12

Giulio Ventura
Elena Benvenuti  Editors

Advances in
Discretization Methods
Discontinuities, Virtual Elements,
Fictitious Domain Methods

Se MA
SEMA SIMAI Springer Series
Series Editors: Luca Formaggia • Pablo Pedregal (Editors-in-Chief)
Jean-Frédéric Gerbeau • Tere Martinez-Seara Alonso • Carlos Parés • Lorenzo Pareschi •
Andrea Tosin • Elena Vazquez • Jorge P. Zubelli • Paolo Zunino

Volume 12
More information about this series at http://www.springer.com/series/10532
Giulio Ventura • Elena Benvenuti
Editors

Advances in Discretization
Methods
Discontinuities, Virtual Elements, Fictitious
Domain Methods

123
Editors
Giulio Ventura Elena Benvenuti
Structural, Building, Geotechnical Eng. Department of Engineering
Politecnico di Torino University of Ferrara
Torino, Italy Ferrara, Italy

ISSN 2199-3041 ISSN 2199-305X (electronic)

SEMA SIMAI Springer Series
ISBN 978-3-319-41245-0 ISBN 978-3-319-41246-7 (eBook)
DOI 10.1007/978-3-319-41246-7
Library of Congress Control Number: 2016950209

© Springer International Publishing Switzerland 2016

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made.

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG Switzerland
Preface

This volume of the SEMA SIMAI Springer Series brings together selected contribu-
tions presented at the international conference “eXtended Discretization MethodS”
(X-DMS), held during September 2015 in Ferrara, Italy. The conference was one of
the thematic conferences supported by the European Community in Computational
Methods in Applied Sciences (ECCOMAS) and also one of the special interest
conferences sponsored by the International Association for Computational Mechan-
ics (IACM). Twelve minisymposia, more than one hundred oral presentations, and
plenary lectures given by eminent personalities in the computational mechanics
research field contributed to the scientific value of the event.
In gathering some of the most interesting contributions at the X-DMS 2015
Conference, the book aims to disseminate ideas and to promote discussion among
researchers with an interest in the development and application of computational
methods in science and technology. In line with this objective, the volume addresses
some of the most advanced discretization methods for the numerical analysis of a
variety of physical problems. In recent years, the efforts of the scientific community
in computational mechanics have especially focused on improving both the overall
computational efficiency and the versatility of the methods, including the addition of
special features of the solution directly in the approximation and/or discretization
space. The results of these efforts can be found in a wide range of computational
methodologies, including partition of unity finite element methods (meshfree,
XFEM, GFEM), virtual element and fictitious domain methods, special techniques
for static and evolving interfaces, multiscale discretization, strong discontinuity
approaches. The selected contributions in this volume recall the main aspects of
some of these methodologies, demonstrating their potentialities and possibilities for
application.
The book is organized into four parts. Part I focuses on the proposals of numerical
schemes for simulations in porous and fractured media; here the challenge is to
handle effectively complex geometries coupled with complex physical problems.
Part II deals with some of the most advanced recent techniques, based on hybrid
and extended discretization methods, for fracture and interface problems. Evolving
fractures in polycrystalline materials, crack lip contact modelling, and procedures

v
vi Preface

for the computation of stress intensity factors are some of the addressed topics.
Part III is devoted to contributions on polygonal and polyhedral methods: these
methods consist in using general polytopes, as opposed to more standard tetrahedra
and hexahedra, for the discretization of partial differential equations. Specialized
forms of the discontinuous Galerkin method and the virtual element method
represent the core of this section. In Part IV, recent advances involving extended
finite element methods and fictitious domain methods are introduced; the goals
are especially to overcome some fundamental problems relating to these methods,
which are basically integration at the element level and ill-conditioning of the
resulting system of equations.
A common feature of all the selected contributions is the direct link between
computational methodologies and their application to different engineering topics.
This mix of theory and application reveals an underlying cooperation between
mathematicians and engineers and highlights the way in which the scientific world
is reacting to the increasing demand for simulations to contribute to the development
of sustainable future technologies in engineering, biomedicine, and environmental
sciences.
The Editors of this volume wish to express their sincere gratitude to all the actors
involved in the X-DMS 2015 Conference, from the Institutions and Sponsors to the
Scientific Committee, the organizers of the minisymposia, the plenary speakers, and
all the participants.
Finally, special acknowledgments are due to Prof. Luca Formaggia, Editor-in-
Chief of the SEMA SIMAI Springer Series, and to the Springer Milan editorial
office for offering the opportunity to compile this volume.

Torino, Italy Giulio Ventura

Ferrara, Italy Elena Benvenuti
May 2016
Contents

Part I Enriched Methods for Flow and Mechanics

in Heterogeneous Media
A Mixed Finite Element Method for Modeling the Fluid
Exchange Between Microcirculation and Tissue Interstitium . . . . . . . . . . . . . . 3
Domenico Notaro, Laura Cattaneo, Luca Formaggia, Anna Scotti,
and Paolo Zunino
On a PDE-Constrained Optimization Approach for Flow
Simulations in Fractured Media . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 27
Sandra Pieraccini and Stefano Scialò
A Review of the XFEM-Based Approximation of Flow
in Fractured Porous Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 47
Bernd Flemisch, Alessio Fumagalli, and Anna Scotti

Part II Enhanced Finite Element Formulations for Fracture

and Interface Problems
Modeling of Fracture in Polycrystalline Materials .. . . . . .. . . . . . . . . . . . . . . . . . . . 79
Steffen Beese, Stefan Loehnert, and Peter Wriggers
eXtended Hybridizable Discontinuous Galerkin (X-HDG) for
Void and Bimaterial Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 103
Ceren Gürkan, Esther Sala-Lardies, Martin Kronbichler,
and Sonia Fernández-Méndez
Crack Lip Contact Modeling Based on Lagrangian Multipliers
with X-FEM. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 123
Yuan Jin, Olivier Pierard, Eric Wyart, and Eric Béchet
Stress Intensity Factors Through Crack
Opening Displacements in the XFEM . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 143
Markus Schätzer and Thomas-Peter Fries

vii
viii Contents

Part III Polygonal and Polyhedral Methods

The Virtual Element Method for Underground Flow
Simulations in Fractured Media . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 167
Matías Fernando Benedetto, Stefano Berrone, and Andrea Borio
Adaptive Discontinuous Galerkin Methods on Polytopic Meshes . . . . . . . . . . 187
Joe Collis and Paul Houston

Part IV Advances in XFEM/Fictitious Domain Methods

Stabilized X-FEM for Heaviside and Nonlinear Enrichments . . . . . . . . . . . . . . 209
Giulio Ventura and Claudia Tesei
An Adaptive Fictitious Domain Method for Elliptic Problems . . . . . . . . . . . . . 229
Stefano Berrone, Andrea Bonito, and Marco Verani
Higher-Order Accurate Integration for Cut Elements
with Chen-Babuška Nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 245
Thomas-Peter Fries
 Part I
Enriched Methods for Flow and Mechanics
in Heterogeneous Media
A Mixed Finite Element Method for Modeling
the Fluid Exchange Between Microcirculation
and Tissue Interstitium

Domenico Notaro, Laura Cattaneo, Luca Formaggia, Anna Scotti,

and Paolo Zunino

Abstract Thanks to dimensional (or topological) model reduction techniques,

small inclusions in a three-dimensional (3D) continuum can be described as one-
dimensional (1D) concentrated sources, in order to reduce the computational cost
of simulations. However, concentrated sources lead to singular solutions that still
require computationally expensive graded meshes to guarantee accurate approxi-
mation. The main computational barrier consists in the ill-posedness of restriction
operators (such as the trace operator) applied on manifolds with co-dimension
larger than one. We overcome the computational challenges of approximating
PDEs on manifolds with high dimensionality gap by means of nonlocal restriction
operators that combine standard traces with mean values of the solution on low
dimensional manifolds. This new approach has the fundamental advantage of
enabling the approximation of the problem using Galerkin projections on Hilbert
spaces, which could not be otherwise applied because of regularity issues. This
approach, previously applied to second order PDEs, is extended here to the mixed
formulation of flow problems with applications to microcirculation. In this way we
calculate, in the bulk and on the 1D manifold simultaneously, the approximation
of velocity and pressure fields that guarantees good accuracy with respect to mass
conservation.

1 Introduction

The ultimate objective of the project is to perform large scale simulations of

microcirculation. In the context of blood flow, the application of geometrical model
reduction techniques plays an essential role, see for example [10, 18]. In particular,
small vessels embedded into a continuum can be described as one-dimensional (1D)
concentrated sources, in order to reduce the computational cost of simulations.

D. Notaro • L. Cattaneo • L. Formaggia • A. Scotti • P. Zunino ()

MOX, Department of Mathematics, Politecnico di Milano, piazza Leonardo da Vinci 32,
20133 Milano, Italy
e-mail: [email protected]; [email protected]; [email protected];
[email protected]; [email protected]

© Springer International Publishing Switzerland 2016 3

G. Ventura, E. Benvenuti (eds.), Advances in Discretization Methods,
SEMA SIMAI Springer Series 12, DOI 10.1007/978-3-319-41246-7_1
4 D. Notaro et al.

Although the coupling of three-dimensional (3D) continua with embedded (1D)

networks arises in applications of paramount importance such as microcirculation,
flow through perforated media and the study of reinforced materials, it has not been
well investigated yet.
Two remarkable examples of methods that were previously proposed to over-
come the challenges of simulating small objects into a continuum are the immersed
boundary methods [15, 17, 22] and the fictitious domain methods [11, 12, 21].
Although they share some similarities with the approach that we pursue here,
they have never been applied for solving coupled partial differential equations on
embedded domains.
In the particular case of microcirculation, many ad-hoc approaches have been
proposed. Since capillaries can be modelled as long and narrow cylindrical vessels,
asymptotic expansions that exploit the large aspect ratio of the channel can be
derived to approximate the fluid exchange from one capillary to the surrounding
tissue. This idea has been successfully exploited to study the microvascular flow in
simple arrays of capillaries [1, 8, 9]. However, vascular networks are characterized
by a complex, possibly irregular geometry. The previous semi-analytic methods may
be hardly applied to realistic configurations. We believe that numerical methods
may override this obstacle. For example, the method of Green’s functions, has been
extensively applied to the study complex vascular networks of tumors [13, 19, 20].
In this work we aim to move away from ad-hoc approaches and cast the micro-
circulation problem into a new unified framework to formulate and approximate
coupled partial differential equations (PDEs) on manifolds with heterogeneous
dimensionality. The main computational barrier consists in the ill-posedness of
restriction operators (such as the trace operator) applied on manifolds with co-
dimension larger than one. Following the approach introduced in [6, 7, 14], we will
overcome the computational challenges of approximating PDEs on manifolds with
high dimensionality gap. The main idea consists of introducing nonlocal restriction
operators that combine standard traces with mean values of the solution on low
dimensional manifolds, in order to couple the problem solution in 3D with the one in
1D. This new approach has the fundamental advantage to enable the approximation
of the problem using Galerkin projections on Hilbert spaces, which could not be
otherwise applied, because of regularity issues.
Within this general framework, the specific objective of this work is to formulate
the microcirculation problem as a system of coupled 1D and 3D partial differential
equations governing the flow through the capillary network and the interstitial
volume, respectively. In order to obtain a good approximation of pressure and
velocity fields, and in particular to satisfy mass conservation, we formulate the
problem in mixed form. Then, we derive a discretization method based on mixed
finite elements. Before moving forward to address applications of the method to
study pathologies related to microcirculation, such as cancer [3, 4, 16], we address
here a thorough validation of the solver based on two benchmark problems.
Mixed Finite Elements for Flow in Microcirculation and Tissue 5

2 Model Set Up

We study a mathematical model for fluid transport in a permeable biological

tissue perfused by a capillary network. The domain where the model is defined is
composed by two parts, ˝ and , denoting the interstitial volume and the capillary
bed respectively. We assume that the capillaries can be described as cylindrical
vessels and  denotes the centerline of the capillary network. The capillary radius,
R, is for simplicity considered to be constant. We decompose the network  into
individual branches i . Branches are parametrized by the arc length si ; a tangent
unit vector i is also defined over each branch, defining in this way an arbitrary
branch orientation. Differentiation over the branches is defined using the tangent
unit vector as @si :D r  i on i , i.e. @si represents the projection of r along i .
The blood flow along each branch is described by Poiseuille’s law for conservation
of momentum and mass:

R2 @pv;i @uiv
uiv D  i ; R2 D gi on i ; (1)
8 @si @si

where gi is the transmural flux leaving the vessel. As a consequence of the

geometrical assumptions, the vessel velocity has fixed direction and unknown scalar
component along the branches, namely uiv D uiv i . We shall hence formulate the
vessel problem using the scalar unknown uv . The governing flow equations for the
whole network  are obtained by summing (1) over the index i.
We consider the interstitial volume ˝ as an isotropic porous medium, described
by the Darcy’s law, namely

1
ut D  IK rpt ; (2)


where ut is the average velocity vector in the tissue, IK D kI is the isotropic

permeability tensor,  is the viscosity of the fluid and pt is the fluid pressure.
The coupled problem for microcirculation and interstitial flow reads as follows
8
ˆ
ˆ ut C rpt D 0 in ˝;
ˆ
ˆ k
ˆ
ˆ
ˆ
ˆ
ˆ
< r  ut  f . pt ; pv /ı D 0 in ˝;
8 @pv (3)
ˆ
ˆ u C D0 in ;
ˆ
ˆ 2 v
ˆ
ˆ R @s
ˆ
ˆ @uv 1
:̂ C f . pt ; pv / D 0 in :
@s R2
For brevity, we avoid to number each equation of systems. In the remainder, we
will refer to single sub-equations within a system using letter numbering from top
6 D. Notaro et al.

to bottom (which does not explicitly appear in the expression of the equation, this
notation will be adopted throughout the entire manuscript), e.g. (3) a to d.
The constitutive law for blood leakage from the capillaries to the tissue is
provided by means of Starling’s law of filtration,

f . pt ; pv / D 2RLp . pv  pN t /; (4)

with
Z 2
1
pN t .s/ D pt .s; /Rd: (5)
2R 0

Before proceeding, we write the equations in dimensionless form. We choose

length, velocity and pressure as primary variables for the analysis. The correspond-
ing characteristic values are: (1) the average spacing between capillary vessels d,
(2) the average velocity in the capillary bed U, and (3) the average pressure in the
interstitial space P. The dimensionless groups affecting our equations are:
R
– R0 D , non-dimensional radius;
d
k P
– t D , hydraulic conductivity of the tissue;
 Ud
P
– Q D 2R0 Lp , hydraulic conductivity of the capillary walls;
U
R04 Pd
– v D , hydraulic conductivity of the capillary bed,
8 U
and the corresponding dimensionless equations read as follows
8
ˆ 1
ˆ
ˆ ut C rpt D 0 in ˝;
ˆ
ˆ t
ˆ
ˆ
ˆ
ˆ
ˆ
< r  ut  Q . pv  pN t / ı D 0 in ˝;
(6)
ˆ
ˆ R02 @pv
ˆ
ˆ uv C D0 in ;
ˆ
ˆ v @s
ˆ
ˆ
ˆ @uv 1
:̂ C Q . pv  pN t / D 0 in :
@s R02
For simplicity of notation, we used the same symbols for the dimensionless
variables, i.e velocities and pressure scaled by U and P, respectively.
Remark 1 Equations (6) b and d, counting from the top, can be combined up to
obtain a more meaningful formulation of the mass conservation law, namely

@uv
r  ut C R02 ı D 0 in ˝; (7)
@s
meaning that the total amount of fluid in the domain ˝ [  must be preserved.
Mixed Finite Elements for Flow in Microcirculation and Tissue 7

Fig. 1 On the left, a simple network made by a single Y-shaped bifurcation. Arrows show the flow
orientation of one inflow branch on the left of the bifurcation point and two outflow branches on the
right. On the right, the discretization of vessels network is shown. The domain has been split into
branches, the flow problem is defined over each branch and compatibility conditions are enforced
at the junction point

Boundary conditions will be specified further on for both the tissue and vessel
problems. The imposition of suitable compatibility conditions at the bifurcations
or branching points of the capillary tree is also necessary to guarantee well
posedness of (6). Specifically, we shall enforce conservation of mass and continuity
of total pressure at junctions. Let us introduce these conditions in a simple Y-
shaped bifurcation network (Fig. 1). Since in the reduced 1D model of the capillary
network the cross-section is supposed to be constant over the whole network, the
conservation of flow rate is equivalent to require that in correspondence of the
junction point xM the inflow velocity u0v is equal the sum of the outflow velocities
u1v ; u2v , namely u0v .xM / D u1v .xM /Cu2v .xM /. Similarly, we require the pressure over
each branch to be the same at the junction, namely p0v .xM / D p1v .xM / D p2v .xM /.
The general case of an arbitrary number of critical points, possibly with different
number of inflow and outflow branches, will be described in Sect. 3. Indeed, it is
important to emphasize that such compatibility conditions will be enforced in a
natural way, at the level of the variational formulation.

3 Variational Formulation

In order to obtain the weak formulation of the tissue interstitium problem, we

multiply Eqs. (6) a and b, counting from the top with sufficiently smooth functions
and integrate over the volume ˝, namely
Z Z
1
ut  vt dx C rpt  vt dx D 0; (8)
˝ t ˝
Z Z
.r  ut / qt dx  Q . pv  pN t / ı qt dx D 0: (9)
˝ ˝
8 D. Notaro et al.

We now apply the Green’s theorem to (8) to obtain an anti-symmetric formulation

of the Darcy’s problem in the tissue:
Z Z Z
1
ut  vt dx  pt .r  vt / dx C pt vt  n d.x/ D 0; (10)
˝ t ˝ @˝
Z Z
.r  ut / qt dx  Q . pv  pN t / ı qt dx D 0: (11)
˝ ˝

As concerns the choice of boundary conditions, for simplicity, we enforce a given

pressure distribution over @˝, namely

pt D gt on @˝; (12)

where gt 2 L2 .@˝/. The weak formulation of the problem in ˝ reads

Z Z Z
1
ut  vt dx  pt .r  vt / dx D  gt vt  n d.x/
˝ t ˝ @˝
Z Z
.r  ut / qt dx  Q . pv  pN t / ı qt dx D 0 :
˝ ˝

For the vessel problem we start giving a general functional framework. At this
point, we only require regularity for vessel velocity and pressure over each branch
separately:

[
N [
N
Vv D H 1 .i / Qv D L2 .i /:
iD1 iD1

The definition of trial and test spaces will be revised in the sequel, in the light of
the particular junction conditions we will chose, while no boundary conditions are
enforced in the definition of the spaces. As for the tissue problem, we multiply
Eqs. (6.c),(6.d) by sufficiently smooth test functions and integrate over :
Z Z
R02 @pv
uv vv ds C vv ds D 0; (13)
 v  @s
Z Z
@uv 1
qv ds C Q . pv  pN t / qv ds D 0: (14)
 @s R02 

The integration by parts is not trivial in this case because the vessel variables pv and
uv may be discontinuous at multiple junctions. Let us treat separately the second
integral of (13) and decompose it over the individual branches i :
Z X N Z Z XN
@pv @pv @vv C
vv ds D vv ds D  pv ds C Πpv vv i ; (15)
 @s iD1 i @s  @s iD1
i
Mixed Finite Elements for Flow in Microcirculation and Tissue 9

C
where  i and i represent the inflow and outflow boundaries of i , according to
the orientation i . Let us define the set of the indexes of junction points:
˚ 
J :D j 2 N W sj 2 ; #.Psj /  2 ;

where Psj is the patch of the j-th junction node, i.e. the collection of all branches
joining at the node, and # indicates the counting measure. We also need the
following disjoint partition of the indexes in Psj . According to the orientation unit
vector i , for any branching point sj we distinguish branches that are entering the
node, whose contribution to mass conservation is positive, from branches who are
leaving the node, whose contribution is negative. The former are branches whose
outflow region coincides with the point sj , while for the latter it is the inflow region:
˚ 
Pjout :D i 2 f1; : : : ; Ng W C
i  fsj g ;

˚ 
Pjin :D i 2 f1; : : : ; Ng W 
i  fsj g ;

for all j 2 J . At this point, the fluid mass conservation at each node can be
expressed as follows
X X
uv jC  uv j
i
D 0; 8j 2 J ; (16)
i
i 2 Pjout i 2 Pjin

where in ; out indicate the collection of inflow and outflow boundaries of the
vessel network, i.e. non junction points where the tangent unit vector is inward-
pointing and outward-pointing, respectively. This collection contains the boundary
points, i.e. the extrema that also belong to @˝, but the inclusion may be strict.
However, in this contribution we do not address the issue of network extrema
belonging to ˝,V i.e. we do not consider immersed tips.
In order to enforce such conditions, we proceed as follows. First, we reformulate
the last term in (15) by isolating the terms relative to inflow junction nodes from
those relative to outflow nodes, namely
2 3
X
N
 C X 6 X X 7 out
Πpv vv i D 4 pv vv jC  pv vv j
i
5 C Πpv vv in :
i i
iD1 j2J i 2 Pjout i 2 Pjin

Here, we have implicitly assumed the trace of . pv vv / over i exists for vv smooth
enough, i.e. the evaluation of the product at the extrema of i makes sense.
Furthermore, we write . pv vv / .sj / D pv .sj / vv .sj / for some point sj 2 . This
is feasible if the trace of the pressure exists. Obviously, a general L2 function is not
N that in particular implies compatibility
sufficient, the natural choice is pv 2 C 0 ./,
of pressure values at the junctions. Indeed, if the pressure is continuous at the
10 D. Notaro et al.

junction, we have

pv j
i
 pv .sj /  pv jC 8i 2 Pjin ; k 2 Pjout 8j 2 J : (17)
k

Under that hypothesis, we finally factorize out the pressure and isolate a term that
corresponds to the junction conditions for the velocity test functions, that is
2 3
X 6 X X 7
pv .sj / 4 vv jC  vv j
i
5:
i
j2J i 2 Pjout i 2 Pjin

Then, we weakly enforce mass conservation into the variational formulation by

multiplying (16) by the pressure test functions qv , which act as a Lagrange multiplier
for this constraint, namely
2 3
X 6 X X 7
qv .sj / 4 uv jC  uv j
i
5:
i
j2J i 2 Pjout i 2 Pjin

Finally, after adding the previous term to Eq. (14), the weak formulation of the vessel
problem reads
Z Z
R02 @vv
ds C Πpv vv 
out
uv vv ds  pv in
 v  @s
2 3
(18)
X 6 X X 7
C pv .sj / 4 vv jC  vv j
i
5 D 0;
i
j2J i 2 Pjout i 2 Pjin
Z Z
@uv 1
qv ds C Q .pv  pN t / qv ds
 @s R02 
2 3
(19)
X 6 X X 7
 qv .sj / 4 uv jC  uv j
i
5 D 0:
i
j2J i 2 Pjout i 2 Pjin

Concerning the boundary conditions for the vessels network, the natural choice is
to enforce a given pressure distributions at the inflow and the outflow of the network,
pv D gv on in [out . The generic regularity requirements for the Dirichlet’s datum
are measurability and square-summability, namely gv 2 L2 .in [ out /. In practice,
we consider a constant pressure drop
Pv D Pout v  Pv :
in

(
Pin
v s 2 in
gv .s/ D (20)
Pout
v s 2 out :
Mixed Finite Elements for Flow in Microcirculation and Tissue 11

Since we are considering the mixed formulation of the problem, we enforce such
condition in a weak natural way.
At this point, we combine (10), (11), (18), (19) to obtain the whole weak formu-
lation of our 3D-1D coupled model of fluid exchange between microcirculation and
tissue interstitium. The variational formulation of problem (6) consists of finding
ut 2 Vt ; pt 2 Qt ; uv 2 Vv ; pv 2 Qv s.t.
8
ˆ 1     
ˆ
ˆ ut ; vt ˝  pt ; r  vt ˝ D  gt ; vt  n @˝ 8vt 2 Vt ;
ˆ
ˆ t
ˆ
ˆ
ˆ
ˆ    
ˆ
ˆ r  ut ; qt ˝  Q . pv  pN t / ı ; qt ˝ D 0 8qt 2 Qt ;
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
< R02    
uv ; vv   pv ; @s vv 
ˆ v hP i
ˆ
ˆ P P
D  Πgv vv 
out
ˆ
ˆ C p v .s / vv j C  vv j 
in 8vv 2 Vv ;
ˆ
ˆ j j i  i i
ˆ
ˆ
i
ˆ
ˆ   1  
ˆ
ˆ @s uv ; qv  C Q p  N
p ; q
ˆ
ˆ R02
v t v 
ˆ
ˆ h i
ˆ P P P
:̂  j qv .sj / i uv jC  i uv ji D 0 8qv 2 Qv :

i
(21)

4 Numerical Approximation

The discretization of problem (6) is achieved by means of the finite element method
that arises from the variational formulation (21) combined with a discretization
of the domain. In particular, one of the advantage of our formulation is that the
partitions of ˝ and  are completely independent. Let us now analyze the two
approximations separately.
We denote with Tt h an admissible family of partitions of ˝N into tetrahedrons K
[
˝N D K;
h
K2Tt

that satisfies the usual conditions of a conforming triangulation of ˝. Here, h

denotes the mesh characteristic size, i.e. h D maxK2Tt h kK , being hK the diameter
of simplex K. Moreover, we are implicitly assuming that ˝ is a polygonal
domain. The solutions of (21) a and b, counting from the top are approxi-
mated using discontinuous piecewise-polynomial finite elements for pressure and
12 D. Notaro et al.

Hdiv -conforming Raviart-Thomas finite elements [2] for velocity, namely

˚ 
Ykh :D wh 2 L2 .˝/ W wh jK 2 Pk1 .K/ 8K 2 Tt h ;
˚ 
RTkh :D wh 2 H ..div; ˝/ W wh jK 2 Pk1 .KI Rd / ˚ x Pk1 .K/ 8K 2 Tt h ;

for every integer k  0, where Pk indicates the standard space of polynomials of

degree  k in the variables x D .x1 ; : : : ; xd /. For the simulations presented later on,
the lowest order Raviart-Thomas approximation has been adopted, corresponding to
k D 1 above. In numerical experiments performed on the 3D problem alone (the test
case is not reported here), we have observed quadratic convergence of the pressure
field and linear convergence of the velocity field.
Concerning the capillary network, we adopt the same domain splitting technique
described at the continuous level, obtaining the following discrete domain:

[
N
h D hi ;
iD1

where hi is a finite element mesh on the one-dimensional manifold i , i.e. a

partition of the i-th network branch made by a sufficiently large number of segments.
The solution of sub-equations (21) c and (21) d, counting from the top, over
a given branch i is approximated using continuous piecewise-polynomial finite
element spaces for both pressure and velocity. Since we want the vessel velocity to
be discontinuous at multiple junctions, we define the related finite element space
over the whole network as the collection of the local spaces of the single branches.
Conversely, the pressure has been assumed to be continuous over the network. We
will use the following families of finite element spaces for pressure and velocity,
respectively:
˚ 
N W wh jS 2 Pk .S/ 8S 2 h ;
XhkC1 ./ :D wh 2 C 0 ./
[
N
WhkC2 ./ :D XhkC1 .i / ;
iD1

for every integer k  0. As a result, we use generalized Taylor-Hood elements

on each network branch, satisfying in this way the local stability of the mixed
finite element pair for the network. At the same time, we guarantee that the
pressure approximation is continuous over the entire network . In particular, for
the numerical experiments shown later on we have used the lowest order, that is
k D 1.
Mixed Finite Elements for Flow in Microcirculation and Tissue 13

The discrete formulation arising from (21) is hence easily obtained by adding the
subscript h to the weak continuous formulation: find ut;h 2 Vht ; pt;h 2 Qht ; uv;h 2
Vvh ; pv;h 2 Qhv s.t.
8
ˆ
ˆ 1     
ˆ
ˆ ut;h ; vt;h ˝  pt;h ; r  vt;h ˝ D  gt;h ; vt;h  n @˝ 8vt;h 2 Vht ;
ˆ
ˆ t
ˆ
ˆ
ˆ
ˆ    
ˆ
ˆ
ˆ
ˆ r  ut;h ; qt;h ˝  Q . pv;h  pN t;h / ı ; qt;h ˝ D 0 8qt;h 2 Qht ;
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
< R02    
uv;h ; vv;h   pv;h ; @s vv;h 
ˆ
ˆ v hP i
ˆ P P
ˆ
ˆ C p .s / v j  v j  D  Œ gv;h vv;h 
out
8vv;h 2 Vvh
ˆ
ˆ j v;h j i v;h i C i v;h i in
ˆ
ˆ
ˆ
ˆ   1  
ˆ
ˆ @s uv;h ; qv;h  C Q pv;h  pN t;h ; qv;h 
ˆ
ˆ 02
ˆ
ˆ R
ˆ
ˆ P hP P i
:̂  q v;h .s j / u v;h j C  u v;h j   D 0 8qv;h 2 Qhv ;
j i i  i i

(22)

where gt;h , gv;h indicate the discrete counterparts of continuous boundary data.
We observe that (22) is a generalized saddle-point problem arising from the
combination of local problems with mass conservation constraints (see also (23)),
such as the mixed formulation of Darcy equation and the incompressible flow
on each network branch with junction conditions. Although, we guarantee local
stability of each block, the global well-posedness is an still an open problem, which
is under investigation.

4.1 Algebraic Formulation

Let us now derive the algebraic form of our discrete problem. We define the number
of degrees of freedom of our discrete (finite) spaces as:
   
Nth :D dim Vht ; Mth :D dim Qht ;
   
Nvh :D dim Vvh ; Mvh :D dim Qhv :

Nh Mh Nh Mh
We denote with f'it giD1
t
 f ti giD1
t
and f'iv giD1
v
 f vi giD1
v
the finite element basis
for Vt  Qt and Vv  Qv respectively. These two sets are completely independent,
h h h h
14 D. Notaro et al.

since the 3D and 1D meshes do not conform. We set:

h h
X
Nt
j j
X
Mt
j j
uht .x/ D Ut 't .x/ ; pht .x/ D Pt t .x/ 8x 2 ˝t ;
jD1 jD1

h h
X
Nv
X
Mv
uhv .s/ D Uvj 'vj .s/ ; phv .s/ D Pjv v
j
.s/ 8s 2 ;
jD1 jD1

j Nh j Mh j Nh j Mh
being Ut D f Ut gjD1 t
; Pt D f Pt gjD1
t
; Uv D f Uv gjD1
v
and Pv D f Pv gjD1
v
; the
degrees of freedom of the finite element approximation. Then, by replacing the
linear combinations within the discrete weak form (22) and using the linearity of
the inner product, from (22) we deduce the following linear system:
2 32 3 2 3
M DTtt
6 tt
O O U
76 t7
F
6 t7
6 76 7 6 7
6 Dtt Btt O Btv 7 6 Pt 7 607
6 76 7 6 7
6 76 7D6 7: (23)
6
4
O O Mvv DTvv  JTvv 7 6U 7
5 4 v5
6Fv 7
4 5
O  Bvt Dvv C Jvv Bvv Pv 0

Standard finite element matrices and right hand sides are defined as follows

1 j i
Mtt 2 RNt Nt ;
h h
ŒMtt i;j :D 't ; 't ˝
t
 
Dtt 2 RNt Mt ;
j h h
ŒDtt i;j :D r  't ; ti ˝
 
Dvv 2 RNv Mv ;
h h
ŒDvv i;j :D @s 'vj ; i
v 
 
ŒMvv i;j :D R02 =v 'vj ; 'vi  Mvv 2 RNv Nv ;
h h

  h
ŒFt i :D  gt;h ; 'it  n @˝ F t 2 RN t ;

ŒFv i :D Œ gv;h 'vi 

out h
in F v 2 RN v :

For the implementation of exchange matrices, namely Btt ; Btv ; Bvt ; Bvv , we
define two discrete operators: the first one extracts the mean value of a generic
basis function of Qht , while the second interpolates between Qht and Qhv . For every
node sk 2 h we define T .sk / as the discretization of the perimeter of the vessel
.sk /, see Fig. 2 for an illustration. For simplicity, we assume that .sk / is a circle
of radius R defined on the orthogonal plane to h at point sk . The set of points
of T .sk / is used to interpolate the basis functions ti . Let us introduce a local
discrete interpolation matrix ˘  .sk / which returns the values of each test function
t on the set of points belonging to T .sk /. Then, we consider the average operator
i
Mixed Finite Elements for Flow in Microcirculation and Tissue 15

Fig. 2 Illustration of the vessel with its centerline h , a cross section, its perimeter .sk / and
its discretization T .sk / used for the definition of the interface operators N vt W Qht ! Qhv and
tv W Qhv ! Qht

N vt W Qht ! Qhv such that qN t D N vt qt . The matrix N̆ vt that corresponds to this

operator belongs to RMv Mt and it is constructed such that each row is defined as,
h h

N̆ vt jk D wT .sk / ˘  .sk / k D 1; : : : ; Mvh (24)

where w are the weights of the quadrature formula used to approximate the integral
Z 2
1
qN t .s/ D qt .s; / R d
2R 0

on the nodes belonging to T .sk /. The discrete interpolation operator tv W Qhv !
Qht returns the value of each basis function belonging to Qht in correspondence of
nodes of Qhv . In algebraic form it is expressed as an interpolation matrix ˘ tv 2
RMv Mt . Using these tools we obtain:
h h

Btt D Q ˘ Tvt MPvv N̆ vt ; (25)

Btv D Q ˘ Tvt MPvv ; (26)
Bvt D Q=R 02
MPvv N̆ vt ; (27)
Bvv D Q=R02 MPvv ; (28)

being MPvv the pressure mass matrix for the vessel problem defined by
   
MPvv i;j
:D j
v ; v :
i

Concerning the implementation of junction compatibility conditions, we intro-

duce a linear operator giving the restriction with sign of a basis function of Vvh over
16 D. Notaro et al.

a given junction node. For a given k 2 J , we define Rk W Vvh ! R such that:

(
j
C 'v .sk / j in hl ^ l 2 Pkout
Rk . 'vj / :D j
(29)
 'v .sk / j in hl ^ l 2 Pkin

for all j D 1; : : : ; Nvh , where the expression “j in hl ” means that the j-th dof is
linked to some vertex of the l-th branch. Note that we are implicitly using the usual
property of Lagrangian finite element basis functions, i.e. that they vanish on all
nodes except the related one. As a consequence, our definition is consistent for all
junction vertexes. Indeed, Rk may only assume values 1; 0; C1 and in particular
j
Rk . 'v / D 0 for all couples of indexes .k; j/ that are uncorrelated. Furthermore,
the definition of Rk can be trivially extended to all network vertexes. Using this
operator, the generic .i; j/ element of Jvv may be computed as follows
X
ŒJvv i;j D  Rk . 'vj / v .sk / :
i
(30)
k2J

5 Numerical Experiments

We validate the mixed-finite element solver through the following test cases,
illustrated in Fig. 3, which have been designed to obtain sufficient generality with a
straightforward interpretation of the results:

Fig. 3 (Left) Computational domain for test-case I. The discrete network h is made by a single
capillary vessel immersed in a unitary slab of tissue interstitium, ˝h . We have used a discretization
step h D 0:05 for both the 1D and 3D problems. (Right) Computational domain for test-case II. The
discrete vessels network h is made by three capillaries joined junction point xM D .0:5; 0:5; 0:5/:
0h entering branch, 1h and 2h exiting branches. The tissue interstitium domain ˝h is a unit cube.
Again, we have used a discretization step h D 0:05 for both the 1D and 3D problems
Mixed Finite Elements for Flow in Microcirculation and Tissue 17

(1) Coupled 3D-1D problem on a single branch;

(2) Coupled 3D-1D problem on a Y-shaped bifurcation.
In this way, we address the two main modeling issues: (1) first, we test the ability
of the computational model to approximate the coupling between 3D and 1D
equations; (2) second, we verify that the assembly of junction conditions works
properly.

5.1 Coupled 3D-1D Problem on a Single Branch

For such a simple setting, we can easily isolate the exchange terms. The 3D-1D
coupled problem is given by (6). In this case, the integration by parts in (6)(iii)
is standard since there are not any junction points. As a consequence, we replace
condition (15) with the following:
Z Z
@pv @vv
ds C Πpv vv 
out
vv ds D  pv in
 @s  @s
Z 1
@vv
D pv ds C pv .1/ vv .1/  pv .0/ vv .0/ :
0 @s

Therefore, we obtain the following linear system:

2 32 3 2 3
6
Mtt DTtt O O 7 6 Ut 7 6 F t 7
6 76 7 6 7
6
6
Dtt Btt O Btv 7 6P 7 6 0 7
76 t 7 D 6 7 :
6 (31)
T 7 6 7 6 7
6
4
O O Mvv Dvv 7 6U 7 6F 7
5 4 v5 4 v5
O  Bvt Dvv Bvv Pv 0

We recall that submatrices in (31) have been defined in Sect. 4.1. Nevertheless,
according to the above expression of vessel boundary term it is possible to specify
the right hand side, namely
2 3
gv;h .0/
6 7
6 0 7
6 7
1 6 :: 7
Fv D Πgv;h 'vi 0  6 : 7 (32)
6 7
6 7
4 0 5
gv;h .1/

being gv;h the discrete counterpart of the vessel boundary datum. In the last equality
we used the fundamental property of finite element basis functions. Note that (31)
equals the generic linear system (23) in the special case Jvv D O.
18 D. Notaro et al.

5.1.1 Numerical Results

For the tissue sample ˝h we use a tetrahedral structured mesh, Th , with charac-
teristic size h D 1=20; the same step has been used for the network discretization
h , resulting in 48,000 elements for the approximation of interstitial volume and 60
elements for the discrete network. We prescribe the following boundary conditions:

pt j@˝ D 0 ; pv .0/ D 1:0 ; pv .1/ D 0:5 : (33)

For the solution of the linear system (31) we developed a C++ code based
on GetFEM++ (see https://home.gna.org/getfem), an open-source general purpose
finite element library. Specifically, we applied the direct solver SuperLU 3.0 (see
http://crd.lbl.gov/verb~xiaoye/SuperLU). Numerical solutions are shown in Fig. 4.
These plots show qualitatively that the definition and implementation of the method
works properly. In order to find a quantitative way to validate our numerical method
we exploit the exact solution proposed by Chapman and Shipley [5] for the single

Fig. 4 Coupling between the vessel and tissue interstitium. Numerical solutions obtained with
mesh size h D 0:05 and parameters t D v D 1, R0 D 1, Q D 1. On the left a double-check for
pressure exchange: (top-left) visualization of tissue and vessel pressures at the transversal medium
plane, (bottom-left) a 3D qualitative representation. On the right the velocity exchange: (top-right)
visualization of tissue and vessel velocities at the axial medium plane, (bottom-right) a 2D view of
the vector field
Mixed Finite Elements for Flow in Microcirculation and Tissue 19

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 5 Capillary pressure as function of arclength s for different vascular permeabilities LOp D
104 ; 2  106 ; 106 ; 5  107 ; 107 ; 108 . To be compared with Fig. 7 in [5]

branch problem. In that work the authors model a fluid flow through the leaky
neovasculature and porous interstitium of a solid tumor, in particular they consider
the simplest case of an isolated capillary immersed in a tumor tissue, giving rise to
the same problem addressed here.
Finally, in order to reproduce numerical results of [5] we choose the non-
dimensional parameters of the problem as follows

R0 D 102 ; t D 4 ; v D  R03 =8LOp ; Q D 2 ; (34)

˚ 
where LOp 2 104 ; 2  106 ; 106 ; 5  107 ; 107 ; 108 is an array of non-
dimensional vascular permeabilities used in the numerical tests of [5]. In Fig. 5 we
represent the capillary pressure as a function of arc-length for different vascular
permeabilities. We can observe perfect agreement with the plots shown in [5]
(not reported here). Moreover, we notice that for the lowest value of the vascular
permeability LOp , corresponding to an almost impermeable vessel, the computational
model predicts a linearly decreasing pressure, in agreement with the Poiseuille
equation that governs the flow. Conversely, for high permeability values there is
a substantial deviation from the linear trend because the leakage dominates over the
axial flow component.
In addition, a sensitivity analysis has been performed to investigate the influence
of the relative position of the 3D and 1D grids. To this purpose, we simulated
three different configurations in which the network vertexes coincide with particular
points of the 3D mesh, as shown in Fig. 6. Numerical results of similar test cases,
where the 1D mesh is slightly shifted to coincide with the location of Fig. 6 (top),
20 D. Notaro et al.

B
C

Ut
2,72e-04

0,00015757

4,31e-05

Z
Y X

Ut
2,72e-04

0,0002255

1,79e-04

Z
Y X

Ut
2,72e-04

0,00022574

1,80e-04

Z
Y X

Fig. 6 (Top panel) Proposed configurations for 1D/3D mesh correlation analysis. The red dots
A; B; C indicate three meaningful configurations w.r.t. the distribution of Raviart-Thomas dof
(arrows). We show below the sensitivity analysis for the 1D/3D mesh coupling. Numerical
solutions have been obtained with mesh size h D 0:1 and parameters t D v D 1, R0 D 1,
Q D 104 . For each of the three configurations A; B; C, we extract the smallest patch of elements
intersected by the 1D mesh (truncated along the axial direction for visualization purposes). The
local velocity field is also displayed together with its magnitude (color scale)
Mixed Finite Elements for Flow in Microcirculation and Tissue 21

are given in Fig. 6 (bottom). These results suggest that when the 1D mesh is not
aligned with edges or faces of the 3D one, the velocity field in the neighborhood of
the vessel looks smooth and symmetric. A similar conclusion holds true also when
the 1D branch lays on the 3D mesh faces. However, a problematic case is observed
when the 1D mesh entirely coincides with edges of the 3D one. In this case, the
velocity field looses cylindrical symmetry around the 1D capillary. A preliminary
and heuristic interpretation of this behavior can be found observing that the Raviart-
Thomas degrees of freedom are located on the element faces. For this reason, the
3D velocity field is not uniquely defined on tetrahedral element edges. When the 1D
source term is exactly located on the element edges, there is an inconsistency in the
approximation of the velocity field.

5.2 Coupled 3D-1D Problem on a Y-Shaped Bifurcation

We aim to validate the imposition of the mass conservation constraint at the junction.
We observe that the conservation of total pressure at the junction xM ,

p0v .xM / D p1v .xM / D p2v .xM /  pv .xM / ; (35)

is automatically ensured thanks to the use of continuous finite elements for the ves-
sels pressure approximation. Conversely, in order to impose the mass conservation
constraint

u0v .xM / D u1v .xM / C u2v .xM / ; (36)

we proceed as in (30). In practice, in the simple Y-shaped configuration, we first

identify the FEM degrees of freedom (dofs) related to the same junction node. For
those dofs we add in some specific entries of the problem matrix C1 for each inflow
branch and 1 for each outflow branch.

5.2.1 Numerical Results

We apply again the SuperLU direct method to solve the linear system (31) and
boundary conditions as in the second test-case, (33).
We notice that both vessels pressure and velocity confirm the expected behavior
along the network: the former is continuous everywhere while the latter is split into
two after the junction. We then conclude that conditions (35) and (36) are fulfilled
also at the numerical level. Furthermore, Fig. 7 (bottom panel) confirms that, also in
this case, the 3D/1D coupling behaves correctly.
Finally, we present in Fig. 8 a comparison against the pressure formulation of the
3D/1D coupled problem (6), previously proposed in [3, 4]. In that work, only the
pt and pv variables were approximated, in particular using piecewise linear finite
22 D. Notaro et al.

Pv Uv
Y 4,996e-01 0,7498 9,999e-01 0,000e+00 0,1634 3,269e-01
X
Z

Pv Pt Ut
0.6 0.3 Uv
0.05 5

0.56 0.175
–0.025 2.5

Z 0.5 0 0 0
Z
Y Y
X X

Fig. 7 (Top) Numerical solution of the 1D vessel problem, .pv;h ; uv;h / obtained with h D 0:05
and unitary parameters v D 1, R0 D 1, Q D 1. As expected the pressure (left) is almost linearly
decreasing between the imposed boundary values 1 and 0; velocity (right) is almost constant over
each branch and it halves after the junction. (Bottom) Visualization of the 3D/1D coupled pressure
and velocity fields

elements. The velocity field was reconstructed a-posteriori, after calculating the
pressure gradients as constant vector functions over each element. By comparing
the results of the new mixed-form model (left column) against those of the pressure-
form model (right column) it appears, as expected, that we lose accuracy in 3D
pressure approximation, because we use piecewise constant approximation instead
of piecewise linears, but we visibly gain a better approximation of the 3D velocity
field. This is ultimately a very important advantage. Indeed, following the work of
[3, 16] we are planning to combine these simulations with mass transport problems
for drug delivery through the microcirculation, where the velocity field is adopted
to model advection.
Mixed Finite Elements for Flow in Microcirculation and Tissue 23

Z
X

Z
X

Y
X

Fig. 8 Comparison between numerical predictions of pressure-form (right column) and mixed-
form (left column) models within identical settings. In both cases we adopted h D 0:05 and
dimensionless parameters R0 D 0:1, t D 1, v D 1, Q D 1

6 Conclusions

We have proposed a mixed finite element formulation for coupled incompressible

flow problems defined on a 1D domain embedded into a 3D porous medium.
Because of the non standard coupling operators based on nonlocal restriction of the
3D solution to the 1D manifold, the definition and implementation of the method is
24 D. Notaro et al.

challenging. We have implemented a non-standard finite element method into a C++

solver and the purpose of this work was to carefully validate it on two benchmark
problems. The numerical solutions feature the expected behavior and confirm the
correct functioning of the code. The mixed finite element approximation for 3D/1D
coupled incompressible flow problems is a significant improvement with respect to
the previously available solver based on the primal pressure formulation, because
the velocity field is a variable of the problem and mass conservation constraints
are directly enforced. Indeed, we are planning to exploit the better approximation
properties of the velocity field, to combine the present solver with mass transport
equations.

References

1. Blake, T., Gross, J.: Analysis of coupled intra- and extraluminal flows for single and multiple
capillaries. Math. Biosci. 59(2), 173–206 (1982)
2. Brezzi, F., Fortin, M.: Mixed and Hybrid Finite Element Methods, Springer Series in
Computational Mathematics, vol. 15. Springer, New York (1991). doi:10.1007/978-1-4612-
3172-1
3. Cattaneo, L., Zunino, P.: A computational model of drug delivery through microcirculation to
compare different tumor treatments. Int. J. Numer. Methods Biomed. Eng. 30(11), 1347–1371
(2014). doi:10.1002/cnm.2661
4. Cattaneo, L., Zunino, P.: Computational models for fluid exchange between microcirculation
and tissue interstitium. Netw. Heterog. Media 9(1), 135–159 (2014)
5. Chapman, S.J., Shipley, R.J., Jawad, R.: Multiscale modeling of fluid transport in tumors. Bull.
Math. Biol. 70(8), 2334–2357 (2008). doi:10.1007/s11538-008-9349-7
6. D’Angelo, C.: Finite element approximation of elliptic problems with dirac measure terms
in weighted spaces: applications to one- and three-dimensional coupled problems. SIAM J.
Numer. Anal. 50(1), 194–215 (2012)
7. D’Angelo, C., Quarteroni, A.: On the coupling of 1d and 3d diffusion-reaction equations:
application to tissue perfusion problems. Math. Models Methods Appl. Sci. 18(8), 1481–1504
(2008)
8. Flieschman, G., Secomb, T., Gross, J.: Effect of extravascular pressure gradients on capillary
fluid exchange. Math. Biosci. 81(2), 145–164 (1986)
9. Fleischman, G., Secomb, T., Gross, J.: The interaction of extravascular pressure fields and fluid
exchange in capillary networks. Math. Biosci. 82(2), 141–151 (1986)
10. Formaggia, L., Gerbeau, J., Nobile, F., Quarteroni, A.: On the coupling of 3d and 1d Navier-
stokes equations for flow problems in compliant vessels. Comput. Methods Appl. Mech. Eng.
191(6–7), 561–582 (2001). doi:10.1016/S0045-7825(01)00302-4
11. Glowinski, R., Pan, T.W., Periaux, J.: A fictitious domain method for Dirichlet problem and
applications. Comput. Methods Appl. Mech. Eng. 111(3–4), 283–303 (1994)
12. Glowinski, R., Pan, T.W., Hesla, T., Joseph, D.: A distributed lagrange multiplier/fictitious
domain method for particulate flows. Int. J. Multiph. Flow 25(5), 755–794 (1999)
13. Hicks, K., Pruijn, F., Secomb, T., Hay, M., Hsu, R., Brown, J., Denny, W., Dewhirst, M.,
Wilson, W.: Use of three-dimensional tissue cultures to model extravascular transport and
predict in vivo activity of hypoxia-targeted anticancer drugs. J. Natl. Cancer Inst. 98(16),
1118–1128 (2006)
14. Lesinigo, M., D’Angelo, C., Quarteroni, A.: A multiscale Darcy-Brinkman model for fluid
flow in fractured porous media. Numer. Math. 117(4), 717–752 (2011)
Mixed Finite Elements for Flow in Microcirculation and Tissue 25

15. Mittal, R., Iaccarino, G.: Immersed boundary methods. Annu. Rev. Fluid Mech. 37, 239–261
(2005)
16. Nabil, M., Decuzzi, P., Zunino, P.: Modelling mass and heat transfer in nano-based cancer
hyperthermia. R. Soc. Open Sci. 2(10), 150447, 1–17 (2015). doi:10.1098/rsos.150447
17. Peskin, C.S.: The immersed boundary method. Acta Numer. 11, 479–517 (2002).
doi:10.1017/S0962492902000077
18. Quarteroni, A., Formaggia, L., Veneziani, A.: Cardiovascular mathematics: modeling and
simulation of the circulatory system. Model. Simul. Appl. 1, 1–512 (2009)
19. Secomb, T., Hsu, R., Braun, R., Ross, J., Gross, J., Dewhirst, M.: Theoretical simulation of
oxygen transport to tumors by three-dimensional networks of microvessels. Adv. Exp. Med.
Biol. 454, 629–634 (1998)
20. Secomb, T., Hsu, R., Park, E., Dewhirst, M.: Green’s function methods for analysis of oxygen
delivery to tissue by microvascular networks. Ann. Biomed. Eng. 32(11), 1519–1529 (2004)
21. Yu, Y., Baek, H., Karniadakis, G.: Generalized fictitious methods for fluid-structure interac-
tions: analysis and simulations. J. Comput. Phys. 245, 317–346 (2013)
22. Zhang, L., Gerstenberger, A., Wang, X., Liu, W.K.: Immersed finite element method. Comput.
Methods Appl. Mech. Eng. 193(21–22), 2051–2067 (2004). doi:10.1016/j.cma.2003.12.044
On a PDE-Constrained Optimization Approach
for Flow Simulations in Fractured Media

Sandra Pieraccini and Stefano Scialò

Abstract In the present work we consider the problem of performing underground

flow simulations in fractured media, following the Discrete Fracture Network (DFN)
model. We will focus on a quite recent approach to the problem, based on a
PDE-constrained optimization formulation, which allows for the use of totally non-
conforming meshes on the network. In this way arbitrarily complex DFNs can be
effectively tackled, without requiring any modification of the geometry of the net-
work. Extended numerical simulations are reported demonstrating the performances
of the proposed method, and highlighting its robustness in handling networks with
hard-to-mesh configurations, such as extremely narrow angles between intersecting
fractures. The problem of advection-diffusion of pollutant species in networks of
fractures is also addressed in a time-dependent framework, using the optimization-
based approach both to derive the Darcy velocity and to solve the transport problem
at each time frame.

1 Introduction

Several critical and up-to-date applications involving underground exploitation,

such as geothermal applications, enhanced oil & gas production, geological storage,
require effective tools for performing underground flow simulations, possibly in
very large scale domains representing basins.
Within this framework, a possible approach for modelling the underground
is given by the Discrete Fracture Network (DFN) model [1, 17, 20], in which
the fractures in the rock matrix are represented as planar polygons intersecting
each other along segments called traces. The surrounding rock matrix is assumed
here to be impervious, so that no flux exchange occurs among the fractures and
the surrounding rock matrix. Nevertheless, flux exchanges occur among fractures
through the traces. The quantity under investigation is the hydraulic head H, which

S. Pieraccini • S. Scialò ()

Dipartimento di Scienze Matematiche, Politecnico di Torino, Corso Duca degli Abruzzi,
24 - 10129 Torino, Italy
e-mail: [email protected]; [email protected]

© Springer International Publishing Switzerland 2016 27

G. Ventura, E. Benvenuti (eds.), Advances in Discretization Methods,
SEMA SIMAI Springer Series 12, DOI 10.1007/978-3-319-41246-7_2
28 S. Pieraccini and S. Scialò

is related to the fluid pressure p, being defined as H D p=.g/ C z, where g is the

gravitational acceleration,  the fluid density and z the elevation. The flow on each
fracture is modelled by means of the Darcy law, and suitable matching conditions
are imposed at fracture intersections, over the whole network, in order to ensure
head continuity and flux balance.
The major complexity in approaching the problem of the effective simulation of
the flow in large networks of fractures lies in the geometrical size and complexity
of the computational domain, characterized by a large number of fractures, with
mutually intersecting traces that might form extremely narrow angles, or might
be extremely close to each other. These hard-to-mesh configurations make the use
of standard approaches based on conforming grids infeasible, and require targeted
methods in order to ease the meshing process. Solutions proposed in [19, 34–37]
make use of the mortar method, sometimes in conjunction with modifications of the
geometry. In [18, 31, 32] problem complexity is tackled resorting to the resolution
of lower-dimensional problems; in [28] an algorithm is designed, suitable for
stochastic analysis in DFNs, capable of detecting and rejecting DFN configurations
too complex to be meshed. Other techniques to solve complex flow configurations
in DFNs can be found in [21, 22, 26, 29, 30].
In rather recent papers [11–13, 15] a new approach has been proposed in order to
tackle any kind of geometrical configuration without requiring any form of mesh
conformity at fracture intersections. The method has a strongly parallel nature
and, thanks to the nonconformity, allows for an independent mesh generation
process on each fracture of the DFN [14], thus overcoming the complexities
encountered by classical approaches. The method can be used with standard finite
element discretizations, but it is especially designed in order to take advantage of
non-standard discretization techniques, such as the XFEM [16] and the VEM [8].
In the present work we briefly sketch a formulation of the flow problem in
networks of fractures (Sect. 2) and we shortly recall the mathematical foundations
of the method and its discrete formulation in Sect. 3. New and extensive numerical
results are presented in Sect. 4 on complex DFN configurations that can be
considered realistic from the point of view of the geometrical complexities. Both
the problem of the computation of the steady-state distribution of the hydraulic
head and the problem of the transient response in advection-diffusion phenomena
are addressed.

2 Problem Formulation

In this section we briefly recall the problem formulation on the network. We refer
the reader to [11, 13] for more details.
Each fracture in the network is modelled as a planar polygon, with arbitrary
dimensions and orientation in space. Let Fi denote the i-th fracture, with i 2 I .
The fractures may intersect each other along a line called trace. Let Sm denote the
PDE-Constrained Optimization Approach for Flow Simulations in Fractured Media 29

m-th trace, with m 2 M . In the sequel we assume that each trace is generated by
exactly two fractures.
In order to relate the trace index m to the fracture indices i and j generating Sm ,
we set Im D .i; j/, being the couple .i; j/ ordered following the convention that i < j.
The DFN ˝ is therefore defined as the set

˝ D [i2I Fi :

Each fracture boundary @Fi is (possibly) divided in a Dirichlet part iD and in a

Neumann part i N . We correspondingly introduce the following sets of fracture
boundaries:

@˝ D [i2I @Fi ; D D [i2I iD ; N D [i2I iN :

Notice that we require D to be non-empty; on the other hand, some of the sets iD
are allowed to be empty. For further reference, we introduce the following sets: we
define

S D [m2M Sm

and 8i 2 I we denote by Si  S the subset of traces belonging to fracture Fi .

Let us introduce 8i 2 I the spaces
n o
Vi D H10 .Fi / D v 2 H1 .Fi / W vj D D 0
i

with dual space Vi0 , and, if iD ¤ ;, let us introduce

n o
ViD D H1D .Fi / D v 2 H1 .Fi / W vj D D HiD ;
i

where HiD is the Dirichlet condition imposed on iD . In the sequel, we adopt the
convention that for functions globally defined on the whole network, a i subscript
denotes the restriction to Fi .
Let Ki .xi / be, for all i 2 I , a symmetric and uniformly positive definite tensor
describing the fracture transmissivity, where xi refers to the local space coordinate
system on Fi . The hydraulic head on each fracture is obtained as follows: find Hi 2
ViD such that 8v 2 Vi
Z Z Z X Z  @Hi 
Ki rHi rvd˝ D qi vd˝ C HiN vj N d C vj d ; (1)
Fi Fi iN i
S2S S
@Si S S
i

1
where HiN 2 H 2 . i N / is the Neumann boundary condition imposed on iN .
The symbol @H
@ i
i
D .niS /T Ki rHi denotes the outward co-normal derivative of the
S
30 S. Pieraccini and S. Scialò

hh ii
@Hi
hydraulic head along the (fixed) unit vector niS normal to trace S, and @Si
denotes
S
2
its jump along Furthermore, qi 2 L .Fi / is a source term on Fi . Note that the last
niS .
term in (1) represents the net flow entering/exiting the fracture through its traces.
In order to ensure continuity of H and flux balance along all the network, the
following matching conditions are imposed at all traces: 8m 2 M , with i; j 2 Im ,

HijSm  Hj j Sm D 0; (2)
"" ## "" ##
@Hi @Hj
C D 0: (3)
@Si m j
@Sm S
Sm m

3 A PDE-Constrained Optimization Approach

We describe here an approach to the solution of problem (1)–(3) first conceived in

[11, 12] and further developed in [8, 13]. In particular, we follow the approach of
the latter reference. In order to ease notation, in the following we assume that the
traces are disjoint; this is a simplifying assumption which can be easily removed,
see [11, Remark Q 2.1]. 1 Q
Let Ui WD S2Si H 2 .S/, U WD i2I Ui . For each m 2 M , with Im D .i; j/,
we define the quantities
"" ## "" ##
@Hi @Hj
Uim WD C˛Hi jSm ; Ujm WD C˛Hj j Sm ; (4)
@Si m j
@Sm
Sm Sm

where ˛ > 0 is a fixed parameter, which will play the role of control variables. Next,
we build the following tuples of control variables:

Ui D ˘
S2Si
UiS 2 Ui ; UD ˘ Ui 2 U
i2I
:

The exact fulfilment of Eqs. (2) and (3) is replaced by the minimization of the
functional
X

J.H; U/ D
Hij  Hjj
2 1 C (5)
Sm Sm 2 H .Sm /
m2M



Sm S  

2

Ui C Uj m  ˛ HijSm C HjjSm
 1 ;
H 2 .Sm /
PDE-Constrained Optimization Approach for Flow Simulations in Fractured Media 31

with i; j such that Im D .i; j/, constrained by Eq. (1), which—taking into
account (4)—rewrites
Z X Z Z
Ki rHi rvd˝ C ˛ HijSm vjSm d D qi vd˝ C (6)
Fi m2M Sm Fi
Z X Z
HiN vj iN d C Uim vj Sm d
iN m2M Sm

8v 2 Vi , 8i 2 I . The problem

min J.H; U/ (7)

subject to (6)

has a unique minimizer which corresponds to the solution of (1)–(3), see [11, 13].
By using a finite element based discretization and following a “first dis-
cretize, then optimize” approach, a finite dimensional counterpart of problem (7)
is obtained. To this aim, we introduce on each fracture a triangular mesh, inde-
pendently of the position of the other fractures and of fracture intersections; we
also introduce a space discretization on the traces. Since the hydraulic head may be
nonsmooth along the traces, a special care is needed for an accurate reproduction
of the solution near the traces. This has been accomplished, in conjunction with
this approach, following several strategies. In [11–13] the Extended Finite Element
method (XFEM) [7, 25] has been used; in [8] the original triangular mesh has been
transformed in a polygonal mesh, using the new Virtual Element Method [2–6]. The
description which follows does not rely on a specific space discretization.
We introduce:
• on each fracture Fi , i 2 I , a (possibly non-standard) finite element basis f'i;k g,
k D 1; : : : ; Ni ;
• on each trace Sm , m 2 M , the set f i;k m
g, k D 1; : : : ; Nim of 1D linear finite
element basis functions.
In the sequel we will use small case letters to denote the discrete variables: in
particular, hi will denote the discrete hydraulic head on Fi and um
i will denote the
discrete control variable on the trace Sm of fracture Fi :
m
X
Ni X
Ni
hi D hi;k 'i;k ; um
i D um
i;k i;k :
m
(8)
kD1 kD1

With a notation overload, hi and um

i will also denote the corresponding vectors of
DOFs, the difference being always clear from the context.
In view of writing the discrete counterpart of problem (7), we introduce for
each fixed i 2 I the vector ui 2 RNSi obtained by concatenating
P (column-wise)
i corresponding to traces Sm 2 Si , thus NSi D
vectors um m
Sm 2Si Ni . We further
32 S. Pieraccini and S. Scialò

F
P column-wise vectors hi , with i 2 I , obtaining
concatenate
NT
a vector
P h 2 RN ,
N D i2I Ni , and vectors ui obtaining a vector u 2 R , N D i2I NSi .
F T

The constraint Eqs. (6) on each fracture rewrite as

Ai hi D qQ i C Bi ui ; (9)

where Ai 2 RNi Ni is the stiffness matrix, matrix Bi 2 RNi NSi collects the integrals
of functions 'i;k , k D 1; : : : ; Ni , against functions f i;k
m
g, k D 1; : : : ; Nim , Sm 2 Si ,
and the vector qQ i 2 R accounts for the source term and the boundary conditions.
Ni

Now we turn our attention to the discrete counterpart of (5). While referring
the reader to [13] for all the details, we give here a brief sketch of the discrete
formulation.
In Eq. (5), we compute J by means of L2 .Sm / norms and plug into (5) defini-
tion (8). The algebraic form obtained for (7) is as follows:

1 T h 
min J.h; u/ WD h G h  ˛hT Bu  ˛uT BT h C uT Gu u (10)
2
s.t. Ah D B u C q; (11)

where A 2 RN N is a block diagonal matrix built from the local stiffness matrices
F F

Ai ; Gh 2 RN N and Gu 2 RN N are sparse matrices defined by suitably

F F T T

assembling the local blocks defined on the fractures, and collect integrals of the
basis functions for h and u, respectively, on the traces, i.e. integrals of the kind:
Z Z
'i;k jSm 'j;` jSm ; m
i;k j;` I
m
(12)
Sm Sm

for either i D j or .i; j/ D Im ; B 2 RN N , B 2 RN N are again sparse matrices

F T F T

defined by suitably assembling the local blocks defined on the fractures, and collect
integrals of functions 'i;k against functions j;k
m
.
Notice that despite Eqs. (10)–(11) are defined on the whole DFN, the computa-
tions can be actually decoupled in local fracture-oriented computations, see [14].
Linearity of the constraints (11) easily allows to eliminate h from J, thus obtaining
the following equivalent unconstrained minimization problem:

O 1
min J.u/ WD uT .B T AT Gh A1 B CGu˛ B T AT B˛BT A1 B /u
2
CqT AT .Gh A1 B ˛B/u
1 T O
WD u Gu C qO T u: (13)
2
PDE-Constrained Optimization Approach for Flow Simulations in Fractured Media 33

3.1 Discretization Strategies

The proposed optimization approach does not require any kind of conformity along
the traces for the meshes on the intersecting fractures, as only integrals of the kind
of (12) need to be computed, that involve the discretization on different fractures.
Typically, the hydraulic head displays a non-smooth behaviour on the fractures, as
jumps in the first order derivatives may occur across the traces, corresponding to the
flux entering or leaving the fracture through its traces. This non-smooth behaviour
can not be correctly reproduced with standard finite elements on nonconforming
meshes, and for this reason unconventional discretization techniques might be
preferred. A possible strategy consists in using the flexibility of the VEM in
handling polygonal meshes to generate a mesh on each fracture that is conforming
to the traces on the fracture but is still nonconforming to the mesh on the intersecting
fractures. The resulting meshing process is independent on each fracture and is
performed by simply splitting the triangular elements of the nonconforming mesh
crossed by the traces in sub-polygons not crossed by the traces. This idea is at
the basis of the approaches described in [8–10]. Another strategy, adopted for the
simulations in the present work, consists in the use of the XFEM, thus adding to the
discrete space suitable basis functions reproducing the irregular behaviour of the
solution. The use of the XFEM in the context of DFN simulations in conjunction
with the optimization approach was proposed since the earliest works on the
subject [11–13], and is also adopted in different frameworks, as for example in
[26, 27]. Some numerical issues need to be addressed for efficient use of the XFEM
in large scale DFN simulations with the proposed approach, mainly concerning
prevention of ill conditioning and the imposition of boundary conditions. Although
these topics are extensively discussed in the literature (see e.g. [24, 33]), targeted
solutions are mandatory in the present context. The use of XFEM might be source
of ill conditioning due to the presence of redundant, linearly (or nearly linearly)
dependent basis functions in the discrete space, as a consequence of the additional
basis functions introduced to catch the behaviour of the solution near the traces.
When parallel or nearly parallel traces intersect the same mesh element, some of
the resulting basis functions might be redundant. Due to the complexity of the
geometry, a preliminary detection of redundant basis functions is not a viable option,
and other strategies need to be performed, such as a fracture-local rank revealing
QR factorization [16]. Also the imposition of Dirichlet boundary conditions can be
performed weakly through the functional, following the approach in [16].

4 Numerical Results

In this Section we report some numerical results aimed at showing the viability,
the performances, and the robustness of the proposed strategy for the simulation of
the flow in complex DFNs. A first set of simulations deals with the computation
34 S. Pieraccini and S. Scialò

of the hydraulic head distribution in realistic networks of varying complexity and

number of fractures, followed by a second set of numerical results reporting some
preliminary results on the resolution of diffusion-advection problems in DFNs with
the optimization-based approach.

4.1 Flow in Complex DFNs

In this Section we deal with the steady-state solution on DFNs with realistic
geometrical configurations, exhibiting challenging critical features such as very
narrow angles between intersecting traces, or very close and almost parallel traces
lying in the same fracture. These characteristics represent the major causes of
failure for simulations based on standard approaches relying on finite elements on
conforming grids, whereas they are easily tackled by the optimization based method
considered here. In order to prove the robustness of the method in dealing with
such hard-to-mesh features, we will compare the performances of the method on
networks with severe geometrical configurations with the results on less complex
networks.
A pool of four networks is considered here, with increasing number of fractures.
In Table 1 we report information concerning the network sizes. Namely, in the
table I # and M # represent the number of fractures and traces in the network,
respectively, and L is the cumulative length of all the traces. The two intermediate
size networks (709fract and 909fract) are stochastically generated, whereas the
609fract and 959fract networks are obtained starting from stochastically generated
networks, removing some fractures in order to eliminate the mentioned hard-to-
mesh geometrical features, thus obtaining simpler geometrical configurations.
The result of this process is shown in Figs. 1 and 2, in which we report
the distribution of the hard-to-mesh features for the considered networks: the
distribution of angles formed by couples of intersecting traces in the same fracture is
shown in Fig. 1, highlighting that the two intermediate size networks 709fract and
909fract have angles approaching zero (the smallest value is less than 0:3ı for both
DFNs), while the 609fract and the 959fract DFNs do not display angles smaller
than 10ı . In Fig. 2 the distribution of distances between couples of non intersecting
traces in the same fracture is reported, underlining again that the 709fract and
909fract DFNs have traces very close to each other, which are not present in the
other two networks.

Table 1 DFN labels and data DFN label I# M# L

concerning network size
609fract 609 1690 2:5922e C 4
709fract 709 3939 6:6206e C 4
909fract 909 7084 1:1798e C 5
959fract 959 4911 7:5822e C 4
PDE-Constrained Optimization Approach for Flow Simulations in Fractured Media 35

10 4 10 4
609 909
709 959

10 3 10 3
# Occurrences

# Occurrences
10 2 10 2

10 1 10 1

10 0 10 0

0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90
Angle[DEG] Angle[DEG]

Fig. 1 Distribution of the angle formed by couples of intersecting traces in a fracture: Left:
609fract and 709fract DFNs; Right: 909fract and 959fract DFNs

105 10
5

609 909
709 959

104 10
4
# Occurrences

# Occurrences

103 10
3

102 10
2

101 10
1

100 10
0

-4 -3 -2 -1 0 1 2 -4 -3 -2 -1 0 1 2
log 10 (Length) log 10 (Length)

Fig. 2 Distribution of the distance between couples of non intersecting traces in a fracture: Left:
609fract and 709fract DFNs; Right: 909fract and 959fract DFNs

A non uniform fracture transmissivity field is prescribed, and the fracture

transmissivity on Fi is given by:
8 4

ˆ
<104 1 C ci;1 e 1C16ci;2 ..Qx1/2 C.Qy1/2 / if i is odd
Ki D  (14)
:̂104 1 C ci;1 e2ci;2 ..Qx1/2 C.Qy1/2 / if i is even

where constants ci;1 2 f102 ; 101 ; 1g and ci;2 2 f0:1; 0:5; 1; 2g are randomly
2x 2y
chosen on each fracture and xQ D and yQ D , being `x , `y characteristic
`x `y
dimensions of the fracture in the x and y-direction, respectively, in the local
reference system.
On all the considered networks, a source and a sink fracture are placed at the top
and the bottom of the domain of interest, respectively, carrying Dirichlet boundary
36 S. Pieraccini and S. Scialò

Fig. 3 DFN geometry and fracture transmissivity distribution (log scale): left to right, top to
bottom: 609fract, 709fract, 909fract, 959fract

conditions, prescribing a value of H D D 1000 on one edge of the source fracture

and H D D 0 on one edge of the sink fracture. All other fracture edges are insulated.
The DFN geometries, along with the distribution of the fracture transmissivities, are
shown in Fig. 3 for the considered networks.
Simulations are performed on three different grids, characterized by the mesh
parameter ı 2 f7; 2; 0:5g, representing the maximum element area of the mesh.
The total number of degrees of freedom (DOFs) for each DFN and for each grid
are reported in Table 2. The saddle point problem is solved through the parallel
iterative solver described in [14], implementing the preconditioned conjugate
gradient method. The preconditioner P is a three-diagonal symmetric positive
definite matrix representing a very rough approximation of the matrix GO in (13),
and has a block diagonal structure, with blocks related to the fractures in the DFN,
[16]. The performances of the preconditioner are shown for the 709fract network
in Table 3, which reports the number of iterations required by the iterative solver
to reach a relative residual (namely, relative to the norm of the right hand side of
the linear system) smaller than a prescribed value , both in the preconditioned
(prec) and in the not preconditioned (not prec) case. It can be seen that the
PDE-Constrained Optimization Approach for Flow Simulations in Fractured Media 37

Table 2 Number of degrees DFN ıD7 ıD2 ı D 0:5

of freedom for each DFN and
mesh parameter 609fract 230,723 552;615 1,681,341
709fract 442,073 941;475 2,496,875
909fract 733,486 1;498;025 3,751,857
959fract 541,888 1;175;673 3,215,716

Table 3 Number of iterations vs 

 D 106  D 107  D 108
ı Not prec Prec Not prec Prec Not prec Prec
7 34,774 14,692 241;974 46,171 388,523 75,480
2 26,204 14,913 106;250 32,097 347,048 81,459
0.5 23,149 14,896 87;542 37,983 324,108 95,563
Comparison of the preconditioned and not preconditioned case on the 709fract DFN

Fig. 4 Detail of the computational mesh with ı D 2 for the 709fract DFN

preconditioner provides a reduction in the number of iterations with respect to the

not preconditioned case up to a factor of about 5 for the coarsest mesh and the
smallest relative residual. It is to remark that, thanks to its block-diagonal structure,
the preconditioner P can be applied inexpensively also in the parallel version of the
algorithm.
A detail of the computational mesh with ı D 2 is shown in Fig. 4 for the 709fract
DFN: it can be noticed that mesh element edges are arbitrarily placed with respect
to the intersections between fractures, as a result of the meshing process that is
performed on each fracture independently of the other fractures and of the traces.
The solution obtained for the 709fract and 959fract networks is reported, as an
example, in Figs. 5 and 6, respectively, on the intermediate size mesh (ı D 2). In
38 S. Pieraccini and S. Scialò

Fig. 5 Solution of the 709fract DFN on the mesh with ı D 2 with iso-h lines

Fig. 6 Solution of the 959fract DFN on the mesh with ı D 2 with iso-h lines

these figures the colouring is proportional to the hydraulic head, and iso-h lines are
also shown in order to highlight gradients in the solution.
In order to assess the robustness of the method in tackling critical geometrical
configurations, typical of realistic DFNs, we compare the performances of the
PDE-Constrained Optimization Approach for Flow Simulations in Fractured Media 39

method in solving the four considered networks, recalling that DFNs 709fract and
909fract have hard-to-mesh features that are not present in the other two networks.
Performances will be evaluated in terms of the number of iterations required by
the iterative solver to reach a given relative residual and in terms of three error
indicators used to determine the quality of the obtained solution. The error indicators
are
head , measuring the mismatch in hydraulic head continuity across the traces,

flux , measuring the flux unbalance across the traces, and

cons , representing the
error in the conservation of the flux in the whole network. They are defined as:
sX

hij  hjj
2
Sm Sm
m2M

head D ;
hmax L
sX

m  

u C um  ˛m hij C hjj
2
i j Sm Sm
m2M

flux D ;
N
˚L
and
ˇP P R ˇ
ˇ ˇ
ˇ i2Iin [Iout Sm 2Si Sm .um
i  ˛m hijSm /ˇ

cons D ;
˚N
where L is the cumulative trace length (see Table 1), hmax is the maximum of the
hydraulic head in the network (that for the considered cases coincides with the
Dirichlet condition on the source fracture) and ˚N is the total flux flowing through
the network. According to the imposed boundary conditions, ˚N can be computed as:
0 1
1 X X Z X X Z
˚N D @ .um
i  ˛m hijSm / 
A
i  ˛m hijSm / ;
.um
2 Sm Sm
i2Iin Sm 2Si i2Iout Sm 2Si

being Iin and Iout the index sets corresponding to source and sink fractures,
respectively. Following from the above definitions, the error indicators are relative
quantities:
head expresses a measure of the error per unitary value of the source
hydraulic head and per unitary trace length;
flux a measure of the error per unitary
total flux and per unitary trace length;
cons a measure of the error per unitary total
flux, this last quantity being independent of the number of traces in the network.
These definitions as relative quantities allow the use of these error indicators to
compare the quality of the solution obtained on DFNs with different geometries
and hydraulic properties. Table 4 reports, for all the considered networks and
grids, the number of iterations required by the preconditioned conjugate gradient
method to reach values of the relative residual lower than three prescribed values
( D 106 ; 107 ; 108 ), whereas values for the three error indicators described
40 S. Pieraccini and S. Scialò

Table 4 Number of iterations vs  for all the considered networks and meshsizes
609fract 909fract
ı  D 106  D 107  D 108 ı  D 106  D 107  D 108
7 14,871 34,623 50,126 7.0 11,075 31,241 120,228
2 13,993 21,601 60,561 2.0 11,599 26,854 169,311
0.5 16,173 25,135 69,331 0.5 11,741 29,241 70,143
709fract 959fract
ı  D 106  D 107  D 108 ı  D 106  D 107  D 108
7 14,692 46,171 75,480 7.0 24,043 42,980 64,869
2 14,913 32,097 81,459 2.0 26,483 47,671 66,662
0.5 14,896 37,983 95,563 0.5 26,408 48,480 77,574

Table 5 Error indicators for the various networks, meshes and relative residuals
609fract 909fract
ı 
head
flux
cons
head
flux
cons
7 106 5.6141e6 3.3316e8 6.4713e4 6.5259e7 1.0583e8 1.0849e3
107 5.6141e6 3.3316e8 6.4457e4 6.5259e7 1.0583e8 1.1002e3
108 5.6141e6 3.3316e8 6.4472e4 6.5259e7 1.0583e8 1.1003e3
106 3.0565e6 1.7465e8 2.9737e5 4.6689e7 7.6415e9 8.9169e4
2 107 3.0564e6 1.7466e8 2.8968e5 4.6689e7 7.6413e9 9.0033e4
108 3.0564e6 1.7466e8 2.9029e5 4.6689e7 7.6414e9 9.0005e4
106 1.9031e6 1.0716e8 2.6639e5 2.6142e7 3.5418e9 1.8584e4
0.5 107 1.9030e6 1.0716e8 2.6829e5 2.6140e7 3.5420e9 1.8786e4
108 1.9030e6 1.0716e8 2.6715e5 2.6140e7 3.5420e9 1.8756e4
709fract 959fract
ı 
head
flux
cons
head
flux
cons
7 106 1.3952e6 7.0877e9 5.7534e3 1.7037e6 2.1318e9 2.5212e3
107 1.3950e6 7.0862e9 5.7521e3 1.7036e6 2.1296e9 2.5354e3
108 1.3950e6 7.0862e9 5.7518e3 1.7036e6 2.1297e9 2.5351e3
106 1.0801e6 3.5207e9 3.4049e4 1.1070e6 1.3376e9 9.6060e4
2 107 1.0800e6 3.5204e9 3.4102e4 1.1067e6 1.3366e9 9.6625e4
108 1.0800e6 3.5204e9 3.4093e4 1.1067e6 1.3365e9 9.6625e4
106 6.6866e7 3.3422e9 7.1786e4 6.4611e7 8.4986e10 3.5035e4
0.5 107 6.6860e7 3.3404e9 7.1523e4 6.4543e7 8.4875e10 3.5232e4
108 6.6860e7 3.3403e9 7.1534e4 6.4542e7 8.4871e10 3.5241e4

above are reported in Table 5, for all the DFNs and grids and in correspondence
of the same values for . Finally, Table 6 shows the value of ˚N computed for
all the network and meshes with the lowest value of the residual and used in the
definition of the error indicators. Looking at Table 4 it can be seen that, for all
considered meshsizes, the number of iterations required by the method for reaching
a prescribed relative residual appears to be, in general, not significantly affected
PDE-Constrained Optimization Approach for Flow Simulations in Fractured Media 41

Table 6 Value of ˚N for the ı 609fract 709fract 909fract 959fract

considered networks vs
7 1:1182e C 1 1:2672e C 1 3.9928 4:5195e C 1
meshsize
2 1:1263e C 1 1:2672e C 1 3.9869 4:4985e C 1
0:5 1:1286e C 1 1:2700e C 1 3.9865 4:4896e C 1

by the presence of critical geometrical features in the network: see for example the
case  D 106 , in which the lowest number of iterations is attained on the “difficult”
909fract network. An exception is provided by the case  D 108 , in which the
909fract network requires an higher number of iterations. Note however that this is
only true on the coarser meshes, and that the value  D 108 is quite demanding,
often corresponding to an over-solving (see later).
The number of iterations appears to be quite independent also of the number of
traces and fractures in DFNs, and independent of the meshsize (and thus of the size
of the linear system), thanks to the use of preconditioning.
Looking at Table 5 it is possible to notice that the values of the error indicators
are very similar in all the considered cases, independently of the geometrical
complexities of the networks. Furthermore, it can be noted that pushing the
resolution of the linear system to values of the residual lower than 106 does not
provide substantial improvements in the quality of the solution, denoting that the
solution rapidly converges to the best solution for a fixed meshsize. A reduction in
meshsize, instead, provides, as expected, an improvement in the solution in all the
considered cases. Looking at Table 6 we can see that, on the contrary, the value of
˚N is quite independent of the mesh.

4.2 Unsteady Flow in DFNs

We end this section with some preliminary data on the resolution of the unsteady
advection-diffusion problem in DFNs with the proposed approach. The optimization
method is used to compute the Darcy velocity in a given DFN, through the
computation of the gradient of the evaluated hydraulic head distribution. At each
time-step of the backward Euler method, the optimization method is then used to
compute the concentration of a passive scalar quantity C (a pollutant) in the network.
Using the same notation for the hydraulic head, at a fixed time frame, let us assume
Ci 2 H1D .Fi /, 8i 2 I , with some prescribed Dirichlet boundary conditions on
D ; furthermore, let us assume, for simplicity of notation, homogeneous Neumann
boundary conditions on N and no source terms. Then we have:

min J.C; U/ WD
X

 

2


Cij  Cjj
2 1 C
UiSm C UjSm  ˛ CijSm C CjjSm
 1
Sm Sm 2 H .Sm / H 2 .Sm /
m2M
42 S. Pieraccini and S. Scialò

such that:
Z Z X Z Z
1
Ci vd˝ C i rCi rvd˝ C ˛ CijSm vjSm d C ˇi rCi vd˝ D

t Fi Fi m2M Sm Fi
Z X Z
1
Ci0 vd˝ C Uim vj Sm d; 8v 2 Vi ; 8i 2 I ; (15)

t Fi m2M Sm

where now functions Uim act as control variables of the minimization problem
for C, i is the diffusivity of the pollutant in Fi , ˇi is the velocity field on Fi ,

t is the time-step, and Ci0 is a term accounting for the solution at the previous
time step. Analogously to what described in Sect. 3, the cost functional J.C; U/
expresses the mismatch in the continuity and the unbalance of the diffusive flux
of the pollutant through the traces of the DFN and it is minimized constrained by
the SUPG-stabilized [23] version of (15) on the fractures, expressing the advection-
diffusion law for the passive scalar Ci . Results are shown for a 59 fracture DFN,
with two different fields of fracture transmissivity: in a first batch of simulations
the fracture transmissivity is set to one for all fractures, whereas, in a second set
of simulations, the fracture transmissivity is a piecewise constant function with a
different value on each fracture. The geometry of the considered network is shown
in Fig. 7, left, where the computational mesh and the hydraulic head distribution
for the uniform transmissivity field case is also shown, whereas in Fig. 7, right, the
transmissivity field in the non homogeneous case is plotted. Figure 8 reports a zoom
of the inflow and outflow fractures for the homogeneous transmissivity field, also
showing vectors of the computed Darcy velocity. The diffusion of the passive scalar

Fig. 7 59 fracture DFN. Left: computational mesh and hydraulic head distribution (uniform
transmissivity field). Right: representation of the non uniform, fracture-constant transmissivity field
PDE-Constrained Optimization Approach for Flow Simulations in Fractured Media 43

Fig. 8 Velocity field in inflow (left) and outflow (right) fractures for the 59 fracture DFN. Uniform
transmissivity field

Fig. 9 Pollutant concentration after 102 and 103 time units for the uniform transmissivity field ((a)
and (b)) and after 103 time units for the non-uniform transmissivity field (c)

after 102 and 103 time units for the uniform transmissivity field is shown in Fig. 9a
and b: the passive scalar flows from the inflow to the outflow boundary, with an
advection-dominated transport, with the diffusivity coefficient for the pollutant set
to 105 . The solution at 103 time units for the non uniform transmissivity field is
also reported for comparison in Fig. 9c, highlighting the different path followed by
the pollutant in the two cases.

Acknowledgements This work has been supported by the Italian MIUR through PRIN research
grant 2012HBLYE4_001 “Metodologie innovative nella modellistica differenziale numerica” and
by INdAM-GNCS.
44 S. Pieraccini and S. Scialò

References

1. Adler, P.M.: Fractures and Fracture Networks. Kluwer Academic, Dordrecht (1999)
2. Ahmad, B., Alsaedi, A., Brezzi, F., Marini, L.D., Russo, A.: Equivalent projectors for virtual
element methods. Comput. Math. Appl. 66(3), 376–391 (2013). doi:http://dx.doi.org/10.1016/
j.camwa.2013.05.015
3. Beirão da Veiga, L., Brezzi, F., Cangiani, A., Manzini, G., Marini, L.D., Russo, A.: Basic
principles of virtual element methods. Math. Models Methods Appl. Sci. 23(1), 199–214
(2013). doi:http://dx.doi.org/10.1142/S0218202512500492
4. Beirão da Veiga, L., Brezzi, F., Marini, L.D.: Virtual elements for linear elasticity problems.
SIAM J. Numer. Anal. 51(2), 794–812 (2013). doi:http://dx.doi.org/10.1137/120874746
5. Beirão da Veiga, L., Brezzi, F., Marini, L.D., Russo, A.: Virtual element method for general
second-order elliptic problems on polygonal meshes. Math. Models Methods Appl. Sci. 26(4),
729–750 (2016). doi:https://doi.org/10.1142/S0218202516500160
6. Beirão da Veiga, L., Lipnikov, K., Manzini, G.: The Mimetic Finite Difference Method for
Elliptic Problems. Modeling, Simulation and Applications, vol. 11. Springer, Cham (2014)
7. Belytschko, T., Black, T.: Elastic crack growth in finite elements with minimal remeshing. Int.
J. Numer. Methods Eng. 45, 601–620 (1999)
8. Benedetto, M.F., Berrone, S., Pieraccini, S., Scialò, S.: The virtual element method for discrete
fracture network simulations. Comput. Methods Appl. Mech. Eng. 280(0), 135–156 (2014).
doi:http://dx.doi.org/10.1016/j.cma.2014.07.016
9. Benedetto, M.F., Berrone, S., Borio, A.: The virtual element method for underground flow
simulations in fractured media. In: Ventura, G., Benvenuti, E. (eds.) Advances in Discretization
Methods. SEMA SIMAI Springer Series, vol. 12. Springer International Publishing, Switzer-
land (2016)
10. Benedetto, M.F., Berrone, S., Scialò, S.: A globally conforming method for solving flow in
discrete fracture networks using the virtual element method. Finite Elem. Anal. Des. 109,
23–36 (2016). doi:http://dx.doi.org/10.1016/j.finel.2015.10.003
11. Berrone, S., Pieraccini, S., Scialò, S.: A PDE-constrained optimization formulation for discrete
fracture network flows. SIAM J. Sci. Comput. 35(2), B487–B510 (2013). doi:http://dx.doi.org/
10.1137/120865884
12. Berrone, S., Pieraccini, S., Scialò, S.: On simulations of discrete fracture network flows with an
optimization-based extended finite element method. SIAM J. Sci. Comput. 35(2), A908–A935
(2013). doi:http://dx.doi.org/10.1137/120882883
13. Berrone, S., Pieraccini, S., Scialò, S.: An optimization approach for large scale simulations of
discrete fracture network flows. J. Comput. Phys. 256, 838–853 (2014). doi:http://dx.doi.org/
10.1016/j.jcp.2013.09.028
14. Berrone, S., Pieraccini, S., Scialò, S., Vicini, F.: A parallel solver for large scale DFN flow
simulations. SIAM J. Sci. Comput. 37(3), C285–C306 (2015). doi:http://dx.doi.org/10.1137/
140984014
15. Berrone, S., Borio, A., Scialò, S.: A posteriori error estimate for a PDE-constrained optimiza-
tion formulation for the flow in DFNs. SIAM J. Numer. Anal. 54(1), 242–261 (2016)
16. Berrone, S., Pieraccini, S., Scialò, S.: Towards effective flow simulations in realistic discrete
fracture networks. J. Comput. Phys. 310, 181–201 (2016). doi:http://dx.doi.org/10.1016/j.jcp.
2016.01.009
17. Cammarata, G., Fidelibus, C., Cravero, M., Barla, G.: The hydro-mechanically coupled
response of rock fractures. Rock Mech. Rock Eng. 40(1), 41–61 (2007). doi:http://dx.doi.
org/10.1007/s00603-006-0081-z
18. Dershowitz, W.S., Fidelibus, C.: Derivation of equivalent pipe networks analogues for three-
dimensional discrete fracture networks by the boundary element method. Water Resour. Res.
35, 2685–2691 (1999). doi:http://dx.doi.org/10.1029/1999WR900118
19. de Dreuzy, J.R., Pichot, G., Poirriez, B., Erhel, J.: Synthetic benchmark for modeling flow in
3D fractured media. Comput. Geosci. 50(0), 59–71 (2013)
PDE-Constrained Optimization Approach for Flow Simulations in Fractured Media 45

20. Fidelibus, C., Cammarata, G., Cravero, M.: Hydraulic characterization of fractured rocks. In:
Abbie, M., Bedford, J.S. (eds.) Rock Mechanics: New Research. Nova Science Publishers Inc.,
New York (2009)
21. Formaggia, L., Fumagalli, A., Scotti, A., Ruffo, P.: A reduced model for Darcy’s problem in
networks of fractures. ESAIM: Math. Model. Numer. Anal. 48, 1089–1116 (2014). doi:http://
dx.doi.org/0.1051/m2an/2013132
22. Formaggia, L., Antonietti, P., Panfili, P., Scotti, A., Turconi, L., Verani, M., Cominelli, A.:
Optimal techniques to simulate flow in fractured reservoir. In: ECMOR XIV-14th European
Conference on the Mathematics of Oil Recovery (2014)
23. Franca, L.P., Frey, S.L., Hughes, T.J.R.: Stabilized finite element methods: I. Application to
the advective-diffusive model. Comput. Methods Appl. Mech. Eng. 95, 253–276 (1992)
24. Fries, T.P.: A corrected XFEM approximation without problems in blending elements. Int. J.
Numer. Methods Eng. 75, 503–532 (2008). doi:http://dx.doi.org/10.1002/nme.2259
25. Fries, T.P., Belytschko, T.: The extended/generalized finite element method: an overview of the
method and its applications. Int. J. Numer. Methods Eng. 84(3), 253–304 (2010). doi:http://
dx.doi.org/10.1002/nme.2914
26. Fumagalli, A., Scotti, A.: A numerical method for two-phase flow in fractured porous media
with non-matching grids. Adv. Water Resour. 62, 454–464 (2013). doi:http://dx.doi.org/10.
1016/j.advwatres.2013.04.001
27. Fumagalli, A., Scotti, A.: An efficient XFEM approximation of Darcy flows in arbitrarily
fractured porous media. Oil Gas Sci. Technol. - Rev. d’IFP Energies Nouv. 69(4), 555–564
(2014). doi:http://dx.doi.org/10.2516/ogst/2013192
28. Hyman, J., Gable, C., Painter, S., Makedonska, N.: Conforming Delaunay triangulation of
stochastically generated three dimensional discrete fracture networks: a feature rejection
algorithm for meshing strategy. SIAM J. Sci. Comput. 36, A1871–A1894 (2014). doi:http://
dx.doi.org/10.1137/130942541
29. Jaffré, J., Roberts, J.E.: Modeling flow in porous media with fractures; discrete fracture models
with matrix-fracture exchange. Numer. Anal. Appl. 5(2), 162–167 (2012)
30. Karimi-Fard, M., Durlofsky, L.J.: Unstructured adaptive mesh refinement for flow in heteroge-
neous porous media. In: ECMOR XIV-14th European Conference on the Mathematics of Oil
Recovery (2014)
31. Nœtinger, B.: A quasi steady state method for solving transient Darcy flow in complex 3D
fractured networks accounting for matrix to fracture flow. J. Comput. Phys. 283, 205–223
(2015). doi:http://dx.doi.org/10.1016/j.jcp.2014.11.038
32. Nœtinger, B., Jarrige, N.: A quasi steady state method for solving transient Darcy flow in
complex 3D fractured networks. J. Comput. Phys. 231(1), 23–38 (2012). doi:http://dx.doi.org/
10.1016/j.jcp.2011.08.015
33. Osher, S., Fedkiw, R.: Level Set Methods and Dynamic Implicit Surfaces. Applied Mathemat-
ical Sciences, vol. 153. Springer, New York (2003)
34. Pichot, G., Erhel, J., de Dreuzy, J.: A mixed hybrid mortar method for solving flow in discrete
fracture networks. Appl. Anal. 89, 1629–643 (2010). doi:http://dx.doi.org/10.1080/00036811.
2010.495333
35. Pichot, G., Erhel, J., de Dreuzy, J.: A generalized mixed hybrid mortar method for solving
flow in stochastic discrete fracture networks. SIAM J. Sci. Comput. 34, B86–B105 (2012).
doi:http://dx.doi.org/10.1137/100804383
36. Pichot, G., Poirriez, B., Erhel, J., de Dreuzy, J.R.: A Mortar BDD method for solving flow
in stochastic discrete fracture networks. In: Domain Decomposition Methods in Science
and Engineering XXI. Lecture Notes in Computational Science and Engineering, pp. 99–112.
Springer, Cham (2014)
37. Vohralík, M., Maryška, J., Severýn, O.: Mixed and nonconforming finite element methods on
a system of polygons. Appl. Numer. Math. 51, 176–193 (2007)
A Review of the XFEM-Based Approximation
of Flow in Fractured Porous Media

Bernd Flemisch, Alessio Fumagalli, and Anna Scotti

Abstract This paper presents a review of the available mathematical models and
corresponding non-conforming numerical approximations which describe single-
phase fluid flow in a fractured porous medium. One focus is on the geometrical
difficulties that may arise in realistic simulations such as intersecting and immersed
fractures. Another important aspect is the choice of the approximation spaces
for the discrete problem: in mixed formulations, both the Darcy velocity and the
pressure are considered as unknowns, while in classical primal formulations, a
richer space for the pressure is considered and the Darcy velocity is computed a
posteriori. In both cases, the extended finite element method is used, which allows
for a complete geometrical decoupling among the fractures and rock matrix grids.
The fracture geometries can thus be independent of the underlying grid thanks to
suitable enrichments of the spaces that are able to represent possible jumps of the
solution across the fractures. Finally, due to the dimensional reduction, a better
approximation of the resulting boundary conditions for the fractures is addressed.

1 Introduction

The simulation of subsurface flow is of great importance for a large number

of applications ranging from the production of energy (oil and gas reservoirs,
geothermal energy) to the management of water resources, or the safe storage of
atomic waste and carbon dioxide. Fractures are present in porous media at a variety
of scales. Large fractures and faults in particular are very relevant for the flow since
they can either act as barriers (in the case of impermeable faults) or preferential
pathways for the flow (in the case of permeable fractures). Opposed to small-scale

B. Flemisch
Department of Hydromechanics and Modelling of Hydrosystems, University of Stuttgart,
Pfaffenwaldring 61, 70569 Stuttgart, Germany
e-mail: [email protected]
A. Fumagalli • A. Scotti ()
MOX, Dipartimento di Matematica, Politecnico di Milano, via Bonardi 9, 20133 Milan, Italy
e-mail: [email protected]; [email protected]

© Springer International Publishing Switzerland 2016 47

G. Ventura, E. Benvenuti (eds.), Advances in Discretization Methods,
SEMA SIMAI Springer Series 12, DOI 10.1007/978-3-319-41246-7_3
48 B. Flemisch et al.

fractures, that can be accounted for by upscaling of the permeability, large features
should be explicitly included in the model to reproduce their non-local effects on
the flow. Thanks to the developments of numerical methods and computing power
direct numerical simulations of fracture networks are replacing or complementing
multi-continua approaches such as dual-porosity/permeability.
Fractures and faults are three-dimensional regions characterized by a different
porosity and permeability with respect to the surrounding porous matrix. However,
thanks to their small aperture compared to the typical length and the size of the
domain, they are usually represented as .N1/-dimensional interfaces immersed
in a N-dimensional matrix. From a computational viewpoint, this avoids the need
for an extremely fine grid to resolve the width of fractures, that are now replaced
by discontinuity surfaces where a suitable reduced .N1/-dimensional problem
is solved and coupled with the surrounding flow. However, the complexity of
geological structures remains one of the main challenges in large-scale numerical
simulations. Indeed, the data for the construction of the model are usually given
as a large number of possibly intersecting surfaces, called horizons, that separate
layers with different mechanic and hydraulic properties, and a set of surfaces that
represent faults and fractures. In this framework, the construction of a grid which
is conforming with all the aforementioned features is a difficult task, [20], whose
outcome could be a grid that is either too refined to be used, or with low quality
elements. Since in these realistic cases the construction of a high-quality grid that
honours the geometry of hundreds or thousands of fractures is a challenging task,
two alternative approaches are possible:
• to develop numerical methods that are accurate and robust even for very distorted
grids, such as the Mimetic Finite Difference Methods [8];
• to allow the fractures to cross a fairly regular and coarse grid in arbitrary ways,
and to employ the eXtended Finite Element Method (XFEM) to account for the
solution discontinuities within elements.
This paper presents the second approach, reviewing the recent literature on the
application of the XFEM to the simulation of flow in fractured porous media,
focusing on single-phase flow in the presence of one or more, possibly intersecting,
fractures.
The XFEM has been successfully used for the simulation of crack mechanics for
a long time, [22, 23, 44], while its application to flow in fractured media is a recent
development. Allowing for non-matching grids with respect to the fracture network
can be advantageous in geological problems since not only it avoids the burden of
computing a conforming grid, but it avoids the need for re-meshing in the case of
uncertain geometry, i.e. one could perform simulations of different scenarios with
different fracture configurations with the same background grid.
The enrichment of the finite element spaces should be able to effectively
represent discontinuities in the pressure and in the flux across fractures: pressure
jumps arise in the case of impermeable interfaces, while a discontinuous flux can be
observed due to the fact that fluid can enter the fractures and flow along them.
Review of XFEM for Flow in Fractured Porous Media 49

The development and the analysis of .N1/-dimensional models for fractures for
single-phase flow have been extensively addressed in [3, 7, 13, 29, 42], where the
fracture flow equations and the proper interface conditions across the fracture have
been first derived, and the continuous and discrete problems have been studied in
their mixed formulation. However, in the aforementioned works, the computational
grid of the porous domain is considered to be matching with the fracture, i.e. the
fracture is the (conforming) interface between two mesh blocks, possibly with
different resolution. Similarly, in more recent works, this type of space discretization
has been employed to describe the flow in faulted sedimentary basins, coupled with
a double-layer model for the fault, see [26, 52]. However, while non-conforming
meshes on the interface could be dealt with by mortaring, this does not allow the
fractures to cut the elements of the grid.
The use of XFEM to deal with fractures as non-matching, immersed interfaces,
was first introduced in 2011, for the single phase case, in [19, 41]. In the former,
the concept of EFEM (Enriched Finite Element Method) is applied to the primal
formulation with suitable enrichments for the pressure, while in the latter the
mixed formulation of the problem is considered, and the authors employ concepts
borrowed from [39] to enrich both the pressure and the Darcy velocity spaces.
In the same years, XFEM have been applied for the discretization of the primal
formulation in [50].
The application to the case of several intersecting fractures is discussed in
[11, 27] in the case of an impervious surrounding medium, with slightly different
coupling conditions at the intersections based on different assumptions on the
fractures permeability. The coupled problem, in the mixed and primal formulation
respectively, are considered in [30, 31, 51]. As concerns the physics of flow in porous
media, some works consider also the case of passive transport of solutes in fractured
porous media, [32], and two-phase flow [34] by means of the XFEM. Moreover, the
coupling of Darcy flow with fracture mechanics (opening, propagation) is addressed
in [37, 48].
Regarding the analysis of the method, the inf-sup stability of XFEM applied to
Darcy flow in porous media has been proven, under suitable conditions, in [19, 21].
This paper is organized as follows. In Sect. 2, we introduce the mathematical
model for single-phase flow in fractured porous media in the equi-dimensional case,
and derive the corresponding hybrid-dimensional, or reduced, model. Intersecting
fractures are considered, as well as the task of assigning boundary conditions for
the fractures. In Sect. 3, we present the numerical discretization techniques for the
problem in primal and dual mixed form, with a focus on the ad-hoc enrichments
at intersections and tips, and on the approximation of coupling terms. Section 4 is
dedicated to solvers for the resulting linear system, in particular to conditioning
issues and the choice of iterative vs. monolithic approaches. Finally, Sect. 5 is
devoted to some concluding remarks and future perspectives.
50 B. Flemisch et al.

2 Governing Equations

In this section, we present the mathematical model of single-phase flow in porous

media, focusing our attention on the description of the fractures. We start consider-
ing the standard Darcy law and mass balance in an equi-dimensional setting, where
the fractures are N-dimensional regions embedded in an N-dimensional porous
matrix. Then we introduce the so called reduced, or hybrid-dimensional, models to
handle fractures as objects of effective lower dimension, intersections and branching
of fractures, and proper boundary conditions to prescribe at the fracture tips and on
the cut matrix boundaries.

2.1 Equi-Dimensional Models

We consider an inert and at rest porous medium which can be modeled as a bounded,
connected, and open set D  RN , ND2 or 3. We assume that the medium is
saturated with a single incompressible fluid phase that is composed of a unique
component, e.g., water. The boundary, which is required to be regular enough for the
forthcoming assumptions, is indicated by @D with outward unit normal nD . Let us
assume that D contains several fractures, that all together constitute a single domain
of spatial dimension N such that  D, which is a possibly unconnected,
open subset of D. The fracture network can also be seen as the union of fracture
branches i . The surrounding porous rock, namely, the remaining part of D, is called
˝ :D D n . The outer boundary of the rock matrix is indicated by @˝ D ˝ \ @D,
while the outer boundary of the fracture network is indicated by @ and defined by
@ :D \ @D. Moreover, the internal part of the fracture boundary, namely, the
interface between the fracture domain and the surrounding medium, is indicated by
 and defined as  :D \ ˝. For each fracture branch, we call its own part of this
internal boundary i . The unit normal, pointing out of into ˝, is indicated by n.
We suppose that for each fracture branch there exists a central axis Oi , which is a
non self-intersecting .N1/-dimensional surface, such that a fracture branch can be
described as
 
di
i D x 2 R W x D s C rni ; s 2 Oi ; jrj <
N
; (1)
2

where di is the aperture of i , which may depend on the curvilinear abscissa s,

and ni is the unit normal associated with the central axis Oi . We assume that the
apertures di are small compared to other characteristic dimensions of the fractures.
With definition (1) the fracture domain is composed by D [ i , but note that i
may intersect each other with a non-null intersection. See Fig. 1 for an illustration
of the aforementioned notation.
Review of XFEM for Flow in Fractured Porous Media 51

Fig. 1 Notation for a general

configuration of fractures in
the equi-dimensional model Γ1 Γ2
γ̂4

γ
Ω
n Γ3
nD

In this work we assume that the fractures are filled by a porous medium
themselves such that Darcy flow takes place in both the rock matrix and fractures.
In the relevant case of open fractures the lubrication model could be used, see for
instance [53].

2.1.1 Dual Formulation

In this part, our objective is to compute the steady-state pressure field p and the
Darcy velocity field, or macroscopic velocity, u in the entire porous domain D. To
this purpose, following for example [9], we employ the law of mass conservation
together with Darcy’s law and, to ease the notation, we assume homogeneous
boundary conditions for the pressure on the whole boundary. The system of
equations for the porous matrix ˝ is given by
8
<r  u D f
ˆ
in ˝,
u C rp D 0 (2a)
:̂
pD0 on @˝,

where the scalar source term f represents a possible volume source or sink and 
denotes the symmetric and positive definite permeability tensor in ˝. To simplify
the notation we consider a permeability tensor that is already scaled with the
viscosity. Coupled with (2a), a similar system of equations can be considered for
the fracture network. The data and unknowns related to the fractures are indicated
with a subscript f. We obtain
8
< r  uf D ff
ˆ in ,
uf C f rpf D 0 (2b)
:̂
pf D 0 on @ .

Following [42] we require that the permeability tensor in the fracture system, for
each fracture branch, can be written as f;i D i;n Ni C i; T i , where the projection
52 B. Flemisch et al.

matrix Ni in the direction normal to Oi and the projection matrix T i in the direction
tangential to Oi are defined as follows:

Ni :D ni ˝ ni and T i :D I  Ni :

To couple the systems (2a) and (2b) we consider the following classical interface
conditions, namely
(
p D pf
on : (2c)
u  n D uf  n

Combining (2a)–(2c), we obtain the strong problem formulation in its dual form.
Problem 1 (Dual Equi-Dimensional Strong Formulation) Find velocity fields
u; uf and pressure fields p; pf such that (2) is fulfilled.
The proof of the well-posedness of Problem 1 in its mixed weak form can be found
in a number of references, such as [15, 24, 46, 49].

2.1.2 Primal Formulation

A common formulation for single-phase porous-media flow is the so-called primal

formulation, that can be obtained inserting Darcy’s law for matrix and fracture
domain, namely, the second lines of (2a), (2b) into the mass-balance equations,
namely, the first lines of (2a), (2b), as well as into the flux-coupling condition,
namely, the second line of (2c). In particular, for the matrix domain ˝, we obtain
(
r  .rp/ D f in ˝,
(3a)
pD0 on @˝,

while for the fracture domain , we have

(
r  .f rpf / D ff in ,
(3b)
pf D 0 on @ ,

coupled by
(
p D pf
on : (3c)
rp  n D f rpf  n

The problem can, in this case, be cast as follows.

Review of XFEM for Flow in Fractured Porous Media 53

(a) (b)

Fig. 2 Model domains with fracture. (a) Equi-dimensional model domain. (b) Hybrid-
dimensional model domain which includes a lower dimensional fracture

Problem 2 (Primal Equi-Dimensional Strong Formulation) Find pressure fields

p; pf such that (3) is fulfilled.
The well-posedness of Problem 2 and its weak form can be found in any standard
textbook on partial differential equations or finite elements (Fig. 2).

2.2 Hybrid-Dimensional Models

In this section, we present the hybrid-dimensional model, or reduced model, in the

case of single fracture dividing the domain in two unconnected parts. We refer to
[2, 6, 19, 25, 42] for a detailed presentation. The derivation of the model is based
on its dual formulation, however, we present also its primal formulation, obtained
with a “post-processing”. For both formulations, we briefly introduce their weak
formulation to be used for the numerical discretization.
During the process, we substitute the fracture by its centre line O and the
surrounding porous medium is enlarged to fill the gap. In practical cases, this step
is very seldom performed since the fracture geometry is directly given as an object
of codimension one. The Darcy equations (2a) for the rock matrix are thus the same
as in the equi-dimensional case and we focus only on the Eqs. (2b) for the fracture.
Since ˝ is split in two parts ˝1;2 , we define n D n1 D n2 , where ni is the outward-
pointing normal of ˝i . Let us introduce the normal and tangential divergence and
gradient on the fracture: given two regular functions a and a, respectively vector-
and scalar-valued, we define

r  a D rn  a C r  a with rn  a :D N W ra and r  a :D T W ra;

ra D rn a C r a with rn a :D Nra and r a :D Tra:
54 B. Flemisch et al.

2.2.1 Dual Formulation

We decompose the Darcy velocity in the fracture into its normal part uf;n :D Nuf
and tangential part uf; :D Tuf , such that uf D uf;n C uf; . We consider first the
conservation equation which is integrated along the normal direction of O for the
fracture aperture d. We obtain a conservation equation in the tangential space of
R d=2
O for the reduced flux uO :D d=2 uf; which involves also the contribution of the
incoming flux from the surrounding porous medium, namely,

r  uO D fO C u  nO in O ; (4)
R d=2
where the reduced source term is defined as fO :D d=2 ff . In (4) we made use of the
jump operator defined as u  nO :D u1  n  u2  n, with an abuse of notation for the
normal n. We consider now the Darcy equation projected on the tangential space of
O and integrated in normal direction for the aperture of the fracture, obtaining

O  pO D 0
uO C r in O ; (5)
R d=2
where pO is the reduced pressure in the fracture, defined as pO :D d1 d=2 pf . In the
previous equation, O is the effective permeability in tangential direction, defined as
O :D df; . Finally, projecting Darcy’s law on the normal space of the fracture and
integrating in normal direction on the first and on the second half of the aperture,
we end up with coupling conditions between the lower-dimensional fracture and the
rock matrix. Using a suitable approximation of the integral of ui  n, as discussed in
[7, 42], we get
(
u1  n C .1  /u2  n D 2O . p1  pO /
on O ; (6)
u2  n C .1  /u1  n D 2O . pO  p2 /

where O is the effective permeability in normal direction of the fracture, defined

as O :D f;n =d. Moreover,  2 .0:5; 1 is a closure parameter related to the
pressure cross profile in the fracture, see the aforementioned works for more details.
Considering (2a) coupled with (4)–(6) we obtain the following hybrid-dimensional
problem.
Problem 3 (Dual Hybrid-Dimensional Strong Formulation) Find .ui ; pi / for
i D 1; 2 and .u;
O pO / such that
8 8
O
< r  ui D fi
ˆ in ˝i , < r  uO D f C u  nO
ˆ in O ,
ui C i rpi D 0 and O  pO D 0
uO C r (7a)
:̂ :̂
pi D 0 on @˝i , pO D 0 on @O ,
Review of XFEM for Flow in Fractured Porous Media 55

with interface conditions

(
u1  n C .1  /u2  n D 2O . p1  pO /
on O : (7b)
u2  n C .1  /u1  n D 2O . pO  p2 /

An alternative form of the interface conditions (7b), introduced in [18], is

(
ffu  nggO D O pO
  on O ; (7b-bis)
0 u  nO D O ffpggO  pO

with 0 D 4=.2  1/ and where we have used the average operators ffpggO :D
1
2
. p1 C p2 / and ffu  nggO :D 12 .u1  n C u2  n/, as well as the jump operator for
the pressure pO :D p1  p2 .
We now introduce the weak formulation of the reduced problem, which will be
useful to present the XFEM in Sect. 3. For a detailed presentation of the suitable
functional spaces refer to [7]. We start by introducing the following bilinear forms
and functionals for the rock matrix
X  
ad .u; v/ :D .Hui ; vi /˝i C O ffu  nggO ; ffv  nggO O
i

C0 O u  nO ; v  nO ; with i D 1; 2;
O

with H :D 1 , O :D 1
O
the inverse of the permeabilities. The bilinear form and
the functional which include the source term and possibly boundary conditions read
X X
bd . p; v/ :D  . pi ; r  vi /˝i and F.q/ :D . fi ; qi /˝i :
i i

The weak formulation for the fracture requires to introduce the following bilinear
forms and functional

aO d .u;
O vO / :D .O u; O O ; bO d . pO ; v/
O v/ O :D  . pO ; r  v/ O q/ :D fO ; qO
O O and F.O
O

with O :D O 1 the inverse of the effective tangential permeability. Finally the

bilinear form which couples the fracture and the surrounding porous medium

cd .u; qO / :D u  nO ; qO :
O

The weak formulation of (7) is given as follows.

56 B. Flemisch et al.

Problem 4 (Dual Hybrid-Dimensional Weak Formulation) Find .ui ; pi / for i D

1; 2 and .u;
O pO / respecting the given boundary conditions such that
( (
ad .u; v/ C bd . p; v/ C cd . pO ; v/ D 0 aO d .u; O C bO d . pO ; v/
O v/ O D0
and ;
bd .q; u/ D F.q/ bO d .Oq; u/ O q/
O  cd .Oq; u/ D F.O

for all test functions v, q, vO and qO defined in their proper spaces.

2.2.2 Primal Formulation

As for the equi-dimensional setting, a primal formulation can be derived by inserting

Darcy’s laws into the mass balance equations in (7a).
Problem 5 (Primal Hybrid-Dimensional Strong Formulation) Find pi for i D
1; 2 and pO such that
( (
r  i rpi D fi in ˝i , O  pO D fO  rp  n
r  r in O ,
O
and
pi D 0 on @˝i , pO D 0 on @O ,
(8a)

with interface conditions (7b-bis) reformulated as

(
 ffrp  nggO D O pO
  on O : (8b)
0 rp  nO D O ffpggO  pO

Proceeding to the weak formulation of the primal problem, we define the matrix
bilinear form
X   
ap . p; q/ D .i rpi ; rqi /˝i C 0 O ffpggO ; ffqggO O C O pO ; qO
O
i

and the fracture bilinear form

  
O  pO ; r qO C 0 O pO ; qO :
aO p . pO ; qO / D r
O O

The coupling between matrix and fracture is accounted for by the bilinear form
 
cp . p; qO / D 0 O ffpggO ; qO O :

This allows to obtain the weak formulation of Problem 5.

Review of XFEM for Flow in Fractured Porous Media 57

Problem 6 (Primal Hybrid-Dimensional Weak Formulation) Find pi for i D

1; 2 and pO respecting the given boundary conditions such that
(
ap . p; q/  cp .q; pO / D F.q/
;
O q/
cp . p; qO / C aO p . pO ; qO / D F.O

for all test functions q and qO defined in their proper spaces.

The analysis of Problem 6 is straightforward and presented in, for example, [50].
However, the situation becomes more involved if the fracture is allowed to end
inside the interior of the matrix, yielding a non-Lipschitz domain ˝. A rigorous
mathematical analysis of this setup is carried out in [37].

2.3 Branching and Intersections

In this part, we present several strategies to model the intersection of fractures.

This is an important aspect, since the complex nature of networks of possibly
heterogeneous fractures requires an appropriate treatment to avoid un-physical
results. In a crossing, however, the different properties of every fracture branch can
overlap and a unique association of properties is not always possible, so that, in
general, new properties have to be defined for the crossing area, based on physical
arguments to be provided by the modeler. If there is a crossing of fractures with very
different permeabilities, one fracture always dominates a crossing from a geological
point of view. For example, if there exists a highly permeable fracture which
becomes intersected over time by an almost impermeable fracture, the crossing
permeability is more likely to be almost impermeable than highly conductive or
averaged. It is then neither a realistic choice to always average the permeabilities
in a crossing nor to neglect the connection between different fractures. In the
forthcoming reduced models, to simplify the notation, we focus our attention on
a single fractures intersection inside the porous domain where several fracture
branches Ok meet. In all the subsequent cases, the model for the flow in the
fractures, in the surrounding rock matrix and the coupling conditions between each
fracture and the corresponding portion of the rock matrix are the same as in (7), or
equivalently (8), but separately for all the pieces. The reduction process, similarly
to the previous part, replaces the equi-dimensional domain,
Tf which represents the
intersecting region, to a single point, which is ip :D nkD1 Ok with nf the number
of participating branches, and introduce a new variable pO I which represents the
pressure in the intersection.
The equi-dimensional setting for a crossing is shown in Fig. 3 on the example
of four intersecting fracture branches. The equi-dimensional model domain can be
58 B. Flemisch et al.

Fig. 3 Crossing with y

intersection geometries and
location of pressure
unknowns in the
equidimensional model

d I x
p pI

p
p
n

decomposed into three different domain types: matrix, fracture and crossing, namely
 
D D [ ˝i [ [ i [ I:
i i

We define the crossing area I with boundaries to the fractures (solid red lines) and
boundaries to the rock matrix (dashed red lines), respectively as

.@I/f;i ..D @ i \ @I and .@I/m ..D @In.[ @ i /:

i

Introducing the green boundaries in Fig. 4 by connecting the appropriate corners

in our crossing region, we get a closed control area for which we can write the mass
conservation equation which gives a relation between the crossing area pressure and
the adjacent fracture pressures pi . For the reduced model, mass conservation implies

X
nf
uO k   k jip D fI ; (9)
kD1

where  k is the unit tangent along the fracture branch Ok , or, in other words,  k jip is
the unit outward normal of Ok at ip . Moreover, fI 2 R is an integrated source term
given for the intersection.
In the following, we will distinguish three different approaches for assigning
boundary/coupling conditions for the fracture branches Ok at the intersection point
ip . The first two admit rather general intersection situations and are mainly suited
Review of XFEM for Flow in Fractured Porous Media 59

Fig. 4 Crossing with

definition of geometrical
parameters inside the crossing
area I
pI

d p

for the primal formulation in connection with assigning degrees of freedom in

the intersection point for every fracture branch: assuming pressure continuity and
Robin-type conditions. The third assumes an X-shaped intersection of four fracture
branches and is especially tailored for the dual formulation in connection with a
lower-dimensional XFEM approach to capture the discontinuity in the intersection.

2.3.1 Assuming Pressure Continuity

If the properties, e.g., permeability and aperture, of the fractures and in the
intersection are equal, or at least comparable, a simple strategy is to impose a
pressure and normal flux continuity. The model is valid also if we can consider the
intersection as a void space, i.e., infinite permeability, or small enough that it can
be neglected. In the latter case the source term at the intersection may be omitted.
Following [4, 5, 10] and the references therein, we require mass conservation (9)
together with pressure continuity

pO k jip D pO I 8k D 1; : : : ; nf : (10)

With (9) and (10), it is possible to eliminate the value pO I of the pressure at the
intersection. Moreover the primal formulation of (10) is straightforward. In some
cases the heterogeneity between fractures could be severe and the aforementioned
model behaves poorly, see [27].

2.3.2 Robin Boundary Conditions

In [51], an alternative to requiring pressure continuity in the intersection has been

proposed which amounts to replace the Dirichlet-type coupling (10) by Robin-type
conditions for each fracture branch.
Considering the equi-dimensional setup from Figs. 3 and 4, we assume that the
Darcy velocity uk associated with the fracture branch k can be prolongated to the
60 B. Flemisch et al.

intersection region I and be defined there as

1
uk jI D I  k . pI  pk /;
`k

where `k is the distance between the crossing point ip and the point @ k \ Ok .
Proceeding to the reduced model and integrating along the green lines in Fig. 4
yields the Robin boundary condition

dk
uO k   k jip D  k > I  k . pO k  pO I /; (11)
`k
ˇ ˇ
where dk D ˇ@ k ˇ is the length of the interface (green) for fracture k within I and
 k the unit outward normal on that interface.
The mass conservation (9) can be rewritten as

X
nf
dk
 k > I  k . pO k  pO I / D fI : (12)
kD1
`k

Conditions (11) and (12) can be easily incorporated into the primal hybrid-
dimensional problem formulation (8).

2.3.3 Dual Formulation for X-Shaped Intersections

In the case of two intersecting fractures with an X-shaped intersection, a mathe-

matically rigorous derivation of coupling conditions is presented in [27, 35]. These
conditions are perfectly suited for incorporation into a dual problem formulation
discretized by XFEM. In this case, we have nf D 4 but we make an explicit use
of the fact that two distinct fractures intersect and associate only one index k with
the two branches of one fracture, see Fig. 5. This allows to formulate for a quantity
qO k associated with fracture k its average ffOqk ggip and jump Oqk ip at the intersection
point ip . The model takes into account the aperture, permeability and angle at the
intersection between fractures as well as the permeability in the intersecting region.

Fig. 5 Geometry and notation for the case of X-shaped intersections

Review of XFEM for Flow in Fractured Porous Media 61

This model allows a pressure and Darcy velocity discontinuity across the inter-
section, where the jumps are computed accounting for the pressure and fluxes from
both fractures. In addition to the mass conservation (9), the coupling conditions are
8
2
ˆ jIj X O ik
ˆ
ˆ
< ffuO k   k ggip D Opi ip
di kD1 dk for i; j D 1; 2; i ¤ j; (13)
ˆ
ˆ O dj
:̂0 O ii uO i   i ip D ffOpi ggip  pO I
di

where dk D dk = sin  and  is the angle between the two fractures, O ij is the
tangential projection along Oj and then Oi of the inverse of the permeability in
the intersection region, namely O ij :D  > 1
i I  j . Note the similarity of (13) to
the interface conditions (7b-bis) of the “full” dual hybrid-dimensional problem.
This allows to directly apply the corresponding XFEM techniques in a lower-
dimensional context. Moreover, the system of equations (13) can be viewed as
a generalization of (10) since the former boils down to the latter providing the
intersection permeability goes to infinity or the dimension of the intersection goes
to zero.

2.4 Boundary Conditions

This section is divided into two parts: Dirichlet boundary conditions for a matrix
boundary that is intersected by fractures and conditions for fracture tips that are
located in the interior of the matrix domain.

2.4.1 Dirichlet Conditions for Fractured Porous Media

Boundary conditions for fractured porous media systems with explicitly modeled
fractures are not easy to define. The simplest choice is to prescribe a constant
pressure along a domain boundary or a linear change, for example for the case
of a hydrostatic pressure distribution. This often does not reflect the highly
heterogeneous structure in the case of fractured porous media systems. That again
leads to a strong influence of the boundary conditions on the solution if the domain is
not chosen large enough. For field scale simulations, one usually obtains pointwise
pressure information from which the best boundary conditions are to be picked.
In [51], a possibility is presented to interpolate pointwise pressure data along a
given boundary including the information of the geometrical position and geological
parameters (aperture, permeability) of the fractures intersecting with this boundary.
In particular, the situation depicted in Fig. 6 is considered. For a boundary
segment !  @˝ that is parametrized by  2 Œ0; 1 , the left and right pressure
values pleft and pright at  D 0 and  D 1 are assumed to be known. The segment
62 B. Flemisch et al.

Fig. 6 Partitioning of the boundary according to intersecting fractures

! is intersected by nf fractures that divide it in nf C 1 parts !i . The goal is

to find a pressure distribution pb on ! that accounts for the presence of the
intersecting fractures by admitting jumps across the fracture-boundary intersections
j . Proceeding analogously to the derivation of Problem 6, one has to find pb such
that pb .0/ D pleft , pb .1/ D pright and

X
f
X
f
n n
 
.b;i rpb ; rqb /!i C O .j / ŒŒpb j ŒŒqb j C 0 .ffpb g j  pf j /ffqb g j D 0;
iD0 jD1
(14)

for all test functions qb . Here, b;i D  > b i  b is the permeability along the boundary
segment, while ŒŒ  j and f  g j refer to the jump and average in the fracture-boundary
intersection j . In order to derive a closed system, one is left with the choice of
the fracture pressures pf j . In [51], two options are discussed. The first one assumes
pf j D f pb g j such that the fracture pressures are indeed an outcome of solving (14).
The second one builds upon expert knowledge to describe explicit values for pf j .
In order to facilitate the solution of (14), it is assumed that pb is piecewise linear
with respect to the unfractured parts !i , namely,

pb ./ D mi  C bi on !i ; (15)

with the two unknown coefficients mi ; bi . By choosing appropriate test functions qb ,

analytical expressions for these coefficients are derived for an arbitrary number of
fractures and the case pf j D f pb g j , see [51].
For example, for one single fracture, nf D 1, the slopes are given by
pright  pleft
m0 D b;0  
;
O
C j!1 j C j!0 j b;0
b;1
C . b;0
b;1
 1/
pright  pleft
m1 D b;1  b;1
:
O C j!0 j C j!1 j b;1
b;0
C .1  b;0 /
Review of XFEM for Flow in Fractured Porous Media 63

In [51], the resulting boundary conditions are shown to be superior to standard

choices. In particular, the transition from the inner part of the domain to the
boundary appears much more natural in the presence of blocking fractures.

2.4.2 Boundary and Coupling Conditions for Fracture Tips

We consider now the situation where parts of the fractures are ending inside the inte-
rior of the matrix domain, namely, in Otip D @O \ ˝. Apart from being challenging
from the mathematical and numerical point of view, the modeling question is what
kind of boundary or coupling conditions should be prescribed at Otip .
An obvious easy choice is to prescribe no-flow conditions across the fracture tip
[7], namely,

uO  jOtip D 0: (16)

In many situations, this condition is well justified by the essential modeling

assumption that the fracture aperture d is small compared to its lateral dimensions.
However, there can be problem settings where (16) may not be accurate enough. In
particular, if the tangential permeability of the fracture is larger than the normal one,
namely, O > O , the flow across the fracture tip could be rather large compared to the
flow over the fracture’s lateral boundaries and should be taken into consideration.
Taking into account the flow across the tip can be achieved by assigning a
corresponding source term f˝ for the matrix domain [50],

f˝ D ıOtip uO  jOtip : (17)

The coupling condition (17) can be complemented by a condition involving the

matrix and fracture pressures. For example, one could aim for

pO jOtip D ŒŒp Otip : (18)

Conditions (17) and (18) are discussed and investigated for the discretized pri-
mal formulation in [50]. While (17) is implemented as a source for the matrix
domain, (18) can be realized as a Dirichlet condition for the fracture. However,
a proper mathematical derivation from the continuous setting as well as thorough
numerical comparisons with (16) are still missing.

3 Numerical Discretization by Means of XFEM

Before we present the numerical approximation of the previous reduced models,

both in dual and primal form, using the extended finite element method (XFEM),
we provide a very brief overview of its historical development and some pointers to
the literature.
64 B. Flemisch et al.

Ideally, one would want to use a mesh that is as structured and axis-aligned
as possible. The standard Galerkin finite-element method, however, cannot handle
discontinuities in the solution except by resolving them through the grid, namely,
by doubling and decoupling the degrees of freedom along the discontinuities.
Coming from the structural-mechanics problem of evolving cracks that leads to
discontinuities in the solution (displacement, stress, strain), an extension to the
standard finite-element scheme was developed, [22, 23, 44], and called “eXtended
Finite Element Method.”
From the more theoretical point, Nitsche’s method, intentionally developed to
handle Dirichlet constraints, evolved to a new possibility to treat interface problems,
[16, 38, 39]. XFEM and Nitsche’s method applied to interface problems are in
this case essentially the same approach. An overview of recent problems where
XFEM methods are investigated is given in [1]. Some works that influence the
following presentation are [23, 40, 44]. XFEM was first used in the fractured
porous media context in [19, 33] for lower dimensional fractures introducing a
discontinuous solution in the matrix, in [11] for lower dimensional fracture networks
having different permeabilities in the network and therefore also discontinuities, and
in [41] for thin heterogeneities (equi-dimensional) which are not resolved directly
with the grid but rather with the XFEM.
In the literature such techniques are very often referred to as “partition-of-unity”
PUFEM and “generalized finite-element methods” GFEM. The difference here is
that those are usually on a global level where XFEM adopts the same techniques on
an element-local formulation. The composite finite element method, first presented
in [36], is a special type of a geometric multi-grid methods and falls therefore in the
category of multi-scale methods.

3.1 Modification and Addition of Basis Functions

In the classical Galerkin finite-element approach, the discrete solution, ph .x/, at

a global point x in space, which lies within an element E, is defined by the sum
over all shape functions associated with this element multiplied by the value of the
corresponding degree of freedom pQ i , cf. for example [14],
X
ph .x/ D bi .x/ pQ i : (19)
i2NE

Here, bi denotes the shape function of the degree of freedom i, NE D fn1 ; : : : ; nr g

denotes the set of standard degrees of freedom of the element E. All matrix elements
which are not cut by a fracture are treated with such a standard finite element
approach.
If an element E is cut by a fracture, additional degrees of freedom pQ je are
introduced. Those elements which are cut by at least one fracture are called enriched
Review of XFEM for Flow in Fractured Porous Media 65

elements. The discrete solution on an enriched element E can be written as

X X
ph .x/ D bi .x/ usi .x/ pQ i C bej .x/ uej .x/ pQ je (20)
i2NE j2Ne;E

Here, Ne;E is the set of enriched degrees of freedoms. To capture discontinuities

in the solution the basis functions are multiplied by discontinuous functions,
where usi denotes the discontinuity functions for the standard degrees of freedom,
while uej denotes the discontinuity functions for the enriched degrees of freedom,
respectively. These functions will be defined below.
In the course of this work, the standard basis at cut elements is chosen to be same
as for uncut elements, i.e., (bi-)linear and the additional shape-functions are chosen
to be of the same type as the standard shape-functions, i.e., also (bi-)linear, bi D bej
if i and j refer to degrees of freedom located at the same vertex. Then, bej denotes the
nodal shape function of an enriched node j. Furthermore, the XFEM concept is here
used in combination with the Ritz-Galerkin approach, i.e., the basis-function space
and the test-function space are equal.
The choice of the discontinuity functions is somehow arbitrary, as long as certain
conditions are fulfilled. One commonly desired goal is to choose the discontinuity
functions such that the resulting enriched basis functions are forced to be zero in all
nodes. On the one hand this leads to a propitious quality: the nodal interpolation is
still guaranteed by the solution in the standard nodes alone, [44]. More importantly
this property yields to enriched basis functions which are completely local with
respect to the cut elements and every basis function has only one discontinuity
(within this element) for every set of additional degrees of freedom. This avoids
blending elements, [28], which have to be introduced otherwise. However, this is
only valid for the special case of a single interface per element. The general, more
complex case of several (intersecting) interfaces is more demanding and discussed
in, for example, [50]. There are many other possible choices for the discontinuity
functions with different properties. For example, they can be chosen such that the
standard basis remains unmodified and the discontinuity is only represented by the
enriched basis or such that the mean of the enriched basis functions is zero.
To become more explicit, we define the discontinuity functions by using the sign
function sgn which is positive one on the side of the positive normal direction and
negative one on the other, as

1 1
usi .x/ ..D jsgn.x/ C sgn.xi /j ; uei .x/ ..D jsgn.x/  sgn.xi /j :
2 2

The second term, sgn.xi /, associates a constant value to every node, so that the
discontinuity function for the original degrees of freedom is one if x and xi lie on
the same side of the interface and zero if they are on different sides, and vice versa
for the additional degrees of freedom. The modified basis functions for this kind
of discontinuity functions are exemplarily shown for the one-dimensional case in
Fig. 7. For this approach, the orientation of the normal vector n of the interface has
to be chosen. This choice is arbitrary.
66 B. Flemisch et al.

Fig. 7 The zero dimensional fracture with local coordinate  divides the one dimensional matrix
element. Solid lines show the two modified basis functions associated to node two (degrees of
freedom two and five), dashed lines the modified basis of degrees of freedom three and six at node
three.

3.2 Primal Formulation with XFEM

The matrix domain ˝ is discretized by nm triangular or quadrilateral elements Ej

m
into Thm D fEj gnjD1 independent of  . The fracture  is discretized with lower-
j f
dimensional elements, Thf D fEf gnjD1 , independent of Thm . We define all elements
E 2 Thm which are totally in or partly belonging to ˝i as Ei D E \ ˝i . All elements
which are not fully included in one sub-domain belong to both. The discrete space
can then be defined as

Qhi D fqh;i 2 C0 .˝i / W qh;i jE\˝i 2 Q1 .Ei /; E 2 Thm g

for quadrilateral elements E, with Q1 being replaced by P1 for triangular elements.

The complete discrete space for the rock matrix domain is then just the product
space of the sub-domain spaces Qhm D Qh1  Qh2 , where the elements cut by a
fracture are contained twice but each with the cut basis. The discrete space for 
reads

Qhf D fOqh 2 C0 . / W qO h jEf 2 Q1 .Ef /; Ef 2 Thf g

so that the combined space is Qh D Qhm  Qhf . This allows to obtain the discrete
formulation of Problem 6.
Problem 7 (Primal Hybrid-Dimensional Discrete Formulation) Find ph D
. ph;1 ; ph;2 / and pO h in subspaces of Qhm and Qhf that respect the given boundary
conditions such that
(
ap . ph ; qh /  cp .qh ; pO h / D F.qh /
;
O qh /
cp . ph ; qO h / C aO p . pO h ; qO h / D F.O

for all test functions qh and qO h defined in proper subspaces of Qhm and Qhf .
Review of XFEM for Flow in Fractured Porous Media 67

3.3 Dual Mixed Formulation

In [19], and in some more recent works such as [30–32] the XFEM is applied to
the dual mixed formulation of the problem in a similar way, but with different
FEM spaces. In particular, the lowest order Raviart-Thomas pair RT0 ; P0 , see
[47, 49], is employed for velocity and pressure, respectively. This is a common
choice in porous media simulations, which guarantees local mass conservation.
In the aforementioned works the domain is discretized by means of a triangular
or tetrahedral grid, however, the method could be generalized to the case of
quadrilateral or hexahedral grids.
The cut mixed finite element ˚ spaces can be defined  as follows. For each
element Em let RT0 .Em;i / D vh jEm;i W vh 2 RT0 .Em / be the restriction of the
standard
˚ RT0 functions to the sub-element Em;i , and analogously let P0 .Em;i / D
qh jEm;i W qh 2 P0 .Em / be the restriction of the standard P0 functions. See Fig. 8 for
a sketch of the restricted basis functions and the corresponding degrees of freedom
in the 2D case.
The discrete velocities and pressure in ˝ are then sought in the following spaces
respectively:

Vh D V1;h  V2;h Qh D Q1;h  Q2;h

where
˚ 
Vi;h D vh 2 Hdiv .˝i / W vh 2 RT0 .Em;i / 8Em 2 Thm
˚ 
Qi;h D qh 2 L2 .˝i / W qh 2 P0 .Em;i / 8Em 2 Thm :

The finite element spaces for the fracture problem, on  , are the standard RT0 -P0
in N1 dimensions, thus, the discrete flux and pressure in the fracture are sought in
the spaces
˚ 
O i;h D vO h 2 Hdiv . / W vO h 2 RT0 .Ef / 8Ef 2 Th
V f
˚ 
QO i;h D qO h 2 L2 . / W qO h 2 P0 .Ef / 8Ef 2 Thf :

Fig. 8 Basis functions for the lowest order Raviart-Thomas pair, restricted to the subdomains ˝i
68 B. Flemisch et al.

In the case of intersecting fractures, one could consider non-matching fracture grids
at the intersection: in this case, a suitable XFEM enrichment should be considered
also in the fractures. We refer to [27] for details.
We can now define V Q h D Vh  V, O and QQ h D Qh  Q
ch , and Wh D VQh Q Q h and
formulate the discrete version of Problem 4.
Problem 8 (Dual Hybrid-Dimensional Discrete Formulation) Find .uh ; uO h ;
ph ; pO h / 2 Wh such that

ah .uh ; uO h ; vh ; vO h / C bh . ph ; pO h ; vh ; vO h /  bh .qh ; qO h ; uh ; uO h / D
F .vh ; vO h ; qh ; qO h / 8.vh ; vO h ; qh ; qO h / 2 Wh

for all test functions qh and qO h defined in proper subspaces of Qhm and Qhf .
The well-posedness of the dual discrete problem has been proven in [19] for
the case of given pressure in the fracture, and in [21] for the fully coupled case.
Particularly relevant is the problem of the inf-sup stability of the extended spaces:
indeed, even if we start from a stable pair the enriched spaces could present
instabilities in some particular configurations. Sufficient conditions on the fracture
geometry and on the underlying grid of the porous medium are given in the two
aforementioned works.

3.4 Fracture Grids and Approximation of the Coupling Terms

Once the finite element spaces for both the fracture and the rock matrix are defined,
one crucial ingredient is the approximation of the coupling term between the two
media. To simplify the presentation, we consider only fracture and matrix grids
which are genuinely non-matching if each fracture element is fully contained in a
N-dimensional element or is contained in a pair of facing N-dimensional elements.
In the important case of matching geometries, a standard technique can be employed
to approximate the coupling term, see for example [29]. The construction of the
fracture grid can be done in two different ways. One possibility is to consider the
fracture grid induced by the intersection between the background mesh and the
interface. This approach avoids the construction of complex interpolation operators,
explained in the sequel, but can be done easily only in the two-dimensional case
and may produce elements with strongly varying aspect ratios. One possibility to
overcome these difficulties consists in using the trace of higher dimensional basis
functions as in the Trace FEM method, coupled with suitable stabilizations, see
[17, 43, 45]. To allow a higher flexibility in the numerical discretization, it is possible
to introduce an interpolation operator M W QO h ! Qh which maps the value of
the pressure in the fracture elements to the corresponding element in the matrix
grid. Note that it has to take into account also fractions of fracture elements. In the
Review of XFEM for Flow in Fractured Porous Media 69

particular case of a piecewise constant approximation for pressure, following [34],

the discrete version of M is a rectangular matrix Mh with entries
Z
ŒMh ij D 1dx;
KO j \Ki

where Ki is the i-th element in matrix mesh and KO is the j-th element in the
fracture mesh. To preserve mass conservation at discrete level, we consider the
approximation of the adjoint operator M W Qh ! QO h as the transpose of Mh .
In [29], the authors note that, in the case of an immersed fracture, if the fracture
grid is too fine compared with the mesh of the matrix, depending also on the
permeability contrast, the solution in the fracture could present oscillations. The
authors suggest a possible explanation which is related to the singularity of the
solution at the fracture tip. The aforementioned work is in the context of non-
matching, but still conforming (i.e. aligned with the grid) discretizations, and the
authors consider a mortar technique to deal with the non-matching elements at
the interface. The same phenomenon is also observed when the XFEM method is
employed. In fact considering the simple domain depicted in Fig. 9, it is possible
to obtain the solutions reported in Fig. 10 where the fracture behaves like a barrier
only in its middle part. The domain is discretized using a structured triangular mesh
where each boundary edge is approximated with n segments, while the fracture is
discretized using m segments. The results show that, also in the case of XFEM, the
oscillations occur when the discretization of the fracture is finer than the rock mesh.
Apart from cases where the mesh is too coarse the oscillations exhibit a frequency
that depends only on the discretization of the outer medium. In this particular case
it can be estimated as n=2 for most of the cases. We notice that this frequency
corresponds to the nth eigenfunction of the problem. Indeed, if n is odd the
solution becomes asymmetrical. It is also interesting that, for fixed n, the amplitude
of the oscillations is constant with increasing m, while if we refine both fracture and

0.25 0.5 0.25

0.47

Fig. 9 Representation of the domain for the oscillation problem. In the fracture we highlight the
pieces with different permeability
70 B. Flemisch et al.

Fig. 10 In the left, the grid sizes of the matrix and the fracture are comparable. In the centre, the
fracture grid is finer than the rock grid, some oscillations are present. In the right, both meshes
are refined maintaining the same ratio of the grid size as the solution in the centre. In this case the
amplitude of the oscillations decreases

Fig. 11 Basis function enrichment around fracture tips. Red circles indicate XFEM enrichment,
green squares enrichment by radial functions. Left: pure XFEM enrichment according to [50].
Right: XFEM and radial functions enrichment as suggested in [37]

medium the amplitude decreases. This phenomena are particularly relevant when
the normal or tangential (or both) permeability change sharply along the fracture.

3.5 Basis Function Enrichment Around Fracture Tips

If a fracture branch ends inside the interior of the matrix domain in the fracture tip
Otip , the question arises on how to enrich the basis functions inside an element Etip
that contains Otip . For the primal formulation, an ad-hoc solution is presented and
used in [50]. In Etip , only those basis functions are enriched that correspond to the
vertices of the element face that is intersected by the fracture, see the left picture of
Fig. 11. While this approach is attractive for its simplicity, it disregards the potential
singularity in the solution at the tip Otip . As described in [37], it is more appropriate
to account explicitly for this singularity by adding radial functions
p p
g1 .r; / D r sin.=2/; g2 .r; / D r cos.=2/;

multiplied with the standard basis functions bi in Etip . Here, r is the distance from
Otip , while  indicates the angle with respect to the tangential fracture direction. This
situation is depicted in the right picture of Fig. 11.
Review of XFEM for Flow in Fractured Porous Media 71

4 Solvers

In this section, we discuss some issues related with the numerical solution of the
system resulting from numerical discretization. In particular, we present the problem
of ill-conditioning in the presence of very small sub-elements, and the possibility of
using iterative strategies in a domain-decomposition framework as an alternative to
a fully monolithic approach.

4.1 Conditioning

Since we are considering an arbitrary position of the interface with respect to

the underlying grid, it is possible that, when an element E is cut into two sub-
elements E1 , E2 , the ratio jE ij
jEj is very small or even zero for some elements/some
configurations of the interface. The case of jEi j D 0, corresponding to a fracture
that contains one or two adjacent vertices of the element, is an exception ruled out
by the assumptions of the XFEM method and should be handled separately, while
the case of small sub-elements can lead to ill-conditioned matrices. In particular,
the condition number can degenerate as .Thm ; O / D minE2G miniD1;2 jE ij
jEj tends
to zero, where G denotes the set of elements that are cut by the interface. This
problem has been investigated for the case of a mixed formulation in [19]. In
this work, a numerical experiment is presented, where, changing the position of
a vertical fracture on a structured grid, the authors obtain smaller and smaller values
of .Thm ; O / and compute the corresponding maximum and minimum eigenvalue of
the matrix C defined as
 
A BT
CD ;
B 0

arising from the discretization with XFEM of the Darcy problem in the cut bulk
medium. While the maximum eigenvalue is approximately constant, the minimum
decreases with . Even if the matrix is not symmetric and positive definite, the ratio
between maximum and minimum eigenvalue

maxi ji .C/j

ic.C/ D
mini ji .C/j

can be used as an indicator of ill-conditioning. In [19] an optimal preconditioner P

for the problem is presented and tested on the same problem, showing that ic.P1 C/
is constant for all . Moreover, a simpler diagonal preconditioner PL can be used
to perform matrix equilibration: in this case the preconditioner is not optimal with
respect to the grid size h, but for a fixed h one still obtains a constant ratio ic.P1
L C/.
72 B. Flemisch et al.

4.2 Iterative Approaches

The coupled problem of flow in a porous matrix and a fracture has been interpreted
in a domain-decomposition framework in [2] under the assumptions of pressure
continuity across the fracture, i.e. the assumptions of permeable fractures. It is
shown how, in the cases of a fracture that cuts the domain ˝ in two disjoint parts ˝i
the problem can be recast as a global equation on the interface O for the unknown pO .
In a more recent work, [42], the concept is generalized to the case of fractures with
arbitrary permeability, including the case of (nearly) impermeable interfaces. Once
again the problem can be formulated as a positive definite problem on the interface
O . In particular, in the case  D 1 the problem simplifies again to a problem for only
one scalar unknown pO ,

O  pO / D ff
S1 . pO ; f1 / C S2 . pO ; f2 / C r  .r (21)

where S1;2 . pO ; f1;2 / are the Robin-to-Neumann operators accounting for the coupling
with the flow problem in ˝1;2 . The interface Eq. (21) can then be solved iteratively.
This approach is meant to provide an efficient method for the solution of the
coupled fracture-medium problem: indeed, when considering a mixed formulation,
the system is not positive definite and in realistic configurations it can be very large.
Therefore, it can be convenient to eliminate some of the unknowns to obtain a
problem that is easier to solve.
In the context of non-matching discretizations, an iterative approach for the
solution of the coupled problem is also discussed in [19], where a similar approach
is used, adding more information and numerical evidence about the convergence of
the procedure. In particular, it is proven that the iterative method converges for some
values of a relaxation parameter ! that must satisfy
( s )
O maxf1 ; 2 g
! & M2O ; where MO D min 1; h :
0 O O

Numerical experiments show that, for small O =O ratios the relaxation parameter
can be very small and convergence is achieved in just few iterations, while for higher
ratios the iterative method converges very slowly.
In the relevant case of a discrete fracture network approximation, namely, where
the rock matrix is supposed to be impervious, an interesting approach is proposed
in [11, 12, and references therein] to solve in an efficient way the global system
of equations. The authors consider the continuous coupling conditions among the
fractures, as presented in Sect. 2.3.1, and a primal formulation of the problem. In
this part, to simplify the notation we assume that the equi-dimensional domain are
the fractures while the one-codimensional domain are the intersections among the
fractures. To decouple the solution on each fracture, the transmission conditions (10)
are imposed in a weak way through an optimization problem, i.e. naming S a strip
Review of XFEM for Flow in Fractured Porous Media 73

of intersecting elements between two fractures solve

8 X

ˆ
<min J.p/ D min
p
2 C 2kffu  ngg k2
S S
S2S
:̂ s.t. p solution in each fracture

where S is the set of all the intersection regions, the norms are defined on proper
spaces and u  n is a suitable reconstruction of the normal flux at the intersection S.
The optimal solution of the minimum problem gives J. p/ D 0 and can be computed
numerically using a gradient method. With this method the linear system, which
couple all the fractures, has much smaller size than the aforementioned approaches
and the computation of the pressure in each fracture is completely parallelizable.

5 Conclusions

In this review paper, we presented several mathematical models and numerical algo-
rithms to simulate single-phase flow in a porous medium containing fractures. Two
main challenges are addressed. First, the fractures play a crucial role in subsurface
flows and should be carefully accounted for to achieve reliable simulations, however
their geometrical and geological data pose several difficulties from a modeling and
discrete point of view. A common approach which is broadly used in the literature is
to consider an hybrid-dimensional model where the fractures are treated as objects
of lower dimension. Second, the position of the fractures may be unknown and
several scenarios are needed to obtain a representative solution of the problem, or to
speed up the simulations: for this reason it is better to allow these fractures to be geo-
metrically decoupled from the surrounding porous medium. In this case, an XFEM
approach can be a valuable option to overcome this requests. Several geometrical
difficulties have been addressed in this paper, such as the treatment of intersections
and tips. To the best of our knowledge, several interesting issues are still open for
further investigation, such as a full three-dimensional setting for general networks
of fractures with XFEM, a physical derivation of a more appropriate condition at the
fracture tip, suitable stabilizations to increase the robustness of the linear solvers and
a deep analysis for the case of vanishing aperture.

References

1. Abdelaziz, Y., Hamouine, A.: A survey of the extended finite element. Comput. Struct. 86(11–
12), 1141–1151 (2008)
2. Alboin, C., Jaffré, J., Roberts, J.E., Wang, X., Serres, C.: Domain decomposition for some
transmission problems in flow in porous media. In: Numerical Treatment of Multiphase Flows
in Porous Media (Beijing, 1999). Lecture Notes in Physics, vol. 552, pp. 22–34. Springer,
Berlin (2000)
74 B. Flemisch et al.

3. Alboin, C., Jaffré, J., Roberts, J.E., Serres, C.: Modeling fractures as interfaces for flow and
transport in porous media. In: Fluid Flow and Transport in Porous Media, Mathematical
and Numerical Treatment: Proceedings of an AMS-IMS-SIAM Joint Summer Research
Conference on Fluid Flow and Transport in Porous Media, Mathematical and Numerical
Treatment, Mount Holyoke College, South Hadley, Massachusetts, 17–21 June 2001, vol. 295,
pp. 13–25. American Mathematical Society, Providence (2002)
4. Alboin, C., Jaffré, J., Roberts, J.E., Serres, C.: Modeling fractures as interfaces for flow and
transport in porous media. In: Fluid Flow and Transport in Porous Media: Mathematical and
Numerical Treatment (South Hadley, MA, 2001). Contemporary Mathematics, vol. 295, pp.
13–24. American Mathematical Society, Providence (2002)
5. Amir, L., Kern, M., Martin, V., Roberts, J.E.: Décomposition de domaine et précondition-
nement pour un modèle 3D en milieu poreux fracturé. In: Proceeding of JANO 8, 8th
Conference on Numerical Analysis and Optimization (2005)
6. Angot, P.: A model of fracture for elliptic problems with flux and solution jumps. C. R. Math.
337(6), 425–430 (2003)
7. Angot, P., Boyer, F., Hubert, F.: Asymptotic and numerical modelling of flows in fractured
porous media. M2AN Math. Model. Numer. Anal. 43(2), 239–275 (2009)
8. Antonietti, P.F., Formaggia, L., Scotti, A., Verani, M., Verzotti, N.: Mimetic finite difference
approximation of flows in fractured porous media. Technical Report, Politecnico di Milano
(2015)
9. Bear, J.: Dynamics of Fluids in Porous Media. American Elsevier, New York (1972)
10. Benedetto, M.F., Berrone, S., Pieraccini, S., Scialò, S.: The virtual element method for discrete
fracture network simulations. Comput. Methods Appl. Mech. Eng. 280(0), 135–156 (2014)
11. Berrone, S., Pieraccini, S., Scialò, S.: On simulations of discrete fracture network flows with
an optimization-based extended finite element method. SIAM J. Sci. Comput. 35(2), 908–935
(2013)
12. Berrone, S., Pieraccini, S., Scialò, S.: An optimization approach for large scale simulations of
discrete fracture network flows. J. Comput. Phys. 256(0), 838–853 (2014)
13. Bonn, W.M., Nordbotten, J.M.: Robust discretization of flow in fractured porous media.
arXiv:1601.06977 [math.NA] (2016)
14. Braess, D.: Finite Elements: Theory, Fast Solvers, and Applications in Solid Mechanics.
Cambridge University Press, Cambridge (2007)
15. Brezzi, F., Fortin, M.: Mixed and Hybrid Finite Element Methods. Computational Mathemat-
ics, vol. 15. Springer, Berlin (1991)
16. Burman, E., Hansbo, P.: Fictitious domain finite element methods using cut elements: II. A
stabilized Nitsche method. Appl. Numer. Math. 62(4), 328–341 (2012)
17. Burman, E., Claus, S., Hansbo, P., Larson, M.G., Massing, A.: CutFEM: discretizing geometry
and partial differential equations. Int. J. Numer. Methods Eng. 104(7), 472–501 (2015)
18. D’Angelo, C., Scotti, A.: A Mixed Finite Element Method for Darcy Flow in Fractured
Porous Media with Non-Matching Grids. Technical Report, MOX, Mathematical Department,
Politecnico di Milano, 2010
19. D’Angelo, C., Scotti, A.: A mixed finite element method for Darcy flow in fractured porous
media with non-matching grids. Math. Model. Numer. Anal. 46(02), 465–489 (2012)
20. Dassi, F., Perotto, S., Formaggia, L., Ruffo, P.: Efficient geometric reconstruction of complex
geological structures. Math. Comput. Simul. 46(02), 465–489 (2014)
21. Del Pra, M., Fumagalli, A., Scotti, A.: Well posedness of fully coupled fracture/bulk Darcy
flow with XFEM. Technical Report 25/2015, Politecnico di Milano (2015). Submitted to:
SIAM Journal on Numerical Analysis
22. Dolbow, J.: An extended finite element method with discontinuous enrichment for applied
mechanics. Ph.D. thesis, Northwestern University (1999)
23. Dolbow, J., Moës, N., Belytschko, T.: Discontinuous enrichment in finite elements with a
partition of unity method. Finite Elem. Anal. Des. 36(3–4), 235–260 (2000)
24. Ern, A., Guermond, J.-L.: Theory and Practice of Finite Elements. Applied Mathematical
Sciences. Springer, New York (2004)
Review of XFEM for Flow in Fractured Porous Media 75

25. Faille, I., Flauraud, E., Nataf, F., Pégaz-Fiornet, S., Schneider, F., Willien, F.: A new fault model
in geological basin modelling. Application of finite volume scheme and domain decomposition
methods. In: Finite Volumes for Complex Applications, III (Porquerolles, 2002), pp. 529–536.
Hermes Science Publishing, Paris (2002)
26. Faille, I., Fumagalli, A., Jaffré, J., Roberts, J.E.: A double-layer reduced model for fault flow
on slipping domains with hybrid finite volume scheme. SIAM: J. Sci. Comput. (2015, in
preparation). hal-01162048
27. Formaggia, L., Fumagalli, A., Scotti, A., Ruffo, P.: A reduced model for Darcy’s problem in
networks of fractures. ESAIM: Math. Model. Numer. Anal. 48, 1089–1116 (2014)
28. Fries, T.: A corrected XFEM approximation without problems in blending elements. Int. J.
Numer. Methods Eng. 75(5), 503–532 (2008)
29. Frih, N., Martin, V., Roberts, J.E., Saâda, A.: Modeling fractures as interfaces with nonmatch-
ing grids. Comput. Geosci. 16(4), 1043–1060 (2012)
30. Fumagalli, A.: Numerical Modelling of Flows in Fractured Porous Media by the XFEM
Method. Ph.D. thesis, Politecnico di Milano (2012)
31. Fumagalli, A., Scotti, A.: Numerical modelling of multiphase subsurface flow in the presence
of fractures. Commun. Appl. Ind. Math. 3(1) (2011)
32. Fumagalli, A., Scotti, A.: A reduced model for flow and transport in fractured porous media
with non-matching grids. In: Proceedings of ENUMATH 2011, the 9th European Conference
on Numerical Mathematics and Advanced Applications. Springer, Berlin (2012)
33. Fumagalli, A., Scotti, A.: An efficient XFEM approximation of Darcy flows in fractured porous
media. MOX Report 53 (2012)
34. Fumagalli, A., Scotti, A.: A numerical method for two-phase flow in fractured porous media
with non-matching grids. Adv. Water Resour. 62(Part C(0)), 454–464 (2013). Computational
Methods in Geologic CO2 Sequestration
35. Fumagalli, A., Scotti, A.: An efficient XFEM approximation of Darcy flow in arbitrarly
fractured porous media. Oil Gas Sci. Technol. - Rev. d’IFP Energies Nouv. 69(4), 555–564
(2014)
36. Hackbusch, W., Sauter, S.A.: Composite finite elements for the approximation of PDEs on
domains with complicated micro-structures. Numer. Math. 75(4), 447–472 (1997)
37. Hanowski, K., Sander, O.: Simulation of deformation and flow in fractured, poroelastic
materials (2016). Preprint, arXiv:1606.05765
38. Hansbo, P.: Nitsche’s method for interface problems in computational mechanics. GAMM-
Mitteilungen 28(2), 183–206 (2005)
39. Hansbo, A., Hansbo, P.: An unfitted finite element method, based on Nitsche’s method, for
elliptic interface problems. Comput. Methods Appl. Mech. Eng. 191(47–48), 5537–5552
(2002)
40. Hansbo, A., Hansbo, P.: A finite element method for the simulation of strong and weak
discontinuities in solid mechanics. Comput. Methods Appl. Mech. Eng. 193(33), 3523–3540
(2004)
41. Huang, H., Long, T.A., Wan, J., Brown, W.P.: On the use of enriched finite element method to
model subsurface features in porous media flow problems. Comput. Geosci. 15(4), 721–736
(2011)
42. Martin, V., Jaffré, J., Roberts, J.E.: Modeling fractures and barriers as interfaces for flow in
porous media. SIAM J. Sci. Comput. 26(5), 1667–1691 (2005)
43. Massing, A., Larson, M.G., Logg, A.: Efficient implementation of finite element methods on
non-matching and overlapping meshes in 3d. arXiv preprint arXiv:1210.7076 (2012)
44. Mohammadi, S.: Extended Finite Element Method. Wiley, New York (2008)
45. Olshanskii M.A., Reusken, A., Grande, J.: A finite element method for elliptic equations on
surfaces. SIAM J. Numer. Anal. 47(5), 3339–3358 (2009)
46. Quarteroni, A., Valli, A.: Numerical Approximation of Partial Differential Equations. Springer
Series in Computational Mathematics, vol. 23. Springer, Berlin (1994)
47. Raviart, P.-A., Thomas, J.-M.: A mixed finite element method for second order elliptic
problems. Lect. Notes Math. 606, 292–315 (1977)
76 B. Flemisch et al.

48. Remij, E., Remmers, J., Huyghe, J., Smeulders, D.: The enhanced local pressure model for the
accurate analysis of fluid pressure driven fracture in porous materials. Comput. Methods Appl.
Mech. Eng. 286, 296–312 (2015)
49. Roberts, J.E., Thomas, J.-M.: Mixed and Hybrid Methods. Handbook of Numerical Analysis,
vol. II, pp. 523–639. North-Holland, Amsterdam (1991)
50. Schwenck, N.: An XFEM-based model for fluid flow in fractured porous media. Ph.D. thesis,
Department of Hydromechanics and Modelling of Hydrosystems, University of Stuttgart
(2015)
51. Schwenck, N., Flemisch, B., Helmig, R., Wohlmuth, B.: Dimensionally reduced flow models in
fractured porous media: crossings and boundaries. Comput. Geosci. 19(6), 1219–1230 (2015)
52. Tunc, X., Faille, I., Gallouët, T., Cacas, M.C., Havé, P.: A model for conductive faults with
non-matching grids. Comput. Geosci. 16, 277–296 (2012)
53. Zimmerman, R.W., Kumar, S., Bodvarsson, G.S.: Lubrication theory analysis of the permeabil-
ity of rough-walled fractures. Int. J. Rock Mech. Mining Sci. Geomech. Abstr. 28(4), 325–331
(1991)
 Part II
Enhanced Finite Element Formulations
for Fracture and Interface Problems
Modeling of Fracture in Polycrystalline
Materials

Steffen Beese, Stefan Loehnert, and Peter Wriggers

Abstract Predicting the behaviour of fracture processes within polycrystalline

microstructures will help to develop more accurate mesoscale material models
and will give insight to effects which can only be measured ex-situ. Therefore a
non-local damage model is introduced and coupled to finite deformation crystal
plasticity. Cracks are represented sharply by using the extended finite element
method in combination with level set techniques. As damage evolves cracks start to
propagate. A new crack propagation algorithm is presented and studied by academic
examples.

1 Introduction

In industrial forming processes metals are subjected to large plastic deformations.

Related to the plastic deformation within the material microstructure pores start to
grow, merge and finally lead to microcracks within the material. Macroscopically
this is recognised as a loss of stiffness of the structural behaviour and a reduced
strength. This effect is usually modelled with continuum damage mechanics as done
by Tvergaard and Needleman in [29]. In the context of finite elements, during the
last decades several theories were established to circumvent the mesh dependent
localization. Most of them, see e.g. [10, 23] or [26], are based on the introduction
of a new degree of freedom for the thermodynamic driving force of the damage.
This contributions follows the work of Reusch et al. [26] who determinate the new
degree of freedom by solving an additional scalar balance equation. Nevertheless,
if damage evolves and is used to predict the initiation or propagation of cracks
the global stiffness matrix becomes ill-conditioned. Furthermore cracks are just
represented in a smeared way and their dimension is related to the mesh size and
some artificial internal length parameter.

S. Beese () • S. Loehnert • P. Wriggers

Institute of Continuum Mechanics, Leibniz Universität Hannover, Appelstraße 11, Hannover,
Germany
e-mail: [email protected]; [email protected];
[email protected]

© Springer International Publishing Switzerland 2016 79

G. Ventura, E. Benvenuti (eds.), Advances in Discretization Methods,
SEMA SIMAI Springer Series 12, DOI 10.1007/978-3-319-41246-7_4
80 S. Beese et al.

To overcome this drawback the damage is transferred to discrete cracks if it

exceeds a material depended threshold value. The representation of fracture is
modeled with the eXtended Finite Element Method (XFEM) according to [4]. In
combination with the level set techniques [5] this numerical tool enables a nearly
mesh independent crack representation. The position of the crack is described
implicitly by level set functions, and the discontinuity within the displacement
field is captured using enrichment functions and additional degrees of freedom.
Since classical crack propagation criteria lose their validity in a finite deformations
context and for inelastic material behaviour an alternative approach is described.
The argumentation for the chosen damage based criterion is that once the crack
is initiated it will propagate if the pores around the crack front will continue to
coalesce. So the mechanisms for crack initiation and propagation are the same. A
very elegant way to construct the crack extension implicitly is the global crack
tracking algorithm proposed in Oliver and Huespe [22] and the level set update
procedures proposed by Gravouil et al. [12] and Moës et al. [19]. Here a combination
of both approaches is presented and a different solution method for the level set
update is suggested. The accuracy of both methods is studied in a similar way Duflot
proposed it in [8].
After the reconstruction of the level set they have to be re-initialised and re-
orthogonalised by the solution of a Hamilton-Jacobi equation. After the propagation
no remeshing is needed, but the enrichment scheme has to be updated in the vicinity
of the crack front. There is also a need for adapting the history variables in these
elements. This becomes necessary because the integration points are determined by
a subcell integration strategy as commonly used in the XFEM since the early work
[18]. Within this strategy the finite element is divided along the interface and the
integration points are placed in these subcells in order to consider the discontinuous
character of the displacement field. This framework is used in combination with a
crystal plasticity model with viscous regularization as proposed e.g. by Steinmann
and Stein [27] but within the context of elastic isomorphism [6, 28] to predict the
evolution of microcracks.

2 Problem Statement

In this section the boundary value problem of a fractured body is elaborated. After
a short review of the balance equation the constitutive framework is presented. By
coupling of the constitutive equations for crystal plasticity with a non-local damage
formulation a natural way of degrading material strength is introduced.
Modeling of Fracture in Polycrystalline Materials 81

2.1 Balance of Linear Momentum for a Fractured Body

A body B  E3 is a subset of the three dimensional Euclidean space and can be

identified by the position X 2 E3 of its material points. At time t D 0 the body
is in its initial configuration .B; 0/ D B 0 and at current time t the deformed
body occupies the current configuration .B; t/ D B t . The deformation  in this
context is a bijective mapping  W E3  R ! E3 . In Fig. 1 a fractured body B is
schematically displayed in the current configuration. Cracks are allowed to be fully
embedded in the domain and to penetrate the surface @B of the body. The crack
surfaces are assumed to be traction free. A surface traction Nt is applied on the
boundary @Bt as Neumann condition and some displacement uN is specified on @But
as Dirichlet boundary condition. If body forces are not present the strong form of
the resulting quasi static boundary value problem defined in the initial configuration
reads:
BVP 1 Find u.X; t/ 2 E3 such that:

DivP D 0 8X 2 B 0
PN D TN 8X 2 @B0
PN D 0 8X 2 0
u D uN 8X 2 @Bu0 :

Here P is the first Piola-Kirchhoff stress, N is the unit normal vector to the boundary
@B 0 and T is the surface traction t with respect to an area element in the initial
configuration.

Fig. 1 Schematic boundary

value problem of cracked
body in current configuration
82 S. Beese et al.

2.2 Constitutive Equations for Finite Deformations Crystal

Plasticity

Due to its strong relevance the modelling of crystal plasticity under finite defor-
mations has a long tradition. Therefore extensive literature exists, and textbooks as
well as overview articles cover this topic. Here just a short introduction to a simple
crystal plasticity model is given. While the basic kinematic setup of continuum
crystal plasticity theories is well established, see e.g. Asaro [1], there are still
many experimental observable features as for example the deformation induced
self organisation of dislocation patterns in dislocation cells, which are not yet fully
understood. The constitutive model is embedded in the framework of elastic material
isomorphism as e.g. motivated by Svendsen [28] or Bertram [6, 7].
Metals have a microstructure built of grains. A grain is modelled as ideal crystal
with inherent oriented crystallographic planes. Plastic deformation is introduced
as simple shear caused by sliding of several of this planes against each other
keeping the crystalline structure unchanged. Because the crystalline symmetry is
not changed the elastic behaviour also remains unchanged. If the elastic behaviour
is assumed to be Hyperelastic, the second Piola-Kirchhoff stress S is defined as:

@ E
SD2 ; (1)
@C
with representing the free energy density and C the right Cauchy-Green tensor.
As commonly done the free energy density is assumed to decompose additively into
an elastic and an isotropic hardening part and is of the form:

WD .C; P;  / D E .C; P/ C iso . / : (2)

Here fP;  g are internal kinematic like variables, with P being the plastic
transformation and  being the accumulated plastic slip. If the elastic behaviour is
not allowed to change during plastic deformation we obtain a material isomorphism
and the dependence of the free energy on the right Cauchy-Green tensor and
the plastic transformation has to be of the form: E .C; P/ D E .P T CP/.
Equation (1) then reads:

@ E .P
T
CP/ @ E .CE /
SD2 D 2P PT ; (3)
@C @CE
„ ƒ‚ …
DWSE

with CE WD P T CP.
As already mentioned, the plastic deformation is the combination of the slip of
all active slip systems ˛. A slip system ˛ is formed by the direction of the slip,
modelled by the Euclidean vector s˛ and the plane ˛ where the slip is acting defined
by its normal vector m˛ . Since they are orthogonal they can be understood as dual
Modeling of Fracture in Polycrystalline Materials 83

vectors of the ˛ slip system:

s˛  m˛ D 0 : (4)

Here s˛ is a covariant and m˛ is a contravariant vector. With this definition and

the rules for the pull-back of co- and contravariant vectors we can obtain the slip
direction and the normal vector in the initial configuration by:

S˛ D F1 s˛
(5)
M˛ D FT m˛ :

The introduction s˛ as covariant and m˛ as contravariant vector ensures their orthog-

onality under every deformation . Nevertheless if plastic deformation evolves,
both, the Lagrangian vectors and the Euclidean vectors will also evolve. Therefore
one can, as depicted in Fig. 2, pull back the slip system to a time independent
isoclinic reference configuration:

SQ ˛ D P 1 S˛ D P 1 F1 s˛
(6)
Q ˛ D P T M˛ D P T FT m˛ :
M

In order to decide if a slip system ˛ is active the definition of the resolved shear
stress  ˛ is necessary. It is defined as the projection of the Cauchy stress  onto
the slip system ˛. With Eq. (6) and the application of the transformation rules for
stresses the resolved shear stress is computed by:
˛
 ˛ WD  W .m˛ ˝ s˛ / D CE SE W .SQ ˛ ˝ M
Q /: (7)
„ƒ‚…
DWME

Fig. 2 Transformation of slip

systems
84 S. Beese et al.

The slip system ˛ is active if the stress  ˛ exceeds the yield limit y˛ . /. This leads
to the definition of the yield surface as:

˚. ˛ ; y˛ / WD max .˚ ˛ /  0 with ˚ ˛ D  ˛  y˛ . / : (8)

˛

For simplicity we assume the same yield limit for all glide systems y˛ . / D y . /
and the isotropic hardening law:

@ iso
y . / D y0 C : (9)
@

To obtain a complete set of constitutive equations an evolution law for the internal
variables fP;  g needs to be defined. By the application of the principle of maxi-
mum dissipation we obtain the evolution equation for the plastic transformation:

X
n
˛
PP D P Q ;
P˛ SQ ˛ ˝ M (10)
˛D1
„ ƒ‚ …
DW

with P being the slip rate of slip system ˛. In order to avoid an active set strategy
we follow the method of visco-plastic regularisation similar to Steinmann and Stein
in [27]. Therefore a constitutive equation for the slip rates is introduced:
m1
˛ < ˛ >
P˛ D P0 ; (11)
y y

with the reference slip rate P0 , the sensitivity exponent m and < x > being
the Macauley operator. The total amount of slip can then be obtained by simple
summation:

X
nsys
P D P˛ : (12)
˛D1

The evolution equations (10) and (12) are integrated by a modified generalised
midpoint rule.

2.3 Continuum Damage Mechanics Coupled to Crystal

Plasticity

The modelling of failure and fracture in ductile materials is often modelled

with the Gurson-Tvergaard-Needleman model (GTN) [29]. This model is micro
mechanically motivated and captures the growth of a single pore in a perfect
Modeling of Fracture in Polycrystalline Materials 85

plastic matrix and additionally pore nucleation and coalescence. In [20] Needleman
successfully studied the influence of pores around the crack tip and their impact to
the fracturing process with the GTN model. Therefore the model is modified in this
section to fit in the constitutive framework presented in Sect. 2.2.
In the GTN-model damage is characterised by the pore volume fraction f . The
load carrying matrix material is reduced by this volume fraction and the yield
strength is decreased. This is done by the introduction of a new yield surface ˚f˛ :
2 
˛ p
˚f˛ WD  1 C 2q1 f  cosh 3q2  .q1 f  /2  0 : (13)
y y

The hydrostatic pressure p affects now the yield surface. While f  is the effective
void volume fraction, the parameters q1 and q2 are just model parameters and cannot
be physically interpreted. The quantity f  takes into account the void coalescence
and nucleation if a material dependent threshold fc is exceeded. It is determined by:
8
<f ; if f  fc
f  . f / WD 1
q1 fc
(14)
:fc C . f  fc / else :
fF fc

This relation considers the effect that the material strength decreases rapidly close
to the final void volume fracture of failure fF . In order to obtain the value of the
internal variable f a new evolution equation has to be introduced. In [29] an additive
split of f in a part fg related to pore growth and fn responsible for pore nucleation is
proposed as:

f D fg C fn : (15)

Originally the material is plastically incompressible, but by using the yield sur-
face (13) a non deviatoric flow rule and thus a volumetric plastic deformation is
introduced. The source of this volumetric deformation is the growth of pores and
the respective evolution equation reads:

P :
fPg D .1  f  /Tr.P 1 P/ (16)

The pore nucleation part of the evolution equation follows a more phenomenological
argumentation. Since it is a statistical process a standard distribution is assumed as:
2
Pfn D p An exp  .  n / P : (17)
2n 2n

This constitutive law for void nucleation is strain controlled by the mean strain for
pore nucleation n and has the variance n . A further extension of the presented
model is the introduction of a stress controlled constitutive law for the pore
86 S. Beese et al.

nucleation as suggested in [20]. To complete the set of constitutive equation the

free energy function is specified as:

1 
.CE ;  / D .TrCEiso  3/ C .ln.det.CE ///2
2 8
 (18)
1 2 exp.h2  /
C H C h1 C  y0 :
2 h2

2.4 Non-Local Extension of the GTN Model

If the boundary value problem BVP 1 with the constitutive equations of Sects. 2.2
and 2.3 is solved with the finite element method the problem of mesh dependent
localization occurs. This was observed quite early by Pietruszczak and Mróz
[24]. This mesh dependency typically results in a strong dependence of the force
displacement curve on the element size. Instead of convergence to a certain material
response, softening materials tend to diverge with decreasing element size. Another
example is the formation of localisation bands. Here the width of the localised
deformation is close to the element size. So if some material microstructure
constrains the width of such localisation this has to be taken into account explicitly.
Several attempts have been made to circumvent the above described behaviour.
Here we focus on the gradient dependent theories which are associated with the
names micromorphic continua [10] or gradient enhanced damage models [23] and
others. All these models rely on the introduction of a new field variable  and its
gradient r and the coupling of the local evolution of an internal variable to these
quantities. These theories only differ in their motivation and derivation, but end up
with similar additional balance equation. Reusch, Svendsen and Klingbeil proposed
in [26] such a theory relating the evolution of the volume void fraction fP of the GTN
model to its non-local equivalent .P This new field variable has to fulfil the equation:

P  `2
 D fP 8X 2 B 0
(19)
r  N D 0 8X 2 @B 0 ;

with homogeneous Neumann boundary condition. Here the local quantity fP forms
the right hand side of the equation and couples the pure displacement based
boundary value problem BVP 1 with the non-local part. The material parameter
` is related to the size of microscopic material feature and can be determined from
the width of a real measured localisation band. In the example Sect. 5.1 a simple test
problem is studied for different mesh resolutions.
Modeling of Fracture in Polycrystalline Materials 87

2.5 Weak Form of the Resulting Boundary Value Problem

To summarize the equation introduced so far we will present the weak form of the
boundary value problem BVP 1 and Eq. (19). This weak form is given in the initial
configuration. By the standard procedure of multiplying this both equations with
some arbitrary functions fıu; ıg, integrating over the entire domain and applying
the divergence theorems one ends up with:
BVP 2 Find u 2 U and  2 H 1 .B/ such that for all ıu 2 V and ı 2 H01 :
Z Z
Gu .u; ıu/ D ıF W P dV  Q
ıuTdA D0
B @B
Z
G .; ı/ D ı. fP  /
P  `2 rı  rdV D 0
B

subjected to the boundary condition:

PN D TN 8X 2 @B0
PN D 0 8X 2 0
u D uN 8X 2 @Bu0
r  N D 0 8X 2 @B 0

is fulfilled. Where the vector spaces U and V are defined as:

n  3 o n  3 o
U D uju 2 H 1 .B/ V D ıujıu 2 H 1 .B/ ; ıu D 0 on @Bu0 :

3 Discretisation of the Boundary Value Problem

with the XFEM

In this section the discretisation of the weak form of the boundary value problem
BVP 2 with the eXtended Finite Element method (XFEM) [4] is described.
The XFEM is used because the displacement field is discontinuous across the
crack surface and this method is capable to handle this accurately with just a
minor additional computational expense.Within the XFEM additional degrees of
freedom are introduced to the nodes in the vicinity of the crack. This enriches the
approximation space of the displacement by some knowledge of the characteristic
features of the solution. Along the crack surface the nodes become jump enriched
because of the presence of a displacement jump in the solution. In linear elastic
fracture mechanics, at the crack front the nodes are usually enriched by the basis
functions of the asymptotic solution. In Fig. 3 the enrichment scheme for a 2D model
88 S. Beese et al.

Fig. 3 Discretised boundary value problem (left) and local coordinate system at the crack tip
(right)

problem is given. It is obvious that the total number of degrees of freedom is just
increased moderately.
The main advantage of the XFEM is the fact, that in case of crack propagation
remeshing is not necessary due to the nearly mesh independent description of the
crack geometry and the displacement discontinuity. Usually cracks are implicitly
modelled by two mutually perpendicular level set functions '1 and '2 . While the
level set '1 represents the signed distance to the crack surface, the second level set
'2 describes the signed distance to the crack front. With this convention the jump
enrichment function is defined as:
(
1; for '1  0
H.X/ WD (20)
1; for '1 < 0 :

Elguedj et al. [9] derived an enrichment basis for the crack front field for the case
of small deformation elasto-plasticity with power law hardening based on the well
known HRR-field. Since there is no analytic solution for the near tip field in case of
finite deformations crystal plasticity we follow Legrain et al. [16] and Loehnert et al.
[17] and just employ one enrichment function f1 . The enrichment function reads:


f1 .r; / D r sin
n
; (21)
2

where r and  describes the crack geometry as seen in Fig. 3 on the right hand side.
The same holds for the three dimensional case where at every point of the crack
front such a local coordinate system is placed. With n chosen to be n  1 the derived
Modeling of Fracture in Polycrystalline Materials 89

strain and stress fields do not have a singularity if r ! 0. With this introduction the
approximated displacement field uh .X/ is computed via:
2 3
X
n X
m
uh .X/ D NI .X/ 4uI C H.X/aI C fj .X/bjI 5
ID1 jD1
3 2
uI (22)
6 a 7
Xn
 6
6
I 7
7 X T
n
D NI NI H NI f1 : : : NI fm 6 b1I 7 D NI u I :
6 7
ID1 4 : : : 5 ID1
bmI

Since the non-local void volume fraction  can develop independently on both
sides of the crack and is assumed to have a peak value at the crack front similar
enrichment functions are used:
2 3
Xn Xm

 h .X/ D NI .X/ 4I C H.X/aI C fj .X/bjI 5 D NI T I : (23)
ID1 jD1

By employing a Bubnov-Galerkin scheme the same Ansatz is used for the test
functions ıuh and ı h . The discretised versions of the weak forms of BVP 2 are
then computed by:

[
ne X
N Z XN Z
Gu .uh ; ıuh / D D ıuI rNI T P dVe  ıuI NI TQ dAe
Be @Be
eD1 ID1 „ ƒ‚ … ID1 „ ƒ‚ …
PuI FextI
(24)
[
ne X
N Z
G .h ; ıh / D D ıI NI T . fP  /
P  `2 rNI T  rdV :
Be
eD1 ID1 „ ƒ‚ …
PI

The nonlinear set of equations is treated by a standard Newton-Raphson proce-

dure. In each iteration the linear system:
    
Fext  Pu K uu K u
u
D (25)
P K u K 


is solved where K ij is the consistent linearisation of the weak form i with respect to
the degree of freedom j with i; j 2 fu; g.
Because of the complexity of real structures we discretise the domain by 10 node
tetrahedral elements with quadratic shape functions. The integration is performed
90 S. Beese et al.

Fig. 4 Integration scheme for an inclined crack (red) within a single 10 node tetrahedral element

by a subcell strategy where the element is split along the discontinuity. A standard
Gauss integration is then used in each subcell as can be seen in Fig. 4.

4 Quasistatic Crack Propagation with the XFEM

The task of propagating cracks is here divided into three steps. After a short
motivation of the crack propagation criterion the crack tracking algorithm is
presented. The section is concluded by embedding the crack propagation algorithm
into a standard load stepping scheme of finite element software.

4.1 Crack Propagation Criterion

Ductile fracture occurs if the material strength is exceeded. The material strength is
measured for a defect free mono crystal and is much higher than the strength of real
materials. For brittle materials in linear elastic fracture mechanics the concept of a
critical energy release rate Gc is used. This concept goes back to the works of Griffith
[13], where Gc is the required energy to form a new Surface
A. This criterion is
based on the assumption of linear elasticity and small deformations and cannot be
transferred to the current framework easily. Therefore we treat crack propagation
Modeling of Fracture in Polycrystalline Materials 91

as the coalescence of voids close to the crack tip. Instead of using an energy based
threshold we propose to use a critical void volume fraction fprop to propagate the
crack. In [20] and [3] Needleman et al. also used the void volume fraction to predict
the crack path. But their approach suffers from mesh depended crack paths and
numerical difficulties when f ! fF because the problem becomes ill posed. With
the here presented strategy we are able to treat cracks sharply and nearly mesh
independent. Following the damage concept the crack will propagate in the direction
P
where voids grow fastest. This direction can be identified to be proportional to r .
In the following sections an efficient algorithm is presented to propagate the crack.

4.2 Reconstruction of the Fracture Surface

The presented algorithms do not depend on the choice of crack propagation

direction, so this direction is substituted by the vector field v.X/. In principle there
are two ways of reconstructing the crack surface after determining the crack velocity
vectors v. The first is an explicit reconstruction of the surface by discretising the
crack front line and computing the velocity vectors v at each point. By joining the
tips of the vectors the new surface is determined. This approach was successfully
implemented in a three dimensional context by Fries et al. [11] or by Holl et al.
[14]. However, for inelastic materials this technique oscillatory behaviour of the
crack front line geometry and after a few propagation steps leads to this results in
unphysical crack geometries.
The second approach is to implicitly track the crack front, by either using the
level set approach as done by Moës and Gravouil in Moës et al. [19] and Gravouil et
al. [12] or by the global crack tracking algorithm suggested by Oliver and Huespe
in [22]. We will present both approaches and study their accuracy in the following
sections.

4.2.1 Level Set Update by Solving a Hamilton-Jacobi Equation

As already described in Sect. 3 the cracks are represented by two level set functions
'1 ; '2 . The evolution of such fields follows the Hamilton-Jacobi equation:

@'i
@t
C vN i  r'i D gN 8X 2 B 0
(26)
'i D 'Ni 8X 2 @B'0 ;

where vN i is the velocity of the i-th level set. Equation (26) holds for different
purposes, just by replacing the velocity vN i or the right hand side gN according to
Table 1. In Barths and Sethians contribution [2] they proposed an explicit method
based on a Petrov-Galerkin formulation for the solution of Eq. (26). The proposed
algorithm is designed for simplex elements. Since we want to deal with higher order
92 S. Beese et al.

Table 1 Velocities and right Purpose vN gN

hand sides for different
purposes '1 propagation v '1 0
'2 propagation v '2 0
sign.'1 /0 r'1
Reinitialisation jjr'1 jj
sign.'1 /0
sign.'1 /r'1
Reorthogonalisation jjr'1 jj
0

elements we derived an implicit algorithm by applying the Galerkin Least Square

(GLS) formulation of Hughes et al. [15] for advection dominated problems. The
weighted residual corresponding to Eq. (26) reads:
Z 
@'
G'G .'; ı'/ D ı' C ı' vN  r'  ı'gdV D 0 : (27)
B0 @t

To this weighted residual the least square approximation ı˘'LS of the residual R is
added:

X
N X
N
R WD NI 'PI C vN  rNI 'I  gN
ID1 ID1

[
ne Z
1 d
G'LS .'; ı'/ WD R2 dV
e
2 d' I B 0

ne Z
[ XN  X N
@'PI
D NI C vN  rNI  .NJ 'PJ C vN  rNJ 'J  gN / dV :
e B0 ID1 @'I JD1
(28)
The Parameter  in Eq. (28) is a stabilisation parameter which depends on the
element size h and whose value is chosen after [2]:

 2  2 ! 12
2 2jjvjj
N
D C : (29)

t h

The resulting weighted residual is in general a non linear equation of the form:

G'GLS D G'G C G'LS ; (30)

where just the Galerkin part is consistently linearised in a Newton-Raphson

scheme. Depending on the used time discretisation technique this equation can be
solved explicitly or implicitly. Because of its higher order accuracy and its easy
implementation we chose a Crank-Nicolson discretisation.
Modeling of Fracture in Polycrystalline Materials 93

Fig. 5 BVP for global crack

tracking

4.2.2 Global Crack Tracking Algorithm

The crack surface is described implicitly by the level set value '1 . Oliver and Huespe
[22] postulate the existence of two linearly independent vector fields T and S which
r'1
are perpendicular to the normal n D jjr' 1 jj
of the searched surface. This situation is
depicted in Fig. 5 for the two dimensional case. The problem is to find some '1 so
that:

T  r'1 D 0
(31)
S  r'1 D 0

is fulfilled. Multiplying both sides of the equation with T or S does not alter the
problem.

.T ˝ T/  r'1 D 0
(32)
.S ˝ S/  r'1 D 0

We now add up these two equation and take the divergence. This will lead to a
boundary value problem with homogeneous Neumann boundary conditions:
BVP 3 Find '1 so that:

Div . ŒT ˝ T C S ˝ S  r'1 / D Div .K  r'1 / D 0 8X 2 B 0

K  r'1 D q 8X 2 B 0
(33)
qN D0 8X 2 @Bq0
' D 'N 8X 2 @B'0

is fulfilled. The Dirichlet bound @B'0 is chosen to be the old crack path with negative
second level set.
94 S. Beese et al.

This simple heat equation like linear boundary value problem can be written in its
discretised weak form:

[
ne X N Z
N X
G'GCT .'; ı'/ D ı'I rNI KrNJ dV'J D 0 : (34)
e ID1 JD1 B0

It should be remarked that the tensor K shows up a rank deficiency of 1. This

can lead to an ill conditioned equation system. Therefore a small multiple " of the
zero eigen space is added to K. The influence of this stabilisation parameter on the
solution is later studied for an example problem. For T and S an adequate choice
needs to be made. Since the crack propagation vector v is tangential to the crack
surface, T is chosen to be equal to v. S is assumed to be the crossproduct between
the eigenvector of the smallest principle strain E1 with v:

TDv
(35)
S D v  E1 :

For the second level set '2 this update technique is not possible because the need for
an adequate Dirichlet boundary condition. Therefore the Petrov-Galerkin scheme is
used to update the second level set.

4.2.3 Accuracy of the Two Methods

In this section we study the accuracy of the describes models for propagating the
crack surface in a simple example similar to the study of Duflot [8]. In a block of
size 1  1  0:25 a horizontal crack is placed. The block is discretised by 9600
tetrahedral elements with quadratic shape functions and the crack velocity field v is
prescribed to enforce a crack path inclined by the angle ˛. In Fig. 6 on the left hand
side the setup is displayed.

Fig. 6 Setup of the test problem (left) and Deviation of prescribed crack path with respect to
stabilisation parameter for ˛ D 30ı (right)
Modeling of Fracture in Polycrystalline Materials 95

Fig. 7 Deviation of prescribed crack path with respect to the propagation angle

In the first test the stabilisation parameter " of the global crack tracking algorithm
is changed and the root mean square error:
v
u n
u1 X
RMS D t .'O i  '/2 (36)
n iD1

of the resulting crack with respect to the expected crack path is evaluated. Here 'Oi
is the signed distance of the analysed method which is compared with the exact one
'. As can be seen in Fig. 6 on the right hand side the error becomes minimal for a
stabilisation parameter of the order " 103 . For smaller parameters the deviation
remains nearly constant. In the ongoing examples we will therefore use " D 103 .
In a second example the angle ˛ is altered in the range 0ı  ˛  70ı . The
level set update is performed by using the implicit Petrov-Galerkin scheme and we
compare these results with those obtained using the global crack tracking algorithm.
In Fig. 7 the root mean square error of both methods with respect to the propagation
angle is displayed. The global tracking algorithm performs better in the entire range.
Over a broad range 0ı  ˛ . 53ı the error of this algorithm is one order of
magnitude smaller than the error of the Petrov-Galerkin scheme. For larger angles
the discrepancy is even up to two orders of magnitude. We therefore recommend to
use the global tracking algorithm for updating the first level set.

4.3 Crack Propagation Algorithm

In this section a simple displacement controlled algorithm for quasi static crack
propagation is presented. After the level set update the crack propagation algorithm
continues in updating the enrichment scheme according to the new level sets. By
96 S. Beese et al.

Fig. 8 Scheme for propagating cracks

doing this the mesh will not be changed but the nodal degrees of freedom close
to the crack front will alter. Also the integration point scheme will change to take
the new crack geometry into account. It is necessary to project the displacement
field described with the old enrichment pattern onto the description using the new
enrichment pattern. Similarly it is also necessary to project the history variables
from the old integration points onto the new integration points in those parts of
the domain where the position of the integration points change. For both tasks a
linear least square problem is solved. Since now neither the yield condition nor
the mechanical equilibrium is fulfilled an equilibrium iteration is performed. After
equilibrium is achieved the test for the critical void volume fraction is performed
again. This procedure is repeated as long as the critical volume fraction is exceeded.
A schematic representation of the algorithm is given in Fig. 8.
Remark 1 It has to be remarked that the update of the history variables and the
projection of the degrees of freedom is still the weak part the proposed algorithm.
Often the trial values for constitutive material routine are to far away from the
solution after a crack propagation step. To solve this issue a more advanced
projection technique for the displacement field and the history variables and a new
strategy for integrating the constitutive equations is under development. For simple
examples and small crack propagation increments the algorithm is stable.

5 Numerical Examples

In this section three examples for the described methods are presented. After
a verification of the non local GTN formulation two simple crack propagation
examples are considered.
Modeling of Fracture in Polycrystalline Materials 97

5.1 Verification of the Non-Local Damage Formulation

In Fig. 9 a polycrystalline block with given dimension is shown. The block is a

composition of Voronoi cells generated with the software Neper [25] and each cell
is representing a grain of the material. The orientation of the grains is distributed
randomly and displayed as different colour in Fig. 9. The crystal lattice is body
centred cubic with 12 slip systems according to Nemat-Nasser et al. [21] and
the material data can be found in Table 2 which correspond to a DC04 steel.
A displacement controlled tensile test without constraining the lateral stretch is
performed and the external force and displacement is recorded. In Fig. 10 the non-
local void volume fraction is shown. In the undeformed configuration on the left

Fig. 9 Setup for mesh dependency test

Table 2 Material and model Parameter Value Parameter Value

parameter
 80;000 MPa q1 1.5
 160;000 MPa q2 1
y0 80 MPa ` 0.03 mm
h1 D h2 1 fc 0.07
H 0 n 0.2
m 3 An 0.5
P0 0.005
fF 0.3
98 S. Beese et al.

Fig. 10 Void volume fraction in undeformed state (left) and in deformed state (right)

Fig. 11 Force displacement curve of a tensile test with different discretisations

hand side one can see that there are grains which promote the damage evolution
for the given load (red and green) and a band of grains (blue) which has a higher
remaining strength. Furthermore the distribution of the void volume fraction is very
smooth which reflects its non-local nature. This test is repeated for 3 different
discretisations where the element size is decreasing. The results are displayed in
Fig. 11 and show a convergence behaviour to a certain load deflection path. The
Modeling of Fracture in Polycrystalline Materials 99

presented formulation is clearly mesh independent. The validation of the model is

ongoing research where a real material testing will be performed for comparison.

5.2 Crack Propagation in Crystals

The next test example considers a homogeneous specimen with a horizontal crack
and dimensions shown in Fig. 12 on the left hand side. The block is treated as single
crystal and a vertical displacement is prescribed on the top surface. In order to
not constrain the lateral contraction symmetric boundary conditions are applied.
The example is uniformly discretised with 2592 tetrahedral elements with quadratic
shape functions and the material and model parameters can be found in Table 2. For
this setup the crack has to propagate horizontally which is in fact the case as can be
seen in Fig. 12 on the right hand side.
By crossing a grain boundary the direction of crack propagation should change
according to the lattice orientation. Since the crack is driven by the volume void
distribution and the strain state the presented model should be able to reflect this
behaviour implicitly. To study this effect the block of the previous examples is now
composed of two grains with different orientation as can be seen in Fig. 13. The
discretisation and the boundary conditions are the same as before. On the right hand
side of Fig. 13 the simulated crack path is shown. Close to the material interface the
direction of the crack propagation changes. The crack grows into the second grain
and proceeds its propagation with a different direction.

Fig. 12 Setup for crystalline crack growth (left) and simulated crack path for tensile test (right)
100 S. Beese et al.

Fig. 13 Setup for inter crystalline crack growth (left) and simulated crack path for tensile test
(right)

6 Conclusions

In this work the entire process of formulating the boundary value problem, deriving
the constitutive equation and the numerical treatment of fracture in polycrystals
is derived. Therefore a non-local extension of the Gurson-Tvergaard-Needleman
model in finite deformations crystal plasticity is introduced. In mesh dependence
study we show that the presented model does not suffer from mesh dependent
localisation.
To represent cracks sharply a transition from damage to fracture is described.
The cracks are modelled with the extended finite element method and two update
techniques for the level sets are investigated. Here the global crack tracking
scheme is recommended for updating the first level set because of its simplicity
and accuracy. For re-initialisation purposes and for updating the second level set
an implicit Petrov-Galerkin scheme is derived. Finally a displacement controlled
crack propagation algorithm is outlined and some open questions are remarked.
The problem of mapping history values after a crack propagation step and the
corresponding issue of updating the constitutive law is addressed. In three examples
the features of the presented model are shown.

Acknowledgements This work is supported by the German Research Foundation (DFG) under
the Transregional Collaborative Research Center SFB/TR73: “Manufacturing of Complex Func-
tional Components with Variants by Using a New Sheet Metal Forming Process—Sheet-Bulk
Metal Forming”.

References

1. Asaro, R.J.: Crystal plasticity. J. Appl. Mech. 50, 921–934 (1983)

2. Barth, T.J., Sethian, J.A.: Numerical schemes for the Hamilton-Jacobi and level set equations
on triangulated domains. J. Comput. Phys. 145, 1–40 (1998)
Modeling of Fracture in Polycrystalline Materials 101

3. Becker, R., Needleman, A., Suresh, S., Tvergaard, V., Vasudevan, A.: An analysis of ductile
failure by grain boundary void growth. Acta Metall. 37, 99–120 (1989)
4. Belytschko, T., Black, T.: Elastic crack growth in finite elements with minimal remeshing. Int.
J. Numer. Methods Eng. 45, 601–620 (1999)
5. Belytschko, T., Moës, N., Usui, S., Parimi, C.: Arbitrary discontinuities in finite elements. Int.
J. Numer. Methods Eng. 50, 993–1013 (2001)
6. Bertram, A.: An alternative approach to finite plasticity based on material isomorphisms. Int.
J. Plast. 15, 353–374 (1999)
7. Bertram, A.: Elasticity and Plasticity of Large Deformations. Springer, Berlin (2008)
8. Duflot, M.: A study of the representation of cracks with level sets. Int. J. Numer. Methods Eng.
70, 1261–1302 (2007)
9. Elguedj, T., Gravouil, A., Combescure, A.: Appropriate extended functions for X-FEM
simulation of plastic fracture mechanics. Comput. Methods Appl. Mech. Eng. 195, 501–515
(2006)
10. Eringen, A.C., Suhubi, E.: Nonlinear theory of simple micro-elastic solids i. Int. J. Eng. Sci. 2,
189–203 (1964)
11. Fries, T.-P., Baydoun, M.: Crack propagation with the extended finite element method and a
hybrid explicit implicit crack description. Int. J. Numer. Methods Eng. 89, 1527–1558 (2012)
12. Gravouil, A., Moës, N., Belytschko, T.: Non-planar 3d crack growth by the extended finite
element and level sets part ii: level set update. Int. J. Numer. Methods Eng. 53, 2569–2586
(2002)
13. Griffith, A.A.: The phenomena of rupture and flow in solids. Philos. Trans. R. Soc. Lond. Ser.
A, Contain. Pap. Math. Phys. Character 221, 163–198 (1921)
14. Holl, M., Rogge, T., Loehnert, S., Wriggers, P., Rolfes, R.: 3d multiscale crack propagation
using the xfem applied to a gas turbine blade. Comput. Mech. 53, 173–188 (2014)
15. Hughes, T.J., Franca, L.P., Hulbert, G.M.: A new finite element formulation for computational
fluid dynamics. VIII: the Galerkin/least-squares method for advective-diffusive equations.
Comput. Methods Appl. Mech. Eng. 73, 173–189 (1989)
16. Legrain, G., Moës, N., Verron, E.: Stress analysis around crack tips in finite strain problems
using the extended finite element method. Int. J. Numer. Methods Eng. 63, 290–314 (2005)
17. Loehnert, S., Mueller-Hoeppe, D., Wriggers, P.: 3d corrected xfem approach and extension to
finite deformation theory. Int. J. Numer. Methods Eng. 86, 431–452 (2011)
18. Moës, N., Dolbow, J., Belytschko, T.: A finite element method for crack growth without
remeshing. Int. J. Numer. Methods Eng. 46, 131–150 (1999)
19. Moës, N., Gravouil, A., Belytschko, T.: Non-planar 3d crack growth by the extended finite
element and level sets part I: mechanical model. Int. J. Numer. Methods Eng. 53, 2549–2568
(2002)
20. Needleman, A., Tvergaard, V.: An analysis of ductile rupture modes at a crack tip. J. Mech.
Phys. Solids 35, 151–183 (1987)
21. Nemat-Nasser, S., Okinaka, T., Ni, L.: A physically-based constitutive model for BCC crystals
with application to polycrystalline tantalum. J. Mech. Phys. Solids 46, 1009–1038 (1998)
22. Oliver, J., Huespe, A.: Continuum approach to material failure in strong discontinuity settings.
Comput. Methods Appl. Mech. Eng. 193, 3195–3220 (2004); Computational failure mechanics
23. Peerlings, R.H.J., de Borst, R., Brekelmans, W.A.M., de Vree, J.H.P.: Gradient enhanced
damage for quasi-brittle materials. Int. J. Numer. Methods Eng. 39, 3391–3403 (1996)
24. Pietruszczak, S., Mróz, Z.: Finite element analysis of deformation of strain-softening materials.
Int. J. Numer. Methods Eng. 17, 327–334 (1981)
25. Quey, R., Dawson, P., Barbe, F.: Large-scale 3d random polycrystals for the finite element
method: generation, meshing and remeshing. Comput. Methods Appl. Mech. Eng. 200, 1729–
1745 (2011)
26. Reusch, F., Svendsen, B., Klingbeil, D.: Local and non-local Gurson-based ductile damage and
failure modelling at large deformation. Eur. J. Mech. A Solids 22, 779–792 (2003)
102 S. Beese et al.

27. Steinmann, P., Stein, E.: On the numerical treatment and analysis of finite deformation ductile
single crystal plasticity. Comput. Methods Appl. Mech. Eng. 129, 235–254 (1996)
28. Svendsen, B.: A thermodynamic formulation of finite-deformation elastoplasticity with hard-
ening based on the concept of material isomorphism. Int. J. Plast. 14, 473–488 (1998)
29. Tvergaard, V., Needleman, A.: Analysis of the cup-cone fracture in a round tensile bar. Acta
Metall. 32, 157–169 (1984)
eXtended Hybridizable Discontinuous Galerkin
(X-HDG) for Void and Bimaterial Problems

Ceren Gürkan, Esther Sala-Lardies, Martin Kronbichler, and Sonia

Fernández-Méndez

Abstract A strategy for the Hybridizable Discontinuous Galerkin (HDG) solu-

tion of problems with voids, inclusions, free surfaces, and material interfaces is
proposed. It is based on an eXtended Finite Element (X-FEM) philosophy with a
level-set description of interfaces where the computational mesh is not required to
fit the interface (i.e. the boundary). This reduces the cost of mesh generation and,
in particular, avoids continuous remeshing for evolving interfaces. Differently to
previous proposals for the HDG solution with unfitting meshes, the computational
mesh covers the domain in our approach, avoiding extrapolations and ensuring the
robustness of the method. The local problem in elements not cut by the interface
and the global problem are discretized as usual in HDG. A modified local problem
is considered for elements cut by the interface. At every cut element, an auxiliary
trace variable on the boundary is introduced, which is eliminated afterwards using
interface conditions, keeping the original unknowns and the structure of the local
problem solver. The solution is enriched with Heaviside functions in case of
bimaterial problems; in case of problems with voids, inclusions, or free surfaces
no such enrichment is required. Numerical experiments demonstrate how X-HDG
keeps the optimal convergence, superconvergence, and accuracy of HDG with no
need of adapting the computational mesh to the interface boundary.

1 Introduction

Among state-of-the-art advanced discretization techniques high-order discontinuous

Galerkin methods are nowadays very popular in computational fluid dynamics. DG
methods inherit the advantages of finite volume methods such as stability through
numerical fluxes and local conservation but they allow for the use of high-order

C. Gürkan () • E. Sala-Lardies • S. Fernández-Méndez

Universitat Politècnica de Catalunya, Barcelona, Spain
e-mail: [email protected]; [email protected]; [email protected]
M. Kronbichler
Technische Universität München, München, Germany
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 103

G. Ventura, E. Benvenuti (eds.), Advances in Discretization Methods,
SEMA SIMAI Springer Series 12, DOI 10.1007/978-3-319-41246-7_5
104 C. Gürkan et al.

approximations with a straightforward implementation of p-adaptivity. Among DG

methods the hybridizable discontinuous Galerkin method (HDG) [4, 5], has been
successfully applied to all kinds of problems, specifically in computational fluid
dynamics (CFD), see, for instance, [6, 19, 20] for its application to the Stokes and
Navier–Stokes equations, or [12, 15, 17] for an efficiency comparison to continuous
finite elements (CFE) in the context of elliptic problems and wave propagation.
HDG inherits all the advantages of high-order DG methods [2, 14, 18, 21],
such as local conservation of quantities of interest, intrinsic stabilization thanks
to a proper definition of numerical fluxes at element boundaries, suitability for
code vectorization and parallel computation, and straight-forward realization of
adaptivity. However, two main characteristics allow HDG to outperform other
DG methods for problems involving self-adjoint operators, namely hybridization
and superconvergence. The hybridization process allows reducing the degrees of
freedom in the final linear system to the nodes in the element faces (sides in 2D),
similarly to static condensation in the context of high-order continuous elements, see
for instance [12]. On other hand, HDG is based on a particular mixed formulation.
When approximating both the primal unknown and its derivatives with polynomials
of the same degree k, convergence rates of order k C 1 in the L2 norm are observed
for both variables. Therefore, a simple element-by-element postprocessing of the
derivatives leads to a superconvergent approximation of the primal variables, with
convergence of order k C 2 in the L2 norm. The superconvergent solution can be
used to define a simple and efficient error estimator, which then can be used to
drive an adaptivity process [11, 13]. When it comes to the efficient solution of
problems with moving boundaries and interfaces there is still work to be done.
In [23], HDG has been applied to the solution of Stokes interface problems,
but always considering computational meshes fitting the interface. In[16], HDG
has been used for the solution of elliptic material interface problems where the
material interface is defined by the computational mesh. There, the mesh needs
to be adapted to properly describe the interface geometry, requiring continuous
remeshing in the case of evolving interfaces. Even though a methodology for the
solution of elliptic problems with meshes not fitting the boundary has been proposed
in [3, 8], extrapolating the solution beyond the elements causes some restrictive
requirements on the distance from the computational mesh to the boundary which
as well limits the practical applicability of the proposal. An alternative strategy
for the HDG solution of interface problems, based on an extended finite element
(X-FEM) philosophy, is proposed here. The extended finite element method is
a clever strategy to treat, for instance, the discontinuities arising at interfaces.
The suitability of the X-FEM strategy for the solution of problems with moving
boundaries and interfaces in the context of continuous finite elements along with
high-order approximation[1, 9] is well-established, see for instance, the overview
paper [10] and the references therein. In X-FEM, interfaces are usually represented
as the zero-level set of a signed distance function, and the original computational
mesh and unknown structure are maintained, with a proper numerical integration at
X-HDG for Void and Bimaterial Problems 105

elements and faces cut by the interface [22]. In the case of material interfaces or
cracks, the solution is enriched to represent weak or strong discontinuities across
interfaces.
Motivated by the demonstrated efficiency of HDG and X-FEM, this works aims
to combine these two frameworks for efficiently solving problems with moving
boundaries and interfaces described by level sets. The method, labeled eXtended
Hybridizable Discontinuous Galerkin (X-HDG), proposes the efficient solution of
problems including voids, inclusions, free surfaces, and material interfaces. The
solution is enriched with Heaviside functions in the case of material interfaces
and, in the case of weak discontinuities, continuity is weakly imposed, emulating
the imposition of continuity across element boundaries in standard HDG. X-
HDG inherits the advantages of X-FEM methods i.e., the computational mesh
is not required to adapt to the interface. This simplifies and reduces the cost of
mesh generation while keeping the computational efficiency, stability, accuracy and
optimal convergence of HDG. Differently to [3, 8], the computational mesh always
covers the domain and, therefore, no extrapolations are required, leading to a more
robust method. Hence, X-HDG keeps the superconvergence and stability properties
of standard HDG, but in accordance with an X-FEM philosophy.
Section 2 presents the detailed X-HDG formulation and discretization for void
problems with Neumann boundary conditions on the interface (i.e. on the boundary
not fitted by the computational mesh). The local problem at elements not cut
by the interface and the global problem are discretized as usual in HDG. The
discretization of the local problem for cut elements is developed in Sect. 2.2. At
every cut element, an auxiliary trace variable on the boundary is introduced, which
is eliminated afterwards using the boundary conditions on the interface, keeping
the original unknowns. The X-HDG formulation is then extended to problems
with Dirichlet interfaces. The application to problems involving both Dirichlet and
Neumann interfaces is straightforward. Section 3 explains the X-HDG formulation
for bimaterial problems mostly referring to the slight differences in the formulation
as compared to standard HDG. Here the solution space is enriched with Heaviside
functions to introduce the discontinuity within the element for those elements cut
by the interface. Elements not cut by the interface are treated with a standard HDG
approach. Finally, numerical examples in Sect. 4 demonstrate the applicability of
the method and that X-HDG keeps the accuracy, optimal convergence, and super-
convergence of HDG. The application of the developed formulation to problems
with evolving boundaries avoids the continuous remeshing to fit the interface, with
the corresponding saving in computational time. In addition, no loss of accuracy
due to the projection of quantities from one mesh to another occurs in this setting.
The additional cost of X-HDG in front of HDG is mainly due to the modification
of the numerical quadrature, since elemental computations for cut elements are
substantially more expensive than those for standard elements. However, it is worth
noting that in practical applications the ratio of cut elements to standard elements is
small, and the overcost is negligible compared to the cost of remeshing.
106 C. Gürkan et al.

2 X-HDG Formulation for Void Problems

Let ˝  Rd be a bounded domain with an interior boundary I (also referred to

as interface) and an exterior boundary @˝ ext WD @˝nI. The following problem is
considered,

r  .r u/ D f in ˝;
r u  n D g on I; (1)
u D uD on @˝ ext ;

where u is the solution,  is a material coefficient, f is a given source term, uD are

prescribed values on the exterior boundary, and g is a prescribed flux on the interior
boundary, i.e. the voids boundary. Neumann boundary conditions are considered
on the interior boundary I. The implementation of Dirichlet boundary conditions
on I is straight-forward, see Remark 2. Here, for simplicity, Dirichlet boundary
conditions are considered on the exterior boundary @˝ ext ; other boundary conditions
at @˝ ext do not add any difficulty, since they are implemented as in standard HDG.
The domain ˝ is assumed to be covered by a finite element mesh with nel
disjoint elements Ki , such that
"n #
[
nel [ el

˝ Ki; Ki \ Kj D ; for i ¤ j, @˝ ext

@ Ki :
iD1 iD1

Note that the mesh fits the exterior boundary @˝ ext , but some elements may be cut
by the interior boundary I, see Fig. 1. The union of all nfc faces i (sides for 2D)
intersecting the domain ˝ is denoted by

[
nel
  [nfc
 
WD @Ki \ ˝ D f \ ˝ :
iD1 f D1

The discontinuous setting of (1) consists of some element-by-element equations

and some global ones. The local element-by-element problems correspond to the
statement of the PDE in (1) with essential boundary conditions in each element Ki ,

Fig. 1 Two examples

of domain with a void: a
circular void boundary and a
straight interface, I black.
The mesh covers the domain
˝ (gray) and fits the exterior
boundary @˝ ext D @˝nI
X-HDG for Void and Bimaterial Problems 107

that is,
9
r  q D f in Ki =
q C r u D 0 in Ki if I \ Ki D ;; (2a)
;
u D uO on @Ki
9
r  q D f in ˝i >
>
=
q C r u D 0 in ˝i
if I \ Ki ¤ ; (2b)
q  n D g on Ii >
>
;
u D uO on @˝i nIi

for i D 1; : : : ; nel , where, for cut elements,

˝i WD ˝ \ Ki ; Ii WD I \ Ki : (3)

In (2), two new variables are introduced. The quantity q represents the flux of u in
order to formulate the PDE as two first order PDEs, and uO corresponds to the trace
of u at the mesh faces. The local problems have been particularized for elements cut
by the interior boundary (2b) and standard elements (2a).
Note that, given the trace uO , which is single valued on the mesh skeleton , the
local problems (2) can be solved in each element to determine the solution u and the
flux q. Thus, the problem now reduces to determine the trace uO , with the so-called
conservativity conditions (also known as global equations), that is, the continuity of
the flux across element boundaries

q  n D 0 on n@˝ ext ; (4)

and the boundary condition, equivalent to the exterior boundary condition in (1),

uO D uD on @˝ ext : (5)

The jump  operator is defined at a face f as

} D }L. f / C }R. f / on f ;

where R. f / and L. f / are numbers of the left and right elements sharing the face, that
is, f D KL. f / \ KR. f / , and the subindex }i denotes the value of function } from
element Ki . In particular, q  n D qL. f /  nL. f / C qR. f /  nR. f / D .qL. f /  qR. f / /  nL. f / .
It is important noting that the continuity of the solution u across is imposed
by the Dirichlet boundary condition in the local problems (2) and the fact that uO is
single valued on .
The discretization of the conservativity condition (4) and the local problems (2),
with the boundary condition (5), leads to the X-HDG formulation.
108 C. Gürkan et al.

The next sections present the details of the X-HDG formulation, stating the
discretization of the local problems for standard and cut elements and the discretiza-
tion of the conservativity condition (4). The local problem at elements not cut by
the interface (2a) and the global problem are discretized as usual in HDG [4, 5],
as recalled in Sects. 2.1 and 2.3. The discretization of the local problem for cut
elements (2b) is developed in Sect. 2.2.
To simplify the presentation, the same notation is used for the numerical
approximation, belonging to the usual finite dimensional spaces of piecewise
polynomials on elements, and the exact solution, that is u, q and uO .

2.1 Local Problem for Standard Elements

This section recalls the standard HDG local problem on an element Ki not cut by
the interface. It corresponds to the discretization of (2a), that is: Given uO 2 h , find
u 2 Pk .Ki /, q 2 ŒPk .Ki / d where Pk denotes the space of polynomials of degree less
or equal to k, such that
Z Z Z
vr  q dV C v.u  uO / dS D vf dV 8v 2 Pk .Ki /;
Z Ki Z @Ki Z Ki
(6)
q  w dV  ur  w dV C  uO w  n dS D 0 8w 2 ŒPk .Ki / d
Ki Ki @Ki

The first equation in (6) can be derived from the first equation in (2a) by applying
integration by parts, replacing the flux by the numerical flux

qO WD q C .u  uO /n; (7)

and undoing the integration by parts. The second equation is obtained from the weak
form of the second equation in (2a), applying integration by parts and replacing the
boundary condition u D uO on the element boundary.
Remark 1 The parameter , appearing in the definition of the numerical flux (7),
is a non-negative stabilization parameter usually taken of order O./. For each
element, it may be taken as a positive constant on all faces, or positive on one
arbitrary face and zero on the rest (single face). Both options lead to stable and
optimally convergent solutions, with superconvergent post-processed solutions. See
for instance [4, 11] for details on the influence of this parameter on the solution
behavior.
The discretization of the local problem (6) leads to a system of equations of the
form
( K i
Auui u C AKuqi qi C AKuOui ƒi D fKu i ;
(8)
AKqui ui C AKqqi qi C AKqOui ƒi D 0;
X-HDG for Void and Bimaterial Problems 109

where ui and qi are the vectors of nodal values of u and q in element Ki , and ƒi is
the vector of nodal values of uO on the n faces of the element (n D 3 for triangles and
n D 4 for tetrahedra). That is,
3 2
uO Fi1
6 7
ƒi WD 4 ::: 5 ; (9)
O Fin
u

where uO f denotes the nodal values of uO on face f , and Fij is the number of the j-th
face of element Ki , Note that the subindices in the A matrices refer to the space for
the weighting function and the test function, respectively.
System (8) can be solved for ui and qi in each element, obtaining the so-called
local solver in the element Ki

ui D UKi ƒi C fKUi ; qi D QKi ƒi C fKQi ; (10)

with
" # " # " # " #
UK i AKuOui fKUi fKu i
1 1
D A ; DA (11)
Q Ki AKqOui fKQi 0

and
" #
AKuui AKuqi
AD :
AKqui AKqqi

That is, for each element, the elemental values of the solution, ui and qi , can be
explicitly expressed in terms of the trace on its faces, ƒi .

2.2 Local Problem for a Cut Element

The X-HDG local problem at an element Ki cut by the interior boundary corresponds
to the discretization of (2b), that is: Given uO 2 h , find u 2 Pk .˝i /, q 2 ŒPk .˝i / d
such that
Z Z Z Z
vr  q dV C v.u  uO / dS C v.u  uQ / dS D
i
vf dV
Z ˝i Z @˝i nIi Z Ii
Z ˝i
(12)
q  w dV  ur  w dV C  uO w  n dS C  uQ w  n dS D 0
i
˝i ˝i @˝i nIi Ii
110 C. Gürkan et al.

for all v 2 Pk .˝i / and w 2 ŒPk .˝i / d , where ˝i D ˝ \ Ki , and uQ i is a new trace
variable approximating the trace of the solution on the interface Ii D I \ Ki , see
Fig. 2. Compared to the weak form for standard elements (6), the X-HDG weak
form for a cut element has two additional terms corresponding to integrals along the
interface Ii , involving the new trace variable uQ i .
The discretization of the local problem (12) leads to the system of equations
  i ˝i i Ii i
A˝ Ii ˝i i Q D f˝
uu C Auu u CAuq q C AuOu ƒ C AuQui u
i
u ;
i

˝i i Ii i
; (13)
A˝ ˝i i Q D 0;
qu u CAqq q C AqOu ƒ C AqQui u
i i

similar to (8), but with three new matrices corresponding to integrals on the
Neumann boundary (marked with the superindex Ii ), and the nodal values for
the new trace variable, uQ i . The local problem is closed by imposing the Neumann
boundary condition on Ii , i.e.,

qO  n D g on Ii :

Using the expression of the numerical flux qO defined in (7), i.e. qO  n D q  n C

.u  uQ i /, the weak form of the Neumann condition in I i is to find uQ i 2 Pk .Ii / such
that
Z Z Z
 
vq
Q  n dS C  vQ u  uQ dS D
i
vg
Q dS 8vQ 2 Pk .Ii /: (14)
Ii Ii Ii

The discretization of the Neumann boundary condition leads to a discrete

equation of the form

AIuQiiu ui C AIuQiiq qi C AIuQiiuQi uQ i D g;

that allows expressing the new trace values uQ i in terms of elemental values

uQ i D Tiu ui C Tiq qi  ti (15)

Fig. 2 X-HDG discretization

in an element cut by the
interior Neumann boundary:
gray for elemental variables,
u and q, black nodes for trace
variable uO and white nodes for
the trace on the Neumann
boundary uQi
X-HDG for Void and Bimaterial Problems 111

with

Tiu D ŒAIuQiiuQi 1 AIuQiiu ; Tiq D ŒAIuQiiuQi 1 AIuQiiq ; ti D ŒAIuQiiuQi 1 g:

Replacing (15) in (13) leads to the final discrete local problem

h i h i
Ii i Ii i ˝i i Ii i
A˝ Ii ˝i ˝i
uu C Auu C AuQui Tu u C Auq C AuQui Tq q C AuOu ƒ D fu C AuQui t
i i i
h i h i
Ii i Ii i ˝i i Ii i
A˝ i i ˝i
qu C AqQui Tu u C Aqq C AqQui Tq q C AqOu ƒ D AqQui t
i

Now, similarly to (8), this system can be solved for ui and qi , obtaining the local
solver in the cut element Ki , i.e., Eq. (10) with
" # " # " # " #
UK i A˝ i
fKUi f˝i Ii i
u C AuQui t
1 uOu 1
D A ; DA (16)
Q Ki A˝
qOu
i
fKQi AIqQui i ti

and
2h ih i3
Ii i Ii i
A˝ Ii
uu C Auu C AuQui Tu
i A˝
uq C AuQui Tq
i

AD4 h i h i5
Ii i Ii i
A˝
qu C AqQui Tu
i A˝
qq C AqQui Tq
i

Note that the structure of the local solver is exactly the same as for non-cut
elements (10), as the internal trace variable uQ i has been isolated and it is not an
unknown of the problem anymore.
Remark 2 In case of Dirichlet boundary conditions imposed over the interface I,
the X-HDG formulation is straightforward. Once the system in (13) is obtained, the
terms AIuQui i uQ i and AIqQui i uQ i are already known so they can simply be moved to the right
hand side of the system to obtain a local solver that looks like:
  i ˝i i
A˝ Ii ˝i i
uu C Auu u CAuq q C AuOu ƒ D f
i ˝i
C fIu i
˝i i
(17)
A˝ ˝i i Ii
qu u CAqq q C AqOu ƒ D fq
i i

where AIuQui i uQ i D fIu i and AIqQui i uQ i D fIq i . Analogous to the Neumann case, (17) can
be solved for ui and qi to obtain the local solver for each element and finally the
problem can be closed by solving global equation defined in detail in Sect. 2.3. Note
that interface problems where some interfaces are subject to Neumann conditions
and others are subject to Dirichlet conditions can be naturally covered by selecting
the appropriate strategy element-by-element.
112 C. Gürkan et al.

2.3 Global Problem

The local problem, both in a standard element or in a cut element, leads to the
local solver (10), that computes of the solution in the element, u and q, in terms of
the trace values at its boundary, uO . Thus, now the problem is reduced to determine
the trace nodal values fuO f gnf D1
fc
on the mesh skeleton . For this purpose the so-
called global problem is stated, which corresponds to the discretization of the
conservativity condition on (4).
Replacing q by the numerical flux (7), the weak form for the trace variable is:
Find uO 2 h such that uO D uD on @˝ ext and
Z Z
vq
O  n dS C 2 vO .fug  uO / dS D 0 8vO 2 h ;

where fg is the mean operator on the faces,

1 
f}g D }L. f / C }R. f / on f :
2
As usual in HDG, the discretization of this equation for every face f leads to an
equation of the form
f ;L f ;L f ;R f ;R
AuO u uL. f / C AuO q qL. f / C AuO u uR. f / . f / C AuO q qR. f / C AuO uO uO f D 0:
f
(18)

Replacing the local solver (10), for the elements KL. f / and KR. f / , in (18) for every
face f , leads to a system of equations involving only the trace variables fuO f gnf D1fc
.
As usual in an HDG code, the implementation of the method involves a loop
over elements. For each element, the matrices and vectors for the local solver (10)
are computed, and the contribution to the Eq. (18) is assembled for each face of
the element. Once the system is assembled for all elements, and Dirichlet boundary
conditions (5) are imposed, the linear system can be solved. Then, given the trace
variables fuO f gnf D1
fc
, the interior values ui and qi , can be computed for each element
using (10). Note that X-HDG keeps the structure of a standard HDG code. The only
difference is the modified local problem on cut elements (13), and the corresponding
matrices in the local solver (16).
Remark 3 An element-by-element postprocessing procedure allows to compute an
X-HDG superconvergent solution. Similarly to standard HDG the superconvergent
solution can be computed in every element Ki as the solution of: Find u 2 PkC1 .˝i /
such that
Z Z
r u  r v dV D  q  r v dV 8v 2 PkC1 .˝i /;
˝i Z ˝i Z
u dV D u dV;
˝i ˝i
X-HDG for Void and Bimaterial Problems 113

with ˝i D Ki for standard elements, and ˝i D Ki \˝ for cut elements. The solution
of this element-by-element computation, u , converges with order k C 2 in the L2
norm. See [4, 7] for details and other possible computations of a superconvergent
solution.

3 X-HDG Formulation for Bimaterial Problems

Let ˝  Rd be a bounded domain divided into two disjoint subdomains, ˝ D

˝1 [ ˝2 ; ˝1 \ ˝2 D ;, with an interface I D ˝1 \ ˝2 . The following bimaterial
problem is considered,

r  .r u/ D f in ˝1 [ ˝2 ;
un D 0 on I;
(19)
r u  n D 0 on I;
u D uD on @˝;

where u is the solution,  is a material coefficient with discontinuous definition

across the interface (that is,  D i in ˝i for i D 1; 2), f is a given source term and
uD are prescribed values at the exterior boundary. The jump  operator is defined at
an interface (material interface or, later, faces between elements), using values from
the domains to the left and right of the interface, } D }1 C }2 ; always involving
the normal vector on the interface.
The partition of ˝ in elements Ki induces a new problem equivalent to (19),
with some element-by-element equations and some global ones. The local problem
at uncut element is the one of standard HDG i.e., equivalent to Eq. (2a), and its
discretization can be found in Sect. 2.1. For elements cut by the interface the local
problem is
9
r  q D f in Ki nI >>
>
q C r u D 0 in Ki nI > >
=
un D 0 on I \ Ki if I \ Ki ¤ ;; (20)
>
q  n D 0 on I \ Ki >
>
>
>
;
u D uO on @Ki

and its X-HDG discretization is stated in Sect. 3.1.

The problem is closed with the conservativity condition (4), corresponding to
the imposition of continuity across element boundaries, and the boundary condition
uO D uD on @˝. Its weak form is the same as for standard HDG, see Sect. 2.1.
114 C. Gürkan et al.

The following discrete spaces for elemental variables, u and q, and for trace
variables, uO , are considered,
˚
V h WD v 2 L2 .˝/ W vjKi 2 Pk .Ki / if Ki \ I D ;;

vjKi 2 Pk .Ki / ˚ HPk .Ki / if Ki \ I ¤ ; ;
˚ 2
(21)
 WD vO 2 L . / W vj
h
O i 2 Pk . i / if i \ I D ;;

vj
O i 2 Pk . i / ˚ HPk . i / if i \ I ¤ ; ;

where H is a Heaviside function that enriches the approximation in cut elements and
on cut faces. The Heaviside function can be defined, for instance, as

1 in ˝1
HD
1 in ˝2

to introduce discontinuities across the interface I.

In every cut element Ki , given the considered enriched space Pk .Ki / ˚ HPk .Ki /,
u is assumed to be discontinuous across the interface I. The continuity of the
solution and the normal flux across the interface is then imposed as usual in HDG by
means of an auxiliary trace variable on the interface, which is eliminated afterwards
keeping the original unknowns.
At cut faces, a discontinuous approximation for the trace uO is also considered
but, in this case, there is no need to impose weak continuity on it. In this way the
resulting formulation is equivalent to a standard HDG formulation on a mesh where
some elements are split (for instance into a triangle and a quadrilateral) to fit the
interface, with the usual face-by-face discontinuous approximation of the trace.
This definition ensures that the proposed X-HDG formulation keeps the super-
convergence and stability properties of standard HDG accordance with an X-FEM
philosophy using the original computational mesh with an enriched approximation
at cut elements and faces.

3.1 Local Problem at a Bimaterial Element

The X-HDG local problem at an element Ki cut by the interface corresponds to the
discretization of (20), that is: Given uO 2 h , find u 2 Pk .Ki / ˚ HPk .Ki /, q 2
ŒPk .Ki / ˚ HPk .Ki / d such that
Z Z Z Z
vr  q dV C  v.u  uO / dS C 2 fv.u  uQ i /g dS D vf dV
Ki nIi @Ki Ii Ki nIi
Z Z Z Z
1
q  w dV  ur  w dV C uO w  n dS C uQ i w  n dS D 0
Ki nIi  Ki nIi @Ki Ii
(22)
X-HDG for Void and Bimaterial Problems 115

for all v 2 Pk .Ki / ˚ HPk .Ki / and w 2 ŒPk .Ki / ˚ HPk .Ki / d , where Ii WD Ki \ I
is the interface segment in the element and uQ i is a new trace variable approximating
the trace of the solution on the interface Ii . The mean operator fg is defined taking
values from the left and right domains sharing the interface or the mesh face, f}g D
1
2 .}1 C }2 /.
The X-HDG local problem (22) can be derived following the standard HDG
rationale for each one of the domains, Ki \ ˝1 and Ki \ ˝2 , and summing the
obtained weak forms. Compared to the weak form for standard elements (6), the
X-HDG weak form for a cut element has two additional terms corresponding to
integrals along the interface Ii involving the new trace variable uQ i and taking into
account the discontinuous nature of the approximation space. These new terms
weakly impose the condition u D uQ i on Ii , ensuring the weak imposition of
continuity of u across the interface.
The local problem is now closed by imposing the conservativity condition across
the interface Ii , Oq  n D 0 on Ii : The weak form of this conservativity condition
in I i can be written as: given u 2 Pk .Ki / ˚ HPk .Ki /, q 2 ŒPk .Ki / ˚ HPk .Ki / d , find
uQ i 2 Pk .Ii / such that
Z Z
 
vq
Q  n dS C 2 vQ fug  fgQui dS D 0 8vQ 2 Pk .Ii /: (23)
Ii Ii

Similarly to the void cut element with Neumann boundary conditions, Eq. (23)
can be used to isolate uQ i in terms of ui and qi . Then, replacing in (22) a local problem
with the same structure as standard HDG (10), but modified matrices, is obtained.

4 Numerical Tests

The performance of the novel X-HDG method is tested on three numerical

examples. The Laplace equation with known analytical solution is solved over a
square domain with a circular void in Sect. 4.1. Boundary condition on the circular
boundary is of Neumann type in the first case and of Dirichlet type in the second.
The accuracy and the convergence of X-HDG for void problems is tested and
compared to HDG with a mesh adapted to the void boundary.
Section 4.2 demonstrates the performance of X-HDG for bimaterial problems
where the domain is divided by a straight interface. On each side of the interface
different materials are considered and the corresponding error analysis is done. As
usual, the performance of X-HDG is compared with the one of HDG by solving
exactly the same problem with HDG settings, i.e, a mesh fitting to the interface.
In all numerical tests, the stabilization parameter is  D 1 on all faces.
116 C. Gürkan et al.

4.1 X-HDG with Voids

The first numerical example assesses the performance of X-HDG with Neumann
boundary conditions imposed at the void interface. The Laplace Eq. (19), with  D
1, is solved over a square domain with a centered circular void with radius 0:41,
˝ D .1; 1/2 nB..0; 0/; 0:41/. Neumann boundary conditions are imposed on the
void boundary I D @B..0; 0/; 0:41/, and Dirichlet
 boundary conditions are imposed
on the exterior boundary @˝nI D @ .1; 1/2 . Dirichlet and Neumann values and
the source term f are set so that the analytical solution is
 
u.x; y/ D exp 0:1 sin.5:1x  6:2y/ C 0:3 cos.4:3x C 3:4y/ :

Figure 3 shows the computational domain and the analytical solution.

Figure 4 presents the computational mesh for X-HDG and for standard HDG,
respectively. For the X-HDG computation, a regular triangular mesh in the square
domain .1; 1/2 is considered, covering the domain ˝ and fitting the exterior
boundary @˝nI. A level set function is used to describe the boundary of the void,
I. Three kinds of elements appear in the computational mesh. Elements inside
the domain (dark gray) are treated as standard HDG elements. For elements cut
by the interior boundary I (light gray), the modified X-HDG local problem is
considered, see Sect. 2.2. The elements that are totally inside the void (white) have
no contribution to the solution, so they are simply disregarded. The computational
mesh for standard HDG is adapted to fit the void boundary, with similar uniform
mesh size.
Figure 5 shows the evolution of the L2 error for decreasing uniform mesh
size, for X-HDG and HDG with degrees k D 2; 3; 4, for both the solution and
the postprocessed superconvergent solution, see Remark 3. X-HDG keeps optimal
convergence with rates close to kC1 for the solution, and kC2 for the postprocessed
solution, with similar levels of accuracy as HDG, while getting rid of adapting the
mesh to the void boundary.

1.4

1.2

0.8

1
0.5 1
0 0.5
–0.5 0
–0.5
–1 –1

Fig. 3 Circular void example: domain ˝ and void boundary I , and analytical solution
X-HDG for Void and Bimaterial Problems 117

Fig. 4 Circular void example: computational mesh for X-HDG and for standard HDG. The X-
HDG mesh is not adapted to the void boundary. Elements in the interior of the domain are colored
in dark gray. Elements in softer gray are elements cut by the interface I . Elements in white are
inside the void, and are not considered in the computation

Fig. 5 Neumann circular void example: convergence plots for X-HDG (left) and HDG (right).
The numbers correspond to the slope of each segment, and they are underlined for postprocessed
solution

Remark 4 The description of the interface is done using a level set function given
by its nodal values on the k-degree mesh used for the solution, and piece-wise k C 1
degree level set function is used for the computation of post-processed solution to
ensure k C 2 convergence rate reached by the postprocessed solution. For standard
HDG this is equivalent to requiring a boundary approximation of polynomial degree
k C 1 properly fitting the description of the boundary for the superconvergent
postprocessing.
Now, Dirichlet boundary conditions imposed over the interface I. The domain,
the boundary conditions on the exterior boundary, and source term definitions are
118 C. Gürkan et al.

Fig. 6 Dirichlet circular void example: convergence plots for X-HDG (left) and HDG (right).
The numbers correspond to the slope of each segment, and they are underlined for postprocessed
solution

kept same as before. Convergence plots are shown in Fig. 6. Again, a one to one
resemblance is observed between the X-HDG and HDG results, keeping optimal
convergence properties of HDG with an unfitted computational mesh.

4.2 Bimaterial X-HDG

Finally, the Laplace equation (19) over a square domain divided into two disjoint
subdomains by a linear interface is solved in order to demonstrate the efficiency of
XHDG formulation for bimaterial problems. The size of our domain is .1; 1/2 and
the interface I is located at x D 0:2031 with the viscosity parameter being 1 D 1
in the left side of the interface i.e., at ˝1 and being 2 D 2:5 in ˝2 as shown in
Fig. 7.
Homogenous Neumann boundary conditions are set on the interface I and
homogenous
  Dirichlet boundary conditions set on the outer boundary @˝nI D
@ .1; 1/2 where the analytical solution reads

5x5 in ˝1 ;
u.x/ D
2x5 C A in ˝2 ;

with jump A D 3.0:2031/5. In order to solve this problem, the procedure explained
in Sect. 3 is followed for elements cut by the linear interface. The standard HDG
formulation is used for elements not cut by the interface.
Figure 8 shows the convergence history of X-HDG and HDG for the bimaterial
interface problem. Starting with an initial mesh with four elements per coordinate
direction, we have done four mesh refinements—each refinement doubling the ele-
ment number on the sides—both for X-HDG and HDG with varying approximation
X-HDG for Void and Bimaterial Problems 119

Fig. 7 Bimaterial X-HDG: computational mesh for X-HDG and for standard HDG after two mesh
refinements. On the left, the X-HDG mesh not adapted to the linear interface boundary I which
shown in black. Elements in white are standard elements in domain one whereas elements in dark
gray are standard elements in domain two. With light gray, elements cut by the interface I are
shown. On the right, HDG mesh fitting to the linear interface is shown. Elements in domain one
are shown in white whereas elements in domain two shown in light gray

Fig. 8 Bimaterial X-HDG: Convergence history of X-HDG and HDG. Slopes for the solution
on each segment are shown in bold, slopes for post-processed solution are shown in bold and
underlined

degree between k D 1 to k D 4. The results verify the X-HDG formulation for

bimaterial interface problems when compared to HDG. Over the fine mesh and for
higher approximation degree we can see the slight effect of matrix ill-conditioning
in both X-HDG and HDG convergence history but this does not prevent any of the
methods to reach optimal k C 1 and k C 2 (super-) convergence rates.
120 C. Gürkan et al.

5 Conclusions

A strategy for the HDG solution of boundary value problems with voids, inclu-
sions, free boundaries, and material interfaces has been proposed. It is based on
an extended finite element (X-FEM) philosophy with a level-set description of
interfaces. With this new strategy, the computational mesh does not need to be
adapted to the interface (i.e.; the boundary), simplifying and reducing the cost of
mesh generation. In case of evolving interfaces, the cost of continuous re-meshing,
and the approximation errors due to the projection from one mesh to another, are
avoided.
Differently to previous attempts for the HDG solution with meshes that do not fit
the boundary, the computational mesh covers the domain in our method, avoiding
extrapolations and ensuring the robustness of the method. The local problem in
elements not cut by the interface and the global problem are discretized as usual in
HDG. A modified local problem is considered for elements cut by the interface. In
every cut element, an auxiliary trace variable on the boundary is introduced, which
is eliminated element-by-element using the boundary conditions on the interface,
keeping the original unknowns and the structure of the local problem solver. In case
of bimaterial problems the solution space is enriched with Heaviside functions to
introduce the discontinuity within the element for the elements cut by the interface
whereas elements not cut by the interface are treated with standard HDG approach.
Although the special treatment for cut elements due to the modified numerical
integration causes a slight increase in CPU time, in practical applications the extra
cost becomes negligible compared to the cost of mesh adaptation or re-meshing.
The proposed method has been compared to standard HDG on three numerical
tests: the solution of a Laplace problem with known analytical solution in a square
domain with a circular void with Dirichlet and Neumann boundary conditions,
and the solution of the Laplace equation over a square domain with two different
materials being divided by a linear interface. Standard HDG computations have been
done on a mesh fitting the voids boundaries and material interfaces for comparison
with X-HDG that uses a regular mesh covering the domain. The numerical tests
assert that X-HDG keeps the HDG optimal convergence rates for the solution,
the gradient, and the post-processed super-convergent solution, without the need
to adapt the mesh to the boundary. In all tests, similar accuracy is observed for a
similar mesh size.

Acknowledgements This work was supported by the DAFOH2 project (Ministerio de Economia
y Competitividad, MTM2013-46313-R), and the Erasmus Mundus Joint Doctorate SEED project
(European Commission).
X-HDG for Void and Bimaterial Problems 121

References

1. Cheng, K.W., Fries, T.P.: Higher-order XFEM for curved strong and weak discontinuities. Int.
J. Numer. Methods Eng. 82(5), 564–590 (2010)
2. Cockburn, B.: Discontinuous Galerkin methods for computational fluid dynamics. In:
Encyclopedia of Computational Mechanics. Fluids, vol. 3, chap. 4. Wiley, New York (2004)
3. Cockburn, B., Solano, M.: Solving convection-diffusion problems on curved domains by
extensions from subdomains. J. Sci. Comput. 59(2), 512–543 (2014)
4. Cockburn, B., Dong, B., Guzmán, J.: A superconvergent LDG-hybridizable Galerkin method
for second-order elliptic problems. Math. Comput. 77(264), 1887–1916 (2008)
5. Cockburn, B., Gopalakrishnan, J., Lazarov, R.: Unified hybridization of discontinuous
Galerkin, mixed, and continuous Galerkin methods for second order elliptic problems. SIAM
J. Numer. Anal. 47(2), 1319–1365 (2009)
6. Cockburn, B., Gopalakrishnan, J., Nguyen, N.C., Peraire, J., Sayas, F.J.: Analysis of HDG
methods for Stokes flow. Math. Comput. 80(274), 723–760 (2011)
7. Cockburn, B., Qiu, W., Shi, K.: Conditions for superconvergence of HDG methods for second-
order elliptic problems. Math. Comput. 81(279), 1327–1353 (2012)
8. Cockburn, B., Qiu, W., Solano, M.: A priori error analysis for HDG methods using extensions
from subdomains to achieve boundary conformity. Math. Comput. 83(286), 665–699 (2014)
9. Dreau, K., Chevaugeon, N., Moës, N.: Studied X-FEM enrichment to handle material interfaces
with higher order finite element. Comput. Methods Appl. Mech. Eng. 199(29–32), 1922–1936
(2010)
10. Fries, T.P., Belytschko, T.: The extended=generalized finite element method: An overview of
the method and its applications. Int. J. Numer. Methods Eng. 84(3), 253–304 (2010)
11. Giorgiani, G., Fernández-Méndez, S., Huerta, A.: Hybridizable Discontinuous Galerkin p-
adaptivity for wave propagation problems. Int. J. Numer. Methods Fluids 72(12), 1244–1262
(2013)
12. Giorgiani, G., Modesto, D., Fernández-Méndez, S., Huerta, A.: High-order continuous and
discontinuous Galerkin methods for wave problems. Int. J. Numer. Methods Fluids 73(10),
883–903 (2013)
13. Giorgiani, G., Fernández-Méndez, S., Huerta, A.: Hybridizable Discontinuous Galerkin with
degree adaptivity for the incompressible Navier–Stokes equations. Comput. Fluids 98,
196–208 (2014)
14. Hesthaven, J., Warburton, T.: Nodal high-order methods on unstructured grids: I. Time-domain
solution of Maxwell’s equations. J. Comput. Phys. 181(1), 186–221 (2002)
15. Huerta, A., Angeloski, A., Roca, X., Peraire, J.: Efficiency of high-order elements for
continuous and discontinuous Galerkin methods. Int. J. Numer. Methods Eng. 96(9), 529–560
(2013)
16. Huynh, L., Nguyen, N., Peraire, J., Khoo, B.: A high-order hybridizable discontinuous
Galerkin method for elliptic interface problems. Int. J. Numer. Methods Eng. 93, 183–200
(2012)
17. Kirby, R., Sherwin, S.J., Cockburn, B.: To CG or to HDG: A comparative study. J. Sci. Comput.
51(1), 183–212 (2011)
18. Montlaur, A., Fernández-Méndez, S., Huerta, A.: Discontinuous Galerkin methods for the
Stokes equations using divergence-free approximations. Int. J. Numer. Methods Fluids 57(9),
1071–1092 (2008)
19. Nguyen, N., Peraire, J., Cockburn, B.: A hybridizable discontinuous Galerkin method for
Stokes flow. Comput. Methods Appl. Mech. Eng. 199(9–12), 582–597 (2010)
20. Nguyen, N.C., Peraire, J., Cockburn, B.: An implicit high-order hybridizable discontinuous
Galerkin method for the incompressible Navier-Stokes equations. J. Comput. Phys. 230(4),
1147–1170 (2011)
21. Peraire, J., Persson, P.O.: The compact discontinuous Galerkin (CDG) method for elliptic
problems. SIAM J. Sci. Comput. 30(4), 1806–1824 (2008)
122 C. Gürkan et al.

22. Sala-Lardies, E., Fernández-Méndez, S., Huerta, A.: Optimally convergent high-order X-FEM
for problems with voids and inclusions. In: ECCOMAS 2012: 6th European Congress on
Computational Methods in Applied Sciences and Engineering. Programme book of abstracts,
September 10–14, 2012, Vienna, Austria, pp. 1–14 (2012)
23. Wang, B., Khoo, B.C.: Hybridizable discontinuous Galerkin method (HDG) for Stokes
interface flow. J. Comput. Phys. 247, 262–278 (2013)
Crack Lip Contact Modeling Based
on Lagrangian Multipliers with X-FEM

Yuan Jin, Olivier Pierard, Eric Wyart, and Eric Béchet

Abstract The eXtended Finite Element Method (X-FEM), developed intensively

in the past 15 years has become a competitive tool for the solution of problems
with evolving discontinuities and singularities. In the present study, we focus
on the application of X-FEM on frictionless contact problems in the context of
fracture mechanics. A promising approach in the literature counting for this problem
consists in applying Lagrangian multipliers. Meanwhile, as pointed out in Ji and
Dolbow (Int J Numer Methods Eng 61:2508–2535, 2004), a naive choice for
Lagrangian multiplier space leads to oscillatory multipliers on the contact surface.
This oscillation results from a non-uniform but mesh-dependent inf-sup condition.
In this work, we adapt the algorithm proposed in Béchet et al. (Int J Numer Methods
Eng 78:931–954, 2009) on crack lip contact by discretizing the displacement field
with both scalar and vector tip enrichment functions (Chevaugeon et al., Int J
Multiscale Comput Eng 11:597–631, 2013). The influence of the tip enrichment
functions on the stability of the formulation is addressed. We show evidences
that the vector enrichment functions can improve the conditioning of the problem
without jeopardizing the simulation accuracy in the presence of contact.

1 Introduction

The eXtended Finite Element Method (X-FEM) [18], which allows for considering
crack geometries within elements, is largely used nowadays to handle crack growth
problems. This approach, which uses the Partition of Unity [17], enriches the
classical basis of shape functions with the so-called enrichment functions to describe
the discontinuity of the displacement field across the crack surface.
In crack propagation simulation, the crack lip contact is important to be con-
sidered especially when dealing with complex mechanical loadings, e.g. presence

Y. Jin • O. Pierard () • E. Wyart

Cenaero, rue des Frère wright 29, 6041 Gosselies, Belgium
e-mail: [email protected]; [email protected]; [email protected]
E. Béchet
Université de Liège, Allée de la Découverte 9, B-4000 Liège, Belgium
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 123

G. Ventura, E. Benvenuti (eds.), Advances in Discretization Methods,
SEMA SIMAI Springer Series 12, DOI 10.1007/978-3-319-41246-7_6
124 Y. Jin et al.

of residual stresses generated during an upstreaming manufacturing process and/or

compressive loading. A promising route proposed in the literature to solve contact
problem in the X-FEM framework is with Lagrangian multipliers. The idea
was firstly introduced by Ji and Dolbow[14]. The authors report that a naive
discretization of the Lagrangian multipliers violates the inf-sup condition, which
determines the stability of the formulation. The problem can be fixed by defining
a reduced Lagrangian multiplier space as proposed in [3, 8, 10, 20]. In these
works, only Heaviside functions are introduced to represent the jump across the
implicit interface. Nevertheless, for crack lip contact, the tip enrichment functions
are necessary for the description of the stress singularity around the crack tip
and thus for the calculation of the stress intensity factors (SIFs). In other words,
these asymptotic functions cannot be omitted in crack propagation simulations.
The algorithm of Lagrangian space construction proposed in [3] was applied to
model crack lip contact with classical scalar tip enrichment functions on structured
meshes in [25]. Meanwhile, the effect of these functions on the stability of the
formulation was not discussed by the authors [25]. The present study aims at
modeling frictionless crack lip contact on unstructured meshes. Both Heaviside
function and tip enrichment functions are used to discretize the displacement field.
The influence of the tip enrichment functions on the stability of the formulation is
specifically addressed.
Besides the advantage of relaxing the constraint on finite element mesh, X-FEM
is reported to be able to reach the optimal convergence rate with first-order shape
functions when the geometrical enrichment strategy is used [2, 15]. However, this
enrichment strategy can result in high condition number of the stiffness matrix. The
conditioning problem prevents applying this method on complex large industrial
problems. Several alternatives are proposed in the literature to address this issue by
applying different types of enrichment functions, e.g. the stabilized GFEM [12, 13],
the vector enrichment functions [6, 23]. In this work, in addition to classical scalar
tip enrichment functions, vector enrichment functions proposed by Chevaugeon
et al. [6] are also applied to model crack lip contact. Their performances in terms of
simulation accuracy and conditioning are compared.
The paper is organized as follows. We recall at first the contact formulation.
Secondly the discretization of the displacement field and the Lagrangian multipliers
is presented. The influence of the discretization on the stability of the formulation is
then discussed. The fourth section consists in a description of the SIF calculation in
the case of crack lip contact. In the last section of the paper, two two-dimensional
test cases are applied to analyze the performance of the proposed strategy.

2 Formulation of the Continuous Problem

Hereinafter, bold symbols denote vectors or tensors. We consider a cracked domain

˝ in R2 . The two sides of the crack are denoted by cC and c respectively
(Fig. 1). Only normal (frictionless) contact condition is considered between cC
Crack Lip Contact Modeling Based on Lagrangian Multipliers with X-FEM 125

Fig. 1 A cracked domain

and c in the present study. In the initial configuration, the two sides coincide.
n (D nC D n ) denotes the outward normal unit vector on cC . The body
 2
is subjected to surface loads g 2 L2 . N / . Then under planar small strain
assumptions, the problem of homogeneous isotropic linear elasticity consists in
finding the displacement field u W ˝ ! R2 satisfying

r   .u/ D 0; in ˝; (1)
u D 0; on D ; (2)
 .u/  n D g; on N ; (3)

where  D .ij /, 1  i; j  2, stands for the stress tensor field and ‘’is a contraction
over a single index.  is defined by the constitutive law for linear elasticity:

 D C W "; (4)
 
where C is the Hooke tensor, ".u/ D ru C ruT =2 represents the linearized
strain tensor field and ‘W’denotes a double contraction.
For a displacement field u, the following notations are adopted to describe the
displacement on the contact surface:

uC D uC C C C
n n C ut t and u D u   
n n C ut t ; (5)

where t is the unit tangent vector of C and uC (resp. u ) is the trace of displacement
on CC (resp. on C ). Then, the frictionless contact condition reads:
 
n .u/  0; u D uC 
n C un  0; n .u/  u D 0; (6)

where n .u/ is the normal pressure on the contact surface.

In what follows, we adopt classical notations for the functional spaces: W m;p .˝/
denotes the Sobolev space of functions v for which all components of mth order
weak derivative Dm .v/ lie in LP .˝/. The associated norm is denoted by k  km;p;˝ .
Following the convention in the literature [8], we note H m .˝/ WD W m;2 .˝/, for
which the index p D 2 is omitted. The components of the solution to problems
126 Y. Jin et al.

 2
(1–6) are then belong to the space V WD fv 2 H 1 .˝/ ; vj D D 0g. As in [1], we
introduce the space:

W D fvn j C W v 2 Vg: (7)

and its dual space W 0 . We note H 1=2 . C / the trace space of H 1 .˝/ restricted to
C and H 1=2 . C / the dual space of H 1=2 . C /. Then we have W  H 1=2 . C / and
W 0  H 1=2 . C /. Let  WD nT   n be the normal surface force on the contact
surface.  then belongs to the space M D f 2 W 0 W h; iW 0 ;W  0; for all 2
W;  0 on C g. The notation h; iW 0 ;W stands for the duality pairing between W 0
and W. The weak formulation of the frictionless contact problem then reads, find
u 2 V and  2 M such that
Z Z Z
 
8u 2 V;  .u/ W " u d˝  u d D g  u d : (8)
˝ C N

Z
8 2 M;  ud D 0: (9)
C

3 Discretization

The discretization of the problem (8)–(9) involves a pair of finite

 element
 spaces
Vh  V and M h  M leading to the discrete formulation: find uh ; h 2 Vh  M h
such that
Z Z Z
   
8uh 2 Vh ;  uh W " uh d˝  h uh d D g  uh d : (10)
˝ C N

Z
8h 2 M h ; h uh d D 0: (11)
C

In this work, we suppose that ˝ is meshed with a family of affine triangular meshes
Th , which is assumed to be quasi-uniform and regular. The choice of the Vh and M h
has to satisfy a uniform inf-sup condition [4] with respect to suitable norms. In the
present study, as in [3, 8], the following mesh-dependent L2 norms are considered:
X
k h k21=2; C WD he k h k20;e ; (12)
e2 c
X 1
k uh k21=2; C WD k uh k20;e ; (13)
e2
h e
c
Crack Lip Contact Modeling Based on Lagrangian Multipliers with X-FEM 127

where he denotes the length of the segment e on C . Then in the present context, the
inf-sup condition reads:
R
C h uh d
9c < 0; independent of h; 8h 2 M ; suph
 ch1=2 k h k0; C
uh 2Vh k uh k1;˝
(14)
where h WD max he .
e
Let Uh and Lh be respectively the vector of the displacement and the Lagrangian
multipliers. Then the algebraic form of the problem (10)–(11) reads:
  
Ah BTh Uh Fh
D : (15)
Bh 0 Lh 0

To evaluate numerically the inf-sup condition, thanks to the work in [3, 5], the
following algebraic equation is introduced:
  
Ah BTh Uh Fh
1 D (16)
Bh k Mh Lh 0

where
R Mh , a symmetric mass matrix, corresponds to the integral on C
( C h h d ). Equation (16) is equivalent to Eq. (15) when k tends to infinity.
We denote ˇhImin the first non-zero eigenvalue of the following problem:

1  T 1 
B M Bh Vh D ˇh Ah Vh : (17)
h h h
0:5
Then the value of c in Eq. (14) is approximated by minhi ˇhImin .

3.1 Discretization of the Displacement Field

Let Eh be the set of elements of Th completely cut by C and Vh the corresponding

intersected edges. We denote Nh  N the nodes of the mesh, Kh  Nh the nodes
of elements in Eh . X-FEM is applied to count for the displacement jump across the
crack surface. The discretized displacement field is expressed as follows:
X X
uh D Ni .x/ui C Nk .x/H.x/ak : (18)
Nh Kh

The Heaviside-like step function H.x/, with a changing sign across the crack,
enriches the nodes in Kh . Classically, it is defined thanks to the signed distance
function to the crack , i.e. the so-called normal level set (see [19, 26]):

H.x/ D sign ..x// : (19)

128 Y. Jin et al.

In addition to Heavisde functions, tip enrichment functions are also applied on the
set of nodes within a prescribed enrichment radius to the crack tip Jh  Nh for
the purpose of describing the stress singularity. In the present study, two types of
tip enrichment functions denoted respectively by scalar enrichment functions and
vector enrichment functions are used. For classical scalar enrichment functions, the
displacement field is approximated by
!
X X X X
u D
h
Ni .x/ui C Nk .x/H.x/ak C Nj .x/ fl .x/bkl ; (20)
Nh Kh Jh l

where
p p p p
f D Πr cos.=2/; r sin.=2/; r sin.=2/ sin ; r cos.=2/ sin./ : (21)

.r; / represents the polar coordinate system associated with the crack tip. The
vector enrichment functions are introduced in [6]:
X X XX
uh D Ni .x/ui C Nk .x/H.x/ak C Nj .x/K˛ .x/cj˛ ; (22)
Nh Kh Jh ˛

where
p
K1 D
r cos.=2/ .k  cos.// Œe1 .x/ C e2 .x/ ; (23)
p p
K2 D r sin.=2/ .k C 2 C cos.// e1 .x/ C r cos.=2/ .k  2 cos.// e2 .x/;
(24)
p
K3 D r sin.=2/e3 .x/: (25)

As indicated in Eqs. (23)–(25), K˛ , ˛ 2 f1; 2; 3g are defined on the local basis ei , i 2

f1; 2; 3g. The vector enrichment functions limit the number of degrees of freedom
per enriched node to 3 (to be compared to the 12 degrees of freedom per node
with scalar enrichment functions). More details of these two types of enrichment
functions are provided in [2, 6].

3.2 Discretization of the Lagrangian Multipliers

As in [8, 10], the Lagrangian multiplier components are defined on the parent nodes
belonging to Kh . They are based on the same linear shape functions Ni .x/ that are
used for the displacement field. The approximation space M h for the multiplier is
then:
0
M h D fh j , h 2 M h g; (26)
Crack Lip Contact Modeling Based on Lagrangian Multipliers with X-FEM 129

where
0
X
M h WD fx 2 Eh ! i Ni .x/; i 2 Rg: (27)
i2Kh

With X-FEM, a naive linear P1  P1 interpolation choice for the displacement

and Lagrangian multipliers respectively is not stable and results in oscillations [14].
An adequate reduction of the Lagrangian multiplier space can efficiently address the
problem [3, 8]. In the algorithm proposed by Béchet et al. [3], a piecewise linear or
constant interpolation is adopted for the Lagrangian multipliers. This interpolation,
which can be considered as between P1 and P0, is denoted by P1 in the present
study. Accordingly, combined with the linear interpolation of the displacement,
the formulation used is denoted by P1=P1 . Another approach was proposed by
Ferté et al. [8]. By releasing some constraint from P1 , the new algorithm of P1
multipliers, referred to P1 , verifies P1  P1 . The formulation is combined with
a piecewise quadratic displacement, which is denoted by P2=P1. Both approaches
are proven mathematically to be able to satisfy the inf-sup condition. Roughly
speaking, with the constructed Lagrangian multiplier space, R it is possible to find
h wh d
wh 2 Vh that h D wh . Then it is sufficient to show Ckwh k1;˝  ch1=2 k
h k0; C . We refer to [3, 8] for more details of this demonstration.
In this work, a linear interpolation on the displacement field is adopted. Accord-
ingly, the algorithm of Béchet is applied to construct the Lagrangian multiplier
space. It is noteworthy that in [3, 8], the implicit interface completely crosses the
simulation domain. However, for contact between crack lips, the contact interface
ends within the simulation domain. Therefore, a special treatment is applied on the
element containing the crack tip. As illustrated in Fig. 2, the element with nodes
i ; i 2 1; 2; 3 contains the crack tip. If the edge q1  q3 supports an equality
qct ct ct

relationship, then a complementary equality relationship is imposed so that the

degree of freedom of the Lagrangian multiplier defined on qct 2 equals the ones on
qct
1 and q ct
3 . Otherwise, since either q ct
3 or q ct
1 is connected to another edge supporting
an equality relationship (see [3]), a single equality relationship is imposed among

Fig. 2 Restriction algorithm P1 defined by Béchet et al. [3]

130 Y. Jin et al.

i ; i 2 1; 2; 3 so that the node connects the other two nodes. In summary, a P0

qct
interpolation is used in this element for Lagrangian multipliers.
Furthermore, in the work of Béchet et al. [3], only the Heavisde function is used
for enrichment. In the present study, the tip enrichment functions given in Eqs. (21)
and (23)–(25) are also applied to describe the stress singularity around the crack
tip. The application of these functions does not violate the inf-sup condition. In
fact, by simply imposing the related degrees of freedom (bkl in Eq. (20) or cj˛ in
Eq. (22)) equal to zero, we can easily find a displacement field wh 2 Vh that satisfies
h D wh . Then the demonstration procedure proposed in [3] proves that with this
displacement discretization the inf-sup condition is respected.

4 Stress Intensity Factors

An accurate calculation of the stress intensity factors (SIFs) is essential for crack
propagation simulation. In the present study, the SIFs are extracted from the X-
FEM solution thanks to the interaction integrals in a domain form [7, 16, 24, 28].
Here, we recall the calculation procedure. For this, two states of a cracked body are
defined:

• state 1 = present state u1i ; ij1 ; "1ij ,

• state 2 = auxiliary state u2i ; ij2 ; "2ij .

Accordingly, the interaction integral I .1;2/ can be calculated as follows:

I  
@u2 @u1
I .1;2/ D W .1;2/ ı1j  ij1 i  ij2 i nj d ; (28)
@x1 @x1

where ı is the Kronecker’s delta and W .1;2/ is the interaction strain energy defined
as
.1/ .2/ .2/ .1/
W .1;2/ D ij "ij D ij "ij : (29)

The direction x1 is defined by the local basis associated with the crack tip (see
Fig. 3). By defining a weighting function q.x/:
8
ˆ
<1
ˆ on ;
qD 0 on C0 ; (30)
ˆ
:̂arbitrary elsewhere;
Crack Lip Contact Modeling Based on Lagrangian Multipliers with X-FEM 131

Fig. 3 Contour integral

around the crack tip

Equation (28) can be reformulated as:

Z  2 1
.1;2/ .1;2/ 1 @ui 2 @ui
I D W ı1j C ij C ij qmj dC
C @x1 @x1
Z  2 1
.1;2/ 1 @ui 2 @ui
 W ı1j C ij C ij qmj dC; (31)
CC CC @x1 @x1

where C is the contour of the Equivalent Domain integral, i.e.

C D C CC C C C C0 ; (32)

and m is the unit outward normal of C as illustrated in Fig. 3. On , we have

m D  n.
I .1;2/ can be written for a general mixed modes (I-II-III) fracture [27]:

2  1 2  KIII
1 2
KIII
I .1;2/ D K K C K 1 2
K C ; (33)
E I I II II


where E D 1 E
2 in the present study (plane strain). E and  denote respectively

the Young’s modulus and the Poisson ratio. To determine KI1 , the auxiliary field can
be chosen as the asymptotic solution of a pure mode I state so that KI2 D 1, KII2 D 0
2
and KIII D 0. Then KI1 can be deduced from Eq. (33) as follows:

E .1;2/ E .1;2/
KI1 D 2
I D I : (34)
2KI 2

.2/
For such auxiliary field, ij D 0 for i D 1; 2 and j D 1; 2 on CC and C .
Accordingly, in case of crack lip contact, the interaction integral I .1;2/ can be
132 Y. Jin et al.

written as
Z " .2/ .1/
#
.1;2/ .1;2/ .1/ @ui .2/ @ui @q
I D W ı1j C ij C ij dA
A @x1 @x1 @xj
Z " .2/
#
.1/ @ui
 ij qmj dC: (35)
CC CC @x1

It is worth emphasizing that different from traction-free cracks, in the case of crack
lip contact, the second term in Eq. (35) cannot be ignored. Details of the calculation
procedure of the SIFs can be found in [21, 22].

5 Numerical Experiments

In this section, two two-dimensional (2D) frictionless contact problems are used
to illustrate the performance of the algorithm described in the previous section.
In these two cases, contact forces along the crack are respectively uniform and
non-uniform compression. The numerical solutions are compared with analytical
values of the Lagrangian multipliers. Plane strain conditions are assumed in both
examples. The Young’s modulus and the Poisson ratio and are taken respectively as
equal to 1 and 0:3.

5.1 Cracked Block Under Uniform Pressure

A uniform pressure .0:01/ is imposed on the top surface of a square sample Œ2  2 ,

whereas the bottom right corner is restrained in both directions and the bottom left
corner in the vertical .y/ one. A crack is inserted into the square sample at y D 0
(from x D 1 to x D 0:6). The test case is illustrated in Fig. 4.

Fig. 4 Uniform pressure

imposed on the top surface of
a 2D square sample with a
crack (red line) inside
Crack Lip Contact Modeling Based on Lagrangian Multipliers with X-FEM 133

(a) (b)

Fig. 5 2D compression test case; displacement field is only enriched with Heaviside functions;
(a) Lagrangian multiplier in the simulation domain; (b) the value of the Lagrangian multiplier on
the crack expressed as a function of the distance to the left extremity of the crack

To solve the above problem, firstly as in [3] only Heaviside functions are used as
enrichment functions to discretize the displacement field. The problem is solved on
an unstructured mesh. 42 equal subdivisions are used along each edge of the square.
The mesh size h is thus about 0:05. Applying the previously presented algorithm,
the resulted Lagrangian multiplier field is illustrated in Fig. 5. With an adequate
Lagrangian multiplier space, the Lagrangian multiplier is constant along the crack
and the value corresponds to the pressure imposed on top surface.
Now, for the same problem, the two types of tip enrichment functions are
included in the displacement field discretization as indicated by Eqs. (21) and (23)–
(25). Geometrical enrichment strategy is applied and the enrichment radius of the tip
enrichment functions equals 0:2. However, the application of the full description of
the displacement field leads to an oscillation of the Lagrangian multiplier along the
crack as illustrated in Fig. 6a. As the same results in terms of Lagrangian multipliers
are obtained with the vector enrichment functions, only the results of the scalar
functions are shown in this section.
The magnitude of the oscillation is around 1 %. Meanwhile it can be limited to
the crack tip by refining the mesh (Fig. 6b). This problem is thus not related to the
instability of the formulation. The numerical inf-sup test (Eq. (16)) is performed
on this test case. The result is given in Fig. 7. The inf-sup value with a naive
Lagrangian multiplier space (with only Heavisde functions) tends to zero, which
indicates a violation of the inf-sup condition. On the other hand, with the adequate
discretization, regardless the type of tip enrichment functions, the value is more or
less constant as the mesh is refined. This observation further shows that the inf-sup
condition is respected with tip enrichment functions. The oscillation is very likely
to be a numerical error of the finite element resolution.
134 Y. Jin et al.

(a) (b)

Fig. 6 Oscillation of the Lagrangian multiplier around the crack tip using the scalar tip enrichment
functions; (a) result on the mesh shown in Fig. 5; (b) comparison of the results on different
meshes(the mesh size h is indicated in the Figure)

Fig. 7 Numerically computed inf-sup value

Up to now, the integration degree (IntDeg) is set to be 5 on the elements

within the enrichment radius. In the present numerical context, the number inte-
gration points used for the triangular elements equals ŒIntDeg C 3/=2 2 while the
number of integration points for the linear elements on the crack surface equals
ŒIntDeg C 1/=2 (Œx refers to the integer part of x). The calculation of the stiffness
matrix (Ah in Eq. (15)) contains the integral of the differential of the shape function
fl , which leads p to the integral of the singular functions containing a combination
of f 1r ; p1r ; 1; r; rg multiplied by harmonic functions. Therefore, more integration
points are needed to improve the calculation accuracy. As shown in Fig. 8, by
doing so, we can considerably decrease the magnitude of the Lagrangian multiplier
oscillation around the crack tip.
Crack Lip Contact Modeling Based on Lagrangian Multipliers with X-FEM 135

Fig. 8 Lagrangian multiplier value along the crack surface using different integration degrees

Besides the Lagrangian multipliers, the calculation of the SIFs is also analyzed.
Since it is a two-dimensional compressive case and only the normal contact is
considered, mode I is predominant with respect to mode II. Accordingly, we focus
on KI in this analysis. The discontinuity should be completely removed when the
contact condition is fully respected. The theoretical solution of KI in this case is
thus zero. As the local basis is aligned with the global basis and the normal to the
crack surface is the axis y, by making use of the asymptotic solution of pure opening
mode I as the auxiliary field, Eq. (35) can be further simplified as
Z " .2/ .1/
#
.1;2/ .1;2/ .1/ @ui @q.2/ @ui
I D W ı1j C ij C ijdA
A @x1 @xj @x1
Z  
1 kC1
C  p qdC; (36)
CC CC 4 2r

where k D 3  4 under the plane strain assumption. Therefore, the paccuracy of the
SIF calculation depends on the integral of the singular function 1= r. The absolute
errors in KI obtained by applying different integration degrees are expressed as a
function of mesh size (h) in Fig. 9. In the present numerical context, for the integral
on the crack surface, the crack tip is considered as a node of the surface mesh. Then
the relative position of the crack tip with respect to the finite element containing it
can have an influence on the integral of the singular function p1r . Here to simplify
the discussion, the analysis shown in Fig. 9 is performed exceptionally on structured
meshes, which do not fit the crack surface.
As expected, the error in KI decreases as more integration points and/or smaller
elements are used. As indicated in Fig. 9,pthe convergence rate equals 0:5, which cor-
responds to the one of the integral of 1= r. This observation shows for this problem
since the calculation of the Lagrangian multiplier is rather accurate especially when
136 Y. Jin et al.

Fig. 9 Evolution of the error

in KI obtained with different
integration degrees as a
function of mesh size

an elevated degree of integration is used (see Fig. 8), the convergence rate
p of KI is
mostly determined by the integration accuracy of the singular function 1= r on CC
and C . Addressing this point with a more adapted integration method, e.g. [2, 9]
will be a short-term perspective of the present study.

5.2 Cracked Block Under Non-Uniform Pressure

In this section, to further test the performance of the proposed algorithm in a more
general case, the problem of a crack emanating from a stress raiser is examined: a
circular hole in an infinite plate (illustrated in Fig. 10). The exact solution of this
problem, given in [11], is expressed in polar coordinates as follows:
 
1 R2 1 3R4 4R2
rr .r; / D 1 2 C 1 C 4  2 cos 2; (37)
2 r 2 r r
 
1 R2 1 3R4
  .r; / D 1C 2  1 C 4 cos 2; (38)
2 r 2 r

1 3R4 2R2
r .r; / D  1  4 C 2 sin 2; (39)
2 r r

where R is the hole radius, 1 is the remote compressive load and a is the crack
length. The problem is solved on a sequence of uniformly refined unstructured
meshes.
To facilitate the implementation of the problem, the exact stress is imposed as
Neumann boundary conditions on a semi-circular plate as illustrated by the dash
line in Fig. 10. The radius of the hole (R) equals 1, while the radius of the circular
Crack Lip Contact Modeling Based on Lagrangian Multipliers with X-FEM 137

σ∞ θ
r σ∞
(0, 0) x

Fig. 10 Uniform pressure imposed on the two sides of an infinite plate with a crack emanating
form a circular hole

Fig. 11 Solution of the displacement of the infinite plate problem

plate equals 3. The crack length (a) and 1 are respectively set to be 1 and 0:01.
Both Heaviside and tip enrichment functions are used to discretize the displacement
field. The enrichment radius equals 0:3. To guarantee the accuracy of the Lagrangian
multiplier calculation, the degree of integration is set be 20 for the elements within
the enrichment radius. The solution of the displacement field is described in Fig. 11.
The evolution of the energy norm error, which is defined by Eq. (40), is described
in Fig. 12.
Z 0:5
 T  
Err D  ex   h D  ex   h d˝ : (40)
˝

In addition to the cracked infinite plate problem, a continuous problem without crack
in the same configuration has also been solved. With both types of tip enrichment
functions, the optimal convergence rate, which is conformed with the continuous
problem, is reached.
138 Y. Jin et al.

Fig. 12 Evolution of the energy norm error of the infinite plate problem

Fig. 13 Profile of the Lagrangian multiplier along the crack obtained on an unstructured mesh
(h D 0:025) compared with the exact solution

The profile of the Lagrangian multiplier along the crack is compared with the
exact solution in Fig. 13. Once again, the numerical solutions with the two types of
tip enrichment functions overlap the exact solution. Since the exact stress field is
not uniform, to evaluate the calculation of the Lagrangian multiplier, the following
L2 error is introduced:
Z 1=2
k e kL2 D .ex  h /2 d : (41)
C
Crack Lip Contact Modeling Based on Lagrangian Multipliers with X-FEM 139

Fig. 14 Evolution of the L2

error of the Lagrangian
multiplier as a function of
mesh size (h)

Fig. 15 The error in KI in

the infinite plate problem
expressed as a function of
mesh size (h)

The evolution of the L2 error as function of the mesh is described in Fig. 14. As
the energy norm error, the errors in the Lagrangian multiplier converge also at the
optimal rate, which is equal to 1.
As in the first test case, mode I is predominant with respect to the other mode
in the infinite plate problem. The absolute error in KI is expressed as a function
of mesh size in Fig. 15. For this case, KI is calculated on the unstructured mesh
and it does not converge uniformly. Meanwhile, with the two types of enrichment
functions, the decrease rate of the absolute error is around 0:5, which shows despite
the non-uniform value of the Lagrangian multipliers, the error in Kp I is mostly under
the influence of the integral accuracy of the singular function 1= r on the crack
surface.
140 Y. Jin et al.

Fig. 16 The evolution of the condition number as a function of mesh size (h)

Finally the condition number of the stiffness matrix, defined by the following
norm kk, is expressed as a function of element size (h) in Fig. 16.
X
k A kD max jAij j: (42)
j
i

As shown in [6], The vector enrichment functions improve considerably the con-
ditioning compared to their scalar counterparts. In this case, despite the difference
in conditioning, the two types of tip enrichment functions lead to the same overall
results. However, this improvement can be essential for more complex problems,
e.g. unilateral contact, multi-crack problem, etc.

6 Conclusions

In the present study, the Lagrangian method is applied to consider crack lip contact
in an X-FEM context. Following conclusions are drawn from the study:
• The algorithm introduced in [3], which leads to a P1=P1 formulation, is adapted
in the present study by imposing a P0 interpolation in the element containing the
crack tip to construct the Lagrangian multiplier space along the crack surface.
• In addition to the Heaviside function, tip enrichment functions are included in
the displacement discretization. We show theoretically and numerically that the
discretization with both scalar and vector tip enrichment functions respects the
inf-sup condition.
• The proposed method leads to a precise calculation of Lagrangian multipliers.
In the case of constant pressure on the contact surface, slight oscillations (less
than 1 %) have been observed around the crack tip. We show evidences that this
Crack Lip Contact Modeling Based on Lagrangian Multipliers with X-FEM 141

oscillation can be considerably reduced by increasing the integration degree.

In the case of variable pressure on the contact surface, the vector and scalar
enrichment functions result in the same overall results. The application of the
vector functions improves considerably the conditioning of the stiffness matrix
compared to scalar ones.
• The calculation of SIFs is adapted in the present work to count for the crack lip
contact.
The present study was conducted in a two-dimensional context using unstruc-
tured meshes, and only frictionless contact is considered. In the future, this work
should be at first extended to three-dimensional structures and then to frictional
contact. Besides, we observe that the calculation of the Lagrangian multiplier and
the stress intensity is tightly related to the integration accuracy. The application of
a more adapted integration strategy within the X-FEM framework will be another
short-term perspective of this work.

Acknowledgements This work is funded by an SBO Project grant 110070: eSHM with AM of
the Agency for Innovation by Science and Technology (IWT). The authors would like to thank
C. Friebel and N. Poletz for the insightful discussion about the implementation of the numerical
framework.

References

1. Amdouni, S., Moakher, M., Renard, Y.: a stabilized lagrange multiplier method for the enriched
finite-element approximation of contact problems of cracked elastic bodies. Comput. Methods
Appl. Mech. Eng. 270, 178–200 (2014)
2. Béchet, E., Minnebo, H., Moës, N., Burgardt, B.: Improved implementation and robustness
study of the X-FEM for stress analysis around cracks. Int. J. Numer. Methods Eng. 8, 1033–
1056 (2005)
3. Béchet, E., Moës, N., Wohlmuth, B.: A stable Lagrangian multiplier space for stiff interface
conditions within the extended finite element method. Int. J. Numer. Methods Eng. 78, 931–
954 (2009)
4. Brezzi, F., Bathe, K.J.: A discourse on the stability conditions for mixed finite element
formulations. Comput. Methods Appl. Mech. Eng. 53, 27–57 (1990)
5. Chapelle, C., Bathe, K.J.: The inf-sup test. Compos. Struct. 47, 537–545 (1993)
6. Chevaugeon, N., Moës, N., Minnebo, H.: Improved crack tip enrichment functions and
integration for crack modeling using the extended finite element method. Int. J. Multiscale
Comput. Eng. 11, 597–631 (2013)
7. Duflot, M.: A study of the representation of cracks with level sets. Int. J. Numer. Methods Eng.
70, 1261–1302 (2007)
8. Ferté E., Massin, P., Moës, N.: Interface problems with quadratic X-FEM: design of a stable
multiplier space and error analysis. Int. J. Numer. Methods Eng. 11, 834–870 (2014)
9. Fries, T.P.: A corrected XFEM approximation without problems in blending elements. Int. J.
Numer. Methods Eng. 75, 503–532 (2008)
10. Geniaut, S., Massin, P., Moës, N.: A stable 3D contact formulation using X-FEM. Eur. J.
Comput. Mech. 16, 259–275 (2007)
11. Giner, E., Tur, M., Tarancón, J.E., Fuenmayor, F.J.: Crack face contact in XFEM using a
segment to segment approach. Int. J. Numer. Methods Eng. 82, 1424–1449 (2010)
142 Y. Jin et al.

12. Gupta, V., Duarte, C.A., Babuška, I., Banerjee, U.: A stable and optimally convergent
generalized FEM (SGFEM) for linear elastic fracture mechanics. Comput. Methods Appl.
Mech. Eng. 266, 23–39 (2013)
13. Gupta, V., Duarte, C.A., Babuška, I., Banerjee, U.: Stable GFEM (SGFEM): improved con-
ditioning and accuracy of GFEM/XFEM for three-dimensional fracture mechanics. Comput.
Methods Appl. Mech. Eng. 289, 355–386 (2015)
14. Ji, H., Dolbow, J.E.: On strategies for enforcing interfacial constraints and evaluating jump
conditions with the extended finite element method. Int. J. Numer. Methods Eng. 61, 2508–
2535 (2004)
15. Laborde, P., Pommier, J., Renard, Y., Salaün, M.: High order extended finite element method
for cracked domains. Int. J. Numer. Methods Eng. 64, 354–381 (2005)
16. Li, F.Z., Shih, C.F., Needleman, A.: A comparison of methods for calculating energy release
rates. Eng. Fract. Mech. 21, 405–421 (1985)
17. Melenk, J.M., Babuška, I.: The partition of unity finite element method: basic theory and
applications. Comput. Methods Appl. Mech. Eng. 1–4, 289–314 (1996)
18. Moës, N., Dolbow, J., Belytschko, T.: A finite element method for crack growth without
remeshing. Int. J. Numer. Methods Eng. 1, 131–150 (1999)
19. Moës, N., Gravouil, A., Belytschko, T.: Non-planar 3D crack growth by the extended finite
element and level sets - Part I: mechanical model. Int. J. Numer. Methods Eng. 53, 2549–2568
(2002)
20. Moës, N., Béchet, E., Tourbier, M.: Imposing Dirichlet boundary conditions in the extended
finite element method. Int. J. Numer. Methods Eng. 67, 1641–1669 (2005)
21. Moran, B., Shih, C.F.: Crack tip and associated domain integrals from momentum and energy
balance. Eng. Fract. Mech. 27, 615–641 (1987)
22. Moran, B., Shih, C.F.: A general treatment of crack tip contour integrals. Int. J. Fract. 35,
295–310 (1987)
23. Pereira, J.P., Duarte, C.A., Jiao, X.: Three-dimensional crack growth with hp-generalized finite
element and face offsetting methods. Comput. Mech. 46, 431–453 (2010)
24. Shih, C.F., Moran, B., Nakamura, T.: Energy release rate along a three-dimensional crack front
in a thermally stressed body. Int. J. Fract. 30, 79–102 (1986)
25. Siavelis, M., Guiton, M.L.E., Massin, P., Moës, N.: Large sliding contact along branched
discontinuities with X-FEM. Comput. Mech. 1, 201–219 (2012)
26. Stolarska, M., Chopp, D.L., Moës, N., Belyschko T.: Modelling crack growth by level sets in
the extended finite element method. Int. J. Numer. Methods Eng. 51, 943–960 (2001)
27. Wyart, E.: Three-dimensional crack analysis in aeronautical structures using the substructured
finite element/extended finite element method. Ph.D Thesis, Universié catholique de Louvain
Faculté des Sciences Appliquées (2007)
28. Yau, J.F., Wang, S.S., Corten, H.T.: A mixed-mode crack analysis of isotropic solids using
conservation laws of elasticity. J. Appl. Mech. 47, 335–341 (1980)
Stress Intensity Factors Through Crack
Opening Displacements in the XFEM

Markus Schätzer and Thomas-Peter Fries

Abstract The computation of stress intensity factors (SIFs) for two- and three-
dimensional cracks based on crack opening displacements (CODs) is presented
in linear elastic fracture mechanics. For the evaluation, two different states are
involved. An approximated state represents the computed displacements in the solid,
which is obtained by an extended finite element method (XFEM) simulation based
on a hybrid explicit-implicit crack description. On the other hand, a reference state
is defined which represents the expected openings for a pure mode I, II and III.
This reference state is aligned with the (curved) crack surface and extracted from
the level-set functions, no matter whether the crack is planar or not. Furthermore,
as only displacements are fitted, no additional considerations for pressurized
crack surfaces are required. The proposed method offers an intuitive, robust and
computationally cheap technique for the computation of SIFs where two- and three-
dimensional crack configurations are treated in the same manner.

1 Introduction

Simulations in linear elastic fracture mechanics pose a challenging task for numer-
ical methods, as they deal with discontinuities and singularities in solids. For
such simulations the efficiency of the standard finite element p method (FEM)
and boundary element method (BEM) is limited by the r behaviour of the
displacements in the vicinity of the crack tip/front. Special elements at the crack
tip/front, like ‘quarter-point’ elements [13, 20, 21], improve the approach. However,
a suitable mesh has to be provided for each crack geometry during propaga-
tion which frequently requires a remeshing. The extended finite element method
(XFEM) [17, 24] treats these discontinuities and singularities within the elements by
additional enrichment functions, wherefore a simple, often structured background
mesh is possible with a completely decoupled description of the crack.

M. Schätzer () • T.-P. Fries

Institute of Structural Analysis, Graz University of Technology,
Lessingstr. 25/II, 8010 Graz, Austria
e-mail: [email protected]; [email protected]

© Springer International Publishing Switzerland 2016 143

G. Ventura, E. Benvenuti (eds.), Advances in Discretization Methods,
SEMA SIMAI Springer Series 12, DOI 10.1007/978-3-319-41246-7_7
144 M. Schätzer and T.-P. Fries

In the XFEM, cracks may be defined in different ways. An implicit crack

description by means of level-set functions, where the crack geometry is described
by the zero level-sets of scalar functions [7, 23]. This description is a standard in
the XFEM, however it has the disadvantage that, in general, the update of level-set
functions after a propagation step can be cumbersome [7, 12, 22]. A hybrid explicit-
implicit crack description [10] combines the advantages of an implicitly defined
crack with those of an explicit representation, e.g. by means of straight line segments
in two dimensions or flat triangles in three dimensions. An explicit crack description
allows a simple update of the crack geometry during the crack propagation as the
update of the crack geometry is achieved by adding new segments. This hybrid
explicit-implicit description is based on three level-set functions j which are
derived from the explicitly defined master configuration, see Sect. 2.1. In this work,
this second representation of the crack geometry is used.
The behaviour of the fracture including the propagation may be based on stress
intensity factors (SIFs). A linear combination of three independent crack modes
which are scaled by the SIFs uniquely describes the situation at the crack front. A
challenging task is the computation of these scaling factors for non-planar crack
geometries with stress-free or loaded crack surfaces. One of the most important
technique is based on the interaction integral [6, 18, 26], which is based on the
energy release rate G. SIFs are then computed by means of auxiliary fields and
the relation of the energy release rate and SIFs. However, for complex three-
dimensional crack geometries the evaluation of this integral can be computationally
expensive and unstable. In this paper, a more intuitive and computationally cheap
method is introduced which allows a straightforward extension to three dimensions
also considering mode III. This is achieved by observing the displacements,
particularly the crack opening displacements (CODs), in the vicinity of the crack
tip/front and their comparison with the expected openings for a pure mode I, II and
III crack. The fitting of CODs has frequently been done in classical FEM simulations
with conforming meshes [5, 20], however, not in an XFEM context.
The paper is organized as follows: Sect. 2 recalls the XFEM approach in linear
elastic fracture mechanics and a hybrid explicit-implicit crack description. In Sect. 3,
the computation of stress intensity factors (SIFs) by crack opening displacements
(CODs) is discussed. Special attention is also given to the location of the fitting
approach and the consideration of the crack front. Numerical results in two and
three dimensions are presented in Sect. 4. Finally, Sect. 5 concludes this paper with
a brief summary.

2 XFEM with a Hybrid Explicit-Implicit Crack Description

In this section, the XFEM approach in linear elastic fracture mechanics with a
hybrid explicit-implicit crack description is shortly discussed. Beside of phase-
field- or thick level-set-models [4, 15, 19], the XFEM offers a popular approach
for linear elastic fracture problems as it allows the approximation of inner-element
discontinuities, singularities, kinks and other non-smooth features within elements
SIFs Through CODs in the XFEM 145

with optimal accuracy. Therefore, a priori knowledge about the solution character-
istics is introduced to the problem using enrichment functions.
Starting point is a domain ˝ with the boundary u where the displacements are
prescribed, and the boundary t where the tractions are prescribed. The domain is
cracked by the (curved) crack path/surface c which may be stress-free or loaded
as in applications of hydraulic fracturing. A hybrid explicit-implicit description of
the crack is used which is explained in more detail in Sect. 2.1. Cracks feature
discontinuous displacements along c and singular stresses at the crack tip/front.
An illustration of the situation is given in Fig. 1 for a two-dimensional case.
Additional enrichment functions which extend the standard finite element
approach incorporate the discontinuities and singularities in the displacement and
stress fields. For linear elastic fracture problems, the enriched approximation of the
XFEM [17] is given by

X X 4 X
X j j
uh .x/ D Ni .x/ui C Ni .x/ step .x/ai C Ni .x/ tip .r; /bi : (1)
I I jD1 J 

The first term of the right side represents the standard finite element approach, where
Ni are the finite element shape functions. The discontinuities in the displacement
field are considered by the additional enrichment function step and the singular
j
stresses at the crack tip/front are considered by the four enrichment functions tip .
Additional degrees of freedom (ai ; bi ) are introduced at the enriched nodes I  and J 
j

[17]. In linear elastic fracture mechanics it is standard to use the Heaviside function
for the step enrichment step [17] and
 
j p  p  p  p 
tip .r; / D rsin ; rcos ; rsin sin; rcos sin (2)
2 2 2 2

for the crack tip/front enrichment. These enrichment functions are based on a polar
coordinate system (r; ) which is aligned with the tangent at the crack tip/front and
has its origin at the crack tip/front. The definition of this coordinate system is based
on three level-set functions which is discussed in Sect. 2.1.

Fig. 1 Situation of a cracked

domain in two dimensions
146 M. Schätzer and T.-P. Fries

2.1 Hybrid Explicit-Implicit Crack Description

The explicit-implicit crack description is recalled in this section. The aim is to

combine an explicit crack description by means of straight line segments in two
dimensions or flat triangles in three dimensions, which simplifies the crack update
during propagation, and an implicitly defined crack by means of level-set functions.
In Fig. 2a, an example for a three-dimensional explicitly defined crack is illustrated.
Simulations with the XFEM typically use level-set functions to determine the
enrichments. The definition of the enriched nodes (I  and J  ) is also based on
level-set functions, wherefore they are derived from the explicitly defined master
configuration. In [10], three level-set functions j are introduced which are the basis
for the crack location and the enrichment functions in the XFEM. The three level-set
functions are defined as follows:
• 1 .x/ is the (unsigned) distance function to the crack path/surface.
• 2 .x/ is the (unsigned) distance function to the crack tip/front.
• 3 .x/ is a signed distance function to the extended crack path/surface.
An illustration of these functions is presented in Fig. 2b–d, where the corresponding
level-sets for a three-dimensional crack geometry are illustrated. Furthermore, these
functions imply two coordinate systems (r,) and (a,b), which are required for
the enrichment functions (Sect. 2) and the reference state for the fitting (Sect. 3.2).
These two coordinate systems are discussed in the following.

2.1.1 Coordinate System (a; b)

The (a,b)-coordinate system is defined by two scalar functions a.x/ and b.x/. The
correlation of the level-set functions and the coordinates is given by

b.x/ D 3 .x/ (3)

and
p
a .x/ D 2 .x/2  3 .x/2 : (4)

Fig. 2 (a) Explicitly defined crack and level-sets: (b) 1 D 1, (c) 2 D 1, (d) 3 D 0I 1I 2I 3 [10]
SIFs Through CODs in the XFEM 147

Fig. 3 (a) Coordinate system (a; b) in two dimensions and (b) zero level-sets of a.x/ and b.x/
implied by the three level-set functions [10]

As a .x/ only describes an unsigned distance its sign has to be adapted for a.x/,
see [10] for further details. Figure 3a shows the (a,b)-coordinate system in two
dimensions and (b) the zero level-sets of a.x/ and b.x/ in three dimensions. This
coordinate system simplifies the localisation of the enriched nodes [10]. However,
the enrichment functions themselves are based on the polar coordinate system (r; ),
wherefore the relation of these two coordinate systems is discussed in the following
section.

2.1.2 Coordinate System (r; )

The polar coordinate system (r; ) can be expressed by the three level-set functions
as shown in [10]. A similar setting is used, for example, in [23]. However, for the
computation of the expected openings in the reference configuration (Sect. 3.2), it
is preferred to express these coordinates by the two scalar functions a.x/ and b.x/.
The radius r is given by
p
r.x/ D a.x/2 C b.x/2 : (5)

For the definition of , several cases must be considered. With the limitation that
b.x/ ¤ 0, the angle .x/ is provided by
8 
ˆ
ˆtan1 b.x/ for: a > 0
ˆ
ˆ a.x/
ˆ
ˆ 1  b.x/
ˆ
<tan
a.x/
C for: a < 0
.x/ D (6)
ˆ
ˆ 
for: a D 0I b > 0
ˆ
ˆ
ˆ2
ˆ
:̂ 3 for: a D 0I b < 0:
2
148 M. Schätzer and T.-P. Fries

Fig. 4 Coordinate system

(r;  ) in 2D implied by the
(a; b) coordinate system [10]

A graphical representation of the polar coordinate system (r; ) in two dimensions

is found in Fig. 4.

3 Stress Intensity Factors Through Crack Opening

Displacements

The proposed method shows the computation of SIFs by observing the displacement
field in the vicinity of the crack tip/front. This method is intuitive and the extension
from two to three dimensions is straightforward and treats mode III as the other
modes. Furthermore, no modification is necessary for curved and loaded cracks. The
approximated displacements, which are extracted from the XFEM simulation, are
compared with the expected displacements for a pure mode I, II and III crack. That
is, two different states are involved which are described in more detail in Sects. 3.1
and 3.2. It is noted that any displacements in the vicinity of the crack tip/front
can be used for the computation of SIFs. However, translations by means of rigid
body motions have to be extracted from these displacements which complicates the
computation. Therefore, we prefer to use relative displacements
u, particularly
the crack opening displacements (CODs), as shown in Fig. 5 which consider rigid
body motions automatically. The relative displacement
uhCOD contains the first two
modes in two dimensions and all three modes in three dimensions, wherefore all
modes can be extracted from this opening. Section 3.3 specifies selected directions
where the best results can be expected. In the following sections, the computation
of the approximated and expected CODs is discussed in more detail. Additionally,
in Sect. 3.3 it is shown how to use these openings for the evaluation of SIFs.

3.1 Approximated State

h.S/
This section shows the computation of the approximated crack opening
uCOD of a
point S on the crack path/surface. Openings of the point S are computed after solving
the linear elastic fracture problem based on Eq. (1) as a post-processing step. The
SIFs Through CODs in the XFEM 149

Fig. 5 Crack opening in two

dimensions

difficulty here is that the enrichment and shape functions have to be evaluated in
a special setting so that CODs are obtained. For that, in a first step a point has to
be found on the implicitly described crack path/surface. It is hardly possible to do
this in a global setting as the level-set functions are only given at the nodes and
are interpolated within the two or three-dimensional elements by the corresponding
shape functions [10].
X
jh D Ni .x/j .xi /; j D 1; 2; 3 : (7)
i

This interpolation automatically leads to curved zero level-sets in general. However,

the following section shows a simplified detection of the zero level-sets. Starting
point is a reference quadrilateral or hexahedral element which is called ‘master ele-
ment’. This is decomposed into simplex elements which are called ‘sub-elements’
and allow for a simple detection of a planar zero level-set. It is noted that a similar
setting of the introduced technique is also used for the integration of the domain,
see e.g. [9, 11].

3.1.1 Localisation of the Zero Level-Sets

Each potentially cut master element, identified by the condition

maxŒ3 .xi /  minŒ3 .xi / < 0; (8)

is subdivided into two linear triangular elements in two dimensions or into six linear
tetrahedral elements in three dimensions to ensure that the zero level-set is piecewise
straight/planar. Then the level-set functions are interpolated within the sub-elements
150 M. Schätzer and T.-P. Fries

Fig. 6 Zero level-set in reference (a) triangular and (b) tetrahedral element

by the standard simplex shape functions NiS .x/. The intersections of the zero level-
set with the element edges are marked with xSi , in three dimensions this is: xSi D
ŒxSi ; ySi ; zSi T : Based on the roots on the sub-element edges, the zero level-set is
interpolated within the sub-element by the shape functions NiS .x/. This leads to a
straight or flat representation of the zero level-set in the sub-element, as illustrated
in Fig. 6. Only three different shapes of the zero level-set are possible: straight lines
in two dimensions and flat triangles or quadrilaterals in three dimensions. These
straight/flat elements represent the crack path/surface and are, hence, called ‘crack
elements’. We associate standard finite element shape functions NiC .x/ to them.

3.1.2 Computation of Approximated CODs

j
The crack elements are the basis for finding the points xS on the crack path/surface
as all points within these elements can be mapped to the sub-element by

j
X
xS D NiC .xj /xSi : (9)
i

j
As the displacement field is discontinuous on the crack surface, the point xS has
to be ‘splitted’ into the two opposite on the crack surface which are infinitesimally
close. We call one point SC on the positive side and the other one S on the negative
side. The sign of the sub-domain is defined by the sign of the third level-set function
j
3 .x/. This splitting is done in the sub-element by the normalized normal vector nS
as illustrated in Eq. (10).
j j j j
xS˙ D xS ˙ "nS with W jjnS jj D 1 (10)
SIFs Through CODs in the XFEM 151

Herein, " defines the shifting magnitude of the point. It should be small enough
j
so that the splitted points are as close as possible. The normalized normal vector nS
is defined by the cross product of two tangential vectors t1 and t2 ,

j t1  t2
nS D ; (11)
jjt1  t2 jj

see Fig. 6. The sign of the normal vector is chosen such that it always points to the
positive sub-domain, as shown in Fig. 7 for some example values of the level-set
function 3 .x/ in the corner nodes. There, the master element is illustrated by the
dashed line and the sub-element by the solid line. The green line/surface represents
the crack element, i.e. the zero level-set within the sub-element. The blue point
illustrates a point on the implicitly defined crack path/surface and the blue line its
normal vector pointing towards the positive sub-domain.
If the level-set functions and local coordinates (a,b) are known in the two
j j
opposite but infinitesimally close points xS˙ their polar coordinates r.xS˙ / and
j
.xS˙ / are determined with Eqs. (5) and (6). Then the approximated displacements
h.SC / h.S / h.S/
uCOD and uCOD are evaluated by Eq. (1). The approximated crack opening
uCOD
is described in the global coordinate system .x; y; z/ by the difference of both
displacements.

h.S/ h.SC / j h.S / j

uCOD .xS / D uCOD .xSC /  uCOD .xS / (12)

For the sake of clarity, the index COD is not written anymore, instead, the index is
used to give information about the current coordinate system, where .x/ describes
the global coordinate system and .a/ the coordinate system mentioned in Sect. 2.1.
This information is used for the evaluation of SIFs in Sect. 3.3 to indicate the current
coordinate system.

Fig. 7 Master and sub-elements in (a) two dimensions and (b) three dimensions and the identified
straight zero level-set
152 M. Schätzer and T.-P. Fries

3.2 Reference State for Pure Mode I, II and III Cracks

In this section, the reference state is discussed which is needed for the evaluation of
the expected CODs for a pure mode I, II and III crack, respectively. For straight or
planar crack geometries, the definition of the reference configuration can be based
on an orthonormal coordinate system which is aligned with the tangent at the crack
tip/front. However, in practical applications, cracks are often curved, especially for
propagating cracks where mode II is dominant. For such problems, the use of an
orthonormal coordinate system is not easily justified. We propose the definition of
a reference state based on the local coordinate system (a,b) which is also valid
for curved cracks. The definition of this coordinate system has been discussed in
Sect. 2.1.1. Figure 8a illustrates a two-dimensional curved crack (bold black line) in
the (a,b)-coordinate system.

3.2.1 Stress Intensity Approach

A satisfactory description of the crack tip/front behaviour is given by a linear

combination of three independent crack modes which are scaled by the SIFs kI ; kII
and kIII . This approach completely describes the state of the displacements, stresses
and strains in the vicinity of the crack tip/front. Under general mixed-mode loadings,
the displacements are given in the (a,b,c)-coordinate system by Anderson [1]:
2 3 2 3
kI p r 
 2
 kII p r 
 2

6ua 7 6 2 2 cos 2   1 C 2sin 2 C 2 2 sin 2  C 1 C 2cos 2 7
6 7 6 7
6 7 6k p r   kII p r  7
6ub 7 D 6 I  2  2 7
6 7 6 2 2 sin 2  C 1  2cos 2  2 2 cos 2   1  2sin 2 7
4 5 4 5
2kIII p r 
uc  2 sin 2
8 (13)
< 3 for plane stress
1C
with  D 2.1C/
E
and D
:3  4 for plane strain.

Fig. 8 Curved crack in the (a,b)-coordinate system in a (a) closed and (b) open setting
SIFs Through CODs in the XFEM 153

Herein, kI ; kII and kIII are the SIFs for mode I  III,  the second Lamé-Constant
and  a material parameter. The parameters r and  describe the polar coordinate
system at the crack tip/front, which has been mentioned in Sect. 2.1.2. As Eq. (13)
shows, displacements out of the plane only exist if mode III is non zero. However,
displacements in direction of a and b consist of a combination of mode I and II for
any 0 < jj < . It is noted that in three dimensions, only plain strain conditions
make sense and each point on the crack front has its own stress SIFs.

3.2.2 Computation of Expected CODs

Starting from the definition of the crack in the (a,b)-coordinate system, the expected
m.S/ j
CODs
ua of the point xS for a pure mode I, II and III are evaluated. The point
j j j
xS and its splitting in two opposite points xSC and xS is the same as in Sect. 3.1 for
j j
which the local coordinates a.xS˙ / and b.xS˙ / are already known. A representation
of the undeformed situation is presented in Fig. 8a. By using the relation of the
coordinates (a,b) and (r,) (Eqs. (5) and (6)) as well as the application of Eq. (13),
m.S˙ / m.S˙ / m.S˙ / m.S˙ / m.S˙ /
the expected displacements ua , with ua D Œua ; ub ; uc T , of
j j
the points xSC and xS are evaluated. Herein, m describes the current crack mode,
m.S˙ /
wherefore, for the expected openings ua , the stress intensity factor km is set to 1
and the others to 0. The expected COD of mode m is then given by the difference of
both points

/ C j /  j

um.S/
a D
um.S
a .xSC / 
um.S
a .xS /: (14)

In Fig. 8b, the expected opening (red line) for a pure mode I is presented in the (a,b)-
j
coordinate system. Furthermore, the expected COD of the point xS in direction of
b (blue line) is illustrated. Note that the expected openings are given in the (a,b,c)-
coordinate system. In the following section, the computation of SIFs based on the
evaluated CODs in the approximated and reference state is discussed.

3.3 Computation of SIFs Using CODs

A comparison of the approximated and expected openings is only possible if both

openings are described in the same coordinate system. Therefore, the approximated
CODs, which are given in the global coordinate system (x,y,z), are transformed into
the local coordinate system (a,b,c). In two dimensions, this is done based on the
154 M. Schätzer and T.-P. Fries

Jacobi-matrix J of the coordinate transformation, hence,

" #
a;x a;y

uha DJ
uhx with J D ; i D 1; 2: (15)
b;x b;y

It is noted that ra and rb are only orthonormal for straight/planar cracks. In three
dimensions there is no need to explicitly define a third function as rc is always
normal to ra and rb. Therefore, rc can be computed by the cross-product of ra
and rb.

rc D ra  rb (16)

With this information of the third direction, Eq. (15) is straightforwardly extended
to the third dimension.
2 3
a;x a;y a;z
6 7

uha D J 
uhx with J D 6 7
4b;x b;y b;z 5 ; i D 1; 2; 3: (17)
c;x c;y c;z

Now that both CODs are available in the same coordinate system, the comparison
of the approximated and expected CODs leads to the following system of equations:

uha D kI 
uIa C kII 
uIIa C kIII 
uIII
a: (18)

However, the impact of the individual SIFs to the displacement components of a

point which is ‘quasi’ on the crack surface with:  D ˙.  "/, is quite different
as Eq. (13) shows. Herein, kI mainly leads to displacements in direction of b, kII to
displacements in direction of a and kIII to displacements in direction of c. Therefore,
SIFs can be directly computed by

kI D
uhb =
uIb I kII D
uha =
uIIa I kIII D
uhc =
uIII
c : (19)

It remains to specify the location of the considered fitting points and the
consideration of the crack front in the context of a numerical simulation.

3.4 Location of the Fitting Points

It is well known that the stress intensity approach is only valid in the vicinity of the
crack tip/front, wherefore the maximum distance of the fitting points to the crack
SIFs Through CODs in the XFEM 155

tip/front is limited. Numerical results indicate that in two dimensions 10 % of the

crack length lc yield satisfactory results. However, in three dimensions there is no
‘classical’ crack length available, wherefore the maximum distance is limited by
10 % of an effective crack length leff
c . This effective crack length is described by the
ratio of the area of the crack surface Ac and the length of the crack front lcf :

Ac
c D
leff : (20)
lcf

j
The limitation of the maximum distance of the fitting point xS to the crack tip/front
can then be expressed as follows:
8
<0:1  lc in 2D
j
r.xS /  (21)
:0:1  leff in 3D.
c

For the evaluation of the partial derivatives of a.x/ and b.x/ it turned out that it is
preferable to use points within sign enriched elements rather than crack-tip enriched
elements. In coarse meshes or small crack geometries, it may happen that such
j j
points are hard to find. Then, the point xS is used where the distance r.xS / is a
minimum. Additionally, there exists a wide range of different crack configurations,
wherefore it is hardly possible to determine a fixed point on the crack path/surface
where the best results are obtained. Therefore, it is proposed to use a number of
points in this scope and compute averaged SIFs which also leads to an increased
robustness.
In Sect. 3.2.1, it has been mentioned that each point of the crack front has its
own SIFs to be determined by Eq. (19). However, for a numerical simulation it is
not necessary to know the SIFs of the whole crack front. As [10] shows, it is often
sufficient to know the behaviour at the explicitly defined nodes on the crack front.
Therefore, the following section investigates the situation at the crack front.

3.5 Consideration of the Crack Front

j
Each point xS on the implicitly defined crack surface generally describes the
behaviour of some related point on the crack front. The localisation of several points
which are representative for one special point at the crack front is quite difficult. It
j
is assumed that the point xS is valid for the point Fj on the crack front where it has
a minimum distance, see Fig. 9a. By using an explicit-implicit crack description,
a propagation is considered by adding new segments. Therefore, SIFs are needed
in the explicitly defined crack front nodes where the direction and distance of the
156 M. Schätzer and T.-P. Fries

Fig. 9 Three-dimensional crack front: (a) Scope of computed SIFs and (b) assignment to crack
front node

propagation is determined. It can be assumed that the change of the SIFs along the
crack front changes slowly for physically justified crack surfaces. Therefore, each
evaluated point Fj is assigned to the explicitly defined crack front node Kn where the
distance is a minimum, see Fig. 9b. It is noted that the distance of Fj and Kn is also
limited by 10 % of the effective crack length, so that only SIFs are considered in
the vicinity of the node. However, this limitation is only used for coarsely described
crack geometries.
In Fig. 9, a three-dimensional crack front (bold black line) is presented with
j
possible points xS on the crack surface (black stars). The assignment of these points
to the crack front is illustrated by the blue dashed lines and the scope of the crack
.K /
front node Kn is illustrated by the red line. The SIFs km n of node Kn related to mode
m is obtained by the average of all SIFs of the points which are assigned to node Kn .
This can be expressed by

X
N .F /
km j
km.Kn / D : (22)
jD1
N

SIFs are now known for each node of the crack front in the explicit crack
description.
SIFs Through CODs in the XFEM 157

4 Numerical Results

In this section, five test cases in linear elastic fracture mechanics are presented in
two and three dimensions. The first four examples investigate the behaviour of static
planar crack configurations with well known analytical or numerical solutions to
show the validity of the proposed method. Herein, SIFs are computed on different
meshes and are compared with the expected solutions. In the results, the ratio of the
computed and the expected SIFs are plotted over the used number of elements. The
first two examples are in two dimensions and refer to externally loaded domains with
different boundary conditions and stress-free crack surfaces. Loaded crack surfaces
are illustrated in examples three and four, where one is in two and the other in
three dimensions. A quasi-static two-dimensional crack propagation is presented
in the last example to show the accuracy of the proposed method also for curved
cracks. Each mesh is based on bilinear 4-node quadrilateral or trilinear 8-node
hexahedral elements. Furthermore, a brittle and isotropic material is used with a
Young’s modulus E D 35 GPa and a Poisson’s ratio  D 0:3 . In two dimensions,
plane stress conditions are assumed. The interest is only in the computation of SIFs
in brittle materials under quasi-static mixed-mode loadings, wherefore no dynamic
effects [14] or cohesive models [16] are considered herein.

4.1 Eccentric Three-Point Bending Test

This test case shows an edge cracked three-point bending test with an eccentric load,
where F D 100 kN and the eccentricity d D 75 cm. The beam is 600 cm long (l)
and 150 cm high (h) and exhibits an lc D 75 cm long initial crack in the middle, see
Fig. 10a. The expected SIFs are given [8] by
2 3 2 p 0
3
3Fl lc FI
6 kI 7 6 h2  3 7
6 7D6 .1 lhc / 2 7 : (23)
4 5 4 p 5
kII 3Fl lc
h2
 F II

0
With FI D 0:4010 and FII D 0:0876 , Eq. (23) leads to the expected SIFs kI D
69:64 and kII D 5:38 . Starting with a coarse mesh and 390 elements as shown in
Fig. 10b, SIFs are computed on seven different meshes, which are generated by a
refinement until a fine mesh with 67.470 elements is obtained. A local refinement
in the vicinity of the crack allows a limitation of the element number. Figure 10c
shows the normalized results of the computed SIFs for mode I (red) and II (blue).
In this test case, the proposed method provides results within 5 % for mode I on
any of the meshes. The results of mode II are a little bit worse, however the achieved
error is also limited to 10 % on the coarse mesh and improves upon refinement.
158 M. Schätzer and T.-P. Fries

Fig. 10 Eccentric three-point bending test in two dimensions (a) geometry parameters, (b) mesh
and (c) results

Fig. 11 Edge cracked rectangular plate (a) geometry parameters and (b) mesh

4.2 Shear Edge Crack

A shear loaded edge cracked rectangular plate is investigated next. The extent of
the plate is given by: h D 7 m, l D 16 m and exhibits an initial crack with a length
lc D 3:5 m, as shown in Fig. 11a. SIFs are computed on 12 different meshes. We
start with a coarse mesh (77 elements), which is locally refined at the crack, as
illustrated in Fig. 11b. This mesh is refined until 46.025 elements are obtained.
SIFs Through CODs in the XFEM 159

Fig. 12 Edge cracked rectangular plate (a) boundary and loading conditions and (b) results

The plate is clamped on the bottom and loaded by a shear traction  D 1 GPa on
the top as Fig. 12a shows. This configuration leads to a mixed-mode loading, where
the SIFs are given with kI D 34 and kII D 4:55 , see e.g. [2]. The obtained results
are illustrated in Fig. 12b, where the red line represents the normalized mode I and
the blue line the normalized mode II SIF.
This test case also achieves good results, where the error of kI is below 5 % for
all used meshes and shows a clear convergence during the refinement. Mode II has
again less accurate results, but also converges upon refinement.
In this two externally loaded test cases with stress-free crack paths, mode I SIFs
were evaluated within an error of 5 % and mode II SIFs within 10 %. The next two
cases investigate loaded crack surfaces which would lead to issues in the ‘classical’
interaction integral. A big advantage of the proposed method is that no modifications
are necessary as the following examples show.

4.3 Loaded Crack in 2D

In this example, the accuracy of the proposed method for loaded crack surfaces
is shown. Therefore, an edge cracked rectangular plate is loaded by a tension
 D 1 GPa and shear traction  D 1 GPa within the crack. The load situation
and boundary conditions are illustrated in Fig. 13a. This test case uses the same
geometry parameters and meshes as in Sect. 4.2. Due to the large l=h ratio it is
assumed that the impact of the boundary conditions is relatively small compared to
160 M. Schätzer and T.-P. Fries

Fig. 13 Loaded crack in two dimensions (a) boundary and loading conditions and (b) results

the behaviour of the crack. This assumption leads to following expected SIFs [25]:
2 3 2  3
p l2c l3c l4c
k
6 I7 6 lc 1:122  0:231 lc
h C 10:550 h2
 21:710 h3
C 30:382 h4 7
6 7D6 q  7:
4 5 4 2 3 5
 lclc 1:122 C 0:561 lhc C 0:085 hc2 C 0:180 hc3
l l
kII
1 h

(24)
By inserting the given geometry parameters into Eq. (24), the SIFs are given by
kI D 9:381 and kII D 4:151 . The ratio of the computed to the expected SIFs are
shown in Fig. 13b, where the red line represents mode I and the blue line mode II.
Figure 13b shows that the proposed method leads for most of the meshes to SIFs
within an error of 5 %. That is, the proposed method allows the computation of
SIFs also for loaded crack surfaces with a similar accuracy as for stress-free crack
surfaces.

4.4 Penny-Shaped Crack

Another test case without crack propagation is considered in three dimensions.

An embedded penny-shaped crack with tension and shear loaded crack surfaces
is examined. The individual components of the loading are given in the global
coordinate system (see Fig. 14b) by  D 1 GPa, x D y D p12 GPa. The crack
surface has a diameter d D 2 m and is explicitly described by 1.536 flat triangles and
the crack front by 64 line segments as shown in Fig. 14a. As shown in Fig. 14c, the
crack is located within a cube-like domain with a side length of 4 m. Displacements
are prescribed to zero at some corner nodes on the bottom. An illustration of
the situation is presented in Fig. 14. For this test case, four different meshes with
SIFs Through CODs in the XFEM 161

(a) (b) (c)

Fig. 14 Penny shaped crack (a) explicitly defined crack surface, (b) situation and (c) mesh

Fig. 15 Penny shaped crack (a) investigated point A and (b) results

trilinear hexahedral elements are used. The number of elements along an edge varies
between 13 and 25. Figure 14c shows an example mesh with 2.197 elements.
In this general example, SIFs vary along the crack front. The expected SIFs for
the whole front are given [25] by
2 3 2 q 3
2
k
6 I 7 6   d
2 7
6 7 6 q 7
6 7 6 4 7
6 kII 7 D 6 d7; (25)
6 7 6 .2/ .cos!/  2 7
4 5 4 q 5
4.1/
kIII .2/ .sin!/  d2

where ! describes the angle between the direction of the resultant shear traction
and the reviewed point. Here, the point A (see Fig. 15a) with an angle of ! D 4 is
observed and all three modes are present there. For that point, Eq. (25) leads to the
following SIFs: kI D 11:28 , kII D 9:39 and kIII D 6:57 . The normalized computed
SIFs of point A are presented in Fig. 15b.
162 M. Schätzer and T.-P. Fries

In this example, mode II and mode III are computed with a similar accuracy as
shown in Fig. 15b. Furthermore, the result shows that with only 173 elements, SIFs
can be computed with an error of less than 10 % and 213 elements lead to results
below 5 %. That is, SIFs are well obtained for a general three-dimensional crack
configuration with stresses on the crack surface without any modifications.

4.5 Crack Propagation in Two Dimensions

Until now, only cases with straight or flat crack geometries have been investigated.
To show the accuracy of the proposed method also for curved cracks, this final test
case shows a mixed-mode crack propagation in a square specimen with the extent
l D 1 m [3, 10], where curved cracks are generally expected. The displacements are
prescribed on the upper and lower side of the domain with 1 mm in direction of the
specified angle ˛, as illustrated in Fig. 16a. These boundary conditions produce an
opening of the crack, wherefore no other loadings are needed. The domain is dis-
cretized by a structured mesh. It is noted that our interest is only in the computation
of SIFs, wherefore no attention is given to the crack velocities. It is assumed that
the crack always propagates by the crack increment da in direction of the maximum
circumferential stresses. By using SIFs and the maximum circumferential stress
criterion, the propagation angle c can be expressed as follows [17]
0 s 1
2
1 @ kI kI
c D 2arctan ˙ C 8A ; (26)
4 kII kII

where the sign depends on the sign of kII .

(a) (b)
70
60
50

40
30
20
10
0
-10
-20
-3
= 15°, 165° -4 0
0
= 30°, 180°
-5
-60
0
-70

= 45°, 135°
= 60°, 120°
= 75°, 105°
= 90°

Fig. 16 Edge crack in a squared plate (a) geometry parameters and supports and (b) results of the
crack propagation
SIFs Through CODs in the XFEM 163

This test case starts with an initial crack of length lc D 0:5 m which is described
by one straight line segment. Then, the boundary conditions are prescribed and
the resulting SIFs are computed by the proposed method. With these SIFs the
propagation angle c is evaluated and a new line segment of the length da D 5 cm in
direction of c is added. This process is performed ten times. To keep the influence
of the element size to a minimum, a fine setting with 101  101 elements is used.
The prescribed direction of the boundary conditions is varied from 15  ˛  165.
In Fig. 16b, the results of the computed crack paths are presented.
As expected, for ˛ D 90 the crack propagates horizontally as for a pure tension
loading where only mode I is relevant. The greater the boundary conditions deviate
from a pure tension loading, the more dominant the impact of mode II is. The
obtained crack paths for the different ˛ are in good agreement to [3] so that it is
concluded that the proposed technique is also working for curved cracks.

5 Conclusions

In the presented work, it is shown how crack opening displacements (CODs) are
used for the computation of SIFs within the XFEM in two and three dimensions.
The use of CODs for computing SIFs provides an intuitive, computationally cheap
and robust method, which works in two and three dimensions in a consistent manner
no matter whether the crack surfaces are stress-free or loaded. The aim is to use this
approach also in simulations of hydraulic fracturing where stresses are exerted by a
fluid on the crack surface.
Two coordinate systems are used to simplify the definition of the enrichment
functions on the basis of level-set functions and describe a reference state, where
pure mode I, II and III openings are evaluated. These coordinate systems consider
curved cracks as well as planar ones. A comparison of the approximated states and
the reference states leads to the SIFs. An assignment of the fitting points to the
explicitly defined crack front nodes provides SIFs for each of those nodes. At the
crack tip/front, it is recommended to have rather small elements which considerably
improves the solutions in general. The accuracy and robustness of the proposed
method has been shown by five examples in two and three dimensions. For each test
case, results were achieved below 10 % error improving upon mesh refinement. All
obtained results are in good agreement with the available analytical, numerical and
experimental solutions so that it is concluded that the proposed method provides
good results for any crack configurations.

References

1. Anderson, T.L.: Fracture mechanics: fundamentals and applications. CRC, Boca Raton (2005)
2. Belytschko, T., Black, T.: Elastic crack growth in finite elements with minimal remeshing. Int.
J. Numer. Methods Eng. 45, 601–620 (1999)
3. Bourdin, B., Francfort, G., Marigo, J.: Numerical experiments in revisited brittle fracture. J.
Mech. Phys. Solids 48, 797–826 (2000)
164 M. Schätzer and T.-P. Fries

4. Cazes, F., Moës, N.: Comparison of a phase-field model and of a thick level set model for
brittle and quasi-brittle fracture. Int. J. Numer. Methods Eng. 103, 114–143 (2015)
5. Chan, S.K., Tuba, I.S., Wilson, W.K.: On the finite element method in linear fracture
mechanics. Eng. Fract. Mech. 2, 1–17 (1970)
6. Dolbow, J., Gosz, M.: On the computation of mixed-mode stress intensity factors in function-
ally graded materials. J. Solids Struct. 39, 2557–2574 (2002)
7. Duflot, M.: A study of the representation of cracks with level sets. Int. J. Numer. Methods Eng.
70, 1261–1302 (2007)
8. Fett, T.: Stress Intensity Factors and Weight Functions for Special Crack Problems, vol. 6025.
FZKA, Karlsruhe (1998)
9. Fries, T.: A corrected XFEM approximation without problems in blending elements. Int. J.
Numer. Methods Eng. 75, 503–532 (2008)
10. Fries, T., Baydoun, M.: Crack propagation with the extended finite element method and a
hybrid explicit-implicit crack description. Int. J. Numer. Methods Eng. 89, 1527–1558 (2012)
11. Fries, T., Belytschko, T.: The extended/generalized finite element method: an overview of the
method and its applications. Int. J. Numer. Methods Eng. 84, 253–304 (2010)
12. Gravouil, A., Moës, N., Belytschko, T.: Non-planar 3D crack growth by the extended finite
element and level sets - Part II: level set update. Int. J. Numer. Methods Eng. 53, 2569–2586
(2002)
13. Gray, L., Phan, A., Paulino, G., Kaplan, T.: Improved quarter-point crack tip element. Eng.
Fract. Mech. 70, 269–283 (2003)
14. Haboussa, D., Gregoire, D., Elguedj, T., Maigre, H., Combescure, A.: X-FEM analysis of the
effects of holes or other cracks on dynamic crack propagations. Int. J. Numer. Methods Eng.
86, 618–636 (2011)
15. Miehe, C., Hofacker, M., Welschinger, F.: A phase field model for rate-independent crack
propagation: robust algorithmic implementation based on operator splits. Comput. Methods
Appl. Mech. Eng.. 199, 2765–2778 (2010)
16. Moës, N., Belytschko, T.: Extended finite element method for cohesive crack growth. Eng.
Fract. Mech. 69, 813–833 (2002)
17. Moës, N., Dolbow, J., Belytschko, T.: A finite element method for crack growth without
remeshing. Int. J. Numer. Methods Eng. 46, 131–150 (1999)
18. Moës, N., Gravouil, A., Belytschko, T.: Non-planar 3D crack growth by the extended finite
element and level sets - Part I: mechanical model. Int. J. Numer. Methods Eng. 53, 2549–2568
(2002)
19. Moës, N., Stolz, C., Bernard, P.E., Chevaugeon, N.: A level set based model for damage growth:
the thick level set approach. Int. J. Numer. Methods Eng. 86, 358–380 (2011)
20. Nejati, M., Paluszny, A., Zimmerman, R.: On the use of quarter-point tetrahedral finite
elements in linear elastic fracture mechanics. Eng. Fract. Mech. 144, 194–221 (2015)
21. Nikishkov, G.P.: Accuracy of quarter-point element in modeling crack-tip fields. Comput.
Model. Eng. Sci. 93, 335–361 (2013)
22. Peng, D., Merriman, B., Osher, S., Zhao, H., Kang, M.: A PDE-based fast local level set
method. J. Comput. Phys. 155, 410–438 (1999)
23. Stolarska, M., Chopp, D., Moës, N., Belytschko, T.: Modelling crack growth by level sets in
the extended finite element method. Int. J. Numer. Methods Eng. 51, 943–960 (2001)
24. Sukumar, N., Moës, N., Moran, B., Belytschko, T.: Extended finite element method for three-
dimensional crack modelling. Int. J. Numer. Methods Eng. 48, 1549–1570 (2000)
25. Tada, H., Paris, P.C., Irwin, G.R.: The Analysis of Cracks Handbook. ASME Press, New York
(2000)
26. Walters, M., Paulino, G., Dodds, R.: Interaction integral procedures for 3-D curved cracks
including surface tractions. Eng. Fract. Mech. 72, 1635–1663 (2005)
 Part III
Polygonal and Polyhedral Methods
The Virtual Element Method for Underground
Flow Simulations in Fractured Media

Matías Fernando Benedetto, Stefano Berrone, and Andrea Borio

Abstract We focus on the problem of performing underground flow simulations

in fractured media. The medium is modelled by means of the so-called Discrete
Fracture Network (DFN) model. Within this framework, we discuss about the use
of the Virtual Element Method (VEM) in performing simulations, and about its
role in facilitating the meshing process. DFN models are characterized by a large
number of planar fractures stochastically generated and placed in a 3D reference
system starting from given probabilistic distribution of space position, dimension,
aspect ratio, orientation and hydrogeological properties. For these reasons, realistic
DFNs are usually characterized by the presence of very complex geometrical
configurations on which it is very difficult to obtain a good quality Finite Element
mesh. In particular, in order to circumvent these difficulties, we consider two
different VEM approaches: an approach based on a totally conforming polygonal
mesh obtained starting from a triangular mesh independently generated on the
different fractures of the system and an approach based on a partially conforming
mesh in which the connection between the fracture meshes is imposed through a
mortaring approach borrowed from the classical domain decomposition methods.
Concerning these two approaches, we discuss their relative advantages and present
numerical results aimed at comparing them.

1 Introduction

The focus of the present contribution is on underground flow simulations in

fractured media. This is a quite relevant issue in several present critical applications,
concerned with the long term analysis of geological systems subject to underground
exploitation.

M.F. Benedetto
Laboratorio de Métodos Numéricos en Ingeniería, Facultad de Ingeniería,
Universidad de Buenos Aires, Av. Las Heras 2214, C1127AAR, Buenos Aires, Argentina
e-mail: [email protected]
S. Berrone () • A. Borio
Dipartimento di Scienze Matematiche, Politecnico di Torino,
Corso Duca degli Abruzzi, 24, 10129 Torino, Italy
e-mail: [email protected]; [email protected]

© Springer International Publishing Switzerland 2016 167

G. Ventura, E. Benvenuti (eds.), Advances in Discretization Methods,
SEMA SIMAI Springer Series 12, DOI 10.1007/978-3-319-41246-7_8
168 M.F. Benedetto et al.

A possible representation of a fractured medium is given by the so-called

Discrete Fracture Network (DFN) model [1, 23, 26]. Assuming that the surrounding
rock matrix is impervious, according to DFN model, the fluid flows along fractures
and through fractures intersections. Most methods proposed in the recent literature
concerning flow simulations in fracture networks ask for some kind of mesh
conformity at fracture intersections (traces), in order to ensure a correct represen-
tation of the flux exchange between fractures. This (possibly partial) conformity
requirement often results in a fully coupled meshing process [25, 30, 36–38, 41],
which is quite often very difficult to be performed or even infeasible, in some
critical configurations, which are nevertheless likely to be encountered, especially
in the framework of uncertainty quantification analysis [18]. To remove these
conformity constraints, an optimization based method was proposed in [14–16, 20]
and developed in [9, 17, 19]. More details on this optimization approach can be
found in [39]. A different approach not involving mesh generation is described in
[24, 34, 35]. More complex models on the fractures can be found in [27–29, 31].
In the present work, we discuss on the use of the rather new Virtual Element
Method for performing simulations in fracture networks. After introducing some
notation in the sequel of this section, we will describe the problem formulation in
Sect. 2; in Sect. 3 we will discuss the use of VEM in this framework, focusing on two
possible approaches. Finally, we propose in Sect. 4 some numerical results aimed at
comparing the two approaches.

1.1 Discrete Fracture Networks

Each fracture in the underground is represented as a (2D) open polygon. In what

follows, we will identify the geological fracture and its representing polygon, and
we will refer also to the polygons as fractures. Accordingly, a DFN ˝ is represented
as a (3D) set given by the union of N fractures Fi , with i 2 I D f1; : : : ; Ng. We
also introduce the set @˝ D [i2I @Fi . Fractures intersections are called traces and
denoted by m , with m 2 M D f1; : : : ; Mg (see Fig. 1). Without loss of generality,
we assume that the set ˝N is connected. For the sake of simplicity, we also assume
that each trace is given by the intersection of precisely two fractures. For future
reference, we introduce the following notation:
• 8i 2 I, we introduce the subset Mi  M of the indices of traces lying on Fi ;
each subset Mi is assumed to be ordered, and we will denote by Mi .k/ the k-th
index of a trace in Mi ;
• 8m 2 M, we introduce the couple Im D .i; j/ of indices such that m D Fi \ Fj ;
the couple is assumed to be ordered in such a way that i  j;
• for each i 2 I and each m 2 Mi , we fix a unit vector nO im normal to m on Fi .
The VEM for Underground Flow Simulations in Fractured Media 169

Fig. 1 Example of two ...........

...........
...........
fractures (Fi and Fj ) Fj ...........
...........
...........
...........
...........
...........
intersecting and generating a ...........
...........
...........
...
trace ( m ) .... .....
... .. .....
..... . ..
...
. .... ..... ..... ...
..
.....
....................... .....
................
................. .. ...... .....
................. ...... ......
..................
.....
.....
................
Fi
...
.
.
..
......
......
...... Ҕm
......
......
......
......
......
..... ...... ......
..... ... ...... ......
..... ...... ......
..... .
... .
........ ......
.....
..... ...... ................
..... .
. ......
.. ..
.. ...
...........
..... . ...... . .......
..... . .
....... .. ..... ....
..... ..
..... ............
..... .............
..... ... ............
..... ...... .............
..... .
.... . ..... .................
..... ......
......................................
.........
.........
.........
.........
.........
.........
.........
.........
.........
.........
.........
.......

1.2 Further Notation

Given an arbitrary set !, we denote by .; /! and kk! the L2 .!/ scalar product and
norm, respectively; furthermore, we let .; /˛;! and kk˛;! denote the H˛ .!/ scalar
product and norm, respectively. For any function defined on ˝, we use a subscript
i to indicate its restriction to fracture Fi . Let D
@˝ be the portion of boundary
on which Dirichlet conditions are imposed, and let hD be the Dirichlet condition
defined on D . We define the functional spaces
˚ 
Vi WD v 2 H1 .Fi / W  D .v/ D 0 8i 2 I ;
˚ 
ViD WD v 2 H1 .Fi / W  D .v/ D hDi 8i 2 I ;
V WD fvW vi 2 Vi 8i 2 Ig ;
˚ 
V D WD vW vi 2 ViD 8i 2 I :

For any segment   Fi , i 2 I, we introduce the trace operator  W H1 .Fi / !

1
H ./ and the notation
2

1 1
h; ˇi WD 1 h; ˇi 1 ; 8 2 H 2 ./ ; ˇ 2 H 2 ./ ;
H 2 . / H 2 . /

1 1
to denote the duality product between H 2 ./ and H 2 ./. Let v 2 V. In order
to simplify the notation, it is convenient to introduce the vectors Mi .v/, 8i 2 I,
the k-th element of Mi .v/ being  Mi .k/ .v/. Furthermore, we introduce the jump
across a trace m as
 
v m WD  m .vi /   m vj ; if Im D .i; j/ ;
170 M.F. Benedetto et al.

and we introduce the symbols vM and vMi to denote the vectors of jumps of v
across all traces in the network, and across traces on Fi , respectively. With the same
purpose, 8i 2 I, we introduce the notation
X Y 1 Y 1
h; ˇiMi WD hm ; ˇm i m ; 8 2 H 2 . m / ; ˇ 2 H 2 . m / :
m2Mi m2Mi m2Mi

Q 1
Finally, for any ˇ;  2 m2Mi H 2 . m / we denote
X
.ˇ; /Mi WD .ˇ; / m :
m2Mi

2 Problem Formulation

We are interested in computing the hydraulic head h D =.g/ C z 2 V D , where

 is the fluid pressure, g the gravitational acceleration,  the fluid density and z
the elevation. The hydraulic head, on each fracture Fi , is modeled by means of the
Darcy law as follows. Let Ki denote the transmissivity on Fi , which we assume to
be constant, and fi D fi .x/ denote the source term on Fi ; notice that both Ki and fi
are functions of the local tangential coordinate system.
The problem on each fracture is: find hi 2 ViD such that, 8vi 2 Vi ,

*  +
D E @hi
.Ki rhi ; rvi /Fi D . fi ; vi /Fi C hNi ;  i N .vi / C ; Mi .vi /
i iN @OnMi Mi Mi
(1)
 
@hi @hi
where @Onim
D Ki rhi  nO im along nO im ;
is the jump of the co-normal derivative @Onim
m
1  
furthermore, iN
@Fi is the Neumann boundary on Fi and hNi 2 H 2 iN is the
Neumann boundary condition. For future reference, we set N D [i2I iN  @˝
and define hN such that hNi is the restriction of hN to Fi .
The problems on each fracture are coupled by the imposition of suitable matching
conditions which guarantee the continuity of the solution and balance of incoming
and outgoing fluxes at each trace: 8m 2 M, with Im D .i; j/, we have
 
h m D  m .hi /   m hj D 0;
   
@hi @hj
C D 0: (2)
@Onim m j
@Onm m
The VEM for Underground Flow Simulations in Fractured Media 171

3 The Virtual Element Method for DFN Simulations

First introduced in [3] and extended in [2, 4–7, 22], the Virtual Element Method
allows the use of any kind of non-degenerate star-shaped polygon to mesh the
spatial domain, even including the possibility of straight angles. In the present
framework, we take advantage from this flexibility to easily build a mesh which,
on each fracture, is locally conforming to the traces. In the following of this section,
we review the use of VEM, focusing on the framework of DFN simulations.

3.1 The VEM Setting in the DFN Framework

Let us fix a fracture Fi . To obtain a locally conforming mesh, we first introduce on

Fi a triangular mesh built independently of trace positions; the triangles are then
cut into polygons by the traces, possibly prolonging the trace segment up to the
nearest mesh edge if it happens to end in the interior of a triangle. Note that in
this latter case the trace tip is kept as a node of the discretization, a new node is
created at the intersection between the prolongation of the trace and the mesh edge,
and therefore two edges are created, with a 180ı angle between them. Let Tıi be
the resulting local mesh. Referring to the configuration already depicted in Fig. 1,
we report in Figs. 2 and 3 an example of the VEM meshes Tıi and Tıj obtained on
the two fractures Fi and Fj , respectively. Furthermore, we notice that all polygons
created with the above procedure are convex, thus satisfying the assumptions in [3],
that require the elements to be star-shaped with respect to a ball. Let Tı D [i2I Tıi ,
being ı a global mesh parameter, e.g. the maximum element diameter. We will use
the symbols Eıi and Vıi to denote the sets of edges and vertices on fracture Fi ,
respectively; similarly, EEıi and VEıi denote sets of edges and vertices on the element
E 2 Tıi , and define Eı D [i2I Eıi , Vı D [i2I Vıi .
Let k 2 N be the VEM order, let i 2 I and consider an element E 2 Tıi . In order
to define the Virtual Element space of order k for the DFN, let us define the local
finite dimensional functional space
˚ 
VıiE WD v 2 H1 .E/ W e .v/ 2 P k .e/ 8e  @E;
vjE 2 P k2 .E/ ; ; (3)

Fig. 2 VEM mesh Tıi on Fi

172 M.F. Benedetto et al.

Fig. 3 VEM mesh Tıj on Fj

and let the virtual elements space on Fi be

˚ 
Vıi WD vı 2 C0 .Fi / W v 2 VıiE 8E 2 Tıi :

Finally, we introduce the following global discrete subspace of V:

Vı WD fvı 2 VW v 2 Vıi 8i 2 Ig : (4)

It can be proved (see [3]) that a possible set of degrees of freedom that uniquely
define a function vı 2 Vı is given by:
• the values of vı at each vertex V 2 Vı ;
• if k > 1, the values of vı at k  1 internal points (e.g. internal Gauss-Lobatto
quadrature nodes) on each
1
R edge e 2 Eı ; 2
• if k > 1 the moments jEj E vh m˛ for j ˛ j  k  2, where ˛ D .˛1 ; ˛2 / 2 N and

˛1 ˛2
x  xE y  yE
m˛ .x; y/ WD ;
hE hE

where .xE ; yE / and hE are the centroid and the diameter of the element E,
respectively. As a basis of Vı , we consider the functions k , k 2 f1; : : : ; Ng which
are lagrangian with respect to the above degrees of freedom.
We now turn our attention to Eq. (1), assuming that the transmissivity coefficient
is constant on each element of the mesh. Following [3, 6], in order to discretize
The VEM for Underground Flow Simulations in Fractured Media 173

this equation using functions in Vı , since the virtual functions are not known in the
interior of the elements, we introduce, for each i 2 I and E 2 Tıi , the operator
˘Er W VıiE ! P k .E/ that associates, to each  2 VıiE , the polynomial ˘Er  such that
8 
ˆ
ˆ Ki r˘Er ; rp E D .Ki r; rp/E 8p 2 P k .E/ ;
<P P
r
E ˘ .V/ D
ˆR V2Vıi E R V2VEıi .V/ if k D 1;
:̂ ˘ r v D v if k > 1:
E E;k ı E ı

We remark that the application of this operator only requires the knowledge of the
degrees of freedom of , through the application of Green’s formula (see [5]). In
addition to ˘Er , for each E 2 Tıi , let SE W VıiE VıiE ! R be a symmetric bilinear form
defined in such a way that there exist two positive constants c and c independent
on E and i such that 8 2 VıiE , if ˘Er  D 0, then

c .Ki r; r/E  SE .; /  c .Ki r; r/E : (5)

With the above ingredients, we define the discrete bilinear form aEı W Vı  Vı ! R
such that, 8v; w 2 Vı ,
   
aEı .v; w/ WD Ki r˘Er v; r˘Er w E C SE v  ˘Er v; w  ˘Er w :

Thanks to (5), we easily find that aEı .v; v/ scales like .Ki rv; rv/E , with scaling
constants independent of E and of the fracture index i. Thus, if we introduce the
fracture-wide bilinear form aıi W Vıi  Vıi ! R such that
X
aıi .v; w/ WD aEı .v; w/ ;
E2Tıi

we have the following property:

9˛ ; ˛  > 0W ˛ .Ki rv; rv/Fi  aıi .v; v/  ˛  .Ki rv; rv/Fi : (6)

Moreover, since the scalar product . f ; vı / is not computable in general if vı is a

virtual function, we define the discrete scalar product
 
. fi ; vıi /ı;Fi WD fi ; ˘Q k0 vıi Fi 8i 2 I ;

where the pseudo-projection ˘Q k0 is defined, as done in [5], by local projections,

r
recovering from ˘E;k vıi the missing information about vıi :
( 
˘Q k0 vıi ; p E D .vıi ; p/E 8p 2 P k2 .E/ ;
8E 2 Tıi ;  0   r 
˘Q k vıi ; p E D ˘E;k vıi ; p E 8p 2 P k .E/ n P k2 .E/ :
174 M.F. Benedetto et al.

Finally, to ease the notation in the following it is convenient to define the global
discrete products
X
aı .v; w/ WD aıi .v; w/ 8v; w 2 Vı ;
i2I
X
. f ; vı /ı WD . fi ; vıi /ı;Fi 8v 2 Vı :
i2I

Remark 1 As suggested in [3, 6], a possible choice for the stabilization term SE is
given by the scalar product between the vectors containing the degrees of freedom of
the two arguments on the element. This choice guarantees property (5) under some
basic regularity assumptions on the triangulation. This choice has been adopted
in [9].

3.2 Formulation of the Problem Towards Domain

Decomposition

For each m 2 M, with Im D .i; j/, we define the function s m W Im ! f0; 1g such that

s m .i/ D 1; s m . j/ D 0;
Q 1
and the bilinear form bi W Mi WD m2Mi H 2 . m /  Vi ! R such that
X
bi .v; / WD .1/s m .i/ h m ;  m .vi /i m ;
m2Mi

in such a way that

X X ˝ ˛
b .v; / WD bi .v; / D m ; v m :
m
i2I m2M

Q 1
We define the functional FW V  M WD m2M H 2 . m / such that

X1 D E
F.v; / WD .Ki rvi ; rvi /Fi  . fi ; vi /Fi  hNi ;  i N .vi /
2 i iN
i2I
      D 
i ; rvi Fi C bi .v; / C bi Ri h ;
C rRi hD ;
The VEM for Underground Flow Simulations in Fractured Media 175

1
where Ri is the lift operator from H 2 .Fi / to H1 .Fi /, i 2 I. It is well known (see
[40]) that solving problem (1)–(2) is equivalent
  to solve the problem of finding
.h; / 2 V D  M such that hi D h0i C Ri hD , h0i 2 Vi , i 2 I and
 
F h0 ;  D min max F .v; / ; (7)
v2V 2M

that, by uniqueness of the solution, implies

   
@hi @hj
m D D ;
@Onim m
j
@Onm m

with Im D .i; j/. Notice that the functional F is made up of local contributions
from each fracture. We will now present two different approaches for discretizing
problem (7).

3.3 A Globally Conforming Approach

A first approach to tackle problem (7), introduced in [10], is to strongly impose

the matching conditions by building a globally conforming mesh. This is easily
achieved by exploiting the capability of the Virtual Element Method to handle
straight angles. Consider the mesh Tı constructed as depicted in Sect. 3.1 and let us
consider an arbitrary trace m , with m 2 M and Im D .i; j/. Then, we add to Tıi the
nodes generated by Tıj on m , and vice-versa. Some polygons belonging to mesh
Tıi (Tıj , respectively) having an edge lying on m , will possibly have such edges
split by the new nodes, the new edges forming a straight angle at their intersection.
In the same configuration of Figs. 1, 2 and 3, we show in Figs. 4 and 5 the globally
conforming VEM meshes on Fi and Fj , respectively.
gc
We will call Tı this new set of globally conforming polygons and discretize
gc
problem (7) with the Virtual Element Method as previously described. Let Vı  V
gc
be the Virtual Element space defined on Tı considering, for each trace node, two
different degrees of freedom, each one associated with one of the two fractures
gc
intersecting there. In other words, we admit that functions in Vı could be
discontinuous on each trace, even though on each trace their degrees of freedom

Fig. 4 Globally conforming

VEM mesh on Fi
176 M.F. Benedetto et al.

Fig. 5 Globally conforming

VEM mesh on Fj

correspond to the same geometrical point. To discretize the space M we define, for
each i 2 I and on each trace m such that m 2 Mi , the finite dimensional space
gc ˚ 
ki ; k D 1; : : : ; N m ;
Mım;i D span m

where N m is the number of interior nodes on m and m

ki is a continuous linear
operator such that
˝  ˛ gc
ki ;  m vıj m D ıij vıi .xk /
m m
8vı 2 Vı ; (8)

being ıij the KroneckerQdelta and xm

k the k-th node on trace m . The space that
gc gc
discretizes M is Mı D m2M Mım , where
gc ˚ 
k W k D ki  kj if Im D .i; j/ ; k D 1; : : : ; N m :
Mım WD m m m m

With this definition we have, for any m 2 M, if Im D .i; j/, 8k 2 f1; : : : ; N m g,

˝ m ˛ ˝ ˛ ˝ m  ˛
k ; vı  m D mki ;  m .vıi / m  kj ;  m vıj D vıi .xk /  vıj .xk /;
m m
m m
The VEM for Underground Flow Simulations in Fractured Media 177

gc
and therefore, enforcing orthogonality of the jumps on traces with respect to Mı is
gc
enough to obtain continuity on Vı . Indeed,
gc
b .vı ; ı/ D0 8 ı 2 Mı ” vı M D 0 :

Let us define, for i 2 I

 
hı D h0ı C Rı hD ;
 
where h0ı 2 Vı and Rı hD is the discrete lifting of the boundary conditions. For
gc

the sake of simplicity we neglect the approximation error of the Dirichlet boundary
condition. The globally conforming discrete solution .h0ı ; ı / 2 Vı  Mı satisfies
gc gc

the virtual element discretization of (7):

 
F h0ı ; ı D mingc maxgc F .vı ; ı/ ;
vı 2Vı ı 2Mı

which leads us to the following saddle point formulation:

8   ˝ ˛
ˆ
ˆ a h0 ; v C b .vı ; ı / D . f ; vı /ı C hN ;  N .vıi / N
< ı ı ı 8vı 2 Vı ;
gc
   
ˆ  C aı Rı hD ; vı
:̂b h0 ;      gc
ı D b Rı h ; ı 8 2 Mı :
D
ı ı

P
The above problem has a unique solution because vı 7! i2I .Ki rvıi ; rvıi /Fi is a
norm on

gc ˚ gc gc 
Wı WD vı 2 Vı W b .vı ; ı/ D0 8 ı2 Mı D
˚ gc 
D vı 2 Vı W vı  m D 0 8m 2 M ;

gc
and therefore aı is coercive on Wı thanks to (6), and that

b .vı ; ı /
8 ı 2 Mı ; sup Dk ı kMı ;
vı 2Vı
gc kvı kVı

being Mı a space of linear operators.

178 M.F. Benedetto et al.

3.3.1 Implementation
˚ gc dim V gc gc
Let k kD1 ı be the Lagrangian basis of Vı and define, for each i 2 I, the fracture
stiffness matrix Ai , such that
 gc gc 
.Ai /kl D aıi ki ; li :
gc
We note that, by (8), if l is a basis function associated to an internal node of the
mesh (always placed on the boundary of a VEM element),
(
 gc  .1/s m .i/ if xl D xm
k;
bi l ; m
k D
0 otherwise;

gc gc
where xl is the node associated to l . If l is a basis function associated to one
of the polygon internal degrees of freedom, bi l ; m k D 0 8i 2 I; m 2 Mi ; k 2
f1; : : : ; N m g. To collect the terms coming from the bilinear form b, we define a
global numbering of the degrees of freedom on all the traces and, for each m 2 M,
with Im D .i; j/, we identify the row vector Bm such that .Bm /k D 1 if the k-th trace
degree of freedom is on trace m and fracture i and .Bm /k D 1 if the k-th trace
degree of freedom is on trace m and fracture j. Then the vector h containing the
degrees of freedom of hı is the solution of
  
A BT h f
D ; (9)
B 0  d

where  is a vector of Lagrange multipliers, f is the vector containing the right-

hand-side terms, d the vector of nodal values of hD on the traces and
0 1
A1 0 1
B A2 C B1
B C B C
A WD B :: C; B WD @ ::: A :
@ : A
BM
AN

Using classical results (see e.g. [33]) it is easily proven that system (9) has a unique
solution. Moreover, this reformulation falls into the framework of domain decom-
position methods [43]; this property is exploited in [10] to devise a preconditioned
one-level FETI method [32] for its solution.
The VEM for Underground Flow Simulations in Fractured Media 179

3.4 A Hybrid Mortar Virtual Element Approach

The second approach we present here, developed in [11], consists in imposing a

weak continuity of the solution, by applying the mortar element method [8, 12, 13].
We consider the locally conforming mesh Tı defined in Sect. 3.1 and, for each
m 2 M, with Im D .i; j/, we introduce a finite dimensional space Mım  L2 . m /
definedQ on the discretization of the trace induced by Tıi . We discretize (7) on Vı 
Mı D m2M Mım , where Vı is the Virtual Element space defined  by (3)–(4). The
mortar formulation of the problem is: find hı D h0ı C Rı hD , with h0ı 2 Vı and
ı 2 Mı such that,
8    
ˆ
ˆ a h0 ; v C b .vı ; ı / D . f ; vı /ı C hN ; vı N
< ı ı ı 8vı 2 Vı ;
  D 
ˆ   a ı R ı h ; vı (10)
:̂b h0 ;  D b R hD ;  8 ı 2 Mı :
ı ı ı ı

In practical implementations, Mım will be a piecewise polynomial space (see

[12, 42]). The well-posedness of (10) is proved in [11] under the hypothesis that
Mı contains the constant functions and under the following regularity assumption.
Assumption 1 There exists a constant  > 0 independent of ı such that, 8E 2 Tı ,
the distance between any two vertices of E is larger than or equal to hE , where hE
is the diameter of E.
P
Indeed, since vı 7! i2I .Ki rvıi ; rvıi /Fi is a norm on

Wı WD fvı 2 Vı W b .vı ; ı/ D0 8 ı 2 Mı g ;

by (6), aı is coercive of Wı . Moreover, by classical arguments (see [21, 40]), since

VEM functions are piecewise polynomials on traces and Assumption 1 allows us
to build a regular triangulation inside each polygon, there exists a constant ˇ > 0
independent of ı such that the following inf–sup condition holds:

b .vı ; ı/
inf sup  ˇ:
ı 2Mı vı 2Mı kvı kVı k ı kMı

An advantage of the present method, as opposite to the globally

 conforming
 case,
@hi
is that ım is now a piecewise polynomial approximation of @Oni , whereas in
m m
the previous case the latter is obtained by evaluating the gradient of the numerical
solution, and this may yield less accurate approximations.
180 M.F. Benedetto et al.

3.4.1 Implementation

As previously sketched, the space Mım , m 2 M, will contain piecewise polynomial

(not necessarily continuous) functions, defined on the discretization of the trace m
induced by Tıi , with Im D .i; j/. Let Nh and N be the total number of degrees of
freedom of h0ı and ı , respectively. The vectors h and  containing the degrees of
freedom are the solution of the system
  
A BT h f
D ;
B 0  d

where A 2 RNh Nh is defined, as in Sect. 3.3.1, as the block-diagonal matrix of the
local stiffness matrices and B 2 RN Nh collects the terms coming from b:

Blk WD b .k ; l/ 8l 2 f1; : : : ; N g; k 2 f1; : : : ; Nh g;

being k the k-th basis function of Vı and l the l-th basis function of Mı . Regarding
the right-hand-side, we have
     
fk WD . f ; k /ı C hN ; k N  aı Rı hD ; l 8k 2 f1; : : : ; Nh g;
   
dl WD b Rı hD ; l 8l 2 f1; : : : ; N g:

4 Numerical Results

While referring the reader to [10, 11] for numerical results that validate the two
proposed approaches, here we focus on a comparison between the two.
We consider two DFNs. The first example is a more simple DFN (Fig. 6a)
presenting all possible configurations (tips, three fractures intersection point and
traces spanning all the fracture) for which we have an analytical solution, reported
in [11]. The second one (Fig. 6b) is a more complex DFN stochastically generated,
containing 27 fractures and all common geometrical complexities. More details can
be found in [10].
The VEM for Underground Flow Simulations in Fractured Media 181

Fig. 6 (a) Geometry of the 3 fracture DFN. (b) Geometry of the 27 fracture DFN

Fig. 7 (a) Non-conformity of the fracture meshes to traces. (b) Detail of non-conformity at
fracture intersection.

Figure 7a displays a possible mesh configuration on the DFN in Fig. 6b. In

Fig. 7b a detail of the previous mesh is reported. We highlight that, as said in
Sect. 3.1, a triangular mesh is built independently on each fracture and then triangles
are cut in along the traces, thus obtaining convex polygons. In the case of the
globally conforming approach (Sect. 3.3), all the mesh points on the traces become
degrees of freedom of the VEM spaces defined on each fracture (each physical point
corresponds to two different degrees of freedom, one for each fracture). Instead,
the VEM-Mortar approach (Sect. 3.4) requires that the points on the traces become
degrees of freedom of the VEM space defined only on the fracture which they
belong to.
182 M.F. Benedetto et al.

3
globally conforming VEM
exact solution
2.5 VEM - Mortar

1.5

0.5

-0.5
0 0.2 0.4 0.6 0.8 1

Fig. 8 Superposition of the fluxes computed by the two methods and the exact solution on a trace
of the DFN in Fig. 6a

To compare the results obtained by the two methods, we focus on the fluxes
computed on traces. Using the Mortar approach with piecewise linear Lagrange
multipliers, we obtain an approximation of the flux at each trace from the solution
ı of the problem. In the globally conforming case, we obtain an approximation of
the flux by post-processing the discrete solution, first projecting it on polynomials
on each VEM element and then computing the jump of the co-normal derivative of
the projection.
In Fig. 8 we compare the fluxes obtained by the conforming and the Mortar
approaches to the exact solution for the trace with the tip (the vertical one) of the
DFN in Fig. 6a. We can see a good agreement between the fluxes.
In Fig. 9 we show this comparison on three different traces of the DFN in Fig. 6b,
from which we see that the two proposed approaches give comparable results.
In the Mortar approach the plot of the fluxes displays evident oscillations, due
1
to the fact that convergence to the exact solution is proven in the H 2 . m /-norm,
8m 2 M. Nevertheless, the quality of the approximation is usually more accurate
and reliable being the result of a direct computation and not of a post-processing
process.
The VEM for Underground Flow Simulations in Fractured Media 183

0.18 0.4

0.16 0.35

0.14 0.3

0.12 0.25

0.1 0.2
0.08 0.15
0.06 0.1
0.04 0.05
0.02 0
0 5 10 15 20 25 30 35 40 45 50 0 10 20 30 40 50 60 70 80

0.25

0.2

0.15

0.1

0.05

0
0 2 4 6 8 10 12 14 16 18

Fig. 9 Superposition of the fluxes computed by the two methods on three traces of the 27 fracture
DFN in Fig. 6b

5 Conclusions

We have considered the application of the Virtual Element Methods to underground

flow simulations. Among the several open and challenging issues stated by these
applications there are the geometrical complexities that can pose very difficult prob-
lems in the mesh generation process. These difficulties are very strong constraints
in the classical Finite Element approach to the simulations. The large flexibility
introduced by the VEM in the mesh generation is a key issue that can largely help
simulations.
We have discussed how it is possible to take advantage from this flexibility
in practical applications. Numerical examples here presented, and more detailed
numerical examples presented in [10, 11], clearly indicate the viability of the
approach.

Acknowledgements This work has been supported by the Italian MIUR through PRIN research
grant 2012HBLYE4_001 “Metodologie innovative nella modellistica differenziale numerica” and
by INdAM-GNCS. Fernando Matias Benedetto was supported by the European Commission
through the Erasmus Mundus Action 2-Strand1 ARCOIRIS programme, Politecnico di Torino.
Stefano Berrone acknowledges the partial financial support received from the European Unions
184 M.F. Benedetto et al.

Seventh Framework Programme (FP7/ 20142016) under Grant Agreement Number 607626
(Safeciti). Project title: “Simulation Platform for the Analysis of Crowd Turmoil in Urban
Environments with Training and Predictive Capabilities”. This publication reflects the views only
of the authors and the Commission cannot be held responsible for any use which may be made of
the information here contained.

References

1. Adler, P.M.: Fractures and Fracture Networks. Kluwer Academic, Dordrecht (1999)
2. Ahmad, B., Alsaedi, A., Brezzi, F., Marini, L., Russo, A.: Equivalent projectors for virtual
element methods. Comput. Math. Appl. 66, 376–391 (2013)
3. Beirão da Veiga, L., Brezzi, F., Cangiani, A., Manzini, G., Marini, L., Russo, A.: Basic
principles of virtual element methods. Math. Models Methods Appl. Sci. 23(01), 199–214
(2013)
4. Beirão da Veiga, L., Brezzi, F., Marini, L.: Virtual elements for linear elasticity problems.
SIAM J. Numer. Anal. 51(2), 794–812 (2013)
5. Beirão Da Veiga, L., Brezzi, F., Marini, L., Russo, A.: The hitchhiker’s guide to the virtual
element method. Math. Models Methods Appl. Sci 24(8), 1541–1573 (2014)
6. Beirão da Veiga, L., Brezzi, F., Marini, L., Russo, A.: Virtual element methods for general
second order elliptic problems on polygonal meshes. Available online at http://arxiv.org/abs/
1412.2646 (2014)
7. Beirão da Veiga, L., Lovadina, C., Mora, D.: A virtual element method for elastic and inelastic
problems on polytope meshes (2015). http://arxiv.org/abs/1503.02042. Available online
8. Belgacem, F.B.: The mortar finite element method with lagrange multipliers. Numer. Math.
84(2), 173–197 (1999)
9. Benedetto, M.F., Berrone, S., Pieraccini, S., Scialò, S.: The virtual element method for discrete
fracture network simulations. Comput. Methods Appl. Mech. Eng. 280(0), 135–156 (2014).
doi:http://dx.doi.org/10.1016/j.cma.2014.07.016
10. Benedetto, M.F., Berrone, S., Scialò, S.: A globally conforming method for solving flow in
discrete fracture networks using the virtual element method. Finite Elem. Anal. Des. 109,
23–36 (2016). doi:http://dx.doi.org/10.1016/j.finel.2015.10.003
11. Benedetto, M., Berrone, S., Borio, A., Pieraccini, S., Scialò, S.: A hybrid mortar virtual element
method for discrete fracture network simulations. J. Comput. Phys. 306, 148–166 (2016).
doi:http://dx.doi.org/10.1016/j.jcp.2015.11.034
12. Bernardi, C., Maday, Y., Patera, A.T.: A new nonconforming approach to domain decompo-
sition: the mortar element method. In: Nonlinear Partial Differential Equations and Their
Applications. Collège de France Seminar, vol. XI (Paris, 1989–1991), Pitman Res. Notes Math.
Ser., vol. 299, pp. 13–51. Longman Sci. Tech., Harlow (1994)
13. Bernardi, C., Maday, Y., Rapetti, F.: Basics and some applications of the mortar element
method. GAMM-Mitteilungen 28(2), 97–123 (2005)
14. Berrone, S., Pieraccini, S., Scialò, S.: A PDE-constrained optimization formulation for discrete
fracture network flows. SIAM J. Sci. Comput. 35(2), B487–B510 (2013). doi:http://dx.doi.org/
10.1137/120865884
15. Berrone, S., Pieraccini, S., Scialò, S.: On simulations of discrete fracture network flows with an
optimization-based extended finite element method. SIAM J. Sci. Comput. 35(2), A908–A935
(2013). doi:http://dx.doi.org/10.1137/120882883
16. Berrone, S., Pieraccini, S., Scialò, S.: An optimization approach for large scale simulations of
discrete fracture network flows. J. Comput. Phys. 256, 838–853 (2014). doi:http://dx.doi.org/
10.1016/j.jcp.2013.09.028
The VEM for Underground Flow Simulations in Fractured Media 185

17. Berrone, S., Pieraccini, S., Scialò, S., Vicini, F.: A parallel solver for large scale DFN flow
simulations. SIAM J. Sci. Comput. 37(3), C285–C306 (2015). doi:http://dx.doi.org/10.1137/
140984014
18. Berrone, S., Canuto, C., Pieraccini, S., Scialò, S.: Uncertainty quantification in discrete fracture
network models: stochastic fracture transmissivity. Comput. Math. Appl. 70(4), 603–623
(2015). doi:http://dx.doi.org/10.1016/j.camwa.2015.05.013
19. Berrone, S., Borio, A., Scialò, S.: A posteriori error estimate for a PDE-constrained optimiza-
tion formulation for the flow in DFNs. SIAM J. Numer. Anal. 54(1), 242–261 (2016)
20. Berrone, S., Pieraccini, S., Scialò, S.: Towards effective flow simulations in realistic discrete
fracture networks. J. Comput. Phys. 310, 181–201 (2016). doi:http://dx.doi.org/10.1016/j.jcp.
2016.01.009
21. Brezzi, F.: On the existence, uniqueness and approximation of saddle-point problems arising
from lagrangian multipliers. Rev. Française Automat. Informat. Recherche opérationelle. Anal.
Numér. 8(2), 129–151 (1974)
22. Brezzi, F., Falk, R.S., Marini, D.L.: Basic principles of mixed virtual element methods. ESAIM
Math. Modell. Numer. Anal. 48(04), 1227–1240 (2014)
23. Cammarata, G., Fidelibus, C., Cravero, M., Barla, G.: The hydro-mechanically coupled
response of rock fractures. Rock Mech. Rock. Eng. 40(1), 41–61 (2007). doi:http://dx.doi.
org/10.1007/s00603-006-0081-z
24. Dershowitz, W.S., Fidelibus, C.: Derivation of equivalent pipe networks analogues for three-
dimensional discrete fracture networks by the boundary element method. Water Resour. Res.
35, 2685–2691 (1999). doi:http://dx.doi.org/10.1029/1999WR900118
25. de Dreuzy, J.R., Pichot, G., Poirriez, B., Erhel, J.: Synthetic benchmark for modeling flow in
3D fractured media. Comput. Geosci. 50(0), 59–71 (2013)
26. Fidelibus, C., Cammarata, G., Cravero, M.: Hydraulic characterization of fractured rocks. In:
Abbie, M., Bedford, J.S. (eds.) Rock Mechanics: New Research. Nova Science Publishers Inc.,
New York (2009)
27. Formaggia, L., Fumagalli, A., Scotti, A., Ruffo, P.: A reduced model for Darcy’s problem in
networks of fractures. ESAIM Math. Modell. Numer. Anal. 48, 1089–1116 (2014). doi:http://
dx.doi.org/0.1051/m2an/2013132
28. Formaggia, L., Antonietti, P., Panfili, P., Scotti, A., Turconi, L., Verani, M., Cominelli, A.:
Optimal techniques to simulate flow in fractured reservoir. In: ECMOR XIV-14th European
Conference on the Mathematics of Oil Recovery (2014)
29. Fumagalli, A., Scotti, A.: A numerical method for two-phase flow in fractured porous media
with non-matching grids. Adv. Water Resour. 62, 454–464 (2013). doi:http://dx.doi.org/10.
1016/j.advwatres.2013.04.001
30. Hyman, J., Gable, C., Painter, S., Makedonska, N.: Conforming Delaunay triangulation of
stochastically generated three dimensional discrete fracture networks: a feature rejection
algorithm for meshing strategy. SIAM J. Sci. Comput. 36, A1871–A1894 (2014). doi:http://
dx.doi.org/10.1137/130942541
31. Jaffré, J., Roberts, J.E.: Modeling flow in porous media with fractures; discrete fracture models
with matrix-fracture exchange. Numer. Anal. Appl. 5(2), 162–167 (2012)
32. Klawonn, A.: FETI domain decomposition methods for second order elliptic partial differential
equations. GAMM-Mitteilungen 29(2), 319–341 (2006). doi:10.1002/gamm.201490036
33. Nocedal, J., Wright, S.J.: Numerical Optimization. Springer, Berlin (1999)
34. Nœtinger, B.: A quasi steady state method for solving transient Darcy flow in complex 3D
fractured networks accounting for matrix to fracture flow. J. Comput. Phys. 283, 205–223
(2015). doi:http://dx.doi.org/10.1016/j.jcp.2014.11.038
35. Nœtinger, B., Jarrige, N.: A quasi steady state method for solving transient Darcy flow in
complex 3D fractured networks. J. Comput. Phys. 231(1), 23–38 (2012). doi:http://dx.doi.org/
10.1016/j.jcp.2011.08.015
36. Pichot, G., Erhel, J., de Dreuzy, J.: A mixed hybrid mortar method for solving flow in discrete
fracture networks. Appl. Anal. 89, 1629–643 (2010). doi:http://dx.doi.org/10.1080/00036811.
2010.495333
186 M.F. Benedetto et al.

37. Pichot, G., Erhel, J., de Dreuzy, J.: A generalized mixed hybrid mortar method for solving
flow in stochastic discrete fracture networks. SIAM J. Sci. Comput. 34, B86–B105 (2012).
doi:http://dx.doi.org/10.1137/100804383
38. Pichot, G., Poirriez, B., Erhel, J., de Dreuzy, J.R.: A Mortar BDD method for solving flow
in stochastic discrete fracture networks. In: Domain Decomposition Methods in Science
and Engineering XXI. Lecture Notes in Computational Science and Engineering, pp. 99–112.
Springer, Cham (2014)
39. Pieraccini, S., Scialò, S.: On a PDE-constrained optimization approach for flow simulations
in fractured media. In: Ventura, G., Benvenuti, E. (eds.) Advances in Discretization Methods.
SEMA SIMAI Springer Series vol. 12. Springer International Publishing, Cham (2016)
40. Raviart, P., Thomas, J.: Primal hybrid finite element methods for 2nd order elliptic equations.
Math. Comput. 31(138), 391–413 (1977)
41. Vohralík, M., Maryška, J., Severýn, O.: Mixed and nonconforming finite element methods on
a system of polygons. Appl. Numer. Math. 51, 176–193 (2007)
42. Wohlmuth, B.I.: A mortar finite element method using dual spaces for the lagrange multiplier.
SIAM J. Numer. Anal. 38(3), 989–1012 (2000)
43. Wohlmuth, B.I.: Discretization Techniques Based on Domain Decomposition. Springer, Berlin
(2001)
Adaptive Discontinuous Galerkin Methods
on Polytopic Meshes

Joe Collis and Paul Houston

Abstract In this article we consider the application of discontinuous Galerkin

finite element methods, defined on agglomerated meshes consisting of general
polytopic elements, to the numerical approximation of partial differential equation
problems posed on complicated geometries. Here, we assume that the underlying
computational domain may be accurately represented by a geometry-conforming
fine mesh Tfine ; the resulting coarse mesh is then constructed based on employing
standard graph partitioning algorithms. To improve the accuracy of the computed
numerical approximation, we consider the development of goal-oriented adaptation
techniques within an automatic mesh refinement strategy. In this setting, elements
marked for refinement are subdivided by locally constructing finer agglomerates;
should further resolution of the underlying fine mesh Tfine be required, then adaptive
refinement of Tfine will also be undertaken. As an example of the application of
these techniques, we consider the numerical approximation of the linear elasticity
equations for a homogeneous isotropic material. In particular, the performance of
the proposed adaptive refinement algorithm is studied for the computation of the
(scaled) effective Young’s modulus of a section of trabecular bone.

1 Introduction

Over the last couple of decades extensive work has been undertaken on the design
and mathematical analysis of numerical methods for the approximation of partial
differential equations (PDEs) based on exploiting general meshes consisting of
polytopic elements, i.e., polygons/polyhedra in two-/three-dimensions, respectively.
In particular, we mention the Polygonal Finite Element Method [34], the Extended
Finite Element Method [21], the Mimetic Finite Difference Method [10, 12, 16],
the Virtual Element Method [11], the Hybrid High Order Method [19, 20], the
Composite Finite Element Method [1, 24–26], and the closely related Agglomerated
Discontinuous Galerkin (DG) method [5–7]. The exploitation of general polytopic

J. Collis • P. Houston ()

School of Mathematical Sciences, University of Nottingham, University Park,
Nottingham NG7 2RD, UK
e-mail: [email protected]; [email protected]

© Springer International Publishing Switzerland 2016 187

G. Ventura, E. Benvenuti (eds.), Advances in Discretization Methods,
SEMA SIMAI Springer Series 12, DOI 10.1007/978-3-319-41246-7_9
188 J. Collis and P. Houston

elements offers great flexibility for mesh generation, and moreover allows for
sequences of nested, successively coarser, meshes to be generated for use within
multi-level solvers, such as multigrid and domain decomposition preconditioners,
cf. [2, 4, 8, 22], for example. We point out that polytopic elements naturally
arise when fictitious domain methods, unfitted methods or overlapping meshes are
employed, cf. [13–15, 28, 30], for example.
The motivation here for employing polytopic elements is very much inspired
by the work undertaken by Hackbusch and Sauter on Composite Finite Element
methods in the articles [25, 26]; for the extension to DG methods, we refer to [1, 24].
Here, polytopic elements allow for the construction of a geometry-conforming
mesh with a very small number of elements, irrespective of the complexity of the
underlying domain; as an example in Fig. 1 we consider a section of trabecular
bone which will be treated in Sect. 5.2 below. Indeed, by removing the need to have
standard-shaped elements, i.e., triangles/quadrilaterals in two-dimensions and tetra-
hedra/hexahedra/prisms/pyramids in three-dimensions, highly complex geometries
may be accurately meshed using a small number of (polytopic) elements. In the
series of articles [1, 24–26], the underlying mesh is constructed based on adaptively
refining an overlapping mesh and removing elements which do not lie inside the
domain until a suitably accurate representation of the given geometry is computed.
By exploiting the underlying tree structure generated by this adaptive refinement
procedure, general polytopic elements are formed by agglomerating elements which
share the same parent; for a review of this approach in the DG setting, we refer to [3].
The generation of such coarse meshes is advantageous from a computational point
of view, in the sense that coarse approximations may be computed very efficiently;

Fig. 1 Section of trabecular bone, cf. [31]

Adaptive Discontinuous Galerkin Methods on Polytopic Meshes 189

however, the accuracy of the resulting numerical solution may be insufficient and
subsequent mesh refinement may be desirable. In the composite finite element
framework adaptive mesh refinement may be undertaken in a simple manner by
constructing finite element partitions of the domain consisting of agglomerated
elements which belong to different levels of the underlying hierarchical tree data
structure; indeed, this approach has been exploited within the articles [23, 24].
In this article we consider an alternative approach whereby the underlying
geometry-conforming mesh may be constructed more generally using, for exam-
ple, standard mesh generation software, rather than the refined-overlapping mesh
approach considered in [23, 24]. The underlying coarse composite mesh is then
constructed using standard graph partitioning algorithms; for this purpose, here we
employ METIS [29]. On the basis of this coarse agglomerated mesh partition we
consider the application of goal-oriented dual-weighted-residual (DWR) a posteriori
error estimation for DG finite element methods; to fix ideas, we focus on the
discretization of the linear elasticity equations for a homogeneous isotropic material.
However, we stress that this approach is completely general and can be applied to
general classes of PDE problems. Once elements have been marked for refinement
on the basis of the DWR error indicators, the agglomerated elements can be
subdivided by simply employing a local graph partitioning algorithm involving
the set of (standard) elements which form the marked polytope. In this way
adaptive refinement can easily and efficiently be implemented without the need for
complicated mesh refinement tree data structures. However, once a refined element
is of the granularity of an element in the background fine mesh, then standard
refinement of the fine mesh may then need to be undertaken. To illustrate the
performance of the proposed agglomeration-based refinement algorithm, we present
a series of numerical examples; in particular, we consider the aforementioned three-
dimensional section of trabecular bone depicted in Fig. 1.
The outline of this article is as follows. In Sect. 2 we introduce, as a prototype
PDE problem, the linear elasticity equations for a homogeneous isotropic material.
Section 3 is then devoted to the formulation of the corresponding DG discretization,
based on employing the symmetric version of the interior penalty method, cf.
[27, 36]. In Sect. 4 we briefly introduce a goal-oriented a posteriori error estimator,
followed by an outline of the design of an appropriate agglomeration-based adaptive
mesh refinement algorithm. The practical performance of the proposed adaptive
refinement strategy is studied in Sect. 5. Finally, in Sect. 6 we summarize the work
presented in this article and draw some conclusions.

2 Model Problem

Given that ˝ is a bounded, connected Lipschitz domain in Rd , d > 1, with boundary

@˝, consider the following linear elasticity equation: find u such that

 r   .u/ D f in ˝; (1)
190 J. Collis and P. Houston

where u D .u1 ; : : : ; ud /> is the displacement and  is the stress tensor for a
homogeneous isotropic material, i.e.,

 .u/ D 2".u/ C r  u I;

I is the d  d identity matrix, ".u/ D 1=2.ru C ru> /, and  and  are the Lamé
coefficients, which satisfy the relation

0 < minf;  C g:

We divide @˝ into the disjoint subsets @˝D , @˝ND , and @˝N whose union is
@˝, with @˝D or @˝ND nonempty and relatively open in @˝. Following [35], we
supplement (1) with the following boundary conditions

u D gD on @˝D ;
u  n D gND on @˝ND ;
(2)
 .u/n  t D 0 on @˝ND ;
 .u/n D gN on @˝N ;

where n and t denote the unit outward normal vector and unit tangential vector(s)
on the boundary @˝.

3 Interior Penalty Discontinuous Galerkin Method

In this section we introduce the DG discretization of the model problem (1), (2)
based on employing the (symmetric) version of the interior penalty method,
cf. [27, 36].
To this end, let T be a subdivision of the computational domain ˝ into disjoint
open polytopic elements  such that ˝N D [2T . N For the purposes of this
article the polytopic mesh T will be constructed based on an agglomeration of
a geometry-conforming fine mesh Tfine consisting of standard element types, i.e.,
triangles/quadrilaterals in two-dimensions, and so on, cf. Sect. 1 above. Further
details concerning the construction of T will be given below in Sect. 4. On the
basis of the (polygonal/polyhedral) mesh T , given the polynomial degree p  1,
we define the corresponding DG finite element space Vp .T / by

Vp .T / D fu 2 ŒL2 .˝/ d W uj 2 ŒPp ./ d ;  2 T g;

where Pp ./ denotes the space of polynomials of total degree p over . As in [17],
the local elemental polynomial spaces employed within the definition of Vp .T / are
constructed in the physical space, without the need to map from a given reference
or canonical frame.
Adaptive Discontinuous Galerkin Methods on Polytopic Meshes 191

Following [17], we define the interfaces of the mesh T to be the set of .d  1/-
dimensional facets of the elements  2 T . To allow for the presence of hanging
nodes/edges, the interfaces of T are defined to be the intersection of the .d  1/-
dimensional facets of neighbouring elements. In two-dimensions, i.e., when d D 2,
the interfaces of a given element  2 T will always consist of .d  1/-dimensional
simplices (line segments). In general, for d D 3, this will not be the case; in this
setting, we assume that each interface of an element  2 T may be subdivided
by a set of co-planar triangles. Thereby, we use the terminology ‘face’ to refer to
a .d  1/-dimensional simplex, which forms part of the boundary (interface) of an
element  2 T .
As in [17, 18], we assume that a sub-triangulation into faces of each mesh
interface is given if d D 3, and denote by F the union of all open mesh interfaces
if d D 2 and the union of all open triangles belonging to the sub-triangulation of
all mesh interfaces if d D 3. We note that this assumption is trivially satisfied in
our setting when T is formed by the agglomeration of a fine mesh Tfine consisting
of simplices. We write F D F I [ F B , where F I denotes the set of all open
.d  1/-dimensional element faces F 2 F that are contained in ˝, and F B is the
union of element boundary faces. Furthermore, we write F B D FDB [ FND B
[ FNB
B B B
where FD , FND , and FN denote the set of boundary faces whose union form
@˝D , @˝ND , and @˝N , respectively. Here, the boundary @ of an element  and the
sets @n@˝, @ \ @˝D , @ \ @˝ND , and @ \ @˝N will be identified in a natural way
with the corresponding subsets of F . Implicit in these definitions is the assumption
that T respects the decomposition of @˝ in the sense that each F 2 F B belongs
to the interior of exactly one of @˝D , @˝ND , or @˝N .
Next, we define average and jump operators. To this end, let  C and   be two
adjacent elements of T , and x be an arbitrary point on the interior face F  @ C \
@  , F 2 F I . Given vector- and matrix-valued functions v and , respectively, that
are smooth inside each element  ˙ , by .v˙ ;  ˙ / we denote the traces of .v; / on F
taken from within the interior of  ˙ , respectively. Then, we introduce the averages
at x 2 F:

ffvgg D .vC C v /=2; ffgg D . C C   /=2:

Similarly, the jumps of v at x 2 F are given by

ŒŒv D vC ˝ n C C v ˝ n  ; ŒŒv D vC  n C C v  n  :

On a boundary face F 2 F B , we set ffvgg D v, ffgg D , ŒŒv D v ˝ n, and

ŒŒv D v  n, where n denotes the unit outward normal vector on the boundary @˝.
With this notation, the symmetric version of the interior penalty DG method is
given by: find uh 2 Vp .T / such that

B.uh ; vh / D `.vh / (3)

192 J. Collis and P. Houston

for all vh 2 Vp .T /. Here, the bilinear form B W Vp .T /  Vp .T / ! R is given by

XZ
B.w; v/ WD  .w/ W ".v/ dx
2T 

X Z 
 ff h .w/gg W ŒŒv C ff h .v/gg W ŒŒw  ˛  ŒŒw W ŒŒv ds
F
F2F I [FDB

X Z
C ˛  ŒŒw ŒŒv ds
F
F2F I [FDB [FND
B

X Z
 ... h .w/n/  n/ .v  n/ C .. h .v/n/  n/ .w  n// ds;
B F
F2FND

and the linear functional ` W Vp .T / ! R is defined by

Z Z Z
`.v/ D f  v dx   h .v/ W gD ˝ n ds C ˛  gD  v ds
˝ @˝D @˝D
Z Z
C ˛  .gD  n/.v  n/ ds  gND .. h .v/n/  n  ˛  v  n/ ds
@˝D @˝ND
Z
C gN  v ds;
@˝N

where  h is the stress tensor defined elementwise.

The non-negative function ˛ is referred to as the discontinuity-penalization
parameter; the precise definition will be given below based on the work undertaken
in our recent article [17]. To this end, following [17] we first introduce the submesh
TQ of elements from T .
Definition 1 Let TQ denote the subset of elements ,  2 T , such that each  2 TQ
can be covered by at most nT shape-regular simplices Ki , i D 1; : : : ; nT , such that

dist.; @Ki / > Cas diam.Ki /=p2 ; and jKi j  cas jj

for all i D 1; : : : ; nT , for some nT 2 N and Cas ; cas > 0, independent of  and T .
With this definition, we recall the following inverse inequality from [17]; we
stress that this result is sharp with respect to both the polynomial order p, and
moreover takes into account .d  k/-dimensional element facet degeneration, where
k D 1; 2; : : : ; d  1.
Adaptive Discontinuous Galerkin Methods on Polytopic Meshes 193

Lemma 1 Let  2 T , F  @ denote one of its faces, and TQ be defined as in

Definition 1. Then, for each v 2 Pp ./, we have the inverse estimate

p2 jFj
kvk2L2 .F/  CINV .p; ; F/ kvk2L2 ./ ; (4)
jj

with
8 ( )
ˆ
ˆ jj
ˆ
< min ;p 2d
; if  2 TQ ;
sup F  j[F j
CINV .p; ; F/ WD Cinv [
ˆ
ˆ jj
:̂ ; if  2 T nTQ ;
sup F  j[F j
[

and [F denotes a d-dimensional simplex contained in  which shares the face F with
 2 T . Furthermore, Cinv is a positive constant, which if  2 TQ depends on the
shape regularity of the covering of  given in Definition 1, but is always independent
of jj= sup F  j[F j (and, therefore, of jFj), p, and v.
[

Based on the inverse inequality stated in Lemma 1, together with the analysis
presented in [17, 27] the discontinuity-penalization parameter ˛ may be defined as
follows.
Definition 2 Let ˛ W F ! RC be defined facewise by
8 n
ˆ
ˆ p2 jFj o
< max CINV .p ; ; F/ ; x 2 F 2 F I ; F  @ C \ @  ;
2f C ;  g jj
˛.x/ D C˛
ˆ p2 jFj
:̂ CINV .p ; ; F/ ; x 2 F 2 F B ; F  @ \ @˝;
jj
(5)
with C˛ > C˛min , where C˛min is a sufficiently large lower bound.

4 Error Estimation and Adaptive Mesh Refinement

As noted above, for the purposes of this article we assume that the computa-
tional mesh T , consisting of general polytopic elements, is constructed based on
agglomerating an underlying fine mesh Tfine . Here, Tfine may be constructed using
a standard mesh generator; for example, we employ Triangle [32] and Tetgen [33]
in two- and three-dimensions, respectively. By employing standard mesh generation
software fine-scale geometry-conforming meshes may be generated; in the case
when the computational domain ˝ is highly complex, then Tfine may consist of a
very large number of elements, cf. [3], for example. With this mind, we employ
METIS, cf. [29], for example, to construct T with a user-defined number of
partitions based on agglomerating elements contained within Tfine . In general the
194 J. Collis and P. Houston

construction of coarse computational meshes will not be sufficient to meet the

accuracy demanded by a user. With this mind, we consider the design of a mesh
refinement algorithm which automatically refines the agglomerates which form
T based on repartitioning elements which possess a large error contribution. To
illustrate this approach, we focus on employing a dual-weighted-residual (DWR)
error estimator, cf. [9], for example, together with the references cited therein.
To this end, given a (linear, for simplicity) target functional J, we recall the
following a posteriori error estimation formula
X
J.u/  J.uh / D `.z  zh /  B.uh ; z  zh /  
2T

for all zh 2 Vp .T /, where  D  .uh ; z  zh / denotes the local elementwise error

indicators on ,  2 T . Furthermore, z denotes the solution of the corresponding
dual/adjoint problem: find z such that

B.v; z/ D J.v/ 8v: (6)

In practice, the error indicators  ,  2 T , are computed based on approximating

the dual solution z by zO on the polytopic finite element mesh T , using polynomials
one degree higher than those employed for the computation of uh ; thereby, zO 2
VpC1 .T /. Writing O  D  .uh ; zO  zh / gives rise to the approximate error
representation formula
X
J.u/  J.uh / O  :
2T

On the basis of the size of the modulus of the (approximate) local error indicators,
i.e., jO  j, the elements in the mesh T are marked for refinement using the fixed
fraction strategy with refinement parameter REF. Once an element  2 T has
been marked for refinement, then assuming that  is formed from the union of a
.l/
set of elements belonging to Tfine  Tfine , l D 0, i.e.,  D [ 0 2S .l/  0 , where

.l/
S  T denotes the set of fine elements which form , then METIS is applied
.l/
to the corresponding graph representation of S to yield a local partition of 
consisting of m agglomerated elements; here, we set m D 2d . In the case when
.l/
m > card.S / for any element  2 T which has been marked for refinement,
.l/
then the elements  0 2 S ,  2 T , are first isotropically refined using standard
.lC1/ .lC1/
adaptive mesh refinement algorithms to yield a new fine mesh Tfine . Once Tfine
.lC1/
has been constructed then new local partitions S of each element  2 T
may be computed and, for those marked for refinement, subsequently subdivided
using graph partitioning techniques. We stress that, assuming Tfine does not require
adaptive mesh refinement to be undertaken, then the refinement of T can be
done in a very straight-forward manner using only graph partitioning algorithms,
Adaptive Discontinuous Galerkin Methods on Polytopic Meshes 195

Algorithm 4.1 Construction and refinement of the agglomerated mesh T

.l/
1: Construct a geometry conforming fine mesh Tfine  Tfine , l D 0, consisting of standard
element shapes.
2: Compute initial agglomerated mesh T  T .k/ , k D 0, consisting of a user-defined number
of elements.
3: Solve (3) for uh 2 Vp .T .k/ /.
4: Compute the numerical approximation zO 2 VpC1 .T .k/ / to the dual/adjoint problem (6).
5: Evaluate the (approximate) element error indicators O  D  .uh ; zO  zh /, zh 2 Vp .T .k/ /, for
.k/
P 2 T .
each
6: if 2T .k/ O  < TOL, where TOL is a user-defined tolerance then
7: STOP
8: end if
9: Set refine_fine_mesh = False
10: Construct the refinement set R based on employing the fixed fraction refinement strategy.
11: for all  2 R do
.l/
12: if m > card.S /, where m D 2d then
0 .l/
13: for all  2 S do
0
14: Refine  and set refine_fine_mesh = True
15: end for
16: end if
17: end for
18: if refine_fine_mesh then
.l/
19: Set l D l C 1 and construct new fine mesh Tfine .
.k/ .l/
20: Recompute agglomeration partition for T relative to the new fine mesh Tfine .
21: end if
22: for all  2 R do
23: Refine  based on computing a subpartition consisting of m elements.
24: end for
25: Set k D k C 1 and construct newly refined agglomerated mesh T .k/ .
26: Goto 3.

without the need to implement complicated tree data structures, which are typically
employed within standard refinement procedures. A summary of the proposed
adaptive algorithm is presented in Algorithm 4.1.

5 Numerical Examples

In this section we present a series of numerical experiments to highlight the practical

performance of the agglomeration-based adaptive mesh refinement algorithm out-
lined in Sect. 4. To this end, following the work presented in the recent article [35],
we consider the evaluation of the (scaled) effective Young’s modulus of a given
structure; the key application presented in Sect. 5.2 is concerned with the modelling
of the section of trabecular bone depicted in Fig. 1.
Throughout this section, we assume that the underlying geometry ˝ is contained
within a d-dimensional cuboid ˝cube , where ˝cube D ˘iD1 d
.xmin
i ; xi
max
/. Writing
@˝cube to denote the boundary of ˝cube , i.e., the planar sides of the cuboid, we
196 J. Collis and P. Houston

define @˝ND D @˝ \ @˝cube and @˝N D @˝n@˝ND ; thereby, @˝D D ;. With this
notation, we set gND D uN D 0:01Ld , Ld D xmax
d  xmin
d , on the top section of @˝ND ,
i.e., where xd D xd , and gND D 0 on all other portions of @˝ND . Furthermore,
max

we set gN D 0 on @˝N . Finally, writing E to denote the Young’s modulus and  the
Poisson ratio, we define our functional of interest by
Z
1 Ld
J.u/ D dd .u/ dx;
E uN j˝cube j ˝

cf. [35].

5.1 Example 1: Two-Dimensional Cross Geometry

In this section we consider a simple two-dimensional cross-geometry enclosed

within the unit square .0; 1/2 ; cf. Fig. 2. Here, we set E D 106 and the Poisson
ratio  D 0:3; thereby, on the basis of a fine mesh calculation we compute an
approximation to the (scaled) effective Young’s modulus J.u/ 0:228481378.
We generate an initial fine mesh Tfine consisting of 111,457 triangular elements;
this is then agglomerated using METIS to generate a coarse polytopic mesh T
comprising 32 elements only, cf. Fig. 2. In Tables 1 and 2 we show the convergence

Fig. 2 Example 1. Initial agglomerated mesh consisting of 32 elements

Adaptive Discontinuous Galerkin Methods on Polytopic Meshes 197

Table 1 Example 1. Adaptive algorithm for p D 1

P
No of Eles No of Dofs J.u/  J.uh / 2T 
O 
32 192 3.794E02 3.142E02 0:83
50 300 2.424E02 1.973E02 0:81
80 480 1.561E02 1.307E02 0:84
128 768 1.075E02 9.145E03 0:85
203 1218 7.681E03 6.793E03 0:88
323 1938 5.535E03 5.100E03 0:92
514 3084 3.885E03 3.670E03 0:94
820 4920 2.661E03 2.563E03 0:96
1308 7848 1.759E03 1.707E03 0:97
2084 12,504 1.143E03 1.112E03 0:97
3316 19,896 7.352E04 7.182E04 0:98
5274 31,644 4.534E04 4.443E04 0:98
8366 50,196 2.717E04 2.680E04 0:99
13,214 79,284 1.565E04 1.551E04 1:00
20,798 124,788 8.965E05 8.946E05 1:00

Table 2 Example 1. Adaptive algorithm for p D 2

P
No of Eles No of Dofs J.u/  J.uh / 2T 
O 
32 384 9.499E03 4.471E03 0:47
50 600 5.955E03 3.318E03 0:56
80 960 3.458E03 2.111E03 0:61
128 1536 2.149E03 1.473E03 0:69
203 2436 1.140E03 8.385E04 0:74
323 3876 5.537E04 4.417E04 0:80
514 6168 2.645E04 2.221E04 0:84
817 9804 1.114E04 9.954E05 0:89
1301 15,612 4.839E05 4.337E05 0:90
2068 24,816 2.051E05 1.839E05 0:90
3277 39,324 8.643E06 7.999E06 0:93
5216 62,592 3.491E06 3.491E06 1:00

history of the proposed agglomeration-based adaptive strategy using a polynomial

order of degree p D 1 and p D 2, respectively, with REF D 20 %. In each case,
we show the number of elements in polytopic mesh T , the number of degrees of
freedom in underlying finite element space Vp .T /, the true error in the (scaled)
effective
P Young’s modulus functional J./, theP computed error representation for-
mula 2T O  , and the effectivity index  D 2T O  =.J.u/  J.uh //. As noted
in [23], here we see that, even on very coarse finite element meshes, the quality of
the computed error representation formula is relatively good, in the sense that the
effectivity indices are not too far away from unity. In particular, accuracy which
198 J. Collis and P. Houston

-1
10
p=1
p=2
-2
10

-3
10
|J(u)-J(u )|
h

-4
10

10 -5

-6
10
10 2 10 3 10 4 10 5
Degrees of Freedom

Fig. 3 Example 1: Convergence of the adaptive mesh refinement algorithm

is sufficient for practical/engineering calculations can be attained with a relatively

small number of degrees of freedom. The results from Tables 1 and 2 are also
shown in Fig. 3; here, we clearly observe the superiority of employing higher-order
elements, in the sense that the error computed with p D 2 is significantly smaller
than the corresponding quantity evaluated for p D 1, when the same number of
degrees of freedom are employed.
Finally, in Fig. 4 we show the agglomerated polytopic meshes generated after
6 and 11 adaptive refinement steps have been performed with p D 1. Here we
observe that the meshes have been refined in the vicinity of the reentrant corners
present in the cross domain ˝ as we would expect. Indeed, the regions away from
these features have remained largely unrefined. Moreover, we note that hanging
nodes are naturally generated through the application of local agglomeration-based
refinement; this is very easily handled within the DG setting.

5.2 Example 2: Modelling Trabecular Bone

Following the work presented in the recent article [35], in this section we consider
the evaluation of the (scaled) effective Young’s modulus of the section of trabecular
bone depicted in Fig. 1. The geometry ˝ represents a cuboidal section of trabecular
bone obtained by X-ray CT scanning of a bone biopsy reconstructed from two-
dimensional slices, cf. [31]. In this section we set the Young’s modulus E D 10 GPa
and the Poisson ratio  D 0:3; in this case, we computed the approximate reference
value J.u/ 0:1236. The initial fine mesh Tfine , cf. Fig. 5a, consists of 1,179,569
tetrahedral elements, which is then agglomerated to generate a coarse polytopic
Adaptive Discontinuous Galerkin Methods on Polytopic Meshes 199

Fig. 4 Example 1. Agglomerated mesh for p D 1 after: (a) 6 refinements, with 514 elements; (b)
11 refinements, with 5274 elements
200 J. Collis and P. Houston

Fig. 5 Example 2. (a) Initial fine mesh; (b) Zoom of the first 30 agglomerated elements (shown
in colour)

mesh T comprising of only 8000 elements. The first 30 elements are depicted in
Fig. 5b. In Tables 3 and 4 we tabulate the results of the proposed adaptive refinement
strategy with p D 1 and p D 2, respectively, as before, with REF D 10 %; cf.,
also, Fig. 6. As in the previous numerical experiment, we again notice that the
effectivity indices  are relatively good, given the coarse nature of the finite element
meshes employed. Indeed, as the mesh is refined, we observe that  improves and
approaches unity. Again, here we observe that a sufficiently accurate (in terms of
engineering accuracy) approximation to the target functional of interest may be
computed with very few degrees of freedom. Finally, in Figs. 7 and 8 we show the
primal and dual displacements, respectively.
Adaptive Discontinuous Galerkin Methods on Polytopic Meshes 201

Table 3 Example 2. Adaptive algorithm for p D 1

P
No of Eles No of Dofs J.u/  J.uh / 2T 
O 
8000 96;000 1.402E01 1.171E01 0:83
13,600 163;200 1.185E01 1.023E01 0:86
22,994 275;928 9.759E02 8.748E02 0:90
38,867 466;404 7.945E02 7.310E02 0:92
65,634 787;608 6.561E02 6.203E02 0:95
110,752 1;329;024 5.497E02 5.267E02 0:96
18,6586 2;239;032 4.452E02 4.374E02 0:98
314,088 3;769;056 3.517E02 3.585E02 1:02

Table 4 Example 2. Adaptive algorithm for p D 2

P
No of Eles No of Dofs J.u/  J.uh / O 
2T 
8000 240;000 6.009E02 3.872E02 0.64
13,600 408;000 4.153E02 3.065E02 0.74
22,962 688;860 2.473E02 2.106E02 0.85
38,808 1;164;240 1.660E02 1.606E02 0.97
65,584 1;967;520 1.253E02 1.253E02 1.00
110,602 3;318;060 8.375E03 a a

a
Indicates that the dual problem was not computed

p=1
p=2

10 -1
|J(u)-J(u )|
h

-2
10

10 5 10 6
Degrees of Freedom

Fig. 6 Example 2: Convergence of the adaptive mesh refinement algorithm

202 J. Collis and P. Houston

Fig. 7 Example 2. Approximate primal solution: (a) u1 ; (b) u2 ; (c) u3

Adaptive Discontinuous Galerkin Methods on Polytopic Meshes 203

Fig. 8 Example 2. Approximate dual solution: (a) z1 ; (b) z2 ; (c) z3

204 J. Collis and P. Houston

6 Conclusions

In this article we have developed an agglomeration-based adaptive mesh refinement

algorithm within the goal-oriented setting for the DG approximation of the linear
elasticity equations for a homogeneous isotropic material. We stress that this PDE
model was chosen simply as a prototype problem; indeed, the work undertaken in
this article naturally generalises to general classes of PDEs. By exploiting general
agglomeration techniques, complicated geometries can be meshed using only a very
small number of polytopic elements. The proposed adaptive refinement strategy
then applies graph partitioning techniques to local elements which are marked for
refinement. In this manner, the underlying mesh is adaptively refined on the basis of
solution accuracy, rather than computing on fine meshes generated for the purposes
of resolving the underlying geometry. This approach has been applied to a simple
two-dimensional problem, as well as a test case involving estimating the (scaled)
effective Young’s modulus for a section of trabecular bone. Extensions of this work
to include, for example, automatic hp-refinement will be undertaken as part of our
programme of future research.

Acknowledgements Joe Collis acknowledges the financial support of the EPSRC under the grant
EP/K039342/1.

References

1. Antonietti, P.F., Giani, S., Houston, P.: hp-Version composite discontinuous Galerkin methods
for elliptic problems on complicated domains. SIAM J. Sci. Comput. 35(3), A1417–A1439
(2013)
2. Antonietti, P.F., Giani, S., Houston, P.: Domain decomposition preconditioners for discontin-
uous Galerkin methods for elliptic problems on complicated domains. J. Sci. Comput. 60(1),
203–227 (2014)
3. Antonietti, P.F., Cangiani, A., Collis, J., Dong, Z., Georgoulis, E.H., Giani, S., Houston, P.:
Review of discontinuous Galerkin finite element methods for partial differential equations on
complicated domains. In: Barrenechea, G.R., Brezzi, F., Cangiani, A., Georgoulis, E.H. (eds.)
Building Bridges: Connections and Challenges in Modern Approaches to Numerical Partial
Differential Equations. Lecture Notes in Computational Science and Engineering. Springer,
Berlin (2015)
4. Antonietti, P.F., Houston, P., Sarti, M., Verani, M.: Multigrid algorithms for hp-version interior
penalty discontinuous Galerkin methods on polygonal and polyhedral meshes (2015, submitted
for publication)
5. Bassi, F., Botti, L., Colombo, A., Rebay, S.: Agglomeration based discontinuous Galerkin
discretization of the Euler and Navier-Stokes equations. Comput. Fluids 61, 77–85 (2012)
6. Bassi, F., Botti, L., Colombo, A., Di Pietro, D.A., Tesini, P.: On the flexibility of agglomeration
based physical space discontinuous Galerkin discretizations. J. Comput. Phys. 231(1), 45–65
(2012)
7. Bassi, F., Botti, L., Colombo, A.: Agglomeration-based physical frame dG discretizations: an
attempt to be mesh free. Math. Models Methods Appl. Sci. 24(8), 1495–1539 (2014)
Adaptive Discontinuous Galerkin Methods on Polytopic Meshes 205

8. Bassi, F., Botti, L., Colombo, A.: h-Multigrid agglomeration based solver strategies for BR2
discontinuous Galerkin discretizations of elliptic problems (2015, Submitted for publication)
9. Becker, R., Rannacher, R.: An optimal control approach to a-posteriori error estimation in
finite element methods. In: Iserles, A. (ed.) Acta Numerica, pp. 1–102. Cambridge University
Press, Cambridge (2001)
10. Beirão da Veiga, L., Droniou, J., Manzini, G.: A unified approach for handling convection
terms in finite volumes and mimetic discretization methods for elliptic problems. IMA J.
Numer. Anal. 31(4), 1357–1401 (2011)
11. Beirão da Veiga, L., Brezzi, F., Cangiani, A., Manzini, G., Marini, L.D., Russo, A.: Basic
principles of virtual element methods. Math. Models Methods Appl. Sci. 23(1), 199–214
(2013)
12. Brezzi, F., Buffa, A., Lipnikov, K.: Mimetic finite differences for elliptic problems. ESAIM
Math. Modell. Numer. Anal. 43(2), 277–295 (2009)
13. Burman, E., Hansbo, P.: An interior-penalty-stabilized Lagrange multiplier method for the
finite-element solution of elliptic interface problems. IMA J. Numer. Anal. 30, 870–885 (2010)
14. Burman, P., Hansbo, P.: Fictitious domain finite element methods using cut elements: I. A
stabilized Lagrange multiplier method. Comput. Methods Appl. Mech. Eng. 199, 2680–2686
(2010)
15. Burman, E., Hansbo, P.: Fictitious domain finite element methods using cut elements: II. A
stabilized Nitsche method. Appl. Numer. Math. 62, 328–341 (2012)
16. Cangiani, A., Manzini, G., Russo, A.: Convergence analysis of the mimetic finite difference
method for elliptic problems. SIAM J. Numer. Anal. 47(4), 2612–2637 (2009)
17. Cangiani, A., Georgoulis, E.H., Houston, P.: hp-Version discontinuous Galerkin methods
on polygonal and polyhedral meshes. Math. Models Methods Appl. Sci. 24(10), 2009–2041
(2014)
18. Cangiani, A., Dong, Z., Georgoulis, E.H., Houston, P.: hp-Version discontinuous Galerkin
methods for advection–diffusion–reaction problems on polytopic meshes. ESAIM Math.
Modell. Numer. Anal. 50(3), 699–725 (2016)
19. Di Pietro, D.A., Ern, A.: Hybrid High-Order methods for variable diffusion problems on
general meshes. C.R. Math. 353, 31–34 (2014)
20. Di Pietro, D.A., Ern, A., Lemaire, S.: A review of Hybrid High-Order methods: formulations,
computational aspects, comparison with other methods. In: Barrenechea, G.R., Brezzi, F.,
Cangiani, A., Georgoulis, E.H. (eds.) Building Bridges: Connections and Challenges in Mod-
ern Approaches to Numerical Partial Differential Equations. Lecture Notes in Computational
Science and Engineering. Springer, Berlin (2015)
21. Fries, T.-P., Belytschko, T.: The extended/generalized finite element method: an overview of
the method and its applications. Int. J. Numer. Methods Eng. 84(3), 253–304 (2010)
22. Giani, S., Houston, P.: Domain decomposition preconditioners for discontinuous Galerkin
discretizations of compressible fluid flows. Numer. Math. Theory Methods Appl. 7(2), 123–
128 (2014)
23. Giani, S., Houston, P.: Goal-oriented adaptive composite discontinuous Galerkin methods for
incompressible flows. J. Comput. Appl. Math. 270, 32–42 (2014)
24. Giani, S., Houston, P.: hp-Adaptive composite discontinuous Galerkin methods for elliptic
problems on complicated domains. Numer. Methods Part. Diff. Equ. 30(4), 1342–1367 (2014)
25. Hackbusch, W., Sauter, S.A.: Composite finite elements for problems containing small
geometric details. Part II: implementation and numerical results. Comput. Visual Sci. 1, 15–25
(1997)
26. Hackbusch, W., Sauter, S.A.: Composite finite elements for the approximation of PDEs on
domains with complicated micro-structures. Numer. Math. 75, 447–472 (1997)
27. Hansbo, P., Larson, M.G.: Discontinuous Galerkin methods for incompressible and nearly
incompressible elasticity by Nitsche’s method. Comput. Methods Appl. Mech. Eng.
191(17–18), 1895–1908 (2002)
28. Johansson, A., Larson, M.G.: A high order discontinuous Galerkin Nitsche method for elliptic
problems with fictitious boundary. Numer. Math. 123(4), 607–628 (2013)
206 J. Collis and P. Houston

29. Karypis, G., Kumar, V.: A fast and highly quality multilevel scheme for partitioning irregular
graphs. SIAM J. Sci. Comput. 20(1), 359–392 (1999)
30. Massing, A.: Analysis and implementation of finite element methods on overlapping and
fictitious domains. Ph.D. thesis, University of Oslo (2012)
31. Perilli, E., Baruffaldi, F.: Proposal for shared collections of X-ray microCT datasets of bone
specimens. In: ICCB03, Zaragoza, Spain, 24–26 September 2003
32. Shewchuk, J.R.: Triangle: engineering a 2D quality mesh generator and Delaunay triangulator.
In: Lin, M.C., Manocha, D. (eds). Applied Computational Geometry: Towards Geometric
Engineering. Lecture Notes in Computer Science, vol. 1148, pp. 203–222. Springer, Berlin
(1996); From the First ACM Workshop on Applied Computational Geometry
33. Si, H.: TetGen, a Delaunay-based quality tetrahedral mesh generator. ACM Trans. Math.
Softw. 41(2), 11:1–11:36 (2015)
34. Sukumar, N., Tabarraei, A.: Conforming polygonal finite elements. Int. J. Numer. Methods
Eng. 61(12), 2045–2066 (2004)
35. Verhoosel, C.V., van Zwieten, G.J., van Rietbergen, B., de Borst, R.: Image-based goal-
oriented adaptive isogeometric analysis with application to the micro-mechanical modeling
of trabecular bone. Comput. Methods Appl. Mech. Eng. 284, 138–164 (2015)
36. Wihler, T.P.: Locking-free adaptive discontinuous Galerkin FEM for linear elasticity problems.
Math. Comput. 75(255), 1087–1102 (2006)
 Part IV
Advances in XFEM/Fictitious Domain
Methods
Stabilized X-FEM for Heaviside and Nonlinear
Enrichments

Giulio Ventura and Claudia Tesei

Abstract One of the drawbacks of the eXtended Finite Element Method and
similar approaches, like the Generalized Finite Element Method, is the problem
of ill-conditioning of the related systems of equations at the solution stage. This
occurs for example in Heaviside function enrichments when the discontinuity is
close to discretisation nodes but also for non-linear enrichment functions used in
conjunction to geometric enrichment domains. In the present work the motivation
of ill-conditioning is analyzed to derive a novel methodology for stabilization, based
on setting proper constraints for the variables. This methodology does not impact
on the initial formulation nor in the element stiffness computation, so that it is very
effective for engineering applications. Results are analyzed in 1D and 3D to show
its performance and properties.

1 Introduction

The success of enriched formulations like XFEM or GFEM [2, 4, 6, 13, 16, 37]
is nowadays well recognized also in commercial softwares [1, 23], implementing
many features of these methods to the end of achieving more accurate solutions
with coarser meshes and independency of discretization on the position of features
of the problem like cracks or singularities. However, some fundamental problems
are still present like quadrature at the element level [14, 24, 26, 27, 31, 32, 34, 35]
and ill-conditioning of the resulting systems of equations. These problem are solved
in practical applications by some techniques that require additional computational
resources or introduce approximations of difficult determination.
The problem of ill-conditioning, although known since the beginning of the
development of these methods, has found limited attention in literature. The main
techniques used for stabilization can be divided into three groups:
• perturbation or preconditioning of the global system of equations. One of the
first contribution based on a perturbation scheme was given by Strouboulis et al.

G. Ventura () • C. Tesei

DISEG, Politecnico di Torino, Corso Duca degli Abruzzi 24, Torino, Italy
e-mail: [email protected]; [email protected]

© Springer International Publishing Switzerland 2016 209

G. Ventura, E. Benvenuti (eds.), Advances in Discretization Methods,
SEMA SIMAI Springer Series 12, DOI 10.1007/978-3-319-41246-7_10
210 G. Ventura and C. Tesei

[30] in the context of GFEM. Approaches based on special preconditioners were

presented by Béchet et al. [5], Menk and Bordas [25], Berger-Vergiat et al. [12],
Waisman and Berger-Vergiat [36], Lang et al. [21] to cite some of the many works
based on this technique;
• element stiffness matrix kernel space correction. This approach, proposed in [22],
is based on analyzing the element stiffness matrix eigenspace and introduce a
threshold and a stabilization factor. The element matrix is modified by adding,
for each eigenvalue needing stabilization, a term given by the stabilization
factor times the outer product of the corresponding eigenvector times itself. This
method proves effective but its implementation is not straightforward;
• the Stable Generalized Finite Element Method (SGFEM), introduced by Babus̆ka
and Banerjee [3], and recently reformulated for higher order shape functions in
[38]. This method is conceptually quite attractive and is based on eliminating
linearly dependent parts in the approximation space by subtracting a polynomial
interpolation of the enrichment function. An application of this method to 3D
fracture has been given in [18].
Finally, a very recently introduced approach, called Orthonormalized General-
ized Finite Element Method (OGFEM) has been introduced in [29]. In OGFEM the
ill-conditioning is prevented by orthonormalizing the basis functions to eliminate
linear dependence in the approximation basis.
The present paper proposes a new method for stabilization that is conceptually
different from the literature approaches and is conceived for simplicity and effective-
ness to have the least impact on implementation and computational cost. This will
be accomplished by introducing suitable constraints on the enrichment variables.
The outline of the paper follows: in Sect. 2 it is first examined the motivation for
ill-conditioning; in Sect. 3 the analytic structure of ill conditioning is derived and
commented. Then the new stabilization technique is introduced in Sect. 4 and the
method is validated by analyzing 1D and 3D results in Sect. 5.

2 Motivation for Ill-Conditioning

Reference will be made in the following to linear elastic solid mechanics problems,
although the discussion can be applied to general enriched finite element discretiza-
tions. Let u be the displacement field, an enrichment function and N the elements
shape functions. The dependence of the functions w.r.t. the position in space will be
omitted for the sake of brevity.
The standard formulation in XFEM/GFEM is:
X X
uD Ni ui C Ni a i (1)
i i
Stabilized X-FEM for Heaviside and Nonlinear Enrichments 211

where the first addend is the standard finite element approximation and the second
adds to the approximation space the enrichment function by partition of unity.
Elements where the enrichment function is not needed will miss the second addend.
Transition from enriched to non-enriched elements may need blending techniques
as illustrated in [15, 34]. The symbols in (1) are defined as follows:
• the sum on i is extended to the nodes of the finite element where the evaluation
point of the displacement function u is located;
• ui are the classical finite element displacement nodal variables;
• ai are the enrichment nodal variables.
When the approximation (1) is introduced in the variational principle stating
the mechanical problem, whenever the second addend will have terms linearly
dependent on terms in the first addend then ill-conditioning of the resulting system
of equations will occur. This ill-conditioning expresses the fact that the same
solution can be represented by linearly dependent sets of variables ui and ai
(indeterminate system of equations). An analytical example will be presented in
the next Section.
This situation may occur in several cases. The most common and intuitive is
when the generalized Heaviside step function H is used as enrichment function for
representing jumps in displacement (e.g. cracks). Let d be the signed distance from
the discontinuity, it is

C1 for d  0
H.d/ D : (2)
1 for d < 0

The generalized Heaviside being a piecewise constant function, it will be linearly

independent of the element shape functions whenever the discontinuity surface
d D 0 will cross inside an element. Strict linear dependence will hold when the
discontinuity will cross a node and numerical ill-conditioning is observed as the
discontinuity will get close to a node. This problem is sometimes solved by moving
the crack trajectory away from nodes or by moving the nodes away from the crack
surface with the consequence of introducing an a priori unknown source of further
approximation [22]. However, not only ill-conditioning takes place with piecewise
constant or linear enrichment functions (spanned by element shape functions). Ill-
conditioning also arises when employing non-linear enrichments in conjunction
with the concept of geometric enrichment area.
The concept of geometric enrichment was first introduced by Ventura et al. in
[33] and then, independently and almost simultaneously, by Laborde et al. [20] and
Béchet et al. [5]. It is recalled that by geometric enrichment it is meant a fixed
domain area where finite elements are enriched (independently of the mesh size),
while in topologic enrichment certain specific elements are enriched so that, when
the mesh size is reduced, the enrichment area shrinks accordingly.
It was observed, in fact, that if a near crack tip enrichment (or dislocation core
enrichment) is confined to the element containing the crack tip (or dislocation core),
as the discretization mesh is refined the solution error in energy increases instead of
212 G. Ventura and C. Tesei

decreasing. On the other hand, if a fixed domain area is enriched, the error decreases
by decreasing the mesh size, as expected [33].
We can explain this behaviour by considering that a singularity is present at the
tip, so that the solution is highly nonlinear in a neighbour of the tip and tends
to flatten away. The enrichment function depicts the solution closely, so that the
element shape functions play the role of adjusting the enrichment to the particular
case at hand. If, refining the mesh, the enriched domain is shrunk, then the high
gradients and nonlinearity of the solution must be represented by polynomials (the
representation space of standard finite elements), with the consequent degradation of
the solution approximation. In contrast, this will not occur for geometric enrichment,
being the enriched domain independent of the mesh size.
To see how this may affect ill-conditioning, consider the graph of the enrichment
function for a dislocation in uniform shear stress considered in [33], Fig. 1. The
function has been analytically derived by Koehler [19] and other reference solutions
can be found in [28]. It is immediate to observe that, while the enrichment function is
markedly nonlinear close to the dislocation core, its graph tends to a linear function

Fig. 1 Graph of enrichment functions for edge dislocation problems. The two graphs represent
the displacement components of the analytic solution [19]
Stabilized X-FEM for Heaviside and Nonlinear Enrichments 213

as the distance from the core is increased. Consequently, there will exist a size of
the geometric enrichment for which the enrichment function will be numerically
linear and the global system of equations will be indeterminate. This consideration
is the keystone of SGFEM [3], where the linear interpolation to the enrichment
function is subtracted from the enrichment space. The same reasoning holds for
other enrichments like near field functions for cracks.

3 Structure of Ill-Conditioning

The new method for stabilization proposed in the present paper is based on the
observation of the analytic structure of the problem in the simple case of a bar in
two cases:
• Heaviside function enrichment, representing the physical problem of a cut
separating the bar into two parts. This is typical of crack faces enrichment;
• linear function enrichment, representing the limit behaviour of a nonlinear
function typical, for example, of a near crack tip enrichment or a near dislocation
core enrichment when far from the tip or core, respectively.
It is interesting to note that ill-conditioning is also observed, for the reasons
stated above, when regularized Heaviside functions are directly used as enrichment
functions according to (1). This, however, does not occur when a proper variational
formulation is adopted, like in [10]. The formulation introduced in [10] and
employed also in [7–9, 11] is in fact numerically stable for any value of the
Heaviside function regularization parameter and size of the enrichment domain.

3.1 Heaviside Function Enrichment

Consider a bar with a cut placed at a distance d from the left end, subjected to
prescribed displacements ı1 and ı2 at the left and right ends, respectively, Fig. 2. To
get more readable results, unit bar length, cross section area and material Young’s
modulus are assumed. The solution of this problem is given by zero strain and
stresses and rigid body motions of the two parts separated by the cut.
Consider a discretisation with one finite element and let

N1 .x/ D 1  x I N2 .x/ D x (3a)


C1 for x  d
.x/ D H.x/ D (3b)
1 for x < d
2
X 2
X
u.x/ D Ni ui C .x/Ni ai (3c)
iD1 iD1
214 G. Ventura and C. Tesei

Fig. 2 A bar with a cut discontinuity and prescribed ends displacement

where the subscript 1 stands for the left end and the subscript 2 for the right
end. Enforcing the boundary conditions by Lagrangian multipliers r1 ; r2 , the total
potential energy functional of the system is
Z
1 L
˘D A.x/".x/ dx C r1 .u.0/  ı1 // C r2 .u.L/  ı2 // (4)
2 0

with the strain and stresses being given by, respectively

du.x/
".x/ D I .x/ D E".x/: (5)
dx

The stationarity conditions w.r.t. the variables .u1 ; u2 ; a1 ; a2 ; r1 ; r2 / yield the

following solutions, depending on the position of the discontinuity

r1 D r2 D 0 8d i.e. constraint reactions = 0 as expected

d ¤ 0 and d ¤ L ! u1 D u2 D ı1 Cı2
2
I a1 D a2 D ı2 ı2
1
(6)
dD0 ! u2 C a2  ı2 D 0I u1 D 2 I a1 D ı2 ı
ı1 Cı2
2
1

ı1 Cı2 ı2 ı1
dDL ! u1  a1  ı1 D 0I u2 D 2 I a2 D 2 :

The solution (6) is unique for d ¤ 0 and d ¤ L (discontinuity along the bar)
and is indeterminate for d D 0 and d D L (discontinuity at the ends of the bar).
In particular, when the solution is indeterminate, a standard variable is linearly
dependent on the corresponding enrichment variable, yielding 11 solutions. Being
H.0/ D 1; H.L/ D C1, (3b), the general structure of the indeterminate equation
is

ui C H.xi /ai  ıi D 0: (7)

Equation (7) is derived from (6) looking at the cases d D 0 and d D L, first equation
after the right arrows.
Stabilized X-FEM for Heaviside and Nonlinear Enrichments 215

3.2 Linear Function Enrichment

Consider a bar subjected to prescribed displacements ı1 and ı2 at the left and

right ends, respectively, Fig. 3. Again, to get more readable results an unit bar
length, cross section area and material Young’s modulus are assumed. Consider the
following linear enrichment function

.x/ D c0 C c1 x: (8)

The solution of this problem is given by linearly varying displacement from ı1

to ı2 and constant strain .ı2  ı1 /=L as if the enrichment function would not be
present. It is useful to mark that this enrichment function has no physical sense, but
it represents a decaying nonlinear enrichment function near the boundary of a fixed
geometric enrichment area.
For better clarity, the cases c0 D c1 D 0 (no enrichment), c0 ¤ 0; c1 D 0
(constant enrichment), c0 D 0; c1 ¤ 0 (linear homogeneous enrichment) and c0 ¤
0; c1 ¤ 0 (general linear enrichment) will be examined.
Following the same steps as in the previous section, replacing the expression
for .x/ with (8), the stationarity conditions yield the following solutions in the
different cases

r1 D ı2  ı1 ; r2 D ı1  ı2 8c0 ; c1 constraint reactions

c0 ¤ 0; c1 D 0 ! u1 C c0 a1  ı1 D 0I u2 C c0 a2  ı2 D 0
c0 D 0; c1 ¤ 0 ! u1 D ı1 I u2 C a2 c1  ı2 D 0I a2 D a1
c0 ¤ 0; c1 ¤ 0 ! c0 u2  .L/.u1  ı1 /  c0 ı2 D 0I
u1 C c0 a1  ı1 D 0I u1 C c0 a2  ı1 D 0:
(9)
As in the case of the Heaviside function enrichment, it is immediate to observe
that the indeterminate equations tie standard and enrichment variables ui and ai .
Consequently, when ill conditioning occurs, it is generated by the fact that infinite
sets of variables ui and ai are solution to the linear system. Then, adding proper

Fig. 3 A bar with a linear enrichment function and prescribed ends displacement
216 G. Ventura and C. Tesei

additional constraint equations for the variables can resolve indetermination, as

shown in the next Section.

4 The New Stabilization Technique

To eliminate the indeterminacy a fictitious constraints can be introduced for the

variables. Starting from the study of Eqs. (6) and (9) several forms of additional
constraints have been studied. For example, in the case of Heaviside function
enrichment, considering (7), an additional constraint equation of the form

ui  H.xi /ai D 0 (10)

has been tested and allows to obtain good results. The rationale in (10) is that it
is an equation always linear independent from (7), as H.xi / shows reversed sign
in the two equations. Although (10) works nicely with Heaviside enrichments, its
analytical study shows that it is not able of eliminating indeterminacy in the case of
nearly linear enrichments.
The most efficient and universal stabilization equation that has been found is
simply biasing to zero the enrichment variables ai . The term “biasing” and not
“setting” is used as, when the indeterminacy holds strictly (i.e. the Heaviside
function has the discontinuity at a node or the additional enrichment function
is linear), then the enrichment variables ai can be set to zero. However, when
indeterminacy is caused by nearly linear enrichment or by an Heaviside function
having the discontinuity in proximity of a node, setting the enrichment variables to
zero will cause approximations of difficult determination.
The ideal stabilization should be such that, if indeterminacy would take place, it
is stabilized. If applied to elements and cases where it is not needed it will leave the
solution unaltered.
Therefore, the enrichment variables ai are biased to zero by adding a quadratic
penalty term to the total potential energy functional as follows
Z 2
1 L
1 X 2
˘D A.x/".x/ dx C r1 .u.0/  ı1 // C r2 .u.L/  ı2 // C ˛ a: (11)
2 0 2 iD1 i

The added term states, in the penalty method light, that the enrichment variables
ai should be zero for each i. Enforcing strictly this condition would produce an
erroneous solution to our problem, and this would happen for high values of the
penalty parameter ˛. However, as the indeterminacy of the system of equations
is to be ruled out, the minimum possible value for ˛ will be chosen to recover a
numerically well conditioned problem.
Therefore, in contrast to penalty methods, where the penalty parameter is set
very high to enforce as more precisely as possible the constraint (leading again to
Stabilized X-FEM for Heaviside and Nonlinear Enrichments 217

ill conditioning), here the constraint is fictitious and the penalty parameter is set the
lowest possible. Physically, this is equivalent to introduce a kind of weak spring for
the enriched variables.
Note that, as the penalty term is not integrated at the element level, the
stabilisation can be applied on the global stiffness matrix and its implementation
is straightforward in any dimensions and for any element type.

5 Method Validation

The proposed methodology is applied to some 1D to 3D examples to show its

effectiveness. In particular, the 1D problems of the bar with a cut and the bar with
linear enrichment will be examined. Then, two 3D examples of a parallelepiped with
a crossing surface discontinuity will be shown to observe the marked improvement
in problem conditioning and the limited approximation introduced by the proposed
technique.

5.1 1D Bar with a Cut

Reference is made to the 1D problem of the bar with a cut of Fig. 2. Although in this
case stabilization is needed only when d approaches 0 or L, it is useful to observe the
influence of the stabilization term in the solution when the discontinuity is placed far
from the ends. If the stabilization will not alter the solution even when not needed,
the problem of distinguishing whether it is needed or not will be automatically ruled
out, and this is an important objective of the present technique in order to drastically
simplify its application. Of course, criteria can be introduced to selectively apply
stabilization and reduce stabilization induced approximation.
The exact solution of the problem of Fig. 2 for ı1 D 0, ı2 D 1 is given by
zero strain, zero displacement on the left of the cut .x < d/, unit displacement on
the right of the cut .x > d/. Figure 4 shows in the three columns the solution for
d D 0, d D 0:3 L, d D L. In the first two rows the plots of displacement and
strain are given for a proper value of the penalty parameter. In the present case,
numerical experiments have shown that a good value for the penalty parameter ˛ is
given by 1=1000 the maximum diagonal term in the stiffness matrix, giving a system
condition number (intended as ratio of the largest to the smallest matrix eigenvalue)
of 4250 when the discontinuity is at one end of the bar.
Finally, the last row in Fig. 4 shows plots of the condition number with ˛. It
can be observed that, when stabilization is not needed .d D 0:3 L/, any value of
alpha does not impact significantly the condition number. On the other hand, when
stabilization is essential to get the problem solution by ruling out indeterminacy,
excessive low values of the penalty parameter cause the condition number to blow
up as the stabilization term tends to disappear.
218 G. Ventura and C. Tesei

Fig. 4 Stabilized solutions for the reference problem of the bar with a cut of Fig. 2. Columns show
three positions of the discontinuity (d D 0; 0:3 L; L/. The first two rows show displacement and
strain for ˛ D 0:001 while the last row plots the system condition number with varying ˛

The solutions of Fig. 4 show displacement practically equal to the exact solution
and the approximation introduced by the present stabilization technique is visible
in the strain plot where, instead of zero strains (exact solution), a small error is
observed, being " below 2‰. In the cases where stabilization is really needed (first
and third column of Fig. 4) the error in strain is 0:5‰. This approximation is quite
good if it is considered that the strain in absence of discontinuity is ı2 =L D 1, so
that the error in strain introduced by the stabilization is of the order of 1‰.
To conclude the study of the present example, plots of the solution in the case
d D 0 (cut at the left end) are given for three values of ˛ in Fig. 5. Solution in
terms of displacement and strain are plotted for ˛ D 1:0; 0:1; 0:01. These values
have been chosen large otherwise the effect in the approximation introduced by
stabilization would not be visible. It can be observed how large values of the
stabilization parameter give a slope to the displacement plot and larger residual
strain.

5.2 1D Bar with Linear Enrichment

Reference is made to the 1D problem of the bar with a linear enrichment function (8)
of Fig. 3 representing the degenerate case of a nonlinear enrichment function that has
become approximately linear in the element. As previously observed, in this case
Stabilized X-FEM for Heaviside and Nonlinear Enrichments 219

Fig. 5 Stabilized solutions for the reference problem of the bar with a cut of Fig. 2 for large values
of the penalty parameter ˛

stabilization is needed for any values of the constants c0 and c1 as the enrichment
function is already spanned by the element shape functions (the enrichment
function is linearly dependent on the shape functions).
The exact solution of this problem for ı1 D 0, ı2 D 1 is given by unit constant
strain and linearly varying displacement from zero at the left end to one at the right
end, for any values of the constants c0 and c1 defining the enrichment function.
Before showing the effectiveness of the method, it is useful to observe the
structure of the linear system involved in this case to set the penalty stabilization
parameter properly. For how the case has been constructed, the matrix of coefficients
of this system is given by the enriched element stiffness matrix plus the two
rows and columns of the Lagrangian multipliers enforcing the Dirichlet (prescribed
displacement) boundary conditions. The system matrix is given by (12), where the
system variables are ordered as follows: u1 ; u2 ; a1 ; a2 ; r1 ; r2 .
220 G. Ventura and C. Tesei

2 3
1 1 c0 c0  c1 1 0
6 1 1 c0 c0 C c1 0 1 7
6 7
6 c2 c2 7
6 c0 c0 c20 C 31 C 2˛ c20  c1 c0  31 c0 0 7
6 7 (12)
6 c 2
4c 2 7
6 c0  c1 c0 C c1 c20  c1 c0  31 c20 C 2c1 c0 C 31 C 2˛ 0 c0 C c1 7
6 7
4 1 0 c0 0 0 0 5
0 1 0 c0 C c1 0 0

Looking at the diagonal terms of the enriched variables part (third and fourth
row/column) they are proportional to the penalty stabilization parameter ˛ and to
the square of the coefficients c0 and c1 , i.e. roughly to the square of the maximum
enrichment function value in the element. Therefore, two considerations can be
made:
• large enrichment function values would cause anyway ill-conditioning. However,
this case is not of interest in applications as nonlinear enrichment functions
usually assume vanishing values far from the features that are to be represented
(for example crack tip enrichments far from the singularity);
• the value of the penalty stabilization parameter must be chosen to be suitably
large compared to the diagonal stiffness terms otherwise its contribution will
have no effect.
In the present example, in contrast to the previous case of the bar with a cut, the
exact solution is always obtained for any value of the stabilization penalty parameter
ranging by several orders of magnitude, Fig. 6. What changes is the system condition
number, that in this case is influenced both by ˛ and by c0 and c1 .
This has been studied in the following Table 1, where the system condition
number is reported for varying values of c0 and c1 in the range 1 : : : 1000 for a
penalty parameter ˛ given by 1=1000 the maximum diagonal term in the stiffness
matrix. When the enrichment coefficients become large compared to element
stiffness terms, ill conditioning is observed anyway as previously pointed out, but
this feature is independent of the proposed stabilization method.

5.3 Three Dimensional Test

The proposed stabilization technique is finally tested on two 3D benchmark

examples, similar to the one dimensional cases where the problem has been studied
analytically. However, while in the 1D cases the stabilization parameter has been
fixed arbitrarily, here a study is presented for selecting the best stabilization penalty
parameter in terms of optimal system conditioning. The two considered cases are:
• a bar with a slant cut with standard Heaviside function enrichment;
• a bar with an orthogonal cut where the enrichment is a regularized Heaviside
function representing the case of quasi linear enrichment.
Stabilized X-FEM for Heaviside and Nonlinear Enrichments 221

Fig. 6 Stabilized solutions for the reference problem of the bar with linear enrichment of Fig. 3

Table 1 Values of the system condition number for varying enrichment coefficients and constant
value of the stabilization parameter
c0 0 1 10 1 100 1 1000 1 1000
c1 0 1 1 10 1 100 1 1000 1000
Condition number 1E3 3E3 1E5 3E4 1E7 3E6 1E9 3E8 2E9

The physical problem is similar to the example in Fig. 2. However, a 3D

parallelepiped is used in place of the 1D bar element. The problems have been
studied by the software X3D developed by the Author. It is a FORTRAN code using
GMSH [17] as pre and post-processor. The code is able of dealing with problems of
several hundred thousand elements with high efficiency and has been developed as
research platform.
In the two examples a linear elastic material with a Young modulus E D 10 GPa
and a Poisson coefficient  D 0 is assumed. The parallelepiped has one end fixed
and the other displaced by 1 mm in the longitudinal direction. Let  D max.Kii /
the value of the largest diagonal element in the stiffness matrix. The penalty
222 G. Ventura and C. Tesei

Fig. 7 A 3D parallelepiped with a slant cutting plane and prescribed displacements at the ends

stabilization parameter is set as


˛D (13)
!
where ! is a scaling factor that will be varied to show how the condition number is
affected by the stabilization. The largest the ! the smallest the stabilization penalty
parameter will be set.
In the first case, a parallelepiped with dimensions 100  100  400 mm is
considered, with a mesh of 14;440 hexahedral elements (19  19  40 elements)
and the discontinuity is a slanted cut. In Fig. 7 the displacement contours at solution
show the physics of the problem.
This problem cannot run without stabilization as the linear system solver exits
with zero pivot error detection and numerically infinite condition number unless a
cutting plane avoiding all element nodes is considered.
To fix the order of magnitude of the condition number, the same problem,
analyzed as linear elastic (without discontinuity), has a system condition number
of 6E C 5.
Stabilized X-FEM for Heaviside and Nonlinear Enrichments 223

Table 2 Values of the system condition number for varying stabilization parameter scaling factor
in the 3D cut example
! 1E C 3 1E C 4 1E C 5 1E C 6
Condition number 2E C 6 2E C 6 4E C 6 4E C 7

Fig. 8 Graph of the regularized Heaviside function (14)

Table 2 summarizes how the system condition number varies with the stabiliza-
tion penalty parameter scaling factor ! when the cut discontinuity is introduced. As
observed in the one-dimensional study, the smallest stabilization penalty parameter
(and therefore the largest !) will yield the lowest approximation error in terms
of residual traction between the two parts separated by the cut. The best result is
obtained for ! D 1E C 5, that is the largest ! value not implying an increase
in condition number. In this case the residual stress due to the introduction of the
penalty stabilization has been found to be 0:02 MPa, that is 8‰ of the stress in the
linear elastic case with no discontinuity (equal to 25 MPa).
This result is quite good and evidences the only limitation of the methodology,
i.e. when very large displacements are present the closing residual stresses due to
the penalty stabilization may be large. However, the methodology is certainly sound
for small displacement mechanics.
The case of the bar with quasi linear enrichment is equal to the previous example
but, instead of using an exact Heaviside function enrichment, its regularized form
has been considered
2
H .d/ D arctan.d=/ (14)

Function (14) is plotted in Fig. 8. For decreasing  values tends to the exact
Heaviside function (2). Far from the discontinuity this function can be regarded
as linear or approximately constant. When this function is inserted as enrichment
function into the XFEM displacement approximation (1) and an enrichment
domain of width
extending over several elements is considered, the expected
224 G. Ventura and C. Tesei

behavior is:
• a well conditioned problem when the enrichment domain width
is small
compared to the regularization parameter . This because for
comparable to 
the function is markedly nonlinear in the enrichment domain;
• a progressively more ill conditioned problem as the enrichment domain width

increases because the enrichment function becomes approximately linear in the

enriched elements.
It is important to mark again that the purpose of this example is to illustrate
how ill-conditioning may arise when using enrichment functions that decay to linear
functions with the distance from features and these functions are directly introduced
in the expression (1). However, when a proper variational formulations is used [8,
10, 11], the ill-conditioning issue may be automatically ruled out.
This problem has been studied on a 100  100  1200 mm parallelepiped
discretized by 9  9  30 elements (2430 hexahedra), same material properties
and boundary conditions as the previous example and a value for the regularization
parameter  D 0:003.
The expected behavior is fully confirmed by numerical experiments, whose
results are reported in Tables 3 and 4, where the enrichment area width
has been
varied in between 50 and 500 mm. For the sake of comparison it is recalled that the
system condition number for the linear elastic case (no regularized discontinuity)
is 2:0E C 6. Figure 9 reports the displacement contours at solution, that are
undistinguishable for all the examined cases.
Table 3 reports the condition number in the case of absence of stabilization. With
respect to the linear elastic case (no discontinuity) the condition number appears
six orders of magnitude larger for
D 50 and diverges to numerical infinity for

D 500.

Table 3 Values of the system condition number for varying enrichment domain width in the 3D
cut with regularized enrichment example and no stabilization

50 100 200 300 400 500

Condition number 8E C 12 7E C 13 1E C 15 8E C 16 4E C 19 1

Table 4 Values of the ! 1E C 3 1E C 4 1E C 5 1E C 6

system condition number for
varying enrichment domain
D 50 2E C 7 2E C 7 2E C 7 5E C 7
width in the 3D cut with
D 100 2E C 7 2E C 7 2E C 7 5E C 7
regularized enrichment
D 200 2E C 7 2E C 7 2E C 7 5E C 7
example and penalty
D 300 2E C 7 2E C 7 2E C 7 5E C 7
stabilization
D 400 2E C 7 2E C 7 2E C 7 5E C 7

D 500 2E C 7 2E C 7 2E C 7 5E C 7
Stabilized X-FEM for Heaviside and Nonlinear Enrichments 225

Fig. 9 Displacement contours for the regularized cut problem

On the other hand, when stabilization is introduced, Table 4, a very limited

increase in the condition number is observed. The condition number remains
practically constant for a wide range of the stabilization parameter scaling factor
! despite of the fact that the enrichment width is varied. This shows the high
effectiveness and robustness of the proposed methodology. Assuming as optimal
value for the stabilization parameter the largest possible value before affecting the
condition number, from Table 4 the value ! D 1E C 5 is derived, similarly to the
previously analyzed case of pure jump and single element enrichment.
Similarly to the previous case of exact Heaviside enrichment, penalty stabiliza-
tion introduces some low spurious stresses at solution. These are plotted in Fig. 10
for the limiting cases of regularisation lengths
D 50 and
D 500. Considering
that when the discontinuity is absent (linear elastic case) the axial stress is equal to
8:3 MPa, from the contours in Fig. 10 an error smaller than 0:5 % is found, that is
acceptable for technical applications. Stabilisation spurious stress can be reduced by
increasing the stabilization parameter ! and accepting some impact on the system
condition number.
226 G. Ventura and C. Tesei

Fig. 10 Stabilisation spurious stresses at solution for

D 50 (left) and
D 500 (right). The
blue area is associated to the enrichment width

6 Conclusions

The addition of a stabilization penalty term biasing to zero the enrichment variables
has proved to be an effective and very simple technique to eliminate indeterminacy
in the system of equations of XFEM and XFEM-like formulations. In the paper
the analytic structure of ill conditioning has been shown and numerical studies
performed on 1D and 3D cases have evidenced a marked improvement in the
condition number of the system of equations. Due to the extreme simplicity of
the approach, the absence of additional computational cost and the possibility of
leaving the basic problem formulation unchanged, the presented method appears a
valid alternative to the existing techniques already present in the literature.

References

1. ABAQUS: ABAQUS Documentation. Dassault Systèmes, Providence (2014)

2. Abdelaziz, Y., Hamouine, A.: A survey of the extended finite element. Comput. Struct. 86,
1141–1151 (2008)
3. Babuška, I., Banerjee, U.: Stable generalized finite element method (SGFEM). Comput.
Methods Appl. Mech. Eng. 201–204, 91–111 (2012)
4. Babuška, I., Banerjee, U., Osborn, J.: Survey of meshless and generalized finite element
methods: a unified approach. Acta Numer. 12, 1–125 (2003)
5. Béchet, E., Minnebo, H., Moës, N., Burgardt, B.: Improved implementation and robustness
study of the X-FEM for stress analysis around cracks. Int. J. Numer. Methods Eng. 64, 1033–
1056 (2005)
6. Belytschko, T., Gracie, R., Ventura, G.: A review of extended/generalized finite element
methods for material modeling. Model. Simul. Mater. Sci. Eng. 17, p. Article number 043001
(2009)
Stabilized X-FEM for Heaviside and Nonlinear Enrichments 227

7. Benvenuti, E.: Mesh–size–objective XFEM for regularized continuous–discontinuous transi-

tion. Finite Elem. Anal. Des. 47, 1326–1336 (2011)
8. Benvenuti, E.: XFEM with equivalent eigenstrain for matrix–inclusion interfaces. Comput.
Mech. 53, 893–908 (2014)
9. Benvenuti, E., Tralli, A.: Simulation of finite-width process zone for concrete-like materials.
Comput. Mech. 50, 479–497 (2012)
10. Benvenuti, E., Tralli, A., Ventura, G.: A regularized XFEM model for the transition from
continuous to discontinuous displacements. Int. J. Numer. Methods Eng. 74, 911–944 (2008)
11. Benvenuti, E., Ventura, G., Ponara, N., Tralli, A.: Variationally consistent extended FE model
for 3D planar and curved imperfect interfaces. Comput. Methods Appl. Math. Eng. 267, 1–22
(2013)
12. Berger-Vergiat, L., Waisman, H., Hiriyur, B., Tuminaro, R., Keyes, D.: Inexact Schwarz-
algebraic multigrid preconditioners for crack problems modeled by extended finite element
methods. Int. J. Numer. Methods Eng. 90, 311–328 (2012)
13. Duarte, C., Kim, D.: Analysis and applications of a generalized finite element method with
global–local enrichment functions. Comput. Methods Appl. Mech. Eng. 197, 487–504 (2008)
14. Ferté, G., Massin, P., Moës, N.: Convergence analysis of linear or quadratic X-FEM for curved
free boundaries. Comput. Methods Appl. Mech. Eng. 278, 794–827 (2014)
15. Fries, T.: A corrected XFEM approximation without problems in blending elements. Int. J.
Numer. Methods Eng. 75, 503–532 (2008)
16. Fries, T.P., Belytschko, T.: The extended/generalized finite element method: an overview of the
method and its applications. Int. J. Numer. Methods Eng. 84, 253–304 (2010)
17. Geuzaine, C., Remacle, J.-F.: Gmsh: A 3-d finite element mesh generator with built-in pre-and
post-processing facilities. Int. J. Numer. Methods Eng. 79, 1309–1331 (2009)
18. Gupta, V., Duarte, C.A., Babuška, I., Banerjee, U.: Stable GFEM (SGFEM): improved con-
ditioning and accuracy of GFEM/XFEM for three-dimensional fracture mechanics. Comput.
Methods Appl. Mech. Eng. 289, 355–386 (2015)
19. Koehler, J.: On the dislocation theory of plastic deformation. Phys. Rev. 60, 397–410 (1941)
20. Laborde, P., Pommier, J., Renard, Y., Salaun, M.: High order extended finite element method
for cracked domains. Int. J. Numer. Meth. Eng. 64, 354–381 (2005)
21. Lang, C., Makhija, D., Doostan, A., Maute, K.: A simple and efficient preconditioning scheme
for heaviside enriched XFEM. Numer. Anal. arXiv:1312.6092v1 (2013)
22. Loehnert, S.: A stabilization technique for the regularization of nearly singular extended finite
elements. Comput. Mech. 54, 523–533 (2014)
23. LS-DYNA: User’s Manual. Livermore Software Technology Corporation, Livermore (2013)
24. Martin, A., Esnault, J.-B., Massin, P.: About the use of standard integration schemes for X-
FEM in solid mechanics plasticity. Comput. Methods Appl. Mech. Eng. 283, 551–572 (2015):
25. Menk, A., Bordas, S.P.A.: A robust preconditioning technique for the extended finite element
method. Int. J. Numer. Methods Eng. 85, 1609–1632 (2011)
26. Mousavi, S.E., Pask, J.E., Sukumar, N.: Efficient adaptive integration of functions with sharp
gradients and cusps in n-dimensional parallelepipeds. Int. J. Numer. Methods Eng. 91, 343–357
(2012)
27. Müller, B., Kummer, F., Oberlack, M.: Highly accurate surface and volume integration on
implicit domains by means of moment-fitting. Int. J. Numer. Methods Eng. 96, 512–528 (2013)
28. Nabarro, F.: Theory of Crystal Dislocations. Dover Publications, New York (1967)
29. Sillem, A., Simone, A., Sluys, L.J.: The orthonormalized generalized finite element method-
OGFEM: efficient and stable reduction of approximation errors through multiple orthonormal-
ized enriched basis functions. Comput. Methods Appl. Mech. Eng. 287, 112–149 (2015)
30. Strouboulis, T., Babuška, I., Copps, K.: The design and analysis of the generalized finite
element method. Comput. Methods Appl. Mech. Eng. 181, 43–69 (2000)
31. Sudhakar, Y., Wall, W.A.: Quadrature schemes for arbitrary convex/concave volumes and
integration of weak form in enriched partition of unity methods. Comput. Methods Appl. Mech.
Eng. 258, 39–54 (2013)
228 G. Ventura and C. Tesei

32. Ventura, G.: On the elimination of quadrature subcells for discontinuous functions in the
extended finite-element method. Int. J. Numer. Methods Eng. 66, 761–795 (2006)
33. Ventura, G., Moran, B., Belytschko, T.: Dislocations by partition of unity. J. Numer. Methods
Eng. 62, 1463–1487 (2005)
34. Ventura, G., Gracie, R., Belytschko, T.: Fast integration and weight function blending in the
extended finite element method. Int. J. Numer. Methods Eng. 77, 1–29 (2009)
35. Ventura, G., Benvenuti, E.: Equivalent polynomials for quadrature in heaviside function
enriched elements. Int. J. Numer. Methods Eng. 102, 688–710 (2015)
36. Waisman, H., Berger-Vergiat, L.: An adaptive domain decomposition preconditioner for crack
propagation problems modeled by XFEM. Int. J. Multiscale Comput. Eng. 11, 633–654 (2013)
37. Yazid, A., Abdelkader, N., Abdelmadjid, H.: A state-of-the-art review of the X-FEM for
computational fracture mechanics. Appl. Math. Model. 33, 4269–4282 (2009)
38. Zhang, Q., Banerjee, U., Babuška, I.: Higher order stable generalized finite element method.
Numer. Math. 128, 1–29 (2014)
An Adaptive Fictitious Domain Method
for Elliptic Problems

Stefano Berrone, Andrea Bonito, and Marco Verani

Abstract In the Fictitious Domain Method with Lagrange multiplier (FDM) the
physical domain is embedded into a simpler but larger domain called the fictitious
domain. The partial differential equation is extended to the fictitious domain
using a Lagrange multiplier to enforce the prescribed boundary conditions on the
physical domain while all the other data are extended to the fictitious domain. This
lead to a saddle point system coupling the Lagrange multiplier and the extended
solution of the original problem. At the discrete level, the Lagrange multiplier is
approximated on subdivisions of the physical boundary while the extended solution
is approximated on partitions of the fictitious domain. A significant advantage of
the FDM is that no conformity between these two meshes is required. However,
a restrictive compatibility conditions between the mesh-sizes must be enforced to
ensure that the discrete saddle point system is well-posed. In this paper, we present
an adaptive fictitious domain method (AFDM) for the solution of elliptic problems in
two dimensions. The method hinges upon two modules ELLIPTIC and ENRICH
which iteratively increase the resolutions of the approximation of the extended
solution and the multiplier, respectively. The adaptive algorithm AFDM is convergent
without any compatibility condition between the two discrete spaces. We provide
numerical experiments illustrating the performances of the proposed algorithm.

S. Berrone
Dipartimento di Scienze Matematiche, Politecnico di Torino, Corso Duca degli Abruzzi,
24 - 10129 Torino, Italy
e-mail: [email protected]
A. Bonito
Department of Mathematics, Texas A&M University, College Station, TX 77843-3368, USA
e-mail: [email protected]
M. Verani ()
MOX-Dipartimento di Matematica, Politecnico di Milano, P.zza L. da Vinci, I-20132 Milano,
Italy
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 229

G. Ventura, E. Benvenuti (eds.), Advances in Discretization Methods,
SEMA SIMAI Springer Series 12, DOI 10.1007/978-3-319-41246-7_11
230 S. Berrone et al.

1 Introduction

In many engineering applications the efficient numerical solution of partial differen-

tial equations on deformable or complex geometries is of paramount importance. In
this respect, one crucial issue is the construction of the computational grid. To face
this problem, one can basically resort to two different types of approaches. In the
first approach, a mesh is constructed on a sufficiently accurate approximation of the
exact physical domain (see, e.g., isoparametric finite elements [8], or isogeometric
analysis [9]), while in the second approach one embeds the physical domain into
a simpler computational mesh whose elements can intersect the boundary of the
given domain. Clearly, the mesh generation process is extremely simplified in
the second approach, while the imposition of boundary conditions requires extra
work. Among the huge variety of methods sharing the philosophy of the second
approach, let us mention here the Immersed Boundary methods (see, e.g., [16] and
the references therein), the Penalty Methods (see, e.g., the seminal work [2]), the
Fictitious Domain/Embedding Domain Methods (see, e.g., [7] and the references
therein). In this paper, we focus on the Fictitious Domain Method with Lagrange
multiplier (FDM) introduced in [11, 12] (see also [1] for the pioneering work
inspiring this approach). In this approach, the physical domain ! is embedded
into a simpler and larger domain ˝ (the fictitious domain), the right-hand side
is extended to the fictitious domain and the boundary conditions on the boundary
of the physical domain are appended through the use of a Lagrange multiplier.
The FDM gives rise to a saddle point problem whose exact solution restricted to
! corresponds to the solution of the original problem. At the discrete level, the
FDM allows the use of structured and uniform meshes in the fictitious domain,
without requiring any conformity between the bulk mesh and the boundary of the
physical domain. This represents a relevant computational advantage. However,
there are two important issues to be taken into account to build numerical techniques
that are able to take advantage of the crucial features of FDM: the choice of the
discrete spaces for the approximation of the solution and the multiplier, and the
construction of the extension of the right-hand side from the physical domain to the
fictitious one. As pointed out in the analysis performed in [11] in the context of finite
element approximation for the solution of elliptic problems with Dirichlet boundary
conditions, the first condition (inf-sup condition) is essential to ensure existence and
uniqueness of the discrete solution, while the second one influences the regularity of
the extended continuous solution, thus impacting on the approximation properties
of the discrete spaces (on this latter topic, see, e.g., [14]). The first condition turns
out to introduce some restrictive compatibility conditions between the mesh-sizes
of the fictitious domain grid and the subdivision of boundary of the physical domain
(needed to approximate the Lagrange multiplier). The second issue can spoil, for
example, the performance of the linear finite element method on uniform meshes
whenever the original solution is sufficiently regular, e.g. H 2 regular, while the
extended solution is less regular.
AFDM for Elliptic Problems 231

In view of the above discussion, the computational effectivity of FDM seems

to be a non trivial issue. However, as shown in the present paper, a judicious use
of adaptivity can allow to overcome the above two obstructions and recover the
full potentiality of FDM, i.e. working with discrete spaces (and meshes) violating
the compatibility conditions and recover even in presence of less regular extended
solutions, the optimal performance of finite elements on uniform grids in presence
of regular solutions. In particular, in this work we present an adaptive fictitious
domain method, named AFDM, based on the use of linear finite elements for the
approximation of the solution and piecewise constants for the approximation of the
Lagrange multiplier. In the spirit of the algorithm provided in [6], the method hinges
upon two modules, ELLIPTIC and ENRICH that iteratively modify the discrete
spaces for the approximation of the extended solution and the multiplier. Our
method is proved to be convergent regardless of the imposition of any compatibility
condition between the two discrete spaces. Similar remarks has been already pointed
out in different contexts by Dahlke et al. [10] and Bänsch et al. [4] (see also [3]
for an abstract discussion of inexact Uzawa methods) and we refer to [13] for the
mathematical study of an adaptive algorithm for the Stokes system, which serves as
a benchmark for saddle point problems. Moreover, preliminary numerical results
show that AFDM is optimal with respect to the number of degrees of freedom
employed to approximate the extended solution and the Lagrange multiplier. In two
dimension, the optimality of the adaptive refinement strategy seems to require an
adaptive strategy to generate the successive subdivision of the fictitious domain,
while uniform or quasi-uniform subdivisions can be used for the boundary mesh.
The outline of the paper is as follows. In Sect. 2 we introduce the fictitious
domain method, while in Sect. 3 we introduce the adaptive fictitious domain
method. Finally, in Sect. 4 we numerically explore the convergence (and optimality)
properties of our algorithm.

2 Fictitious Domains Method

Let ! be a bounded domain of R2 with boundary  . To make the presentation

simpler, we assume that ! is a polygon. We are interested in employing the
fictitious domain method to solve the following model problem: let f 2 L2 .!/,
find u 2 H01 .!/ such that


u D f in ! ; (1)
uD0 on  : (2)
232 S. Berrone et al.

The fictitious domain formulation of problem (1)–(2) hinges on a square or

rectangular domain ˝ with boundary WD @˝ and such that !  ˝. It reads:
1
for any L2 -extension fQ of f to ˝, find .Qu; / 2 H01 .˝/  H  2 . / such that
Z Z
r uQ  rv  h; vi D fQv dx 8v 2 H01 .˝/ ; (3)
˝ ˝
1
h; uQ i D 0 8 2 H  2 . / ; (4)

where for v 2 H01 .˝/, its restriction to  is understood in the sense of traces, and
1 1
h; i denotes the duality pairing between H  2 . / and H 2 . / (recall that  is a
closed curve). Using an integration by parts formula, it is immediate to verify (see,
e.g., [12]), that the fictitious domain formulation (3)–(4) is equivalent to original
formulation (1)–(2) where
h @Qu i
D (5)
@n 

is the jump of @Q@n across  and n denotes the unit normal exterior to !.
u

The regularity of uQ and  depends on the extension chosen for f and the domain
3
!. In the worst case, uQ 2 H 2  .˝/ for any  > 0 and  2 L2 . / satisfies  2
1
H 2 .i /, for every straight line i composing  (see [11]). In what follows we will
drop the symbol Q if no confusion arises.
In [11], the Babuska-Brezzi’s theory is used to guarantee
R that problem (3)–(4)
is well posed. In particular, the bilinear form .u; v/ 7! ˝ ru  rv is coercive on
the set fv 2 H01 .˝/ W h; vi D 0 8 2 H 1=2 . /g and that the following inf-sup
condition holds: there exists a constant  > 0 such that

hv; i
inf sup : (6)
v2H 1 .˝/ kk  21 kvkH 1 .˝/
1
2H  2 . / 0 H . /

3 Adaptive Fictitious Domain Method

In this section we present our adaptive fictitious domain method (AFDM) based
on Uzawa iterations. In Sect. 3.1, we describe the infinite dimensional version of
the algorithm, whereas in Sect. 3.2 we introduce its adaptive finite dimensional
counterpart.
AFDM for Elliptic Problems 233

3.1 Infinite Dimensional Fictitious Domain Algorithm

We start this section by describing the infinite dimensional fictitious domain

algorithm for solving (3)–(4). It consists of Uzawa-type successive iterations: Given
1
˛ > 0 and 0 2 H  2 . / we seek, for j  1,
Z Z
uj 2 H01 .˝/ W ruj  rv D f v C hj1 ; vi 8v 2 H01 .˝/ ; (7)
˝ ˝
1 1
j 2 H  2 . / W .j ; / D .j1 ; /  ˛ .uj ; / 8 2 H  2 . / (8)
1
where we denote by .; / the scalar product in H  2 . / and where we used the
1 1
identification of L2 . / and have with a slight abuse of notation H 2 . /  H  2 . /.
1 1 1
The Schur complement operator S W H  2 . / ! H 2 . /  H  2 . / defined as

S D u ; (9)

where u 2 H01 .˝/ is the solution to

Z
ru  rv D h; vi 8v 2 H01 .˝/ :
˝

The operator S is symmetric and positive definite [12] and is instrumental in the
analysis of the Uzawa iterations. In fact, (7)–(8) can be written using S as
1
j D .I  ˛S/j1 C ˛uf in H  2 . /; (10)

where uf 2 H01 .˝/ is given by

Z Z
ruf  rv D fv 8v 2 H01 .˝/:
˝ ˝

It is immediate to verify that if 0 < ˛ < 2=kSk 1 1 , then

L .H  2 . /;H  2 . //

ˇ WD kI  ˛Sk 1 1 <1 (11)

L .H  2 ;H  2 /

and thus the infinite dimensional fictitious domain algorithm is convergent.

3.2 Adaptive Finite Dimensional Fictitious Domain Method

We now introduce our adaptive finite dimensional fictitious domain algorithm

(AFDM) which iteratively builds a sequence of nested finite dimensional spaces to
achieve a reduction of the approximation error between each iterative step. We start
234 S. Berrone et al.

with an initial conforming subdivision T0 of ˝ made of triangles and an initial

subdivision of  (made of segments). We assume that

for each T 2 T0 ; TV \  is connected, (12)

which is automatically satisfied upon assuming that the initial subdivision T0

is sufficiently fine to capture the interface and could be enforced via uniform
refinements without affecting the asymptotic performances of the algorithm. From
now on, j  0 will always denote the AFDM iteration counter. We denote by Tj
and Sj the j-th conforming partitions of ˝ and  made of triangles and segments,
respectively. The diameters of the elements T 2 Tj and ` 2 Sj are denoted
hT WD diam.T/ and h` WD diam.`/, respectively. We emphasize that the two
partitions built by the adaptive algorithm described below are mutually independent
and in particular no compatibility conditions between the two partitions is required.
This is a crucial difference from the results in [11] (see Remark 2 below). The
shape regularity constant of a generic subdivision T is maxT2T hTT , where T is the
diameter of the largest ball inside T. The shape regularity constant of a sequence of
subdivision fTi gi0 is

hT
sup max :
i0 T2Ti T

Associated with conforming partitions T of ˝ and S of  , we introduce the

finite dimensional spaces
˚ 
VT WD v 2 C0 .˝/ W vjT 2 P 1 .T/ 8T 2 Tj \ H01 .˝/

and

MS WD fw 2 L2 . / W wj` 2 P 0 .`/ 8` 2 S g;

where for k 2 N, P k .D/ is the set of polynomials of degree k in D. In the following,

we set Vj WD VTj and Mj WD MSj .
The finite dimensional adaptive fictitious domain algorithm relies on two sub-
routines described now.

3.2.1 The Module ELLIPTIC

The module ELLIPTIC adaptively constructs approximations Uj of the exact

solution uj to (7). To describe this procedure, we let uj 2 H01 .˝/ satisfying
Z Z Z
ruj  rv D f vC j1 v 8v 2 H01 .˝/ ; (13)
˝ ˝ 

for a given j1 2 Mj1 .

AFDM for Elliptic Problems 235

1
In contrast to (7) where j1 2 H  2 . /, we note that j1 belongs to a
finite dimensional subspace of L2 . /. Moreover, we observe that (13) is a weak
formulation of


uj D f in !; 
uj D f in ˝ n !;
 
@uj
Œuj D 0 on ; D j1 on ; (14)
@n 
uj D 0 on @˝:

If "j stands for an adjustable error tolerance, then the module ELLIPTIC,

.Tj ; Uj / D ELLIPTIC.Tj1 ; j1 ; "j /;

constructs adaptively a refined mesh Tj of Tj1 such that the solution of the discrete
elliptic problem
Z Z Z
Uj 2 Vj W rUj  rV D f VC j1 V 8V 2 Vj ; (15)
˝ ˝ 

approximate uj within in the prescribed tolerance "j , i.e.

kr.uj  Uj /kL2 .˝/  "j : (16)

The adaptive ELLIPTIC module iterates a classical strategy of the type

SOLVE  ESTIMATE  MARK  REFINE

until condition (16) is satisfied (see, e.g., [15]).

The following upper bound for the error of any inner-iterate is instrumental in
ESTIMATE. A corresponding lower bound is also valid and we refer to [5] for their
proofs.
Proposition 3.1 (Upper Bound for ELLIPTIC) Let uj 2 H01 .˝/ be the solution
to (13) and Uji 2 Vi be the discrete Galerkin solution to (15) associated with any
refinement T i of Tj1 . Assume that the initial subdivision satisfy (12). Then, there
exists a positive constant K  only depending on T i through its shape-regularity
constant such that the following a posteriori error estimate holds
X 2
kr.uj  Uji /k2L2 .˝/  K  i .T/ (17)
T2T i
236 S. Berrone et al.

where
2 X
i .T/ WD h2T kRiT k2L2 .T/ C hT kJei k2L2 .e/ C hT kj1 k2 2 V
;
L .T\ /
e edge of T

with
( @Uji
Π@n e  j1 on e \ ;
RiT WD . f C
Uji /jT ; Jei WD @Uji
(18)
Π@n e on e n :

Here Œ: e is the jump across the edge e and TV is the topological interior of T.
Remark 1 (Elliptic Estimator) We note that the non-standard terms containing j1
in the estimator i .T/ above both measure the discrepancy from the exact relation
 D Π@u
@n  , see (5). In particular, within an element T, rUj is continuous so that
i

" #
@Uji
kj1 kL2 .T\
V Dk  j1 kL2 .T\
V ;
/ @m V
/
T\

@U i @Uji
where Π@mj T\
V denotes the jump of @m across TV \  and m denotes one of the
normal to TV \  .

3.3 The Module ENRICH

 
Let Vj be the restriction of functions in Vj . We denote by ˘j W Vj ! Mj the
orthogonal L2 -projection onto Mj . The module

Sj D ENRICH.Sj1 ; Uj ; "j / ; (19)

constructs a refinement Sj of Sj1 such that

k.I  ˘j /Uj k 1  C"j (20)

H  2 . /

for a given C independent of j.

The quantity k.I  ˘j /Uj k  21 is not computed exactly but estimated as
H . /
follows. Standard interpolation error estimates together with a trace estimate and
the stability of the L2 -projection yield
R R
 .I  ˘j /Uj   .I  ˘j /Uj .  ˚/
k.I  ˘j /Uj kH 1=2 . / D sup D sup
2H 1=2 . / kkH 1=2 . / 2H 1=2 . / kkH 1=2 . /
AFDM for Elliptic Problems 237

1=2 1=2
kh` .I  ˘j /Uj kL2 . / kh` .  ˚/kL2 . /
 sup
1=2
2H . / kkH 1=2 . /
1=2
 Ckh` .I  ˘j /Uj kL2 . /  C.max h` /1=2 k.I  ˘j /Uj kL2 . /
`2Sj

 C.max h` /kUj kH 1=2 . /  C.max h` /kUj kH 1 .˝/ ;

`2Sj `

where ˚ 2 Mj and C is a generic constant independent of j. Now the stability

estimate kUj kH 1 .˝/  Ck f kL2 .˝/ imply

k.I  ˘j /Uj kH 1=2 . /  C.max h` /k f kL2 .˝/ : (21)

`

Based on this, the ENRICH routine refines recursively all the elements of Sj1
"
until they all have a mesh-size smaller than k f k j2 . The resulting mesh is output of
L .˝/
ENRICH since it satisfies (20).

3.4 The Module UPDATE

The discrete Lagrange multiplier is updated by the module UPDATE

j D UPDATE.Tj ; Sj ; j1 ; Uj ; ˛/ ; (22)

which computes according to (8)

j 2 Mj W j D j1  ˛˘j Uj : (23)


We note that as Uj 2 Vj its restriction to  is an element of Vj (by construction).

3.5 The AFDM Algorithm

We are now in a position to detail the iterative structure of AFDM. Each iteration
of the algorithm consists of an inner solver employing ELLIPTIC in place of (7),
followed by an application of the module ENRICH and by an update of the multiplier
performed by the module UPDATE.
238 S. Berrone et al.

ADAPTIVE FICTITIOUS DOMAIN METHOD (AFDM)

Given the initial triangulations T0 and S0 , and
the parameters ˛, "0 > 0, 0 <  < 1 set j D 1 and
iterate:
1. Select any function 0 2 M0 .
2. Update "j "j1 .
3. Compute .Tj ; Uj / D ELLIPTIC.Tj1 ; j1 ; "j /.
4. Enrich Sj D ENRICH.Sj1 ; Uj ; "j /.
5. Compute j D UPDATE.Tj ; Sj ; j1 ; Uj ; ˛/.
6. Update j j C 1.
7. Go to step (2).

The algorithm AFDM is convergent [5] as reported in the theorem below.

Theorem 1 (Convergence) Let ˛ > 0 be such that (11) holds and assume that
the initial subdivision satisfy (12). Let .Tj ; Sj ; Uj ; j / be the sequence of meshes
and subdivision produced by AFDM . There exist positive constants C1 and ı < 1
depending on the shape regularity constant of fTj gj0 such that

kr.u  Uj /kL2 .˝/ C k  j k 1  C1 ı j : (24)

H  2 . /

Remark 2 (Compatibility Condition) In [11] the authors obtain a priori error esti-
mates for the finite element approximation of (3)–(4) under the assumption that the
mesh size of the bulk triangulation is sufficiently large compared to the mesh size
of the boundary triangulation, i.e. there holds a compatibility condition between
the discrete spaces. This latter is a crucial requirement to prove the validity of
a discrete inf-sup condition and thus the existence of the discrete solution. We
emphasize that the convergence of our adaptive algorithm holds without enforcing
any compatibility condition, thus possibly violating, the discrete inf-sup condition
(see [4, 10] for similar results in different contexts).
The algorithm AFDM described above never ends until a stopping criterion is
appended. In principle, we can resort to (24) to obtain an a priori estimate for
the necessary number of iterations to reach a prescribed tolerance. However, this
strategy is not implementable as the constant ı appearing in (24) is not accessible
in practice. For this reason, we now provide an a-posteriori error estimate for the
quantity kr.u  Uj /kL2 .˝/ C k  j1 k  21 which can be employed to stop
H . /
AFDM while guarantying a prescribed approximation error.
1
Proposition 3.2 (Upper Bound for AFDM) Let .u; / 2 H01 .˝/  H  2 . / be the
solution to (3)–(4) and f.Uj ; j1 /g be the sequence of approximations produced by
AFDM. Assume that the initial subdivision satisfy (12). Then there exists a constant
AFDM for Elliptic Problems 239

C2 depending on the shape regularity constant of Tj such that

 
kr.u  Uj /kL2 .˝/ C k  j1 k 1  C2 Tj C Sj1 ; (25)
H  2 . /

where
X 1=2
Tj WD j .T/2 (26)
T2Tj
 X 1=2
Sj1 WD h` kr Uj k2L2 .`/ (27)
`2Sj1

and
X
j .T/2 WD h2T kRT;j k2L2 .T/ C hT kJe;j k2L2 .e/ C hT kj1 k2 2 V
;
L .T\ /
e edge in T

with
(
@U
Π@nj e  j1 on e \ ;
RT;j WD . f C
Uj /jT ; Je;j WD @U (28)
Π@nj e on e n :

Remark 3 (Boundary Indicator) In view of the results contained in [17], the

indicator Sj1 measures the mismatch between the trace of Uj on  and the exact
homogeneous Dirichlet boundary condition (2).

4 Numerical Results

In this section we illustrate numerically the convergence properties of the AFDM

algorithm and investigate numerically its optimality studied in [5]. Before present-
ing the numerical results we discuss some details regarding the implementation of
AFDM.

4.1 Implementation Issues

Data structures for representing the (two dimensional) triangular bulk mesh of the
fictitious domain ˝ and the (one dimensional) boundary mesh of  are organized
as binary trees starting from the 0-th level meshes T0 and S0 , respectively. The
initial bulk mesh T0 is a regular mesh of the domain ˝ constructed ignoring the
conformity with  . The initial boundary mesh S0 is made of the edges of  ,
240 S. Berrone et al.

........
....... ..
....... ....
level 3 ........ Added Boundary Point ............................ ...
... ...
....... ...
level 2 Boundary Points ............................ ....
..
level 1 ........ . . . . ...
....... ...
....... ..
level 0 ....
.......... ..........
. ..
.
.
....... ... .
........ . . . ...
....... . . ...
........... ........ ..
......
...... . . . ....
............. ................ T ...... . . ...
....... .... ................... .............. .............. ...... ..
....... ... ................................... .. ... ..... ...... .. ...
......... ..... .. ...... .......
... ...
... ..... ... . ......... ...... . .... ...
... ... . .
. ... .... ..... ...... . ...
... ... ... ... . .... ..... ......
...... . . .
.
..
... ... .
.. .... .. ... ..... ...... . .
...
... .... ... ... ... .... . .....
.. ...... ...
... ... ..
. ..... . .. . .. ... .....................
. ......
...... ...
... .. .... ... . ... .
...
...
. ...... ..
... ... ... ........ ...................... ...... .
...
... .. .. ............ ......
... ...... ...
g
..... .. ...
........ .
......... ........................... ...... ...
...
....
g
Fig. 1 Example of boundary refinement: (top-left) four different levels of refinement of the
boundary mesh and (bottom-left) associated bulk mesh with local refinement (dotted line); (right)
boundary points produced by the refinement of the bulk elements

which is assumed to be a polygonal curve, see Fig. 1(left). The refinement of the
triangular bulk mesh is performed by employing the longest edge splitting. The
refined elements are labeled as non-active and the newly created elements become
active. Notice that additional refinements of neighboring elements may be needed
in order to ensure conformity.
The refinement strategy advocated for the boundary mesh is more involved. On
each boundary edge, we consider the points given by the intersection between the
edge itself and the triangular mesh elements. We refer to these points as to boundary
points. Whenever a refinement of a boundary edge is needed, we split the edge at the
closest boundary point to the edge mid-point, see Fig. 1(right). In Fig. 1(bottom-left)
we sketch three successive refinement of the horizontal portion of  . In Fig. 1(top-
left), we report the meshes resulting from the first four steps of a uniform refinement
of the boundary elements. In case no boundary nodes are available, we perform
successive refinements of the triangle containing the edge until one boundary nodes
becomes available. However the two children produced by refinement of the triangle
are labeled as non-active unless needed by the bulk mesh refinement procedure.
In our implementation the elements of the mesh S are always determined by
the intersection between the boundary  and a triangle (not necessarily active)
belonging to the infinite binary tree with root T0 . In view of this construction, the
levels of refinement to which the elements in S and in T belong to are completely
independent. Clearly, the above construction of S adds some geometric restrictions
on the set of meshes we can deal with. However, the advantage at the computational
level is quite remarkable because all the computer operations required by the
refining process, as well as all the numerical computation of integrals involving the
interaction between bulk and boundary objects (mesh elements or functions) can be
always performed at the level of the single element (or of the binary tree stemming
from it) without affecting the binary data structure of the neighboring elements.
AFDM for Elliptic Problems 241

For the sake of presentation in the rest of this section we employ the following
notation. We set euj D u  Uj and ej D   j . Moreover, we denote by k  k0 the
L2 .˝/ or L2 . /-norm depending on the context and by j  j1 the H 1 .˝/-seminorm.

4.2 Test Case: L-Shape Domain

We consider the L-shaped domain ! D .1 1/2 n .1 0/2 with boundary  . We
are interested in the following model problem: we choose f 2 L2 .!/ such that, after
introducing polar coordinates .r; /, the solution to (1)–(2) is

u.r; / D h.r/r2=3 sin.2=3. C =2//

where

w.3=4  r/ r2 if r > 0
h.r/ D ; w.r/ D
w.r  1=4/ C w.3=4  r/ 0 else.

The fictitious domain formulation of problem (3)–(4) is obtained by embedding !

in the square domain ˝ D .1 1/2 with boundary D @˝, see Fig. 2, and extend
5
f by zero outside !. It is not difficult to see that u 2 H 3  .!/, for any  > 0 and
the exact Lagrange multiplier is

2
D h.r/ r1=3 :
3
In the following, we explore the convergence and optimality properties of
AFDM algorithm. The AFDM algorithm is applied with the following parameters:
˛ D 0:5,  D 0:95 and different values of "0 , namely "0 D 1:0; 0:5; 0:25; 0:1
(see Sect. 3 for the precise meaning of the parameters). In the sequel, we report the
results obtained by AFDM. The outer iteration of AFDM is stopped when

Tj C Sj1 <  45 "0 :

In Fig. 2 we display the initial and the final mesh together with two intermediate
adaptively refined meshes, while in Fig. 3 (left) we report the final discrete solution
U plotted on the fictitious domain and in Fig. 3 (right) we collect the graphs of
final discrete Lagrange multiplier  together with the exact multiplier  and the
restriction of U on the boundary  . A close inspection of the figures reveals that
the algorithm AFDM correctly approximates the exact pair .u; /. As the value of ˇ
is unknown, an a priori choice for  is impossible. However, as already observed
in [4], a practical choice for  seems to be  D 0:95. Indeed, larger values of 
clearly guarantee the boundedness of the inner iterations of ELLIPTIC, but at the
expense of an increased number of outer iterations. Values of  close to 0:95 ensure
242 S. Berrone et al.

(a) 1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0

−0.2 −0.2

−0.4 −0.4

−0.6 −0.6

−0.8 −0.8

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1

(b) 1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0

−0.2 −0.2

−0.4 −0.4

−0.6 −0.6

−0.8 −0.8

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1

Fig. 2 Adaptive meshes produced by AFDM. (a) Initial mesh T0 (left) and mesh at iteration 12
(right), (b) mesh at iteration 18 (left) and final mesh (right)

0.5
1

−1
0
−2

−3

−4 −1
−0.5 −5 −0.8
1 1 −1 −0.6
0.5 0.5 −0.8
−0.6 −0.4
0 0
−0.4 −0.2
−0.5 −0.5 −0.2
−1 −1 0 0
Az: -24 EI: 66

Fig. 3 Left: discrete solution U in the fictitious domain. Right: zoom on  at the reentrant corner
of the discrete solutions U (red),  (blue) and exact multiplier  (green)
AFDM for Elliptic Problems 243

Table 1 Computed rates of convergence of the true errors and of the error estimators Tj and
Sj1 with respect to the total number of dofs #Tj C #Sj1 for different values of "0
keuj k0 jeuj j1 kej1 k 1 ;j1 Tj Sj1
2

"0 D1 0:9073 0:5257 0:4279 0:5201 0:6892

"0 D 0:5 0:9347 0:5447 0:4232 0:5080 0:7072
"0 D 0:25 0:9019 0:5368 0:4104 0:5062 0:7491
"0 D 0:1 0:8819 0:5584 0:3924 0:4980 0:7966

Table 2 Growing rate r of #Tj C #Sj1 #Tj #Sj1

the total number of dofs
(second column), bulk dofs "0 D1 1:8951 1:9415 0:9895
(third) and boundary dofs "0 D 0:5 1:9326 1:9595 0:9928
(fourth) with respect to "j , i.e. "0 D 0:25 2:0156 2:0319 1:0240
r such that #./j 2 O ."rj / "0 D 0:1 2:2928 2:3096 1:0831

an optimal decay rate of the H 1 -error. This motivates the initial choice of setting
 D 0:95 for running the numerical tests.
In the following, we further explore the optimality properties of AFDM and see
their dependency on the parameter "0 . In particular, in Table 1 we report the rates
of convergence with respect to the total number of degrees of freedom, of the
1
true errors for euj in L2 .˝/, H 1 .˝/, the error for ej1 in an approximate H  2 . /
computed as a weighted L2 norm
 X 1=2
kej1 k 1 kej1 k 1 ;j1 WD h` k  j1 k2L2 .`/ ;
2 2
`2Sj1

and the error estimators Tj and Sj1 .

From the third and fourth columns we infer that the total error kr.uUj /kL2 .˝/ C
kj1 k  21 decreases to zero roughly as .#Tj C#Sj1 /0:5 , the optimal decay
H . /
for piecewise approximations of u. This is corroborated by the results in columns
five and six which are thus in agreement with Proposition 3.2.
In Table 2 we collect the growing rates of the total number of dofs #Tj C #Sj1 ,
of the bulk dofs #Tj and of the boundary dofs #Sj1 . The results shows that optimal
convergence is dictated by the regularity of the original non-extended solution u.
Finally, collecting the results of Tables 1–2, we realize that both the total error and
the error indicator Tj C Sj1 decay as "j .

5 Conclusions

We introduced an Adaptive Fictitious Domain Method (AFDM). The core of

the method is based on two modules, ELLIPTIC and ENRICH that iteratively
modify the discrete spaces for the approximation of the extended solution and the
244 S. Berrone et al.

multiplier, respectively. Numerical results show that AFDM is convergent, regardless

of the imposition of any compatibility condition between the two discrete space.
Moreover, preliminary tests seem to suggest the optimal behaviour of AFDM. This
last topic requires further investigations, which is the topic of [5].

Acknowledgements This work has been partially supported by the Italian MIUR through
PRIN research grant 2012HBLYE4_001 “Metodologie innovative nella modellistica differenziale
numerica”, by INdAM-GNCS and by the National Science Foundation grant DMS-1254618.

References

1. Babuška, I.: The finite element method with Lagrangian multipliers. Numer. Math. 20, 179–
192 (1972/1973)
2. Babuška, I.: The finite element method with penalty. Math. Comput. 27, 221–228 (1973)
3. Bacuta, C.: A unified approach for Uzawa algorithms. SIAM J. Numer. Anal. 44(6), 2633–
2649 (2006)
4. Bänsch, E., Morin, P., Nochetto, R.H.: An adaptive Uzawa FEM for the Stokes problem:
convergence without the inf-sup condition. SIAM J. Numer. Anal. 40(4), 1207–1229 (2002)
5. Berrone, S., Bonito, A., Verani, M.: An adaptive fictitious domain method for elliptic problems:
convergence and optimality (2016, in preparation)
6. Bonito, A., DeVore, R.A., Nochetto, R.H.: Adaptive finite element methods for elliptic
problems with discontinuous coefficients. SIAM J. Numer. Anal. 51(6), 3106–3134 (2013)
7. Börgers, C., Widlund, O.B.: On finite element domain imbedding methods. SIAM J. Numer.
Anal. 27(4), 963–978 (1990)
8. Ciarlet, P.G.: The Finite Element Method for Elliptic Problems. Classics in Applied
Mathematics, vol. 40. Society for Industrial and Applied Mathematics (SIAM), Philadelphia
(2002). Reprint of the 1978 original [North-Holland, Amsterdam; MR0520174 (58 #25001)]
9. Cottrell, J.A., Hughes, T.J.R., Bazilevs, Y.: Isogeometric Analysis: Toward Integration of CAD
and FEA. Wiley, Hoboken (2009)
10. Dahlke, S., Dahmen, W., Urban, K.: Adaptive wavelet methods for saddle point problems—
optimal convergence rates. SIAM J. Numer. Anal. 40(4), 1230–1262 (2002)
11. Girault, V., Glowinski, R.: Error analysis of a fictitious domain method applied to a Dirichlet
problem. Japan J. Indust. Appl. Math. 12(3), 487–514 (1995)
12. Glowinski, R., Pan, T.-W., Périaux, J.: A fictitious domain method for Dirichlet problem and
applications. Comput. Methods Appl. Mech. Eng. 111(3–4), 283–303 (1994)
13. Kondratyuk, Y., Stevenson, R.: An optimal adaptive algorithm for the Stokes problem. SIAM
J. Numer. Anal. 46(2), 747–775 (2008)
14. Mommer, M.S.: A smoothness preserving fictitious domain method for elliptic boundary-value
problems. IMA J. Numer. Anal. 26(3), 503–524 (2006)
15. Nochetto, R.H., Veeser, A.: Primer of adaptive finite element methods. In: Multiscale and
Adaptivity: Modeling, Numerics and Applications. Lecture Notes in Mathematics, vol. 2040,
pp. 125–225. Springer, Heidelberg (2012)
16. Peskin, C.S.: The immersed boundary method. Acta Numer. 11, 479–517 (2002)
17. Sacchi, R., Veeser, A.: Locally efficient and reliable a posteriori error estimators for Dirichlet
problems. Math. Models Methods Appl. Sci. 16(3), 319–346 (2006)
Higher-Order Accurate Integration for Cut
Elements with Chen-Babuška Nodes

Thomas-Peter Fries

Abstract The higher-order accurate numerical integration of geometries that are

implicitly defined by level-set functions is considered. A higher-order background
mesh is employed providing an interpolation of the level-set function by Lagrangian
shape functions. The integration may take place on the zero-level set or in the
domains defined by the sign of the level-set function. This work is a follow-up of
Fries and Omerović (Int J Numer Methods Eng, doi:10.1002/nme.5121). Herein, it
is shown that special distributions of the element nodes, which are optimized for
integration, yield significantly better results than equally-spaced nodes. Different
error norms are proposed which allow to investigate the accuracy of general implicit
geometries in two and three dimensions.

1 Introduction

The integration of implicit geometries is a standard task in many numerical methods

such as fictitious domain methods [3, 8, 18] and extended finite element methods
(XFEM) [2, 9, 19]. In the context of the level-set method [23, 27], the boundaries
of the geometry or the interfaces inside are described by the zero-level set of a
scalar function in the domain. The level-set values are typically given at the nodes
of a background mesh which does not consider for the position of the interfaces
or boundaries. Herein, we assume a higher-order background mesh composed by
Lagrangian elements which imply interpolation functions for the level-set function
inside the elements. The zero-level sets are typically curved and arbitrarily cut the
background elements.
It is well-known that equally spaced nodes in the higher-order Lagrange elements
of the background mesh feature problems for interpolations [24, 25, 29], in particular
they suffer from the Runge phenomenon. In [4, 5], optimal interpolation points are
provided for standard elements in one, two, and three dimensions. Herein, these
special node distributions are investigated and it is confirmed that the resulting

T.-P. Fries ()

Institute of Structural Analysis, Graz University of Technology, Lessingstr. 25/II, 8010 Graz,
Austria
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 245

G. Ventura, E. Benvenuti (eds.), Advances in Discretization Methods,
SEMA SIMAI Springer Series 12, DOI 10.1007/978-3-319-41246-7_12
246 T.-P. Fries

integration points are considerably more accurate than with equally spaced element
nodes as used in [10].
Two standard integration tasks occur in the context of the integration of implicit
geometries: The integration on the zero-level set or on the two sides of the zero-level
set. For example, in fictitious domain methods an integration on the zero-level set is
necessary to account for boundary conditions and an integration on one side of the
zero-level set to integrate the governing equations of a model in weak form. In the
XFEM, an integration on the zero-level set may be needed to account for interface
conditions and on the two sides of the interface to evaluate volume terms in the weak
form.
A unified treatment of these two integration tasks in two and three dimensions
was proposed by the authors of this work in [10] for equally spaced element nodes.
This work is a follow-up which (1) extends the algorithm to Lagrange elements
with special node distributions [4, 5], (2) extends the iterative method to locate
positions on the zero-level set, and (3) proposes new error norms suitable for general
implicit geometries. In [10], results have only been presented based on circles in two
dimensions and spheres in three dimensions, and some of the suggested error norms
are not useful for general level-set functions. Furthermore, this work may also be
seen as a summary of the elaborate description proposed by the authors in [10].
Approaches for the integration of elements with internal boundaries and inter-
faces may be distinguished based on the fact whether they rely on a decomposition
of the cut elements into sub-elements or not. For the first class, the standard
approach is to recursively decompose a cut element into polygonal sub-cells until
the desired accuracy is reached [1, 7, 20]. This typically leads to a very large number
of integration points in the context of higher-order approximations [7, 13, 15, 28, 33].
It was already noted in [6, 10, 14, 26] that a decomposition into sub-elements
with curved, higher-order edges or faces is a strategy which consistently enables
to obtain higher-order accurate integration rules with only a modest number of
integration points. Of course, the effort for generating these integration points
is larger than for the polygonal approaches with reduced accuracy. The other
class of approaches is built by methods that do not decompose the cut elements,
the interested reader is referred to [21, 22, 31, 32]. Compared to the element
decomposition techniques, it seems that these approaches are often less intuitive and
general. Their extendability to three dimensions, higher-order accuracy, and general
integrands is not guaranteed.
The paper is organized as follows: The two main sections describe the integration
in two and three dimensions, respectively. In each section, the meshing of the zero-
level set by higher-order interface elements inside the cut reference background
elements is described. This is called reconstruction and is sufficient for the
integration on zero-level sets. For the integration on the two sides of the zero-level
set, sub-elements are defined on the two sides which feature a higher-order side
coinciding with the reconstructed interface element. The blending function method
is used for this purpose [11, 12, 30]. The paper concludes in Sect. 4.
Integration for Cut Elements with Chen-Babuška Nodes 247

2 Integration in Two Dimensions

Let us assume a two-dimensional domain ˝ 2 R2 which is discretized by Lagrange

elements of order m˝ . These elements build the background mesh and their edges
do not align with potential interfaces and boundaries. The element nodes are
either equally spaced in the corresponding reference elements or they feature a
special distribution optimized for interpolation. In [4], such nodes are given by
Chen and Babuška in one-dimensional elements and triangles, see Figs. 1 and 2a,
respectively. For quadrilaterals, the points are easily determined by building tensor-
products of the one-dimensional distributions, see Fig. 2b. We label elements with
the special nodes “Chen-Babuška elements” in this work, referring to the authors of
[4, 5]. The shape functions of the Lagrange elements with equally spaced nodes or
Chen-Babuška nodes share some important properties: (1) they feature Kronecker-
ı property, (2) they are C0 -continuous across element boundaries, (3) they are
negative in some parts of the element (yet bounded for Chen-Babuška elements
but unbounded for equally spaced nodes). It is important to note that the Runge
phenomenon is a well-known problem of Lagrange elements with equally spaced
nodes but is largely reduced for Chen-Babuška elements.
There exists a continuous scalar function  .x/ for all x 2 ˝, called the level-set
function. The zero-level set is defined as

0 D fx 2 ˝ W  .x/ D 0g ; (1)

which is a (curved) one-dimensional line in two dimensions. This line may be

interpreted as an interface which cuts the domain into two non-overlapping regions
˝ C and ˝  based on the sign of  .x/; this is frequently the case in applications
of the XFEM. It may also define the boundary of a domain, then ˝ C may be

(a) (b)
linear linear

quadratic quadratic

cubic cubic

4th order 4th order

5th order 5th order

6th order 6th order

7th order 7th order

8th order 8th order

−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1

Fig. 1 Equally spaced nodes and Chen-Babuška nodes in one-dimensional intervals. (a) Equally
spaced nodes. (b) Chen-Babuška nodes
248 T.-P. Fries

(a) (b) (c)

(d) (e) (f)

Fig. 2 Triangular and quadrilateral Chen-Babuška elements of different order. (a) Cubic, (b) 4th
order, (c) 6th order, (d) cubic, (e) 4th order, (f) 6th order

the domain and ˝  the outer region which is neglected in a simulation; this is
frequently the case in applications of fictitious domain methods.
The level-set function  .x/ is replaced by its discrete counterpart
X
 h .x/ D Ni˝ .x/  i (2)
i2I

where I is the set of all nodes in ˝, Ni˝ .x/ are the higher-order shape functions,
and i D  .xi / are the nodal level-set values.
It is obvious that the zero-level set may cut arbitrarily through the elements of
the background mesh. It is then a standard task, e.g. in applications of the XFEM
and fictitious domain methods, to properly integrate on the two sides of the zero
level-set or right on the zero-level set. The focus is now on one element cut by the
zero-level set.
The procedure proposed in [10] is to first approximate the zero-level set of  h .x/
by means of a higher-order, typically curved, one-dimensional interface element
in the reference background element. This is called “reconstruction” and involves
finding roots of the higher-order level-set function through an iterative procedure. It
is then simple to integrate on the zero-level set in the domain ˝ by first mapping
the nodes of the curved interface element to the physical background element and
then mapping one-dimensional Gauss points to this interface element, see Fig. 3.
Integration for Cut Elements with Chen-Babuška Nodes 249

s
physical
)
x(r
background
element
y

x
r a
reference reference
background element interface element

Fig. 3 Sketch of the integration on the zero-level set. The element nodes of the curved 1D interface
element are located in the reference background element and mapped to the physical element.
Then, standard Gauss points in 1D (red crosses) are mapped to this interface element

b
s
a )
x(r
y
sub-elements
reference

b
physical
background
x element

a reference r
background element
(with mapped interface and sub-elements)

Fig. 4 This sketch demonstrates the integration in the cut elements. The reconstructed curved
interface element cuts the reference background element into special sub-elements with one higher-
order side. There, standard Gauss points (red crosses) are located and mapped to the reference
background element and further to the physical element

For the integration on the two sides of the zero-level set, the background refer-
ence element is decomposed into sub-elements. Therefore, also the reconstructed
interface element is needed and special sub-elements are defined which feature
one higher-order side coinciding with the interface element. The blending function
mapping is employed for this purpose [11, 12, 30]. Standard Gauss points in two
dimensions are mapped to the special sub-elements in the reference background
element and further to the physical background element, see Fig. 4.
It is obvious that higher-order elements may feature very complicated topologies
of zero-level sets. For example, element edges may be cut several times, or the zero-
250 T.-P. Fries

(a) (b) (c)

Fig. 5 Some complicated zero-level sets in quadrilateral elements corresponding to invalid level-
set data. Note that after recursive refinements, valid level-set data is obtained in the refined
elements. (a) Some edges are cut more than once. (b) More than two edges are cut. (c) The element
is cut although none of the edges are cut

(a) (b) (c)

1 3
1 1 T 1 4
2
3 2
3 Q1
2 3
T T3 2
2
2 3
1 3
3 2 2
T T1
3 T4 3 Q2
T1
T 2 2
4
1 2 3 1 1 2 3 1 4 1

Fig. 6 Topologically different cut situations for valid level-set data in triangles and quadrilaterals.
(a) Triangle, (b) quadrilateral, top. 1, (c) quadrilateral, top. 2

level set is completely inside an element, see Fig. 5. Therefore, a recursive procedure
may be required (typically only in very few elements) until a “valid topology” inside
the element is obtained. By “valid” we refer to the situation where (1) each element
edge is only cut once, (2) the overall number of cut edges must be two, and (3) if
no edge is cut then the element is completely uncut. Whether these conditions are
fulfilled is checked based on a sample grid in the reference element. If they are not
matched, the reference element is recursively refined and the described procedures
apply to the refined elements. For an overview, see Fig. 5 and for details, see [10].
Provided that valid level-set data is obtained, the zero-level set cuts the reference
element in a limited number of topologically different cases, see Fig. 6. Only the
sign at the corner nodes of the reference element determines the situation. A cut
triangle is decomposed into one sub-triangle and one sub-quadrilateral; the latter
is further decomposed into three sub-triangles. For cut quadrilaterals, two different
topologies are possible. Topology 1 is obtained when two neighboring edges are
cut (resulting into four sub-triangles). Topology 2 is obtained when two opposite
edges are cut leading to two sub-quadrilaterals. Note that the sub-elements feature
one curved side which coincides with the reconstructed interface element.
Integration for Cut Elements with Chen-Babuška Nodes 251

2.1 Reconstruction of Zero-Isolines in 2D Background

Elements

In order to approximate the zero-level set in the reference background element

by (curved) interface elements of the desired order m , element nodes have to be
located on the zero-level set. Therefore, an iterative procedure is required for each
node. Many alternatives are possible, being different e.g. in the starting values and
search paths. In [10], a Newton-Raphson based method was suggested using straight
search paths. It was demonstrated that the issue of starting points and corresponding
search paths has important consequences on the resulting accuracy of the integration
schemes. Most importantly, starting values in two-dimensional elements should at
least be provided by a linear reconstruction of the zero-level set. This is the straight
line between the intersection points of the zero-level set and the element edges, see
Fig. 7. For inner nodes of higher-order interface elements, good starting guesses are
found by mapping the higher-order element nodes to the linear interface element.
This works analogously for equally spaced nodes and Chen-Babuška nodes. In
Fig. 7, the starting points for a Chen-Babuška element of order 5 are shown for
different scenarios in triangles and quadrilaterals.
It is now necessary to address the question how to move towards suitable
positions on the zero-level set from the starting points on the linear reconstruction.
Two approaches relying on straight search paths are proposed in [10]: approach
1 uses the normal direction to the linear interface element (Fig. 8a), approach 2
uses the gradient of the level-set function  .x/ at the starting point (i.e. the path
perpendicular to each contour line (Fig. 8b). Approach 3 is new and employs a
search path which is not straight for the whole iterative procedure. It rather uses the
gradient of  .x/ at each intermediate position during the iteration. All approaches
yield quadratic convergence rates of the iterative procedure but, of course, to
(slightly) different positions of the element nodes on the zero-level set.

+ + + + +
(a) (b) (c)

+ − + − − −

Fig. 7 Starting values for the root finding. The thick black line is the exact zero-level set of  h .x/.
The intersections with the element edges provide a straight line (linear reconstruction). Interface
elements of the desired order m (with equally spaced nodes or Chen-Babuška nodes) are mapped
to this linear element. (a) Triangle, (b) quadrilateral, top. 1, (c) quadrilateral, top. 2
252 T.-P. Fries

(a) (b) (c)

Fig. 8 Different approaches for search paths at the starting points on the linear reconstruction. (a)
Approach 1, (b) approach 2, (c) approach 3

Mathematically, the algorithm is described as follows. The task is to find the

root of  h .r/ in the reference element, i.e. some position on the zero-level set.
The starting point of the iterative procedure is labeled r0 . The algorithm for all
approaches considered here is based on the following iteration:
 
 h ri
r iC1
Dr 
i
 N: (3)
r .ri /  N

In approach 1, N is the normal vector of the linear reconstruction which is equal

 0  In approach 2, N is the gradient of  at each starting
for all starting points. point,
hence
  N D r r . Finally, approach 3 uses the current gradient of  at r i
, i.e. ND
r ri . When all nodes of the higher-order interface element are determined in
the reference background element, it is a simple task to map them to the physical
background elements using the isoparametric concept.
It is now clear how to reconstruct higher-order interface elements in the reference
and physical background elements which approximate the zero-level set with the
expected accuracy. The integration on the zero-level set is straightforward when
mapping standard Gauss points in one-dimensional intervals to the curved interface
elements, see Fig. 3.

2.2 Integration in Cut 2D Background Elements

For an integration in a cut element, the element is decomposed into special sub-
elements which feature one higher-order side that coincides with the reconstructed
interface element, see Fig. 4. A mapping of standard Gauss points in triangles and
quadrilaterals to these special sub-elements has to be defined. The blending function
mapping is used for this purpose; we refer to [10] for details and a unified assessment
for the situation in different element types. The situation of Chen-Babuška elements
is extended straightforwardly to the description therein. Also the issue of “consistent
Integration for Cut Elements with Chen-Babuška Nodes 253

integration rules” in triangles applies just as well: It is crucial to use tensor-product

Gauss rules of quadrilaterals also in the triangular sub-elements (collapsing at the
node opposite to the higher-order side) in order to avoid a negative impact on
the accuracy due to the properties of the blending function mapping in triangles.
Namely, the Jacobian is not smooth at the node opposite to the higher-order side.

2.3 Numerical Results in Two Dimensions

Convergence studies are presented for the integration in two-dimensional domains:

Firstly, on the zero-level set of  .x/ and, secondly, in the region where  .x/ < 0,
i.e. ˝  . The domain is discretized by higher-order background meshes with equally
spaced nodes or Chen-Babuška nodes in each element, respectively. It is noted
that also the reconstructed interface elements may feature (independently of the
background elements) equally spaced nodes or Chen-Babuška nodes. Therefore,
four different setups are distinguished in Table 1, see also Fig. 9.
It was shown in [10] that the accuracy of the integration is directly related
to the order of the background and interface elements, so that it is natural to
choose identical orders for both, i.e. m˝ D m . Following [10], triangular and
quadrilateral elements are investigated here, each in a Cartesian arrangement where
the element edges align with the physical coordinate system (“undeformed mesh”)
and in a deformed mesh where the elements are perturbed, see Fig. 10. Two different

Table 1 Combinations of equally spaced nodes and Chen-Babuška nodes in background and
interface elements
Background elements Interface elements
Case 1: ES-ES Equally spaced nodes Equally spaced nodes
Case 2: ES-CB Equally spaced nodes Chen-Babuška nodes
Case 3: CB-ES Chen-Babuška nodes Equally spaced nodes
Case 4: CB-CB Chen-Babuška nodes Chen-Babuška nodes

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

Fig. 9 Equally spaced (ES) and Chen-Babuška (CB) nodes in background elements and recon-
structed interface elements (of order 5). (a) Case 1: ES-ES. (b) Case 2: ES-CB. (c) Case 3: CB-ES.
(d) Case 4: CB-CB
254 T.-P. Fries

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

Fig. 10 Deformed and undeformed background meshes composed by higher-order triangular and
quadrilateral elements. The red and blue lines are the exact zero-level sets of 1 .x/ and 2 .x/,
respectively. (a) Undef., quad., (b) def., quad., (c) undef., tri., (d) def., tri.

level-set functions are considered:

p
1 .x/ D x2 C y2  r (4)
p
2 .x/ D x2 C y2  R ./ (5)

with r D 0:7123, R ./ D 0:5 C 0:1  sin .8/ and  .x/ D atan .y=x/. See Fig. 10
for a visualization of the zero-level sets of 1 and 2 . The zero-level set of 1 is a
circle with radius r and is frequently used in the literature, e.g. [10], and 2 is similar
to [16].
In the convergence studies, f6; 10; 20; 50; 100; 200g elements are used per dimen-
sion. We use Gauss rules in 1D (for the integration on the zero-level set 0 ) and
2D (for the integration in ˝  ) with a rather high order of 11 which is kept constant
independently of the order of background and interface elements. Results are studied
in different error norms which are all related to integrating a function f .x/ either on
0 or in ˝  .

2.3.1 Integration on Zero-Isolines

Four different (relative) error norms are introduced for the integration on zero-level
sets in two dimensions.
ˇ X ˇ ˇ ˇ Z
ˇ ˇ ˇ ˇ
"1 D ˇ wi  I1 ˇ= I1 with I1 D 1 d ; (6)
i 0
ˇ X ˇ ˇ ˇ Z
ˇ ˇ
" f D ˇ wi  f .xi /  If ˇ= ˇIf ˇ with If D f .x/ d ; (7)
i 0
ˇ X ˇ ˇ ˇ
ˇ ˇ
" f h D ˇ wi  f1D
h
.xi /  If ˇ= ˇIf ˇ ; (8)
1D
i
ˇ X ˇ ˇ ˇ
ˇ ˇ
" f h D ˇ wi  f2D
h
.xi /  If ˇ= ˇIf ˇ ; (9)
2D
i
Integration for Cut Elements with Chen-Babuška Nodes 255

where xi are the integration points in the interface elements with integration weights
wi . The function f .x/ is defined as

f .x/ D 1=2  x C 1=4  y C x2 C 2  y3 : (10)

For f1Dh
.x/, f .x/ is prescribed at the nodes of the higher-order interface elements
and interpolated inside the interface elements. Analogously, f2D h
.x/ is the higher-
order interpolation based on the background elements. For the error norm " f , the
function (10) is evaluated exactly at the integration points xi . For the error norms
" f h and " f h , the function is interpolated at xi by the higher-order shape functions in
1D 2D
the interface or background elements, respectively. It is noted that, in contrast to the
nodes of the interface elements, the integration points at xi are only approximately
on the zero-level set of  h .x/.
Hence,  h .x/ and f h .x/ only approximate  .x/ and f .x/, respectively. Fur-
thermore, the interface elements only approximate the zero-level set of  h .x/.
The numerical integration based on Gauss points only approximates the analytical
integration. Nevertheless, although several approximations are involved, it is shown
below that integration rules with optimal accuracy are recovered for integrands over
f .x/ in domains defined by the level-set function  .x/.

Study 1: Variation of Search Paths

The deformed quadrilateral background mesh is used with Chen-Babuška nodes in

the background and interface elements (case 1). The influence of the search paths
as described in Sect. 2.1 is investigated. Results for " 1 are shown in Fig. 11 for
2 .x/. Alternatives 1 and 2 were straight search paths based on the normal vector
of the linear reconstruction or the gradient of the level-set function, respectively.
Alternative 3 used the gradient at the intermediate positions throughout the iter-
ative procedure to locate roots. It is seen that alternative 1 is slightly superior
than 2 and 3. Because it is also particularly simple, it is used as a default in the
remainder of this work.

0
(a) 0
(b) 0
(c)
10 10 10
−2 −2 −2
10 10 10
−4 −4 −4
10 10 10
−6 −6 −6
10 10 10
−8 −8 −8
10 10 10
order = 1 order = 1 order = 1
−10 order = 2 −10 order = 2 −10 order = 2
10 10 10
order = 3 order = 3 order = 3
−12 order = 4 −12 order = 4 −12 order = 4
10 order = 5 10 order = 5 10 order = 5
−14 order = 6 −14 order = 6 −14 order = 6
10 10 10
−2 −1 −2 −1 −2 −1
10 10 10 10 10 10

Fig. 11 Results for different search paths. (a) Alternative 1. (b) Alternative 2. (c) Alternative 3
256 T.-P. Fries

Study 2: Variation of Error Norms

The different error norms are studied next, keeping the parameters of the previous
study unchanged. The search directions are according to alternative 1. Results are
seen in Fig. 12. It is noted that functions are integrated with optimal accuracy
on the zero-level set, no matter whether exact function values are prescribed or
interpolations in background or interface elements are used at the integration points.

Study 3: Variations of Elements

Above, results were obtained with deformed quadrilateral background meshes. The
aim is now to show that equivalent results are obtained also for triangular elements
and independently of deformed or undeformed background meshes, see Fig. 10.
Results for " 1 are shown in Fig. 13 and confirm this assumption.

0
(a) 0
(b)
10 10
−2 −2
10 10
−4 −4
10 10
−6 −6
10 10
−8 −8
10 10
order = 1 order = 1
−10 order = 2 −10 order = 2
10 10
order = 3 order = 3
−12 order = 4 −12 order = 4
10 order = 5 10 order = 5
−14 order = 6 −14 order = 6
10 10
−2 −1 −2 −1
10 10 10 10

0
(c) 0
(d)
10 10
−2 −2
10 10
−4 −4
10 10
−6 −6
10 10
−8 −8
10 10
order = 1 order = 1
−10 order = 2 −10 order = 2
10 10
order = 3 order = 3
−12 order = 4 −12 order = 4
10 order = 5 10 order = 5
−14 order = 6 −14 order = 6
10 10
−2 −1 −2 −1
10 10 10 10

Fig. 12 Convergence in different error norms. (a) " 1 , (b) " f , (c) " f h , (d) " f h
1D 2D
Integration for Cut Elements with Chen-Babuška Nodes 257

0
(a) 0
(b)
10 10
−2 −2
10 10
−4 −4
10 10
−6 −6
10 10
−8 −8
10 10
order = 1 order = 1
−10 order = 2 −10 order = 2
10 10
order = 3 order = 3
−12 order = 4 −12 order = 4
10 order = 5 10 order = 5
−14 order = 6 −14 order = 6
10 10
−2 −1 −2 −1
10 10 10 10

0
(c) 0
(d)
10 10
−2 −2
10 10
−4 −4
10 10
−6 −6
10 10
−8 −8
10 10
order = 1 order = 1
−10 order = 2 −10 order = 2
10 10
order = 3 order = 3
−12 order = 4 −12 order = 4
10 order = 5 10 order = 5
−14 order = 6 −14 order = 6
10 10
−2 −1 −2 −1
10 10 10 10

Fig. 13 Convergence for deformed/undeformed background meshes with triangular/quadrilateral

elements. (a) Quad., undef., (b) quad., def., (c) tri., undef. (d) tri., def.

Study 4: Variation of Equally Spaced Nodes and Chen-Babuška Nodes

Again, we consider the deformed quadrilateral background mesh. All parameters are
as in study 1 where the search paths are according to alternative 1. Equally spaced
and Chen-Babuška nodes in the background and interface elements are systemati-
cally compared. The four possible combinations have already been described above.
Results for " 1 are shown in Fig. 14. It is seen that the use of Chen-Babuška nodes
yields significantly better results: the convergence rates are optimal in all cases yet
the magnitude of the error varies with a factor of up to 10. For the best results,
Chen-Babuška nodes are used in the background and interface elements.
258 T.-P. Fries

0
(a) 0
(b)
10 10
−2 −2
10 10
−4 −4
10 10
−6 −6
10 10
−8 −8
10 10
order = 1 order = 1
−10 order = 2 −10 order = 2
10 10
order = 3 order = 3
−12 order = 4 −12 order = 4
10 order = 5 10 order = 5
−14 order = 6 −14 order = 6
10 10
−2 −1 −2 −1
10 10 10 10

0
(c) 0
(d)
10 10
−2 −2
10 10
−4 −4
10 10
−6 −6
10 10
−8 −8
10 10
order = 1 order = 1
−10 order = 2 −10 order = 2
10 10
order = 3 order = 3
−12 order = 4 −12 order = 4
10 order = 5 10 order = 5
−14 order = 6 −14 order = 6
10 10
−2 −1 −2 −1
10 10 10 10

Fig. 14 Convergence in " 1 for equally spaced nodes or Chen-Babuška nodes in the background
and interface elements, respectively. (a) Case 1: ES-ES. (b) Case 2: ES-CB. (c) Case 3: CB-ES.
(d) Case 4: CB-CB

2.3.2 Integration in Cut 2D Elements

Three different (relative) error norms are introduced for the integration in ˝  :
ˇ X ˇ ˇ ˇ Z
ˇ ˝ˇ ˇ ˝ˇ
"˝
1 D ˇ wi  I 1 ˇ= I1 with I1˝ D 1 d˝; (11)
i ˝
ˇ X ˇ ˇ ˇ Z
ˇ ˇ
"˝
f Dˇ wi  f .xi /  If˝ ˇ= ˇIf˝ ˇ with If˝ D f .x/ d˝; (12)
i ˝
ˇ X ˇ ˇ ˇ
ˇ ˇ
"˝
fh D ˇ wi  f h .xi /  If˝ ˇ= ˇIf˝ ˇ ; (13)
i

where xi are 2D integration points in the special sub-elements and wi the corre-
sponding weights. The standard setting from above is used: a deformed quadrilateral
background mesh, search paths according to alternative 1, Chen-Babuška nodes in
the background and interface elements, and the level-set function 2 . Results in
Integration for Cut Elements with Chen-Babuška Nodes 259

0
(a) 0
(b) 0
(c)
10 10 10
−2 −2 −2
10 10 10
−4 −4 −4
10 10 10
−6 −6 −6
10 10 10
−8 −8 −8
10 10 10
order = 1 order = 1 order = 1
−10 order = 2 −10 order = 2 −10 order = 2
10 10 10
order = 3 order = 3 order = 3
−12 order = 4 −12 order = 4 −12 order = 4
10 order = 5 10 order = 5 10 order = 5
−14 order = 6 −14 order = 6 −14 order = 6
10 10 10
−2 −1 −2 −1 −2 −1
10 10 10 10 10 10

Fig. 15 Convergence in different error norms. (a) "˝ ˝ ˝

1 , (b) "f , (c) "f h

0
(a) 0
(b)
10 10
−2 −2
10 10
−4 −4
10 10
−6 −6
10 10
−8 −8
10 10
order = 1 order = 1
−10 order = 2 −10 order = 2
10 10
order = 3 order = 3
−12 order = 4 −12 order = 4
10 order = 5 10 order = 5
−14 order = 6 −14 order = 6
10 10
−2 −1 −2 −1
10 10 10 10

0
(c) 0
(d)
10 10
−2 −2
10 10
−4 −4
10 10
−6 −6
10 10
−8 −8
10 10
order = 1 order = 1
−10 order = 2 −10 order = 2
10 10
order = 3 order = 3
−12 order = 4 −12 order = 4
10 order = 5 10 order = 5
−14 order = 6 −14 order = 6
10 10
−2 −1 −2 −1
10 10 10 10

Fig. 16 Convergence in "˝ fh

for equally spaced nodes or Chen-Babuška nodes in the background
and interface elements, respectively. (a) Case 1: ES-ES. (b) Case 2: ES-CB. (c) Case 3: CB-ES.
(d) Case 4: CB-CB

the three different error norms are shown in Fig. 15. Optimal convergence rates
are achieved in all cases. Next, equally spaced nodes and Chen-Babuška nodes
are varied in the background and interface elements. Results are given in Fig. 16
and confirm the findings from above: Chen-Babuška nodes yield significantly better
results than equally spaced nodes.
260 T.-P. Fries

We obtained very similar results for different background elements in deformed

and undeformed configurations. Also, 1 yields equivalent results so that conver-
gence curves are not shown here for brevity. It is noted that for 1 , I1˝ is the exact
area of a circle and I1 the circumference.

3 Integration in Three Dimensions

The two-dimensional situation is now extended to the three-dimensional case.

Hence, a three-dimensional higher-order background mesh built by Lagrange
elements with equally-spaced nodes or Chen-Babuška nodes is considered. The
element faces do not align with boundaries or interfaces, which are surfaces in 3D.
In [5], Chen-Babuška nodes are given in tetrahedra, see Fig. 17a–c. For hexahedra
as shown in Fig. 17d–f, the nodes are obtained by tensor-products from one-
dimensional distributions as given in [4]. The shape functions associated to these
elements feature the same properties as mentioned in Sect. 2.
The zero-level set 0 of the level-set function  .x/, cf. Eq. (1), is a (curved)
surface in three dimensions, typically defining boundaries or interfaces. As above,

(a) (b) (c)

(d) (e) (f)

Fig. 17 Tetrahedral and hexahedral Chen-Babuška elements of different order. (a) Cubic, (b) 4th
order, (c) 6th order, (d) cubic, (e) 4th order, (f) 6th order
Integration for Cut Elements with Chen-Babuška Nodes 261

(a) (b)

Fig. 18 Topologically different cut situations for valid level-set data in tetrahedra. (a) Topology
1, (b) topology 2

the level-set function is replaced by its interpolant  h .x/, cf. Eq. (2). Again, the aim
is to integrate on 0 or inside the domains ˝ C and/or ˝  .
Starting point is the reconstruction of the zero-isosurface, i.e. the meshing by
higher-order interface elements. This is first described for tetrahedra. We assume
that valid level-set data is present, at least after some recursive refinements [10].
That is, the topological cut situation is defined based on the sign at the corner
nodes of the tetrahedra. Then, only two different topologies are possible. Topology
1 is obtained when three edges are cut, the iso-surface is then a curved triangular
element in the reference tetrahedron, and the two volumetric sub-elements are a sub-
tetrahedron and a sub-prism. For topology 2, four elements edges are cut yielding
a curved quadrilateral interface element and cutting the tetrahedron into two sub-
prisms. The situation is shown in Fig. 18.
It is noted that hexahedral elements lead to a much larger number of topological
cases even when only the signs at the corner nodes are considered. In a low-
order context, these topologies are addressed by the marching cube algorithm [17],
however, only for the detection and definition of the zero-level set. This algorithm
does not address the decomposition into volumetric sub-elements for an integration
on the two sides of the zero-level set. Therefore, it is advisable to first decompose
a hexahedral element into tetrahedra and then proceed in each tetrahedron as
mentioned below.
262 T.-P. Fries

3.1 Reconstruction of Zero-Isosurfaces in 3D Background

Elements

For the approximation of the zero-level set of  h .x/ by higher-order interface ele-
ments, the corresponding element nodes have to be located on the zero-isosurface.
The algorithm of the iterative procedure of Sect. 2.1, see Eq. 3, is straightforwardly
extended to the three-dimensional case. The three alternatives for the search paths
apply also in 3D.
An important difference, however, is the definition of starting points in the
tetrahedra. It turns out that linear reconstructions, see Fig. 19a and d, do not
provide good starting points for the iteration. The iteration often fails already for
rather simple cases (modest curvature of the zero-level set in the tetrahedron)
initiating (undesired) recursive refinements. It is much better to use a higher-order
reconstruction based on face reconstructions of the tetrahedron. That is, for topology
1, a quadratic triangle element is generated by applying the 2D-reconstructions of
Sect. 2.1 on the element faces, see Fig. 19b. For topology 3, a cubic Serendipity
element is computed by face reconstructions, see Fig. 19e. Once these intermediate
higher-order reconstructions are obtained, it is simple to map two-dimensional
triangular or quadrilateral elements (with equally spaced nodes of Chen-Babuška
nodes) onto them, see Fig. 19c and f. They provide excellent starting points for the
root search. For the search paths, one may adapt the three alternatives for two-
dimensional reconstructions, see Sect. 2.1. For alternative 1, it is noted that the
normal vector at the starting points is no longer constant due to the curvature of
the intermediate reconstruction. All alternatives converge quadratically towards a
position on the zero-level set defining a node of the interface element.
For the integration on 0 , it is simple to map the obtained element nodes in
the reference background element to the physical elements using the iso-parametric
concept. Then, standard Gauss points in triangles and quadrilaterals are mapped to
the physical interface elements. This follows the principle shown in two dimensions
in Fig. 3. Gram’s determinant is used to adjust the integration weights.

3.2 Integration in Cut 3D Background Elements

For the task to integrate in the three-dimensional domains ˝ C and/or ˝  , the

cut elements have to be decomposed into volumetric sub-elements. As seen in
Fig. 18, possible sub-elements are tetrahedra with one higher-order triangular side
and prisms where the higher-order side is either a quadrilateral or a triangle. The
higher-order side always matches with the reconstructed interface element. In [10],
the mapping from the special (reference) sub-elements (where 3D Gauss points are
used) to the cut reference background elements is defined based on the blending
function mapping. These integration points are then mapped further to the physical
background elements based on the isoparametric concept. The issue of “consistent
Integration for Cut Elements with Chen-Babuška Nodes 263

(a) (b) (c)

(d) (e) (f)

Fig. 19 Reconstruction in tetrahedra. (a) and (d) show starting points based on a linear recon-
struction which turned out to be less robust. Therefore, for topology 1, a quadratic triangle and for
topology 2, a cubic Serendipity element are defined by 2D-reconstructions on the faces, see (b) and
(e). The desired higher-order element is mapped onto these interface elements in order to provide
(highly accurate) starting points for the root finding, see (c) and (f). (a) Top. 1, linear. (b) Top. 1,
face recon. (c) Top. 1, starting points. (d) Top. 2, linear. (e) Top. 2, face recon. (f) Top. 2, starting
points

integration rules” which consider the special properties of the mapping is pointed
out in [10]; it restricts the selection of possible Gauss rules in the (reference) sub-
elements.
Note that the blending function mapping may lead to negative Jacobians even
if the reconstruction of the interface element, defining the higher-order side of
the sub-elements, was successful. The Jacobians must therefore be checked and
a recursive refinement may be necessary to obtain valid mappings. We find the
following reasons for recursive refinements: (1) to obtain “valid” level-set data, (2)
to provide suitable roots inside the reference background element, and (3) to obtain
non-negative Jacobians for the blending function mapping.
264 T.-P. Fries

3.3 Numerical Results in Three Dimensions

Numerical results are now presented for the integration of three-dimensional

geometries. First for integrations on 0 , then for integrations in ˝  . The procedure
is similar to the studies in two dimensions, see Sect. 2.3. For the search paths
of the nodes of the interface elements, we restrict ourselves to the one which is
based on the normal vector of the intermediate reconstruction. Note that in 3D,
the intermediate reconstruction is already higher-order (quadratic triangle elements
or cubic Serendipity elements) so that the normal vectors are not constant for all
starting points as in 2D. The gradient-based alternatives were leading to worse
results and are, therefore, omitted. We consider the following two level-set functions
in our studies:
p
1 .x/ D x2 C y2 C z2  r (14)
2 .x/ D 1 .x/ C 0:1  Œcos .2  x/ C cos .2  y/ C cos .2  z/ (15)

with r D 0:7123. The zero-isosurfaces of these two level-set functions are visualized
in Fig. 20. There, also the undeformed Cartesian background mesh is shown,
deformed configurations in three dimensions are neglected here. For the integrand,
the function

f .x/ D x2 C y2 C 1=2  cos.z/ (16)

is chosen.
Hexahedral and tetrahedral background meshes are studied. In the convergence
studies, f2; 4; 6; 10; 14; 20; 30; 50; 100g elements are used per dimension. Of course,
especially for the coarse meshes recursive refinements are necessary in a larger
percentage of the elements than in the fine meshes. These refinements are not
considered for the element spacing h in the convergence plots shown herein, leading
to some wiggles in the convergence behavior.

(a) (b)

Fig. 20 Zero-isosurfaces of 1 and 2 in Œ1; 1 3 according to Eqs. (14) and (15), respectively. (a)
Zero-isosurface of 1 , (b) zero-isosurface of 2
Integration for Cut Elements with Chen-Babuška Nodes 265

3.3.1 Integration on Zero-Isosurfaces

The error norms of Eqs. (6)–(9) are straightforwardly extended to the three-
dimensional case, leading to " 1 , " f , " f h , " f h which are integrated on the zero-
2D 3D
isosurface 0 . Results are shown in Fig. 21 for hexahedral background meshes
and 2 . Comparing the upper and the lower figures of Fig. 21, respectively,
shows that Chen-Babuška points in the elements again yield to better results than
equally spaced nodes. Furthermore, comparing the left and right figures in Fig. 21,
respectively, shows that the error converges similar in all error norms. We only show
a representative subset of the results for brevity. Note that, most importantly, the
convergence rates are higher-order yet not optimal in three dimensions. This is in
agreement with [10]. There, this issue was further traced back to the fact that the
boundaries of the interface elements are strictly enforced to be on the element faces
of the background elements.

(a) (b)
100 100

10-2 10-2

10-4 10-4

10-6 order = 1 10-6 order = 1

order = 2 order = 2
order = 3 order = 3
10-8 order = 4 10-8 order = 4
order = 5 order = 5
order = 6 order = 6
10-10 10-10
10-2 10-1 10-2 10-1

(c) (d)
10 0 10 0

10-2 10-2

10-4 10-4

10-6 order = 1 10-6 order = 1

order = 2 order = 2
order = 3 order = 3
10-8 order = 4 10-8 order = 4
order = 5 order = 5
order = 6 order = 6
10-10 10-10
10-2 10-1 10-2 10-1

Fig. 21 Convergence for the integration on 0 in three dimensions. (a) Case 1: ES-ES, " 1 . (b)
Case 1: ES-ES, " f h . (c) Case 4: CB-CB, " 1 . (d) Case 4: CB-CB, " f h
3D 3D
266 T.-P. Fries

3.3.2 Integration in Cut 3D Elements

The error norms of Eqs. (11)–(13) are straightforwardly extended to the three-
dimensional case. However, as seen for the integration on the zero-level sets, the
convergence behavior is independent of these norms so that results are only shown
for "˝
fh
. Furthermore, we restrict ourselves to the case of 1 , where the integration is
performed in a ball with radius r. The advantage is that there are analytical results
available for the integration so that numerical reference solutions are not needed.
For "˝ ˝ 3 ˝ ˝
1 follows the volume of a ball, I1 D =3    r , and for "f and "f h follows
4

If˝ D 2    .sin.r/  r  cos.r// C 4=105    r7 when f is defined by Eq. (16). Results

are seen in Fig. 22 and confirm the findings from above.

(a) (b)
100 100

10-2 10-2

10-4 10-4

10-6 order = 1 10-6 order = 1

order = 2 order = 2
order = 3 order = 3
10-8 order = 4 10-8 order = 4
order = 5 order = 5
order = 6 order = 6
10-10 10-10
10-1 10-1

(c) (d)
100 100

10-2 10-2

10-4 10-4

10-6 order = 1 10-6 order = 1

order = 2 order = 2
order = 3 order = 3
10-8 order = 4 10-8 order = 4
order = 5 order = 5
order = 6 order = 6
10-10 10-10
10-1 10-1

Fig. 22 Convergence in "˝ fh

for the integration in ˝  in three dimensions. (a) Case 1: ES-ES,
hexa. (b) Case 1: ES-ES, tetra. (c) Case 4: CB-CB, hexa. (d) Case 4: CB-CB, tetra
Integration for Cut Elements with Chen-Babuška Nodes 267

4 Conclusions

The use of special element nodes as proposed by Chen and Babuška in [4, 5] in the
context of higher-order accurate integrations of implicit geometries in two and three
spatial dimensions is investigated. Two standard integration tasks are studied, the
integration on zero-level sets and in the sub-domains implied by the sign of the level-
set function. In a first step, the algorithm relies on the reconstruction of the zero-level
set of the interpolated level-set function in terms of higher-order interface elements.
In the second step, sub-elements are defined which feature a curved side which
coincides with the higher-order interface element. The blending function mapping
is used to map standard Gauss points into these special sub-elements.
The underlying algorithm was introduced and described in detail in [10]. Herein,
it is extended to the use of Chen-Babuška elements. It is found that this improves the
error by a factor of typically 5 to 10 for the same number of elements, i.e. the same
computational effort. Curved geometries are studied in two and three dimensions
and convergence studies are made for (1) different element types, (2) Cartesian or
perturbed background meshes, and (3) different search directions. Optimal results
have been obtained in two dimensions independently of the element type and state
of the background meshes. The search directions based on the normal vector of the
linear reconstructions of the zero-level set performs better than those based on the
gradient of the level-set function. In three dimensions, the results are higher-order
accurate yet sub-optimal. In [10], this was at least partly traced back to the fact
that the sub-elements are strictly enforced to be inside the background reference
elements. The errors are checked in different norms. Some error norms also involve
interpolated integrands instead of exact values at the integration points and it is
found that this does not reduce the accuracy.
The proposed integration scheme allows for a consistent integration in higher-
order accurate numerical methods, e.g. in the frame of fictitious domain methods
and extended finite element methods.

References

1. Abedian, A., Parvizian, J., Düster, A., Khademyzadeh, H., Rank, E.: Performance of different
integration schemes in facing discontinuities in the finite cell method. Int. J. Comput. Methods
10, 1–24 (2013)
2. Belytschko, T., Black, T.: Elastic crack growth in finite elements with minimal remeshing. Int.
J. Numer. Methods Eng. 45, 601–620 (1999)
3. Burman, E., Hansbo, P.: Fictitious domain finite element methods using cut elements: I. A
stabilized Lagrange multiplier method. Comput. Methods Appl. Mech. Eng. 199, 2680–2686
(2010)
4. Chen, Q., Babuška, I.: Approximate optimal points for polynomial interpolation of real
functions in an interval and in a triangle. Comput. Methods Appl. Mech. Eng. 128, 405–417
(1995)
268 T.-P. Fries

5. Chen, Q., Babuška, I.: The optimal symmetrical points for polynomial interpolation of real
functions in the tetrahedron. Comput. Methods Appl. Mech. Eng. 137, 89–94 (1996)
6. Cheng, K., Fries, T.: Higher-order XFEM for curved strong and weak discontinuities. Int. J.
Numer. Methods Eng. 82, 564–590 (2010)
7. Dréau, K., Chevaugeon, N., Moës, N.: Studied X-FEM enrichment to handle material interfaces
with higher order finite element. Comput. Methods Appl. Mech. Eng. 199, 1922–1936 (2010)
8. Düster, A., Parvizian, J., Yang, Z., Rank, E.: The finite cell method for three-dimensional
problems of solid mechanics. Comput. Methods Appl. Mech. Eng. 197, 3768–3782 (2008)
9. Fries, T., Belytschko, T.: The extended/generalized finite element method: an overview of the
method and its applications. Int. J. Numer. Methods Eng. 84, 253–304 (2010)
10. Fries, T., Omerović, S.: Higher-order accurate integration of implicit geometries. Int. J. Numer.
Methods Eng. doi:10.1002/nme.5121 (2015, accepted)
11. Gordon, W., Hall, C.: Construction of curvi-linear co-ordinate systems and applications to
mesh generation. Int. J. Numer. Methods Eng. 7, 461–477 (1973)
12. Gordon, W., Hall, C.: Transfinite element methods: blending function interpolation over
arbitrary curved element domains. Numer. Math. 21, 109–129 (1973)
13. Laborde, P., Pommier, J., Renard, Y., Salaün, M.: High-order extended finite element method
for cracked domains. Int. J. Numer. Methods Eng. 64, 354–381 (2005)
14. Legay, A., Wang, H., Belytschko, T.: Strong and weak arbitrary discontinuities in spectral finite
elements. Int. J. Numer. Methods Eng. 64, 991–1008 (2005)
15. Legrain, G., Chevaugeon, N., Dréau, K.: High order X-FEM and levelsets for complex
microstructures: uncoupling geometry and approximation. Comput. Methods Appl. Mech.
Eng. 241–244, 172–189 (2012)
16. Li, Z.: A fast iterative algorithm for elliptic interface problems. SIAM J. Numer. Anal. 35,
230–254 (1998)
17. Lorensen, W., Cline, H.E.: Marching cubes: a high resolution 3d surface construction
algorithm. SIGGRAPH Comput. Graph. 21, 163–169 (1987)
18. Mittal, R., Iaccarino, G.: Immersed boundary methods. Annu. Rev. Fluid Mech. 37, 239–261
(2005)
19. Moës, N., Dolbow, J., Belytschko, T.: A finite element method for crack growth without
remeshing. Internat. J. Numer. Methods Eng. 46, 131–150 (1999)
20. Moumnassi, M., Belouettar, S., Béchet, É., Bordas, S., Quoirin, D., Potier-Ferry, M.: Finite
element analysis on implicitly defined domains: an accurate representation based on arbitrary
parametric surfaces. Comput. Methods Appl. Mech. Eng. 200, 774–796 (2011)
21. Mousavi, S., Sukumar, N.: Numerical integration of polynomials and discontinuous functions
on irregular convex polygons and polyhedrons. Comput. Mech. 47, 535–554 (2011)
22. Müller, B., Kummer, F., Oberlack, M.: Highly accurate surface and volume integration on
implicit domains by means of moment-fitting. Internat. J. Numer. Methods Eng. 96, 512–528
(2013)
23. Osher, S., Fedkiw, R.: Level Set Methods and Dynamic Implicit Surfaces. Springer, Berlin
(2003)
24. Phillips, G.: Interpolation of Approximation by Polynomials. Springer, Berlin (2013)
25. Rivlin, T.: An Introduction to the Approximation of Functions. Dover, New York, NY (1981)
26. Sala-Lardies, E., Fernández-Méndez, S., Huerta, A.: Optimally convergent high-order X-FEM
for problems with voids and inclusion. In: Proceedings of the ECCOMAS 2012, Vienna,
Austria (2012)
27. Sethian, J.: Level Set Methods and Fast Marching Methods, 2nd edn. Cambridge University
Press, Cambridge (1999)
28. Stazi, F., Budyn, E., Chessa, J., Belytschko, T.: An extended finite element method with higher-
order elements for curved cracks. Comput. Mech. 31, 38–48 (2003)
29. Szabados, J., Vertesi, P.: Interpolation of Functions. World Scientific, Singapore (1990)
30. Szabó, B., Düster, A., Rank, E.: The p-Version of the Finite Element Method, chap. 5, pp. 119–
139. Wiley, Chichester (2004)
Integration for Cut Elements with Chen-Babuška Nodes 269

31. Ventura, G.: On the elimination of quadrature subcells for discontinuous functions in the
eXtended finite element method. Int. J. Numer. Methods Eng. 66, 761–795 (2006)
32. Ventura, G., Gracie, R., Belytschko, T.: Fast integration and weight function blending in the
extended finite element method. Int. J. Numer. Methods Eng. 77, 1–29 (2009)
33. Zi, G., Belytschko, T.: New crack-tip elements for XFEM and applications to cohesive cracks.
Int. J. Numer. Methods Eng. 57, 2221–2240 (2003)