You are on page 1of 15

Ships and Offshore Structures

ISSN: 1744-5302 (Print) 1754-212X (Online) Journal homepage: https://www.tandfonline.com/loi/tsos20

Parameter determination of double-ellipsoidal


heat source model and its application in the multi-
pass welding process

Guangming Fu, Jijun Gu, Marcelo Igor Lourenco, Menglan Duan & Segen F.
Estefen

To cite this article: Guangming Fu, Jijun Gu, Marcelo Igor Lourenco, Menglan Duan & Segen
F. Estefen (2015) Parameter determination of double-ellipsoidal heat source model and its
application in the multi-pass welding process, Ships and Offshore Structures, 10:2, 204-217, DOI:
10.1080/17445302.2014.937059

To link to this article: https://doi.org/10.1080/17445302.2014.937059

Published online: 10 Jul 2014.

Submit your article to this journal

Article views: 646

View Crossmark data

Citing articles: 11 View citing articles

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=tsos20
Ships and Offshore Structures, 2015
Vol. 10, No. 2, 204–217, http://dx.doi.org/10.1080/17445302.2014.937059

Parameter determination of double-ellipsoidal heat source model and its application


in the multi-pass welding process
Guangming Fua,∗ , Jijun Gub , Marcelo Igor Lourencoa , Menglan Duanb and Segen F. Estefena
a
Ocean Engineering Department, COPPE, Federal University of Rio de Janeiro, Rio de Janeiro, Brazil; b Offshore Oil/Gas
Research Center, China University of Petroleum (Beijing), Beijing, China
(Received 27 November 2013; accepted 16 June 2014)

The parameters of the heat source model have significant influence on the temperature field and sequentially affect the residual
stress field. In this paper, a neural-network programme based on the Levenberg–Marquardt algorithm is developed to predict
the parameters of Goldak’s double-ellipsoidal heat source model. The analytical solution of the heat conduction equation
based on the double-ellipsoidal heat source is obtained by integrating a series of instant point heat sources over the volume of
the ellipsoidal heat source. The transient temperature distribution and the sizes of the molten pool are obtained under various
welding processes by using the analytical method. Then, a neural-network programme is employed to train and predict the
heat source parameters. These results of temperature and molten pool size obtained by the numerical simulation with the
predicted heat source parameters are calibrated by the published experimental results. The numerical results show a good
agreement with the experimental measurements. Finally, the developed Levenberg–Marquardt neural network is employed to
predict the heat source parameters in the multi-pass welding process in the laboratory. By comparing the finite element (FE)
numerical results with experimental results, the heat source parameters have been successfully identified in the multi-pass
welding process.
Keywords: heat source model; neural network; temperature distribution; experimental calibration

1. Introduction location are confined, the non-destructive measure method,


Welding technology is widely used in the construction of such as neutron and X-ray diffractions, cannot be applied
onshore and offshore structures due to the high productiv- conveniently. Therefore, the numerical simulation method
ity during the assembly process. However, during the weld- calibrated by experimental measurements becomes a fea-
ing process, non-uniform heat expansion in the weld bead sible solution (Smith and Smith 2009a; Muránsky et al.
and the adjacent parent material can induce severe residual 2012). When performing an accurate numerical simula-
stresses and distortions. The residual stresses give a larger tion, the parameters of heat source model which have sig-
contribution to the final stress field (Brickstad and Josefson nificant influence on temperature distribution and molten
1998). Severe welding distortion, residual stress and reliev- pool shape should be initially determined (Smith and Smith
ing residual stress process can influence the performance 2009b; Joshi et al. 2013). Joshi et al. (2013) evaluated the
of welded structural members, such as tensile strength and geometric parameters of Goldak’s double-ellipsoidal heat
impact energy especially after the repair welding (Aloraier source model by using SYSWELD (2006) in the gas metal
et al. 2013). Moreover, the phase transformation occurred arc welding (GMAW) process. The generated new mate-
in the weld metal and the adjacent region of the weld line rial temperature-dependent properties in the fusion region
will contribute to the residual stresses (Heinze et al. 2012), were employed for the evaluation of geometric parameters
particularly in the multi-pass welding process due to prior of heat source. The results show good matching with the
weld metal subject to several thermal cycles and some lo- experimental measurements.
cations undergo two or more phase transformations (Deng In this paper, a neural-network programme based
and Murakawa 2006). To minimise the negative influences on the Levenberg–Marquardt algorithm (Levenberg 1944;
of welding on the structure fabrication quality, investiga- Marquardt 1963) is developed to predict the geometric pa-
tion of welding-induced residual stresses and distortions rameters of Goldak’s double-ellipsoidal heat source model.
has become a major concern in the recent years. The solution of the heat conduction equation based on the
Furthermore, since the dimensions of the components double-ellipsoidal heat source model is obtained by su-
of actual structures are usually very large or the spaces of the perposing a series of instant point heat sources over the


Corresponding author. Email: fu@lts.coppe.ufrj.br


C 2014 Taylor & Francis
Ships and Offshore Structures 205

volume of the heat source model. The transient tempera- and rear quadrants of the moving heat source model are
ture profiles and the sizes of the molten pool are obtained described in Equations (2a) and (2b), respectively,
under various welding processes by using this analytical
method. Then, the results are employed as the training sam-  

ples of the Levenberg–Marquardt neural network. These 6 3ff ηQ 3x 2 3y 2 3z2
numerical results of temperature and molten pool size based qf (x, y, z) = √ exp − 2 − 2 − 2
af bcπ π af b c
on the predicted heat source model are calibrated by pub-
(2a)
lished experimental measurements. Finally, the developed
Levenberg–Marquardt neural network is employed to pre- √  
6 3fr ηQ 3x 2 3y 2 3z2
dict the heat source parameters in the multi-pass welding qr (x, y, z) = √ exp − 2 − 2 − 2
ar bcπ π ar b c
process in the laboratory. The finite element (FE) numerical
(2b)
results show a good agreement with the test measurements,
and the heat source parameters have been successfully iden-
tified in the multi-pass welding process. where af and ar are the front and the rear semi-axes of the
heat source, respectively. The fractions of deposited heat, ff
and fr , represent the heat apportionments of the heat flux in
2. Heat source model
the front and rear quadrants, respectively, whereff + fr =
Since Rosenthal (1941) first developed the travelling point 2. Considering the continuity of the volumetric heat source,
heat source model to investigate the temperature distribu- 2af
ff = af +a and fr = af2a+ar r are assumed (Lundback 2003).
r
tion which formed the basis for the most subsequent studies
on weld heat flow, several heat source models were devel-
oped to obtain an accurate heat transfer analysis in the
welding process by later researchers (Pavelic et al. 1969; 3. Heat transfer analysis and
Goldak et al. 1984). Levenberg–Marquardt neural network
Goldak et al. (1984) developed two different heat source 3.1. Heat transfer analysis
models, semi- and double-ellipsoidal heat source models
In the thermal analysis, for an isotropic material, the tran-
which are widely adopted due to the accuracy in several of
sient temperature field T of the welded plate is a function
welding processes. The formulation of the semi-ellipsoidal
of time t and the spatial coordinates (x, y, z):
heat source model as shown in Figure 1(a) can be expressed
as
 
√   ∂ 2T ∂ 2T ∂ 2T ∂T
6 3f ηQ 3x 2 3y 2 3z2 k + + + q = ρc (3)
q(x, y, z) = √ exp − 2 − 2 − 2 (1) ∂x 2 ∂y 2 ∂z2 ∂t
abcπ π a b c

where x, y and z are the local coordinates of the ellipsoidal where q, ρ, c and k are the heat flux, the material density,
model, η is the arc efficiency, Q is the power input, a, b and the specific heat and the thermal conductivity, respectively.
c are the semi-axes of the heat source model. T = T (x, y, z, t) represents the temperature at the point
However, some results with the semi-ellipsoidal heat (x, y, z) at time t.
source show that the temperature gradients are not as steep Rosenthal (1941) first developed the travelling in-
as the experimentally observed in front of the heat source; stant point heat source model to obtain the temperature
however, in rear of heat source, they are steeper than ex- distribution:
perimental measurements. To correct this, Goldak et al.
(1984) proposed a new heat source model, “double el-  R 2

lipsoidal heat source model” which combined two semi- exp − 4πκ(t−t )
dTt  = δQdt  (4)
ellipsoidal heat source models as shown in Figure 1(b). ρc[4π κ(t − t  )]3/2
The double-ellipsoidal heat source can provide more accu-
rate results by comparing the experimental measurements,
especially in the low penetration welding process. Later, where κ = ρc k
is the thermal diffusivity, δQ is the heat
numerous welding-related research works were carried out source, (x , y  , z ) is the location of the heat source δQ at


to investigate the welding residual stress in various welding a time of t  and R 2 = (x − x  )2 + (y − y  )2 + (z − z )2 is


methods, such as gas metal arc welding (GMAW) (Malik the distance to the heat source.
et al. 2008; Joshi et al. 2010), laser welding (Kong et al. Nguyen et al. (1999) applied the semi- and double-
2011; Bannour et al. 2012), metal inert gas welding (MIG) ellipsoidal heat source in a semi-infinite body based on
(Price et al. 2008) and flux cored arc welding (Aloraier and the instant point solution through integration over the vol-
Joshi 2012). The power density distributions of the front ume of heat source. The solution of temperature field by
206 G. Fu et al.

Figure 1. Goldak’s heat source models: (a) Semi-ellipsoidal model (Muránsky et al. 2012) and (b) double-ellipsoidal heat source model.
(This figure is available in colour online.)

employing the double-ellipsoidal heat source is described Fachinotti et al. (2011) proposed a more accurate
by the following equation: solution:

√  ∞ ∞ ∞ T (x, y, z, t) = T0
1 6 3Qdt 
dTt  = √
4 ρcaf ar π π [4π κ(t − t  )]3/2 −∞ −∞ −∞ √  t exp − 3y 2 3z2
− 12κ(t−t

3 3Q 12κ(t−t  )+b2  )+c2
  + √
(x − x  )2 + (y − y  )2 + (z − z )2 ρcπ π 0 12κ(t − t  ) + b2 12κ(t − t  ) + c2
× exp −
4π κ(t − t  )
   × [ff Af (1 − Bf ) + fr Ar (1 − Br )]dt  (6)
ff 3x 2 3y 2 3z2
× exp − 2 − 2 − 2  
af af b c 3(x−vt  )2
exp −
12κ(t−t  )+ai2  ai  
  where Ai = √ , Bi =erf 2
√ √
x−vt
12κ(t−t  )+ai2 κ(t−t  ) 12κ(t−t  )+ai2
fr 3x 2 3y 2 3z2
+ exp − 2 − 2 − 2 dx  dy  dz (5) and i can be replaced by f and r for the front and rear
ar ar b c regions, respectively.
Ships and Offshore Structures 207

In the present study, the MATLAB programme is em- 3.2. The Levenberg–Marquardt neural network
ployed to capture the maximum size of the molten pool The Levenberg–Marquardt algorithm (Levenberg 1944;
(half-width and depth) based on the analytical solution of Marquardt 1963) provided a numerical solution to the prob-
the semi-infinity body subjected to the three-dimensional lem of minimising a non-linear function. In the artificial
(3D) moving heat sources. The size of the molten pool and neural-network field, this algorithm is suitable for train-
the corresponding geometric parameters of the heat source ing small- and medium-sized problems. Other algorithms
model are recorded as the input data of the neural network. have already been developed for the neural-network learn-
The following assumptions are considered in the current ing, such as the error back propagation algorithm, which
study: is regarded as the largest breakthrough for neural-network
learning. However, it is limited by its slow convergence.
The Levenberg–Marquardt algorithm interpolates the gra-
• For minimising the computer time consumption, the dient descent algorithm and the Gauss–Newton algorithm.
size of the rear part of the ellipsoidal heat source is It is more robust than the Gauss–Newton algorithm, since
four times the front part (ar = 4af ). The size of the in many cases it can converge well, even if it starts far
semi-axes in a transverse direction equals the front off the final value. It inherits the stability advantage of the
part (b = af ). Therefore, only two parameters are gradient descent algorithm and the speed advantage of the
calculated. Gauss–Newton algorithm.
• In order to reduce the time consumption, the physical Suppose that we have a function V (x) which we want
properties (c = 600 J/kg/◦ C; k = 29 J/m/s/◦ C and to minimise with respect to the parameter vector x, then
ρ = 7820 kg/m3 ) reported by Nguyen et al. (1999) Newton’s method would be
are employed in the analytical calculations of the
fusion zone (FZ) size. In the numerical simulation,
the temperature independent thermo-properties are x = −[∇ 2 V (x)]−1 ∇V (x) (7)
employed.
where ∇ 2 V (x) is the Hessian matrix and ∇V (x) is the gra-
dient. If we assume that V (x) is a sum of squares function,
A subroutine is implemented in MATLAB to cal-
culate the maximum size of the molten pool based on
Equation (6). The subroutine is used to determine the ana-

lytical solution for the Goldak’s model geometric parame- V (x) = ei2 (x) (8)
ters for 138 combinations of welding parameters, as shown i=1

in Table 1, in which the welding speed range is from 1 to


7 mm/s, the voltage range is from 10 to 50 Volts and the Then, it can be shown that
current range is from 100 to 500 Amperes. These results are
finally used to train a neural-network algorithm, and later
the neural network is employed to determine the geometric ∇V (x) = J T (x)(x) (9)
parameters of heat source model for any welding procedure
in the considered range. ∇ 2 V (x) = J T (x)J (x) + S(x) (10)

Table 1. The training samples of the Levenberg–Marquardt neural network.

Depth and half-


width – analytical

Speed (mm/s) Voltage (V) Current (A) af (mm) c (mm) (mm)

1 10 500 4 4 8 8
1 30 300 6 6 10 10
1 50 300 6 4 12.5 12
4 50 100 4 2 6 5.5
4 30 300 2 6 8 9
4 30 500 2 2 11 11
7 50 300 6 2 9 8
7 50 500 2 2 11 11
7 50 100 2 6 4.5 5.5
208 G. Fu et al.

Table 2. The calibrations of the Levenberg–Marquardt neural network.

Depth and half-width

Velocity (mm/s) Voltage (V) Current (A) (mm) af (mm) c (mm) af – L.M (mm) c – L.M (mm)

1 10 300 5.5 5.5 4 4 4 4


1 30 300 10.5 10.5 4 4 4 4
7 50 500 11 11 6 6 6 8
1 30 100 5.5 5.5 4 4 4 4
1 30 500 13 13 4 6 4 4
1 50 300 13 13 4 6 4 4
1 50 500 15 14.5 6 4 6 4
1 50 500 15.5 15.5 4 4 4 4
4 10 500 6 6 4 4 4 4
4 50 500 13.5 13 6 4 6 6

where J (x) is the Jacobian matrix The Levenberg–Marquardt algorithm can be considered a
⎡ ⎤ trust-region modification to Gauss–Newton.
∂e1 (x) ∂e1 (x) ∂e1 (x) The Levenberg–Marquardt neural network is developed
⎢ ∂x1 ...
⎢ ∂x2 ∂xn ⎥ ⎥ based on the Labview in the current study. The analytical re-
⎢ ∂e2 (x) ∂e2 (x) ∂e2 (x) ⎥ sults (128 groups) are adopted as the training samples of the
⎢ ⎥
⎢ ...
∂xn ⎥
J (x) = ⎢ ∂x1 ∂x2 ⎥ (11) Levenberg–Marquardt neural network. The last 10 groups
⎢ .. .. .. .. ⎥ are employed to calibrate the Levenberg–Marquardt neural
⎢ . . . . ⎥
⎢ ⎥ network. The predicted results indicate a good agreement
⎣ ∂eN (x) ∂eN (x) ∂eN (x) ⎦
... with the analytical results as shown in Table 2.
∂x1 ∂x2 ∂xn

and
4. Finite element model validation

N
S(x) = ei (x)∇ 2 ei (x) (12) The finite element model of transient temperature has
i=1
been developed based on ABAQUS (2010) and FORTRAN
subroutine. In numerical model, the Goldak’s double-
For the Gauss–Newton method, it is assumed that S(x) ≈ 0, ellipsoidal heat model is adopted to simulate the power
and Equation (7) becomes input, and the subroutine is generated to simulate the weld-
ing process. The same available data reported in the pub-
x = [J T (x)J (x)]−1 J T (x)e(x) (13) lished literatures (Goldak et al. 1984; Nguyen et al. 1999)
are employed in the FE model, as shown in Table 3, and the
The Levenberg–Marquardt modification to the Gauss– temperature profiles are calculated by the numerical method
Newton method is which combines the subroutine and the predicted geomet-
ric parameters of Goldak’s heat source model. The finite
x = [J T (x)J (x) + μI ]−1 J T (x)e(x) (14) element model is replicated with the size of 200 mm ×
100 mm × 100 mm in the present study, as shown in
The parameter μ is multiplied by some factor (β) when- Figure 2. The parameters of the double-ellipsoidal heat
ever a step would result in an increased V (x). When a step source model are predicted by the Levenberg–Marquardt
reduces V (x), μ is divided by β. Notice that when μ is large, neural-network programme as shown in Table 4.
the algorithm becomes steepest descent (with step 1/μ), In Goldak et al.’s (1984) work, the experimental welding
while for small μ the algorithm becomes Gauss–Newton. process parameters are: voltage 32.9 V, welding current

Table 3. Input data for predicting the parameters of heat source.

Voltage (V) Current (A) Velocity (mm/s) Half-width (mm) Depth (mm)

Goldak et al. (1984) 30.9 1170 5 14 14


Nguyen et al. (1999) 26 230 5 6.4 3.65
Ships and Offshore Structures 209

Figure 2. Mesh of the finite element model. (This figure is avail-


able in colour online.)

1170 A, welding velocity 5 mm/s and welding efficiency


95%. Considering af = 20 mm and c = 22 mm, according
the first assumption, the rear quadrant of heat source is
four times the front part, ar = 80 mm. In the transverse Figure 3. Results of FZ and HAZ sizes with parameters predicted
direction, b = af = 20 mm. by Levenberg–Marquardt neural network. (This figure is available
Based on the numerical simulation with the predicted in colour online.)
geometric parameters of heat source model, the sizes of
the FZ and heat-affected zone (HAZ) (x1 = 12 mm and source model. It can be concluded that the temperatures
x2 = 20.4 mm), are captured as shown in Figure 3. The are overpredicted in the high-temperature regions in the
results are in good agreement with published experimental current FEM-1. However, in the regions below 1400 ◦ C
measurements (x1 = 14 mm and x2 = 21 mm) by Goldak which provide the largest contributions to the residual stress
et al. (1984). The errors between the numerical results and fields, the current FEM-2 results are in better agreement
experimental measurements are 14% and 3%, respectively. with the experimental measurements than the results of
Goldak et al. (1984) also suggested the parameters current FEM-1. The current FEM-2 results also show a
of heat source (af = 15 mm,ar = 30 mm, b = 20 mm and good agreement with the Krutz’s numerical results.
c = 20 mm) to simulate this welding process and the re- The welding process performed by Nguyen et al. (1999)
sults are presented in Figure 4. The sizes of FZ and HAZ is also replicated to calibrate the Levenberg–Marquardt
are captured as x1 = 15 mm and x2 = 21 mm, respectively. neural network in the present study. The welding voltage is
By comparing the experimental results (x1 = 14 mm and 26 V, the welding current is 230 A and the welding velocity
x2 = 21 mm), the error is 7.1% of the size of FZ. The size is 5 mm/s. The parameters of the heat source model are
of the HAZ was accurately predicted based on Goldak’s predicted by the Levenberg–Marquardt neural network as
proposal. af = 4 mm, ar = 16 mm, c = 12 mm and b = 4 mm. The
Temperature distributions along the transverse direction sizes of the FZ in the transverse direction (x1) and ver-
in the middle of the specimen are plotted in Figure 5. The tical direction (x2) are captured numerically as shown in
current FEM-1 represents the numerical simulation based Figure 6. By comparing with the experimental measure-
on the heat source parameters suggested by Goldak et al. ments performed by Nguyen et al. (1999), there are 6.25%
(1984). The current FEM-2 indicates the numerical results and 17.2% differences in transverse and vertical directions,
based on the current predicted geometric parameters of heat respectively.
By comparing the experimental results presented by
Table 4. Predicted parameters of heat source model. Goldak et al. (1984) and by Nguyen et al. (1999), the devel-
oped Levenberg–Marquardt neural network can reasonably
af ar b c predict the parameters of heat source model with minimum
(mm) (mm) (mm) (mm)
CPU time consumption. The current numerical method-
For Goldak’s model 20 80 20 22 ology combined with the subroutines can provide accept-
For Nguyen’s model 4 16 4 12 able temperature distribution by comparing the experimen-
tal measurements.
210 G. Fu et al.

Figure 6. Comparison of FZ and HAZ sizes between FEM and


experimental results proposed by Nguyen et al. (1999). (This
figure is available in colour online.)
Figure 4. Results of FZ and HAZ sizes with the parameters
provided by Goldak et al. (1984). (This figure is available in colour
online.)
sions of 1200 mm × 500 mm × 19 mm with the groove an-
gle of 20º and the root opening of 6 mm. The plates are
5. Applications in the multi-pass welding process jointed with the single-electrode welding process in five
filling layers. The MIG manual procedure is employed in
5.1. Welding procedure the first two welding passes and the submerged arc weld-
The experiments are performed using the ASTM
A131grade DH36 ferric steel plate typically employed in
the shipbuilding industry. The chemical composition of DH Table 5. Chemical compositions of DH36 steel plate (mass %).
36 is shown in Table 5 and the thermo-physical proper-
C Si Mn P S Ti Nb
ties are reported by Camilleri et al. (2013), as shown in
Figure 7. The laboratory test joints two plates with dimen- 0.160 0.340 1.450 0.025 0.020 0.022 0.025

Figure 5. Temperature distributions on the top of the work piece


in transversal direction. (This figure is available in colour online.) Figure 7. Thermo-mechanical properties.
Ships and Offshore Structures 211

Table 6. Parameters of welding procedures.

Wire diameter Wire speed Welding speed Inter-pass temperature


Pass Process (mm) Current (A) Voltage (V) (m/min) (m/min) (◦ C)

1 MIG 1.2 160 28.9 4.8 0.16 –


2 MIG 1.2 140 28.1 4.8 0.11 70
3 SAW 4.0 530 31.0 2.0 0.26 32
4 SAW 4.0 563 31.0 1.9 0.26 109
5 SAW 4.0 560 30.0 2.3 0.27 170

ing (SAW) process is performed automatically in the last −20, 0, 20 and 40 mm from the centreline of the plates as
three welding passes. The detailed welding procedures are shown in Figure 8. The temperatures are captured by the
indicated in Table 6. data logging system as illustrated in Figure 9.
The K-type thermocouples are directly mounted in the
middle of the plate (9.5 mm from the upper surface) in the
vertical direction prior to the welding process and connected 5.2. Numerical procedures
directly to the multi-channel data logging system. In this The numerical investigation includes two-dimensional (2D)
way, the temperature profile during the welding process and 3D FE models of the present welded plate, as shown in
at any specified time can be conveniently stored for later Figure 10. The 8-node quadratic heat transfer quadrilateral
data processing and comparison with numerical results. The (DC2D8) and 8-node linear heat transfer brick (DC3D8) are
thermocouples TC-1, TC-2, TC-3, TC-4 and TC-5 are at employed in 2D and 3D numerical models, respectively. The
20, 20, 24, 30 and 30 mm from the welding line and −40, double-ellipsoidal heat source model is employed and the

Figure 8. Experimental set-up and thermocouple arrangement. (This figure is available in colour online.)
212 G. Fu et al.

Figure 9. Experimental set-up: (a) thermocouple position and (b) data logging system. (This figure is available in colour online.)

geometric parameters of the heat source model are predicted model. Heat loss due to convection is assumed as shown in
by the Levenberg–Marquardt neural-network programme, Equation (15). The heat loss on the surface due to radiation
as shown in Table 7. to the surrounding region is governed by Equation (16),
The thermal boundary conditions, including the con- following the Stefan–Boltzmann law:
vection and radiation to the environment from all ex-
posed surfaces, are considered in the current numerical
qc = −h(T − T0 ) (15)
Ships and Offshore Structures 213

Figure 10. Finite element model: (a) 2D generalised plain strain and (b) 3D finite element model. (This figure is available in colour
online.)

  in Figure 7. An artificially increased thermal conductivity


qr = eσ (T + 273.15)4 − (T0 + 273.15)4 (16) is adopted for temperatures above the melting point. The
thermal effects due to solidification of the molten pool are
considered by taking into account the latent heat effects;
where qc and qr are the heat fluxes across the surface, the latent heat is assumed as 270 kJ/kg in the present work.
h is a reference film coefficient, T is the temperature at The solid and liquid temperatures are assumed as 1440 and
this point on the surface, T0 is the ambient temperature, 1505 ◦ C (Joshi et al. 2010), respectively.
e is the emissivity of the surface and σ is the Stefan–
Boltzmann constant as 5.67 × 10−8 Wm−2 K−4 , and the
5.3. Results and discussions
ambient temperature is assumed as 25 ◦ C. The film coeffi-
cient and emissivity are defined as 20 Wm−2 K−1 and 0.85, The temperature distributions during the welding process
respectively. at each thermocouple TC-1, TC-2, TC-3, TC-4 and TC-5
The thermo-mechanical properties of DH 36 are sim- are illustrated in Figure 11. The comparisons of the tem-
ilar to the EN 10025 S355J2G3 (Camilleri et al. 2013; perature field between the experimental measurements and
Wang et al. 2013). The temperature-dependent material 3D numerical simulations at several thermocouples in the
properties from the SYSWELD (2006) material database multi-pass welding process are also plotted in Figure 11.
are employed in the present numerical simulation as shown The temperature distributions show a good agreement

Table 7. Predicted parameters of Goldak’s heat source model.

Front part af Rear part ar Half-width b Depth c


Passes (mm) (mm) (mm) (mm)

Pass-1 4 16 4 12
Pass-2 4 16 4 12
Pass-3 10 40 10 16
Pass-4 10 40 10 16
Pass-5 10 40 10 16
214 G. Fu et al.

(a)
(b)

(c) (d)

(e)

Figure 11. Temperature distributions at various thermocouples in different welding passes: 3D (solid line) and experimental measurements
(dotted line): (a) first welding pass, (b) second welding pass, (c) third welding pass, (d) fourth welding pass and (e) fifth welding pass.
(This figure is available in colour online.)

between the 3D numerical simulations and the experimental merical simulations at thermocouple TC-3 in five differ-
measurements. ent welding passes. These results show a good agree-
Figure 12 shows the temperature comparisons be- ment between numerical simulations and experimental
tween the experimental measurements, 2D and 3D nu- measurements.
Ships and Offshore Structures 215

Figure 12. Comparisons of temperature distribution at TC-3 in various welding passes for 2D numerical model, 3D numerical model
and experimental measurements: (a) first welding pass, (b) second welding pass, (c) third welding pass, (d) fourth welding pass and
(e) fifth welding pass. (This figure is available in colour online.)
216 G. Fu et al.

6. Conclusions References
The sizes of the FZ and HAZ are calculated by the analyti- ABAQUS. 2010. Example problems manual version 6.10-1.
cal method for several welding processes and the assumed Rhode Island (RI): Hibbitt, Karlsson and Sorensen, Inc.
Aloraier AS, Joshi S. 2012. Residual stresses in flux cored arc
parameters of Goldak’s heat source model. These captured welding process in bead-on-plate specimens. Mater Sci Eng
results are employed as the training and testing samples for A. 534:13–21.
the Levenberg–Marquardt neural network. The numerical Aloraier AS, Joshi S, Price JW, Alawadhi K. 2013. Mate-
method combined with subroutine and predicted param- rial properties characterization of low carbon steel using
eters of heat source are calibrated. Finally, this method is TBW and PWHT techniques in smooth-contoured and u-
shaped geometries. Int J Press Vessel Piping. 111–112:269–
employed to predict the parameters of the heat source model 278.
in the experimental test. The following conclusions can be Bannour S, Abderrazak K, Mhiri H, Palec GL. 2012. Effects of
drawn. temperature-dependent material properties and shielding gas
on molten pool formation during continuous laser welding
of AZ91 magnesium alloy. Opt Laser Technol. 44(8):2459–
(1) The current FE model combined with subroutines 2468.
can be employed to predict the temperature distri- Brickstad B, Josefson B. 1998. A parametric study of residual
bution and also can provide reasonable geometric stresses in multi-pass butt-welded stainless steel pipes. Int J
shapes of molten pool both in transversal and ver- Press Vessel Piping. 75(1):11–25.
Camilleri D, McPherson N, Gray TG. 2013. The applicability of
tical directions.
using low transformation temperature welding wire to mini-
(2) The numerical simulation of the welding process mize unwanted residual stresses and distortions. Int J Press
coupled with the predicted geometric parameters Vessel Piping. 110:2–8.
of heat source is performed. The temperature distri- Deng D, Murakawa H. 2006. Prediction of welding residual stress
bution obtained by the numerical simulation shows in multi-pass butt-welded modified 9Cr-1Mo steel pipe con-
sidering phase transformation effects. Comput Mater Sci.
a good agreement with the experimental measure-
37:209–219.
ments performed by Goldak et al. (1984). More- Fachinotti VD, Anca AA, Cardona A. 2011. A method for the
over, the present numerical results of the sizes solution of certain problems in least squares. Int J Numer
of FZ and HAZ are more accurate than those Method Biomed Eng. 27(4):595–607.
obtained based on the parameters suggested by Goldak J, Chakravarti A, Bibby M. 1984. A new finite ele-
ment model for welding heat source. Metall Mater Trans B.
Goldak et al. (1984). By comparing experimental
15B:299–305.
results in Goldak et al.’s (1984) and Nguyen et al.’s Heinze C, Schwenk C, Rethmeier M. 2012. Numerical calculation
(1999) works, the developed Levenberg–Marquardt of residual stress development of multi-pass gas metal arc
neural network can reasonably predict the parame- welding. J Constr Steel Res. 72:12–19.
ters of heat source model with minimum CPU time Joshi S, Hildebrand J, Aloraier AS, Rabczuk T. 2013. Character-
ization of material properties and heat source parameters in
consumption. The proposed numerical methodol-
welding simulation of two overlapping beads on a substrate
ogy combined with the subroutines can obtain an plate. Comput Mater Sci. 69:559–565.
acceptable temperature distribution by comparing Joshi S, Semetay C, Price JW, Nied HF. 2010. Weld-induced
the experimental measurements. residual stresses in a prototype dragline cluster and compari-
(3) The comparisons of the temperature fields between son with design codes. Thin-Walled Struct. 48(2):89–102.
Kong F, Ma J, Kovacevic R. 2011. Numerical and experimental
the experimental measurements and FE numerical
study of thermally induced residual stress in the hybrid laser-
simulations show that the proposed neural-network GMA welding process. J Mater Process Technol. 211:1102–
programme can reasonably predict the parameters 1111.
of heat source model in the multi-pass welding Levenberg K. 1944. A method for the solution of certain problems
process. in least squares. Q Appl Math. 5:164–168.
Lundback A. 2003. Finite element modeling and simulation of
welding of aerospace components [master’s thesis]. Sweden:
Acknowledgements Lulea University of Technology.
Malik AM, Qureshi EM, Dar NU, Khan I. 2008. Analysis of cir-
Special thanks to the technical team from the Subsea Technology cumferentially arc welded thin-walled cylinders to investigate
Laboratory – COPPE/UFRJ. the residual stress fields. Thin-Walled Struct. 46(12):1391–
1401.
Marquardt D. 1963. An algorithm for least-squares estimation of
Funding nonlinear parameters. SIAM J Appl Math. 11(2):431–441.
The authors acknowledge gratefully ANP, CNPq and FINEP Muránsky O, Smith MC, Bendeich PJ, Holden TM, Luzin V, Mar-
[PROFAB - grant number 01.10.810.00] of Brazil, the National tins RV, Edwards L. 2012. Comprehensive numerical analysis
Basic Research Program of China (973 Program) [grant num- of a three-pass bead-in-slot weld and its critical validation us-
ber 2011CB013702]; the National Natural Science Foundation of ing neutron and synchrotron diffraction residual stress mea-
China [grant number 11302264], [grant number 51109185] for surements. Int J Solids Struct. 49:1045–1062.
financial support. Guangming Fu acknowledges the financial sup- Nguyen N, Ohta A, Matsuoka K, Suzuki N, Maeda Y. 1999.
port from the China Scholarship Council. Analytical solutions for transient temperature of semi-infinite
Ships and Offshore Structures 217

body subjected to 3D moving heat sources. Weld Res. 3:265s– Smith MC, Smith AC. 2009a. NeT bead-on-plate round robin:
274s. comparison of residual stress predictions and measurements.
Pavelic V, Tanbakuchi R, Auyehara O. 1969. Experimental Int J Press Vessel Piping. 86:79–95.
and computed temperature histories in gas tungsten arc Smith MC, Smith AC. 2009b. NeT bead-on-plate round robin:
welding of thin plates. Weld J Res Suppl. 48(7):295s– comparison of transient thermal predictions and measure-
305s. ments. Int J Press Vessel Piping. 86:96–109.
Price JW, Ziara-Paradowska A, Joshi S, Finlayson T, Semetay C, SYSWELD. 2006. Engineering guide of training and toolbox.
Nied H. 2008. Comparison of experimental and theoretical France: ESI Group.
residual stresses in welds: the issue of gauge volume. Int J Wang Y, Wang L, Di X, Shi Y, Bao X, Gao X. 2013. Simulation
Mech Sci. 50(3):513–521. and analysis of temperature field for in-service multi-pass
Rosenthal D. 1941. Mathematical theory of heat distribution dur- welding of a sleeve fillet weld. Comput Mater Sci. 68:198–
ing welding and cutting. Weld J 20(5):220s–234s. 205.

You might also like