You are on page 1of 18

Error Estimation and Adaptivity in Explicit

Nonlinear Finite Element Simulation of Quasi-Static Problems


Kjell M. Mathisen, Odd S. Hopperstad, Knut M. Okstad and Torodd Berstad
Department of Structural Engineering
Norwegian University of Science and Technology
N{7034 Trondheim, Norway

Abstract

ing towards automatic adaptive solution of nonlinear


problems is the ever decreasing cost of computation
together with the recent developments of more robust
error estimators.
The objective of mesh adaptivity is to obtain a FE
mesh which is optimal in the sense that the computational costs involved are minimal under the constraint
that the error in the solution is below a certain prescribed limit. Since the computational e ort can be
linked to the number of unknowns in the FE mesh
the task is to nd a mesh with a minimum number
of unknowns or nodes for a given error tolerance. In
general, adaptive methods rely on error indicators and
error estimators which can be computed a priori or a
posteriori. Two main types of a posteriori estimates
exist today:

A program module for error estimation with application to nonlinear nite element (FE) analysis of shell
structures is coupled with the adaptive solution procedure in the explicit FE code LS-DYNA. The error
estimation module provides estimates of the local and
global errors and element-level re nement indicators.
Hence, selective re nement of the mesh in areas where
the local error is relatively large compared with a userde ned tolerance is made possible. Furthermore, the
relative global error is estimated giving a measure of
the overall accuracy of the FE model. Projection-type
error estimators based on the L2 -norm of the stress
vector and the accumulated plastic strain are used to
predict the discretization error by comparison of the
FE solution with an improved C 0-continuous solution
obtained by the SPR-method.
Three example problems including both material and
geometric nonlinearities are provided. The numerical
results show that the error estimates capture phenomena such as di use necking and local buckling, and give
meshes with high resolution in areas with large deformations or high stress gradients.

1. Residual based estimators in which the error can


be estimated by some norm of the residual of
the governing equations, see Babuska and Rheinboldt 1] and Babuska et al. 2].
2. Recovery based estimators in which the error is
estimated by taking the di erence between a recovered (post-processed higher order) solution and
the FE solution, see Zienkiewicz and Zhu 3, 4, 5].

1 Introduction

Based on the computed errors a new re ned mesh can


be constructed which yields a better approximate solution. To obtain an optimal mesh it is desirable to
design the mesh such that the error contributions of
the elements are equidistributed over the mesh.
Both the residual based and the recovery based error estimators are frequently used estimators for elastic problems. According to Rheinboldt 6] and Stein
et al. 7] the introduced error estimators for linear elastic problems may be extended into the geometrically
nonlinear regime if they are applied to the linearized
problem. When considering explicit nonlinear nite
element analysis in general, theoretical proofs concerning the convergence that were established in the elastic range are no longer valid. Since we are considering
quasi-static problems only, a rst step in order to justify the use of recovery-based Zienkiewicz{Zhu error

Today, industry problems in engineering are associated with complex geometries, nonlinear constitutive
behavior, contacts, large deformations, etc. For such
problems only numerical methods, like the FE method,
yield solutions of the associated mathematical model.
Since numerical methods yield approximate solutions
only, it is necessary to control the errors inherent in the
method. During the last two decades much research
has been devoted to development of error estimators
and adaptive FE strategies for solving linear problems.
When considering nonlinear problems adaptive solution strategies become even more appealing, particularly for problems where the optimal mesh changes
continuously throughout the analysis, since the CPUtime consumption when solving for such problems may
become prohibitively large. Another reason for mov1

estimators 4] is to neglect the errors due to the discretization in time.


In prior studies, an overview of key aspects which
have to be considered in an adaptive nonlinear solution
procedure has been presented along with the implementation of an h-adaptive mesh re nement technique
into a general purpose implicit non-linear FE program
for solving geometrically non-linear shell type problems involving linear elastic materials 8, 9, 10]. In the
case of history-dependent materials, there exist only a
few contributions on adaptive strategies for historydependent non-linear problems. Johnson et al. 11]
have developed an adaptive strategy for small strain
elasto-plasticity using the Hencky model. However,
due to restrictions in the Hencky model, this method is
generally not applicable to standard elastic-plastic and
elastic-viscoplastic history-dependent problems. Other
contributions involving history-dependent materials include the work by Cheng and Kikuchi 12], Ortiz and
Quigley 13], Lee and Bathe 14], Peric et al. 15, 16],
Wriggers and Scherf 17], Gallimard et al. 18] and
Barthold et al. 19, 20].
In this paper, the objective is to present e cient (although not theoretically justi ed) techniques for controlling the discretization error in quasi-static simulations by explicit nonlinear nite element calculations.
It is demonstrated that Zienkiewicz{Zhu type error estimates can be used to provide accurate resolution of
the current critical zone being formed.

are written in rate form they must be integrated


in time with an appropriate integration scheme.
The choice of the integration scheme performed at
every step (explicit, implicit) is fundamental for
the accuracy.
In this study, only estimation of spatial discretization
error is considered. Two major questions should be
answered while analyzing a solution obtained with a
given FE mesh:
1. What is the precision of the computed solution
in global (structural) and local (elemental) sense,
i.e., how big is the error?
2. How to obtain a solution within a prescribed accuracy at a minimum computational cost?
Error estimation and adaptive mesh re nement give
the answer to the two questions above. The aim is to
estimate the error of the FE method, using already
computed data. A posteriori estimates employ the
solution obtained by the FE analysis in addition to
some a priori assumptions about the solution. Without an a posteriori error estimate there is no reliable
way of judging the quality of the solution. If the analyst's intuition fails when designing the FE mesh, it
is also liable that it fails when evaluating the precision
of the computed solution. Therefore the philosophy of
an adaptive process is that the computer (code) decides, on the basis where to re ne the mesh and does
so automatically.
There are two major tasks, closely tied together, that
should be realized in order to obtain the answers to the
two preceding questions:
1. Establish a posteriori local and global error estimates which accurately and reliably characterize
the error in the FE solution according to some
speci ed norm.
2. Design an adaptive procedure for automatic mesh
re nement in order to achieve a solution with a
prescribed accuracy in an e ective and economic
manner.
The rst task is addressed below, while the latter is
considered in the next Section.

2 Error Estimation
2.1 Preliminaries

Among the various sources of error that can be identied in explicit FE simulation of quasi-static problems,
the following are of great importance:
1. The discretization error : Spatial discretization errors are inevitably connected with the application
of the FE method, because of the discrete representation of the computational domain. In that
way any FE mesh is associated with this type of
error. For the particular case of linear FE analysis of a regular problem the solution converges
towards the exact one and the error decreases to
zero when the element size decreases.
2. Error in the choice of the constitutive model : It
gives the di erence between the real behavior of
the material and the one represented by the constitutive law. In addition we may have errors in
the numerical values of the chosen constitutive law
parameters.
3. Error due to numerical integration of the constitutive law : This type of error is characteristic for
material nonlinear analysis. As material equations

2.2 Error norms

The FE solution, in terms of the displacements uh and


stresses h , di er from the from the corresponding exact values, u and . The di erence de nes the pointwise errors
(1)
eu = u ; uh
h
e =
;
(2)
These point-wise de nition of errors are generally di cult to specify, and various integral measures are more
2

conveniently adopted. One of the most common of such These two features of the FE solution may constitute
measures is the energy norm
useful indication to estimate the discretization error.
In order to construct error estimators based on these
1 12
0Z
indications we must recover some higher order solution.
(3) A simple and very e cient a posteriori error estikekE = @ eT C;1 e dV A
mator for practical engineering analysis is proposed by
A more direct measure is the so-called L2 -norm, which Zienkiewicz and Zhu 3] and thus frequently referred
can be associated with the error in any quantity. Thus, to as the ZZ-estimator. The basic idea consists in the
for the displacements and stresses, respectively, the er- use of a continuous stress eld, , which eliminates
the stress jump across the element interfaces (see Figrors in L2 -norm are
ure 1b), and is (at least) one order more accurate than
0Z
1 12
the FE stress eld, h . In the ZZ-estimator the pointT
@
A
(4) wise error (2) is approximated by the di erence bekeukL2 =
eu eu dV
tween the higher order solution and the FE solution,
0Z
1 12
viz.
e =
; h
(8)
ke kL2 = @ eT e dV A
(5)
By substituting the estimate e into (3), we thus obtain
the ZZ-error estimator in global energy norm
The latter expression di ers from the energy norm only
0Z
112
by a weighting matrix C;1 .
;
;
Although the norms (3){(5) here are de ned on the ke kE = @
; h T C;1 ; h dV A (9)
whole domain, , we note that the square of each one
can be obtained by summing element contributions, i.e.
Obviously, the quality and reliability of the ZZnel
X
2
2
estimator
depend on the accuracy of the recovered sokek = keke
(6)
lution
.
Zienkiewicz and Zhu 21, 22] show that if
e=1
the
recovered
stress eld is superconvergent, the ZZwhere keke represents the contribution from element e
error
estimator
will always be asymptotically exact in
and is obtained by integrating over the elemental subenergy
norm.
domain, e , instead of .
Although the value of the energy-norm error does
have a physical meaning (indeed, it measures the ab- 2.4 Error estimators for nonlinear materials
solute error in strain energy), it may be easier to inter- Two error estimators of ZZ-type are considered herein.
pret relative quantities, especially for L2-norms. For Each estimator measure the discontinuity across the
this purpose we adopt the relative percentage error
elements in some quantity that is derived from the apFE solution. In the present study, the folk
e
k
= kuk 100 %
(7) proximate
lowing criteria have been used in the initial selection of
error estimators for comparative evaluation.
where kuk denotes the corresponding norm of the ex- potential
These
criteria
state that the error estimator must be:
act solution of the problem. The percentage error (7)
can be evaluated for the whole domain and for each a) Valid for the material models considered.
elemental sub-domain.
b) Valid for large displacements (and rotations) but
small strains.
2.3 Zienkiewicz{Zhu error estimation
Usually, the FE solution is less smooth than the exact c) Easy to compute (i.e., with additional necessary
computations kept to a minimum).
one. In other words, it has some de ciency:
The degree of the interpolation polynomials of the d) Based on results available from the analysis code
(LS-DYNA).
elements may be insu cient to represent the exact
solution which may be a higher order polynomial In the literature several possible error estimates have
or some other function.
been considered. Tetambe et al. 23] tested several L2 The derivatives of the FE solution (strains and norms the L2 -norms in stresses, total strains, equivstresses) are discontinuous across the element in- alent total strain and total strain rate, together with
terfaces (see Figure 1a), as C 0 interpolation func- the rate energy norm. Further Peric, Yu and Owen 15]
tions are usually employed for the FE approxima- tested error estimates based on incremental plastic
tion of displacements, while the derivatives of the work, plastic dissipation and the energy norm. With
LS-DYNA we are limited to using error estimates based
exact solution generally are continuous.
3

Inter-element discontinuity

a)

b)
Figure 1: Stress eld on a four-element patch. a) discontinuous FE eld, b) recovered C 0-continuous eld.
of the FE method. Hinton and Campbell 24] pointed
out that the FE method minimizes the error in stresses
in the sense of weighted least squares. This implies
that the FE stresses oscillate within the element and
at some special locations, the superconvergent points of
the stress eld, the convergence rate of the FE stresses
is O(hp+1 ), which exceeds the general theoretical global
rate, O(hp ).
The rst demonstration of superconvergence is due
to Barlow 25]. It is also discussed by Hinton 24] and
later by Barlow 26] for distorted elements. For quadrilateral nite elements, the superconvergent points for
stresses and strains coincide with the reduced Gauss
quadrature points, also referred to as Barlow points.
Linear elements have thus one superconvergent point
each located at the centroid (see Figure 2). The Barlow points are generally the best sampling points for
stresses whereas the element nodes are the worst, as
the derivatives of the shape functions tend to behave
poorly near the border of the interpolation region.
Let vh denote some piece-wise continuous eld, i.e.,
vh 62 C 0( ) but vh j e 2 C 0 ( e), which is computed by
the analysis code and that we want to recover an improved version of. The corresponding recovered eld,
which we denote v , is expressed as a polynomial
v = Pa
(13)

on either rate form quantities, such as strain rate or


plastic strain rate, or an estimate based on stresses
or accumulated plastic strain. Therefore, we focus on
the L2 -norm errors in stresses and accumulated plastic
strain. These are also found to be good error estimates
by the authors mentioned above.
Adopting the ZZ-approach, the point-wise errors in
accumulated plastic strain, p, may be approximated by
= p ; ph
(10)
We then de ne the following two error estimators:
1. L2 -norm error in stresses
ep

0Z
; ;
ke kL = @
2

2.

L2 -norm error

; ;

1
dV A (11)
1
2

in accumulated plastic strain

0Z
1
;
2
kep kL = @ p ; ph dV A
2

1
2

(12)

2.5 Superconvergent Patch Recovery

The ZZ-error estimator was originally used in combination with global least-squares t methods and direct
nodal averaging for obtaining the higher order eld.
More recently, a local least-squares t approach, popularly denoted superconvergent patch recovery (SPR)
has proven to be more e cient and to give more accurate results. Herein, we use the SPR-approach in its
original form as proposed by Zienkiewicz and Zhu 4],
but extended to shell problems.
The SPR-procedure utilizes the concept of superconvergent points. These are certain locations within each
element where the computed FE solution has higher accuracy than elsewhere. The rate of convergence is also
higher in these points than at other locations. The superconvergence property is known from the early days

a)

b)

Figure 2: Examples of nodal patches. ( ) nodal points


( ) the patch node (node de ning the nodal patch) ( )
result sampling points (superconvergent points).
4

a)

b)

c)

Figure 3: Nodal patches associated with a node ( ) on the exterior boundary. a) Node connected to exterior
element boundaries only. b) Node connected to one interior element boundary. c) Node connected to two
interior element boundaries.
where P is a matrix of polynomial terms and a is a
vector of unknown coe cients. The polynomial expansion (13) should (at least) be of the same order as the
basis functions used to interpolate the primary variables in the FE problem. Using rst order elements, P
should therefore contain the bi-linear monomials, i.e.

2
P1
P = 4 ...

Pn

An important issue regarding implementation of the


SPR-methods is the local coordinate system in which
the polynomials for v are expressed, as the conditioning of the resulting equation system may be poor for
some choices of coordinate system. This problem is
addressed by Kvamsdal and Okstad 27], and the procedure proposed therein is also used in this study.

3
5 Pi = 1 x y xy ] (14)

2.6 Global recovered eld

with n denoting the dimension of the vector v .


The polynomial (13) is de ned over small patches
of elements surrounding each nodal point. The coefcients a are then determined from a least squares t
of the eld v to the values of vh at the superconvergent points within the elements in the patch, i.e., we
minimize the following functional with respect to a

When result recovery is used for error estimation, we


are normally interested in recovered values at elementinterior points (integration points) rather than nodal
values. Since a speci c element belongs to more than
one patch, the patch recovery does not provide a unique
result value at such points. In order to construct a
global recovered eld, Blacker and Belytschko 28] proposed to conjoin the polynomialexpansions (13), for all
n
sp
X; h
T; h
F (a) =
v i ; v (xi ) vi ; v (xi ) (15) patches containing the actual element using the nodal
shape functions as weighting functions. Adopting this
i=1
approach, the recovered eld within a bi-linear element
where xi is the coordinates of the i'th superconvergent e is evaluated through
point, vhi is the computed value of vh at that point,
4
and nsp denotes the total number of superconvergent
X
ve (x) =
N a (x) va (x)
(16)
points in the patch. The minimization of (15) leads
a=1
to a small linear system of equations which is assembled and solved on each patch of elements. Typical
examples of such patches are shown in Figure 2. Note where N a (x) denotes the bi-linear shape function asthat patches are established for interior nodes only, as sociated with element node a and va (x) is a local renodes on the exterior boundary rarely are connected to covered eld on the form (13). If the element node a is
enough elements1 . Neither are patches established for an interior node, va (x) is evaluated on the patch of elnodes along shell intersections, since some stress com- ements surrounding that node. For nodes lying on the
ponents should not be continuous across such intersec- exterior boundary, va (x) is instead evaluated on the
tions, e.g., when two perpendicular shell surfaces meet patch (or patches) associated with the other node(s)
the normal stress in one shell surface is transferred into that are connected to node a through an interior element boundary, see Figure 3. If more than one patch
a transverse shear stress in the other surface.
1 With rst order elements containing one superconvergent in this manner are connected to a boundary node (Figpoint each, there must be at least 3 elements in the patch in ure 3c), the corresponding values for va computed on
order to be able to recover a linear eld over the patch.
each patch are averaged.
5

tk

= tk;1 +

De ne initial state

= 0, t = 0, t0 = tbirth

) 0 , r0, s0

?
t = t+ t
?HH
H
HHH t > ttotal HH Yes - Exit
HH
?No
Time integration
) t , rt, st

?HH
H
t ; tk = t HH
HH No
and
HH
HHHt tdeath
HH
?Yes

k=k+1

= tk;1 +

=k+1

Map solution variables


rt and st from the old
to the new mesh

6
t = tk

6
HYesHH

Estimate the global error, ,


and calculate local re nement
indicators, e , at time t

?H
HHH No
H
max
HHH
HH
?HYesH
H
No
HHHTwo-pass ? HH
HH
Yes
?
t=t

tk

HHHTwo-pass ? HH No
HH
6
Re ne the mesh

Figure 4: Schematic overview of the adaptive solution procedure in LS-DYNA.

3 Adaptivity within LS-DYNA

In an adaptive simulation the user speci es a birth


time, tbirth , at which the adaptive remeshing begins
the death time, tdeath , at which the adaptive remeshing ends and a time interval, t, between each check
for adaptive re nements. In addition, the maximum
number of re nement levels, nmax the minimum element size, hmin and of course the prescribed error
tolerance, max , are speci ed. Explicit FE analyses are
usually characterized by very small (but many) time
steps, t. However, the computational cost for each
step is modest when comparing with an implicit anal-

3.1 Adaptive solution procedure

Figure 4 gives an algorithmic overview of the built-in


adaptive solution procedure in LS-DYNA 29]. Here,
t is the load parameter at time t, rt denotes the primary unknowns of the FE problem (e.g., nodal displacements) at time t, whereas st represents the other
solution-dependent state variables stresses, strains,
etc. The total duration of a simulation is denoted ttotal
whereas the time step is denoted t.
6

mesh is re ned only one level per adaptive step, regardless of the value on e . The only two values of e
that we are able to satisfy exactly is therefore 1:0 (no
sub-division) and 2:0 (sub-division). Herein, we choose
an in-between solution and subdivide all elements for
which e 1:5.

ysis. It is therefore not desirable to check for adaptive


remeshing at every time step, and typically t t.
LS-DYNA o ers two types of adaptivity a one-pass
strategy and a two-pass strategy. In the latter variant,
the solution procedure jumps back to the time instant
tk (see Figure 4) at which the previous error control was
performed, after the adaptive step is completed. Note
that this backward jump is performed also if the estimated error is within the tolerance and not only when
the mesh has been re ned. Thus, the two-pass strategy
ends up calculating every time step between tbirth and
tdeath twice and is therefore considerably more expensive than the one-pass strategy.

4 Numerical Studies
To demonstrate the use of error estimation in adaptive
FE plasticity analysis we consider three elastic-plastic
example problems a tensile test, stretch bending of
a square hollow beam, and axial crushing of a thinwalled square pro le. All simulations are performed
with LS-DYNA using the Belytschko{Lin{Tsay shell
element 35] with single-point in-plane integration, ve
integration points through the thickness and sti nessbased hourglass control. The material is modeled as
an isotropic elastic-plastic material, adopting the von
Mises yield criterion, an isotropic strain hardening rule,
and the associated ow rule in the plastic domain. The
strain hardening of the material is de ned by

3.2 Adaptive re nement strategy

The error estimation and calculation of element-wise


re nement indicators is herein performed by a separate program module 30, 31] in which the error estimation and recovery procedures described in the previous
section have been implemented. The basic principle
for adaptive mesh re nement is to obtain a mesh with
equidistribution of the estimated error. Such a mesh is
`optimal' as it provides results of a given accuracy with
a minimum number of degrees of freedom. In the socalled h-methods the estimated error is used to predict
the size of the elements in the re ned mesh through
the expression
hnew
= hold
(17)
e
e = e
old
where he denotes the characteristic size of element
e in the existing (old) mesh, and hnew
is the desired
e
size of the new elements in the region covered by that
old element. The mesh re nement parameter, e , is
computed based on the theoretical convergence rate of
the error and some mesh optimality criterion.
Herein, we adopt the strategy presented in 32, 33],
where one tries to minimize the number of elements
in the optimal mesh with the constraint that the total
error should be equal to the prescribed tolerance. For
smooth problems this optimality criterion is equivalent
to claiming equally distribution of the error over the
elements. The resulting expression for the re nement
indicator reads
e

1=(q+1)
1=q
max

n
X
el

e=1

2=(q+1)

!1=2q

Y0

+ Q1 (1 ; exp(;C1 p))
+ Q2 (1 ; exp(;C2 p))

(19)

where Y is the von Mises equivalent stress, p denotes


the accumulated plastic strain, and Y0, Qk and Ck are
material constants.
For each test we carry out one reference simulation
using a ne uniform mesh, and three adaptive simulations with characteristics as given in Table 1. The
computational cost of the adaptive simulations with
respect to the reference simulation is given in Table 2
for the three problems. Tables 3 and 4 summarize the
adaptivity control parameters and the material parameters, respectively, that are used for each example problem. The prescribed tolerances p max and max that
are used to guide the adaptive simulations, are here
selected based on the estimated relative global errors
of the corresponding reference simulation.
Table 1: Adaptive simulations.

(18)

Simulation
Error
Adaptive
No.
Estimator Strategy
1
One-pass
p
2
Two-pass
p
3
Two-pass

where q denotes the theoretical convergence rate of the


elements and e is the contribution from element e
to thepestimated
Pnel 2. relative global relative error, , i.e.,
=
e=1 e
The adaptive mesh re nement, based on the computed re nement indicators e , is performed by pure
subdivision of the existing elements into four new elements introducing hanging nodes between the re nedand un-re ned elements. Details on the mesh re nement procedure may by found in 34]. Note that the

Table 2: Normalized CPU-time of the simulations.


Problem
Ref. No. 1 No. 2 No. 3
Tensile test
1.0 0.89 2.04 0.87
Stretch bending test 1.0 0.96 1.61 3.27
Axial crushing test
1.0 0.66 0.85 1.13

Table 3: Control parameters for adaptivity.

Problem

ttotal

s]
Tensile test
35
Stretch bending test 0.008
Axial crushing test
20
Problem

tbirth

s]
0
0
0

tdeath

s]

s]
0.5
0.0001
0.4

1
1
1

nmax

4
4
4

hmin

p max

mm]
0
0
0

max

%]
10
10
20

%]
15
10
25

Table 4: Material constants used in the simulations.


E

Y0

MPa]
Tensile test
72000 0.33
Stretch bending test 72000 0.33
Axial crushing test 70000 0.33

MPa]
300
300
75

Q1

C1

MPa]
300 8
300 8
75.6 25

Q2

C2

MPa]
0
0
0
0
200 1.9

200
70

65

65
25

25
0,02 A
2.7
1

12.5

40

R15
A

Figure 5: The tensile test: Geometry of the test specimen.


LS-DYNA is an explicit dynamic FE code using the
conditionally stable central di erence method for time
integration of the discrete equations of motions. In
general, stability requires time steps in the order of
micro seconds, and simulation of quasi-static problems a)
is practically di cult. One possible way to overcome
this problem is to increase the loading rate in the simulation provided that inertia and rate e ects still are
negligible in the response. Another method is to scale
up the mass to increase the stable time step. The rst
method is adopted for the stretch bending simulation
whereas mass scaling is used in the simulations of the b)
tensile test and the axial crushing test.
Figure 6: The tensile test: a) Fine mesh used in the
reference simulation. b) Initial mesh of the adaptive
4.1 Tensile Test
simulations.
Figure 5 shows the geometry of the test specimen. In
the simulations, the relative velocity of the bolts is pre- Figure 7 shows the deformed geometry at four di erscribed, increasing the velocity smoothly from zero to ent displacement levels for the simulation using the refa constant level in order to avoid spuriously high fre- erence mesh (Figure 6a) and for the three adaptive simquency oscillations in the solution. Figure 6 depicts the ulations. The force{displacement curves for the four
FE mesh used in the reference simulation, and the ini- simulations are compared in Figure 8.
tial mesh of the adaptive simulations. The two meshes The estimated relative global errors in the reference
contain 1316 and 256 elements, respectively.
solution is plotted against the displacement level in
8

a)

b)

c)

d)

Figure 7: The tensile test: Deformed geometry at displacement levels 12.5, 16.5, 20.5 and 24.5 mm]. a)
Reference solution. b) Adaptive simulation No. 1. c) Adaptive simulation No. 2. d) Adaptive simulation No. 3.
Figure 9. Both error estimates are less than 10% in the
plastic domain prior to necking, but is signi cantly
higher than p . The relative global error increases ( p
more strongly than ) as strain localization occurs in
the specimen's gauge section. Hence indicating that
even local errors in the solution may be detected by
studying the evolution of the relative global error.
The peak in p , occurring when the specimen is
mainly in the elastic domain, is due to the normalization procedure. The relative global error is the sum
over all elements of the ratio of the estimated error
to the norm of the FE (or improved) solution. In the
start of the deformation process the accumulated plastic strain is small or even zero, and the result may be
large values of the relative error, even if the absolute
error is small.

Figure 8 shows that all simulations give the same


force{displacement curve up to necking. In the descending part of the force{displacement curve, the reference solution and adaptive simulation No. 2 compare
well, while adaptive simulations No. 1 and 3 predict
higher load levels.
The element distribution (Figure 7) and the total
number of elements (Figure 10) during the simulation
di er signi cantly between the di erent schemes. In all
cases, the element mesh is re ned in the area around
the loading point, where the stress gradients are large
and the stress state is multi-axial. Both error estimates
capture the necking of the specimen and resolve the
high gradients in the localization zone when two-pass
adaptivity (No. 2 and 3) is used { this is not the case
for one-pass adaptivity (No. 1). For the adopted values
9

1600

30

1200

Number of elements

Force (kN)

40

20
Ref.
10

No. 1

800

400

No. 1
No. 2

No. 2

No. 3

No. 3

0
0

10

15

Displacement (mm)

20

10

15

Displacement (mm)

20

25

Figure 10: The tensile test: Number of elements in the


Figure 8: The tensile test: Force{displacement curves. mesh during the adaptive simulations.
50

Global relative error (%)

40

30

20

10

0
0

10

15

Displacement (mm)

20

Figure 9: The tensile test: Estimated relative global


errors.
of max and p max , the error estimator p gives the
highest mesh density and CPU time, see Table 2.

4.2 Stretch Bending Test


Figure 11 shows a simpli ed model of the experimental stretch bending set-up presented by Welo et al. 36].
The model consists of a rigid bending die with radius
50 mm, and an extruded aluminum square hollow section with nominal wall thickness 2 mm, width 20 mm
and length 440 mm. The model is symmetric about

the plane x = 220 mm] and about the plane y = 0.


The pro le is supported along the y-axis where the displacements in the x-, y- and z -directions and the rotations about the x- and z -axes are xed, whereas the
rotation about the y-axis is free. The pro le is exible with elastic-plastic material properties between the
plane x = 135 mm] and the plane x = 220 mm]. Elsewhere it is rigid. The pro le is loaded by gradually
moving the bending die downwards. The maximum
displacement in the z -direction is 60 mm. Figure 12
shows the FE mesh used in the reference simulation
and the initial mesh of the adaptive simulations. Symmetry is taken into account in the model and only one
quarter of the bending die and the extruded pro le is
discretized. The reference model consists of 1420 shell
elements, whereas the initial mesh of the adaptive simulations consists of 384 shell elements. The default
version of the surface-to-surface contact algorithm in
LS-DYNA is used in the simulations. Friction between
the bending die and the pro le is neglected.
As for the tensile test, one reference simulation and
three di erent adaptive simulations with varying error
estimator and adaptive strategy are conducted. The
results are presented in Figures 13{16, respectively,
as ange sagging against die displacement, deformed
mesh at 60 mm die displacement, estimated relative
global errors against die displacement (reference simulation only) and nally the total number of elements
against die displacement.
The ange sagging is de ned (see Figure 11) as the
displacement of point B relative to point A after stretch
bending. Figure 13 shows that the sagging developments predicted in the reference solution and the adaptive solutions are in good agreement. Figure 15 shows

10

C
L

C
L
2

50

z
x

18

A, B
135

85
A

220

PLANE

CROSS-SECTION

Figure 11: The stretch bending test: Geometry, boundary and loading conditions.
2.5
Ref.
No. 1

2.0

Sagging (mm)

a)

No. 2
No. 3

1.5

1.0

0.5

b)
Figure 12: The stretch bending test: a) Fine mesh used
in reference simulation. b) Initial mesh of the adaptive
simulations.
that the relative global errors are less than 10% in the
major part of the forming operation, and in this case
the two error estimates give relatively consistent values
of the relative global error. The element distribution
(Figure 14) and the total number of elements during
the simulation (Figure 16) are almost the same in the
one-pass and two-pass adaptive simulations using the
error estimator p , i.e., adaptive simulations Nos. 1
and 2. The required number of elements is signi cantly
higher in adaptive simulation No. 3 in which is used,
and this is also re ected by a signi cant increase in the
computational cost as shown in Table 2.

0.0
0

20

40

Displacement (mm)

Figure 13: The stretch bending test: Flange sagging


against die displacement.

4.3 Axial Crushing Test


Figure 17 shows the specimen geometry, support and
loading conditions in the axial crushing test of thinwalled square aluminum extrusions 37, 38]. The FE
meshes used in the reference and adaptive simulations
are depicted in Figure 18. The specimen is clamped
at the lower end, free at the top, and loaded quasistatically in compression. The load and the relative
displacement between the base and the crosshead are

11

a)

b)

c)

d)

Figure 14: The stretch bending test: Deformed geometry at 60 mm die displacement. a) Reference solution. b)
Adaptive simulation No. 1. c) Adaptive simulation No. 2. d) Adaptive simulation No. 3.
12

16

Relative global error (%)

12

0
0

20

40

Displacement (mm)

Figure 15: The stretch bending test: Estimated relative global errors for the reference simulation.
1000

Number of elements

800

600

No. 1

400

No. 2
No. 3
200
0

20

40

Displacement (mm)

60

Figure 16: The stretch bending test: Number of elements in the mesh during the adaptive simulations.
measured.
Owing to symmetry, only one quarter of the test setup is modeled. The specimen is modeled using shell
elements. A rigid body modeled by brick elements is
used to impose the load at the free end of the pro le.
The rigid body is given a constant velocity during the
deformation. The specimen is modeled using respectively 240 and 2400 shell elements in the coarse and
ne model. Initial imperfections are introduced near
the top section of the model to trigger a symmetric
buckling mode, see Figure 18.

The results from the reference solution are given in


Figures 19 and 21 where the absorbed energy and the
relative global errors are plotted against the displacement of the rigid body. In this case, the relative global
errors di er signi cantly: the error estimators p and
give relative global errors of approximately 12% and
20{25%, respectively.
The results from the three adaptive simulations are
presented in Figures 19, 20 and 22, in terms of absorbed energy against displacement, deformed mesh at
200 mm displacement, and number of elements against
displacement, respectively.
The energy absorption and the deformation mode
are similar in the reference solution and the adaptive
solutions. The required numbers of elements are signi cantly higher in adaptive simulations Nos. 1 and 3
then in No. 2. Hence, in this case the CPU time (see
Table 2) increases by about 30% only, when going from
one pass to two-pass adaptivity, using the error estimator based on plastic strain.

5 Concluding Remarks
A program module for error estimation of nonlinear
FE shell analysis was coupled with the adaptive solution procedure in the explicit FE code LS-DYNA.
Automatic re nement of the mesh in areas where the
element level error exceeds a user-de ned tolerance is
thus made possible. The discretization error is predicted by comparing the FE solution with an improved
C0-continuous solution obtained by the superconvergent patch recovery (SPR) method. Error estimators
based on the L2-norm of the stress vector and the accumulated plastic strain are adopted. The adaptive
procedure in LS-DYNA is based on mesh enrichment
where a single `parent' quadrilateral element is split
into four equally sized `sibling' elements. The data
transfer between the old deformed mesh and the new
mesh is trivially solved by letting the sibling elements
inherit the state of the parent element. One-pass and
two-pass adaptivity are available. One-pass adaptivity
means that the analysis is restarted with the enriched
mesh from the time step of the current error estimation.
In two-pass adaptivity, the analysis is always restarted
from the time step of the previous error estimation,
even if the mesh remains unaltered. In practice, this
means that the simulation is run through twice.
Example problems involving material and geometric nonlinearities and quasi-static loading conditions
were solved with LS-DYNA. It was found that both
the stress- and strain-based error estimators produced
meshes with high resolution in areas with high stress
gradients or large deformations. It is however di cult to compare the accuracy and e ectivity of the
two error estimators since they predict di erent relative global errors. In addition, the mutual relationship

13

80

Jack
80

Fs

} h = 1.8 - 2.0 - 2.5

AA6060-T4
AA6060-T6

Geometry of test specimens

310

3
50

Static

Section

Plan

Loading and support conditions

Figure 17: The axial crushing test: Geometry, boundary and loading conditions.

Energy (Nm)

6000

4000

Ref.

2000

No. 1
No. 2
No. 3
0
0

40

80

120

Displacement (mm)

160

200

Figure 19: The axial crushing test: Absorbed energy


versus displacement.
between relative global errors may change during the
simulation. One-pass adaptivity is obviously more effective than two-pass adaptivity, but its robustness is
b)
a)
less since abrupt changes in the response (i.e., due to
geometric or material instabilities) between two error
Figure 18: The axial crushing test: a) Fine mesh used checks may be missed. However, a modi ed adaptive
in the reference simulation. b) Initial mesh of the adap- procedure in which the analysis is restarted from the
tive simulations.
time step of the previous error check only when the
14

a)

b)

c)

d)

Figure 20: The axial crushing test: Deformed geometry at 200 mm displacement. a) Reference solution. b)
Adaptive simulation No. 1. c) Adaptive simulation No. 2. d) Adaptive simulation No. 3.
15

ered in this paper, a relative global error between 10


and 20 percent seems to be a good choice.

Relative global error (%)

25

20

Acknowledgments

15

This work was supported nancially by Hydro Raufoss


Automotive Research Centre as. The support is highly
acknowledged.

References

10

1] I. Babuska and W. C. Rheinboldt. A{posteriori


error estimates for the nite element method. In-

ternational Journal for Numerical Methods in Engineering, 12:1597{1615, 1978.

0
0

40

80

120

Displacement (mm)

160

20

Figure 21: The axial crushing test: Estimated relative


global errors for the reference simulation.
4000

3] O. C. Zienkiewicz and J. Z. Zhu. A simple error estimator and adaptive procedures for practical engineering analysis. International Journal for
Numerical Methods in Engineering, 24:337{357,
1987.

No. 1
No. 2

Number of elements

3000

2] I. Babuska, T. Strouboulis, and C. S. Upadhyay.


A model study of the quality of a posteriori error
estimators for linear elliptic problems. Error estimation in the interior of patchwise uniform grids
of triangles. Computer Methods in Applied Mechanics and Engineering, 114:307{378, 1994.

No. 3

4] O. C. Zienkiewicz and J. Z. Zhu. The superconvergent patch recovery and a posteriori error estimates. Part 1: The recovery technique. Inter-

2000

national Journal for Numerical Methods in Engineering, 33:1331{1364, 1992.

1000

5] O. C. Zienkiewicz and J. Z. Zhu. The superconvergent patch recovery and a posteriori error estimates. Part 2: Error estimates and adaptivity.

0
0

40

80

120

Displacement (mm)

160

200

Figure 22: The axial crushing test: Number of elements


in the mesh during the adaptive simulations.
current error check leads to mesh enrichment is readily
implemented. Such a procedure will always give the
same accuracy and robustness as two-pass adaptivity
and in many cases be almost as e ective as one-pass
adaptivity. The main problem inherent in the adaptive simulation is the user-de ned element-level error
tolerance. It is not easy to specify a reasonable value
for this tolerance without any a priori knowledge of the
problem at hand. Such knowledge may be achieved by
running a non-adaptive reference solution using a relatively coarse mesh. However, in the problems consid16

International Journal for Numerical Methods in


Engineering, 33:1365{1382, 1992.

6] W. C. Rheinboldt. Error estimates for nonlinear nite element computations. Computers and Structures, 20:91{98, 1985.
7] E. Stein, C. Carstensen, B. Seifert, and S. Ohnimus. Adaptive nite element analysis of geometrically nonlinear plates and shells, especially buckling. International Journal for Numerical Methods
in Engineering, 37:2631{2655, 1994.
8] K. M. Okstad and K. M. Mathisen. Towards
automatic adaptive geometrically nonlinear shell
analysis. Part I: Implementation of an h {adaptive
mesh re nement procedure. International Journal
for Numerical Methods in Engineering, 37:2657{
2678, 1994.

9] K. M. Okstad. Adaptive Methods for Nonlin- 20] F.-J. Barthold, M. Schmidt, and E. Stein. Error Indicators and Mesh Re nements for Finiteear Finite Element Analysis of Shell Structures.
Element-Computations of Elastoplastic DeformaDr. Ing. dissertation, Department of Structural
tions. In Proc. of abstracts 10th Nordic SemiEngineering, The Norwegian Institute of Technolnar on Computational Mechanics, pages 12{37,
ogy, Trondheim, Norway, 1994.
Tallinn, Estonia, October 1997.
10] K. M. Mathisen and K. M. Okstad. Interactive{
Adaptive Geometrically Nonlinear Analysis of 21] J. Z. Zhu and O. C. Zienkiewicz. SuperconverShell Structures. Engineering with Computers,
gence recovery technique and a posteriori error
12:63{83, 1996.
estimators. International Journal for Numerical
Methods in Engineering, 30:1321{1339, 1990.
11] C. Johnson and P. Hansbo. Adaptive nite element methods in computational mechanics. Com- 22] O. C. Zienkiewicz and J. Z. Zhu. Adaptivity
puter Methods in Applied Mechanics and Engiand mesh generation. International Journal for
neering, 101:143{181, 1992.
Numerical Methods in Engineering, 32:783{810,
1991.
12] J. H. Cheng and N. Kikuchi. A mesh rezoning technique for nite element simulations of 23] R. P. Tetambe, S. M. Yunus, C. Rajakumar, and
metal forming processes. International Journal
S. Saigal. Examination of Flux Projection-Type
for Numerical Methods in Engineering, 23:219{
Error Estimators in Nonlinear Finite Element
228, 1986.
Analysis. Computers and Structures, 54:641{653,
13] M. Ortiz and J. J. Quigley. Adaptive Mesh Re ne1995.
ment in Strain Localization Problems. Computer
Methods in Applied Mechanics and Engineering, 24] E. Hinton and J. S. Campbell. Local and global
smoothing of discontinuous nite element func90:781{804, 1991.
tions using a least squares method. International
14] N. S. Lee and K.-J. Bathe. Error indicators and
Journal for Numerical Methods in Engineering,
adaptive remeshing in large deformation nite el8:461{480, 1974.
ement analysis. Finite Elements in Analysis and
Design, 16:99{139, 1994.
25] J. Barlow. Optimal stress locations in nite element models. International Journal for Numerical
15] D. Peric, J. Yu, and D. R. J. Owen. On error estiMethods in Engineering, 10:243{251, 1976.
mates and adaptivity in elastoplastic solids: Applications to the numerical simulation of strain lo- 26] J. Barlow. More on Optimal Stress Points, Recalization in classical and Cosserat continua. Induced Integration, Element Distortions and Error
ternational Journal for Numerical Methods in EnEstimation. International Journal for Numerical
gineering, 37:1351{1379, 1994.
Methods in Engineering, 28:1487{1504, 1989.
16] D. Peric, Ch. Hochard, M. Dutko, and D. R. J. 27] T. Kvamsdal and K. M. Okstad. Error Estimation
Owen. Transfer Operators for Evolving Meshes in
based on Superconvergent Patch Recovery using
SmallStrain Elasto{Plasticity. Computer Methods
Statically Admissible Stress Fields. International
in Applied Mechanics and Engineering, 137:331{
Journal for Numerical Methods in Engineering,
344, 1996.
42:443{472, 1998.
17] P. Wriggers and O. Scherf. Adaptivity Finite Element Methods for Contact Problems in Plastic- 28] T. Blacker and T. Belytschko. Superconvergent
patch recovery with equilibrium and conjoint inity. In Proceedings of the 4th International Conterpolant enhancements. International Journal
ference on Computational Plasticity, pages 787{
for Numerical Methods in Engineering, 37:517{
807, Barcelona, Spain, March 1995.
536, 1994.
18] L. Gallimard, P. Ladeveze, and J. P. Pelle. Error
Estimation and Adaptivity in Elastoplasticity. In- 29] Livermore Software Technology Corporation. LSDYNA Keywords User's Manuals. Version 940,
ternational Journal for Numerical Methods in En1997.
gineering, 39:189{217, 1996.
19] F.-J. Barthold and E. Stein. Error Estimation and 30] K. M. Okstad and T. Kvamsdal. AFEM | AdapMesh Adaptivity for Elasto-Plastic Deformations.
tive Finite Element Module, Maintenance Manual.
In Proceedings of the 5th International ConferTechnical report, SINTEF Applied Mathematics,
ence on Computational Plasticity, pages 597{603,
Computational Engineering, Trondheim, Norway,
Barcelona, Spain, March 1997.
March 1997.
17

31] O. S. Hopperstad, K. M. Mathisen, K. M. Okstad, T. Berstad, and I. Tiller. Error Estimation


in Plastic Forming Simulations. Technical Report
R-21-97, Department of Structural Engineering,
Norwegian University of Science and Technology,
Trondheim, Norway, May 1997.
32] P. Ladeveze, J.-P. Pelle, and Ph. Rougeot. Error estimation and mesh optimization for classical
nite elements. Engineering Computations, 8:69{
80, 1991.
33] P. Coorevits, P. Ladeveze, and J.-P. Pelle. An
automatic procedure with a control of accuracy for
nite element analysis in 2D elasticity. Computer
Methods in Applied Mechanics and Engineering,
121:91{120, 1995.
34] T. Belytschko, B. L. Wong, and E. J. Plaskacz.
Fission{fusion adaptivity in nite elements for
nonlinear dynamics of shells. Computers and
Structures, 33:1307{1323, 1989.
35] T. Belytschko, J. I. Lin, and C. S. Tsay. Explicit
algorithms for the nonlinear dynamics of shells.
Computer Methods in Applied Mechanics and Engineering, 42:225{251, 1984.

36] T. Welo, F. Paulsen, and K. E. Moen. On the Use


of Numerical Simulation for Predicting Bendability and Geometric Tolerances of AA7xxx Extrusions Formed in Stretch Bending. In Proceedings of
Formability'94, pages 24{27, Ostrava, Czech Republic, October 1994.
37] M. Langseth and O. S. Hopperstad. Static and
Dynamic Axial Crushing of Square Thin-Walled
Aluminium Extrusions. International Journal of
Impact Engineering, 18:949{968, 1996.
38] T. Berstad. Material Modelling of Aluminium for
Crashworthiness Analysis. Dr. Ing. dissertation,
Department of Structural Engineering, Norwegian
University of Science and Technology, Trondheim,
Norway, 1996.

18

You might also like