9th International LS-DYNA Users Conference Metal Forming (2)

A Grid-based Adaptive Scheme for the Three-Dimensional

Forging and Extrusion Problems with the EFG Method
H. S. Lu and C. T. Wu
Livermore Software Technology Corporation
Livermore, CA 94551
[email protected], [email protected]

Abstract
Procedures to adaptively refine meshes are emerging as an important tool for improving
accuracy and efficiency in large deformation and fracture analysis. Comparing to the mesh-
based adaptive method, the grid-based adaptive mesh-free method has several built-in
advantages including naturally conforming in shape functions, smoothed interpolations in
surface construction and state variable transfer, and the results are less sensitive to the
unstructured grids. In this paper, a grid-based adaptive scheme is proposed for the Element Free
Galerkin method in the large deformation analysis of three-dimensional forging and extrusion
simulations. To precisely account for the effect of kernel functions during adaptive procedure,
we consider the meshfree adaptivity in the framework of arbitrary Lagragian-Eulerian meshfree
method with an operator-split time integration. A grid-based interpolation scheme adopted from
the meshfree approximation is developed for the state variable transfer after mesh refinement to
improve conservation and monotonic properties. Several industrial problems have been solved
and compared to the existing numerical methods.

1. Introduction
Adaptive procedure has become an important tool for structural analysis, especially in problems
with large deformation, moving interface, and local behavior, where the optimal mesh
dramatically changes during the deformation process. Several complex issues, such as error
indicator, mesh refinement, state variable transfer, are involved in the adaptive procedures. In
this work, we mainly address the transfer of state variables between the successive meshes,
which is crucially important for preserving the accuracy, convergence properties in solving the
problems with history-dependent materials. Several important aspects should be considered for
the state variable transfer, such as consistency with the constitutive equations, requirement of
equilibrium, compatibility of the state variables at gauss points with the state variables at nodal
points, and minimization of the numerical diffusion, etc. Various transfer schemes can be found
in the literatures [2, 5, 10, 11] for finite element method. The approaches used in [5, 11] are very
similar. The state variables in the old mesh are first projected to nodes, and then the values at
new integration points are interpolated from the old mesh. A different approach is proposed in
[10]. The state variables are transferred from old gauss points to the new gauss points by local
interpolation function. This direct transfer does not require nodal projection, but the continuity of
state variable along the element boundary is not ensured. Recently developed approximation
theories such as moving least square approximation (MSL) [1], reproducing kernel
approximation (RKPM) [8] allow the interpolation functions to be constructed entirely in terms
of arbitrarily placed nodes without the used of an element connectivity. This makes the direct
transfer of information between the subsequent meshes with a desirable consistency and
accuracy possible, especially in the direct transfer of state variable at integration point during the
adaptive procedure. In this work, a local interpolant with a desirable consistency condition,

17-33
Metal Forming (2) 9th International LS-DYNA Users Conference

which is based on the modified moving-least squares approximation (or the modified
reproducing kernel approximation), is constructed to obtain the state variables at new integration
points from old integration points.

Recent developments in mesh-free methods add an additional dimension to computational

mechanics [1, 3, 8, 14, 15]. Those methods do not rely on the conventional grid approach to
define approximation functions. In comparison with conventional finite element methods, the
characteristics of smoothness and naturally conforming of the approximation, exemption from
meshing, and higher convergence rate and the easy of nodal insertion and deletion have make
mesh-free methods attractive alternative numerical techniques for nonlinear analysis of industrial
applications. Although mesh-free method has been extensively investigated and applied to large
deformation analysis in the past few years [1, 3, 6], its application to forging and extrusion
problems that are beyond the Lagrangian description is still a challenging problem due to the
issues in moving surface representation, volumetric locking and domain integration. In this work
we take the meshfree adaptivity [9, 13, 15] in the framework of arbitrary Lagrangian-Eulerian
meshfree method [12]. By doing so, the adaptive procedure is separated into kernel function
redefining due to the material deformation and mesh refinement based on the error indicator. An
advection algorithm is employed to account for the effect of kernel function change in the
adaptive procedure. The refinement of grids is accomplished by a delaunay-type algorithm with a
grid distance control. A special mesh-free interpolation is used to build a complete three-
dimensional geometric and topologic representation to preserve the volume during the adaptive
procedure.

The layout of this paper is as follows. In section 2, moving least square approximation, one of
the most popular meshfree approximations, is reviewed. Meshfree adaptivity with adaptive
Lagrangian particles and Eulerian kernel is discussed in Section 3. In section 4, the transfer of
state variables between the subsequent meshes is presented. Two numerical examples are given
in Section 4.

2. Meshfree Approximation
The discrete moving least square (MLS) approximation of a function u ( x ) , denoted by u h ( x ) , is
constructed by a combination of the monomials as

n
u h ( x ) = ∑ H i ( x )bi ( x ) ≡ H T ( x )b( x ) (1)
i =1

where n is the order of completeness in this approximation, the monomial H i ( x ) are basis
functions, and bi ( x ) are the coefficients of the approximation.

The coefficients bi ( x ) at any point x are depending on the sampling points x I that are collected
by a weighting function wa ( x - x I ) . This weighting function is defined to have a compact
support measured by ‘a’, i.e., the sub-domain over which it is nonzero is small relative to the rest
of the domain. Each sub-domain ΔΩ I is associated with a node I. The most commonly used sub-
domains are disks or balls.

17-34
9th International LS-DYNA Users Conference Metal Forming (2)

In this development, we employ the cubic B-spline kernel function as the weighting function:

⎧ 2 3 ⎫
⎪ 2 ⎛⎜ x − x I ⎞ ⎛ x − xI
⎟ + 4⎜
⎞
⎟
x − xI 1⎪
⎪ − 4⎜ ⎟ ⎜ ⎟ for 0 ≤ ≤ ⎪
⎪ 3 ⎝ a ⎠ ⎝ a ⎠ a 2⎪
⎪ 2 2 ⎪
⎪ 4 ⎛ x − x I ⎞⎟ ⎛⎜ x − xI ⎞
⎟ −4⎜
⎛ x − xI ⎞ 1 x − xI ⎪
w a ( x − x I ) = ⎨ − 4⎜ +4 ⎟ for < ≤ 1⎬ (2)
⎜ ⎟ ⎜ a ⎟⎠ 3 ⎜⎝ a ⎟⎠
⎪3 ⎝ a ⎠ ⎝ 2 a ⎪
⎪ ⎪
⎪ ⎪
⎪ 0 otherwise ⎪
⎪⎩ ⎪⎭

The moving least-squares technique consists in minimizing the weighted L2 -Norm

2
NP
⎡n ⎤
J = ∑ wa ( x − x I )⎢∑ H i ( x )bi ( x ) − u( x I )⎥ (3)
I =1 ⎣ i =1 ⎦

where NP is the number of nodes within the support of x for which the weighting function
wa ( x - x I ) ≠ 0 .

Equation (3) can be written in the form

J = ( Hb − u )T W a ( x )( Hb − u ) (4)
where

u T = (u 1 , u 2 , u NP ) (5)
⎡ {H ( x 1 )} ⎤ T

⎢ ⎥
H =⎢ ⎥ (6)
⎢{H ( x )}T ⎥
⎣ NP ⎦
{H ( x i )}T = {H 1 ( x i ),… H n ( x i )} (7)
W a = diag [wa ( x − x 1 ), , wa ( x − x NP )] (8)

To find the coefficients b, we obtain the extremum of J by

∂J = M [ n ] ( x )b( x ) − B( x )u = 0 (9)
∂b
where M [ n ] ( x ) is called the moment matrix of wa ( x − x I ) and is given by

M [ n] ( x ) = H T W a ( x )H (10)

B( x ) = H T W a ( x ) (11)

17-35
Metal Forming (2) 9th International LS-DYNA Users Conference

So we have
−1
b( x ) = M [n] ( x )B( x )u (12)

For M [ n ] ( x ) to be invertible, the support of wa ( x - x ) needs to be greater than a minimum size

that is related to the order of basis functions.

Using the solution of Equations (1), (10), (11) and (12), the EFG approximation is obtained by

NP
u h ( x ) = ∑ ΨI ( x )u I (13)
I =1

where the EFG shape functions Ψ I ( x ) are given by

−1
ΨI ( x ) = H T ( x ) M [n] ( x ) B ( x ) (14)

and ΨI ( x ) are nth-order complete, i.e.

NP

∑ Ψ ( x) x
I =1
I
p
1I x 2qI x3rI = x1p x 2q x3r for p + q + r = 0, n (15)

In general, mesh-free shape functions ΨI do not possess Kronecker delta properties of the
standard FEM shape functions, i.e.

ΨI ( x J ) ≠ δ IJ (16)

This means that the coefficients of the inetrpolant are not the same as the nodal value, i.e.

u h ( xI ) ≠ uI (17)

3. Adaptive Lagrangian Particles with Eulerian Kernel

Without using the explicit mesh, meshfree method is more suitable for the adaptive procedure
than the conventional finite element method. Particles can be easily added and removed without
complications in the data structure. The accuracy and smoothness of meshfree solution reduce
the numerical diffusion during the transfer of state variable between the subsequent meshes.
Beside the mesh refinement, kernel functions (weighting functions) associated with the particles
also need to be redefined in the procedure of adaptivity. To precisely account for the effect of
kernel functions, we consider the meshfree adaptivity in the framework of arbitrary Lagrangian-
Eulerian meshfree method with an operator-split time integration. As shown in Figure 1, the
adaptive procedure is separated as redefining kernel functions in the old mesh due to the material
deformation, and mesh refinement to create new mesh.

17-36
9th International LS-DYNA Users Conference Metal Forming (2)

Lagrangian Eulerian Mesh

substep tn substep tn refinement tn

Figure 1 Adaptive Procedure in Meshfree Method

Three coordinates, the Lagrangian or material coordinate X , the spatial coordinate x , and the
referential coordinate χ , and the three time derivatives of a function f , the material derivative
∂f ( X , t ) ∂f ∂f ( x, t )
f = , the spatial derivative = , and the referential derivative
∂t X ∂t ∂t x
∂f ( χ , t )
f = are defined respectively. The relationship between the material and referential
∂t χ
time derivative can be written as:

∂f
f = f + c ⋅ ∇f = f + c i (18)
∂xi

Where c is the relative velocity between the material v = x and the referential vˆ = x velocity.

c = v − vˆ (19)

The referential momentum equation of the transient problem can be written as

σ ij , j + ρbi = ρ (vi + c j vi , j ) in Ω (20)

Where σ ij is the stress field, ρ is mass density, and bi is the body force. The boundary and initial
conditions are given as u i = u i on Γu , σ ij n j = hi on , u i ( X ,0) = u i0 ( X ) , vi ( X ,0) = vi0 ( X ) , and

v i ( X ,0) = vi0 ( X ) respectively. The corresponding weak form reads

∫Ω
ρδu i v i dΩ + ∫ ρδu i c j vi , j dΩ + ∫ δu i , j σ j ,i dΩ = ∫ ρδu i bi dΩ + ∫ δu i hi dΓh
Ω Ω Ω Γh
(21)

Substituting the meshfree approximation into Equ.(21) to obtain the matrix form

17-37
Metal Forming (2) 9th International LS-DYNA Users Conference

M v + Cv + Ku = f ext (22)

Where
M IJ = ∫ ρ ΨI ( χ )ΨJ ( χ )dΩ
Ω

C IJ = ∫ ρc jψ I ( χ )ψ J , j ( χ )dΩ (23)
Ω

K IJ = ∫ B I DB J dΩ
Ω
ext
=
∫ ρψ I ( χ )bdΩ + ∫ ψ I ( χ )hdΓh
f I
Ω Γh

B is the strain-displacement matrix

An operator-split with a Lagrangian substep followed by an Eulerian substep is adopted to

integrate the equations in time if the adaptive procedure is invoked. During the Lagrangian
substep, the material and referential coordinate systems are chosen as the same, and the
convective velocity c is trivially zero, i.e. v = vˆ . In the Eulerian substep, the kernel function is
redefined due to the current material deformation. The solution variables have to be transferred
with an advection algorithm between the same mesh due to the different kernel functions.

f = f LAG
+ Δf χ
≈ f LAG
+ d ⋅ ∇f (24)

LAG
Where d is the total displacement during the convective step. f is the quantity f from the
Lagrangian substep.

Unlike in ALE finite element method, the new state variable value after convection can be
naturally obtained due to the smoothness of the shape function by

∂ψ I
f = f LAG
+ d ⋅ (∑ fI ) (25)
I ∂χ χ = χ LAG +θd

Where θ is a parameter between 0 and 1

4. State Variable Transfer

In recent years mesh-free methods has been developed by employing new approximation
theories such as moving least square approximation (MSL) [1], reproducing kernel
approximation (RKPM) [8] that allow the shape functions to be constructed entirely in terms of
arbitrarily placed nodes without the used of an element connectivity in the classical sense. This
makes the direct transfer of information between the subsequent meshes with a desirable
consistency possible, especially in the direct transfer of state variable at integration point during
the adaptive procedure as shown in Figure 2. In this work, a local interpolant with a desirable
consistency condition, which is based on the modified meshfree appxiomation, is constructed to
directly obtain the state variables at new integration points from old integration points. The
numerical diffusion near the boundary, especially in areas undergoing rapid transients, such as
the contact boundary part needs to be limited during state variable transfer. A special treatment
with using the smoothness of meshfree approximation is adopted to minimize the effect of the

17-38
9th International LS-DYNA Users Conference Metal Forming (2)

numerical diffusion. The state variables at nodal and integration points for the new mesh can be
interpolated from the old mesh as
f ( x ) = ∑Ψˆ I ( x ) f I (26)
I =1

Where ψˆ ( x ) is the modified meshfree approximation functions [3, 4, 7]

As a test problem we take the following equation

d du
− ( )+u = g on x ∈ (0,1) (27)
dx dx

with the boundary conditions u (0) = u (1) = 0 . By choosing function g such that the exact
solution of the problem reads

sinh 4 x
u = x2 − (28)
sinh 4

As shown Figures 3 and 4, the recovery solution has the same accuracy and convergence rate as
the EFG numerical solution.

5. Numerical Examples
Two forging examples are analyzed by using the proposed adaptive EFG method. The adaptive
procedure is triggered by a given time interval, and the grid distribution is controlled by the
specific maximum and minimum grid space. Error indicator for adaptive EFG method will be
discussed in the other paper.

5.1 Wheel Forging

The wheel forging problem as shown in Figure 5 (a) is first tested by using the Lagrangian EFG
method. It can not finish the job because the mesh is too coarse to capture the local behavior in
areas undergoing rapid transient. The problem is analyzed by employing the adaptive procedure
with increasing nodes from 5827 to 13661. The distribution of effective plastic strain in the final
deformation is plotted in Figure 5 (b). During the adaptive procedure, the monotonic property of
the internal variables such as the effective plastic strain can be maintained. This is due to the
linear consistency of the interpolation function during the state variable transfer. The volume and
the reaction force are given in Figures 6 and 7 respectively for EFG method and adaptive EFG
method. The adaptive EFG result is comparable with the EFG result.

5.2 Forging of Piston Housing

The piston housing as shown in Figure 8 (a) is simulated in this example. The final deformation
with material flash is given in Figure 8 (b). There are two important issues, suck-in defect Figure
9 (a) and material flash Figure 9 (b), in this simulation. To successfully capture suck-in defect in
the adaptive simulation, a fine mesh (or a local fine mesh) and an accurate state variable transfer

17-39
Metal Forming (2) 9th International LS-DYNA Users Conference

Gauss points in the old mesh Gauss points in the new mesh

Figure 2 direct state variable transfer between two subsequent meshes

EFG solution
Recovery solution
Analytic solution
solution

Figure 3 The comparison of numerical, recovery and analytic solutions

-1.5

-2

-2.5
Log (Error norm)

EFG solution

-3 Interpolation solution

-3.5

Slope 1.68
-4

-4.5
-3 -2.5 -2 -1.5 -1 -0.5 0
Log (Nodal space)

Figure 4 The convergence rate in the first derivative of solution

17-40
9th International LS-DYNA Users Conference Metal Forming (2)

(a) Problem description (b) Contour plot of plastic strain

Figure 5 Problem statement and contour plot of plastic strain

Figure 6 Volume change in the simulation

Figure 7 Reaction force

17-41
Metal Forming (2) 9th International LS-DYNA Users Conference

(a) Problem description (b) Final deformation

Figure 8 Problem statement and final deformation

(a) Suck-in defect (b) Material flash

Figure 9 Suck-in defect and material flashes

Figure 10 Contour plot of the effective plastic strain

17-42
9th International LS-DYNA Users Conference Metal Forming (2)

Figure 11 Contour plot of the effective plastic strain

Figure 12 Reaction force

Figure 13 Volume change

17-43
Metal Forming (2) 9th International LS-DYNA Users Conference

are required. The proposed modified meshfree interpolation can effectively transfer the state
variables between subsequent meshes without losing the high order information. The improved
EFG method can handle the material flash effectively. Figures 10 and 11 give the distribution of
the effective plastic strain at different simulation time respectively. The reaction force and
volume change during the simulation are plotted in Figures 12 and 13 respectively.

Acknowledgements
We would like to thank Dr. John Watton of Alcoa Technical Center for providing the test
examples and helpful suggestions.

References
[1] T. Belytschko, Y.Y. Lu and L. Gu, “Element-free Galerkin methods,” Int. J. Numer. Methods Engrg. 37 (1994)
229-256.
[2] B. Boroomand, and O.C. Zienkiewicz, “Recovery Procedures in Error Estimation and Adaptivity. Part II:
Adaptivity in Nonlinear Problems of Elasto-Plasticity Behaviour”, Computer Methods in Applied Mechanics and
Engineering, Vol. 176, 127-146, 1999.
[3] J. S. Chen, C. Pan, C. T. Wu and W. K. Liu, “Reproducing Kernel Particle Methods to Large Deformation
Analysis of Nonlinear Structures,” Computer Methods in Applied Mechanics and Engineering, Vol. 139, p. 195-
227, 1996.
[4] Chen, J. S., Han, W., You, Y., Meng, X"A Reproducing Kernel Method with Nodal Interpolation Property," Int.l
Journal for Numerical Methods in Engineering, Vol. 56, (2003) 935-960.
[5] N. S. Lee and K. J. Bathe, Error Indicators and Adaptive Remeshing in Large Deformation Finite Element
Analysis, Finite Elem. Anal. Des. 16 (1994) 99-139
[6] S. Li and W. K. Liu, “Meshfree and Particle Methods and Their Applications,” Applied Mechanics Review, Vol.
55, pages 1-34, 2002.
[7] W. K. Liu, W. Han, H. Lu and S. Li, Reproducing Kernel Element Method, Part I Theoretical Formulation,
Comp. Methods in Applied Mechanics and Engineering, 193: 933–951, 2004.
[8] W. K. Liu, S. Jun, and Y. F. Zhang, “Reproducing Kernel Particle Method”, Int. J. Numer. Methods Fluids, 20
(1995) 1081-1106.
[9] H. Lu, and J. S. Chen, “Adaptive Meshfree Particle Method,” Lecture Notes in Computational Science and
Engineering, Vol. 26, pp. 251-267, 2002.
[10] M. Ortiz and J.J. Quigley, IV, “Adaptive Mesh Refinement in Strain Localization Problems”, Comp. Methods
in Applied Mechanics and Engineering, 90, 781–804, 1991.
[11] D. Peric, Ch. Hochard, M. Dutko, and D.R.J. Owen, “Transfer Operators for Evolving Meshes in Small Strain
Elasto-Plasticity”, Comp. Methods in Appl. Meh. Engrg., 137, 331–344, 1996.
[12] J. –P. Ponthot, T. Belytschko, “Arbitrary Lagrangian-Eulerian Formulation for Element-Free Galerkin
Method”, Comp. Meths. in Applied Mechanics and Engineering, 152, 19–46, 1998
[13] T. Rabczuk, T. Belytschko, “Adaptivity for structured meshfree particle methods in 2D and 3D,” International
Journal for Numerical Methods in Engineering, Vol. 63, p. 1559-1582, 2005.
[14] C.A. Duarte, and J.T. Oden, “An hp adaptive method using clouds”, Comp. Methods in Appl. Meh. Engrg.,
139, 237-262, 1996.
[15] J.M. Melenk, and I. Babuska, “The Partition of Unity Method”, Comp. Methods in Appl. Meh. Engrg., 139,
289-314, 1996.

17-44