Optimization of Laser Welding Process Parameters of Stainless Steel 316L Using FEM, Kriging and NSGA-II

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

Advances in Engineering Software 99 (2016) 147–160

Contents lists available at ScienceDirect

Advances in Engineering Software


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

Optimization of laser welding process parameters of stainless steel


316L using FEM, Kriging and NSGA-II
Ping Jiang, Chaochao Wang, Qi Zhou∗, Xinyu Shao, Leshi Shu, Xiongbin Li
The State Key Lab of Digital Manufacturing Equipment and Technology, School of Mechanical Science and Engineering, Huazhong University of Science and
Technology, Wuhan 430074, China

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

Article history: Laser welding process parameters have significant effects on the welding bead profile and quality of the
Received 22 January 2016 welding joint. This paper proposes an integration method of process parameters optimization using finite
Revised 24 May 2016
element method (FEM), Kriging metamodels and nondominated sorting genetic algorithm II (NSGA-II) in
Accepted 5 June 2016
laser welding for stainless steel 316L. The process parameters in this study are laser power (LP), welding
Available online 10 June 2016
speed (WS) and laser focal position (LF). Firstly, a three-dimensional thermal finite element model is de-
Keywords: veloped to obtain the simulated results of bead width (BW) and depth of penetration (DP). Then, Kriging
Laser welding metamodels are constructed to reflect the relationship between input process parameters and output re-
Process parameters optimization sponses. Finally, NSGA-II is used to search for multi-objective Pareto optimal solutions. In addition, the
Finite element method main effects and contribution rates of multiple process parameters on welding bead profile are analyzed.
Kriging The results of verification experiments indicate that the optimal process parameters are effective and
NSGA-II
reliable for producing expected welding bead profile.
© 2016 Elsevier Ltd. All rights reserved.

1. Introduction steels. Zhou et al. [6] optimized process parameters of galvanized


steel in laser brazing using genetic algorithm (GA) and ensemble of
Laser welding technology is widely used in the field of auto- metamodels (EMs). Although these optimization methods can ob-
motive, aerospace, shipbuilding, energy, electronics and medical tain the desired process parameters, the required laser welding ex-
industries. The advantages of laser welding include precise weld- periments is a time-consuming and costly work.
ing energy control, small heat-affected zone, low heat distortion, Finite element method (FEM), as a promising way to investigate
high welding speed and deep penetration [1]. In the laser welding welding process, has been widely used in computer simulation
process, welding process parameters (laser power, welding speed, field. Shanmugam et al. [7,8] predicted the weld bead geometry by
laser focal position, design of welded joints) play decisive role on FEM in laser spot welding of AISI 304 stainless steel sheet of thick-
welding bead profile (bead width, depth of penetration and bead ness 2.5 mm. Han et al. [9] set a three-dimensional finite element
reinforcement). It is very important to determine process parame- model to analyze the temperature distribution weld bead shape in
ters that are highly relevant to impact welding bead profile. How- laser beam welding process. Recently, researchers combined pro-
ever the relationships between the process parameters and bead cess parameters optimization with FEM. Bag et al. [10] developed
profile cannot be expressed explicitly, which makes it impracti- a three-dimensional heat-transfer model using FEM and coupled
cal to determine the optimal process parameters intuitively. Mean- with GA for identification of uncertain input parameters for the gas
while, trial and error methods often lead to sub-optimal solution tungsten arc welding process. Acherjee et al. [11] developed math-
and to some extent may result in a tremendous waste of resources. ematical models which were based on simulation results of FEM
To grasp the relationships between the welding process param- to investigate the influence of process parameters on output re-
eters and the bead performance, polynomial response surface (PSR) sponses. Acherjee et al. [12] used an FEM-RSM integrated approach
methods, artificial neural networks (ANN), Radial basis function to obtain optimal process parameters for simultaneous laser trans-
(RBF) have been widely studied [2–4]. Khan et al.. [5] developed mission welding of polycarbonates. Wang et al. [13] used ANN to
mathematical models between parameters and weld characteristics establish the relationship between welding process parameters and
in process parameters optimization for laser welding of stainless welding bead profile based on FEM for laser penetration welding,
and the model is optimized by GA. Mohammad al [14] predicted
the weld performance using FEM and ANN and obtained the opti-

Corresponding author. Fax: +86 27 87543074. mal welding parameters by multi-objective optimization in friction
E-mail address: qizhouhust@gmail.com (Q. Zhou).

http://dx.doi.org/10.1016/j.advengsoft.2016.06.006
0965-9978/© 2016 Elsevier Ltd. All rights reserved.
148 P. Jiang et al. / Advances in Engineering Software 99 (2016) 147–160

stir welding. Islam et al. [15] used PSR to study the relative in- Table 1
Heat source model parameters.
fluence of different parameters on welding distortion based on a
three-dimensional finite element model and obtained optimal so- Parameter name Value (mm) Parameter name Value (mm)
lution with the goal of minimizing the welding distortion using GA
a 0.90 H 2.17
in arc welding. b 1.20 h 1.12
To some extent, these methods could improve the efficiency c1 0.98 c2 1.52
of optimization and obtain optimal process parameters. However,
PSR belongs to local modeling methods, which is only suitable for
local optimization problems, while machine learning techniques metamodels can be used for process parameters optimiza-
(e.g. RBF and ANN) need a large number of training sampling tion, otherwise, go to Step 4.
points to ensure their accuracy [16,17]. Kriging metamodel, as a Step 6 Implement NSGA-II to obtain the optimum process pa-
global surrogate modeling technique, has been widely used in the rameters. During the NSGA-II optimization, the fitness values
field of structural optimization due to its capability of interpolat- assigned to the populations are set according to predicted
ing the sampling points and filtering noisy data [18,19]. On the value using Kriging metamodels.
other hand, nondominated sorting genetic algorithm II (NSGA-II) Step 7 Output the optimum process parameters and verify their
is an efficient technique for multi-objective optimization problems. validity through experiments.
It makes use of fast non-dominated sorting approach, elitist strat-
egy, and a crowded comparison operator to create Pareto optimal 3. Finite element simulation
solutions [20]. In this paper, a three-dimensional thermal finite el-
ement model is developed to simulate the distribution of temper- In this study, a three-dimensional thermal finite element model
ature field that can determine welding bead profile. Kriging meta- is developed to simulate the distribution of temperature field in
models are used to construct the relationship between input pa- welding process. Schematic diagram of laser welding process is
rameters and output response. Then, main effects and contribution shown in Fig. 2. Due to complex changes in the actual welding pro-
rates of multi-parameters on bead profile are analyzed. NSGA-II is cess, simplifying assumptions are made as follows [21,22]:
used to obtain the optimal solutions with desirability function. In
addition, experiments verify the effectiveness of optimal solutions. (1) The initial temperature of the weldment is 300 K.
The summary of Kriging metamodels and NSGA-II are provided in (2) The finite element model is heat conduction model, which
the “Appendix”. the formation of keyhole is not considered. The thermal
The remaining sections of the paper are as follows. In Section 2, transfer of weldment is of solid-solid, and there is no chem-
the proposed approach is introduced. Then, a thermal finite ele- ical reaction, agitation or oxidation in the melt pool.
ment model is developed in Section 3. In Section 4, Kriging meta- (3) Heat conduction between the weldment and the test table
models are constructed and the main effects of multiple param- are not considered. There is only heat convection between
eters on the bead profile are analyzed. The multi-objective opti- the weldment boundary and air. The radiation heat transfer
mization problem of laser welding process parameters is presented is coupled to heat convection.
in Section 5, followed by a conclusion in Section 6.
3.1. Constructing thermal finite element model

2. Methodologies of process parameters optimization Fig. 3 shows the mesh is built for finite element simulation. The
finite element model takes half of weldment by the centerline of
In this section, a method combining Kriging metamodel and welding bead. The computational domain has a dimension of 6 ×
NSGA-II is proposed where Kriging metamodel is used to estab- 4 × 3 mm3 . Simulation results prove that the size of mesh fully
lish the relationships between the input welding process param- meet the region of welding pool, and temperature can be reduced
eters and welding bead profiles; NSGA-II is used to facilitate de- to 300 K at the boundary of the model. The size of mesh is 0.1 mm
sign space exploration and global optimum search. The overall op- and consists of 66,729 elements.
timization process using FEM, Kriging and NSGA-II methods is de- The most commonly used heat sources of laser welding sim-
picted in the flowchart of Fig. 1. The details steps are as follows: ulation have a Gaussian distribution [23,24]. In this study, the
body heat source model combined by double-ellipsoids, rotating-
Step 1 Determine the laser welding process parameters opti- Gaussian and cone is developed for FEM simulation, shown in
mization problem including the target function, the amount Fig. 4. The popular double ellipsoid heat source model, first
of parameters and parameter ranges. The input variables in- proposed by Goldak et al. [25], is mainly used to control the
clude laser power (LW), welding speed (WS) and laser focal weld width, and the rotating-Gaussian body heat source model
position (LF). The output parameters are depth of penetra- is adopted to control the weld depth [10]. Meanwhile the circu-
tion (DP) and bead width (BW). The objective is to obtain lar truncated cone heat source model is embedded to rotating-
the maximized DP and minimized BW. Gaussian body heat source model for accurately controlling the pa-
Step 2 Construct a three dimensional thermal finite element rameters of laser focus position.
model to investigate the distribution of temperature field of The parameters of the heat source model are closely related to
laser welding process. welding bead profile. The parameters of the heat source model are
Step 3 Check whether the desired level of accuracy for FE model listed in Table 1. In the thermal finite element simulation of laser
is achieved or not. If yes, the constructed finite element welding process, the most basic criteria is the law of energy con-
model can be used for numerical simulation, otherwise, ad- servation. During the simulation process, the value of temperature
just the finite element model according to heat sources. at each node obtained by the transient heat conduction equation
Step 4 Adopt optimal Latin hypercube for design of experiment which is shown below [26]:
(DOE) and simulated the finite element model to obtain the        
∂T ∂T ∂ ∂T ∂ ∂T ∂ ∂T
output responses at the generated sample points. ρCP −v = k + k + k +Q
Step 5 Check whether the desired level of accuracy for Kriging ∂t ∂x ∂x x ∂x ∂y y ∂y ∂z z ∂z
metamodels is achieved or not. If yes, the obtained Kriging (1)
P. Jiang et al. / Advances in Engineering Software 99 (2016) 147–160 149

Fig. 1. Flowchart of the proposed approach.

Laser head

Stainless steel Direction o


fw elding Weld bead

z
y x

Fig. 2. Schematic diagram of laser welding process.

Fig. 3. Mesh used for finite element model.


150 P. Jiang et al. / Advances in Engineering Software 99 (2016) 147–160

Table 2
Chemical composition of stainless steel 316L.

Chemical elements C Si Cr Ni Mo Mn Cu S

Composition (%) 0.0019 0.35 17.07 11.95 2.04 1.68 0.14 0.007

3.2. Discussion of simulation results


z
The finite element simulation for laser welding process is calcu-
lated by FLUENT 13.0 software in which about 15.5 h were totally
c2 y spent for computation. The three-dimensional transient tempera-
c1 ture distribution at the welding interface after 5 ms is shown in
Fig. 5. Fig. 5(a) shows isothermal surface with all temperatures.
a x Fig. 5(b) shows the shape of molten pool with isothermal sur-
face which is having temperature above the fusion temperature of
the material (T > 1616 K). As shown in Fig. 5, a series of isother-
b mal surface is expressed as temperature changes near the melt
pool in three-dimensional transient temperature contours. It can
H be observed that the maximum temperature is 3500 K, which is
h far higher than fusion temperature of stainless steel. The tempera-
ture gradient of front half melt pool is larger than the latter part,
and temperature gradient drops away from the center of the light
spot. As shown in Fig. 6, the shape of molten pool can be directly
observed. Fig. 6(a) shows the whole three-dimensional transient
Fig. 4. Heat source model. temperature contours by mirror symmetry. Fig. 6(b) shows whole
isothermal surface with temperature above the fusion temperature
Table 3 (T > 1616 K) by mirror symmetry.
Thermal properties of stainless steel
316L.
3.3. Verification of finite element model with experiments
Thermal properties Value
Fig. 7 demonstrates the laser welding equipment used in this
Density (g/cm3 ) 7.91
Solid temperature (K) 1616 study. The laser welder adopted here is IPG YLR-40 0 0 ytterbium-
Liquid temperature (K) 1718 doped fiber laser. The maximum average power of the laser is
40 0 0 W, which is a continuous laser. Laser is delivered through the
optical fiber to the laser welder head. The laser welding header is
where x, y and z are the coordinate values of the Cartesian coor- installed on the robot ABB IRB4400. The angle between the vertical
dinate system, T is the temperature(K), t is the time (s), ρ is the direction of the weldment and laser beam is 8°, and shielding gas
material density, CP is the specific heat capacity, kx , ky and kz are is argon with flow rate of 1.0 m3 /h. Fig. 8 shows the dimensions
the thermal conductivities in the x, y, z directions, Q is heat gener- of weldment and schematic of welding bead profile. The dimen-
ation rate, v is the velocity of welding process. sion of specimens is 200 mm × 100 mm × 3 mm. The bead width
The initial condition can be defined as: (BW) and depth of penetration (DP) are selected to measure weld-
ing bead profile, as shown in Fig. 8. The laser focal position (LF) is
T (x, y, z, 0 ) = T0 (2) shown in Fig. 9. The LF is the distance between focal point and top
The boundary condition can be expressed as: surface of the weldment. When the focal point is at the top surface
of the weldment, the LF is defined as 0, and above the top surface,
∂T   the value of LF is negative, otherwise, it is positive [28].
kn − q + h(T − T0 ) + σ ε T 4 − T04 = 0(x, y, z ) ∈ S t > 0 (3)
∂n Three randomly selected process parameters are executed to
where S is the surface boundary, kn is the thermal conductivity, obtain the macrograph of weldment by laser welding experiment.
h is the coefficient of convection heat, σ is a Stefan–Boltzmann As shown in Fig. 10(a), the welding bead profile is marked with
constant for radiation, the value is 5.67 W/(m2 K4 ), ɛ is the heat red dotted line in experiments results; corresponding simulation
radiation coefficient, T is the material surface temperature, T0 is result of welding bead profile which marked with solid black line
the initial temperature (300 K), q is the heat flux of boundary. (T = 1616 K) is shown in Fig. 10(b). The errors between experimen-
The chemical composition of stainless steel 316L is shown in tal results and simulation results are shown in Table 5. It can be
Table 2. The physical properties and parameters of the stainless observed that the simulation results have a good agreement with
steel 316L are shown in Table 3. Thermo-physical properties of ma- the experimental results. In the other words, the accuracy of three-
terial will change as the temperature is increasing during weld- dimensional thermal finite element model can satisfy the require-
ing process, especially when the temperature between liquidus and ments of simulation.
solidus [27]. Therefore, the changes of thermo-physical properties
for stainless steel 316L need to be set in finite element model. The 4. Design of experiments and construction of Kriging
specific heat capacity and thermal conductivity change with tem- metamodels
perature are shown in Table 4. The temperature dependent ma-
terial properties of stainless steel 316L are calculated by JMatPro 4.1. Design of experiments
software, which can be used to compute the material properties
based on chemical composition of material within desired temper- Three laser welding process parameters, the laser power (LP),
ature range. welding speed (WS) and laser focal position (LF), are selected as
P. Jiang et al. / Advances in Engineering Software 99 (2016) 147–160 151

Table 4
Temperature dependent thermal properties of stainless steel 316 L.

Thermal properties Temperature (K)

300 800 1300 1616 1660 1718 20 0 0 2500 30 0 0 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

Fig. 5. Transient temperature contours after 5 ms.

Fig. 6. Whole transient temperature contours by mirror symmetry after 5 ms.

Table 5
The validation of thermal finite element model.

No. LP (kW) WS (m/min) LF (mm) Parameters Simulated value Experiment value Error (%)

1 2.40 3.50 −1 BW (mm) 1.388 1.631 14.90


DP (mm) 2.087 2.113 1.23
2 2.30 2.59 −2 BW (mm) 1.676 1.717 2.39
DP (mm) 2.805 2.749 2.04
3 2.70 3.41 −2 BW (mm) 1.664 1.702 2.23
DP (mm) 2.558 2.569 0.43
152 P. Jiang et al. / Advances in Engineering Software 99 (2016) 147–160

Fig. 7. Laser welding equipment.

BW
Stainless Steel Weld Bead
100mm

Position of welding

DP

200mm 3mm

Fig. 8. The dimensions of weldment and schematic of welding bead profile.

Table 6
The ranges of welding process
parameters.
Laser Head Parameter MAX MIN

LP (kW) 3.5 2.0


WS (m/min) 3.5 2.5
LF (mm) 0 −2

The method of optimal Latin hypercube is adopted for design


of experiments to obtain the relationship between input parame-
Focal Position

ters (LP, WS and LF) and output response (BW and DP) [29]. The
+ ranges of input parameter variables are shown in Table 6. Sixteen
sets of sample points which were obtained by the method of opti-
0 mal Latin hypercube and the simulation results of output response
- for these sample points are listed in Table 7.

4.2. Construction of Kriging metamodels

In this paper, Kriging metamodels are used to construct the


complex nonlinear relationship between process parameters and
welding bead profile. This procedure is run in Matlab R2010b and
Fig. 9. Illustration of position for laser focal.
function “dacefit” in the DACE toolbox is used to upload input and
output parameters. For the Kriging metamodels, a zero-order poly-
input parameter variables. Two parameters of welding bead profile, nomial regression function regpoly0 is applied for the global part
the bead width (BW) and depth of penetration (DP) are selected as and a Guassian correlation function is employed for the local de-
the output response. partures. Each variable has a θ value with the initial value being
P. Jiang et al. / Advances in Engineering Software 99 (2016) 147–160 153

Fig. 10. Comparison of experimental results and simulation results.

set to 1 and the range is from 1e−4 to 200. MSE of 1 × 10−10 be- results of PSR, RBF and Kriging metamodel, the minimum rela-
tween predicted and simulated values is set as termination condi- tive error results among the three metamodeling approaches are
tion. shown in bold. The average error of Kriging metamodels for BW
16 sets of sample points (nos. 1–16) are used to establish Krig- and DP are 2.70% and 1.89%, the maximum error of BW and DP
ing metamodels. The constructed Kriging metamodels for DP and are 8.13% and 5.64%, as shown in Table 10. It can be observed
BW are demonstrated in Fig. 11. Six sets of sample points (nos. 1– that the Kriging metamodel is the best among the three meta-
6) are randomly selected to verify the accuracy of Kriging meta- modeling methods for predicting welding bead profile, and the
models. Comparison of predicted values and simulated values is range of relative error is acceptable. Therefore, the Kriging meta-
shown in Fig. 12. There is a good agreement between predicted models are reliable and can be used for predicting welding bead
values and simulated values. Table 8 demonstrates the predicted profile.
154 P. Jiang et al. / Advances in Engineering Software 99 (2016) 147–160

BW (mm) BW (mm)
1.450 1.500

1.490
1.500 1.400 1.520

1.500 1.480
1.400 1.350
1.480 1.470
BW (mm)

BW (mm)
1.300
1.300
1.460 1.460
1.200
1.250 1.440 1.450
1.100 1.420
1.440
1.200
1.000 1.400
0 3.5 1.430
-0.5 3.5 1.150 3.5
3 1.420
LF -1 LP
(m 3 ) 3 )
m) -1.5 /min 1.100 (kw 2.5 /min 1.410
W S (m )
WS (m
-2 2.5 2 2.5

(a) Kriging metamodels for BW (LP=2.0kw) (b) Kriging metamodels for BW (LF=0)
BW (mm) DP (mm)
1.600

1.550 2.700
1.700 2.800

1.500
1.600 2.600
2.600
BW (mm)

DP (mm)
1.500 1.450
2.400 2.500
1.400 1.400
2.200
2.400
1.300 1.350

1.200 2.000
0 1.300 0 2.300
-0.5 3.5 -0.5 3.5
LF -1 3 1.250 LF -1 3
(m (m 2.200
m) -1.5 m) -1.5 2.5
2.5 w) kw)
-2 2 LP (
k
-2 2 LP (

(c) Kriging metamodels for BW (WS=2.6 m/min) (d) Kriging metamodels for DP (WS=2.6 m/min)
DP (mm) DP (mm)

2.800

2.350 2.300 3.000


2.700
2.300 2.800
DP (mm)

DP (mm)

2.250 2.250 2.600 2.600

2.200 2.400 2.500

2.150 2.200 2.200


2.400
2.100 2.000
3.5 0
3.5 2.150 -0.5 2.300
3 3.5
LP -1
LF
(kw 2.5 3 in) 3 )
) (m/m
(m
m) -1.5 (m /min 2.200
2 2.5 WS -2 2.5 WS
(e) Kriging metamodels for DP (LF=0) (f) Kriging metamodels for DP (LP=2.0kw)

Fig. 11. Kriging metamodels for BW and DP.

1550 3.000

1.475
Simulated value(mm)

Simulated value(mm)

2.800

1.400
2.600

1.325

2.400
1.250

2.200
1.175

1.100 2.000
1.100 1.175 1.250 1.325 1.400 1.475 1.550 2.000 2.200 2.400 2.600 2.800 3.000
Predicted value(mm) Predicted value(mm)
(a) BW (b) DP

Fig. 12. Comparison of predicted values and simulated values.


P. Jiang et al. / Advances in Engineering Software 99 (2016) 147–160 155

Fig. 13. Main effects of each process parameters on welding bead profile.

Table 7 The main effects of each process parameter on welding bead


The simulation results of 16 sets of sample points.
profile are shown in Fig. 13. As shown in Fig. 13(a), BW is almost
No. LP (kW) WS (m/min) LF (mm) BW (mm) DP (mm) in proportion to factor LP. BW has an upward trend as LP increases.
Besides, as factor LF and WS increases, BW will descend firstly and
1 2.60 3.0 0 1.502 2.238
2 2.70 3.0 −1 1.352 2.697
then ascend slightly. Fig. 13(b) shows DP has a downward trend
3 3.40 3.2 −1 1.509 2.718 with factor LF increasing except the initial small period. As factor
4 2.90 2.5 −1 1.508 2.738 LP and WS increases, DP will ascend firstly and then descend from
5 2.30 3.5 −1 1.135 2.650 the middle of the interval of factor LP and WS.
6 3.30 3.2 0 1.456 2.147
Besides the above analysis of the main effects of input pro-
7 3.20 2.8 0 1.454 2.184
8 2.20 2.6 0 1.437 2.187 cess parameters on output responses, the contribution rates of the
9 3.50 2.7 −1 1.531 2.733 process parameters and their interactions contribution rates to
10 3.10 2.8 −2 1.575 2.867 welding bead profile are also discussed in Fig. 14. The blue bar
11 3.00 3.4 −2 1.444 2.700 (positive contribution rates) indicates a positive correlation be-
12 2.80 3.4 −1 1.298 2.693
tween the process parameters and output response, while the red
13 2.00 2.9 −1 1.131 2.638
14 2.10 3.3 −1 1.087 2.635 bar means a negative correlation. LP-WS, LF-LP and LF-WS in-
15 2.50 2.6 −2 1.335 2.802 dicate the interaction contribution between two input parame-
16 2.40 3.1 −2 1.466 2.835 ters and output response respectively. In addition, the magnitude
of the bar demonstrates its degree of importance to the bead
profiles.
As illustrated in Fig. 14(a), LP has the largest positive contribu-
4.3. Main effects and contribution rates analysis of multi-parameters tion rate (a bit more than 30%) to BW, and LP-WS, LF and WS have
on bead profile the most significant negative contribution rate to BW with a per-
centage of approximately −17.5%, −17% and −15%, respectively. In
The main effect of a factor on the response is defined as the Fig. 14(b), it can be observed that factor LF has the most significant
average response of all the tests of a factor. Analysis of variance positive contribution rate to DP (about 41.5%), and LP-WS has neg-
(ANOVA) is an effective method used for obtaining the main effect ative contribution rates to BW with a percentage of approximately
of factors on responses [33]. In this part, ANOVA is adopted to an- −12%. LP-WS, LP, WS and LF-LP affect DP with a percentage of all
alyze the main effects of welding process parameters (LP, WS and under 4%, no matter positively or negatively.
LF) on welding bead profile (BW and DP).

Table 8
Predicted results of PSR, RBF and Kriging metamodels.

No. Simulated value (mm) PSR RBF Kriging

Predicted value (mm) Relative error Predicted value (mm) Relative error Predicted value (mm) Relative error

1 BW 1.457 1.547 6.16% 1.494 2.51% 1.433 1.67%


DP 2.186 2.295 4.98% 2.154 1.44% 2.153 1.51%
2 BW 1.370 1.462 6.73% 1.527 11.45% 1.443 5.34%
DP 2.686 2.903 8.09% 2.885 7.41% 2.838 5.64%
3 BW 1.530 1.538 0.55% 1.484 3.03% 1.523 0.47%
DP 2.751 2.602 5.43% 2.725 0.95% 2.741 0.35%
4 BW 1.487 1.470 1.11% 1.499 0.79% 1.493 0.44%
DP 2.722 2.596 4.62% 2.734 0.46% 2.717 0.17%
5 BW 1.381 1.260 8.76% 1.157 16.23% 1.493 8.13%
DP 2.181 2.277 4.42% 2.070 5.11% 2.260 3.64%
6 BW 1.154 1.252 8.47% 1.164 0.88% 1.152 0.17%
DP 2.645 2.584 2.29% 2.648 0.11% 2.645 0.02%
156 P. Jiang et al. / Advances in Engineering Software 99 (2016) 147–160

Fig. 14. Contribution rates of the process parameters to welding bead profile.

Table 9
Settings of NSGA-II parameters. 2.90
Parameter Value
2.85
Population size 100
Maximum iterations 500
Elite count 2
Pareto fraction 0.4 2.80
DP(mm)

Crossover fraction 0.8


Mutation fraction 0.2
Function tolerance 10E−06 2.75
Scaling fitness function Rank

2.70
5. Multi-objective optimization using NSGA-II

2.65
5.1. Mathematical model of parameters optimization

In the study, the objective of parameters optimization is min- 2.60


imum BW and maximum DP. Therefore, it can be regarded as 1.00 1.10 1.20 1.30 1.40 1.50 1.60
multi-objective optimization problem. The program is run in Mat- BW(mm)
lab R2010b. The mathematic model of multi-objective problem can
Fig. 15. Pareto-optimal front for BW and DP obtained with NSGA-II.
be developed as below:
Objective : Minimize BW(LP, WS, LF)
Maximize DP(LP, WS, LF ) (4) corresponding DP is 2.62 mm, respectively. In general the desired
Subject to : 2.0 ≤ LP ≤ 3.5, 2.5 ≤ WS ≤ 3.5, −2 ≤ LF ≤ 0 BW value is obtained at the expense of a satisfied DP value.
Since there are trade-offs among these two objectives, the op-
timization problem in Eq. (4) generally has a set of optimum so- 5.3. Discussion about the results of verification experiments
lutions. These solutions are optimum in the Pareto sense, that is,
there is no optimum better than other designs in all objectives. Four solutions in the Pareto-optimal front are selected and
In this section, NSGA-II is implemented to solve the laser welding carried out to verify the effectiveness of the optimal results.
process parameters optimization problem and obtain the Pareto- Fig. 16 shows the welding bead profile and dimension of these
optimal front. The settings for the NSGA-II in this two optimiza- four verification experiments. The optimal parameters and cor-
tion problem are listed in Table 9. Fitness values of BW and DP in responding outputs responses are listed in Table 10. As illus-
NSGA-II optimization process are predicted by Kriging metamodel trated in Table 10, the maximum relative error of BW is 5.63%.
constructed in the previous steps. For DP, the minimum relative error is 11.77%, while the rela-
tive error of two optimal results compared with the experiment
5.2. The optimization results result are 28.40% and 28.09%, attributing to the minor differ-
ences between practical process parameters and optimums, non-
Fig. 15 plots the obtained Pareto-optimal front for BW and DP. zero gap, etc. It can indicate that the prediction accuracy of pro-
Each point in Fig. 15 illustrates a specific optimal solution, whose posed approach can meet the requirement of process parameters
corresponding process parameters can be adopted according to optimization.
the requirements of the design makers. It can be concluded from In terms of the weld appearances, it can be observed in
Fig. 15 that all optima are non-inferior with each other because Fig. 16 that the weld appearances four optimal results are beau-
there are trade-offs among the BW and DP, e.g., the maximum DP tiful and few defects exist in the welding bead. For compari-
obtained is nearly 2.87 mm, while the corresponding BW is deteri- son, welding bead profile of three random process parameters is
orated obvious. The minimum BW obtained is 1.07 mm, while the also depicted in Fig. 17. It can be seen that there exist porosity,
P. Jiang et al. / Advances in Engineering Software 99 (2016) 147–160 157

Fig. 16. Weld appearances and dimensions of weld cross-sections obtained from verification experiments.

Table 10
Results of the verification experiments.

No. LP (kW) WS (m/min) LF (mm) Result DP (mm) BW (mm)

1 2.6 3.3 −1 Optimal result 2.680 1.240


Experiment result 2.373 1.314
Relative error (%) 12.94 5.63
2 2.1 3.3 −1 Optimal result 2.635 1.087
Experiment result 2.057 1.112
Relative error (%) 28.09 2.25
3 2.2 2.7 −2 Optimal result 2.807 1.305
Experiment result 2.186 1.312
Relative error (%) 28.40 0.56
4 2.6 2.9 −2 Optimal result 2.867 1.451
Experiment result 2.565 1.410
Relative error (%) 11.77 2.89

Fig. 17. Weld appearances and cross-sections obtained from random parameters.
158 P. Jiang et al. / Advances in Engineering Software 99 (2016) 147–160

Fig. 18. Microstructures of weld zone and interface between weld zone and base materials for optimal result.

collapse, spatter, undercut and root humping in welding bead, between low expense and high accuracy. The method will provide
which can affect the quality of welding joints highly. an appropriate method of process parameters optimization to im-
It is important to point out that some invisible defects which prove the efficiency and quality of laser welding process in auto-
lead to reduce the quality of welding joints may exist in the op- motive industry.
timal welding bead To further confirm the quality of the welding
bead from optimal results, the microstructures of the optimal re- Acknowledgments
sults are obtained using metallographic microscope KEYENCE VHX-
10 0 0C at 10 0x and 200x, shown in the Fig 18. As illustrated in This research has been supported by the National Basic
Fig 18, the direction of the columnar grain is varied. Besides, the Research Program (973 Program) of China under grant no.
fusion zone between the weld zone and base material is narrow, 2014CB046703, the National Natural Science Foundation of China
which is beneficial for the quality of welding joints. Therefore, the (NSFC) under grant no. 51505163, No. 51323009 and No. 51421062,
obtained microstructures of the optimal results conform to the re- and the Fundamental Research Funds for the Central Universities,
quirements. HUST: grant no. 2014TS040. The authors also would like to thank
the anonymous referees for their valuable comments.
6. Conclusions
Appendix
In this study, an integrated methodology by combining FEM,
Kriging metamodels and NSGA-II is proposed for process param- A.1. Background of Kriging metamodels
eters optimization in laser welding for stainless steel 316L. In ad-
dition, the main effects and contribution rates of multiple process Kriging metamodels, an interpolative Bayesian metamodeling
parameters on welding bead profile have been analyzed. It is obvi- technique, was first developed to aid geologists estimating min-
ous that the approach can effectively produce the expected weld- eral deposits by taking only a few samples rather than unearthing
ing bead profile and provide a reliable guidance for laser welding whole plots of land [30]. It combines a global model plus local de-
experiments. Following conclusions can be drawn from the above viations:
investigation:
y (x ) = f (x ) + Z (x ) (5)
1. The constructed three-dimensional thermal finite element
where y(x) is a known polynomial function, z(x)is a stochastic pro-
model is validated with actual laser welding experiment, and
cess with mean zero and nonzero covariance. The nonzero covari-
the simulated results show good agreement with experiment
ance z(x) is given by:
results.
    
2. The constructed Kriging metamodels between process param- Cov Z xi , Z x j = σ 2 R([R(xi ,x j )] ) (6)
eters (LP, WS and LF) and welding bead profile (BW and DP)
have a small relative error with the simulated results. where R is the correlation matrix. R(xi ,xj ) is the correlation func-
3. The ANVOA of input parameters on responses show that (a) LP tion between any two sample points xi andxj . In this paper, we use
and WS have significant effect on BW, while these have less the Gaussian correlation function of the form:
influence on DP. (b) LF has the most positive contribution rate  

n   2
to DP, while its effect on BW negatively. R(x , x ) = exp
i j
θ i
k xk − xkj (7)
4. Verification experiments of obtained optimal process param- k=1
eters have proved that welding bead profile of experimental
results is in good agreement with optimization results, and where n is the number of design variables, θ k are the unknown
the weld quality of optimized solutions are better than un- correlation parameters to be determined.
optimized solutions. Predicted estimates yˆ(x ) at untried values of x are calculated
by:
In the future study, multi-fidelity models will be adopted into

laser welding process parameters optimization to make a trade-off yˆ(x ) = βˆ + rT (x )R−1 y − f βˆ (8)
P. Jiang et al. / Advances in Engineering Software 99 (2016) 147–160 159

Fig. 19. Illustration of NSGA-II procedure.

where y is the column vector of length n that contains the sam- References
ple data of the responses, f is a column vector of length n that is
filled with ones when f(x) is taken as a constant, rT (x) is the corre- [1] Tsirkas S, Papanikos P, Kermanidis T. Numerical simulation of the laser welding
process in butt-joint specimens. J Mater Process Tech 2003;134:59–69.
lation vector between an predicted point x and the sample points [2] Benyounis KY, Olabi AG, Hashmi MSJ. Effect of laser welding parameters on the
{x1 , x2 , . . . , xN } described as: heat input and weld-bead profile. J Mater Process Tech 2005;164-165:978–85.
      T [3] Ruggiero A, Tricarico L, Olabi AG, Benyounis KY. Weld-bead profile and costs
rT (x ) = R x, x1 , R x, x2 , . . . , R x, xN (9) optimisation of the CO2 dissimilar laser welding process of low carbon steel
and austenitic steel AISI316. Opt Laser Tech 2011;43:82–90.
The βˆ in Eq. (4) is estimated using the following equation: [4] Olabi AG, Casalino G, Benyounis KY, Hashmi MSJ. An ANN and Taguchi algo-
rithms integrated approach to the optimization of CO2 laser welding. Adv Eng
βˆ = ( f T R−1 f )−1 f T R−1 y (10)
Softw 2006;37:643–8.
[5] Khan MMA, Romoli L, Fiaschi M, Dini G, Sarri F. Experimental design ap-
proach to the process parameter optimization for laser welding of marten-
The estimated variance of the output model can be calculated
sitic stainless steels in a constrained overlap configuration. Opt Laser Tech
by: 2011;43:158–72.

T
[6] Zhou Q, Rong Y, Shao X, Jiang P, Gao Z, Cao L. Optimization of laser brazing
y − f βˆ R−1 y − f βˆ onto galvanized steel based on ensemble of metamodels. J Intell Manuf 2016.
doi:10.1007/s10845-015-1187-5.
σˆ 2 = (11) [7] Shanmugam NS, Buvanashekaran G, Sankaranarayanasamy K, Ramesh Kumar S.
N
A transient finite element simulation of the temperature and bead profiles of
The unknown parameters θ k in Eq. (3) obtained using maxi- T-joint laser welds. Mater Des 2010;31:4528–42.
mum likelihood estimation can be formulated as [18]: [8] Siva Shanmugam N, Buvanashekaran G, Sankaranarayanasamy K. Some studies
on weld bead geometries for laser spot welding process using finite element
N ln(σˆ 2 ) + ln|R| analysis. Mater Des 2012;34:412–26.
max (θk ) = − (12) [9] GuoMing H, Jian Z, JianQang L. Dynamic simulation of the temperature field of
θk > 0 2 stainless steel laser welding. Mater Des 2007;28:240–5.
[10] Bag S, De A. Development of a three-dimensional heat-transfer model for the
where both σˆ 2 and |R| are the functions of θ k . gas tungsten arc welding process using the finite element method coupled
with a genetic algorithm-based identification of uncertain input parameters.
A.2. Background of NSGA-II Metall Mater Trans A 2008;39:2698–710.
[11] Acherjee B, Kuar AS, Mitra S, Misra D. Modeling of laser transmission contour
welding process using FEA and DoE. Opt & Laser Technol 2012;44:1281–9.
Genetic algorithm searches the whole design space for global [12] Acherjee B, Kuar AS, Mitra S, Misra D. Modeling and analysis of simultaneous
optimum through selected design points using random genetic op- laser transmission welding of polycarbonates using an FEM and RSM combined
approach. Opt Laser Technol 2012;44:995–1006.
erations. It can be extended to multi-objective optimization prob- [13] Wang X, Zhang C, Li P, Wang K, Hu Y, Zhang P, et al. Modeling and optimiza-
lems and the corresponding method is called multi-objective ge- tion of joint quality for laser transmission joint of thermoplastic using an arti-
netic algorithm (MOGA). The conventional MOGA is based on Non- ficial neural network and a genetic algorithm. Opt Laser Eng 2012;50:1522–32.
[14] Shojaeefard MH, Akbari M, Asadi P. Multi objective optimization of friction stir
Dominated Sorting in Genetic Algorithms (NSGA), which is a very
welding parameters using FEM and neural network. Int J Precis Eng Manuf
effective algorithm but has been generally criticized for its high 2014;15:2351–6.
computational complexity, lack of elitism and the need for speci- [15] Islam M, Buijk A, Rais-Rohani M, Motoyama K. Process parameter optimization
of lap joint fillet weld based on FEM–RSM–GA integration technique. Adv Eng
fying the sharing parameter. A modified version, NSGA-II proposed
Softw 2015;79:127–36.
by Deb et al., which has a better sorting algorithm, incorporates [16] Vafeiadis T, Diamantaras KI, Sarigiannidis G, Chatzisavvas KC. A comparison of
elitism and no sharing parameter needs to be chosen a priori, machine learning techniques for customer churn prediction. Simul Model Pr
have been successfully applied in engineering optimization in re- Theory 2015;55:1–9.
[17] Acar E. Effect of error metrics on optimum weight factor selection for ensem-
cent years [14,31,32]. ble of metamodels. Expert Syst Appl 2015;42:2703–9.
Fig. 19 illustrates the NSGA-II procedure. NSGA-II starts with an [18] Jiang P, Wang J, Zhou Q, Zhang X. An enhanced analytical target cascading and
initialized populationPt . Then the population generates offsprings kriging model combined approach for multidisciplinary design optimization.
Math Probl Eng 2015;2015:1–11.
Ot from crossover and mutation operators via selection, crossover [19] Zhou Q, Shao X, Jiang P, Cao L, Zhou H, Shu L. Differing mapping using ensem-
and mutation operators. Once the offsprings are obtained, the pop- ble of metamodels for global variable-fidelity metamodeling. CMES: Comput
ulation with the current population and current offsprings are Model Eng Sci 2015;106:323–55.
[20] Deb K, Pratap A, Agarwal S, Meyarivan T. A fast and elitist multiobjective ge-
combined into one set and sorted according to non-domination netic algorithm: NSGA-II. Evol Comput, IEEE T on 2002;6:182–97.
and crowding distance. Finally, the new population Pt+1 can be ob- [21] Wang R, Lei Y, Shi Y. Numerical simulation of transient temperature field
tained with the best N individuals in the combined set. More de- during laser keyhole welding of 304 stainless steel sheet. Opt Laser Technol
2011;43:870–3.
tails of NSGA-II can refer to Deb et al. [20].
160 P. Jiang et al. / Advances in Engineering Software 99 (2016) 147–160

[22] Esfahani MRN, Coupland J, Marimuthu S. Numerical simulation of alloy com- [28] Anawa EM, Olabi AG. Using Taguchi method to optimize welding pool of dis-
position in dissimilar laser welding. J Mater Process Technol 2015;224:135–42. similar laser-welded components. Opt Laser Technol 2008;40:379–88.
[23] Goldak J, Bibby M, Moore J, House R, Patel B. Computer modeling of heat flow [29] Schilt S, Tombez L, Tardy C, Bismuto A, Blaser S, Maulini R, et al. An experi-
in welds. Metall Trans B 1986;17:587–600. mental study of noise in mid-infrared quantum cascade lasers of different de-
[24] Bag S, Kiran MD, Syed MA, De A. Efficient estimation of volumetric heat source signs. Appl Phys B 2015;119:189–201.
in fusion welding process simulation. Weld World 2012;56:88–97. [30] Sacks J, Welch WJ, Mitchell TJ, Wynn HP. Design and analysis of computer
[25] Goldak J, Chakravarti A, Bibby M. A new finite element model for welding heat experiments. Stat Sci 1989:409–23.
sources. Metall Trans B 1984;15:299–305. [31] Poirier JD, Vel SS, Caccese V. Multi-objective optimization of laser-welded
[26] Wang X, Chen H, Liu H, Li P, Yan Z, Huang C, et al. Simulation and optimization steel sandwich panels for static loads using a genetic algorithm. Eng Struct
of continuous laser transmission welding between PET and titanium through 2013;49:508–24.
FEM, RSM, GA and experiments. Opt Lasers Eng 2013;51:1245–54. [32] Chaki S, Bathe RN, Ghosal S, Padmanabham G. Multi-objective optimisation of
[27] Islam M, Buijk A, Rais-Rohani M, Motoyama K. Simulation-based numerical op- pulsed Nd:YAG laser cutting process using integrated ANN–NSGAII model. J
timization of arc welding process for reduced distortion in welded structures. Intell Manuf 2015. doi:10.1007/s10845- 015- 1100- 2.
Finite Elem Anal Des 2014;84:54–64.

You might also like