Anda di halaman 1dari 18

INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING

Int. J. Numer. Meth. Engng (in press)


Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/nme.2083

Adaptive mesh technique for thermalmetallurgical numerical


simulation of arc welding processes
M. Hamide, E. Massoni and M. Bellet,
Ecole des Mines de Paris, CEMEF, UMR CNRS 7635, BP 207, Sophia Antipolis Cedex 06904, France

SUMMARY
A major problem arising in finite element analysis of welding is the long computer times required for a
complete three-dimensional analysis. In this study, an adaptative strategy for coupled thermometallurgical
analysis of welding is proposed and applied in order to provide accurate results in a minimum computer
time. The anisotropic adaptation procedure is controlled by a directional error estimator based on local
interpolation error and recovery of the second derivatives of different fields involved in the finite element
calculation. The methodology is applied to the simulation of a gastungsten-arc fusion line processed on
a steel plate. The temperature field and the phase distributions during the welding process are analyzed
by the FEM method showing the benefits of dynamic mesh adaptation. A significant increase in accuracy
is obtained with a reduced computational effort. Copyright q 2007 John Wiley & Sons, Ltd.
Received 28 August 2006; Revised 22 March 2007; Accepted 29 March 2007
KEY WORDS:

finite elements; welding; heat transfer; phase transformation; mesh adaptation; anisotropic
metric; error estimation

1. INTRODUCTION
The accuracy of a numerical solution obtained by the finite element method depends on the spatial
discretization of the physical domain. In general, the desired element sizes in different directions
are influenced by the physical and geometrical features of the problem which can vary significantly
in time and space. In many physical problems, including welding, the solution exhibits anisotropic
features creating a demand for elements which are aligned with the solutions anisotropy. In
realistic cases, the information required to compute the desired solution field to an acceptable
level of accuracy is unknown a priori. An efficient approach to overcome this difficulty consists
in applying an adaptative procedure in which the errors arising from spatial discretization are
controlled within a specified tolerance. An anisotropic adaptative procedure modifies the mesh in
such a way that the local mesh resolution becomes adequate in all directions.
Correspondence

to: M. Bellet, Ecole des Mines de Paris, CEMEF, UMR CNRS 7635, BP 207, Sophia Antipolis
Cedex 06904, France.
E-mail: michel.bellet@ensmp.fr

Copyright q

2007 John Wiley & Sons, Ltd.

M. HAMIDE, E. MASSONI AND M. BELLET

The concentrated heat input that appears in most welding applications requires a refined discretization in the neighborhood of the molten region below the moving electrodes, where strong
axial and transverse thermal gradients prevail. In addition, some induced solid-state phase change
often generate residual gradients of phase fractions. The capture of such thermal and metallurgical
gradients requires some kind of remeshing capability in order to continuously maintain or regenerate
a finely discretized region moving with the heat source. The initial work on an automated remeshing strategy for welding applications was performed by Lindgren et al. [1]. Their work included
remeshing of a moving region but did not use any error estimation to guide the remeshing scheme
and control the accuracy of the solution produced. They prescribed the refinement/coarsening in
the input file so that a smaller distance to the source gave smaller elements. The size of the element
behind the heat source was also predetermined. Recently, Runnemalm and Hyun [2] proposed an
adaptative strategy that evaluates both the thermal and the mechanical error distribution using a
ZienkiewiczZhu error estimator [3]. It is combined with a hierarchic remeshing strategy using a
so-called graded element. In this approach, the directionality of the error estimation is ignored,
resulting in isotropic adaptative remeshing.
In the present paper, following the approach initiated by Fortin [4] and Alauzet et al. [5], we place
particular emphasis on the anisotropic mesh adaptation process generated by a directional error
estimator based on the recovery of the second derivatives of the different fields involved in the finite
element solution. The goal of this approach is to achieve a mesh-adaptative strategy minimizing
the directional error estimation in the mesh. As shown in this paper, this approach allows us to
refine the mesh, stretch and orient the elements in such a way that, along the adaptation process,
accurate controlled solutions are obtained while keeping the number of unknowns affordably low.
The organization of the paper is as follows. Section 2 introduces the numerical model that is used
to solve and describe the welding process. In this paper, the analysis is limited to coupled thermal
metallurgical simulations of welding. Section 3 presents the overall anisotropic mesh adaptation
procedure: the anisotropic error estimator, together with the procedures to get the recovered Hessian
matrix are described. In this section we discuss the Hessian strategy and review the concept of a
mesh metric field. Finally, in Section 4, the application of different anisotropic adaptative strategies
to welding simulations is presented and the results obtained are discussed.
2. WELDING ANALYSIS
During welding, the interaction of the heat source and the material leads to rapid heating, melting,
and the formation of the weld pool. When the heat source moves away, the weld pool cools and
solidifies. Depending on the welded alloys, as the temperature decreases, various solid-state phase
transformations take place resulting in the final microstructure of the weldment. The properties of a
weldment, such as strength, ductility, toughness, and corrosion resistance are significantly affected
by its microstructure. Thus, it is important to understand the temperature and microstructure
evolution during welding: this is the purpose of the present model. In the next two sections, a brief
overview of the thermalmetallurgical simulations of welding is given. The next two paragraphs
give a brief overview of the thermalmetallurgical model.
2.1. Heat transfer
Assuming thermal equilibrium at the microscopic scale, the transient heat transfer in a multiphase
solid continuum is governed by the following volume averaged equation (for more details see
Copyright q

2007 John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

ADAPTIVE MESH TECHNIQUE IN THE 3D THERMALMETALLURGICAL SIMULATION

Appendix A):


 
* H

g k j Hk
(T ) =
gik
*t
k=2,N i>1
k>1
i=k

(1)

k= j

where g denotes the volume fraction and the index k denotes either the liquid (k = 1) and (k = 2, N )
for different metallurgical phases that may exist in the solid state (see next section). The volumetric
enthalpy function H is defined as the integral of the heat capacity with respect to temperature:
 T
H (T ) =
c p  d + gl L v
(2)
0

where gl is the liquid volume fraction,  the density, c p is the specific heat, and L v denotes the
latent heat of fusion/solidification per unit of volume.
The term on the right-hand side of Equation (1) includes latent heat effects associated with solid
phase transformations. The enthalpy change associated with the i j transformation is equal to:
Hi j = gi (H j Hi ).
This formulation permits then to take into account the energy changes associated with the
solid-state phase changes, while taking advantage of the stability and robustness of the enthalpy
formulation for the liquidsolid phase change.
2.2. Solid-state phase transformation model
A series of phase transformations take place in both the fusion zone (FZ) and the heat-affected
zone (HAZ) during welding of low alloyed steels. A typical microstructural history of the
FZ is (ferrite, pearlite) austenite liquid austenite (ferrite, pearlite, bainite, martensite),
while a typical microstructure evolution in the HAZ corresponds to (ferrite, pearlite)
austenite (ferrite, pearlite, bainite, martensite). In the HAZ, the transformation during heating
(austenization) is of importance because it affects the kinetics of phase transformations during
cooling.
During heating, the calculation of austenite formation for an arbitrary thermal evolution is based
on the Leblond model for low carbon steel [6]. The rate of transformation is described according
to the expression:
eq

g a =
(T ) =
eq

ga (T ) =

ga (T ) ga (T )
(T )

(3)

(T Ae3 )


(4)

T Ae1
Ae3 Ae1

(5)

The transformation is described by Equation (3), where ga is the volume fraction of austenite and
eq
g a its time derivative, Ae3 the transformation equilibrium temperature, ga the phase equilibrium
defined in Equation (5),  is a function of temperature T , given by Equation (4) and  and  are
material constants.
Copyright q

2007 John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

M. HAMIDE, E. MASSONI AND M. BELLET

The model assumes that austenite starts to form when the temperature is above the austenite
transformation equilibrium temperature Ae1 . In the present work, the influence of grain size on
further transformations during cooling is neglected. Therefore, the evolution of grain size during
austenization is not modeled.
During cooling, the simulation of diffusive metallurgical transformations is based on TTT (time
temperature transformation) diagrams with a Scheils additivity rule for incubation or nucleation
and the JohnsonMehlAvrami (JMA) equation for growth. The use of the TTT diagram in the
case of a non-isothermal transformation is done considering its decomposition into successive
isothermal incremental transformations.
For incubation, the prediction of the beginning of the transformation of austenite into ferrite
pearlite or bainite is achieved using the so-called additivity principle, following Scheils rule, as
suggested by Denis et al. [7] and Fernandes et al. [8]. The considered transformation is supposed
to start when
 t
dt
=1
(6)
0 (T )
where (T ) is the incubation time to transform the fraction g isothermally at temperature T .
For growth, a JMA law (Equation (7)) is used to compute the fraction of ferrite, pearlite or
bainite transformed:
gk (t, T ) = 1 eb(T )t

n(T )

(7)

where gk describes the fraction transformed in an isothermal reaction as a function of time t and
b, n are temperature-dependent coefficients, to be determined from TTT diagrams (cf. [7, 8] for
details).
The calculation of the martensitic transformation is based on KostinenMarburger equation [9,
Equation (8)] and is dependent on the maximum austenite fraction gamax and temperature. It is
assumed that the transformation starts at the martensitic start temperature, Ms ,  being a material
constant which may depend on: composition, cooling rate, stress state
gm (t, T ) = gamax (1 e(Ms T ) )

(8)

3. MESH ADAPTATION
Adaptative finite elements problems are generally based on isotropic, a posteriori error estimates.
Basically, a posteriori error estimators can be classified into two types. The first of these was
introduced by Babuska and Rheinboldt [10, 11] and is based on evaluating the residuals of the
approximate solution. Therefore, it strongly depends on the problem operator. The second approach,
proposed by Zienkiewicz and Zhu [3], estimates the error in gradient-based norms, using some
recovery process (nowadays often called ZZ error estimators).
Recently, error estimators have been proposed for anisotropic meshes [12], the goal being
to reach a given level of accuracy with fewer vertices. The anisotropic interpolation estimates
introduced in [13] were used, together with a ZienkiewiczZhu error estimator to approach the
error gradient.
In the present paper, we place particular emphasis on the anisotropic mesh adaptation process
generated by a directional error estimator based on the recovery of the second derivatives of the
finite element solution (the so-called Hessian strategy).
Copyright q

2007 John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

ADAPTIVE MESH TECHNIQUE IN THE 3D THERMALMETALLURGICAL SIMULATION

3.1. The a posteriori error estimator


To obtain directional information of the error we use the Hessian strategy [5, 14], a method in which
the fields second derivatives are used to extract information on the error distribution. The Hessian
can be computed from any scalar component of the solution fields, in our case, the temperature or
the different phase fractions. As shown further, this directional information can be converted into
a mesh metric field which prescribes the desired element size and orientation.
Let us present the method as briefly as possible, since the details can be found in References
[5, 15, 16]. Denoting u the exact solution of the consistent scalar field and u h its discretization,
we use an indirect approach to estimate the error u u h . It has been proved that for elliptic
problems, the finite element error can be bounded by the interpolation error (Ceas lemma [17]):
u u h cu h u

(9)

where h u is the linear interpolate of u on the mesh and c is a constant independent of the current
mesh.
Here, we assume that this relation still holds in the class of problems envisaged. Actually, similar
analysis based on the interpolation error shows (practically) that the link between the interpolation
error and the approximation error is even stronger than the bound given by Ceas lemma [17].
Therefore, the interpolation error appears a reasonable way of defining an error estimate according
to [4].
The function u being supposed sufficiently smooth can be expanded into a Taylor series. Then
the interpolation error has a upper bound proportional to the second derivatives of u [14, 17]. To
express this upper bound, let us consider the following assumptions and notations:

The function u is regular enough and is associated with the solution of our welding problem.
K = [a, b, c, d] denotes a tetrahedron of the finite element mesh.
The P1 interpolation of u on element K is an affine function on K .
h u coincides with u at the vertices of K .

To bound the error u h u, we use its Taylor expansion with integral rest at a vertex of K
(for instance a) with respect to any interior point x in K :
 1
(u h u)(a) = (u h u)(x) + xa (u h u)(x) +
(1 t)(ax [Hu (x + txa)]ax) dt
0

where Hu denotes the second derivative of the variable u. Let us assume now that the maximal
error is achieved at point x (closer to a than to b, c, or d), so (u h u)(x) = 0, then we have
(xa (u h u)(x)) = 0. As h u coincides with u at the vertices of K ((u h u)(a) = 0), we
have
 1
e(x) =
(1 t)(ax [Hu (x + txa)]ax) dt
(10)
0

Let a be the projection of a on the tetrahedron surface, according to the direction ax (a is located
on the face opposite to a). There exists a positive real number  such that ax = aa. As a is closer
to x than any other vertex of K , then 3/4 [5]. Then, we obtain


9

1


|e(x)| 
(1 t)(aa [Hu (x + txa)]aa) dt
(11)
16 0

Copyright q

2007 John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

M. HAMIDE, E. MASSONI AND M. BELLET

Finally
|e(x)|

9
max |(aa [Hu (y)]aa)|
32 yK

(12)

Relationship (12) is not useful in practice as the bound depends on the extremum x which is not
known a priori. However, it can be reformulated as follows [5]:
e K  c max max(v |Hu (x)|v)
xK

(13)

where c = 9/32, v is any vector joining two interior points in K and |Hu | is the absolute value of
the Hessian matrix of the solution (i.e. consisting of absolute eigenvalues).
The bound of the previous relation is difficult to compute. As we can replace all vectors included
in K with a combination of the edges of K , a new upper bound error can finally be obtained [5]:
e K  c max max (e |Hu (x)|e)
xK

(14)

where e denotes one of the six tetrahedron edges.


The Hessian strategy involves the computation of the symmetric matrix of second derivatives
that can be decomposed as |Hu | = R||R T , where R is the eigenvectors matrix and  = diag(k ) is
the diagonal matrix of eigenvalues. The directions associated with the eigenvectors k are referred
to as principal directions and the eigenvalues k are then equivalent to the second derivatives along
the local principal directions.
3.2. Metric definition
Performing anisotropic mesh adaptation requires a way of defining the desired element size distribution over the domain. Mesh metric tensors are used to represent an anisotropic mesh size field
defining the desired mesh anisotropy at a point (see, for example, [5, 18]). The concept of a mesh
metric field is used to represent the desired size field as a tensor over the domain.
The Hessian strategy is based on the idea that a high magnitude of an eigenvalue implies a high
error in the direction associated with the corresponding eigenvector, so a small element size would
be desired in this direction. Conversely, a low eigenvalue magnitude in a particular eigendirection
suggests that the element size could be large along this direction.
To achieve a suitable mesh resolution in different directions, a uniform distribution of local
errors is applied in the principal directions which leads to ch 2k k = , where  is the user specified
tolerance for the error and h k is the desired size in the kth principal direction. So, the edges e of
the adapted mesh must check  = c(e Me).
The stated goal of the mesh adaptation algorithm is to yield a mesh with regular elements in
the metric space where each edge e must satisfy the following relation (see, for example, [18]):
=1
e Me

(15)

A mesh with all its edges satisfying the above relationship is commonly referred to as a unit mesh.
A mesh metric tensor M is then obtained at each node by calculating a scaled eigenspace of the
T , where R is the eigenvector matrix and 
= (c/) is the
recovered Hessian matrix as M = R R
diagonal matrix of scaled eigenvalues.
Truncation values h min and h max are specified to limit mesh sizes. One reason for truncating the
element size, in terms of edge lengths, is to avoid singular metrics. For example, it is necessary
Copyright q

2007 John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

ADAPTIVE MESH TECHNIQUE IN THE 3D THERMALMETALLURGICAL SIMULATION

Figure 1. Two-dimensional schematic of the intersection of mesh metric tensors represented by ellipses.

to apply h max in case of a null or quasi-null eigenvalue in the direction where the solution does
not vary. The modified eigenvalues of the Hessian matrix then become



1
1
c
|k |, 2
(16)
k = min max
, 2

h max
h min
where  is the prescribed error.
When several variable fields u are considered concurrently, the previous approach leads to a
metric for each variable and we chose to take the intersection of the different metrics. In practice,
the intersection of metrics is achieved by the simultaneous reduction of two quadratic forms which
is a valid operation since the metric tensors are positive definite, for details see [5]. We illustrate
the procedure from a geometrical point of view in Figure 1. It can be shown that this process
allows one to compute a common basis for the two quadratic forms that can be used to determine
the ellipsoid with maximum volume contained in the geometrical intersection of the two candidate
ellipsoids. The ellipsoid representing the final intersected metric is the one with the maximum
volume contained in the common volume of all the candidates and therefore respects the size
requirements of the different metrics.
3.3. Relative error
To combine various variables together to construct a metric, it is often reasonable to consider a
relative bound on the error. In order to have dimensionless error, we define the following estimation:




 u h u 
|Hu (x)|


c max max e
e
(17)
 |u|

xK eE k
|u|

,K
where |u|
= max(|u|, u, ) with  is a cut-off to avoid numerical difficulties (in case of a
null or quasi-null value of u).
3.4. Hessian evaluation
To compute the Hessian matrix Hu of the P1 field u, we reconstruct in two steps the second
derivatives at each node P by using the computed solution from the patch S of all elements K
surrounding node P. This is done as follows.
Copyright q

2007 John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

M. HAMIDE, E. MASSONI AND M. BELLET

In the first step, we recover the gradient at node P by taking the volume weighted average of
gradients on elements in the patch S surrounding the node P. Note that u being a P1 field, its
gradient is constant on each element K :



1
h (u h )(P) = 
|K |(u h )|K
(18)
K S |K | K S
It can also be noticed that this is equivalent to a lumped-mass approximation of a least-squares
reconstruction of the gradient for linear elements.
In the second step, the same procedure is now applied to the P1 field h (u h ) to obtain the
recovered Hessian matrix:





1
*u h
*

(Hu )i j (P) =
h
(19)
|K |
*x j
*xi
K S |K | K S
3.5. Transfer of variables
State variables should be transferred from the old to the new mesh after remeshing. In the context
of thermalmetallurgical simulation of welding, these variables consist of enthalpy and phase
fractions (liquid and solid). Since they are defined at nodes, a direct interpolation can be used.
For each node of the new mesh, its location in the old mesh is determined (element and local
isoparametric coordinates). The new values at this node are then obtained by interpolation in the
element of the old mesh. It should be noted that the temperature field can be determined, in a
second step, from the new value of the enthalpy and phase fraction.

4. APPLICATION TO WELDING SIMULATION


4.1. Welding conditions and material properties
We consider a thick plate of A508 steel, the dimensions of which are given in Figure 2(a). The
temperature-dependent thermophysical properties are given in Figure 3.
The welding parameters chosen for this analysis are as follows. Welding process: gastungstenarc welding (GTAW); welding current I = 150 A, welding voltage V = 10 V and traveling speed

Figure 2. (a) Geometry of the specimen (dimensions in mm) and comparison point A and (b) finite element
reference mesh of one half of the plate.
Copyright q

2007 John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

ADAPTIVE MESH TECHNIQUE IN THE 3D THERMALMETALLURGICAL SIMULATION

Figure 3. Thermophysical properties used in analysis (SI units).

of 1 mm s1 . The weld efficiency is assumed to be  = 0.65. The associated heat input, I V ,


moving with the electrode, is simulated by a simple surface flux with uniform distribution within
a disc of radius 5 mm.
4.2. Finite element model
For the simulation study, only one-half of the plate is analyzed due to symmetry. The boundary
conditions of the welding process incorporate heat transfer boundary conditions. The symmetry surface is defined as under adiabatic boundary conditions. On all other surfaces, boundary conditions of convection and radiation with the environment are applied with a convective
coefficient h = 12 W m2 K1 and emissivity coefficient  = 0.75 and an external temperature
Text = 25 C.
To evaluate the efficiency of our adaptative procedure, we first obtain results on a fine mesh
(Figure 2(b)). The mesh size along the electrode path has been fixed after a preliminary study
involving different meshes of different mesh sizes in this region. The value 1 mm for the mesh
size has been fixed after checking the stationarity of the temperature solution with mesh size. The
result obtained is then used for purpose of comparison. Several simulations have been performed
with adaptative remeshing. The calculations differ by the spatial discretization, all other conditions
being identical. Two types of simulation with remeshing have been carried out:
Thermal-driven mesh adaptation: The adaptative technique is based on the thermal error distribution.
Thermometallurgical-driven mesh adaptation: The automatic mesh refinement is based on both
thermal and metallurgical error distributions.
4.3. Thermal-based mesh adaptation
The reference FE model without remeshing consists of 14 329 nodes and 68 891 linear tetrahedral
elements and is presented in Figure 2(b). As indicated above, the minimum mesh size is 1 mm
along the electrode path, the maximum mesh size being 10 mm. The initial FE model used in
calculations with remeshing consists of 6842 tetrahedral elements (1683 nodes).
Copyright q

2007 John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

M. HAMIDE, E. MASSONI AND M. BELLET

Figure 4. Thermal-based mesh adaptation ( = 0.01): (a) anisotropic FEM mesh; (b) zoom on refined
zone with anisotropic remeshing; (c) temperature distribution at time 95 s (K); and (d) zoom on
refined zone with isotropic remeshing.

The remeshing is performed at each time step (dt = 1 s). See Figures 2(b) and 4(a) for the FE
mesh. As expected, the adaptative remeshing generates more refined elements in the neighborhood
of the thermal source and coarser elements in its trail. It can also be seen in Figure 4 that anisotropic
elements aligned with the heat flow are created around the FZ. It should be noted that the aspect
ratio of the elements reach values from 1 to 10 in this region (the allowed aspect ratio for this
simulation was h max / h min = 10). The minimum mesh size is 1 mm.
As shown in Table I, the calculation on the fine reference mesh (without remeshing) required
6 h and 25 min of CPU time. Two calculations with anisotropic remeshing have been performed,
one with a prescribed error  = 0.01 and a second one with  = 0.005. The CPU time was 1 h and
1 min for  = 0.01 and 1 h and 52 min for  = 0.005. The final number of elements in the second
calculation is much higher than in the first one with the same truncation element size values. As
expected, the calculation with  = 0.005 generates a larger refined zone in the neighborhood of the
thermal source than the calculation with  = 0.01.
Copyright q

2007 John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

ADAPTIVE MESH TECHNIQUE IN THE 3D THERMALMETALLURGICAL SIMULATION


imp

Table I. Refinement parameters (Nbe denotes the number of elements, h max the prescribed maximum
imp
element size, h min the prescribed minimum element size, h max the maximum element size, h min the
minimum element size).

Fine reference mesh


Coarse reference mesh
Anisotropic adapted mesh,  = 0.01
Isotropic adapted mesh,  = 0.01
Anisotropic adapted mesh,  = 0.005
Isotropic adapted mesh,  = 0.005

imp

imp

Initial
Nbe

Final
Nbe

h min
(mm)

h max
(mm)

h min
(mm)

h max
(mm)

CPU time

68 891
11 439
6842
6842
6842
6842

68 891
11 439
5866
10 685
11 012
46 906

1
2
1
1
1
1

10
10
10
10
10
10

1
2
0.95
0.95
0.9
0.9

10
10
11.5
11.5
10.6
10.6

6 h 25 min
58 min
1 h 1 min
1 h 57 min
1 h 52 min
4 h 19 min

Note: Calculation run on a Pentium 4 PC, 2 GHz with 2 Gb RAM.

Figure 5. Thermal-based mesh adaptation ( = 0.01): (a) temperature evolution at Point A and
(b) temperature distribution at time 95 s (K).

In order to demonstrate the efficiency of the remeshing procedure due to its anisotropic (i.e.
directional) character, we performed additional comparisons between anisotropic and isotropic
remeshings, using the same objective accuracy  and the same truncation values h min and h max .
The results are reported in Table I. It can be seen that the number of elements required to produce
the same level of interpolation error is significantly different between isotropic and anisotropic
meshes. With  = 0.01, the anisotropic mesh uses only 5866 elements, about half of the 10 685
elements that are required by the isotropic mesh. The computation time is reduced by a factor 2,
when using remeshing. A similar trend is found when prescribing a more stringent accuracy
level ( = 0.005). A zoom on both meshes near heat source (Figure 5(b) and (d)) reveals that
the elements of the isotropic mesh in the region are small and almost equilateral, whereas the
anisotropic elements are stretched orthogonally to the temperature gradient.
An example of temperature distribution is given in Figure 5(c). The temperature evolution
at Point A in the different analyses is shown in Figure 5(a). The first observation from the
plots in Figure 5(a) is that the results are significantly smoother in the time domain than in
the spatial domain, as shown in Figure 5(b). This illustrates that the spatial noise associated with
Copyright q

2007 John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

M. HAMIDE, E. MASSONI AND M. BELLET

Figure 6. Thermal-based mesh adaptation ( = 0.01): (a) bainite distribution at time 95 s and (b) time
evolution of phases proportions at Point A.

Figure 7. Evolution of the temperature difference T = |T Tref |, at Point A, where Tref is the temperature
obtained on the reference fine mesh.

the Hessian recovery does not globally pollute the solution in time, suggesting that the primary
fields (temperature and phase fractions) remain unaffected (Figure 6). The adaptative technique
makes the FEM mesh much denser, so that the temperature distribution is more accurate than with
coarse mesh (see Figures 5 and 7).
We observe in Figure 5(a) that the temperature evolution Point a A converges to the reference
temperature evolution when reducing the prescribed error . From Table I and Figure 5(a), it can
be seen that for a comparable accuracy of the results, the use of an adaptative meshing procedure
reduces CPU costs by a factor 6. This shows the efficiency of the proposed approach.
4.4. Thermometallurgical-based mesh adaptation
Comparing Figures 4 and 8 a clear difference between the obtained two meshes when using
only the thermal error distribution (Figure 4(a)) or both the thermal and the metallurgical error
Copyright q

2007 John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

ADAPTIVE MESH TECHNIQUE IN THE 3D THERMALMETALLURGICAL SIMULATION

Figure 8. A thermometallurgical-based mesh adaptation ( = 0.01): (a) FEM


mesh and (b) zoom on refined zone.
Table II. Refinement parameters and results for thermalmetallurgical adaptation.

Fine reference mesh


Adapted mesh, error 0.01

imp

imp

Initial Nbe

Final Nbe

h min (mm)

h max (mm)

CPU time

68 891
6842

68 891
15 816

1
1

10
50

6 h 25 min
2 h 22 min

Note: Calculation run on a Pentium 4 PC, 2 GHz with 2 Gb RAM.

Figure 9. A thermometallurgical-based mesh adaptation ( = 0.01): (a) temperature evolution at Point A


converges and (b) temperature distribution at time 95 s (K).

distributions is evidenced. It is shown in Figure 4(a) that the automatic mesh refinement using
temperature as error indication produced an elliptical zone in the vicinity of the FZ. A distinct
behavior is found when guiding the mesh adaptation with both phase proportion (in the present
Copyright q

2007 John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

M. HAMIDE, E. MASSONI AND M. BELLET

Figure 10. A thermometallurgical-based mesh adaptation ( = 0.01): (a) time evolution of phases
proportions at Point A and (b) bainite distribution at time 95 s.

Figure 11. Profiles of bainite volume fraction in a cross-section located at X = 0.095 m at time 250 s.

case, the bainite volume fraction) and temperature. In this latter case, the mesh is much denser in
the wake of the heat source in order to provide a better representation of steep gradients of phase
fraction. It can be seen that the thermometallurgical-driven mesh generation produces significantly
more elements than the thermal-driven one (see Table II, Figures 4 and 8). This is due to the
residual gradients of phases fractions that remain in the welded component after welding. The
thermal-driven remeshing creates a much lower number of elements in the model. This is of course
due to a smoother gradient field in the thermal analysis and also because the plate cools down to
a uniform temperature.
Comparing Figures 10(a) and 6(b), it can be seen that the phase time evolutions in a given
point are not significantly affected by mesh adaptation, in agreement with the same trend for
temperature (Figures 9(a) and 5(a)). Regarding the spatial distribution of the phases, the impact
Copyright q

2007 John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

ADAPTIVE MESH TECHNIQUE IN THE 3D THERMALMETALLURGICAL SIMULATION

is much more significant, as expected. Figure 11 shows the residual profile of the bainite volume
fraction in a transverse section, using different meshes. It can be seen that the phase distribution
is much more accurate than with an adaptation based only on the thermal error distribution. The
gradients of phase fraction are better described than with the reference fine mesh. The coupled
thermalmetallurgical results in an optimal description of the distribution of the different phases.
This result is extremely important in terms of the prediction and assessment of the quality of
weldments, for which an accurate determination of the phase fractions prevailing in the HAZ is a
key factor. Beyond the simple thermalmetallurgical approach considered here, this result is also
valuable in view of further thermalmetallurgicalmechanical calculations which are necessary to
predict the risk of failure during welding and the residual stresses and deformations.

5. CONCLUSION
In this paper, adaptive remeshing procedures have been presented and applied in the context of
coupled thermalmetallurgical simulation of welding. The method is based on anisotropic mesh
adaptivity dictated by directional error estimators. Those estimators, based on the Hessian recovery
of P1 field, are used to construct a mesh metric field that provides information on the local mesh
resolution desired in different directions. The method allows to easily combine metric tensors for
variables of different types and nature.
In practice, the calculation results show that the temperature field and the distributions of phase
fractions with adaptive mesh converge to the results obtained with reference mesh. For the case
tested here, the calculation time comparison shows that the adaptive mesh technique can reduce
the calculation time by almost one-third. It also reduces the data-storage requirement substantially.
For some applications, both points are key factors in determining whether a successful FE simulation
can be completed or not.
Larger savings may be expected for application with longer welds as the zone associated with
large gradients will be smaller relative to the total length of the weld. In the framework of thermalmetallo-mechanical simulation, the current logic for deciding the size of the elements in relation
to thermal and metallurgical fields should be combined with mechanical fields: this should be
studied in future works. Future studies will incorporate a moving mesh strategy based on error
estimation to reduce the number of remeshing steps and accelerate the efficiency of the adaptative
method.

APPENDIX A: VOLUME AVERAGING FOR HEAT TRANSFER


In the welding context, metallurgical transformations highly depend on temperature history. Conversely, the impact of phase transformations (liquidsolid and solidsolid) on heat transfer should
be considered.
The application of the spatial averaging method to the equation of energy conservation in a
elementary representative volume (REV) of the multiphase material [19], yields for each phase k
that may exist in the REV:
*(gk k h k )
+ (gk k h k vk ) (qk ) = Q k
*t
Copyright q

2007 John Wiley & Sons, Ltd.

(A1)

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

M. HAMIDE, E. MASSONI AND M. BELLET

where gk denotes the volume fraction of phase k (k = 1, N ), k the density of phase k, h k its
enthalpy per unit mass, vk its intrinsic average velocity, qk the intrinsic average heat flux vector
in phase k, and Q k the heat exchange rate with other phases.
Summing these equations for the different phases k = 1, N , and assuming a uniform temperature
on the REV and the Fourier law for heat conduction, we get, using the convention of summation
for repeated indices:
*(gk k h k )
+ (gk k h k vk ) (gk k Tk ) = 0
*t

(A2)

where  denotes the heat conductivity.


Neglecting advection effects, and denoting H the enthalpy per unit volume, we get
*(gk Hk )
(gk k Tk ) = 0
*t

(A3)

Noting now that *Hk /*t = (*Hk /*T )*T /*t = (c p )k *T /*t, and denoting  = gk k the average
heat conductivity and c p  = gk (c p ) the average heat capacity, this leads to
*gk
*T
+ C p 
(Tk ) = 0
*t
*t

(A4)

Let us define now some notations regarding the different phase changes that may occur in the
REV. For each phase k, the rate of change of the volume fraction can be expressed by

*gk 
gik
g k j
=
*t
i=k
j=k

(A5)

in which gik denotes the rate of transformation of phase i into phase k, using the following
convention: gik >0 when phase i is partially transformed into phase k, and gik = 0 otherwise.
In what follows, we will separate the fusion/solidification from the other phase changes occurring
in the solid state. The liquid phase will then be identified by the index k = 1, and the different
solid phases by k = 2, N . Equation (A4) then becomes




 

gi1
g 1 j H1 +
gik
g k j Hk
i=1

1= j

k=2,N

i=k

k= j

*T
(Tk ) = 0
*t






gi1
g 1 j H1 +
g 1k
g k1 Hk
+c p 

i=1

1= j

k=2,N

k=1

 

*T

g k j Hk + c p 
(Tk ) = 0
gik
*t
k=2,N i>1
k>1
i=k

Copyright q

1=k

2007 John Wiley & Sons, Ltd.

(A6)

(A7)

k= j

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

ADAPTIVE MESH TECHNIQUE IN THE 3D THERMALMETALLURGICAL SIMULATION

The two first terms deal with fusion or solidification, while the third one encompasses solid-state
phase transformations only. Rearranging the two first terms, denoting now the liquid phase with
the index l 1, and putting the term of solid-state transformations on the right-hand side, we get



 
*T
c p 
gi1
gik (Hl Hk ) (Tk )
+
*t
i=k
k=2,N i=1


 

g k j Hk
gik

k=2,N

i>1
i=k

(A8)

k>1
k= j

The first term can be reasonably approximated by L v *gl /*t, with L v the latent heat of fusion per
unit volume. We then obtain

c p 

 

*T
*gl

g k j Hk
+ Lv
(Tk ) =
gik
*t
*t
k=2,N i>1
k>1
i=k

(A9)

k= j

For the sake of simplification, we approximate the two first terms by the time derivative of H , a
function of the temperature only, which is defined as follows and can be seen as a pseudo-average
enthalpy:
 T
H (T ) =
c p  d + gl L v
(A10)
0

In this expression, T0 is a reference temperature and the averaging of the heat capacity c p  is
defined a priori, that is with fixed predetermined volume fractions of the different phases for each
temperature. This approximation finally yields:


 
*H

g k j Hk
(A11)
(T ) =
gik
*t
k=2,N i>1
k>1
i=k

k= j

REFERENCES
1. Lindgren LE, Haggblad HA, McDill JMJ, Oddy AS. Automatic remeshing for three-dimensional finite element
simulation of welding. Computer Methods in Applied Mechanics and Engineering 1997; 147:401409.
2. Runnemalm H, Hyun S. Three-dimensional welding analysis using an adaptive mesh scheme. Computer Methods
in Applied Mechanics and Engineering 2000; 189:515523.
3. Zienkiewicz OC, Zhu JZ. A simple error estimator and adaptive procedure for practical engineering analysis.
International Journal for Numerical Methods in Engineering 1987; 24:337357.
4. Fortin M. Estimation derreur a posteriori et adaptation de maillages. Revue europeenne des e lements finis 2000;
9(4):467486.
5. Alauzet F, Frey PJ, George PL. Anisotropic mesh adaptation for RayleighTaylor instabilities. European Congress
on Computational Methods in Applied Sciences and Engineering (ECCOMAS), Jyvaskyla, Finland, June 2004.
6. Leblond JB, Devaux J, Devaux JC. A new kinetic model for anisothermal metallurgical transformations in steels
including effect of austenite grain size. Acta Metallurgica 1984; 32:137146.
Copyright q

2007 John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

M. HAMIDE, E. MASSONI AND M. BELLET

7. Denis S, Farias D. Mathematical model coupling phase transformations and temperature evolutions in steels. ISIJ
International 1992; 32:316325.
8. Fernandes F, Denis S, Simon A. Mathematical model coupling phase transformation and temperature evolution
during quenching of steels. Materials Science and Technology 1985; 1:838844.
9. Kostinen DP, Marburger RE. A general equation prescribing extent of austenitemartensite transformation in
pure ironcarbon alloys and carbon steels. Acta Metallurgica 1959; 7:417426.
10. Babuska I, Rheinboldt WC. Error for adaptive finite element method. SIAM Journal on Numerical Analysis 1978;
15:736754.
11. Babuska I, Rheinboldt WC. A posteriori error estimators in the finite element method. International Journal for
Numerical Methods in Engineering 1978; 12:15971615.
12. Picasso M. An anisotropic error indicator based on ZienkiewiczZhu error estimator: application to elliptic and
parabolic problems. SIAM Journal on Scientific Computing 2003; 24:13281355.
13. Babuska I, Aziz AK. On the angle condition in the finite element method. SIAM Journal on Numerical Analysis
1976; 13:214226.
14. Kunert G. Toward anisotropic mesh construction and error estimation in the finite element method. Numerical
Methods for Partial Differential Equations 2002; 18:625648.
15. DAzevedo EF, Simpson B. On optimal triangular meshes for minimizing the gradient error. Numerische
Mathematik 1991; 59(4):321348.
16. Garcia NP, Anglada MV, Crosa PB. Directional adaptive surface triangulation. Computer-Aided Design 1999;
16:107126.
17. Ciarlet PG. Basic error estimates for elliptic problems. In Finite Element Methods, Ciarlet PG, Lions JL (eds).
Handbook of Numerical Analysis, vol. 2. North-Holland: Amsterdam; 17352.
18. Gruau C, Coupez T. 3D tetrahedal unstructured and anisotropic mesh generation with adaptation to natural and
multidomain metric. Computer Methods in Applied Mechanics and Engineering 2005; 194:49514976.
19. Rappaz M, Bellet M, Deville M. Numerical Modeling in Materials Science and Engineering. Springer Series in
Computational Mathematics. Springer: Berlin, 2003.

Copyright q

2007 John Wiley & Sons, Ltd.

Int. J. Numer. Meth. Engng (in press)


DOI: 10.1002/nme

Anda mungkin juga menyukai