ECCM-2001

European Conference on
Computational Mechanics
June 26-29, 2001
Cracow, Poland

PROGRESS ON PRACTICAL METHODS OF ERROR ESTIMATION

FOR ENGINEERING CALCULATIONS
J. Tinsley Oden and Serge Prudhomme,
The Texas Institute for Computational and Applied Mathamatics
The University of Texas at Austin, Austin, TX 78712, U.S.A.
email: [email protected], [email protected]

Tim Westermann and Jon Bass

Altair Engineering Inc.
Austin, Texas, U.S.A.
email: [email protected], [email protected]

Key words: Goal-oriented error estimation, Displacement error, Error in stress, Upper and
lower bounds, Dual problem, Residual method.
Abstract. In the present paper, the goal-oriented error estimation method is applied to several
complex engineering analysis of three-dimensional elastic solids. In the examples considered,
the errors are estimated with respect to displacement or stress components at a given point in
the solid. The numerical results indicate that the method can be used effectively for complex
engineering applications.

J.T. Oden, S. Prudhomme, T. Westermann, J. Bass

1 Introduction
The use of a posteriori error estimation as a means for verification of accuracies of computer
simulations has been a topic of research for nearly two decades. Until recently, such error estimates have been global in nature, giving bounds on global approximation errors in energy-type
norms. Extensions of global estimation methods to procedures for estimating errors in local
quantities of interest has been described in recent papers [1, 2]. However, applications of these
new techniques have been confined to very simple model problems involving scalar-valued
functions on one- and two-dimensional domains.
In the present investigation, we review the theory of goal-oriented error estimation and discuss
applications to several complex engineering analysis of three-dimensional elastic solids. These
examples indicate that the goal-oriented methods can be used effectively in quite complex engineering applications.

2 Model problem and approximation error

Let be a bounded open domain in
 , 
 , with Lipschitz continuous boundary .
The region is occupied by a linearly elastic material body in static equilibrium under the


 and surface tractions "!  , where "!$#% .
action of body forces 
The portion of the boundary on which displacements are prescribed is denoted "& , such that
"&(' )!* , and
354 "&(+("!,.-  . The stress / , related to the displacement 0 by the constitutive
equation /1%2
7680:9680<; , satisfies inside the linearly elastic the momentum equations:

= 6?>@/1AB

in

DC

(1)

Here is the elasticity tensor, which is assumed to satisfy the standard symmetry and uniform ellipticity conditions. Moreover, the displacement and stress are subjected to the boundary
conditions:
E
0F.GH
on "&I
C
(2)
/J>LKFMB
on "!
Let N be a positive integer and O an open region in
 ( O may stand for the whole domain ,

for a subdomain of or for a portion of the boundary). Denote by PRQS7O the Sobolev space of


functions in 7O with distributional derivatives of order TN in 7O as well. The space of
admissible displacements U is therefore defined as:

UVXWZY[,P 2 

\ Z]

YF.G

on "&I^_

(3)

the boundary values being understood in the sense of traces of P 2 -functions. The classical
variational form associated with the elasticity problem consists in finding the displacement 0*
U such that:
`


0 \Y MaRbY
cBY[dUe
(4)
2

ECCM-2001, Cracow, Poland

where

0 Y

aRbY




fHgh6iY

4j3

680lkHm"

f ghn>LYokp9 f
q@r l>LYlksN


The bilinear form t>u L> is symmetric positive-definite on

`
w wyx
so-called energy norm,

|YI Y

{z

(5)

U vU

C

and induces a norm, the

(6)

x a partition of , i.e. }~ is a finite collection of open elements i , F

y }CLCL~ C  denote
. The partition is assumed to satisfy <u_ , and +o5- whenever
is different from . The size of an element is measured in terms of its diameter. We
associate with the partition the parameter .8 I_ .
In order to approximate the solution 0 , one constructs a finite element space of hierarchical
Let

shape functions such as


~
W@8P 2  ] MG

on

and "


%  2 _
RS( cd}~j^

(7)

where is the affine mapping from the master element to the element in the partition,

and B 
is the space of polynomial functions defined on of degree at most . We
employ here hierarchical shape functions based on the integrated Legendre polynomials [3, 4].
These shape functions can be classified into nodal, edge, face and interior bubble functions in
three dimensions. We associate with the partition }~ the parameter n.u I . Then, the
~L , or, for the sake of simplicity in the notation, U ~ , is defined as:
finite element space U

C

U ~  ~  #JU

Using the classical Galerkin method, a finite element approximation

`
the solution of

(8)

0)~U ~

C
cBY[dU ~
The numerical error in the approximation shall be defined as iM0 = 0)~ ,
dU

of



0"~ \Y M
 aRbY

is given as
(9)

3 Goal-oriented error estimates

Rather than estimating the numerical error in the energy norm, we aim here at evaluating the
error with respect to a quantity of interest. The quantity of interest represents the goal of the
computational analysis, and is supposed to be defined mathematically as a bounded linear functional on the space of trial displacements U . Many quantities of interest actually fall into that

category. Let  denote such a functional. Therefore, our goal is to predict the quantity 0 .
The numerical error in which we are now interested in is given by the quantity R
, which,
by linearity of  , reduces to:




yC
M0 = 0)~ .0 = 0)~ b
3

(10)

J.T. Oden, S. Prudhomme, T. Westermann, J. Bass

We briefly explain how to estimate  . The reader is referred to [2, 5] for more details.

The starting point is to find a function which relates the quantity  to the residual
obtained from the momentum equation. This function, the so-called influence function, is shown
to be the solution of the dual problem:

and is approximated in U

bY \


bY

cBY*RUe

by the dinite element solution ~RU



bY \~ 
 bY

C
cBY*RU ~

(11)
of the discrete dual problem:
(12)

As for the error in the solution, we introduce the error in the approximate influence function as
M = ~ . Then follows the remarkable result

wx
w
wx

 w

=
=
(13)
 bs 
 n7NLs yN 2 V
 2 NLD9N 2  2 NL N 2 

which establishes the relation between the error quantity  and the energy norms of linear
combinations of and . In Equation (13), N defines a scaling factor in
. Several methods to


obtain bounds on errors in the energy norm are now available in the literature. In the following
numerical experiments, we shall use a simple method described in [6].

In order to derive an estimate of and bounds on b , we suppose that the following global error

estimates are available:

w
wyx

2 yw x T upp
(14)
C
2 T
low T
upp
Then, it is straightforward to show that the quantities l ow and u pp , defined as:





 C
=
=

2  u pp 

2  low

(15)
u pp V2  upp 
l ow 2  low 

produce respectively a lower and an upper bounds on b , i.e.

C

T1b T upp

(16)
low


An estimate of  can be derived by taking the average of l ow and u pp , that is:

yC
  low
9 upp

(17)
est

w wyx
w
wyx

Finally,
w wythe
x value
w ofwyx N is chosen so that the quantities NZ and N 2 have same amplitudes,
 N 2 , which implies that:
i.e. NZ
w wyx
wyx C
w
(18)
Nh

w
wx
w
wx

=
N  and NL
N .
Such a choice of N is justified because it minimizes the quantities NL9

low

w LN 9N
NL = N

ECCM-2001, Cracow, Poland

HyperSolve
beam_hex_grf

04/26/01 22:46:52

Altair Engineering

P-Levels: 1 2 3 4 5 6 7 8

Figure 1: Mesh topology and boundary conditions for cantelever beam.

4 Numerical experiments
The main purpose of the present paper is to provide an account on the performance of goaloriented error estimation for three-dimensional elasticity problems. Our estimates and error
bounds are verified by comparing the results with a refined and/or enriched mesh which is
believed to provide a nearly exact solution.
We consider here as quantities of interest pointwise displacements and stresses. By poinwise
quantities, we actually mean to take local averages using mollifying kernels defined in the form:


 

GH


 2 _t =

if 
otherwise

(19)


g | m Hk m: .
It follows that the quantity of interest corresponding to the displacement at a point h
will
The constant

corresponds to a normalizing factor and is chosen such that

be represented by the linear functional:


 
SC

b  f g y = < | k
m:b

(20)

Likewise, for the stress  at a given point , we shall consider the quantity of interest:



yC

b  f g y = ZkHmb

(21)

The parameter can be prescribed by the user and varies according to the problem and quantity
of interest. In most cases, is chosen of the order of the mesh size .
4.1 The cantilever beam

In the first example, we consider a cantilever beam with dimension  GH  subjected to
a uniform load. The beam is fixed at JG . The mesh topology and boundary conditions are
C
shown in Fig. 1. The material properties are set to GjG GPa and eG  where E is the
Youngs modulus and the Poissons ratio.
5

J.T. Oden, S. Prudhomme, T. Westermann, J. Bass

Mesh
Elements

R
v

v vR
Z vFv v 

v v

Estimated error
Estimate Lower Upper
0.2350 0.2012 0.2688
0.1219 0.0950 0.1487
0.0364 0.0217 0.0516
0.0072 0.0017 0.0126

Exact error


 = s~
0.3871
0.1706
0.0542
0.0154

Table 1: Error estimates and bounds for the quantity of interest (displacement
C C C 
 G G
G G in the cantilever beam).
Influence function for W in x-direction

HyperSolve

beam_hex_grf

04/26/01 23:09:02

Altair Engineering

Influence function for W in y-direction

HyperSolve

beam_hex_grf

04/26/01 23:09:50

at


Altair Engineering

1.6e-06

5.0e-05

4.0e-05
1.2e-06
3.0e-05
8.0e-07
2.0e-05
4.0e-07
1.0e-05

0.0e+00

0.0e+00

-1.0e-05
-4.0e-07
-2.0e-05
-8.0e-07
-3.0e-05

-4.0e-05

-1.2e-06
X

-5.0e-05

-1.6e-06

Figure 2: Influence function corresponding to the displacement at ?

in m -direction (right) in -direction.
4.1.1

G C HG G C C G 

(left)

Displacement at a point

We estimate here the error in the displacement (in direction ) at the free end-point 
C C C 
 G G
G G . We select a sequence of meshes which are uniformly refined by dividing each
element into two new elements in one or several directions. Estimates and bounds on the error
in the displacement are shown in Table 1.
The exact errors are computed here by substracting the discrete displacement from the dis
placement obtained on the refined mesh v
spectral order 2. The displacement at
 G C G
G C C G  on the mesh of reference reads G C v j . with
We observe in Table 1 that the bounds and
estimates, although very accurate, all slightly underestimate the exact errors. We attribute this
behavior to the fact that the global estimates of the solution and influence function in the energy
norm use edge bubble functions only and hence are rather coarse. Finally, we show in Fig. 2 the
m - and -component of the corresponding influence function.
4.1.2

Stress at a point

C 

In the next example, we estimate the error in the stress component  at  GH G y G .

This point is located exactly at the center of the upper face of the beam, as it can be observed
from the shape of the influence function shown in Fig. 3. As before we collect in Table 2 the
estimates of the errors in the quantity of interest. This time, the estimates are rather accurate, but

ECCM-2001, Cracow, Poland

Mesh


vv vF
Z vFivFv 

 v v

Estimated error
Exact error

Estimate Lower Upper
 = B~
-308
-1444
828
-535
-169
-2179 1840
-200
-52
-2476 2371
-57
-17
-1672 1637
-20

Table 2: Error estimates and bounds for the quantity of interest (Stress  at <5
in the cantilever beam).
Influence function for stress Sxx in x-direction

HyperSolve

beam_hex_grf

04/26/01 22:51:01

Altair Engineering

Influence function for stress Sxx in z-direction

HyperSolve

beam_hex_grf

04/26/01 22:56:30

C C 
GH G y  G
Altair Engineering

0.35

0.2

0.30

0.1
0.0

0.25

-0.1
0.20
-0.2
0.15
-0.3
0.10
-0.4
0.05
-0.5
0.00
-0.6
-0.05

-0.7

-0.10

-0.15

-0.8
-0.9

Figure 3: Influence function corresponding to the stress

m -direction (right) in -direction.

at

C C C 
  HG G   G

(left) in

the bounds have the tendency to diverge from the exact error. This behavior was also observed
in [2] for two-dimensional problems.
We finally show in Fig. 4 the distribution of the elementwise contributions to the error in the
displacement and stress component to be used as refinement indicators. We clearly observe the
error in the displacement at the free-end is sensitive to the inaccuracy of the solution at the
fixed-end, as this can be expected. On the other hand, we notice that the error in the stress has a
more local origin.
4.2 The hoop problem
The hoop problem is concerned with the simulation of a portion of a tire under a specific
loading, as shown in Fig. 5. The blue arrows show the direction in which the hoop is held fixed
while the orange arrows indicate a uniformly distributed load. The elements of the mesh are of
order 2 along the surface of the tire and linear in the transversal direction and the provided mesh
consists of 200 elements.

Z H

Here, we analyze the stress component  at *  j and show the errors for a
sequence of meshes. In this case, as shown in Table 3, the bounds for the quantity of interest

J.T. Oden, S. Prudhomme, T. Westermann, J. Bass

Contributions to the error in qoi for W

HyperSolve

beam_hex_grf

04/26/01 23:05:11

Altair Engineering

Contributions to the error in qoi for Sxx

HyperSolve

beam_hex_grf

04/26/01 23:00:20

Altair Engineering

3.5e-04

3.0e-04

4
2.5e-04

3
2.0e-04

1.5e-04
1.0e-04

5.0e-05

0.0e+00

-1

-5.0e-05

-2
-1.0e-04

-3
-1.5e-04

-4

-2.0e-04
Y

-2.5e-04

-3.0e-04

-5
-6

Figure 4: Distribution of the contributions to the error in the quantity of interest (left) for the

C C C 
C C C 
displacement at 5 G GH G G (right) for the stress at <? GH G y G .
Mesh and Loads

HyperSolve

Hoop hex model

04/27/01 12:45:46

Altair Engineering

3068
3089
3088
3087

XZ

Figure 5: Mesh topology and boundary conditions for the hoop problem.
are more accurate as compared to the beam problem. We also show in Fig. 6 the corresponding
influence function and the distribution of the contributions to the error in the stress component.
4.3 The control arm
In the final example, we consider a control arm whose geometry and loading are shown in
Fig. 7. Once again, we are interested here in the stress component  whose distribution we
show in Fig. 7 as well. The goal is to estimate the error in the computed stress at location

( GH y G on the given mesh. We found the following estimate and bounds:


C
 A
est
 j C
low
  = GC  G
A
upp

to be compared with the exact error in the stress   G obtained by using a refined mesh with higher spectral order. We also show a plot of the influence function and the

ECCM-2001, Cracow, Poland

Mesh
Elements
Initial mesh
Refine in (zeta)
Refine in (xi,zeta)

Estimated error
Exact error
Estimate Lower Upper

(1,2)
0.303
-3.066 3.673
1.526
(1,2)
0.307
-3.149 3.763
1.421
(1,2)
0.211
-4.125 4.547
0.579

Table 3: Error estimates and bounds for the quantity of interest (Stress  at 
for the hoop problem).
Influence function for stress Sxx in x-direction

HyperSolve

Hoop hex model

04/27/01 13:02:22

Altair Engineering

Contributions to the error in qoi for the stress

HyperSolve

Hoop hex model

04/27/01 12:47:55

Altair Engineering

3.5

3.0

2.5

2.0

1.5

1.0

0.5

0.0

-0.5

-1.0
X
Z

Z H   

-1.5

-2.0

XZ

-1

-2

Figure 6: Distribution of the contributions to the error in the quantity of interest (left) for the


Z

displacement at 5
H  (right) for the stress at <?  .
distribution of the error in the stress component in Fig. 8. On the latter, we observe that the contributions are not just local and that remote locations do contribute as well. An adapted mesh
based on these elementwise contributions would then control the error in the quantity of interest
by taking care of all contributing sources, near and remote.

5 Conclusions
In this paper, we have applied the goal-oriented error estimation technique to the case of threedimensional linear elasticity problems. We have concentrated on estimating pointwise errors in
displacement and stress components. The results obtained on the various examples shown here
are very promising. In all cases, the error was predicted with very good accuracy. We would like
to point out that relatively simple estimates of the error in the energy norm were used in those
numerical experiments, which, to our own surprise, provided us with very reasonable estimates
and bounds on the error.
Acknowledgements This research was supported by a contract with the General Motors Research and Development Center. This support and the encouragement of Dr. Mark Botkin of
GM are greatfully acknowledged.

J.T. Oden, S. Prudhomme, T. Westermann, J. Bass

Control arm

HyperSolve

Control arm model

05/02/01 12:24:35

Altair Engineering

Stress Sxx

HyperSolve

Control arm model

05/02/01 12:31:51

4000

250

3500

200

Altair Engineering

150

3000

100
2500
50
2000
0
1500
-50
1000
-100
500

-150

-200
Z

P-Levels: 1-500
2 3 4 5 6 7 8

-250

Figure 7: The control arm problem: (left) Geometry and loading (right) Distribution of the stress
component  .
Influence function

HyperSolve

Control arm model

05/02/01 12:28:03

Altair Engineering

Error contribution for the quantity of interest

HyperSolve

Control arm model

05/02/01 12:41:46

0.4

Altair Engineering

4000

0.2
3500
0.0
3000

-0.2
-0.4

2500
-0.6
2000

-0.8
-1.0

1500

-1.2
1000
-1.4
-1.6

500

-1.8
0
-2.0

Z
X

X
-2.2

Y
-500

Figure 8: The control arm problem: (left) Influence function in m -direction corresponding to the

pointwise stress Zm at 5 GH y G (right) Distribution of the error in the stress component

Lm at 5 GH y G .

References
[1] S. Prudhomme and J.T. Oden. On goal-oriented error estimation for elliptic problems:
Application to the control of pointwise errors. Comp. Meth. in Appl. Mech. and Eng., 176,
313331, (1999).
[2] J.T. Oden and S. Prudhomme. Goal-oriented error estimation and adaptivity for the finite
element method. Computers Math. Applic., 41(56), 735756, (2001).
[3] B. Szabo and I. Babuska. Finite Element Analysis. John Wiley & Sons, (1991).
[4] J.T. Oden. Optimal h-p finite element methods. Comp. Meth. in Appl. Mech. and Eng.,
112, 309331, (1994).

10

ECCM-2001, Cracow, Poland

[5] M. Ainsworth and J.T. Oden. A Posteriori Error Estimation in Finite Element Analysis.
John Wiley & Sons, (2000).
[6] S. Prudhomme and J.T. Oden. Simple techniques to improve the reliability of a posteriori
error estimates for finite element approximations. ECCN 2001, Cracow, Poland (2001).

11