You are on page 1of 80

XFEM 2011

II European Conference on eXtended Finite Element


Cardiff University, United Kingdom
Cardiff, June 29-30 July 1, 2011

A Geometrically Non-linear, Discontinuous Solid-Like Shell Element


A. Ahmed, F.P. van der Meer, L.J. Sluys
Delft University of Technology, The Netherlands, {A.Ahmed,F.P.vandermeer,L.J.Sluys}@tudelft.nl

Introduction

Shell structures are frequently used in many fields of engineering due to their structural efficiency and
light weight constructions. However, like other structures they are also susceptible to various types
of defects and damages such as initiation and propagation of cracks, global and local buckling etc.,
which may impair structural soundness. Correct predictions of failure in such structures is fundamental.
There has been growing interest in the industrial and defense community, in developing reliable and
computationally efficient numerical tools, able to represent different failure mechanisms. These tools are
necessary for the design and analysis of new or existing structures and materials.
Traditionally, analysis of shell structures is performed with standard shell elements, where the three
dimensional body is reduced to a two dimensional body embedded in a three dimensional eucledian
space. The kinematic quantities are defined at the shells mid-surface. However, such models do not
allow damage analysis at lower level of observation e.g meso-scopic level. In this spirit, solid shells
and solid-like shell elements were formulated to perform a complete three dimensional analysis of thin
walled structures. Such shell models, like the one by [1], on one hand incorporate the kinematics of
shells directly by linear interpolation of 3D kinematic relations and remove contradictory assumptions
originated due to dimensional reduction from 3D to 2D in conventional shell models. On the other hand,
they remove the poisson thickness locking effect commonly found in volume elements, when employed
for thin shell applications, due to large aspect ratios.
Failure analysis of thin walled structures demands numerical models which allow simultaneous treatment
of material and geometrical nonlinearity and are able to represent the micro-mechanical behavior at meso
or macro-scopic level of observation. This work is concerned with the development of a discontinuous
solid-like shell (DSLS) finite element model which takes into account mesh independent modeling of
initiation and propagation of cracking phenomena in combination with large deformations and structural
stability.

Kinematics of DSLS

The discontinuous shell finite element is based on a solid-like shell theory [1]. The solid-like shell element is derived from a purely displacement based description of the kinematics of a shell. Hence,
no rotational degrees of freedom are associated with the element. This alleviates the problems associated with standard shell elements, when modeling interfaces between two material layers or cracking
phenomena and/or connecting it with other 3D finite elements.

2.1

Discontinuous displacement field

In order to formulate the discontinuous shell theory in a consistent way, we introduce a curvilinear coordinate system (, , ), which automatically preserves objectivity. The shell element consists of a shell
top surface, a bottom surface and a degenerated shell mid-surface. We consider an eight noded solid-like
1

  







 

 




 

 

 


 

 
  

 






 
  


 


  



  




 
   

 


 


  

Figure 1: Cracked element replaced by a pair of partially active elements and phantom nodes
shell element with three translational degrees of freedom per node. In addition to these, there are also
four internal degrees of freedom at the four corners of the shell mid-surface, which can be condensed
out at element level. This leads to an easy implementation in a standard finite element software. These
internal degrees of freedom describe the internal stretching of an element and avoid thickness locking
phenomena in finite element computations. The displacement field of a solid-like shell is constructed
using top surface displacement field ut , bottom surface displacement field ub , and thickness stretch parameter . The displacement field of a solid-like shell in a curvilinear coordinates system (, , ) is then
written as
(, , ) = uo (, ) + u1 (, ) + (1 2 )u2 (, )

(1)

where uo is the mid-surface displacement,


uo (, ) = [ut (, ) + ub (, )] /2
u1 is the displacement of shells director
u1 (, ) = [ut (, ) ub (, )] /2
u2 denotes internal stretching of an element, which is collinear with the deformed shell director, d and a
function of thickness stretch parameter,
u2 (, ) = d(, )
Let us now consider that a body with domain o contains an internal surface oc , such as a crack, which
divides the domain into two sub-domains oA and oB . This incorporates a strong discontinuity in the
displacement field of the body. To allow for mesh independent modeling of propagating cracks in a
finite element framework, we follow [2], where an element crossed by a discontinuity is replaced by
two partially active elements with the addition of phantom nodes. We call these two new elements A
and B with their active domains represented by o,elem
and o,elem
such that o,elem = o,elem
o,elem
,
B
B
A
A
2

see figure 1. The superscript elem represents a particular element in a finite element mesh crossed by a
discontinuity.
The displacement field is defined only over the active domains of each element.

A (X) X oA
(X) =
(2)
B (X) X oB
The displacement field is continuous over the active domains and discontinuous over oc . Since the two
newly added elements are similar to uncracked shell elements, we can readily write the displacement
field for the pair of elements as:
e = uoe + u1e + (1 2 )u2e

e = A, B

(3)

Since the two elements do not share any of their nodes, the displacement fields of both elements are
independent of each other. This incorporates the desired discontinuity in the approximation field, which
results in a discontinuous shells mid-surface, shells director and most importantly a discontinuous
stretching of the element on each side of a crack. The displacement jump is defined as = B A .
In order to incorporate inelastic material behavior near the crack tip zone, we follow the work by [3] for
cohesive fracture, where such nonlinearities are captured by a cohesive constitutive law which relates the
interface tractions to the jump in the displacement field. As a result of this the independent displacement
fields of the pair of elements will now only interact in the cohesive zones.

2.2

Strain measure

We use the Green Lagrange strain tensor E as a suitable strain measure for large deformation problems
in terms of a contravariant basis Gi as
Ee = Ei j|e Gi G j
with
Ei j|e

1
(Gi , j + G j ,i + ,i , j )e
2

(4)

where Gi are contravariant base vectors in the reference configuration.


Note that due to the independent definitions of displacement fields in pairs of elements, the strain field of
a cracked element resulted in an independent definition of strain fields in each element. Thus the model
allows for different strain conditions on both sides of the discontinuity, which is essential for modeling
phenomena like local buckling, which can also occur on only one side of the crack.

Equilibrium

We start with the strong form of the equilibrium equation in the reference configuration as
DIV P + b = 0 in o \ oc
where P is the first Piola-Kirchoff stress tensor. b is the body force with respect to the reference volume
and defined as b = Jb. Where J is the Jacobian , J = det(F) and b is the body force in the current
configuration.
The corresponding boundary conditions of the problem are
u = u
P nto
PB noc,B
PA noc,A

= t
= tc,B
= tc,A

at ou

(5)

at to

(6)

at oc

(7)

at oc

(8)
oc

Note that equations (7),(8) enforce traction continuity across the discontinuity surface
in the reference configuration, where tc,B = tc,A = tc are the cohesive tractions acting on the crack surface oc and
3

noc,B = noc,A = noc , an outward normal of oB .


Defining as the compatible variation of the displacement field, the weak form of equilibrium is obtained as

o \oc

o : P +

oc

tc

to

o \oc

b = 0

Since the term o : P is energetically equivalent to the work-conjugate term E : , the above equilibrium equation can be written as

o \oc

E : +

oc

tc

to

o \oc

b = 0

Where is the second Piola-Kirchoff stress tensor. The above equation defines the principle of virtual
work. Next using the additive property of integrals and taking first A = 0 and then B = 0, the
following two variational statements are obtained:

oA

EA : +

oB

oc

EB : +

A tc
oc

B tc

o
t,A

A t

o
t,B

B t

oA

A b = 0

(9)

oB

B b = 0

(10)

Numerical Example




A cohesive crack growth analysis is performed in combination with large displacement/rotation analysis
of a shell structure. Figure 2 shows the semi-cylindrical shell with an initial crack. The cylinder is
subjected to an end pinching force at the middle of the free-hanging circumferential periphery.



  
 
   









 


Figure 2: Model and deformed shape of a pinched semi-cylinder with crack growth

References
[1] H. Parisch, A Continuum Based Shell Theory For Non-Linear Applications, International Journal for Numerical
Methods in Engineering, 38: 1855-1883, 1995
[2] A. Hansbo, P. Hansbo, A Finite Element Method for the Simulation of Strong and Weak Discontinuities in Solid
Mechanics, Computer Methods in Applied Mechanics and Engineering, 193: 3523-3540, 2004
[3] G.N. Wells, L.J. Sluys, A New Method for Modeling Cohesive Cracks using Finite Elements, International
Journal for Numerical Methods in Engineering, 50: 2667-2682, 2001

XFEM 2011

II European Conference on eXtended Finite Element


Cardiff University, United Kingdom
Cardiff, June 29-30 July 1, 2011

A stabilized Lagrange multiplier method for the XFEM approximation of contact problems in elastostatics
S. Amdouni1 , P. Hild2 , V. Lleras3 , Y. Renard1
1
2
3

INSA Lyon, France, {Saber.Amdouni,Yves.Renard}@insa-lyon.fr


Universit de Franche Comt, France, patrick.hild@univ-fcomte.fr
Universit de Montpellier 2, France, vlleras@math.univ-montp2.fr

Introduction

To make the finite element methods more flexible, Mos, Dolbow and Belytschko [7] have introduced
in 1999 the XFEM (eXtended Finite Element Method) which allows to perform finite element computations on multi-cracked domains by using meshes of the non-cracked domain. In the original method, the
asymptotic displacement is incorporated into the finite element space multiplied by the shape function
of a background Lagrange finite element method. In this paper, we deal with a variant, introduced in [8],
where the asymptotic displacement is multiplied by a cut-off function. After numerous numerical works
developed in various contexts of mechanics, the first convergence results with a priori error estimates of
order h (h denotes the discretization parameter) under H 2 regularity of the regular part of the solution (in
linear elasticity) were recently obtained in [8]. Some remarkable numerical work has been performed on
a posteriori error estimation for XFEM : by derivative recovery technique [2, 10] or by residual [5].
Few works deal with contact and XFEM and most particularly two main methods are used to formulate contact problems: penalty method [4] and Lagrange multiplier method [3, 6]. To overcome the
drawbacks of the two main methods (choice of the penalization parameter or verification of the BabuskaBrezzi condition), we consider in this work a mixed finite element method which does not require an infsup condition. Such methods which provide stability of the multiplier by adding supplementary terms in
the weak formulation have been originally introduced in [1]. Among the previous works, anyone treats
the error estimates with the contact for the XFEM method. The rapid uptake of the XFEM method by
industry will require adequate error estimation tools to be available to the analysts. Our purpose in this
paper is to carry out the convergence analysis of a stabilized Lagrange multiplier method for the XFEM
approximation of contact problems in elastostatics.

Formulation of the contact and continuous problem

We consider a cracked elastic body occupying a domain in R2 . The contact part of the boundary,
C , represents the cracks location which, for the sake of simplicity, we assume it to be a straight line
segment. In order to deal with the contact between the two sides of the crack as a contact between two
elastic bodies, we use C+ and C to represent each of the two sides of the crack. We consider the
problem of plane linear elasticity: find the displacement field u : R2 satisfying
div (u) + f = 0

in ,

(1)

in ,

(2)

u = 0

on D ,

(3)

(u)n = g

on N .

(4)

(u) = C(u)

The conditions describing the unilateral contact without friction on C are:

un 0, +
n (u) = n (u) = n (u) 0, n (u) un = 0, t (u) = t (u) = 0,

(5)

where un is the jump of the normal displacement across the crack C . We need to introduce the
following spaces:

2
V = v H 1 () : v = 0 on D

M = H 1/2 (C ) :
d 0 for all H 1/2 (C ), 0 a.e. on C .
C

The mixed formulation of the unilateral contact problem without friction (1)(5) consists then in finding
u V and M such that

vn d = f v d +
g v d,
v V,
(u) : (v) d
C

(6)

( )un d 0,
M.
C

The discrete problem

The extended finite element space Vh is defined as

Vh = vh (C ())2 : vh =

ai i +

iNh

iNhH

bi Hi + c j Fj ,
j=1

ai , bi , c j R2 V.

is a C2 cut-off function such that there exists 0 < r0 < r1 with r the distance to the crack tip :

if r r0 ,
1
(0, 1) if r0 < r < r1 ,
(r) =

0
if r r1 .

On the one hand the discontinuous functions along the crack are constructed using the Heaviside function
H(x) = +1

if (x x ) n+ 0, 1

otherwise,

where
denotes the position of the crack tip. On the other hand the singular functions at the crack
tip are constructed using the four singular basis functions: {Fj (x)}1 j4 are the Westergaard functions
defined in polar coordinates located at the crack tip by

F = Fj (x), 1 j 4 =
r sin , r cos , r sin sin , r cos sin ,
2
2
2
2

and i are scalar-valued shape functions associated with classical finite-element method of order k at the
node of index i, Nh denotes the set of all node indices, and NhH denotes the set of nodes indices enriched
by the function H(), i.e., nodes indices for which the support of their shape function is completely cut
by the crack.
Let x0 , ..., xN be given distinct points lying in C . These nodes form a one-dimensional family of meshes
of C denoted by T H . The mesh T H allows us to define:

M0H = H L2 (C ) : H|(x ,x ) P0 (xi , xi+1 ); H 0 on C ,


(7)
i i+1

M1H = H C(C ) : H|(x ,x


i

i+1 )

P1 (xi , xi+1 ), H 0 on C .

(8)

Let Rh be an operator from Vh onto L2 (C ) which approaches the normal component of the stress vector
on C .
2

The discrete problem is then: find uh Vh and H M H such that

h
h
H
H
h
h

(u ) : (v ) d vhn d + ( Rh (u ))Rh (v )d

C
C

= f vh d +
g vh d,

(H H )u d +
(H H )(H R (uh ))d 0,
C

hn

vh V h ,

(9)

H M H ,

where is defined to be constant on each element T as = 0 hT where 0 > 0 is independent of h and H.

Convergence analysis

Let us define for any v (H 1 ())2 and any L2 (C ) the following norm:
|(v, )| =

(v) : (v) d + 1/2 2L2 (C )

1/2

The usual norm of (H s (D))2 is denoted by s,D .


Now the following proposition gives an error estimate for the stabilized mixed finite-element approximation (9) of the unilateral contact problem. We first consider the case of nonpositive discontinuous
piecewise constant multipliers (7).
Theorem 4.1 Let (u, ) be the solution to problem (6). Assume that ur (H 2 ())2 . Let 0 be small
enough and let (uh , H ) be the solution to the discrete problem (9). Then, we have

1
1

u uh , H C hu us 2, + h 2 H 2 1 ,C + H 3/4/2 (u3/2, + 1 ,C ) .
2

Now, we focus on the case of nonpositive continuous piecewise affine multipliers given by (8).

Theorem 4.2 Let (u, ) be the solution to problem (6). Assume that ur (H 2 ())2 . Let 0 be small
enough and let (uh , H ) be the solution to the discrete problem (9). Then, we have

1
1

u uh , H C(hu us 2, + (H 2 + h1/2 ) 1 ,C + H 2 u3/2, ).


2

Numerical tests

= ]0, 1[ ] 0.5, 0.5[


The numerical tests are performed on a non-cracked square domain defined by
and the considered crack is the line segment C = ]0, 0.5[ {0}. Three degrees of freedom are blocked
in order to eliminate the rigid motions. In order to have both a contact zone
and a non-contact zone

0
between the crack lips, we impose the following body vector force f(x, y) =
.
3.5x(1 x)y cos(2x)
In the following figure, we present numerical tests of the stabilized and non stabilized unilateral contact
problem for the following, differently enriched, finite-element methods: P2 /P1 , P2 /P0 , P1 b/P1 , P1 /P1 ,
P1 /P0 .
The curves in the non stabilized case are given on the left of the figure. The P1 /P1 method is not plotted
because it does not work without stabilization. The P2 /P1 and P1 /P0 versions generally work without
stabilization even though an uniform inf-sup condition cannot be proven. Except the P1 /P0 method, the
rate of convergence in the L2 (C )-norm is optimal but there are very large oscillations. For the P1 /P0
method, the rate of convergence in L2 (C )-norm is not optimal (of order 0.42). It seems that there is a
locking of contact pressure at the crack.
On the right of the figure, we can see that the Barbosa Hughes stabilization eliminated the locking of
solution for the P1 /P0 method.

(a) without stabilization

(b) with stabilization

Figure 1: Error in L2 (C )-norm of the contact stress

Conclusion

In this paper we have formulated the contact problem on a crack of a domain. We have used the enrichment strategy introduced in [8] and the stabilized Lagrange multiplier method, namely the BarbosaHughess concept. We have obtained a priori error estimates following different discrete contact conditions and the numerical simulations with the C++ finite element library Getfem++ [9] are in good
agreement with the theoretical results. Indeed the convergence rates are optimal.

References
[1] H.J.C. Barbosa and T.J.R. Hughes, Circumventing the Babuska-Brezzi condition in mixed finite element approximations of elliptic variational inequalities, Computer Methods in Applied Mechanics and Engineering,
97: 193-210, 1992.
[2] S. Bordas and M. Duflot, A posteriori error estimation for extended finite elements by an extended global
recovery, International Journal for Numerical Methods in Engineering, 76: 1123-1138, 2008.
[3] S. Geniaut, P. Massin and N. Mos, A stable 3d contact formulation for cracks using XFEM, Revue Europenne
de Mcanique Numrique, Calculs avec mthodes sans maillage, 16: 259-275, 2007.
[4] A. Khoei and M. Nikbakht, Contact friction modeling with the extended finite element method (XFEM), Journal
of Materials Processing Technology, 177: 58-62, 2006.
[5] P. Hild, V. Lleras and Y. Renard, A residual error estimator for the XFEM approximation of the elasticity
problem, submitted.
[6] N. Mos, J. Dolbow and T. Belytschko, An extended finite element method for modelling crack growth with
frictional contact, Computer Methods in Applied Mechanics and Engineering, 190: 6825-6846, 2001.
[7] N. Mos, J. Dolbow and T. Belytschko, A Finite Element Method for Crack Growth without Remeshing,
International Journal for Numerical Methods in Engineering, 46 (1): 131-150, 1999.
[8] S. Nicaise, Y. Renard and E. Chahine, Optimal convergence analysis for the extended finite element method,
to appear in: International Journal for Numerical Methods in Engineering.
[9] J. Pommier and Y. Renard, Getfem++, an open source generic C++ library for finite element methods,
http://www-gmm.insa-toulouse.fr/getfem.
[10] J.J. Rodenas, O.A. Gonzales-Estrada and J.E. Tarancon, A recovery-type error estimator for the extended
finite element method based on singular plus smooth stress field splitting, International Journal for Numerical
Methods in Engineering, 76: 545-571, 2008.








10
9
8
7
6
5
4
3
2
1
0
1
2

1.5

0.5

0.5

1.5

900
800

900
h=3.3 mm, max=0.24mm, min=0.1mm;
h=1.2 mm, max=0.09mm, min=0.02mm
h=0.8 mm, max=0.06mm, min=0.02mm

700

800

1000

600

600
500

500

400

400

300

300

200

200

100

100
0
0

700

0.2

0.4

0.6

0.8

0
0

0.2

0.4

0.6

0.8

180

0.9

170

0.8
0.7

160

0.6

150

0.5

140

0.4
0.3

130

0.2

120
110
90

0.1

100

110



           
























 













 

  

  

 

 

















 






















XFEM 2011

II European Conference on eXtended Finite Element


Cardiff University, United Kingdom
Cardiff, June 29-30 July 1, 2011

Simulation of Shear Deformable Plates using Meshless Maximum Entropy Basis Functions
J. S. Hale1 , P. M. Baiz Villafranca2
1
2

Department of Aeronautics, Imperial College London, j.hale09@imperial.ac.uk


Department of Aeronautics, Imperial College London, p.m.baiz@imperial.ac.uk

Introduction

Plate theories such as the Mindlin-Reissner, or shear deformable, and Kirchoff, or classical, have seen
varied and wide use throughout engineering practice to simulate the mechanical response of structures
that are far larger in their planar dimensions than through their thickness [2]. Meshless methods such as
the EFG method [1] have been used to construct approximation spaces for the solution of the plate and
shell governing equations.
This extended abstract is designed to give background information to supplement the main presentation.
It begins with outlining the more mathematical aspects of the Reissner-Mindlin model, its relationship
with the Kirchoff model and the well-known shear locking problem. It then goes on to give a review
of existing methods to alleviate locking in the literature for both FE and Meshless methods. Finally
an overview of the meshless Maximum Entropy shape functions is given [9, 10]. Initial results to be
presented suggest that Maximum Entropy shape functions have the potential to match the performance
of MLS based methods whilst being significantly easier to implement.

The Reissner-Mindlin Plate Model

The Reissner-Mindlin Plate Model [2] can be derived from the full 3D equations of elasticity by making
a set of geometrical and kinematical assumptions. The key reason behind making these assumptions is
that we can trade integration over the domain R3 for a cheaper integration over the domain 0 R2
representing the mid-plane of the plate. The Reissner-Mindlin assumptions are as follows [2]:
Geometry The elasticity problem domain has one thin dimension, in this case in the x3 direction.
The plate mid-surface is then described by the domain 0 R2 and the thickness by a function
t : 0 (0, ). Therefore the whole domain R3 can be written as [2]:

(x1 , x2 , x3 ) R3 : (x1 , x2 ) 0 , x3 [t(x1 , x2 )/2, t(x1 , x2 )/2]


(1)

Mechanics Plane stress assumptions apply 33 = 0

Kinematics We assume that the displacement vector u : R3 can be written in the form [2]:
u(x1 , x2 , x3 ) = (z1 (x1 , x2 ) 1 (x1 , x2 )x3 , z2 (x1 , x2 ) 2 (x1 , x2 )x3 , z3 (x1 , x2 ))

(2)

Weak Form of the Reissner-Mindlin Plate Problem: Find the transverse deflection and rotations
(z3 , ) V3 R such that [2]:
t3

L() : () d + t

(z3 ) (y3 ) d =

p3 y3 d (y3 , ) V3 R

L[] D [(1 ) + tr()I]


E
D=
12(1 2 )
Ek
=
2(1 + )

(3a)
(3b)
(3c)
(3d)

where is Poissons ratio, E is Youngs modulus and k = 5/6 is a shear correction factor accounting
for the difference in shear energy produced by the assumed displacement field and the full equations
of elasticity. The exact specification of the Dirichlet or essential boundary conditions on d leads to
different function spaces for V3 and R .

Shear Locking in the Thin Plate Limit

The thin plate limit describes the asymptotic behaviour of the Reissner-Mindlin plate model as t 0.
On an intuitive level, we would assume that given that the Reissner-Mindlin theory describes thick plate
behaviour, as t 0 the solution for a given t (zt3 , t) should tend towards that of a model describing thin
plate behaviour, such as the Kirchoff model, with solution (z03 , 0 ).
We begin by re-writing Reissner-Mindlin plate problem in the following equivalent bilinear form with
loading p3 scaled with t3 to ensure the problem is well behaved as t 0 : Find the transverse deflection
and rotations (z3 , ) (V3 , R ) such that [2]:
a(, ) + t2 b(z3 , y3 ) = f (y3 ) (y3 , ) V3 R

(4)

The locking problem can be demonstrated as follows [2]. The obvious (but nave) choice is to introduce
the approximation space for the solution (zh3 , h ) of the piecewise linear Lagrangian finite elements P1
such that zh3 V3h V3 and h R h R . The clearest case is if we apply hard clamped boundary
conditions on d such that z3 = 0 x d and = 0 x d . We know that in the thin plate limit t 0
that h = z3h . If h = 0 on the boundary, then using P1 elements h must be zero everywhere in
. Thus z3h = 0 everywhere in also, and our finite element problem has converged to (0, 0). This
behaviour is known as shear locking.

Alleviating Locking

The vast majority of successful treatments of locking in the finite element literature are treated through
the application of mixed variational formulations. Popular techniques that can be viewed as mixed formulations include (but are not limited to) reduced integration methods and the Mixed Interpolation of
Tensorial Components (MITC) elements. In mixed formulations the shear stress vector = (xz , yz ) is
treated as an independent variable [2]:
= t2 (z3 )

(5)

giving the equivalent mixed problem as: Find the transverse deflection, rotations and transverse shear
stresses (z3 , , ) (V3 , R , S ) such that [2]:
2

a(, ) + b(, y3 ) = f (y3 )


(z3 ) d

d = 0 (y3 , , ) (V3 , R , S )

(6a)
(6b)

The above mixed formulation is known to be well-behaved in the limiting case as t 0, however at the
cost of introducing two extra variables representing the transverse shear stresses into the problem.
2

The MITC family [3] of elements works by defining a discrete reduction operator that takes values in the
discretised rotation space Rh R and maps them into the discretised shear space Sh [3]:
Rh : Rh Sh

(7)

a(h , h ) + t2 b(z3h Rh (h ), y3h Rh (h )) = f (y3 ) (y3h , h ) V3h Rh

(8)

so the original weak form can be re-cast in terms of the displacement variables only: Find the transverse
deflection and rotations (z3 , ) (V3h , Rh ) such that [3]:
In theory the operator Rh alleviates the shear locking by allowing the Kirchoff constraint to be satisfied
for choices other than = 0 and z3 = 0 as t 0. In practice this is achieved by a cleverly chosen tying
of the two transverse shear strains to the usual nodal displacements at the edge of the element [3].
In meshless methods a variety of approaches has been undertaken to eliminate locking in MindlinReissner plates. This overview is by no means exhaustive:
Higher Order Monomial Basis This approach is amongst the most widely used eg. [4, 5] to eliminate
locking. By increasing the monomial basis to higher orders m > 2 in the approximation, locking is
gradually eliminated because the displacement and rotation space V3 R can better approximate
the Kirchoff constraint z3h = h . However, locking is typically only eliminated satisfactorily
by using monomial basis of order m > 4 [4]. The higher order basis can be introduced either
intrinsically with meshless shape functions, or extrinsically, using PU concepts.
Matching Rotation and Displacement Spaces This approach was originally introduced in [6]. By approximating the displacement space V3 using the meshless shape functions i and the rotation
space R using the derivatives of the shape functions i then the Kirchoff constraint is met exactly. However as pointed out in [7] using the derivatives of the shape functions to approximate R
produces a linearly dependent stiffness matrix.
Mixed Methods It is possible to construct the spaces (Vh , Rh , Sh ) are constructed using MLS shape
functions solve for each field directly. This eliminates locking, however it produces non-positive
definite system matrices with more unknowns than a displacement approach.
Nodal Integration Techniques This technique can be seen as a form of reduced integration, and therefore shares some of the same issues such as spurious modes. These can be alleviated with techniques such as curvature smoothing [8].

Maximum Entropy Basis Functions

Maximum Entropy (MaxEnt) basis functions are one of the most recent developments in the construction
of meshless approximation schemes [9, 10]. A brief overview of their mathematical formulation and
properties is given here.
Consider a set of n nodes X. Each node i has a position xi Rd associated with it. The convex hull of
the set of nodes X is denoted D conv(X). For a function u(x) : D R, the numerical approximation
uh (x) can be written in terms of a set of shape functions i : D R and values ui at the nodes X [10]:
n

uh (x) = i (x)ui

(9)

i=1

Typically we wish the shape functions to satisfy the well known partition of unity condition as well as
first order consistency:
x D,

i (x) = 1

i (x)xi = xi

i=1

(10)

i=1

These two conditions alone are not enough to specify a unique approximation scheme. To this end,
Shannons concept of informational entropy is introduced. The Shannon entropy S(p) of a discrete
probability distribution with n events xi with probabilities pi is [ref]:
n

S(p) = pi ln pi
i=1

(11)

The principle of maximum entropy was proposed by Jaynes;


p

(12)

This analogy naturally implies that the shape functions are always positive:
i (x) 0 x D, i = 1, . . . , n

(13)

Therefore we can find a unique set of shape functions by maximising the entropy S() subject to the
constraints outlined above [9]:

max S() = i ln i

Rd+

i=1

i = 1

i=1

xD

i xi = x

i=1

xD

(14)

The objective function S() is strictly convex on RN+ and the two constraints are affine, so the above
problem can be solved using standard duality methods from the field of convex optimisation [10].
The above problem creates globally supported shape functions. To compute compactly supported shape
functions we use a more general form of entropy measure that includes a prior distribution wi that estimates i [10]:

n
i
H(, w) = pi ln
(15)
wi
i=1

The prior distribution can take the form of cardinal spline functions or Radial Basis Functions (RBF)
such as the Gaussian associated with each node.
The primary advantage of the MaxEnt scheme over standard MLS scheme is that it produces shape
functions with a weak Kronecker-delta property [10]. This makes the imposition of Dirichlet boundary
conditions trivial as in the Finite Element method without resorting to modified variational forms [12].

References
[1] T. Belytschko et al., Element-free Galerkin methods, International Journal for Numerical Methods in Engineering 37(2):229256, 1994.
[2] D. Chapelle and KJ. Bathe, The Finite Element Analysis of Shells: Fundamentals, Springer, 2003.
[3] KJ. Bathe, F. Brezzi, and S.W. Cho, The MITC7 and MITC9 Plate Bending Elements, Computers & Structures
32(3):797814. 1989.
[4] O. Garcia, E.A. Fancello, C.S. de Barcellos and C.A. Duarte. hp-Clouds in Mindlins thick plate model. International Journal for Numerical Methods in Engineering, 47(8): 13811400, 2000.
[5] Choi, Y. Kim, S., Bending analysis of Mindlin-Reissner plates by the element free galerkin method with
penalty technique. Journal of Mechanical Science and Technology, 17(1), 6476, 2003.
[6] B. M. Donning and W. K. Liu. Meshless methods for shear deformable beams and plates. Computer Methods
in Applied Mechanics and Engineering, 152(12): 4771, 1998.
[7] C. Tiago and V. Leitao. On the procedures to eliminate shear locking in meshless methods, ECCOMAS Thematic Conference on Meshless Methods 2005.
[8] D. Wang and J.S. Chen. Lockingfree stabilized conforming nodal integrationfor meshfree MindlinReissner
plate formulation. Computer Methods in Applied Mechanics and Engineering, 193(1214):10651083, 2004.
[9] M. Arroyo and M. Ortiz. Local maximum-entropy approximation schemes: a seamless bridge between finite
elements and meshfree methods. International Journal for Numerical Methods in Engineering, 65 :21672202,
2006.
[10] N. Sukumar and R. W. Wright. Overview and construction of meshfree basis functions: from moving least
squares to entropy approximants. International Journal for Numerical Methods in Engineering, 70 :181205,
2007.
[11] C. E. Shannon. A mathematical theory of communication. The Bell Systems Technical Journal, 27:379423,
1948.
[12] S. Fernandez-Mendez and A. Huerta. Imposing essential boundary conditions in mesh-free methods. Computer Methods in Applied Mechanics and Engineering, 193(12-14):12571275, 2004.

XFEM 2011

II European Conference on eXtended Finite Element


Cardiff University, Cardiff, United Kingdom,
June 29-30 July 1, 2011

Numerical calculation of physical fields in homogeneous media with


isolated heterogeneous inclusions
S. Kanaun, ITESM-CEM, Mexico, kanaoun@itesm.mx

Introduction

Calculation of static and dynamic fields of various nature (electrical, thermic, elastic) in heterogeneous
materials consisting on a homogeneous matrix phase and a set of isolated inclusions is an important problem of the material science. An efficient numerical method of the solution of these problems is based on
their reduction to volume integral equations. If such a reduction is possible, the fields inside the regions
occupied by the inclusions become principal unknowns of the problem. This is the main advantage of
the integral equation method over the Finite Element Method in application to the considering problems.
In the FEM, the fields in the matrix and in the inclusions are equivalent unknowns, and the solution is to
be constructed in the total region. Another class of the problems where the integral equations are more
appropriate tools for the numerical analysis is scattering of electromagnetic or elastic waves on isolated
inclusions in an infinite homogeneous media. A well-known drawback of the integral equation method
is that the matrices of the corresponding discretized problems are usually non-sparse and have large dimensions (if high precision of the numerical solutions is required), and the solution of linear algebraic
systems with such matrices has high computational cost.
In this work, a method of the numerical solution of a wide class of 3D-integral equations of mathematical
physics is proposed. For discretization of these equations, Gaussian radial approximating functions are
used. The theory of approximation by Gaussian and other similar functions were developed in the works
of Mazya and Schmidt (see the summary monograph [1]), and the usage of these functions has the
following advantages.
Actions of the integral operators of many integral equations of mathematical physics on such
functions are presented in closed analytical forms. Thus, the time of calculation of the elements
of the matrices of the discretized problems is essentially reduced in comparison with the methods
that incorporate conventional approximating functions and require numerical integration for the
calculation of the elements of these matrices.
Initial "geometrical" information required in the method is only the coordinates of the approximating nodes but not detailed sizes and shapes of the mesh cells. Thus, the method is meshfree in
fact.
For regular lattices of the approximating nodes, the matrix of the discretized problem has the
Toeplitz structure. Hence, the Fast Fourier Transform (FFT) technique may be used for the calculation of matrix-vector products in the process of iterative solution of the discretized problem.

Integral equation for the electric field in a homogeneous medium with


isolated inclusions

As an example, consider an infinite homogeneous medium with dielectric properties described by a tworank symmetric tensor Ci0j . The medium contains isolated inclusions that occupy a region V , and the
dielectric properties inside V are defined by the tensor Ci j (x), x is a point of the 3D-medium. The tensor
of the dielectric properties of the medium with the inclusions is presented in the form Ci j (x) = Ci0j +
Ci1j (x)V (x),where V (x) is the characteristic function of the region occupied by the inclusions: V (x) = 1
if x V , V (x) = 0 if x
/ V. Let an incident electric field E0 (x,t) = U0 (x)eit propagate in the medium
and be scattered on the inclusions. Here is frequency, t is time. In this case, the electric field E(x,t)
in the medium with the inclusions has the form E(x,t) = U(x)eit , and its amplitude U(x) satisfies the
following integral equation (see e.g., [2])
Ui (x)

Gi j (x x )C1jk (x )Uk (x )V (x )dx = Ui0 (x).

(1)

This equation serves for a set of inclusions as well as for an isolated inclusion in an infinite homogeneous
medium. For an isotropic matrix phase (Ci0j = c0 i j ), the kernel Gi j (x) of the integral operator in this
equation takes the form
Gi j (x) = k02 g(x)i j + i j g(x), g(x) =

eik0 |x|
, k 2 = 2 c0 .
4c0 |x| 0

(2)

Equation (1) was considered by many authors and is in essence the equation for the field U(x) inside
the inclusion (in the region V ). The field outside V may be reconstructed from the same eq (1) if U(x)
inside V is known. The existence and uniqueness of the solution of eq (1) was considered in [2]. Similar
integral equations may be written for static electric fields and static and dynamic elastic fields in the
medium with inclusions. In the last cases the unknowns are strain or stress tensors inside the inclusions.

Numerical solution of the integral equations (1)

For the numerical solution of eq (1), let us take a cuboid W that contains the region V occupied by the
inclusions and cover W by a regular grid of approximating nodes at points x(r) (r = 1, 2, ..., N). The
solution of eq (1) in W may be approximated by the Gaussian functions centered at the node points x(r)
of the grid

N
1
|x|2
(r)
Ui (x) Ui (x x(r) ), (x) =
exp

.
(3)
Hh2
(H)3/2
r=1
(r)

Here N is the total number of the nodes in W , Ui are unknown coefficients of the approximation, h is
distance between neighbor nodes, H is an undimensional parameter of the order 1. In [1], eq (3) is called
the "approximate approximation" because its error does not vanish when h 0. But the non-reducing
part of the error has the order of exp(2 H) and may be neglected in practical calculations.
Upon substitution of (3) into eq (1) we obtain:
N

Ui (x) i j (x x(r) )C jk Uk = Ui0 (x),


i j (x) =

1(r)

(r)

(4)

s=1

1(r)

Gi j (x x )(x )dx , Ci j

= Ci1j (x(r) ).

(5)

It follows from eq (2) for G(x) that the tensor (x) is presented in the form
i j (x) = k02 0 (x)i j + i j 0 (x), 0 (x) =

g(x x )(x )dx ,

(6)

where the function g(x) is defined in eq (2). The integral 0 (x) is calculate in a close analytical form as
follows ( = |x|/h, 0 = k0 h)

H2 + 4i0
h2
iH0 2
iH0 + 2

0 (x) =
exp 0
Er f c
e2i0 Er f c
.
(7)
8c0
4
2 H
2 H
2

Here Er f c (z) is the complimentary error function of a complex argument:


2
Er f c (z) = 1 Er f (z), Er f (z) =

et dt.

(8)

In the case of the electro static problem (0 0), the equation for 0 (x) is essentially simplified and
takes the form

h2

0 (x) =
Er f
.
(9)
4c0
H
(r)

The system of linear algebraic equations for the unknowns Ui in eq (3) follows from eq (4) if the latter
is satisfied at all the nodes x(r) (the Collocation Method). Note that if the nodes of the approximation
(r)
x(r) constitute a cubic grid, the coefficients of the approximation Ui coincide with the values of Ui (x) at
(r)
(r)
the corresponding nodes (Ui = Ui (x ), r = 1, 2, ..., N) (see [1]). As a result, we obtain the following
(r)
system of linear algebraic equations for the coefficients Ui
N

Ui i j C jk Uk = Ui
(r)

(r,s) 1(s)

(s)

0(r)

, r = 1, 2, ..., N;

(10)

s=1

(r,s)

i j

0(r)

= i j (x(r) x(s) ), Ui

= Ui0 (x(r) ),

(11)

where i j (x) is defined in eq (6). The matrix form of these equations is


(I B)X = F,

(12)

where I is a unite matrix of the order (N N), and the forms of the matrix B and the vectors of the unknowns X and of the right hand side F are reconstructed from eq (10). It is seen from eqs (8) and (9) that
the element of the matrix B are calculated fast. Nevertheless, the matrix B is non-sparse and may have
large dimensions if high accuracy is required. For algebraic systems with such matrices, only iterative
methods are efficient. Convergence of the minimal residue method in the case of the electromagnetic
wave scattering problem is proved in [2]. In this method, the vector of approximate solution is multiplied
by the matrix B at every step of the iteration process, and such a product is an expensive computational
operation. If, however, a regular grid of approximating nodes is used, the volume of the calculations is
reduced substantially. In this case, the matrix B has the Teoplitze structure, and the Fourier transform
technique can be used for the calculation of the matrix-vector products. Application of the fast Fourier
transform (FFT) algorithm essentially accelerates the iterative process. In addition, one has to keep in
the computer memory not all the matrix B of the dimensions (N N) but only one column and one row
of this matrix. The details of the FFT algorithm for the calculation of the matrix-vector products in the
case of equations similar to (10) are described in [4], [5].

Numerical results

The conversion of the method was studied in the case of electrostatic problems in [3], and for an electromagnetic wave scattering problem in [6]. The cases of 2D and 3D elasticity were considered in [4] and
[7]. Comparison with the exact and numerical solutions existed in the literature show that the method
converges for any finite contrast in the inclusion and matrix properties. The computational time depends
on the number of approximating nodes N and the contrast in the phase properties. For N about 106 107
the CPU time of a work station with the processor velocity 3GHz is about 1 hour for vector problems
and about 2 hours for tensor problems (elasticity).

Conclusions

The proposed method is an efficient tool for the numerical solution of a wide class of volume integral
equations of mathematical physics. It allows calculating electric, termo and elastic fields inside heterogeneous inclusions of arbitrary shapes embedded in a homogeneous medium. The problem of large number
3

of approximating nodes that appears in the case of 3D-problems is overcome by using Gaussian approximating functions and regular grids of approximating nodes when the FFT algorithms may be used by
the iterative solution of the discretized problem. The application of the method to the solution of the homogenization problems for composite materials of regular and random microstructures was considered
in [7], [8].
Yet another advantage of the method is that it is well suited for parallel computations. For instance, the
blocks of the B-matrix in eq (12) and products of these block- matrices with vectors may be calculated
independently and at the same time. This is another advantage of the method and a source of accelerating
the computational process.

References
[1] V. Mazya, G. Schmidt, Approximate Approximation, Mathematical Surveys and Monographs,141, American
Mathematical Society, Providence, 2007.
[2] A. B. Samokhin, Integral Equations and Iterative Methods in Electromagnetic Scattering, VSP, Utrecht,
Boston, Koln, Tokyo, 2001.
[3] S. Kanaun, S. Babaii Kochekseraii, A numerical method for the solution of 3D-integral equations of electrostatic theory based on Gaussian approximation functions, Applied Mathematics and Computation, 184:754768, 2007.
[4] S. Kanaun, Fast calculation of elastic fields in a homogeneous medium with isolated heterogeneous inclusions,
International Journal of Multiscale Computational Engineering, 7(4): 263-276, 2009.
[5] G. Golub, C. Van Loan, Matrix Computations, The Johns Hopkins University Press, 3d ed, 1996.
[6] S. Kanaun, Scattering of monochromatic electromagnetic waves on 3D-dielectric bodies of arbitrary shapes,
Progress In Electromagnetics Researche B, 21: 129-150, 2010.
[7] S. Kanaun, E. Pervago, Combining self-consistent and numerical methods for the calculation of elastic fields
and effective properties of 3D-matrix composites with periodic and random microstructures, International
Journal of Engineering Sciences, 49(5): 420-442, 2011.
[8] S. Kanaun, Calculation of electro and thermo static fields in matrix composite materials of regular or random
microstructures, International Journal of Engineering Sciences, 49(1): 41-60, 2011.

XFEM 2011

II European Conference on eXtended Finite Element


Cardiff University, United Kingdom
Cardiff, June 29-30 July 1, 2011

An eXtended Finite Element Method for Maxwells equations


F. Lefvre1 , S. Lohrengel1 , S. Nicaise2
1

Universit de Reims Champagne-Ardenne, France, {francois.lefevre,stephanie.lohrengel}@univ-reims.fr


de Valenciennes et du Hainaut Cambrsis, France, serge.nicaise@univ-valenciennes.fr

2 Universit

Introduction

EXtended Finite Element Methods have gathered much interest in the domain of fracture mechanics in
the last ten years since they are able to simulate the behaviour of the displacement field in regions with
cracks using meshes that do not need to fit the crack geometry.
In this communication, a new eXtended Finite Element Method is presented based on edge finite elements which are commonly used in the discretization of Maxwells equations (see e.g. [4]). Contrary
to classical eXtended Finite Element Methods which are based on nodal Lagrange Finite Elements, the
degrees of freedom for edge finite elements are related to the edges (and not to the nodes) of the mesh.
The simulation of the electromagnetic field near cracks is important in electromagnetic testing which
is a form of non destructive testing in order to detect defects in conducting structures, as for example
aircraft wings. Classical edge finite elements on a geometry fitting mesh are able to handle this situation
provided the mesh is sufficiently refined near the crack tip. This results generally in heavy computations.
Here, according to the XFEM strategy, we take into account the singularity and the discontinuity of the
field in the discretization space defined on a structured mesh independent of the crack geometry.

2
2.1

EXtended Finite Edge Elements


The model problem

In this communication, we focus on a simple model problem which is given in a two-dimensional translation invariant setting. The notations curl and curl distinguish between the scalar and vector curl operators, namely,

E2 E1

t
curl E =

and curl =
,
x1 x2
x2 x1
for any field E(x) = (E1 (x1 , x2 ), E2 (x1 , x2 ))t and any scalar function = (x1 , x2 ).
In a time-harmonic setting, the variation in time of all fields is given by eit , where > 0 is the
frequency of the vibrations. In terms of the electric field, the problem reads after elimination of the
magnetic field,
curl 1 curl E 2 E = iJ
in ,
(1)
where and denote, respectively, the magnetic permeability and the electric permittivity, is the
materials conductivity and J represents the applied current density. The parameter is defined by
2
2
= + i
. The domain R is the region of interest containing the crack. More precisely, let Q R
be an open set and = Q its boundary. Consider a closed segment = {sx + (1 s)x | s [0, 1]} in
Q, called the crack, such that x Q and x . Then, is defined as the domain outside the crack, i.e.
= Q \ .

We supplement (1) with the following boundary conditions


E n = 0 on ,

1 curl E n = 0 on ,

(2)
(3)

)t

where n = (n1 , n2 denotes the outer unit normal vector at the boundary of . Here, (2) is the condition
of a perfect conductor outside Q and (3) amounts to saying that no current passes through the crack.
In order to perform a discretization of (1) by means of finite elements, let us make precise the functional
framework of the variational formulation. Let L2 () be the space of square-integrable functions on
with norm 0, and scalar product (, ). For m N, we denote by H m () the standard Sobolev-space
of functions with square-integrable derivatives up to the order m.
The natural space of fields with finite electromagnetic energy is given by

H (curl; ) = u L2 ()2 | curl u L2 () .


Taking into account the boundary condition (2), we are thus led to seek the solution E in

H0, (curl; ) = u H (curl; ) | u n = 0 on ,

such that

(1 curl E, v) 2 ( E, v) = i(J, v) v H0, (curl; ).

(4)

The strategy of XFEM needs to make precise the behaviour of the solution of (4) near the crack. Notice
that in general two kinds of singularities occur: one near the crack tip, the other near x . For the sake
of convenience, we will assume that the only singularities which have to be taken into account, are those
near the crack tip. This is the case for example for small opening angles at x or when the crack hits at
a right angle. We refer to [3] for the general case.
The following theorem relies partially on results from [1]. It has been proven in [3] in a slightly different
form where the regular part Er of the electric field only belongs to H 1 ()2 .
Theorem 1 Assume that J L2 (Q)2 such that div J L2 (Q). The solution E of problem (4) satisfies
div E L2 () and E n = 0 on .

Further, E splits into a regular part Er H 2 ()2 and a singular part deriving from scalar potentials:
E = Er +

cm (Sm ).

(5)

m=1

Here, = (r) denotes a regular cut-off function with respect to x depending only on the distance r to
the crack tip x . The singular functions Sm are defined by

(2m 1)
Sm (r, ) = r(2m1)/2 sin
2
in local coordinates (r, ) near the crack tip.

2.2

Discretization

Following [3], we now define the discretization space for XFEM-edge elements. Whereas nodal finite
elements of type P1, Q1, etc. are related to the nodes of a mesh, classical edge elements are based on
its edges. Hence, consider a triangulation Th of the polygonal domain Q and denote E the set of its
(oriented) edges. Notice that Th is independent from the crack configuration. With a given edge e E ,
we associate the linear form

le (v) = v te ds
(6)
e

where te denotes the unit vector in the direction of e. The local approximation space of classical twodimensional edge elements is given by

x2
RK = a + b
| a R2 , b R
x1
2

on any element K of the mesh. Notice that RK is a proper subspace of the space of vector-valued
polynomials of degree less or equal to 1. We then define the space of classical edge elements by XFE
h =
Vect(we | e E ) where we are the shape functions associated with the degrees of freedom (6) and the
space RK . It is worth noticing that classical edge elements lead to a conforming discretization, meaning
that XFE
h H (curl; Q).
Following the ideas of nodal XFEMs [2], we introduce the set EH of enriched edges: an edge e belongs to EH if the crack passes through the support of the corresponding shape function we . The
discretization space of the XFEM-edge elements is then defined as follows:
XXFEM
= XFE
h
h Vect(Hwe | e EH ) Vect((Sm ) | m = 1, 2),

(7)

where H = H(x) denotes a function of Heaviside-type with respect to the crack . According to the
enrichment strategy, we can prove that the discretization by means of XFEM-edge elements is still conforming:
XXFEM
H (curl; ).
h

Further, we deduce from the estimates in [3] that the method based on (7) converges with optimal rate

O (h). More precisely, under the assumptions of Theorem 1, and if Eh XXFEM


denotes the approximah
tion of the solution of (4) by XFEM-edge elements, we have

E Eh 0, + curl (E Eh ) 0, ChEr 2, .

(8)

Here, C > 0 denotes a constant independent from E and the mesh size parameter h, and 2, is the
standard norm in the Sobolev-space H 2 ()2 .

2.3

Numerics

In this section, we will provide several numerical tests which illustrate the performance of XFEM-edge
elements. The non-cracked domain Q is chosen to be the square [0.5, 0.5] [0.5, 0.5] and is discretized with a stuctured mesh. We consider a crack supported by the x1 -axis, = [, 0.5] {0} with
104 . This allows to test the method in the most current configuration where the crack tip lies in
the interior of a single triangle. Further, no singular behaviour of the electric field occurs at the point x
since hits the boundary under a right angle. We choose to take into account only the first singular
field, (S1 ); the generalization to two or more singular fields is straightforward. Then, the discrete
problem can be written as a block linear system,


A C
Ur
F

= .
(9)
C T as
us
fs
Here, A and F denote, respectively, the (sparse) matrix and the right-hand side corresponding to finite
element shape functions {we | e E } {Hwe | e EH }. C is a column matrix coupling finite element
shape functions and the singular field (S1 ). Finally, as and fs are terms involving only the singular
field and the given data J (for fs ). The block structure of the linear system (9) allows to rewrite the
discrete problem in order to solve two linear systems involving the same sparse matrix A:
AU = F
AS = C

us =

fs Ct U
as Ct S

Ur = U us S

We have tested the XFEM-edge elements in various situations where an exact solution is known. In
Figure 1, the performance of XFEM-edge elements is compared to a discretization by classical edge
elements on a geometry-fitting mesh. The electromagnetic parameters of this academic
testcase are

= 1, = 1, = 1 and = 1. Here, the singular solution was given by E(r, ) = r1/2 sin 2 and has
the strongest singularity and the regular solution refers to a solution with H 2 -regularity. We notice that
in all cases, the XFEM-edge elements converge with nearly optimal rate. They perform much better than
classical edge elements in the presence of strong singularities (which is the general case). In the case of
a regular or nearly regular solution, the influence of the cut-off function and its well-known "pollution
3

Singular solution

0.5

edgeFEM (slope = 0.55)

edgeFEM (slope = 1.06)

1.2
1.3

log10(error)

0.8
0.9

10

edgeXFEM (slope = 0.96)

1.1

0.7

log (error)

Regular solution

edgeXFEM (slope = 0.89)

0.6

1
1.1

1.4
1.5
1.6
1.7

1.2

1.8

1.3

1.9

1.4
1.8

1.6

1.4

1.2

log10(h)

0.8

0.6

2
1.8

0.4

1.6

1.4

1.2

log10(h)

0.8

0.6

0.4

Figure 1: Convergence rate of classical and eXtended edge elements.


effect" leads to slightly higher constants in the error estimate. Variants of the actual edge-XFEM could
probably overcome these undesirable effects.
The second test is an eigenvalue problem. It does concern the case of a cavity resonator containing a
crack and consists in seeking > 0 and E, solution of (4) with J = 0 and = 0 under the additional
constraint that div E = 0 in . Figure 2 shows that the results of XFEM-edge elements are in good
concordance with those of the reference method.
XFEMedge elements

25

20

eigenvalues

eigenvalues

20

15

10

0
0.1

Classical edge elements

25

15

10

0.15

0.2

0.25

0.3

mesh size

0.35

0.4

0
0.1

0.45

0.15

0.2

0.25

0.3

mesh size

0.35

0.4

0.45

Figure 2: Computation of eigenvalues by classical and eXtended edge elements.

Conclusion

In this communication, we have presented a new eXtended Finite Element Method based on edge finite
elements to discretize the two-dimensional time-harmonic Maxwell equations. The method allows to
obtain optimal convergence rate and was shown to yield performing results both in the case of a source
problem and the eigenvalue problem. The adaptation of the method to more involved models in electromagnetism as for example eddy current or transient problems, should be straightforward. The extension
to a three-dimensional setting needs an investigation of the singularities near the crack, and will be part
of future work.

References
[1] M. Costabel and M. Dauge, Singularities of electromagnetic fields in polyhedral domains, Arch. Rational
Mech. Anal., Vol. 151: 221276, 2000.
[2] N. Mos, J. Dolbow, and T. Belytschko, A finite element method for crack growth without remeshing, Int. J.
Numer. Meth. Engng., 46: 131150, 1999.
[3] F. Lefvre, S. Lohrengel, and S. Nicaise, An eXtended Finite Element Method for 2D edge elements, submitted.
[4] J.-C. Ndlec, Mixed finite elements in R3 , Numer. Math., Vol. 35: 315341, 1980.

XFEM 2011

II European Conference on eXtended Finite Element


Cardiff University, United Kingdom
Cardiff, June 29-30 July 1, 2011

Computational homogenization for multiscale crack modelling


V.P. Nguyen1 , O. Lloberas-Valls1 , M. Stroeven1 , L.J. Sluys1
1

Delft University of Technology, Faculty of Civil Engineering and Geosciences, The Netherlands,
{V.P.Nguyen,O.LloberasVall,M.Stroeven,L.J.Sluys}@tudelft.nl

1 Introduction
Computational homogenization (CH) [1] has emerged as a valuable tool to model heterogeneous materials in which all relevant heterogeneities at the fine scale can be taken into account to model structures at
the coarse scale. In this contribution, we present a continuous-discontinuous CH scheme to model failure of heterogeneous quasi-brittle materials. According to this CH scheme, constitutive behaviour of the
bulk material as well as that of the cohesive crack are determined from nested microscopic finite element
computations. A macro-crack is initiated when the microscopic sample shows strain localization. The
method is an extension of the discontinuous CH scheme reported in [2]. The continuous-discontinuous
homogenization scheme is more accurate than the discontinuous CH scheme for problems in which energy dissipated prior to microscopic localization is significant. Herein, we only introduce the main ideas
of the method. Details can be found in [3].

2 A continuous-discontinuous CH method
2.1 Macroscale and microscale models
t
tM
tM
dM

uM
Figure 1: A two dimensional heterogeneous solid containing a cohesive crack.
Let us consider a two-dimensional solid M shown in Fig. 1. The solid contains a cohesive crack dM .
The discrete equation for quasi-static equilibrium reads
int
bulk
coh
fext
M fM fM fM

fext
M

where
is
fbulk
M and the

the external force vector, fint


M is the internal force
cohesive force vector fcoh
M . They are given by

fbulk
M

BT M d
1

fcoh
M

(1)

vector that consists of the bulk force vector

dM

NT tM d

(2)

where M is the macroscale Cauchy stress tensor, and tM is the cohesive traction across the crack dM .
To model cohesive cracks independently of the finite element discretization, we use the phantom node
method [5]. Only Heaviside enrichment is used, the cracks thus grow element-wise.
Behaviour of the macroscopic bulk is given by the following multiscale constitutive equation
M M m

M DM : M

(3)

where M is the macroscale strain tensor, m is the microscale stress tensor and DM , a fourth order tensor,
is the homogenized bulk tangent. Computation of M and DM given M (at a bulk integration point) is
done via the so called bulk/continuous computational homogenization procedure in [4].
The behavior of the macroscale cohesive cracks is described using a multiscale cohesive law that reads
tM uM m

tM TM uM

(4)

where uM is the displacement jump across the macro-crack and TM , a second order tensor, is the
homogenized cohesive tangent. Providing the displacement jump uM at one cohesive integration point,
the corresponding traction tM and cohesive tangent TM are computed using the crack homogenization
procedure given in [2].
The microscale model is a w h rectangular domain where the heterogeneities are explicitly resolved,
see Fig. 1. Microscopic failure is modelled with a non-local isotropic damage model. We refer to [3] for
details.

2.2 Micro-macro coupling


Figure 2 summarizes the procedure of the continuous-discontinuous CH scheme for three node triangle
elements. When localization is detected at one bulk RVE, a macro-crack is inserted at the corresponding
macro-element. Given the jump uM , to compute tM , the following system of equations
ext
fint
m fm uM

(5)

uR w lC tM uM u dam

(6)

ext
fint
m , fm

is solved. In the above,


are the microscopic internal and external force vectors; l is the averaged
width of the microscopic localization band; C is the projection on the crack plane of a fourth-order compliance tensor which is determined as the homogenized bulk tangent of the localized RVE that subjects
an unloading path [3]; uR is the displacements of the right edge of the RVE and u dam is a compatibility displacement. The first term of the RHS of Eq. (6) represents the linear part of the microscopic
displacement.

3 Numerical examples
3.1 RVEs existence test
A uniaxial tension bar discretized into three bilinear four node quadrilateral elements is considered, Fig.
3. Three micro models of size 10 10, 15 15 and 20 20 mm2 are investigated. The volume fraction
of aggregates, its range of radius, thickness of the ITZ and material parameters of the microstructural
constituents can be found in [3]. A plane strain condition is assumed. The existence of an RVE is
verified by the convergence of the load-displacement response with increasing the size of the microsamples. Figure 4 plots the load-displacement curves and the corresponding homogenized tractionseparation curves. As can be seen, increasing the size of the micro-samples makes the macroscopic
responses converging. Note that the 10 10 mm2 sample is too small to be statistically representative for
the microstructure, that is why its solution deviates somewhat from the solutions of the larger samples.
It is emphasized that for random heterogeneous materials the load-displacement curves cannot be on top
of each other or very close to each other.

cohesive RVEs

M Dun M

bulk GP

cloned

cohesive GP
bulk RVE

localized RVE

Figure 2: The fully coupled multiscale homogenization scheme. Initially bulk Gauss points (GPs) are
coupled with bulk RVEs. When localization occurs in one bulk RVE, a macro-crack is inserted at the
corresponding macro-element. Henceforth, cohesive GPs are coupled to RVEs that are cloned from the
localized RVE. The bulk GPs of the cracked macro-element follow a secant unloading path. Dotted lines
denote periodic boundary conditions.

20 20
15 15
100

10 10

10

10

100

Figure 3: RVE existence test: macro mesh (left) and micro samples (right). Length unit is mm. The
dotted line is the crack to be inserted upon satisfaction of a failure criterion.

16

10x10
15x15
20x20

0.12
traction [N]

12
load [N]

0.16

10x10
15x15
20x20

0.08

0.04

0.1

0.2

0.3

0.4

displacement [mm] (x 10-2)

0.1

0.2

0.3

0.4

jump [mm] (x 10-2)

(a) load-displacement curves

(b) homogenized traction-jump curves

Figure 4: Objectivity of the macroscopic response with respect to different adopted micro-samples.

10

rigid plate

10

10 10

Figure 5: FE meshes of the multiscale crack model: macro-mesh (left) and micro-mesh (right).
1.50

reference
original FE2
full FE2

1.25

load [N]

1.00
0.75
0.50
0.25
0.00
0.0

0.5

1.0

1.5

2.0

displacement [mm] (x10-3)

Figure 6: Reference solution, the original multiscale solution and the new fully multiscale solution.

3.2 Verification test


Considering the example given in Fig. 5. The continuous-discontinuous CH solution is denoted by a
full FE2 scheme while the discontinuous CH solution [2] is denoted by the original FE2 scheme. The
reference solution is obtained with the continuous CH scheme [4]. Figure 6 shows that the proposed
continuous-discontinuous CH scheme can capture the reference continuous solution while the discontinuous CH method cannot.

4 Conclusion
We presented a continuous-discontinuous CH method for failure modelling of random heterogeneous
quasi-brittle material. The proposed CH scheme allows to define a representative volume for random
heterogeneous quasi-brittle materials that show strain localization.

References
[1] P. M. Suquet, Local and global aspects in the mathematical theory of plasticity, In A. Sawczuk and
G. Bianchi, editors, Plasticity today: modelling, methods and applications, pages 279310, London,
1985. Elsevier
[2] V.P. Nguyen, O. Lloberas-Valls, M. Stroeven and L.J. Sluys, Homogenization-based multiscale crack
modelling: from micro-diffusive damage to macro-cracks, Computer Methods in Applied Mechanics
and Engineering, 200(9-12): 1220-1236, 2011
[3] V.P. Nguyen, M. Stroeven and L.J. Sluys, A continuous-discontinuous multiscale method for modelling cohesive cracks in random heterogeneous quasi-brittle materials, Engineering Fracture Mechanics, 2011. Submitted for publication.
[4] V. Kouznetsova, W. A. M. Brekelmans and F. P. T. Baaijens, An approach to micro-macro modeling
of heterogeneous materials, Computational Mechanics, 27(1):3748, 2001.
[5] J.H. Song, P.M.A. Areias and T. Belytschko, A method for dynamic crack and shear band propagation with phantom nodes, International Journal for Numerical Methods in Engineering, 67(6):868
893, 2006.











a) fibre enrichement

b) fibrematrix bond relation

matrix

thin fibre

mechanical
system

elastic

debonding

frictional

(stmax , ttmax )

tt

tt

st

(stres , ttres )

enrichment
function

unloading branch
dfib

st

dogbone shaped specimen

fibre distributions

20

100

distribution 1

20

140

clamped into
the supports

100

distribution 2

45

70 45

70

[mm]


forcedisplacement plot

damage plots

25

plain matrix

distribution 1

distribution 2

F [kN]

20
15
10
5
0

0.2

0.4

experimental
plain matrix
distribution 1
distribution 2
0.6 0.8 1 1.2 1.4
u [mm]

1.6

0.95
0

0.5

75

11.25

reinforcement

50
18.75

150

150

18.75

[mm]

0.13% smooth reinforcement

3.5

experimental
numerical

3.5

F [kN]

F [kN]

2
1.5

0.5

0.5
0

0.05

0.1

0.15 0.2
u [mm]

0.25

0.3

0.35

0.26% smooth reinforcement

0.05

0.1

0.15 0.2
u [mm]

0.25

0.3

0.35

0.3

0.35

0.26% ribbed reinforcement

experimental
numerical

3.5

4.5
4

3.5
F [kN]

2.5
F [kN]

2.5

1.5

2
1.5

3
2.5
2
1.5

experimental
numerical

0.5
0

experimental
numerical

2.5

0.13% ribbed reinforcement

4.5

0.5
0

0.05

0.1

0.15 0.2
u [mm]

0.25

0.3

0.35

0.05

0.1

0.15 0.2
u [mm]

0.25

XFEM 2011

II European Conference on eXtended Finite Element


Cardiff University, United Kingdom
Cardiff, June 29-30 July 1, 2011

A Fast and Accurate Solver base on Cartesian Grids for Shape Optimization Problems
J.J. Rdenas, E. Nadal, J. Albelda and M. Tur
Centro de Investigacin en Tecnologa de Vehculos (CITV)
Universidad Politcnica de Valencia, E-46022-Valencia, Spain
jjrodena@mcm.upv.es, ennaso@upvnet.upv.es, jalbelda@mcm.upv.es, matuva@mcm.upv.es

Introduction

Optimization processes are divided into two levels. The higher level, controlled by the optimization
algorithm and the lower level controlled by a numerical method, such as the FEM in structural shape optimization. This one is used to evaluate the objective value and the level of satisfaction of the restrictions,
with the required accuracy to ensure the convergence of the optimization process.
The discretization error of each individual must be controlled to ensure the convergence of the optimization process [1]. This justifies the use of adaptive procedures to ensure that the prescribed accuracy level
is obtained with the lower computational cost. This paper proposes the use of a a Cartesian Grid (CG)
FE code to reduce the mesh burden and to allow for the use of efficient hierarchical data structures [2].
These techniques have an important disadvantage: there is a considerable reduction of the stress accuracy
along the boundary [3]. To account for this we have use the recovery technique SPR-C [4]. We also
propose the use of the recovered fields in the Zienkiewicz & Zhu error estimator [5] to guide the
h-adaptive process.
In an optimization process some parts of the geometry do not change along the optimization process.
Since a CG code has been used it is easy to share information between individuals, reusing information
previously evaluated in other individuals.

2
2.1

Cartesian Grid (CG) FEM


General aspects

A CG solver is characterized by the mesh-geometry independence. The analysis procedure starts intersecting the mesh with the geometry in order to capture the geometrical information. In this process,
elements are classified into three groups in relation with the boundary: a) external which are not considered in the FE analysis, b) internal which are considered as in a standard FE analysis and c) intersected
which have a part into the domain. To accurately represent the geometry in intersected elements, we apply a Delaunays triangulation at intersected elements to generate integration subdomains. The number
of triangles depends on the curvature of the boundary .
Because of the regularity of CGs and the use the hierarchical structure described in [2], stiffness matrix,
elementary force vector, intersection and integration information may be shared between all internal
elements. Boundary elements repeated in the h-adaptive analysis also share its information.

2.2

Refinement procedure

The implementation uses two refinement criteria: a geometrical criterion and a solution-based criterion.

With the geometry-based refinement the size of the mesh is automatically adjusted to the curvature of
the boundary to create the first FE analysis mesh, as represented in Figure 1.
Initial mesh

1st analysis mesh,


geometrically adapted

Initial mesh

1st analysis mesh,


geometrically adapted

Figure 1: Comparison between the initial mesh and the mesh geometrically adapted.
Once the mesh is geometrically adapted, the h-adaptive process considers the minimization of the error
in energy norm to refine the mesh. This error is estimated by using the ideas of Zienkiewicz & Zhu [5]:

ees 2

h T D1 h d

(1)

where h is the stress field obtained from the FE solution, is the the recovered solution and D is the
Hooks tensor.

Stress recovery procedure

In order to improve the accuracy level, mainly along the boundary, we use the SPR-C technique. This
improved technique enforces the satisfaction of the equilibrium and compatibility equations in the recovered stress field. This fact guaranties a high quality recovered field, even at the elements along the
boundary. This field will be used both for the h-adaptive process and as the standard stress field output.
Since the meshes are regular, only a small amount of patch configurations is possible. In a 2D analysis,
there are only 19 possible configurations for internal patches. This fact allows us to pre-calculate these
19 systems of equation for all internal patches and, thus, to essentially reduce the computational cost of
the recovery process only to the contour patches.

Use of informations between individuals

One of the characteristics of the CG-FEM is its independence of the geometry. This allows us to use
the same initial mesh for all individuals in the recovery process. Elements located over a boundary
curve that was set to fixed during the optimization process, may be repeated in several geometries.
This allows us to share the intersection information, the stiffness matrix, the integration information,
the elementary force vector, etc... between these elements, thus avoiding reevaluations of previously
calculated information.

Numerical examples

R
Some optimization problems have been carried out with the implemented solver (CG-FEM) and ANSYS
11 as a solver. In order to control the global accuracy of the solution it is defined the relative error in
energy norm as follows:
ees 2
2
(2)
2
h
u ees 2

where uh is the energy norm:
h 2
u h T D1 h d
(3)

5.1

Pipe under internal pressure

In order to check the proposed methodology, an optimization procedure of a problem with known solution
has been performed. This problem is a pipe under internal pressure. The data, the design variables and
the analytical solution is shown in Figure 2. The external boundary
is represented by using a cubic spline,

defined by points 1, 2, and 3. In this Figure we define r x2 y2 , arctanyx y k R0 a, where


x y are global coordinates. Considering a yield stress Sy 2 106 , the optimal shape is an external
circular surface of radius R0 1067033824461. The optimal area is Aopt 69787307715081.

E 101 106 03 a 5 P 09 106

R2
P1
ur
1 2r 0
2
Ek 1
r

R20
P
r 2
1 2
k 1
r

R20
P
t 2
1 2
k 1
r

xx r cos2 t sin2

yy r sin2 t cos2

xy r t sincos
P
z 2 2
k 1

Figure 2: Pipe under internal pressure. Model and analytical solution.


We ran the problem for several accuracy levels (1, 2, 5, 10)% to check the influence of the discretization error in the area error. Figure 3a shows the evolution of the caculated area in the optimization
process where the black line is the exact optimal solution. There is a relation between the prescriber error
and the area error.
150

CG-FEM sharing
CG-FEM
R
ANSYS
11

Time (s)

Area

100

104

1%
2%
5%
10%

50

2
1

100

50

150

Iteration

100

50

150

Iteration

b) Computational cost of the proposed methodology


R
(CG-FEM sharing and CG-FEM) and ANSYS
11

a) Area evolution during the optimization process for several values of

Figure 3: Pipe under internal pressure. Area and time evolution in the structural optimization process.
Figure 3b displays analysis times using different strategies. The figure shows the time improvement
R
when a CG solver is used instead a standard FE solver such as ANSYS
11 and the effect of sharing
information between different individuals. Observe that a reduction of a 52% in comparison with the
R
time required time by ANSYS
11 is finally obtained. This is more significant if we take into account
R
that the proposed methodology has been fully implemented in Matlab
2009b-64bit.

Conclusions

We have used an CG-FEM solver with a hierarchical data structure to analyze the individuals proposed
by an evolutionary optimization algorithm. Information is easily shared and reused by means of the
proposed methodology allows, which also reduces mesh burden. This considerably reduces the total
computational cost of the FE analysis and, hence, of the optimization process. A comparison with a
commercial FE code has clearly shown these advantages. The quality of the recovered stress fields
obtained by the SPR-C technique has adequately driven the optimization process.

Acknowledgments

This work has been carried out within the research project DPI2010-20542 of the Ministerio de Ciencia
e Innovacin (Spain). The financial support of the FPU program (AP2008-01086), the funding from
FEDER, Universitat Politcnica de Valncia and Generalitat Valenciana are also acknowledged. s

References
[1] G. Bugeda, JJ. Rdenas, J. Albelda and E. Oate, Control of the finite element discretization error during the
convergence of structural shape optimization algorithms, International Journal for Simulation and Multidisciplinary Design Optimization, 3 (3):363-369, 2009
[2] JJ. Rdenas, JE. Tarancn, J. Albelda, A. Roda and FJ. Fuenmayor, Hierarchical properties in elements obtained by subdivision: a hierarchical h-adaptivity program, Adaptive Modeling and Simulation, In P Dez and
NE Wiberg, pages 420-435, 2005
[3] E. Burman and P. Hansbo, Fictitious domain finite element methods using cut elements: I. A stabilized Lagrange multiplier method, Preprint, 2009
[4] E. Nadal, S. Bordas, JJ. Rdenas, JE. Tarancn and M. Tur, Accurate Stress Recovery for the Two-Dimensional
Fixed Grid Finite Element Method, Proceedings of the Tenth International Conference on Computational
Structures Technology, paper 256, 2010
[5] OC. Zienkiewicz and JZ. Zhu, A simple error estimation and adaptive procedure for practical engineering
analysis, International Journal for Numerical Methods in Engineering, 24:337-357, 1987

XFEM 2011

II European Conference on eXtended Finite Element


Cardiff University, United Kingdom
Cardiff, June 29-30 July 1, 2011

Two-phase flows using the XFEM: Modeling of transport phenomena


on fluid interfaces
H. Sauerland1 , T.-P. Fries1
1

Chair for Computational Analysis of Technical Systems, CCES, RWTH Aachen, Germany, {sauerland,fries}@cats.rwth-aachen.de

1 Introduction
The modeling of transport phenomena on arbitrary and even moving interfaces is a topic of increasing
interest in the scientific community. It is directly related to the solution of PDEs on hypersurfaces
(manifolds). The field of applications for such kind of problems is widespread and can be found in
fluid mechanics [2], computer graphics [3], material science [4] and biology [5]. We are considering the
transport of insoluble surfactants on two-fluid interfaces in this paper.

Flow field

Advection/deformation
of the interface

Variable surface tension

Surfactant transport
on the interface

Figure 1: The interaction of the different fields involved.


Figure 1 depicts the considered situation schematically. The flow field and interface representation are
modeled in a coupled manner. Both influence the transport of the surfactants on the interface. Surfactants
give rise to gradients in the surface tension. This leads to tangential forces along the interface and
thereby couples the surfactant transport with the flow field. Though, in this work, we will not consider
an influence of the surfactants on the flow field.
An important prerequisite for the accurate modeling of such interface transport phenomena is an accurate
description of the flow field and the interface movement. Hence, we apply the extended finite element
method (XFEM) [1] in order to account for the discontinuous nature of the field variables in two-phase
flows. Thereby, we obtain an accurate implicit interface description based on the level-set method. The
interface transport is modeled with an implicit approach, too.

2 Modeling of two-phase flows


The fluid velocity and pressure p in both phases j 1 2 are governed by the instationary, incompressible and isothermal Navier-Stokes equations:


in j 0 T
(1)
j
t
0
in j 0 T
(2)
Here, j and are the density of the
respective fluid
phase and the external body forces. The stress tensor
is defined by p p j T , with j being the dynamic viscosity of the respective
fluid phase.
The interface d is implicitly described using the level-set method. We define the scalar level-set field
as a signed-distance function, i.e. the level-set value at an arbitrary point is the shortest distance to the
interface. In order to account for the interface motion the level-set transport equation is solved:

t 0 in 0 T
t

(3)

Due to the density and viscosity differences between the phases, a kink is present in the velocity and
pressure field across the interface. Additionally considering surface tension effects even leads to a jump
in the pressure. As a consequence of the implicit interface description, these discontinuities across
the interface occur inside elements. The XFEM offers the possibility to appropriately account for this
situation by locally enriching the pressure approximation space:
ph Ni pi Ni t i t ai
iI

(4)

iI

ph is the approximated pressure field, Ni is the standard finite element function for node i, pi is
the unknown of the standard finite element part at node i, I is the set of all nodes in the domain, t is
the global enrichment function, ai is the unknown of the enrichment at node i and I is the nodal subset
of the enrichment. We are using the so-called sign-enrichment:

1 : 0
t signt
(5)
0 : 0

1 : 0

3 Scalar transport on the interface


The mathematical description of transport processes is rather simple. However, solving the resulting
PDEs on arbitrary, moving manifolds is not straightforward. There exist explicit and implicit approaches,
see e.g. [6,9]. Using an explicit surface mesh for the discretization requires a generalization of the differential operators in the differential equation in order to account for the arbitrary hypersurface. The
calculation of normal vectors of the interface or the treatment of topological changes is difficult in this
case. Implicit approaches typically extend the scalar off the interface to the whole domain. This obviously increases the dimension of the problem, but the transport equation can then be solved in the whole,
fixed computational domain.
We adopt an implicit approach by Adalsteinsson and Sethian [6] to model scalar advection and diffusion
on implicit interfaces. Assume, the scalar quantity to be transported, G, is given in the whole domain .
The transport of G in is defined as:

G
G G H G G
t

(6)

v w 2
u w 2
u v 2

nx

ny

n
y z
x z
x y z

v u
w u
w v

nx ny

nx nz

ny nz
x y
x z
y z

2
2 G 2 G 2
2 G 2 G 2
G 2 G 2
2 nx
2 ny
2 nz
G
2
2
2
y
z
x
z
x
y
2 G
2 G
2 G
2
nx ny 2
nx nz 2
ny nz
xy
xz
yz
T
velocity: u v w diffusion coefficient: H

normal vector: nx ny nz T curvature:


Due to the signed-distance property of the level-set field, the normal vector components can be computed
directly as the normalized gradient of . The first term on the right hand side in (6) is an advection term,
the second term accounts for the change in the hypersurface shape and the last term provides diffusion.

3.1 Extension of the scalar quantities off the interface


The approach in the previous section is based on the assumption that the scalar G is known throughout
the whole domain . From the physical point of view G is, however, only available on the interface d .
In order to extend this quantity off the interface to we use a technique first applied by Chen et al. [8].
The original motivation was the construction of global velocity fields for problems where only interface
velocities are defined. We apply the slightly modified equation given in Chessa et al. [7]:
signG 0

in

G G

(7)

on d

(8)

The solution of Eq. (7) corresponds to an extension of G off the interface orthogonal to the level-set
field (cf. Fig. 2). The interface d hereby serves as a inflow boundary. That is, if we maintain a smooth
level-set field this also leads to a smooth extended scalar field.
d
2

1
d

Figure 2: Extending G off the interface orthogonal to [7].

Figure 3: Applying the boundary condition for


the extension off G .

It needs to be discussed how we apply the boundary condition (8) on an implicitly given interface. We
adopt a simple approach which turned out to be sufficiently accurate:
1. Find the intersections int
j of d with the element edges.
2. For each cut element node i I find min i int
j .
j

3. Prescribe Gi G

(cf. Fig. 3).

4 Numerical Results
We consider the surfactant transport on a rising n-butanol droplet in water. The initial spherical droplet
has a radius r 00015 m and physical properties 1 845422 kg/m3 , 1 0003281 kg/m/s and surface
3

tension coefficient 000163 kg/s2 . The surrounding phase is defined by 2 986506 kg/m3 and
2 0001388 kg/m/s. We choose a uniform initial surfactant distribution on the interface G 05 and
neglect diffusion on the interface H 0. 40 40 80 hexahedral elements are used to discretize the
rectangular computational domain spanning 0006 0006 0012 m. A time span of 002 s is observed
using a time step t 00005 s.

(a) t 0000 s

(b) t 0005 s

(c) t 0010 s

(d) t 0020 s

Figure 4: Rising droplet with surfactant concentration.


Fig. 4 shows the clustering of the surfactants at the bottom of the drop as it rises.

5 Conclusion
We successfully developed a 3D two-phase flow solver that includes the simulation of transport phenomena on the interface. The XFEM framework serves as a good basis, as discontinuities across the
interface are accurately accounted for. In order to keep the advantage of an implicit interface description,
we employ an implicit technique for the interface transport, too. We show how the scalar quantity on the
interface is extended to the whole domain. The approach given in [6] is used to govern the scalar transport of the globalized quantity. A typical application of the approach is given by means of the simulation
of surfactant transport on a rising droplet.

References
[1] N. Mos, J. Dolbow and T. Belytschko, A Finite Element Method for Crack Growth without Remeshing,
International Journal for Numerical Methods in Engineering, 46 (1): 131-150, 1999
[2] H. Liu and Y. Zhang, Phase-field modeling droplet dynamics with soluble surfactants, J. Comput. Phys., 229
(24): 9166-9187, 2010
[3] M. Bertalmio, L.-T. Cheng, S. Osher and G. Sapiro, Variational Problems and Partial Differential Equations
on Implicit Surfaces, J. Comput. Phys., 174 (2): 759-780, 2001
[4] K. Deckelnick, C. Elliott and V. Styles, Numerical diffusion-induced grain boundary motion, Interfaces Free
Bound., 3: 393-414
[5] C.H. Leung and M. Berzins, A computational model for organism growth based on surface mesh generation,
J. Comput. Phys., 188: 75-99, 2003
[6] D. Adalsteinsson and J.A. Sethian, Transport and diffusion of material quantities on propagating interfaces
via level set methods, J. Comput. Phys., 185: 271-288, 2003
[7] J. Chessa, P. Smolinski and T. Belytschko, The extended finite element method (XFEM) for solidification
problems, Int. J. Numer. Meth. Engng., 53 (8): 1959-1977, 2002
[8] S. Chen, B. Merriman, S. Osher, P. Smereka, A Simple Level Set Method for Solving Stefan Problems, J. Comput. Phys., 135: 8-29, 1997
[9] S. Ganesan and L. Tobiska, A coupled arbitrary Lagrangian-Eulerian and Lagrangian method for computation of free surface flows with insoluble surfactants, J. Comput. Phys., 228: 2859-2873, 2009

XFEM 2011

II European Conference on eXtended Finite Element


Cardiff University, United Kingdom
Cardiff, June 29-30 July 1, 2011

Self-affine Scaling of Simulated Intergranular Cracks in Brittle Polycrystals


Z. Shabir1 , L. Ponson2 , A. Simone1
1 Faculty of Civil Engineering and Geosciences, Delft University of Technology
P.O. Box 5048, 2600 GA Delft, The Netherlands, {z.shabir, a.simone}@tudelft.nl
2 CNRS and Universit Paris 06, UMR 7190, Institut Jean Le Rond dAlembert
F-75005 Paris, France, ponson@dalembert.upmc.fr

Introduction and method of analysis

The statistical analysis of fracture surfaces is a valuable tool for understanding failure mechanisms in
engineering materials [1]. Self affinity of the fracture surfaces has been established through a significant
amount of experimental work [24] and is described by the scaling relation hx hx, where hx
is the height distribution of a crack profile, is a scaling factor, and is a parameter, known as Hurst
or roughness exponent, which characterizes the roughness of the cracked surface. The Hurst exponent
remains invariant under a scale transformation. With reference to the experimental characterization of
fracture in 3D brittle materials, Milman et al. [5] found a Hurst exponent 3D 04 015 for the fracture surface in tungsten single crystals and highly oriented pyrolitic graphite. For materials exhibiting
intergranular brittle fracture, Boffa et al. [6] reported an exponent equal to 05. Ponson et al. [7] have
observed the self-affine properties of cracked surfaces of compacted brittle granular materials and reported 3D 046 005. From the theoretical side, fracture surfaces have been modeled as the trace left
by the crack front as it propagates through an elastic solids with random spatial fluctuations of fracture
energy, leading to a value of roughness exponent th
3D 039 close to the experimental values [8, 9]. For
crack propagation in 2D systems, the value of the roughness exponent is a priori different, since failure
consists now of the propagation of the crack tip in a 2D plane, contrary to the motion of the crack line
in 3D settings [10]. In these geometries, experimental studies have led to values close to 2D 06 07
(see e.g. Ref. [11]), but have been mainly devoted to systems with rather complex microscopic failure
mechanisms such as viscoelasticity and microcracking. As a result, the connexion with the theoretical
models issued from Fracture Mechanics that predict rather smaller values (th
2D 050) [12, 13] is not
clear.
This study has been designed in order (i) to clarify the discrepancy between theory and experiments on
the morphology of crack profiles in 2D setting, and (ii) to understand the role played by the microstructure of polycrystalline materials on their fracture surfaces during brittle and intergranular failure. For
these reasons, we have simulated intergranular cracks in brittle polycrystalline materials. Intergranular
crack paths are obtained with a Generalized Finite Element Method (GFEM) for polycrystals [14]. This
method provides an accurate description of the stress field in a polycrystalline material and yields reliable crack paths [15]. In our approach, a polycrystal is obtained by the superposition of a polycrystalline
topology on a simple background finite element mesh which does not need to conform to the local features of the topology. This superposition is automatically handled using the partition of unity property of
finite element shape functions hence automating the mesh generation stage. A sequential polycrystalline
analysis approach is used to handle our large simulations as discussed in Section 3. Resulting crack
profiles are analyzed using the height-height correlation function in section 4. The estimated value of the
Hurst exponent is compared with recent experimental and theoretical findings.

W 2 4200 m

W 2

2W 3

W 140

( 6 lgb )

process
zone

W 3
( 265 lgb )

28lgb

Figure 1: Geometry and boundary conditions for the notched specimen employed in the simulations.

Test setup and material

The geometry and boundary conditions of the test setup are reported in Figure 1. The material parameters
are taken to be representative of an average polycrystalline alumina, Al2 O3 , with Youngs modulus E
3846 GPa and Poissons ratio 0237. Plane strain analyses are performed under the assumption
of small elastic strains and rotations considering quasi-static loading conditions. A centroidal Voronoi
tessellation algorithm is used to generate 18 different topologies of 3140 grains each. These topologies
are embedded in the process zone shown in Figure 1. The average grain size is 20 m, with an average
grain boundary length lgb 1062 m. The linear elastic isotropic grains are separated from each other
by means of cohesive grain boundaries which follow the Xu-Needleman cohesive law [16] incorporating
secant unloading and reloading behavior. According to Shabir et al. [15], intergranular crack paths in
brittle polycrystals are not influenced by key cohesive law parameters. Following this argument, the
mode-I fracture energy GIc and the maximum normal cohesive strength max are set equal to 396 Jm2
and 06 GPa, respectively. With this set of cohesive law parameters, relatively coarse meshes can resolve
the cohesive law along grain boundaries.

Sequential polycrystalline analysis

Crack paths are generated using a sequential polycrystalline analysis approach. With this approach,
large crack propagation simulations can be carried on by considering a suitable number of cheaper subsimulations with a considerable saving in terms of computational effort. A typical example is shown in
Figure 2(a) with the first and last simulations reported in Figures 2(b) and (c), respectively. The blue
boxes in these figures show the area where the crack is allowed to propagate within a sub-simulation.
Each box is termed process window. Two different levels of mesh refinement are used in each of these
sub-simulations. The mesh in the process window is characterized by a level of refinement according
to [15], whereas the region outside is refined by providing at least two elements along each grain boundary. When a crack touches the end of a process window, the simulation is stopped and the resulting crack
profile is saved. The next sub-simulation is launched considering the saved crack profile from the previous simulation. A new crack tip is defined by reducing the length of the loaded crack profile such that the
new tip is now at a position where the cohesive strength of the previous simulation would be negligible.
In other words, we make sure that nonlinear processes are accurately captured by enclosing nonlinear
regions with a process window. We have found that an overlap of 25lgb between two consecutive
process windows satisfies this requirement with the current choice of cohesive law and parameters. Unlike classical adaptive refinement approaches, the only information that is transferred from simulation to
simulation is the crack path.
We have compared our approach with a monolithic approach on a series of polycrystalline topologies
and the resulting crack paths are identical. Moreover, with reference to the load-displacement curves,
if all the curves obtained with the sequential analysis approach were joined together by removing the
pre-peak branch of each sub-simulation except the first one, then the resulting curve would follow the

traction free crack


crack obtained from analysis
process window enclosing
highly refined region

direction of crack propagation


process window

(a) division of a simulation with 323 cracked GBs in front of notch tip into 8 sub-simulations
highly refined region inside process window

at least 2 elements along each GB outside process window

(b) first simulation


traction-free crack

crack obtained from analysis

(c) last simulation

Figure 2: Sequential polycrystalline analysis of a 3140-grain topology.


curve obtained with the monolithic approach very closely. As for the stress field, we could hardly find
any difference.

Roughness characterization of the crack profiles

In order to characterize the statistical properties of the numerical fracture surfaces, we analyze 18 crack
profiles characterized by 315 cracked grain boundaries obtained from 3140-grain samples each
sample corresponds here to different polycrystalline arrangements. We use the height-height correla12
tion function defined as hx hx x hx2x . Its variations are represented in Figure 3 in
logarithmic scales. It follows a power-law, reminiscent of self-affine properties, characterized by the
roughness exponent 050 in a rather large range of length scales, from the typical length of the grain
boundaries ( 10 m) to a fraction of the system size ( 800 m Lf 4, where Lf is the fracture
ligament length equal to W 3).
Such a value of the Hurst exponent is smaller than the experimental findings 2D 06 07 obtained in
2D materials such as paper sheets [11]. However, such materials are known to give rise to microcracking
and damage ahead of the main crack during crack propagation while in the polycrystalline materials considered here failure mechanisms are perfectly brittle. As a result, those two different values of roughness
exponent are in rather good agreement with the scenario proposed in Ref. [17] where brittle and damage
th
failure of disordered 2D materials are expected to give rise to th
2D 05 and 2D 05, respectively.
Why brittle intergranular failure of polycrystalline materials studied here gives rise to 050, and to
which extent our findings can be understood within the framework developed in Refs. [12, 17] are still
open questions.

Acknowledgements
This research is supported by the Higher Education Commission, Pakistan.

References
[1] E. Bouchaud. Scaling properties of cracks. Journal of Physics: Condensed Matter, 9(21):43194344, 1997.

102

101
h [ m]

050

100

101
101

d0 55 m
100

101

800 m
102

103

104

x [ m]

Figure 3: Hurst exponent estimated on crack profiles obtained from 18 3140-grain topologies.
[2] B. B. Mandelbrot, D. E. Passoja, and A. J. Paullay. Fractal character of fracture surfaces of metals. Nature,
308(19):721722, 1984.
[3] K. J. Mly, A. Hansen, E. L. Hinrichsen, and S. Roux. Experimental measurements of the roughness of
brittle cracks. Physical Review Letters, 68(2):213215, 1992.
[4] L. Ponson, D. Bonamy, and E. Bouchaud. Two-dimensional scaling properties of experimental fracture
surfaces. Physical Review Letters, 96(3):035506, 2006.
[5] V. Yu. Milman, R. Blumenfeld, N. A. Stelmashenko, and R. C. Ball. Comment on experimental measurements of the roughness of brittle cracks. Physical Review Letters, 71(1):204, 1993.
[6] J. M. Boffa, C. Allain, and J. P. Hulin. Experimental analysis of fracture rugosity in granular and compact
rocks. The European Physical Journal - Applied Physics, 2(3):281289, 1998.
[7] L. Ponson, H. Auradou, M. Pessel, V. Lazarus, and J. P. Hulin. Failure mechanisms and surface roughness
statistics of fractured fontainebleau sandstone. Physical Review E, 76(3):036108, 2007.
[8] D. Bonamy, L. Ponson, S. Prades, E. Bouchaud, and C.Guillot. Scaling exponents for fracture surfaces in
homogeneous glass and glassy ceramics. Physical Review Letters, 97:135504, 2006.
[9] L. Ponson. Crack propagation in disordered materials: How to decipher fracture surfaces. Annales de
Physique, 32(1):1128, 2007.
[10] D. Bonamy and E. Bouchaud. Failure of heterogeneous materials: A dynamic phase transition? Physics
Reports, 498(11):144, 2011.
[11] M. Mallick, P. P. Cortet, S. Santucci, S. G. Roux, and L.Vanel. Discrepancy between sub-critical and fast
rupture roughness: A cumulant analysis. Physical Review Letters, 98:255502, 2007.
[12] E. Katzav, M. Adda-Bedia, and B. Derrida. Fracture surfaces of heterogeneous materials: A 2d solvable
model. Europhysics Letters, 78(4):46006, 2007.
[13] I. Ben-Dayan, E. Bouchbinder, and I. Procaccia. Random and correlated roughening in slow fracture by
damage nucleation. Physical Review E, 74:146102, 2006.
[14] A. Simone, C. A. Duarte, and E. Van der Giessen. A Generalized Finite Element Method for polycrystals with
discontinuous grain boundaries. International Journal for Numerical Methods in Engineering, 67(8):1122
1145, 2006.
[15] Z. Shabir, E. Van der Giessen, C. A. Duarte, and A. Simone. The role of cohesive properties on intergranular
crack propagation in brittle polycrystals. Modelling and Simulation in Materials Science and Engineering,
19(3):035006, 2011.
[16] X.-P. Xu and A. Needleman. Numerical simulations of fast crack growth in brittle solids. Journal of the
Mechanics and Physics of Solids, 42(9):13971434, 1994.
[17] E. Bouchbinder, J. Mathiesen, and I. Procaccia. Roughening of fracture surfaces: The role of plastic deformation. Physical Review Letters, 92:245505, 2004.

 





 




 



 
 

   

 
 

  

 









   

 


 


 





 











 





 


















 
 





 

 

 !

" !

  

 

#$% !

()
(
(*
(
(+
()
(

()
( *
( +
(
% ( +
% ( *
% ()

&



 '
!%
 !

" !

()
(
(*
(
(+
()
(

&

#$% !
()
( *
( +
(
% ( +
% ( *
% ()



 ''
!%

XFEM 2011

II European Conference on eXtended Finite Element


Cardiff University, United Kingdom
Cardiff, June 29-30 July 1, 2011

A framework for fracture modeling using implicitly defined enrichments over Ck partitions of unity simultaneously based on finite elements and meshfree nodes
Diego Amadeu F. Torres1 , Clovis S. de Barcellos2 , Paulo de Tarso R. Mendona3 , Andresa Freitas
Group of Mechanical Analysis and Design,
Department of Mechanical Engineering, Federal University of Santa Catarina, Brazil,
1 diego.amadeu@gmail.com, 2 clovis@grante.ufsc.br, 3 mendonca@grante.ufsc.br

1 Introduction
This work is concerned with the development of a procedure for to build discontinuous enrichments for
ansatz spaces using boolean R-functions [1] on arbitrary polygonal supports. A Ck -continuous Partition
of Unity [2] simultaneously based on finite elements and meshfree nodes [3] is built in such a way that
the position of the meshfree nodes and its supports are defined to receive the discontinuous and singular
enrichments. Thus, the topology of the discontinuity can be effectively constructed since all the closed
semi-analytic sets may be represented implicitly by real valued functions with guaranteed differential
properties [4]. The Shepard approximants capability of to couple different partitions of unity is exploited
and the width of the support of the discontinuous enrichments may to vary in order to improve the
precision of the unknowns fields, aiming to circumvent the problems related to the transition layer and
blending elements, as discussed in [5] and [6]. Moreover, the integrations are performed only in global
coordinates, avoiding errors due distortion of meshes. The purpose is to model discontinuities for twodimensional problems of Linear Elastic Fracture Mechanics in such a fashion that the error inherent to
C0 -continuous approximations and due the lack of accuracy on the transition layer between enriched and
non-enriched portions of the domain can be reduced.

2 Boolean R-functions and implicitly defined sets


Construction of a single function that takes on the prescribed values and/or derivatives on some collection
of points sets which may contain isolated points, bounded or unbounded curves, as well as surfaces and
regions of arbitrary topology can be an issue for interpolation methods.
In context of geometric modeling, any object defined by a predicate on primitive halfspaces can also
be represented by a single implicit function. For implicitization, it may be mean the process of constructing such function for free-form parametric curves or surfaces and amounts to extrapolating the data
prescribed on curve or surface into the whole space.
The implicit functions approximates the distance from a given point to the locus of points where some
data my be prescribed. Implicit representation of a point set by zeros of some real valued function is not
constrained by the incidence, regularity, or the topology of the sets being interpolated. Such functions
may be normalized to behave as smooth approximations to distance functions and may be constructed
automatically from other geometric representations [4]. A general algorithmic approach to constructing
implicit functions can be derived using the theory of R-functions [7], which are real-valued functions
whose sign is completely determined by the signs of its arguments. Basically, R-functions are evaluated
and differentiated using set operations. For example, a R-function or with two arguments, f1 and f2 , is
defined as
1

k
f1 k0 f2 := f1 + f2 + f12 + f22 f12 + f22 2

(1)

where k is a positive integer, and the function is analytic everywhere except at the origin, where it is k
times differentiable.
This approach can be understood as a generalization of the classical inverse distance weighting interpolations for scattered data [8], where the interpolating function is constructed from pointwise values fi
with weighting functions Wi , such that f (x) = ni=1 fi Wi (x). It is interesting that the weight functions
be positive continuous functions satisfying the interpolation conditions Wi (x j ) = i j and form a partition
of unity. The evaluation of the weighting function Wi may be performed with

Wi (x) =

j
nj=1; j=i d j (x)
j
n
n
k=1 j=1; j=k d j (x)

(2)

where d j (x) is the Euclidean distance from point x to the node xi .The behavior of the interpolating functions at the node is controlled by the exponents j . When 0 < j 1 the interpolant is not differentiable
at the ith node and for values j > 1 the function is differentiable j 1 times at the ith node. The
satisfaction of prescribed values of derivatives at the ith node can be achieved by a slight modification
the weighting functions, as described in [1].
Therefore, such approach may be used to define some internal features of the domain such cracks, interfaces, holes between others. In other words, in case of Linear Elastic Fracture Mechanics, the crack
topology may be implicitly represented over a portion of the domain probably influenced by the discontinuity and singularity.
For instance, [9] shows that non-continuously differentiable functions when using level set representation
lead to auxiliary fields that do not satisfy the hypothesis of the divergence theorem, which is used to
derive the domain form of the J integral, and therefore giving inaccurate SIFs values.
Thus, the enrichment constructed with aid of R-functions may be differentiable in such a way that the
approximated stress field be not discontinuous due the enrichment.

3 Ck continuous approximation functions


In addition to the interest in assuring smoothness of the enrichment functions, it would be an issue the
smoothness of the Partition of Unity which will be enriched.
In this work, the extrinsic enrichment is adopted because discontinuity does not exist when different enrichments functions are used at different discrete points even though requiring additional global degrees
of freedom.
The Partition of Unity (PoU) with arbitrary continuity can be described as a special instance of the hpcloud in which each cloud has an arbitrary polygonal boundary and the weighting functions are built
from cloud edge functions. The extension of Edwards procedure [10] by Duarte [2] is explored for
convex and non-convex supports (clouds) with the aid of the R-functions.
The choice of the weighting functions Wi is quite arbitrary but they must satisfy some conditions which
will impart a great influence on the approximation process. Firstly, they must have the required continuity
k, and, secondly, the resulting PoU should have reasonable integrability properties, and be able to, at least,
represent a constant solution. The cloud edge functions play therefore an essential role as part of the PoU
building process, that is, they must have at least the desired Ck continuity, k, with k being an arbitrary
integer. The continuity of the PoU is the same as the weight function.
In this approach, a two-dimensional open bounded domain, R 2 (x), is partitioned in a linear trianN
gular mesh,{Ke }NE
e=1 (NE being the number of elements Ke ), using N nodes with coordinates {x j } j=1 .
For each of these nodes one considers as cloud j , j = 1, ..., N, the union of the triangular elements
which share it so that N clouds is an open covering, N , of the domain . Basically, the procedure shares
the same features as the standard finite element method regarding the domain partition and integration
procedures.

4 Coupling between meshfree and finite element-based functions


Transition layer between portions enriched and non-enriched with singular functions is an issue concerning the rate of convergence [5] [11].
In order to balance the weight of the functions around some feature introduced by enrichment, it is the
purpose of this work to use superimposed meshfree nodes. For these nodes, the corresponding cloud is
defined by a set of background elements and the couple of element-based and meshfree PoU is performed
using the Shepard equation.
fe
mf
According [3] for each finite element node x j , j I f e , and each meshfree node x j , j Im f , a partition
of unity N j may be defined as

N j (x) =

W jf e (x)

, se j I f e

fe

I f e (x) W (x) + Im f (x) Wm f (x)

W jm f (x)

fe

mf

(3)

, se j Im f

I f e (x) W (x) + Im f (x) W (x)

where I f e (x) = I f e : W (x) = 0 and Im f (x) = Imf : W (x) = 0 . Let Im f , f e = I f e Im f


denote the index set of all nodes in the domain . Then, the set N j , j Im f , f e , with Im f , f e = I f e Im f
constitutes a partition of unity subordinate to the open covering Im f , f e .
The procedure makes the transition between a finite element PoU and any other type of PoU quite natural,
flexible and transparent. The only requirement is that the union of their supports completely covers the
computational domain. Exponential convergence of the resulting coupled approximation is demonstrated
in [3]. Moreover, this approach facilitates the implementation of Dirichlet boundary conditions.

5 Numerical integrations
Even though the process for constructing continuous PoU be quite versatile, allowing for free-form
clouds, the need for the so-called background grid for domain integration in meshfree methods suggests
the cloud definition using triangular meshes be very appropriate. Thus the finite element mesh is used
to build the meshfree functions and enforce that the support of all functions coincide with the support of
corresponding global functions defined on the same mesh.
The use of smooth approximation functions appears to be particularly effective in dealing with large
deformation problems since the distortion of elements does not causes lose of approximation capability
of the functions. Moreover, as neither the PoU nor enrichment functions are defined in natural domains,
the integration is performed only using global coordinates.
In this work, for integration of conventionally enriched PoU with monomials and singular functions
away the singular points is used the symmetric quadrature of Wandzura [12] which has been tested in
computation of plate models, as discussed in [13]. For cells containing a singular point a modified
Lachat-Watson transformation in conjunction with a 8-nodes Serendipity mapping moving the mid-side
nodes towards the singular point furnishes good integrations if compared against the results of [14].

6 Conclusions
The methodology proposed is been implemented and its performance will be investigated using error
estimation in global and goal-oriented measures. It should be noted that the continuity may be exploited
in implicit residual algorithms for error estimators since there are no stress jumps at element boundaries.
First results show that the continuity (see figures 1 and 2 for a comparison between standard C0 and
continuous functions) provided by the approximation functions may impart benefits in the computation
of propagation parameters as SIFs and J-integral. Moreover, the enrichment built as described allows
using only geometric parameters as the crack tip position, angle and advance, in crack propagation
modeling based on Linear Elastic Fracture Mechanics.

Component y of strain

140

120

120

100

100

80

80

60

60

40

40

20

20

20

0
0

Figure 1:

20

40
0.05

Standard C0

60

Component y of strain

140

80

20

100

0
0

0.1

finite element functions

20

40
0.05

60

80

100

0.1

Figure 2: Continuous partition of unities

References
[1] V. L. Rvachev, T. I. Sheiko, V. Shapiro and I. Tsukanov, Transfinite interpolation over implicitly defined sets,
Computer Aided Geometric Design, 18 (3): 195-220, 2001
[2] C. A. Duarte, D. J. Kim and D. M. Quaresma, Arbitrarily smooth generalized finite element approximations,
Computer Methods in Applied Mechanics and Engineering, 196 (1-3): 33-56, 2006
[3] C. A. Duarte, D. Q. Migliano and E. B. Baker, A technique to combine meshfree- and finite element-based
partition of unity approximations, Structural Research Series 638. Department of Civil and Environmental
Engineering - UIUC, 2005
[4] V. Shapiro, I. Tsukanov, Implicit functions with guaranteed differential properties, in: Proc. 5th ACM Symposium on Solid Modeling and Applications, 258-269. Ann Arbor, 1999
[5] P. Laborde, J. Pommier, Y. Renard and M. Salan, High-order extended finite element method for cracked
domains, International Journal for Numerical Methods in Engineering, 64 (3): 354-381, 2005
[6] E. Bchet, H. Minnebo, N. Mos and B. Burgardt, Improved implementation and robustness study of the XFEM for stress analysis around cracks, International Journal for Numerical Methods in Engineering, 64 (8):
1033-1056, 2005
[7] V. L. Rvachev, Methods for logic algebra in mathematical physics. Naukova Dumka, 1974 (in Russian)
[8] D. Shepard, A two-dimensional interpolation function for irregularly spaced data, in: Proc. 23rd ACM National Conference, 517-524. New York , 1968
[9] M. Duflot, A study of the representation of cracks with level sets, International Journal for Numerical Methods
in Engineering, 70 (11): 1261-1302, 2007
[10] H. C. Edwards, C finite element basis functions, Report 96-45, The University of Texas at Austin, 1996
[11] G. Ventura, R. Gracie, T. Belytschko, Fast integration and weight function blending in the extended finite
element method, International Journal for Numerical Methods in Engineering, 77 (1): 1-29, 2009
[12] S. Wandzura, H. Xiao, Symmetric Quadrature Rules on a Triangle, Computer and Mathematics with Applications, 45 (12): 1829-1840, 2003
[13] P. T. R. Mendona, C. S. de Barcellos, D. A. F. Torres, Analysis of anisotropic Mindlin plate model by
continuous and non-continuous GFEM, to appear in Finite Element in Analysis and Design
[14] K. Park, J. P. Pereira, C. A. Duarte, G. H. Paulino, Integration of singular enrichment functions in the generalized/extended finite element method for three-dimensional problems, International Journal for Numerical
Methods in Engineering, 78 (10): 1220-1257




   

 

       

   


   
    


 

XFEM 2011

II European Conference on eXtended Finite Element


Cardiff University, United Kingdom
Cardiff, June 29-30 July 1, 2011

Mesoscopic Modelling of Masonry using Embedded Weak


Discontinuities based on Partitions of Unity
B. Vandoren12 , K. De Proft2
1
2

XIOS University College, Belgium, bram.vandoren@xios.be (corresponding author)


Hasselt University, Belgium, kurt.deproft@uhasselt.be

Introduction

The modelling of masonry has been a popular topic within computational mechanics for some years now.
Two major groups of modelling approaches can be distinguished: macroscopic and mesoscopic [1]. In
the macroscopic approach the joints and bricks are homogenized to one orthotropic material. The main
advantage of this method is that not much computational effort is needed to calculate large structures.
However, the obtained crack path is less detailed. This drawback can be alleviated by the use of mesoscopic models. In this approach, joints and bricks are modelled by separate entities. Classically, the
joints are incorporated by interface elements, situated on the boundaries of the continuum brick elements
[1, 2]. When a critical state is reached in a joint, a strong discontinuity (i.e. a jump in the displacement
field) is introduced in the interface.
An alternative way to incorporate strong discontinuities is the partition of unity method [3, 4, 5]. Within
this method, nodes are locally enhanced to enrich the solution with discontinuous modes. This concept
was applied to masonry by De Proft et al. [6] and will be extended in this paper by the incorporation of
weak discontinuities. A weak discontinuity introduces a jump in the strain field, allowing for failure to
localise in a zone with finite width [7, 8, 9]. The thickness of this failure is in this case linked to the joint
thickness.

2
2.1

Partition of unity concept for weak discontinuities


Displacement decomposition

The displacement field of a body crossed by a weak discontinuity (Figure 1) is obtained by:
u = u + H w u

(1)

in which u and u denote the regular and enhanced displacement field, respectively. H w is a unit ramp
function [10], defined by:

if x
0

H w =
(2)
if x w
+

1
if x +

Figure 1: Body crossed by a weak discontinuity

2.2

GFEM discretisation

In this work, the Generalized Finite Element method has been adopted to model the discontinuities
[11]. The unit ramp function (Equation (2)) is used as an enhanced basis: its value equals unity for
a point inside a masonry brick. When a support of a node is crossed by a weak discontinuity (i.e.
joint), an enhanced set of degrees of freedom is added to the solution field of that node. Consequently,
each brick possesses its own set of enhanced degree of freedom. Special care has to been taken in the
implementation of a meshgenerator, to prevent linear dependancy of the enhanced basis functions [11].
In the current model, two linear quadrilateral elements are used to model one brick.

3
3.1

Nonlinear modelling
Material model

The stone behaviour remains linear elastic during the simulations, whereas the joint behaviour is governed by an exponential damage evolution law [12]:

=0

if < 0
(3)
= 1 0 (1 ) + (0 )
if 0

where determines the brittleness of response and and 0 are illustrated in Figure 2. The loading
function , expressed in terms of strain invariants, is derived from the Drucker-Prager model [13]:

1 fc ft I1,
3 fc + ft J2,
=
+
(4)
2 fc 1 2
2
fc 1 +
in which ft and fc are the mortar tensile and compressive strength, respectively.

3.2

Energy Release Control

The global equilibrium path is traced using an energy release constraint function [14]:
1
g = fT (0 a a0 )
2

(5)

in which represents the enforced energy dissipation during a loadstep, f is a vector containing the
prescribed loads, is a load scaling factor and a is a vector containing the displacement field. The
subscript 0 refers to the converged values of the previous loadstep.

Figure 2: Exponential damage evolution law

Numerical examples

In order to demonstrate the potential of the developed masonry model, a three point bending test (Figure
3) and shear wall analysis (Figure 4) have been carried out. Typical load-displacement curves and crack
paths are recovered in these simulations.

Figure 3: Load-CMOD curve and deformed mesh for a three point bending test

Figure 4: Load-displacement curve and deformed mesh for a shear wall test

Conclusions

In this contribution, a mesoscopic masonry model is developed in which joints are modelled by weak
discontinuities. The discontinuities are incorporated using the Generalized Finite Element Method. A
Drucker-Prager damage model is used to describe the nonlinear behaviour of the joints. An energy
release constraint function is adopted to trace the global equilibrium path. The performance of the
masonry model is demonstrated by the simulation of a three point bending and shear wall test.

References
[1] P. B. Loureno. Computational Strategies for Masonry Structures (PhD Thesis). Technische Universiteit Delft,
1996.
[2] J. V. Alfate and J. R. de Almeida. Modelling Discrete Cracking on Masonry Walls, Masonry International, 17
(2): 83-93, 2004.
[3] T. Belytschko and T. Black. Elastic Crack Growth in Finite Elements with Minimal Remeshing, International
Journal for Numerical Methods in Engineering, 45 (5): 601-620, 1999.
[4] G. N. Wells and L. J. Sluys. A New Method for Modelling Cohesive Cracks Using Finite Elements, International
Journal for Numerical Methods in Engineering, 50 (12): 2667-2682, 2001.
[5] K. De Proft. A Combined Experimental-Computational Study to Discrete Fracture of Brittle Materials (PhD
Thesis). Vrije Universiteit Brussel, 2003.
[6] K. De Proft, K. Heyens and L. J. Sluys. Mesoscopic Modelling of Masonry Failure, Proceedings of the ICE Engineering and Computational Mechanics, 164 (EM1): 41-46, 2010.
[7] T. Belytschko, J. Fish and B. E. Engelmann. A Finite-Element with Embedded Localization Zones, Computer
Methods in Applied Mechanics and Engineering, 70 (1): 59-89, 1988.
[8] L. J. Sluys and A. H. Berends. Discontinuous Failure Analysis for Mode-I and Mode-II Localization Problems,
International Journal of Solids and Structures, 35 (31-32): 4257-4274, 1998.
[9] M. Jirsek. Comparative Study on Finite Elements with Embedded Discontinuities, Computer Methods in
Applied Mechanics and Engineering, 188 (1-3): 307-330, 2000.
[10] J. Oliver, M. Cervera and O. Manzoli. Strong Discontinuities and Continuum Plasticity Models: the Strong
Discontinuity Approach, International Journal of Plasticity, 15 (3): 319-351, 1999.
[11] A. Simone, C. A. Duarte and E. Van der Giessen. A Generalized Finite Element Method for Polycrystals
with Discontinuous Grain Boundaries, International Journal for Numerical Methods in Engineering, 67 (8):
1122-1145, 2006.
[12] R. de Borst, M. G. D. Geers, E. Kuhl and R. H. J. Peerlings. Enhanced Damage Models for Concrete Fracture,
Computational Modelling of Concrete Structures - Proc. EURO-C 1998 Conf., 1998.
[13] T. J. Massart, R. H. J. Peerlings and M. G. D. Geers. Mesoscopic Modeling of Failure and Damage-induced
Anisotropy in Brick Masonry, European Journal of Mechanics a-Solids, 23 (5): 719-735, 2004.
[14] M.A. Gutirrez. Energy Release Control for Numerical Simulations of Failure in Quasi-brittle Solids, Communications in Numerical Methods in Engineering, 20: 19-29, 2004.

XFEM 2011

II European Conference on eXtended Finite Element


Cardiff University, United Kingdom
Cardiff, June 29-30 July 1, 2011

Imperfect interfaces with XFEM: new developments in enrichment


function
Q.Z. Zhu1 , J. Yvonnet1 , J.F. Shao2 , Q.C. He1
1
2

Universit Paris-Est, France, {qizhi.zhu,j.yvonnet,qi-chang.he}@univ-paris-est.fr


Universit Lille 1, France, jianfu.shao@univ-lille1.fr

Introduction

Discontinuous problems with imperfect interfaces are often encountered in mechanics and engineering
science, and their influences on the mechanical behavior of materials and structures has long been recognized. Among all the imperfect interface models in elastostatic context, the spring-layer interface
model and the coherent interface model are the two most widely used ones. Initially provoked from
phenomenological reasons, these two interface models have been derived by using asymptotic methods
through an equivalent replacement of the three-phase configuration by a two-phase configuration (Figure
1). However, analytical results available are still limited to some very simple cases.
In the context of the extend finite element method (XFEM) [1][2], numerical techniques have been developed for modelling imperfect interfaces with complex geometries [3][4]. The present work is devoted to
further studies on this topic. For each of the aforementioned interface models, the weak form of the initial
problem and an appropriate enrichment function are delivered. Particularly, a unified material-dependent
formulation linking two specific functions is proposed. Moreover, a simple numerical application to
spring-layer imperfect interface problem is presented.



 





 




Figure 1: Equivalent replacement of the perfectly bonded three-phase configuration by a two-phase


configuration involving imperfect interface

Problem statement

Let 1 and 2 be two adjoining bulk phase in a solid, as shown in Figure 1(right). The interface
between 1 and 2 , denoted by , is imperfect and geometrically characterized by the zero level-set
of a function : 3 : x 3 x 0. Assuming that is continuously differentiable and
its gradient x 0 for all x , the unit normal vector field nx defined on and directed from
2 toward 1 is calculated by
ni x

i x

x
1

(1)

As one of the significant advantages in XFEM-framed modelling and simulation, meshes are not needed
to conform to the geometry of discontinuities. Usually, we first generate a regular mesh and then introduce therein the discontinuity via its level set function. Four types of elements may arise in this process,
as those shown in Figure 2.
The key step in numerical modelling by XFEM consists in applying appropriate functions to describe the
discontinuity of physical field by means of superimposing one or several enrichments to the conventional
continuous interpolation, such that [2]
ne

me

i1

r1

ux Ni xai Nr xr xbr

(2)

for an elasticity problem. Above, Ni x are the shape functions for a standard finite element. The
functions r x, known as enrichment functions, influence significantly computational convergence and
computational accuracy especially in the zones close to discontinuities.

  
      



   


   

    
    

Figure 2: Four probable types of elements generated by the interface

Interface models and enrichment functions

In this part, we discuss two widely-used imperfect interface models and their modelling by XFEM.

3.1

Spring-layer imperfect interfaces

Among all the models describing imperfect interfaces, the spring-layer interface model is probably the
simplest one and also the most widely used one. According to this model, the displacement vector ux
is discontinuous across an interface while the traction vector tx n is continuous across the same
interface and proportional to the displacement vector jump
u 0

t 0

t Ku

(3)

where the second order tensor denotes the local interface stiffness. The weak form of the problem
reads [4]

tudS
(4)
u : ud tudS

To describe the strong discontinuity across the interface, we propose the following enrichment function

1
r x sign x 1 sign x signr
(5)
2
where x ni1 Ni xi and r is the value of the level set at node xr . It follows the displacement jump
ux

ne

Nr xbr

(6)

r1

The function (5) possesses the salient features: i) it is node-dependent; ii) it contains three parts where

the coefficient 12 allows no factor to be involved


in (6), the first sign
x provides the required discontinuity across the interface, and the third term 1 sign x signr describes the relative position of
integration point x and node xr ; iii) it allows removing numerical issues of some blending elements [4].
2

3.2

Coherent imperfect interfaces

The coherent imperfect interface model is appropriate for modelling the interface effect in nanomaterials
and nanostructures. This model stipulates that the displacement vector is continuous across an interface
whereas the stress vector suffers a jump across the same interface, which must comply with the YoungLaplace equation
u 0 t 0 divs s t
(7)
The weak formulation corresponding to this model takes the form [3]

u : ud

divs s uudS

tudS

(8)

We propose the following enrichment function for coherent imperfect interfaces:

1
r x x 1 sign x signr
2

(9)

According to the level-set method, the term x , which replaces the first sign x of (5), allows guaranteeing the continuity in displacement. Again the issues of blending elements can be removed automatically.
Note that when one of element edges in 2D case or one of element sides in 3D case locates on the interface, the function x ni1 Ni x i ni1 Ni xi , which were proposed by Mos et al. [7] for
perfect interface problem and followed in [3] for coherent interface problem, takes zero for all integration points inside such elements and therefore fails in describing the discontinuity of traction across the
interface segment 12 (shown in Figure 2) shared as one common side by the triangular linear elements
e124 and e128 . In this sense, the proposed function (9) can remedy the above shortcoming.

3.3

Unified formulation of enrichment functions

Above, enrichment functions for spring-layer interfaces and coherent interfaces are proposed separately.
A linear combination of the two specific functions (5) and (9) leads to a unified formulation
r x

1
1 x 2 sign x 1 sign x signr
2

(10)

which involves two coefficients 1 and 2 for which we propose the following material-dependent forms
1

min0 1 2

min1 2

max1 2
max0 1 2

(11)

in which 1 and 2 , scalar-valued or norms of tensorial quantities, represent the material properties of
the matrix and inclusion phases, respectively, while 0 characterizes that of the initial thin interphase
which has been replaced by the interface. Obviously, as expected, i) when the interphase is very soft
(0 and 0 2 ), 1 0 and 2 1, the function (10) reduces to (5); ii) when the interphase
is very rigid (0 1 and 0 2 ), 1 1 and 2 0, the function (10) reduces to (9). As it is
the case in mathematical derivation, the contrast between 1 and 2 is assumed to be comparable.

A simple numerical example

As shown in Figure 3, the composite occupying the unit square domain is composed of the matrix and
a circular rigid inclusion. The contrast in shear modulus is chosen as 2 1 15. The interface
between the bulk phases is assumed to be imperfect and can be described by the spring-layer model.
For numerical computation, the domain is first discretized by a regular mesh of triangular elements. The
interface is then introduced into the regular mesh via the level set function

f x y x2 y2 r
(12)
The circular inclusion domain centered at 05 05 has the radius r 03. The boundary conditions are
such that ux x 0 0 and uy y 0 0. The structure is subjected on the side x 1 to a uniform
displacement ux x 1 0001.
3

The displacement field over the deformed configurations are reported in Figure 3 (left). As expected, we
observe a pronounced displacement jump in the loading direction. The smoothness of the crescent displacement discontinuity shows that the enrichment function and the relevant numerical treatment allow
describing the field of the imperfect interface problem quite well (otherwise, the field of displacement
jump will be saw-toothed). The right one in Figure 3 shows the distributions of the normal stress along
some chosen lines passing the center and inclined at different angles with respect to the x-axis. We notice that the required condition on continuity of traction vector across the interface is completely satisfied,
which justifies the robustness and efficiency of the proposed numerical treatments.

Figure 3: Displacement distributions on deformed structures (left) and distributions of the normal stresses
along some lines inclined at different angles with respect to the loading direction (right)

Conclusion

The present study together with the previous works [3][4][5][6] confirm the suitability and performance
of the extended finite element method in numerical modelling and simulation of discontinuous problems
with typical imperfect interfaces. Enrichment functions proposed here are of unified nature and show
geometric and physical consistencies. Along this line, problems with complex geometries of imperfect
interfaces will be handled in future work.

References
[1] T. Belytschko and T. Black, Elastic crack growth in finite elements with minimal remeshing. International
Journal for Numerical Methods in Engineering, 45: 601-620, 1999
[2] N. Mos, J. Dolbow and T. Belytschko, A Finite Element Method for Crack Growth without Remeshing,
International Journal for Numerical Methods in Engineering, 46 (1): 131-150, 1999
[3] J. Yvonnet, H. Le Quang, QC. He, An XFEM/level set approach to modelling surface/interface effects and to
computing the size-dependent effective properties of nanocomposites. Computational Mechanics, 42: 119-131,
2008
[4] QZ. Zhu, ST. Gu, J. Yvonnet, JF. Shao, QC. He, Three-dimensional numerical modelling by XFEM of springlayer imperfect curved interfaces with application to linearly elastic composite materials. International Journal
for Numerical Methods in Engineering, DOI: 10.1002/nme.3175, 2011
[5] J. Yvonnet, QC. He, C. Toulemonde, Numerical modelling of the effective conductivities of composites with
arbitrarily shaped inclusions and highly conducting interface, Composites Science and Technology, 68: 28282825, 2008
[6] J. Yvonnet, QC. He, QZ. Zhu, JF. Shao, A general and efficient computational procedure for modelling the
Kapitza thermal resistance based on XFEM, Computational Materials Science, 50: 1220-1224, 2011
[7] N. Mos, N. Cloirec, P. Cartraud, JF. Remacle, A computational approach to handle complex microstructure
geometries. Comput Methods Appl Mech Eng, 192: 3163-3177, 2003