also used to verify the obtained robust optimum.
2. Overview of the proposed approach
The goal of the proposed approach is to identify the best possi-
ble process parameters with minimum angular distortion in fiber
laser keyhole welding under process parameter uncertainty.
Fig. 3 depicts the framework of the proposed approach, which con-
sists of three components: numerical welding simulation, Gaussian
process modeling, and robust process parameter optimization.
The first component of the framework begins with geometric
modeling of the structure. Then, thermal and mechanical analyses
are performed. Actual fiber laser keyhole welding experiments are
conducted to validate the finite element model. The details of the
numerical simulation are presented in Section 3.
Since the angular distortion can visually reflect the degree of
deformation of the workpiece and is easy to be measured, it will
be taken as the quantitative metric in this work. In fiber laser key-
hole welding, it is usually expected that the ratio of the weld pen-
etration to the width (P/W) is larger than a pre-defined threshold
value. Hence, P/W is taken as a constraint. An illustration of P/W
is shown in Fig. 2b.
In the second component, the optimal Latin hypercube sam-
pling approach (OLHS) is introduced to generate evenly distributed
samples in the design domain. Then, numerical simulations are
carried out to obtain the welding angular distortion at these sam-
ple points. P/W at these sample points are simulated by the previ-
ously developed thermal finite element model [14]. The Gaussian
process model is constructed for the relationship between process
parameters and output responses. Finally, the accuracy of the
Gaussian process model is verified by introducing local and global
accuracy metrics. The details of how to build and verify the
Gaussian process model will be presented in Section 4.
In the third component, the robust process parameter optimiza-
tion is formulated. Uncertainty quantification is performed to
obtain the means and standard deviations of angular distortion
and P/W. Finally, the gradient-based sequential quadratic pro-
gramming (SQP) is used to search the robust optimum. Section 5
will present a detailed description of the robust process parameter
optimization procedure.
896 Q. Zhou et al. / Applied Thermal Engineering 129 (2018) 893–906
Table 1
The chemical composition of stainless steel 316L (in weight) [32].
Chemical elements C Si
Composition (%) 0.0019 0.35
Grid refinement region
Symmetry plane
Fig. 4. Illustration of the 3D finite element model.
3. Finite element modeling and simulation
In this study, a three-dimensional thermo-mechanical finite
element model is developed to simulate the angular distortion of
the fiber laser keyhole welding. To obtain P/W in fiber laser key-
hole welding, the previously developed thermal finite element
model [14] is used. Assumptions are as follows,
(1) The properties of material obey the Mises yield criterion.
(2) The yielding behavior of the plastic zone is subject to the
plastic flow criterion and strengthening criterion.
(3) Elastic strain, plastic strain, and temperature strain are
inseparable.
(4) The impacts of sticky and creep are not considered during
stress changing procedure.
(5) The physical and mechanical properties of the material vary
with the temperature. The material properties are isotropic.
Table 1 shows the chemical composition in weight percent of
the base metal. According to numerous tests and surveys
[10,13,30], the angular distortion of the fiber laser keyhole welding
is mainly influenced by three welding process parameters: the
laser power (LP), the welding speed (WS), and the laser focal posi-
tion (LFP). Specifically, low laser power and high welding speed go
against the flow of molten metal, which can result in the incom-
plete fusion, the imperfection of filled groove or even spatter [4];
Excessive large focal position will reduce the laser absorptivity. It
can cause the issues of root concavity and incompletely filled
groove [31]. Therefore, the following settings are used for welding
process parameters.
LP  NðuLP ; r2LP Þ; WS  NðuWS ; r2WS Þ; 2 ðmmÞ 6 LFP 6 0 ðmmÞ
2000 ðWÞ 6 uLP 6 3000 ðWÞ; rLP ¼ 4; 41:67 ðmm=sÞ 6 uWS
6 58:33 ðmm=sÞ; rWS ¼ 0:3
where the uncertainty process parameters, LP and WS, follow a nor-
mal distribution and are assumed to be independent.
Cr Ni Mo Mn Cu S
17.07 11.95 2.04 1.68 0.14 0.007
3.1. Geometric modeling
Fig. 4 shows the mesh employed for finite element simula-
tion. With the symmetry of the model, the size of the finite ele-
ment model takes half of the true welding workpiece to improve
the computational efficiency in welding simulation. The compu-
tational domain has a dimension of 100  40  3 mm3 . The weld-
ing heat source is applied on the central or symmetry plane to
simulate the fiber laser keyhole welding process. Since the
changes of temperature and stress fields are complex at the
weld-forming region, the grids in this region are refined. As
the distance from the weld-forming region increases, the tem-
perature field changes and the stress field is weaken. A proper
coarser mesh is adopted to effectively improve the computa-
tional efficiency. The model contains 44,646 four-node hexagonal
elements and 36,308 nodes.
In the actual welding process, the workpiece is welded without
any fixture. In the simulation, the model may be deformed because
of the expansion and contraction of the material. Therefore, it is
necessary to limit the degrees of freedom for some key nodes to
prevent rigid displacement of the model. As illustrated in Fig. 4,
the degrees of freedom in z and y directions at point A are limited.
At point B, a constraint on the degrees of freedom in the z direction
is applied. The constraint on the degrees of freedom in the x direc-
tion is applied to line L1. A symmetry constraint is also applied to
the symmetric plane.
3.2. Thermo-mechanical analysis
Computationally modeling laser welding is a coupled thermo-
mechanical problem. It contains several coupling phenomena,
such as nonlinear heat flow, weld pool physics, non-linear mate-
rial behavior at high temperatures, thermal deformation, and
mechanical distortion [8]. However, in the computational model-
ing of laser welding, the general idea is to adopt a weakly cou-
pled model, in which a simplified heat input model is used to
replace the physics in the weld and the mechanical analysis is
performed independently of the thermal analysis. This method
is computationally effective and useful when the main purpose
is to study transient temperature, distortion, and stress fields
other than the complex physical properties of metal. Since the
goal of this study is to investigate welding-induced distortion,
a sequential coupling method is adopted. Only the influence of
the welding temperature field on the stress field is considered
and the influence of the welding stress field on the temperature
field is ignored.
3.2.1. Thermal analysis
In the process of laser heat transfer, obtaining the welding tem-
perature field is a nonlinear transient heat transfer problem. The
law of conservation of energy is the most basic criterion in thermal
analysis of laser welding. Therefore, during the thermal analysis of
weld pool, the force and displacement are ignored, and only the
ð1Þ energy is considered. In this step, the value of the temperature at
each node is obtained by the instantaneous heat conduction equa-
tion given as
 Q. Zhou et al. / Applied Thermal Engineering 129 (2018) 893–906 897

3.2.2. Mechanical analysis

After calculating the temperature field of fiber laser keyhole
welding, the thermal analysis units of the model are transformed
into the corresponding structural units. Then the temperatures of
the nodes are put into the structural model as thermal loading to
calculate the stress–strain field.
The stress-strain relationship of the material in the case of elas-
ticity and plasticity is [13]

a=0.90 fdrg ¼ ½Dfdeg  fCgdT ð5Þ

b=1.20 where ½D is an elastic or plastic matrix, fCg is a temperature-related
c1=0.98 matrix, and fdeg is the total strain including elastic strain, plastic
c2=1.52 strain, and thermal strain.
H=2.17 In the thermo-elastic finite element analysis of fiber laser key-
h=1.12 hole welding, the welding transient temperature field analysis
serves as the basis for the calculation of stress field. The tempera-
Fig. 5. Heat source model [14]. ture increment is loaded on each element of the finite element
model to obtain the displacement increment of each node on the




 unit. The strain increment can be expressed as [10]
@T @T @ @T @ @T @ @T
qC P v ¼ kx þ ky þ kz þQ e
fdeg ¼ ½Dfddg
e
ð6Þ
@t @x @x @x @y @y @z @z
ð2Þ e e
where fdeg is the strain increment of each unit and fddg is the dis-
placement increment of the unit node.
where q is the density of the material, C P is the specific heat
capacity, and T is the instantaneous temperature. kx , ky , and kz are 3.3. Discussion of simulation results
thermal conductivities in the directions of x, y, and z. Q is the rate
of heat production within each unit volume, and v denotes the The temperature field of fiber laser keyhole welding is divided
welding speed. into two processes, the heat source loading process and the cooling
The initial condition can be expressed as process. The heat source loading process contains 160 time steps
with there are 5 sub-steps involved in each. The cooling process
Tðx; y; z; 0Þ ¼ T 0 ð3Þ
contains 840 steps. The time integral method is used to calculate
the heat balance equation. Fig. 6 shows the cloud plot of the tran-
where T 0 is the initial temperature (T 0 ¼ 300 K).
sient temperature field at a typical process parameter
The boundary condition can be expressed by
(LP ¼ 3000 W;WS ¼ 50 mm=s;LFP ¼ 1 mm). Because the welding
@T speed of fiber laser keyhole welding is faster than those of tradi-
kn  q þ hðT  T 0 Þ þ reðT 4  T 40 Þ ¼ 0 ðx; y; zÞ 2 S t > 0 ð4Þ tional welding techniques, the smaller heat affected zone can
@n
result in a faster cooling process.
where S is the boundary to be calculated, kn is the thermal conduc- Once the analysis results of temperature field of each step are
tivity on the boundary surface S, h is the convection coefficient, e is obtained, these results will be placed into the calculation model
the radiation heat transfer coefficient, and q is the heat flux on the of the stress field. Then, the angular distortion can be obtained
boundary surface S. by adopting the time integral method. Fig. 7 shows the cloud plot
Typically, the complex physics of heat generation or weld pool of the distortion after cooling. As illustrated by Fig. 7, bending dis-
is simplified considerably and replaced by a heat input model. In tortions occur along both sides of the welding bead. The distortions
this work, the previously developed body heat source model, made at the edge of the sheet are significantly larger than those of the
by combining double-ellipsoids, rotating-Gaussian, and a cone welding region. This is because during the loading process, the
[14], is adopted for FEM simulation. The heat source model is illus- extremely high heat input makes the temperature of the fusion
trated in Fig. 5. zone significantly higher than the ones in other areas of the work-
Thermo-physical properties of the material will change as the piece. The thermal expansion of the material is hindered in produc-
temperature increases during the welding process, especially when ing plastic distortion. In the subsequent cooling process,
the temperature between liquidus and solidus [15]. Therefore, shrinkages with varying degrees occur. Uneven transverse shrink-
thermo-physical properties of stainless steel need to be set in the age finally results in angular distortion, whose direction is perpen-
finite element simulation. The specific heat capacity and thermal dicular to the welding bead.
conductivity change with temperature are listed in Table 2. The
temperature dependent material properties of stainless steel 3.4. Model validation
316L are calculated by JMatPro software, which can be used to
compute the material properties based on chemical composition The laser welding experiment was conducted on a 3 mm-thick
of material within desired temperature range. stainless steel 316L plate. To eliminate the interference from an

Table 2
Temperature dependent thermal properties of stainless steel 316L.

Thermal properties Temperature (K)

300 800 1300 1616 1660 1718 2000 2500 3000 3500
Conductivity (W=m K) 15.91 22.46 29.01 33.15 33.67 30.62 35.07 42.95 50.83 58.71
Specific heat (J=g K) 0.45 0.66 0.65 0.79 1.12 0.78 0.79 0.81 0.83 0.85
898 Q. Zhou et al. / Applied Thermal Engineering 129 (2018) 893–906
Heating process
t=0.8s
Cooling process
t=30s
Fig. 6. The cloud plots of the transient temperature field during heating and cooling processes.
oxidation film and prevent the welding bead from being polluted
by oil, the workpiece had been pretreated and degreased with ace-
tone before welding. Fig. 8 demonstrates the laser welding setup
used in this work. The laser welder utilized here was a
ytterbium-doped fiber laser device (IPG YLR-4000) with a maxi-
mum average power of 4000 W. The continuous laser travels
through the optical fiber to the laser welder head. The laser header
is installed on the robot ABB IRB4400. A focusing lens with the
focal length of 250 mm was placed in the laser welder head. The
focused laser irradiated on the specimen. The radius of the light
spot on the surface of the specimen was about 0.3 mm. The angle
between the vertical direction of the weldment and the laser beam
was set to be 8⁰. Argon with a flow rate of 1.0 m3/h was utilized as
the shielding gas during the welding.
The height vernier caliper shown in Fig. 9 was used to measure
the maximum welding angle distortion of the welded specimen.
To obtain the angle distortion of the welded specimen, the following
procedures were performed. First, the surfaces of the welded speci-
men and measuring pawl were cleaned using anhydrous alcohol.
Then, the front, middle and rear positions of the welded specimen
were marked. The measuring pawl of the height vernier caliper
was placed on the surface of the three marked points to read the
angle distortions. Finally, the maximum angle distortion among
these three points will be selected as the angle distortion of welded
specimen.
Since the angular distortion is highly dependent on the thermal
cycles of various locations of the weld and the shape and size of the
Heating process
t=1.6s
Cooling process
t=700s
transverse section of the weld, the validation of the calculated
weld profile by experimentally observed transverse section under
different process parameters was performed first. Five sample
points were randomly generated and laser welding experiments
were conducted. The comparisons of the calculated weld profile
with experimentally observed transverse section are summarized
in Fig. 10. As can be seen in Fig. 10, the welding bead profiles of
the sample points are in good agreement with the validation
experiments.
To make a comparison between the simulation results and
experimental results of angle distortions, three sample points were
randomly selected and conducted by laser welding experiment.
The experimental results, together with the corresponding simula-
tion results of angle distortions, are shown in Fig. 11. The relative
errors between experimental and simulation results are shown in
Table 3. It can be observed that the simulation results agree well
with the experimental results. That is, the three-dimensional ther-
mal finite element model is reliable and can be used in the subse-
quent process.
4. GP modeling
4.1. Design of experiments
The aim of the design of experiment (DOE) is to decrease the
effects of experimental errors on the responses, while allowing
 Q. Zhou et al. / Applied Thermal Engineering 129 (2018) 893–906
Fig. 7. The cloud plot of the displacement after cooling.
engineers to build an approximation function more efficiently. Sev-
eral DOE approaches, such as the Hammersley sequence sampling
design (HSSD), uniform design (UD), faces centered design (FCD),
and optimal Latin Hypercube sampling (OLHS), are available to
generate sample points that can provide a good coverage of the
design space. In this study, the OLHS is adopted to spread the
Fig. 8. Laser welding equipment.
Welding workpiece
Workbench
Fig. 9. Angular distortion measurement platform.
899
points throughout the design space. The OLHS is a modified LHS
where the combination of factor levels for each factor is optimized,
rather than randomly combined. OLHS provides the designer the
freedom of selecting the number of designs to run for an available
computational budget [31]. Specifically, the OLHS developed by Jin
et al. [35], where an enhanced stochastic evolutionary algorithm to
evaluate the maximin distance criterion of points in the searching
space to obtain space filling sample points, is used. The generated
sample points and corresponding simulation results of the angular
distortions (d) and P/W are summarized in Table 4.
4.2. Construction and validation of GP model
4.2.1. Fundamental of GP model
The Gaussian process is a constrained regression approach to
construct surrogate models [30–32]. GP models have been increas-
ingly used in design optimization and uncertainty quantification
areas. They can not only be applied to fit a wide range of functional
forms, but can also provide the confidence interval for a predicted
value at unobserved points. The obtained predicted deviation can
illustrate the local uncertainty of the prediction. The higher the
prediction deviation is, the less certain the prediction will be,
and vice versa. In this section, a brief introduction of the GP model
is presented. More details can be found in Ref. [30].
Generally, a GP model can be expressed as
f ðxÞ  GPðmðÞ; kð; ÞÞ ð7Þ
Height vernier
caliper
900 Q. Zhou et al. / Applied Thermal Engineering 129 (2018) 893–906

BW=1.717mm BW=1.676mm 1# BW=1.702mm BW=1.664mm

DP=2.558mm
DP=2.569mm
DP=2.805mm
DP=2.749mm

·
1# 2#

(a) 1# P=2300W,WS=43.17mm/s, LFP=-2 (b) 2# P=2700W,WS=56.83mm/s, LFP=-2

BW=1.631mm BW=1.388mm
BW=1.165mm BW=1.294mm

DP=2.113mm

DP=2.087mm
DP=2.082mm
DP=2.024mm

3# 4#

(c) 3# P=2330W,WS=51.67mm/s, LFP=0 (d) 4# P=2400W,WS=58.33mm/s, LFP=-1

BW=1.258mm
DP=2.065mm

5#

(e) 5# P=2170W,WS=52.00mm/s, LFP=0 (f) Summary of the comparison results

Fig. 10. The comparisons of the calculated weld profile with experimentally observed transverse section.

where mðxÞ denotes the mean function and kðx; x0 Þ ¼ covff ðxÞ; f ðx0 Þg f p ðxp Þ at other sampling points, xp ¼ fxp1 ; xp2    xpmp g. Subscripts o
represents the covariance function between the values of function f and p denote observations and predictions, respectively. According
at two locations x and x0 . mðxÞ is usually expressed as hðxÞb. Here, to the definition of the GP model, the joint Gaussian distribution
hðxÞ denotes a row vector of regression functions and b denotes a for f o and f p ðxp Þ can be written as
column vector of regression coefficients. " #   " #!
For a Gaussian correlation function rðx  x0 Þ ¼ fo mðxo Þ Ko K Top
n P o N ; ð9Þ
u 2
exp  j¼1 hj ðxj  x0j Þ , the corresponding covariance function f p ðxp Þ mðxp Þ K op Kp
can be given as where mðxo Þ ¼ Hb is a mean vector from observation. The
( ) covariance matrix is obtained in a similar way by evaluating the
X
u
kðx; x0 Þ ¼ r2 exp  hj ðxj  x0j Þ2 ð8Þ covariance function kð; Þ and it is partitioned to sub-matrices K o ,
j¼1 K p , and K op , corresponding to the training set covariance, test set
covariance, and training-test set covariance, respectively. When f o
where r is the process standard deviation determining the overall is known, the posterior distribution of f p ðxp Þ conditional on f o is still
magnitude of the variance and h1 ; . . . ; hu are the roughness parame- a joint Gaussian, which can be obtained with standard Bayesian
ters representing the rate at which the correlation between f ðxÞ and inference procedure as
f ðx0 Þ decays to zero as x and x0 diverge. b, r and h1 ; . . . ; hu are con-
  
sidered as hyperparameters. To obtain the hyperparameters, the f p ðxp Þjf o  N Eðf p ðxp Þjf o Þ; cov f p ðxp Þ; f p ðx0p Þjf o ð10Þ
maximum likelihood estimate (MLE) is usually adopted [33]. Once
the values of hyperparameters are selected, they will be considered where the mean value Eðf p ðxp Þjf o Þ and covariance
 
as the fixed parameters in the subsequent process. cov f p ðxp Þ; f p ðx0p Þjf o can be calculated as
o o o
Let f o ¼ ff 1 ; f 2    ; f mo g be the responses at a set of sampling ^ þ K op K 1 ðf  HbÞ
^
Eðf p ðxp Þjf o Þ ¼ hðxp Þb o o ð11Þ
points, xo ¼ fx1 ; x2    ; xomo g. The objective is to predict the values
o o
 Q. Zhou et al. / Applied Thermal Engineering 129 (2018) 893–906 901

1# WD=620 m WD=681 m

2# WD=620 m WD=665 m

3# WD=480 m WD=470 m

Fig. 11. Comparison between experimental and simulation results.

Table 3
Verification results of the finite element model.

NO. LP (W) WS (mm/s) LFP (mm) d (um) Relative error (%)

Simulation value Experiment value
1 2500 57.17 2.0 681 620 8.96
2 3100 41.67 1.0 665 620 6.77
3 3200 55.00 2.0 470 480 2.13

Relative error = |(Simulation result  Experimental result)/Simulation result|  100%.

Table 4 ^ ¼ ðHT K 1 HÞ1 HT K 1 f is the maximum likelihood estimate

where b
Simulation results of the angular distortion and P/W. o o o
of b.
NO. LP (W) WS (mm/s) LFP (mm) d (um) P/W
1 2870 56.17 2 601 1.752 4.2.2. Accuracy verification of GP model
2 2350 54.50 0 441 1.682
Based on the sample data listed in Table 4, the GP model is con-
3 2160 53.50 2 461 1.682
4 2740 44.33 2 575 1.655 structed. Fig. 12 plots the three-dimensional (3D) surfaces of the
5 2900 51.83 1 572 1.643 GP model for the angular distortion. Two error metrics, relative
6 2580 54.00 1 519 1.616 maximum error (RMAE) and relative mean error (RMEE), are intro-
7 2970 48.17 0 533 1.585 duced to measure the accuracy of the constructed GP model. Given
8 2000 47.50 1 396 1.567
a test set with N test sample points, these two error metrics can be
9 2520 55.67 2 536 1.565
10 2060 51.33 0 379 1.556 calculated as
11 2030 49.17 1 402 1.550

y  y^i

12 2940 55.17 0 548 1.540 RMAE ¼ max

i
; i ¼ 1; . . . ; N ð13Þ
13 2260 53.00 1 453 1.523 yi

14 2710 52.50 0 510 1.521

15 2130 57.33 1 424 1.512 N

!
16 2650 44.83 0 499 1.508 1 X

^i Þ

ðyi  y

RMEE ¼ ð14Þ
17 2680 46.50 1 537 1.507 N i¼1
yi

18 2550 56.67 1 513 1.506

19 2610 50.83 1 524 1.500
where yi is the actual result at the i
th
^i is the pre-
test point and y
20 2320 47.00 1 467 1.498
th
21 3000 45.50 1 590 1.490 dicted response from GP model at the i test point. RMEE reflects
22 2840 42.17 1 563 1.489 the global accuracy of the metamodel, whereas RMAE reveals the
23 2480 41.67 1 501 1.486
24 2450 43.33 2 523 1.477
local accuracy. The lower the values of RMEE and RMAE are, the
25 2810 58.33 1 588 1.453 more accurate the model is.
26 2100 42.67 1 419 1.452 To calculate the two error metrics, 6 additional random samples
27 2770 50.33 2 581 1.449 within the design domain are generated and simulated. Table 5
28 2390 48.67 0 450 1.434
summarizes the accuracy results of the GP model. It can be con-
29 2230 43.83 0 414 1.410
30 2290 57.83 1 461 1.400 cluded from Table 5 that the GP model has a desirable prediction
31 2190 46.00 2 468 1.400 performance in terms of both local and global accuracy.
32 2420 49.67 2 517 1.388

  4.3. Effects and contribution rates of process parameters on angular

cov f p ðxp Þ; f p ðx0p Þjf o ¼ K p  K op K 1 T
o K op distortion
T 1
þ ðh ðxÞ  HT K 1 T 1
T T T
o K op Þ ðH K o HÞ ðh ðxÞ The main effects of process parameters (LP, WS, and LFP) on
 HT K 1 T
o K op Þ
angular distortion are analyzed and plotted in Fig. 13(a). As can
be seen in Fig. 13(a), the angular distortion will decrease with
ð12Þ
the increase of the LFP. This is because when LFP exists, the laser
902 Q. Zhou et al. / Applied Thermal Engineering 129 (2018) 893–906

energy on the sample surface will be dispersed in a region whose
radius is larger than the radius of the spot, resulting in a decrease
in laser power density. As a result, the laser penetration ability and
scope will decline, together with a decrease in angular distortion.
To validate this observation, laser welding experiments were con-
ducted. The OLHS is used to generate 12 evenly distributed sam-
ples in the design domain. Then, laser welding experiments were
conducted on these samples and the maximum angle distortions
of the welded specimen were recorded. The experimental results
of the angular distortion for investigating the main effect of LFP,
together with the plot of the main effect of LFP on angular distor-
tion based on experimental data are provided in the Appendix A.
The LP demonstrates a reverse effect on the angular distortion
compared to that of LFP. This occurs because thermal energy input
will increase with laser power, resulting in a larger thermal elastic
Angular deformation(um)
plastic distortion. The angular distortion is not so sensitive with
respect to WS. This is attributed to two main reasons: (a) Uncer-
tainties exist in the simulation model, i.e., the differences between
the simulation and experimental results on angular distortion; (b)
The angular distortion is robust in the settings of the design span
for the welding speed.
Based on the constructed GP model, the contribution rates of
the process parameters and their interactions contribution rates
to angular distortion are also analyzed and summarized in
Fig. 13(b). In Fig. 13(b), the positive contribution rates indicate
that the corresponding output response will increase with an
increase in the discussed process parameters and decrease other-
wise. The magnitude of the bars demonstrated its degree of
importance to the bead geometrical characteristics. The symbol
\ indicates the interaction contribution rate to the output per-

P/W

WS WS
(mm (mm
/s) ) )
LP(W /s) LP(W
(a) 3D surface for angular deformation with LFP=0mm (d) 3D surface for P/W with LFP=0mm
Angular deformation(um)

P/W

WS WS
(mm (mm
) )
/s) LP(W /s)
LP(W
(b) 3D surface for angular deformation with LFP=-1mm (e) 3D surface for P/W with LFP=-1mm
Angular deformation(um)

P/W

WS WS
(mm (mm
) )
/s)
LP(W
/s) LP(W
(c) 3D surface for angular deformation with LFP=-2mm (f) 3D surface for P/W with LFP=-2mm

Fig. 12. The constructed GP model for angular distortion and P/W.
 Q. Zhou et al. / Applied Thermal Engineering 129 (2018) 893–906 903

Table 5 [34]. In this work, c is equal to 2, which represents a probability

Accuracy validation results of the GP model. of 95.5%.
Error metrics GP model for angular distortion GP model for P/W For simplicity, the uncertainty quantification process is pre-
RMAE 0.0371 0.0446
sented for angular distortion, a similar process can be used for P/
RMEE 0.0184 0.0281 W. yd ðX; VÞ is used to denote the angular distortion considering
both process parameters and GP model uncertainties, where V rep-
resents the uncertainty from GP model. Using the statistical
formance for different combination types of process parameters. method, the uncertainties of process parameters and GP model
As observed in Fig. 13(b), LP has the most significant positive are treated equally and the mean and variance value of the func-
contribution rate to angular distortion, which accounts for nearly tion for angular distortion can be derived as,
70% of the total part. LFP has the most significant negative con-
tribution rate to angular distortion (about 22%), indicating that it
lðdðXÞÞ ¼ E½yd ðX; VÞ
will be very helpful to decrease the magnitude of angular distor- ¼ E½yd ðx þ w; VÞ
tion by selecting a larger LFP. In terms of the interactions ¼ E½E½yd ðx þ w; VÞjV
ð16Þ
between process parameters, the interactions contribution rates ¼ E½yd ðx þ wÞ
of LP and LFP exhibits the strongest interactions (about 3.2%), Z
which is about twice as much as that of LFP \ WS. WS has pos- ¼ yd ðx þ wÞpðwÞdw
w
itive contribution rates to angular distortion, while its contribu-
tion is not obvious.
r2d ðdðXÞÞ ¼ Var½yd ðX; VÞ
 
¼ E y2d ðX;VÞ  E½yd ðX;VÞ2
5. Robust optimization results and discussion   
¼ E E y2d ðX; VÞjV  E½E½yd X; VÞjV 2
5.1. Optimization formulation and uncertainty quantification h i
¼ E Var½yd ðX;VÞjV  þ E½yd ðX; VÞjV 2
The goal of this study is to obtain the optimal solution where (1)  E½E½yd ðX; VÞjV 2
the variations of angular distortion are not sensitive to the uncer- h i
tainty in process parameters and the GP model, (2) the constraint ¼ E½Var½yd ðX; VÞjV  þ E E½yd ðX; VÞjV 2
of P/W is not violated due to perturbation of the process parame-
ters, even under a worst-case situation, and (3) optimality, with  E½E½yd ðX; VÞjV 2
Z Z
the above-mentioned robustness targets achieved, the optimum
¼ Var½yd ðx þ w;VÞjV pðwÞdw þ ðyd ðx þ wÞÞ2 pðwÞdw
also leads to the best objective value. With this aim, the robust w w
optimization mathematical model can be expressed as Z 2
 ðyd ðx þ wÞÞpðwÞdw
Minimize : f ðXÞ ¼ lðdðXÞÞ þ crðdðXÞÞ w
ð15Þ Z
Subject to; gðXÞ ¼ lðP=WðXÞÞ  crðP=WðXÞÞ P 1:5 ¼ s2 ðyd ðx þ wÞÞpðwÞdw þ Var ½yd ðx þ wÞ
w
where the design variable X ¼ ½LP; WS; LFP consists two parts, the ð17Þ
deterministic part x and the uncertainty part w. Since LFP has no
uncertainty, its corresponding uncertainty part is equal to zero. where s ðyd ðx þ wÞÞ is the mean square error for the GP prediction.
2

lðf ðXÞÞ and rðf ðXÞÞ are the mean and variance of the angular defor-
mation, respectively. lðP=WðXÞÞ and rðP=WðXÞÞ are the mean and
variance of the ratio of weld penetration to the width (P/W), respec-
tively. c is a constant value, which reflects the risk attitude of the
designer. A larger c implies a more conservative attitude towards
the uncertainties. When f ðXÞ is a normal process, the choice of c
corresponds to a different confidence level of prediction intervals
Angular distortion ( µ m)
Since a GP model is constructed, yd ðx þ wÞ can be obtained at min-
imal cost, The Monte Carlo Integration method is introduced to cal-
culate Eqs. (16) and (17).
pðwÞ is the joint distribution of w. When the random parame-
ters w are independent and the marginal probability density func-
tion pwi are given, the joint probability density function pðwÞ can
be calculated as

low high low high low high

LFP (mm) LP (W ) WS (mm / s )

Factors
(a) Effects of process parameters on angular distortion (b) Contribution rates of process parameters on
angular distortion

Fig. 13. Effects and contribution rates of process parameters on angular distortion.
904 Q. Zhou et al. / Applied Thermal Engineering 129 (2018) 893–906

robust optimum are smaller than the deterministic ones. More

importantly, the confirmed constraints value violates the con-
Robust objective function value

straint limit for a deterministic solution. This implies that ignoring

the uncertainty may result in an infeasible solution. On the other
hand, the confirmed constraints value at the robust solution can
safely satisfy the constraint limits. In summary, the developed
robust process parameter approach can be effectively used for
reducing weld-introduced distortion in fiber laser keyhole welding
under uncertainty in process parameters.

6. Conclusions

In this work, a robust process parameters optimization method

is developed to reduce the welding-induced angular distortion in
Iteration numbers fiber laser keyhole welding. The proposed methodology has been
successfully applied to identify the promising process parameters
Fig. 14. The iteration curve of objective function values.
in fiber laser keyhole welding on 316L stainless steel. The following
conclusions can be drawn:
Table 6
The deterministic and robust optimization solutions. (1) Simulation results of welding-induced distortion from the
developed three-dimensional thermal-mechanical FEM
LP (W) WS LFP dðXÞ P/W
show good agreement with those obtained through laser
(mm/s) (mm) (lm)
welding experiments.
Deterministic optimization 2499.08 47.83 0 467.96 1.50
(2) Since performing a computer simulation is cheaper and
Robust optimization 2219.79 53.48 0 412.99 1.56
more efficient than conducting a physical laser welding
experiment, this method can significantly reduce manufac-
turing cost and speed up the time-to-market.
Table 7 (3) Process parameter optimization design based simulation
The validation results for deterministic and robust optimization solutions. models always require a loop of design-simulation-
Deterministic Robust redesign. The direct use of these simulations to evaluate a
optimization optimization large numbers of design alternatives is impractical when
lðdðXÞÞ ðlmÞ 471.00 412.00 exploring the design space for an optimum is necessary. A
rðdðXÞÞ ðlmÞ 1.414 0.876 GP model with high local and global prediction accuracy
lðdðXÞÞ þ 2rðdðXÞÞ ðlmÞ 473.828 413.651 can replace the simulation model to reduce the computa-
lðP=WðXÞÞ 1.439 1.541 tional cost of simulation runs. Besides, a GP model also can
rðP=WðXÞÞ 0.007 0.001
lðP=WðXÞÞ  2rðP=WðXÞÞ 1.426 1.539
provide a confidence interval in prediction for un-sampled
process parameters.
(4) By analyzing the effects and contribution rates of the process
Y
d
pðwÞ ¼ pwi ðwi Þ ð18Þ parameters to angular distortion, it is found that LP has the
i¼1 most significant positive contribution rate to the angular
distortion and LFP has the most negative contribution rate
to the angular distortion, while the effects of WS on angular
distortion is relatively weak.
5.2. Optimization results and verification
(5) Verification of the obtained promising process parameters
illustrates that the mean and variance values of the angular
In this work, the gradient-based SQP is adopted to solve the
distortion are compared with the validation result, with the
above robust optimization problem. The SQP will be terminated
relative error as 0.16% and 1.35%, respectively. More
when the relative distance between the optimal process parame-
importantly, with the process parameters uncertainty taken
ters of two successive iterations is below 1%, or the pre-specified
into consideration, the proposed approach can generate a
maximum number of function evaluations is reached. Then the
robust solution that is both optimum and safe which satis-
obtained promising process parameters will be exported and veri-
fies constraint limits subject to perturbation of the uncertain
fied using the Monte-Carlo simulation.
process parameters.
Fig. 14 shows the iteration curve of the objective function
values in robust optimization. The obtained robust process param-
Overall, the proposed methodology provides a promising way to
eters and the corresponding angular distortion are listed in Table 6.
identify favorable process parameters for reducing welding-
As a comparison, the deterministic optimization solution is also
induced distortion in fiber laser keyhole welding. As a part of
recorded in Table 6. To verify the actual robustness of the obtained
future work, the consideration of other sources of uncertainties,
process parameter, 10 Monte Carlo samples of the random vari-
e.g. variations of material properties and finite elemental model
ables for the deterministic and robust solutions are plugged into
uncertainty (the difference between simulation and experiment),
the three-dimensional thermo-mechanical finite element simula-
will be investigated.
tor. The Monte Carlo samples and corresponding values of the
angular distortion and P/W are summarized in Appendix B. As
illustrated in Table 7, the mean and variance values of the angular Acknowledgements
distortion at the optimal process parameter are compared with the
validation result, with the relative error as 0.16% and 1.35%, respec- This work was supported by the National Natural Science Foun-
tively. The variances of the angular distortion and P/W at the dation of China (NSFC) under Grant No. 51775203, No. 61703385
 Q. Zhou et al. / Applied Thermal Engineering 129 (2018) 893–906 905

and No. 51705182, National Basic Research Program (973 Program) Table B.2
of China under Grant No. 2014CB046703, the Fundamental Research Monte Carlo Samples and corresponding responses for the robust optimum.

Funds for the Central Universities, HUST: Grant No. 2016YXMS272. No. LP (W) WS (mm/s) LFP (mm) dðXÞ ðlmÞ P/W
1 2215.43 53.30 0 411 1.542
Appendix A 2 2219.92 53.71 0 412 1.540
3 2222.00 53.42 0 413 1.541
4 2224.19 53.75 0 413 1.540
In this Appendix, the experimental results of the angular distor- 5 2225.96 53.25 0 413 1.543
tion for investigating the main effect of LFP are summarized in 6 2220.13 53.06 0 411 1.542
Table A.1. The main effect of LFP on angular distortion in the design 7 2213.82 53.06 0 412 1.543
domain (2 ðmmÞ 6 LFP 6 0 ðmmÞ) is analyzed and plotted in 8 2216.82 53.63 0 411 1.540
9 2215.54 53.43 0 411 1.541
Fig. A.1. 10 2219.79 53.48 0 412 1.541

Table A.1
Experimental results of the angular distortion for investigating the main effect of LFP.
Appendix C. Supplementary material
NO. LP (W) WS (mm/s) LFP (mm) d (um)
1 2800 50.50 1 620 Supplementary data associated with this article can be found, in
2 2000 53.83 1 540 the online version, at https://doi.org/10.1016/j.applthermaleng.
3 3400 46.17 0 620 2017.10.081.
4 2500 57.17 2 440
5 2700 43.83 0 320
6 3500 48.33 1 620 References
7 3200 55.00 2 680
8 3000 52.83 0 720 [1] N. Chakraborty, The effects of turbulence on molten pool transport during
9 3100 41.67 1 700 melting and solidification processes in continuous conduction mode laser
10 2900 47.17 2 640 welding of copper–nickel dissimilar couple, Appl. Therm. Eng. 29 (2009) 3618–
11 2300 51.67 0 700 3631.
12 2200 49.50 2 640 [2] Q. Zhou, Y. Rong, X. Shao, P. Jiang, Z. Gao, L. Cao, Optimization of laser brazing
onto galvanized steel based on ensemble of metamodels, J. Intell. Manuf.
(2016).
[3] G. Sproesser, Y.-J. Chang, A. Pittner, M. Finkbeiner, M. Rethmeier, Life Cycle
Assessment of welding technologies for thick metal plate welds, J. Clean. Prod.
108 (2015) 46–53.
[4] Z. Gao, X. Shao, P. Jiang, L. Cao, Q. Zhou, C. Yue, Y. Liu, C. Wang, Parameters
Angular distortion ( µ m)

optimization of hybrid fiber laser-arc butt welding on 316L stainless steel

low
Fig. A.1. The main effect of LFP on angular distortion based on experimental data.
Appendix B
The Monte Carlo samples and corresponding angular distortions
are listed in Tables B.1 and B.2.
Table B.1
Monte Carlo Samples and corresponding responses for solution for the deterministic
optimum.
No. LP (W)
1 2501.23
2 2506.42
3 2490.05
4 2502.53
5 2500.36
6 2493.85
7 2497.35
8 2500.45
9 2513.39
10 2510.16
using Kriging model and GA, Opt. Laser Technol. 83 (2016) 153–162.
[5] S. Saravanan, K. Raghukandan, N. Sivagurumanikandan, Pulsed Nd: YAG laser
welding and subsequent post-weld heat treatment on super duplex stainless
steel, J. Manuf. Process. 25 (2017) 284–289.
[6] H. Wei, Y. Zhang, L. Tan, Z. Zhong, Energy efficiency evaluation of hot-wire
laser welding based on process characteristic and power consumption, J. Clean.
Prod. 87 (2015) 255–262.
[7] M. Aissani, S. Guessasma, A. Zitouni, R. Hamzaoui, D. Bassir, Y. Benkedda,
Three-dimensional simulation of 304L steel TIG welding process: contribution
of the thermal flux, Appl. Therm. Eng. 89 (2015) 822–832.
[8] M. Islam, A. Buijk, M. Rais-Rohani, K. Motoyama, Process parameter
optimization of lap joint fillet weld based on FEM–RSM–GA integration
high technique, Adv. Eng. Softw. 79 (2015) 127–136.
[9] Y. Rong, Y. Huang, G. Zhang, Y. Chang, X. Shao, Prediction of angular distortion
LFP(mm) in no gap butt joint using BPNN and inherent strain considering the actual
bead geometry, Int. J. Adv. Manuf. Technol. 86 (2016) 59–69.
[10] Y. Rong, Y. Huang, J. Xu, H. Zheng, G. Zhang, Numerical simulation and
experiment analysis of angular distortion and residual stress in hybrid laser-
magnetic welding, J. Mater. Process. Technol. 245 (2017) 270–277.
[11] A. Bachorski, M. Painter, A. Smailes, M.A. Wahab, Finite-element prediction of
distortion during gas metal arc welding using the shrinkage volume approach,
J. Mater. Process. Technol. 92 (1999) 405–409.
[12] S. Ahir, N. Patel, K. Patel, Experimental Investigation of welding distortion of
Austenitic Stainless Steel 316 in TIG Welding (2015).
[13] Y. Rong, G. Zhang, Y. Huang, Study on deformation and residual stress of laser
welding 316L T-joint using 3D/shell finite element analysis and experiment
verification, Int. J. Adv. Manuf. Technol. (2016) 1–9.
[14] P. Jiang, C. Wang, Q. Zhou, X. Shao, L. Shu, X. Li, Optimization of laser welding
process parameters of stainless steel 316L using FEM, Kriging and NSGA-II,
Adv. Eng. Softw. 99 (2016) 147–160.
[15] M. Islam, A. Buijk, M. Rais-Rohani, K. Motoyama, Simulation-based numerical
optimization of arc welding process for reduced distortion in welded
WS (mm/s) LFP (mm) dðXÞ ðlmÞ P/W structures, Finite Elem. Anal. Des. 84 (2014) 54–64.
[16] R. Sudhakaran, V.V. Murugan, P. Sivasakthivel, Optimization of process
47.43 0 471 1.455
parameters to minimize angular distortion in gas tungsten arc welded
48.74 0 472 1.428 stainless steel 202 grade plates using particle swarm optimization, J. Eng.
48.05 0 469 1.440 Sci. Technol. 7 (2012) 195–208.
47.81 0 471 1.436 [17] A. Narwadkar, S. Bhosle, Optimization of MIG welding parameters to control
48.05 0 471 1.439 the angular distortion in Fe410WA steel, Mater. Manuf. Process. 31 (2016)
47.77 0 469 1.438 2158–2164.
47.80 0 470 1.437 [18] D. Deng, H. Murakawa, Prediction of welding distortion and residual stress in a
48.28 0 471 1.440 thin plate butt-welded joint, Comput. Mater. Sci. 43 (2008) 353–365.
48.26 0 473 1.438 [19] M.S. Choobi, M. Haghpanahi, M. Sedighi, Prediction of welding-induced
48.26 0 473 1.439 angular distortions in thin butt-welded plates using artificial neural
networks, Comput. Mater. Sci. 62 (2012) 152–159.
906 Q. Zhou et al. / Applied Thermal Engineering 129 (2018) 893–906
[20] D. Deng, FEM prediction of welding residual stress and distortion in carbon
steel considering phase transformation effects, Mater. Des. 30 (2009) 359–366.
[21] D. Deng, H. Murakawa, W. Liang, Prediction of welding distortion in a curved
plate structure by means of elastic finite element method, J. Mater. Process.
Technol. 203 (2008) 252–266.
[22] W. Liang, D. Deng, S. Sone, H. Murakawa, Prediction of welding distortion by
elastic FE analysis using inherent deformation estimated through inverse
analysis, J. Int. Inst. Weld. 49 (2005) 30.
[23] P. Yi, X. Du, Y. Kan, L. Peng, X. Lai, Modeling and experimental study of laser
welding distortion of thin metallic bipolar plates for PEM fuel cells, Int. J.
Hydrogen Energy 40 (2015) 4850–4860.
[24] J. Wang, H. Zhao, J. Zou, H. Zhou, Z. Wu, S. Du, Welding distortion prediction
with elastic FE analysis and mitigation practice in fabrication of cantilever
beam component of jack-up drilling rig, Ocean Eng. 130 (2017) 25–39.
[25] Y.H. Manurung, M.S. Sulaiman, S.K. Abas, G. Tham, E. Haruman, Investigation
on welding distortion of combined butt and T-joints with 9-mm thickness
using FEM and experiment, Int. J. Adv. Manuf. Technol. 77 (2015) 775–782.
[26] J. Song, J. Shanghvi, P. Michaleris, Sensitivity analysis and optimization of
thermo-elasto-plastic processes with applications to welding side heater
design, Comput. Methods Appl. Mech. Eng. 193 (2004) 4541–4566.
[27] T. Vafeiadis, K.I. Diamantaras, G. Sarigiannidis, K.C. Chatzisavvas, A comparison
of machine learning techniques for customer churn prediction, Simul. Model.
Pract. Theory 55 (2015) 1–9.
[28] E. Acar, Effect of error metrics on optimum weight factor selection for
ensemble of metamodels, Expert Syst. Appl. 42 (2015) 2703–2709.
[29] Q. Zhou, X. Shao, P. Jiang, Z. Gao, C. Wang, L. Shu, An active learning
metamodeling approach by sequentially exploiting difference information
from variable-fidelity models, Adv. Eng. Inform. 30 (2016) 283–297.
[30] C.E. Rasmussen, Gaussian processes in machine learning, in: Advanced
Lectures on Machine Learning, Springer, 2004, pp. 63–71.
[31] Y. Wang, B. Chaib-draa, KNN-based Kalman filter: an efficient and non-
stationary method for Gaussian process regression, Knowl.-Based Syst. 114
(2016) 148–155.
[32] L. Pu, D. Qi, L. Xu, Y. Li, Optimization on the performance of ground heat
exchangers for GSHP using Kriging model based on MOGA, Appl. Therm. Eng.
118 (2017) 480–489.
[33] F. Bachoc, Cross Validation and Maximum Likelihood estimations of hyper-
parameters of Gaussian processes with model misspecification, Comput. Stat.
Data Anal. 66 (2013) 55–69.
[34] S. Zhang, P. Zhu, W. Chen, P. Arendt, Concurrent treatment of parametric
uncertainty and metamodeling uncertainty in robust design, Struct.
Multidiscip. Optim. 47 (2012) 63–76.
[35] R. Jin, W. Chen, A. Sudjianto, An efficient algorithm for constructing optimal
design of computer experiments, J. Stat. Plan. Infer. 134 (2005) 268–287.