Computers and Structures 152 (2015) 1–10

Contents lists available at ScienceDirect

Computers and Structures

journal homepage: www.elsevier.com/locate/compstruc

Locally equilibrated stress recovery for goal oriented error estimation in

the extended finite element method
O.A. González-Estrada a,c,⇑, J.J. Ródenas b, S.P.A. Bordas a, E. Nadal b, P. Kerfriden a, F.J. Fuenmayor b
a
Institute of Mechanics and Advanced Materials (IMAM), Cardiff School of Engineering, Cardiff University, Queen’s Buildings, The Parade, Cardiff CF24 3AA, Wales, UK
b
Centro de Investigación de Tecnología de Vehículos (CITV), Universitat Politècnica de València, E-46022 Valencia, Spain
c
School of Mechanical Engineering, Universidad Industrial de Santander, Bucaramanga, Colombia

a r t i c l e i n f o a b s t r a c t

Article history: Goal oriented error estimation and adaptive procedures are essential for the accurate and efficient eval-
Received 11 December 2013 uation of finite element numerical simulations that involve complex domains. By locally improving the
Accepted 14 January 2015 approximation quality, for example, by using the extended finite element method (XFEM), we can solve
expensive problems which could result intractable otherwise. Here, we present an error estimation tech-
nique for enriched finite element approximations that is based on an equilibrated recovery technique,
Keywords: which considers the stress intensity factor as the quantity of interest. The locally equilibrated supercon-
Goal oriented
vergent patch recovery is used to obtain enhanced stress fields for the primal and dual problems defined
Error estimation
Recovery
to evaluate the error estimate.
Quantities of interest Ó 2015 Elsevier Ltd. All rights reserved.
Error control
Mesh adaptivity

1. Introduction
In continuum mechanics, stresses and strains are usually the
main quantities to describe the behaviour of a component under
certain loads. However, when the component is affected by a crack
these parameters are not sufficient to properly describe the beha-
viour of the component in the Linear Elasticity (LE) framework.
For instance, under the LE assumptions, the stress field at the crack
tip will take infinite values, and in the surroundings of the crack tip,
due to the high stress value, the small deformations assumption
does not hold. Linear Elastic Fracture Mechanics (LEFM) assump-
tions are considered valid for brittle fracture [1,2]. One characteris-
ing parameter used to give a more realistic description of the
behaviour around the crack tip is the stress intensity factor (SIF),
which focuses in the local stress state at the crack tip [2] and can
be considered an energy-based quantity. Hence, to properly
describe the behaviour at the crack tip in LEFM is necessary to accu-
rately evaluate the SIF. It results interesting to evaluate an error
measure for the SIF to be able to control its level of accuracy [3,4].
Since the beginning of the use of numerical simulations many
methods have been developed to control the discretisation error
of finite element approximations, mostly based on the evaluation
⇑ Corresponding author at: School of Mechanical Engineering, Universidad
Industrial de Santander, Bucaramanga, Colombia.
E-mail address: estradaoag@cardiff.ac.uk (O.A. González-Estrada).
of global error in energy norm of the Finite Element (FE) solution.
These methods can be broadly classified in residual based [5],
recovery based [6] and dual analysis [7,8]. The numerical results
in [9–11] showed that a recovery technique with a standard super-
convergent patch recovery (SPR) [12], applied in problems with
smooth solution, was more robust than the residual estimates con-
sidered. However, a more interesting approach is to control the
error in a particular quantity relevant for the design process [13–
16]. This quantity could be defined as a bounded functional that
describes the displacement or stresses in a given area of the
domain, or for the case of fracture mechanics, the SIF that charac-
terises the crack. This approach, referred to as goal oriented, is usu-
ally based on the use of duality techniques that involve the
formulation of an adjoint or dual problem directly related to the
quantity of interest (QoI). Residual methods have been frequently
used to evaluate the error in quantities of interest although exam-
ples involving recovery techniques can be found in [17,3], and con-
sidering dual analysis in [18]. In [19] an enhanced version of the
SPR technique was used to obtain accurate estimations of the error
in different QoI in the context of linear elasticity problems solved
with the FEM. In [3], recovery and residual based estimates of
the error in evaluating the J-integral (directly related to the SIF)
for finite element (FE) approximations in the context of LEFM were
presented. The authors showed that the most accurate error esti-
mates for the J-integral were obtained with the recovery based
estimators.

http://dx.doi.org/10.1016/j.compstruc.2015.01.015
0045-7949/Ó 2015 Elsevier Ltd. All rights reserved.
2 O.A. González-Estrada et al. / Computers and Structures 152 (2015) 1–10

Refs. [16,15] showed that the error in the quantity of interest Neumann and Dirichlet conditions are applied, and CC to the free
can be expressed in terms of errors in energy norm, and that if the- traction surface describing a crack such that @ X ¼ CN [ CD [ CC
se errors in energy norm can be bounded we could also bound the and CN \ CD \ CC ¼ ;. We denote by b the body loads, t the trac-
error in the quantity of interest. On the other hand, it is usually dif- tions imposed along CN and r0 ; e0 the initial stresses and strains.
ficult to obtain guaranteed error bounds of the quantities of inter- The displacement field u is the solution of the problem given by
est while maintaining the accuracy of the estimate. The need of
such a bound is also arguable in an engineering context as the
LT r þ b ¼ 0 in X; ð1Þ
reliability of an a posteriori error estimate, which is quantified by Gr ¼ t on CN ; ð2Þ
its local effectivity, can be verified beforehand on a number of Gr ¼ 0 on CC ; ð3Þ
practical cases. Here, we are interested in increasing the effectivity u¼0 on CD ; ð4Þ
of the error estimate used to guide adaptive algorithms rather than
eðuÞ ¼ Lu in X; ð5Þ
error bounding.
In the context of LEFM, the extended finite element method r ¼ DðeðuÞ  e0 Þ þ r0 in X; ð6Þ
(XFEM) [20] has been successfully used to enrich the finite element where L is the differential operator for linear elasticity, and G is the
approximation in order to represent the particular features of projection operator that projects the stress field into tractions over
cracks, namely, the discontinuity along the crack faces and the sin- any boundary, with n the outward unit normal to CN , such that
gularity at the crack tip. This method helps to overcome some of    
the difficulties when modelling crack propagation, such as the @=@x 0 @=@y nx 0 ny
LT ¼ ; G¼ ; ð7Þ
need for remeshing to obtain conforming meshes to the crack 0 @=@y @=@x 0 ny nx
topology. Recent advanced numerical approaches for the purpose
D is the matrix of the linear constitutive relation for stress and
of XFEM and fracture analysis also include Mixed Discrete Least
strain. We consider an homogeneous Dirichlet boundary condition
Squares Meshless (MDLSM) [21], and XIGA [22]. Error estimators
in (4) for simplicity.
in energy norm for XFEM and other partition of unity methods
The problem expressed in its variational form is written as:
have been proposed in [23–26] using recovery techniques, and in
2
[27,4,28] using the residual approach. A goal oriented approach Find u 2 V such that 8v 2 V ¼ fv j v 2 ½H1 ðXÞ ; v jCD ¼ 0g :
for enriched finite element approximations based on the constitu- Z Z Z Z
tive relation error has been presented in [29]. In [30] goal oriented eðuÞT Deðv ÞdX ¼ v T bdX þ vT tdC þ eðv ÞT De0 dX
X X CN X
error estimators based on the explicit residual method were intro- Z
duce for the XFEM framework. In [31], adaptive techniques based  e ðv Þr0 dX:
T
ð8Þ
on energy norm and goal oriented error estimation have been X
investigated for enriched finite element approximations. To evaluate the SIF, noted as K, it is common practice to use the
In this paper, we propose a goal oriented error estimation tech- interaction integral in its Equivalent Domain Integral (EDI) form.
nique for XFEM approximations that is based on the enhanced There are different expressions already available to evaluate EDI
recovery technique previously presented in [25,26] and the consid- integrals for singular problems. In this work, we are going to con-
eration of the SIF, typical of LEFM, as the quantity of interest. One sider the method based on extraction functions, as shown in [33]:
of the key features of the recovery-based error estimators is that 2 3
Z uaux
x q;x
" #
the solution is recovered patch-wise in a basis richer than the raux
xx q;x þ rxy q;y
aux
1
one used for the FE approximation. As shown in [23,32], when K¼ rT 6
4 uaux
y q;y
7
5  uT dX;
C XI rxy q;x þ raux
aux
yy q;y
XFEM is used, the basis used for the recovery should include the uaux aux
y q;x þ ux q;y
singular terms, which is not common in standard recovery tech-
ð9Þ
niques. Therefore, error estimates in quantities of interest will also
aux aux
require a careful consideration of the singular character of the where u ; r are the auxiliary fields used to extract the SIFs in
XFEM solution, and the use of extended recovery approaches mode I or mode II and C is a constant that is dependent on the geo-
becomes a necessity to obtain accurate estimates. To improve the metry and the loading mode. q is an arbitrary C 0 function that
quality of the recovered stresses for the primal and dual problems, defines the extraction zone XI which takes the value of 1 at the sin-
and therefore, the accuracy of the error estimate, we consider equi- gular point and 0 at the boundary C. q;x and q;y are the derivatives of
librium constraints locally in patches of elements and the splitting the function q with respect to x and y.
of the recovered stress field into singular and smooth parts, which is
the fundamental idea in the recovery process to describe the singu- 2.1. Discrete problem using XFEM
lar behaviour of the solution.
The paper is organised as follows. In Section 2, we introduce the Let us consider a finite element approximation of u denoted as
problem under consideration and its corresponding enriched uh . In the XFEM formulation [20], the approximation is usually
approximation. In Section 3, we show useful analytical definitions enriched with two types of enrichment functions by means of
of QoI for the enforcement of equilibrium conditions. The recovery the partition of unity: (i) a Heaviside function H to describe the dis-
process is described in Section 4. We discuss the formulation of the continuity of the displacement field along the crack, in the set of
dual problem when considering the stress intensity factor as the
nodes Icrack whose support is intersected by the crack and (ii) a
quantity of interest in the goal oriented approach. Numerical
set of branch functions F ‘ to represent the asymptotic behaviour
results are provided in Section 5 and conclusion are drawn in
Section 6. of the stress field near the crack tip, in the set of nodes Itip whose
support contains the singularity. The XFEM displacement interpo-
2. Problem statement and XFEM for LEFM lation in a 2D model reads:

!
In this section, we introduce the 2D LEFM problem. We denote X X X X
4
h ‘
by u the displacement, by r the Cauchy stress and by e the strain, u ðxÞ ¼ Ni ðxÞai þ Ni ðxÞHðxÞbi þ Ni ðxÞ F ‘ ðxÞci ;
i2I i2Icrack i2Itip ‘¼1
all these fields defined over the domain X  R2 , of boundary denot-
ed by @ X. CN and CD refer to the parts of the boundary where the ð10Þ
 O.A. González-Estrada et al. / Computers and Structures 152 (2015) 1–10 3

where Ni denotes the classical shape functions associated with node purpose, error estimators measured in the energy norm might be
i and a; b; c are the unknown coefficients. Note that we have four utilised to estimate the error in a particular quantity of interest
coefficients c‘ multiplying each of the terms for the F ‘ functions [13]. In this section we show how to a modified version of the Zien-
used in this paper for the 2D case, defined as [20]: kiewicz and Zhu (ZZ) error estimator presented in [6] is used with
  the SPR-CX recovery to evaluate the error in quantities of interest.
pffiffiffi / / / /
fF ‘ ðr; /Þg  r sin ; cos ; sin sin /; cos sin / : ð11Þ A common approach to evaluate the error in QoI involves the
2 2 2 2
use of duality techniques which solve two different problems. A
Considering the enriched finite-dimensional subspace V h  V primal problem, which is the problem at hand as shown in (8),
spanned by locally supported finite element shape functions, we and a dual problem used to extract information on the QoI. Thus,
we shall explain the formulation of the dual problem.
solve for a discrete solution uh 2 V h of the variational problem in
Consider the primal problem given in (8) and its approximate
(8) such that 8v 2 V h :
Z Z finite element solution uh 2 V h  V. Let Q : V ! R be a bounded
h T linear functional representing some quantity of interest, acting
eðu Þ Deðv ÞdX ¼ rT ðuh ÞD1 rðv ÞdX
X X on the space V of admissible functions for the problem at hand.
Z Z Z
We are interested in estimating the error in the functional Q ðuÞ
¼ v T bdX þ v T tdC þ eðv ÞT De0 dX when calculated using the value of the approximate solution uh :
X CN X
Z
 eðv ÞT r0 dX: ð12Þ QðuÞ  Q ðuh Þ ¼ Q ðu  uh Þ ¼ Q ðeÞ: ð17Þ
X
To evaluate Q ðeÞ, the standard procedure is to solve the auxil-
Once we solve the XFEM problem formulated in (12), we can iary or dual problem
obtain an approximation to the exact value of the stress intensity Z
factor K  using ~ 2 V such that 8v 2 V;
Find u eðv ÞT Deðu~ ÞdX ¼ Qðv Þ; ð18Þ
2 3 X
Z uaux
x q;x
" #
1 h T6 7 h T
r aux
xx q;x þ r aux
xy q;y which can be seen as the variational form of an auxiliary mechan-
K ¼  ðr Þ 4 aux
uy q;y 5  ðu Þ dX;
C XI r aux
xy q;x þ r aux
yy q;y ical problem used to extract information of the QoI. The dual dis-
uaux aux
y q;x þ ux q;y
placement field u ~ 2 V vanishes over CD . Test function v is a
ð13Þ virtual displacement. Field r~ ¼ Dðeðu~ Þ  ~e0 Þ þ r ~ 0 and ~e0
~ 0 , where r
h h
where r ¼ Dðeðu Þ  e0 Þ þ r0 is the XFEM stress solution. are known initial stress and strain, can be interpreted as a mechan-
ical stress field. The left-hand side of (18) is the work of internal
3. Error in quantities of interest forces of the auxiliary mechanical problem and Q ðv Þ is the work
of an abstract external load.
We consider the same finite element space used in the primal
Traditionally, the discretisation error, defined as e :¼ u  uh in
problem to look for an approximation of u ~ 2 V such that the prob-
the absence of other types of errors, is measured as the error in
lem is
energy norm. Now we will show the expressions used to evaluate
Z
the error in energy norm which will serve to introduce the error
~ h 2 V h such that 8v 2 V h ;
Find u eðv ÞT Deðu~ h ÞdX ¼ Q ðvÞ: ð19Þ
estimation in Quantities of Interest in this section. X
To quantify the error introduced by the discretisation a com-
To obtain an exact representation for the error Q ðeÞ in terms of the
mon approach is to use the energy norm of e defined as:
Z solution of the dual problem we substitute v ¼ e in (18) and, con-
kek2 ¼ eðeÞT DeðeÞdX: ð14Þ ~h 2 V h:
sidering the Galerkin orthogonality, for all u
X Z
Using the constitutive relation and introducing the error in the QðeÞ ¼ eðeÞT Deðe~ÞdX; ð20Þ
X
stress field er :¼ r  rh the previous expression can be written as:
Z where e~ :¼ u~u ~ h is the discretisation error of the dual problem
2 T 1 (18). We can obtain an expression in terms of the mechanical
kek ¼ er D er dX: ð15Þ
X stresses using the constitutive relation:
Z
Whereas the exact field u is in general unknown, it is possible to
obtain an estimate of the error by means of the approximation QðeÞ ¼ eTr D1 e
~r dX; ð21Þ
X
introduced in [6] in the context of FE elasticity problems:
Z where e~r :¼ r ~ r ~ h is the stress error of the dual problem and
2  
 
 T 1
kek  er D er dX; ð16Þ h h
r~ ¼ Dðeðu~ Þ  ~e0 Þ þ r~ 0 the finite element stress field.
X

where er is the approximated stress error defined by er :¼ r  rh , 3.2. Recovery-based error estimate in QoI
being r the recovered stress field. Local element contributions are
also obtained from (16) considering the domain of the element Xe . The error in the QoI in (21) is related to the errors in the FE
approximations uh and u ~ h . Thus, we can select from the set of
3.1. Error estimation in QoI available procedures to estimate the error in the energy norm a
technique to obtain estimates of the error in the QoI. Considering
The goal of many numerical computations is to control a speci- expressions (16) and (21) we can derive an estimate for the error
fic design parameter, thus, it results natural to formulate the error in the QoI which reads
in terms of such quantity instead of in energy norm. In this case, as Z
indicated in Section 2, the quantity that describes the behaviour of QðeÞ  E ¼ ðer ÞT D1 ðe
~r ÞdX; ð22Þ
X
a component in the surroundings of the crack tip is the SIF, then we
are interested in controlling the accuracy of the SIF instead of where the approximate dual error is e ~r ¼ r ~ h and r
~  r ~  is the
traditional measures such as the error in energy norm. For this recovered auxiliary stress field. Here, we expect to have a sharp
4 O.A. González-Estrada et al. / Computers and Structures 152 (2015) 1–10

estimate of the error in the QoI if the recovered stress fields are 4. The recovery technique
accurate approximations to their exact counterparts.
The recovered stress fields can be computed in many ways, for In this work we consider the SPR-CX recovery technique, intro-
example, by using the SPR technique as explained in [12]. To obtain duced in [25], which is an enhancement of the error estimator
accurate representations of the exact stress fields for the primal introduced in [37], to recover the solutions for the primal and dual
and dual solutions, we propose the use of the locally equilibrated problems. The technique incorporates the ideas in [38] to guaran-
recovery technique described in Section 4. This technique, which tee locally on patches the exact satisfaction of the equilibrium
is an enhancement of the SPR, enforces the fulfilment of the inter- equations, and the extension in [25] to singular problems.
nal and boundary equilibrium equations locally on patches. For Let us define the field rh such that we subtract the initial stress
problems with singularities the stress field is also decomposed into and strain from the field rh :
two parts: smooth and singular, which are separately recovered.
Two remarks have to be made. First, the analytical expressions rh ¼ rh  r0 þ De0 ; ð28Þ
that define the loads for the dual problem are obtained from the h
and perform the recovery on r . Then, the recovered field is
interpretation of the functional Q in terms of tractions, body loads,
initial stresses and strains. Second, to enforce equilibrium condi- r ¼ r þ r0  De0 ; ð29Þ
tions during the recovery process along the boundary of the domain where r  h
is the recovered field that corresponds to r .
of interest (DoI) used to define the QoI, we consider it as an internal
interface. We use different polynomial expansions on each side of 4.1. Splitting concept
the boundary and enforce statical admissibility of the normal and
tangential stresses as it will be explained in Section 4. Different techniques have been used to account for the singular
part during the recovery process [23,25]. Here, following the ideas
3.3. Analytical definition of the dual problem
in [25], for singular problems we can consider that the exact stress
field r allows to be decomposed into two stress fields, a smooth
The recovery technique presented in [25] and used in this con-
field rsmo and a singular field rsing :
tribution requires that the mechanical equilibrium must be made
explicit in order to recover the dual stress field. Thus, the right- r ¼ rsmo þ rsing : ð30Þ
hand side of (18) is interpreted as the work of mechanical external h
This idea is to decompose the XFEM stress field r into two
forces, and the analytical expression of these forces is derived,
depending on the quantity of interest: parts, one smooth rh h
smo and another singular rsing . The SPR-CX tech-
Z nique will use a different recovery procedure for each one of these
~ 2 V such that 8v 2 V :
Find u eðvÞT Deðu~ ÞdX ¼ Q ðv Þ parts. The purpose of this splitting is to account for the singular
X
Z Z Z part of the stress field in the recovery process, which is necessary
¼ ~ Xþ
v T bd v T ~tdC þ eðv ÞT D~e0 dX in the XFEM framework.
X
Z
CN X For the recovery of the singular part, the expressions which
T describe the asymptotic fields near the crack tip are used. To eval-
 eðvÞ r~ 0 dX: ð23Þ
X uate r
sing we first obtain estimated values of the stress intensity

The problem in (23) is solved using a FE approximation with test factors K I and K II using a domain integral (13) method based on
h
and trial functions in V . The finite element solution is denoted by extraction functions [33,39]. Notice that the recovered part r sing

~h 2 V h.
u is an equilibrated field as it satisfies the equilibrium equations.
Such derivations were presented in [34–36]. Here, we only Once the field r sing has been evaluated, an FE-type approxima-

recall the one that is of interest in LEFM, the SIF. In this case, tion (discontinuous) to the smooth part rh
smo can be obtained sub-
Q ðuÞ ¼ K. Eq. (9) can be written as: tracting r
sing from the raw FE field:
2 3
Z
uaux
x q;x rh h
 r
1 smo ¼ r sing : ð31Þ
6 7
Q ðuÞ ¼ K ¼ ðrÞT  4 uaux
y q;y 5
XI C Then, the recovered field of the smooth part, r smo , is evaluated
uu q;x þ uaux
aux
x q;y applying the enhancements of the SPR technique presented in [38]

" aux #
1 rxx q;x þ rxy q;y
aux
to rh
smo . Finally, the recovered stress field can be evaluated with the
 ðuÞT  dX: ð24Þ
C raux
xy q;x þ ryy q;y
aux
expression:

Comparing (23) with (24) we can rewrite (24) as a function of initial r ¼ r 
smo þ rsing : ð32Þ
~
strains ~e0 and body loads b: The concept of stress splitting and the recovery of the singular
Z and smooth parts is used locally on patches, as described next in
T T~
Q ðuÞ ¼ K ¼ rðuÞ ~e0 þ ðuÞ bdX: ð25Þ Section 4.2.
XI

Thus, if we replace u with the vector of arbitrary displacements 4.2. Recovery of the smooth stress field
v , the quantity of interest can be evaluated from
Z Z The SPR-C (C stand for constraints) is a SPR-based stress recov-
Q ðv Þ ¼ rðv ÞT ~e0 dX þ ~ X:
v T bd ð26Þ ery process that introduces a set of constraints to fulfil at each
XI XI
patch of elements equilibrium and compatibility conditions. This
Hence, the initial strains and the body loads per unit volume technique will be applied to the smooth part of the stress field rh
smo .
that can be applied in the dual problem to extract the SIF are In the SPR-C technique, as in the original SPR technique, we
2 3 define a patch P ðJÞ as the set of elements connected to a vertex
uaux
1 q;1
" #
1 ~¼1 raux
11 q;1 þ r21 q;2
aux
node J. On each patch, a polynomial expansion for each one of
~e0 ¼  64 uaux
2 q;2
7
5; b : ð27Þ
C C raux
12 q;1 þ r22 q;2
aux
the components of the recovered stress field is expressed in the
uaux aux
2 q;1 þ u1 q;2 form:
 O.A. González-Estrada et al. / Computers and Structures 152 (2015) 1–10 5

r^ 
k ðxÞ ¼ pðxÞak k ¼ xx; yy; xy; ð33Þ
where p represents a polynomial basis and ak are unknown coeffi-
cients. Usually, the polynomial basis is chosen equal to the non-ex-
tended finite element basis for the displacements. A least squares
approximation to the values of rh smo evaluated at the integration
points of the elements within the patch, xG 2 P ðJÞ , is used to evaluate
the coefficients ak .
For the 2D case, the recovered stress field coupling the three
stress components reads:
8  9 2 38 9
< rxx ðxÞ >
^
> = pðxÞ 0 0 > < axx >
=
r^  6 7
r^ 
smo ðxÞ ¼ yy ðxÞ ¼ PðxÞA ¼ 4 0 pðxÞ 0 5 ayy :
>
: ^  > > >
rxy ðxÞ ; 0 0 pðxÞ
:
axy
;

ð34Þ
Fig. 1. Equilibrium conditions along internal boundaries.
Note that in contrast to the basic SPR, the SPR-C technique uses
a continuous least squares approach to obtain the coefficients A,
according the functional
Z
2
F ðJÞ ðAÞ ¼ ðPA  rh
smo Þ dX: ð35Þ
P ðJÞ

4.2.1. Equilibrium conditions

Constraint equations are introduced via Lagrange multipliers
into the functional defined in (35) on each patch, in order to
enforce the satisfaction of the:

 Internal equilibrium equation: The constraint equation for the

internal equilibrium in the patch is defined as:
8xj 2 P ðJÞ LT r
^ ðJÞ T ^ int
smo ðxj Þ þ L ðr0 ðxj Þ  De0 ðxj ÞÞ þ bðxj Þ :¼ c ðxj Þ ¼ 0;

ð36Þ
^
where bðxÞ is a polynomial least squares fit of degree p  1 to the
actual body forces bðxÞ, being p the degree of the recovered Fig. 2. Westergaard problem. Infinite plate with a crack of length 2a under uniform
tractions r1 (biaxial) and s1 . Finite portion of the domain X0 , modelled with FE.
stress field r ^ ðJÞ int
smo . We enforce c ðxj Þ at a sufficient number of j
non-aligned points (nie) to guarantee the exact representation
^ Optimizing functional (38) we obtain a linear system of equa-
of bðxÞ, e.g. three points for a linear representation. This proce-
tions to evaluate the coefficients A. To enforce equilibrium condi-
dure will add one equation per point j to the linear system to
tions along internal boundaries (e.g. bimaterial problems,
solve at each patch.
problems with zones subjected to different body forces, etc.), we
 Boundary equilibrium equations: We use a point collocation
consider different polynomial expansions on each side of the
approach to impose the satisfaction of a polynomial approxima-
boundary and enforce the statical admissibility condition imposing
tion to the tractions along the Neumann boundary intersecting
equilibrium along this boundary. Suppose that we have a patch
the patch. The constraint equation reads
intersected by CI such that Xe ¼ X1;e [ X2;e for intersected ele-
T ments, as shown in Fig. 1. To enforce equilibrium conditions along
8xj 2 CN \ P ðJÞ Gr
^ ðJÞ
smo ðxj Þ þ GL ðr0 ðxj Þ  De0 ðxj ÞÞ
ext CI we define the stresses r ^  ^ 
X1 ; rX2 at each side of the internal
þ Gr
sing  tðxj Þ :¼ c ðxj Þ ¼ 0: ð37Þ
boundary. Then, the boundary equilibrium along CI given the
ext
We enforce c ðxj Þ in nbe ¼ p þ 1 points along the part of the prescribed tractions tCI ¼ ½tx t y T is:
boundary crossing the patch. In the case that more than one
boundary intersects the patch, only one curve is considered in ^ 
Gðr ^ 
X1 jCI  rX2 jCI Þ ¼ tCI : ð39Þ
order to avoid over-constraining.
The same procedure can be used for patches intersected by the
 Compatibility equations: ccmp ðxj Þ is only imposed in the case
crack. In this case, we could consider the traction-free condition
that p P 2 in a sufficient number of non-aligned points. r ^ 
smo along the crack faces or define a different prescribed condition
directly satisfies ccmp for p ¼ 1. depending on the configuration.
The continuity of the recovered field is obtained by using a par-
Thus, the Lagrange functional enforcing the constraint equa- tition of unity procedure [40] to weight the stress fields obtained
tions for a patch P ðJÞ can be written as from the patches formed at the vertex nodes of the element. The
X
nie
 int  Xnbe
 ext  field r is interpolated using linear shape functions N ðJÞ associated
LðJÞ ðA; kÞ ¼ F ðJÞ ðAÞ þ kint
i c ðxi Þ þ kext
j c ðxj Þ
with the nv vertex nodes and adding the contributions of the
i¼1 j¼1
smooth and singular parts as in (32), such that
X
nc
þ kcmp
k ðccmp ðxk ÞÞ: ð38Þ X
nv

k¼1 r ðxÞ ¼ N ðJÞ ðxÞðr

^ ðJÞ ^ ðJÞ
smo ðxÞ þ rsing ðxÞÞ: ð40Þ
J¼1
This functional will expand the linear system to solve at each patch
to include the additional equations to enforce the internal and ^ 
Here, to reduce the computational cost, r sing is evaluated over the
boundary equilibrium and the compatibility equation. set of points near the area of influence of the singularity, defined
6 O.A. González-Estrada et al. / Computers and Structures 152 (2015) 1–10

technique presented above. For that purpose we define the effec-

tivity index of the error estimator h as:
E
h¼ ; ð41Þ
Q ðeÞ
where Q ðeÞ denotes the exact error in the quantity of interest, and E
represents the evaluated error estimate. We can also represent the
effectivity in the QoI defined as

Q ðuh Þ þ E
hQoI ¼ ; ð42Þ
Q ðuÞ
and the relative error in the QoI for the exact and estimated error
jQðeÞj jE j
gQ ðeÞ ¼ ; gE ¼ : ð43Þ
jQ ðuÞj jQ ðuh Þ þ E j

5.1. Westergaard problem – FEM solution

Fig. 3. Domain of interest for the extraction of the stress intensity factor.

Let us consider the Westergaard problem [25,41] of linear

by the splitting radius, as shown in [25]. After evaluating the equi-
librated recovered fields on each patch r ^ ðJÞ , we use (40) to obtain a
continuous field. This process introduces a lack of equilibrium
Pv
s ¼ nJ¼1 rN ðJÞ r
^ ðJÞ when evaluating the divergence of the internal
equilibrium equation, as explained in [37,26].
5. Numerical results
In this section we consider numerical examples for 2D problems
with exact analytical solution to evaluate the performance of the
elastic fracture mechanics for which the exact analytical solution
is known. The Westergaard problem corresponds to an infinite
plate loaded at infinity with biaxial tractions rx1 ¼ ry1 ¼ r1
and shear traction s1 , presenting a crack of length 2a as shown
in Fig. 2. Combining the externally applied loads we can obtain dif-
ferent loading conditions: pure mode I, pure mode II or mixed
mode.
The numerical model corresponds to a finite portion of the
domain (a ¼ 5 and b ¼ 10 in Fig. 2). The applied projected stresses
for mode I are evaluated from the analytical Westergaard solution
[41]:

Fig. 4. Equivalent forces at nodes for the primal (left) and dual (right) problems.
 O.A. González-Estrada et al. / Computers and Structures 152 (2015) 1–10 7

300

200

100

Fig. 5. FE (left) and recovered (right) ryy for the primal problem.

x 105
20

15

10

Fig. 6. FE (left) and recovered fields (right) ryy for the dual problem.

"

#
I r1 / / a2 / / Material parameters are Young’s modulus E ¼ 107 and Poisson’s
r x ðx; yÞ ¼ pffiffiffiffiffi x cos  y sin þ y 2 m sin  n cos ;
ratio m ¼ 0:333. We consider loading conditions in pure mode I
jt j 2 2 jt j 2 2
"

# with r1 ¼ 100 and s1 ¼ 0, and pure mode II with r1 ¼ 0 and
r / / a2 / / s1 ¼ 100. We assume plane strain conditions.
rIy ðx; yÞ ¼ p1ffiffiffiffiffi x cos  y sin  y 2 m sin  n cos ;
jt j 2 2 jt j 2 2 In the numerical analysis, we use a geometrical enrichment

defined by a circular fixed enrichment area Bðx0 ; r e Þ with radius
a2 r / /
sIxy ðx; yÞ ¼ y 2 p1ffiffiffiffiffi m cos þ n sin re ¼ 2:5, with its centre at the crack tip x0 as proposed in [42].
jtj jt j 2 2
Bilinear elements are considered in the models, using a sequence
ð44Þ of uniformly refined meshes. For the numerical integration of stan-
and for mode II: dard elements we use a 2 2 Gaussian quadrature rule. We use a
"

# 5 5 quasipolar integration in the subdomains of the element con-
II s1 / / a2 / / taining the crack tip [42]. We do not consider correction for blend-
r x ðx; yÞ ¼ pffiffiffiffiffi 2 y cos þ x sin  y 2 m cos þ n sin ;
ing elements. Methods to address blending errors are proposed in
jt j 2 2 jt j 2 2

[43–46].
a2 s / /
rIIy ðx; yÞ ¼ y 2 p1ffiffiffiffiffi m cos þ n sin ; To evaluate the stress intensity factor K we use an EDI tech-
jt j jt j 2 2 nique [33]. For the primal problem we consider a square plateau
"

#
s / / a2 / / function q centred at the crack tip, as shown in Fig. 3. q ¼ 1 for
sIIxy ðx; yÞ ¼ p1ffiffiffiffiffi x cos  y sin þ y 2 m sin  n cos ; the domain defined by an inner square with side length 6 and
jt j 2 2 jt j 2 2
q ¼ 0 for the part of the domain outside the outer square with side
ð45Þ length 8, q is interpolated in-between the two squares. This plateau
where the stress fields are expressed as a function of x and y, with function is also used to define the subdomain Xi when extracting
origin at the centre of the crack. The parameters t; m; n and / are the quantity of interest in the dual problem. As the dual problem
defined as is also a singular problem we have to evaluate a second stress
intensity factor. In this case, we use a plateau function such that
2
t ¼ ðx þ iyÞ  a2 ¼ ðx2  y2  a2 Þ þ ið2xyÞ ¼ m þ in; q ¼ 1 for all nodes inside a square with side length 4.9 and q ¼ 0
otherwise.
m ¼ ReðtÞ ¼ Reðz2  a2 Þ ¼ x2  y2  a2 ;
ð46Þ In Fig. 4 we represent the equivalent nodal forces used to solve
n ¼ ImðtÞ ¼ ðz2  a2 Þ ¼ 2xy; the primal and dual problems. For the dual problem the vector of
2 forces is constructed using the discrete approximation of the dual
/ ¼ ArgðtÞ ¼ Argðm  inÞ with / 2 ½p; p; i ¼ 1: function. The Dirichlet boundary constraints are the same for both
For the problem analysed, the exact value of the SIF is given by models. For the dual problem, we can see that the forces are dis-
pffiffiffiffiffiffi pffiffiffiffiffiffi tributed in the nodes located in the domain of interest. For the
K I;ex ¼ r1 pa K II;ex ¼ s1 pa: ð47Þ recovery of the primal and dual fields, we perform the splitting
8 O.A. González-Estrada et al. / Computers and Structures 152 (2015) 1–10

Table 2
Stress intensity factor K II as QoI.

dof E Q ðeÞ h hQOI Ea ha hQoI a

351 2.2886 1.1273 2.0302 1.00735 5.6052 4.9724 1.02832

1289 0.2827 0.2902 0.9744 0.99995 2.2203 7.6517 1.01221
4973 0.07 0.0689 1.0147 1.00001 1.1078 16.0677 1.00657
19,637 0.0168 0.0159 1.0554 1.00001 0.5502 34.6562 1.00338
a
Results using the standard SPR recovery.

Fig. 7. Evolution of the effectivity index h considering the SIF as quantity of interest
under mode I and mode II loading conditions.

Table 1
Stress intensity factor K I as QoI.

dof E Q ðeÞ h hQOI Ea ha hQoI a

351 2.5264 2.1144 1.1948 1.00261 13.7089 6.4835 1.07333

1289 0.4822 0.5146 0.9369 0.99979 6.3880 12.4124 1.03715
4973 0.1140 0.1216 0.9376 0.99995 3.1756 26.1141 1.01932
19,637 0.0267 0.0278 0.9617 0.99999 1.5789 56.7646 1.00981
Fig. 8. Evolution of the effectivity index h for the SPRCX and SPR. Mode I and mode
a
Results using the standard SPR recovery. II loading conditions.

of stresses and enforce internal equilibrium, boundary equilibrium

x 10−6
and the compatibility equation.
The yy-component of the stress field for the raw FE and the 4
recovered solutions is represented in Fig. 5. The enrichment area
is indicated with a circle. In Fig. 6 we show the same results for 3
the dual problem. Notice how the recovery procedure smoothes
the stresses along the interface of the domain of interest. As the
dual problem is also characterised by the crack, we have to evalu- 2
ate the corresponding stress intensity factor and perform the singu-
lar + smooth decomposition of the stress field. 1
Fig. 7 shows the evolution of the effectivity index h as we
increase the number of degrees of freedom (dof). We consider as
quantities of interest the two SIFs characterising two different
loading conditions, i.e. mode I and mode II. We can see that for x 10−3
both quantities the error estimator yields effectivities close to
the optimal value h ¼ 1.
In Tables 1 and 2 we indicate the values for the estimated, E, 20
and exact, Q ðeÞ, errors, the global effectivity index h and the effec-
tivity for the quantity of interest hQoI using the proposed recovery
technique and the standard SPR (denoted with y). The magnitude 10
of the exact error is accurately captured by the estimated obtained
with the SPR-CX, which is clearly reflected in the good effectivity
index for both loading modes. As expected, the effectivity in the 0
quantities of interest hQoI is highly accurate. For the SPR, although
the value E decreases as we increase the number of dof, the error
estimate is not as accurate as the estimate obtained with the
Fig. 9. Distribution of the estimated error considering the error in energy norm
SPR-CX and does not decrease as fast as the exact error, loosing
(top) and the error in the quantity of interest E (bottom).
asymptotic exactness. The SPR does not consider the splitting of
the singular stresses, giving less accurate results close to the crack
tip, and does not enforce equilibrium conditions in the primal and by the technique does not converge to the exact error [32]. This
dual recovered fields, which results in a poorer description of the behaviour is similar for the two loading modes.
stresses close to the boundaries and the interface of the domain In Fig. 9 we represent the distribution of the estimated error for
of interest. the second mesh of the sequence for the error in energy norm kees k
Fig. 8 compares the results of the proposed SPR-CX recovery and the error considering the quantity of interest E. This error dis-
with the standard SPR technique. In particular, the SPR cannot tribution might guide the refinement during the adaptivity proce-
properly recover singular fields, thus, the error estimate provided dure. The approach based on energy norm estimates that the most
 O.A. González-Estrada et al. / Computers and Structures 152 (2015) 1–10
critical part is located in the vicinity of the singular point whilst the
goal oriented approach also considers the domain where the infor-
mation of the QoI is extracted.
6. Conclusions and future work
We have presented a locally equilibrated recovery procedure for
goal oriented error estimation in XFEM. We have considered as the
design parameter the generalised stress intensity factor that char-
acterises the solution of singular problems in the context of linear
elastic fracture mechanics. The technique proposes the use of a
stress recovery that provides locally equilibrated stress fields for
both the primal and the dual problem.
To formulate the dual problem we consider the linear equiva-
lent domain integral representing K to obtain the applied loads of
the dual FE approximation. To perform the recovery of the primal
and dual solutions we consider three main ideas: (i) enforcement
of the internal equilibrium equation, (ii) enforcement of boundary
equilibrium and (iii) splitting of the stress field into singular and
smooth parts.
The proposed technique has been tested with problems under
different loading conditions. The obtained results show that the
error estimator accurately captures the exact error in the eval-
uation of the stress intensity factor.
Acknowledgements
This work was supported by the EPSRC grant EP/G042705/1 ‘‘In-
creased Reliability for Industrially Relevant Automatic Crack
Growth Simulation with the eXtended Finite Element Method’’.
Stéphane Bordas also thanks partial funding for his time provided
by the European Research Council Starting Independent Research
Grant (ERC Stg Grant Agreement No. 279578) ‘‘RealTCut Towards
real time multiscale simulation of cutting in non-linear materials
with applications to surgical simulation and computer guided sur-
gery’’. This work has been carried out within the framework of the
research project DPI2010-20542 of the Ministerio de Ciencia e
Innovación (Spain). The financial support of the FPU program
(AP2008-01086), the funding from Universitat Politècnica de
València and Generalitat Valenciana (PROMETEO/2012/023) are
also acknowledged.
References
[1] Anderson TL. Fracture mechanics: fundamentals and applications, vol. 2. Boca
Ratón (Florida): CRC Press; 1995.
[2] Gross D, Seelig T. Fracture mechanics with an introduction to micromechanics.
2nd ed. Springer; 2011.
[3] Rüter M, Stein E. Goal-oriented a posteriori error estimates in linear elastic
fracture mechanics. Comput Methods Appl Mech Eng 2006;195(4–6):251–78.
[4] Panetier J, Ladevèze P, Chamoin L. Strict and effective bounds in goal-oriented
error estimation applied to fracture mechanics problems solved with XFEM. Int
J Numer Methods Eng 2010;81(6):671–700.
[5] Babuška I, Rheinboldt WC. A-posteriori error estimates for the finite element
method. Int J Numer Methods Eng 1978;12(10):1597–615.
[6] Zienkiewicz OC, Zhu JZ. A simple error estimator and adaptive procedure for
practical engineering analysis. Int J Numer Methods Eng 1987;24(2):337–57.
[7] Ladevèze P, Leguillon D. Error estimate procedure in the finite element method
and applications. SIAM J Numer Anal 1983;20(3):485–509.
[8] Almeida Pereira OJB, Moitinho de Almeida JP, Maunder EAW. Adaptive
methods for hybrid equilibrium finite element models. Comput Methods
Appl Mech Eng 1999;176(1–4):19–39.
[9] Babuška I, Strouboulis T, Upadhyay CS. 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. Comput Methods Appl Mech
Eng 1994;114(3–4):307–78.
[10] Babuška I, Strouboulis T, Upadhyay CS, Gangaraj SK, Copps K. Validation of a
posteriori error estimators by numerical approach. Int J Numer Methods Eng
1994;37(7):1073–123.
[11] Babuška I, Strouboulis T, Upadhyay CS. A model study of the quality of a
posteriori error estimators for finite element solutions of linear elliptic
9
problems, with particular reference to the behaviour near the boundary. Int
J Numer Methods Eng 1997;40(14):2521–77.
[12] Zienkiewicz OC, Zhu JZ. The superconvergent patch recovery and a posteriori
error estimates. Part 1: the recovery technique. Int J Numer Methods Eng
1992;33(7):1331–64.
[13] Ainsworth M, Oden JT. A posteriori error estimation in finite element
analysis. Chichester: John Wiley & Sons; 2000.
[14] Paraschivoiu M, Peraire J, Patera AT. A posteriori finite element bounds for
linear-functional outputs of elliptic partial differential equations. Comput
Methods Appl Mech Eng 1997;150(1–4):289–312.
[15] Ladevèze P, Rougeot P, Blanchard P, Moreau JP. Local error estimators for finite
element linear analysis. Comput Methods Appl Mech Eng 1999;176(1–
4):231–46.
[16] Oden JT, Prudhomme S. Goal-oriented error estimation and adaptivity for the
finite element method. Comput Math Appl 2001;41(5–6):735–56.
[17] Cirak F, Ramm E. A posteriori error estimation and adaptivity for linear
elasticity using the reciprocal theorem. Comput Methods Appl Mech Eng
1998;156(1–4):351–62.
[18] Moitinho de Almeida JP, Almeida Pereira OJB. Upper bounds of the error in
local quantities using equilibrated and compatible finite element solutions for
linear elastic problems. Comput Methods Appl Mech Eng 2006;195(4–
6):279–96.
[19] González-Estrada OA, Nadal E, Ródenas JJ, Kerfriden P, Bordas SPA, Fuenmayor
FJ. Mesh adaptivity driven by goal-oriented locally equilibrated
superconvergent patch recovery. Comput Mech 2013;53(5):957–76.
[20] Moës N, Dolbow J, Belytschko T. A finite element method for crack growth
without remeshing. Int J Numer Methods Eng 1999;46(1):131–50.
[21] Amani J, Saboor Bagherzadeh A, Rabczuk T. Error estimate and adaptive
refinement in mixed discrete least squares meshless method. Math Problems
Eng 2014;2014:1–16.
[22] Ghorashi S, Valizadeh N, Mohammadi S. Extended isogeometric analysis for
simulation of stationary and propagating cracks. Int J Numer Methods Eng
2012;89:1069–101.
[23] Bordas SPA, Duflot M. Derivative recovery and a posteriori error estimate for
extended finite elements. Comput Methods Appl Mech Eng 2007;196(35–
36):3381–99.
[24] Duflot M, Bordas SPA. A posteriori error estimation for extended finite
elements by an extended global recovery. Int J Numer Methods Eng 2008;76:
1123–38.
[25] Ródenas JJ, González-Estrada OA, Tarancón JE, Fuenmayor FJ. A recovery-type
error estimator for the extended finite element method based on
singular+smooth stress field splitting. Int J Numer Methods Eng 2008;76(4):
545–71.
[26] Ródenas JJ, González-Estrada OA, Díez P, Fuenmayor FJ. Accurate recovery-
based upper error bounds for the extended finite element framework. Comput
Methods Appl Mech Eng 2010;199(37–40):2607–21.
[27] Strouboulis T, Zhang L, Wang D, Babuška I. A posteriori error estimation for
generalized finite element methods. Comput Methods Appl Mech Eng 2006;
195(9–12):852–79.
[28] Gerasimov T, Rüter M, Stein E. An explicit residual-type error estimator for Q1-
quadrilateral extended finite element method in two-dimensional linear
elastic fracture mechanics. Int J Numer Methods Eng 2012;90(April):1118–55.
[29] Panetier J, Ladevèze P, Louf F. Strict bounds for computed stress intensity
factors. Comput Struct 2009;87(15–16):1015–21.
[30] Rüter M, Gerasimov T, Stein E. Goal-oriented explicit residual-type error
estimates in XFEM. Comput Mech 2013;52(2):361–76.
[31] Pannachet T, Sluys LJ, Askes H. Error estimation and adaptivity for
discontinuous failure. Int J Numer Methods Eng 2009;78(5):528–63.
[32] González-Estrada OA, Ródenas JJ, Bordas SPA, Duflot M, Kerfriden P, Giner E.
On the role of enrichment and statical admissibility of recovered fields in a-
posteriori error estimation for enriched finite element methods. Eng Comput
2012;29(8):1–28.
[33] Szabó BA, Babuška I. Finite element analysis. New York: John Wiley & Sons;
1991.
[34] Ródenas JJ. Goal oriented adaptivity: Una introducción a través del problema
elástico lineal. Technical report, CIMNE, PI274, Barcelona, Spain; 2005.
[35] González-Estrada OA, Ródenas JJ, Nadal E, Bordas SPA, Kerfriden P.
Equilibrated patch recovery for accurate evaluation of upper error bounds in
quantities of interest. Adaptive modeling and simulation. In: Aubry D, Díez P,
Tie B, Parés N, editors. Proceedings of V ADMOS 2011. CINME: Paris; 2011.
[36] Verdugo F, Díez P, Casadei F. Natural quantities of interest in linear
elastodynamics for goal oriented error estimation and adaptivity. Adaptive
modeling and simulation. In: Aubry D, Díez P, Tie B, Parés N, editors.
Proceedings of V ADMOS 2011. CIMNE: Paris; 2011.
[37] Díez P, Ródenas JJ, Zienkiewicz OC. Equilibrated patch recovery error
estimates: simple and accurate upper bounds of the error. Int J Numer
Methods Eng 2007;69(10):2075–98.
[38] Ródenas JJ, Tur M, Fuenmayor FJ, Vercher A. Improvement of the
superconvergent patch recovery technique by the use of constraint
equations: the SPR-C technique. Int J Numer Methods Eng 2007;70(6):705–27.
[39] Giner E, Tur M, Fuenmayor FJ. A domain integral for the calculation of
generalized stress intensity factors in sliding complete contacts. Int J Solids
Struct 2009;46(3–4):938–51.
[40] Blacker T, Belytschko T. Superconvergent patch recovery with equilibrium and
conjoint interpolant enhancements. Int J Numer Methods Eng 1994;37(3):
517–36.
10 O.A. González-Estrada et al. / Computers and Structures 152 (2015) 1–10
[41] Giner E, Fuenmayor FJ, Baeza L, Tarancón JE. Error estimation for the finite
element evaluation of GI and GII in mixed-mode linear elastic fracture
mechanics. Finite Elem Anal Des 2005;41(11–12):1079–104.
[42] Béchet E, Minnebo H, Moës N, Burgardt B. Improved implementation and
robustness study of the X-FEM for stress analysis around cracks. Int J Numer
Methods Eng 2005;64(8):1033–56.
[43] Chessa J, Wang H, Belytschko T. On the construction of blending elements for
local partition of unity enriched finite elements. Int J Numer Methods Eng
2003;57(7):1015–38.
[44] Gracie R, Wang H, Belytschko T. Blending in the extended finite element
method by discontinuous Galerkin and assumed strain methods. Int J Numer
Methods Eng 2008;74(11):1645–69.
[45] Fries T. A corrected XFEM approximation without problems in blending
elements. Int J Numer Methods Eng 2008;75(5):503–32.
[46] Tarancón JE, Vercher A, Giner E, Fuenmayor FJ. Enhanced blending elements
for XFEM applied to linear elastic fracture mechanics. Int J Numer Methods
Eng 2009;77(1):126–48.