Crack propagation

on highly heterogeneous composite

materials
Copyright c 2008 by Miguel Patrı́cio, Eindhoven, The Netherlands.
All rights are reserved. No part of this publication may be reproduced, stored in a
retrieval system, or transmitted, in any form or by any means, electronic, mechanical,
photocopying, recording or otherwise, without prior permission of the author.

A catalogue record is available from the Eindhoven University of Technology Library

ISBN: 978-90-386-1367-3
 Crack propagation
on highly heterogeneous composite
materials

PROEFSCHRIFT

ter verkrijging van de graad van doctor aan de

Technische Universiteit Eindhoven, op gezag van de
Rector Magnificus, prof.dr.ir. C.J. van Duijn, voor een
commissie aangewezen door het College
voor Promoties in het openbaar te verdedigen
op dinsdag 23 september 2008 om 16.00 uur

door

Miguel José Patrı́cio Dias

geboren te Coimbra, Portugal

Dit proefschrift is goedgekeurd door de promotor:

prof.dr. R.M.M. Mattheij

Copromotor:
dr. M.E. Hochstenbach
Acknowledgment

This thesis is the product of research that was carried out at the Eindhoven University
of Technology in the period ranging from October 2004 to August 2008. I find great
pleasure in expressing my gratitude to the people who have contributed to this work.

I would like to start by thanking my promotor prof.dr. (Bob) Mattheij for offering me the
opportunity to join a very interesting research project and for believing in me always. I
will never forget how much I have learned from the many interesting conversations we
had about maths, history, languages and life in general. I am sincerely grateful to my
co-promotor dr. (Michiel) Hochstenbach for all the useful comments, discussions and
the occasional game playing. Also, I greatly appreciate the kindness of prof.dr. (Bert)
de With. Our conversations were very enlightening for me.

I am indebted to dr.ir. (Fons) van de Ven for the many times I went to him with a
question. My thanks go to dr. (Sorin) Pop, ms. Enna van Dijk, dr. (Jos) Maubach and all
the other members of CASA for their great help and friendship.

It’s been great working and having fun with Mark van Kraaij, with whom I’ve had
the privilege of sharing an office. I would also like to mention Remo, Dragan, Vincent,
Evgeny, Willem, Yves, Kakuba, Kamyar and all the colleagues and friends that I’ve been
fortunate to meet. I greet all the friends that I’ve kept throughout the years at home and
abroad, in particular João, António, Mikael, Ana and Sandra. Thank you for bringing
joy and balance to my life.

Most of all I am grateful to my family, my brother Ricardo, my sister-in-law Marta and

my lovely niece Sofia. The loving support of my parents has been beyond words. I am
blessed to have them in my life and I would like to dedicate this thesis to them.
Contents

1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Problem setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Thesis layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Mathematical modelling of linear elastic materials 11

2.1 Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.1 Kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.2 Conservation laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.3 Constitutive equations . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 The problem of linear elasticity . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.1 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.2 Plane stress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3 The mechanics of crack propagation 23

3.1 Elastic failure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Fracture parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2.1 Global approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2.2 Local approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Fracture criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.4 Numerical aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.4.1 Stress intensity factors . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.4.2 Simulation of crack growth . . . . . . . . . . . . . . . . . . . . . . . 37

4 Homogenisation for periodic structures 41

4.1 One-dimensional model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.1.1 Multiple scales method . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.1.2 Recovering the heterogeneities . . . . . . . . . . . . . . . . . . . . . 47
4.1.3 Numerical example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2 Homogenisation for elasticity . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2.1 Main result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2.2 Layered composites . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2.3 Numerical aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
viii Contents

5 Domain decomposition 67
5.1 One dimensional example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.1.1 Substructuring iterative methods . . . . . . . . . . . . . . . . . . . . 68
5.1.2 Alternating Schwarz methods . . . . . . . . . . . . . . . . . . . . . . 71
5.2 The elasticity problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.2.1 The Steklov-Poincaré equation . . . . . . . . . . . . . . . . . . . . . 76
5.2.2 Convergence analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.3 Discrete solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

6 Hybrid approach 91
6.1 One-dimensional problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.1.1 The algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.1.2 Behaviour of the error . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.2 The elasticity problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.3 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.3.1 One-dimensional example . . . . . . . . . . . . . . . . . . . . . . . . 101
6.3.2 Layered elastic materials . . . . . . . . . . . . . . . . . . . . . . . . . 103

7 Fracture of composites 109

7.1 Behaviour of the SIFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.1.1 Effects of the local structure . . . . . . . . . . . . . . . . . . . . . . . 110
7.1.2 Computational aspects . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.2 Crack paths in layered materials . . . . . . . . . . . . . . . . . . . . . . . . 117

Bibliography 125

Index 131

Summary 133

Samenvatting 135

Curriculum vitae 137

Chapter 1

Introduction

1.1 Motivation

Engineering structures are designed to withstand the loads they will be subjected to
while in service. Large stress concentrations are avoided and a reasonable margin of
security is taken to ensure that values close to the maximum admissible stress are never
attained. However, material imperfections that arise at the time of production or us-
age of the material are unavoidable and must be taken into account. Indeed, there are
many unfortunate examples of situations where microscopic flaws have caused seem-
ingly safe structures to fail, cf. [15, 53, 95].

In the past, when a component of some structure exhibited a crack, it was either repaired
or simply retired from service. Such precautions are nowadays often deemed unneces-
sary, not possible to enforce, or may prove too costly. On one hand, the safety margins
assigned to structures have to be smaller, due to increasing demands for energy and
material conservation. On the other hand, the detection of a flaw in a structure does not
automatically mean that it is not safe to use anymore. This is particularly relevant for
expensive materials or components of structures whose usage it would be inconvenient
to interrupt.

In this setting fracture mechanics plays a central role, as it provides useful tools allow-
ing an analysis of materials that exhibit cracks. The goal is to predict whether and in
which manner failure might occur. The origins of this branch of science in the western
world can be traced back to at least as early as the work of Leonardo da Vinci, cf. [77,86].
He did a study of fracture strength of materials using a device described in the Codex
Atlanticus [57,89], as represented in Figure 1.1. The experience consisted of suspending
a basket on a wire of a given thickness and length and allowing it to be slowly filled
with sand from an adjacently suspended hopper. A spring placed on the opening of the
hopper ensured that the sand would stop dropping once the wire broke. The weight of
2 Introduction

the basket with sand provided a measure of the tensile strength of the wire. By repeat-
ing the experience for wires of different lengths, da Vinci observed that shorter wires
supported greater weight than longer wires. This conflicts with the classical theory of
mechanics of materials that states that the stress in each unit of the wire should be the
same regardless of its length. An explanation for this can be provided by the likelihood
of larger fluctuations in the width of longer wires or in other words, by the idea that a
chain is only as strong as its weakest link. The width of a presumably homogeneous
wire actually varies with the position along it and the wire breaks in the region where
its width is smallest.

Figure 1.1: Basket (b) hanging on a wire (a), being filled up by sand falling from a
hopper (c).

More than a century later Galileo investigated the influence of the size of structures in
fracture, cf. [16,23,32]. Much more recently, the English aeronautical engineer Alan Grif-
fith was able to theorise on the failure of brittle homogeneous linear elastic materials,
see [34]. The classic model of a linear elastic material had been given by Robert Hook’s
work in the late XVII century. He related the force applied to a perfect spring to the ex-
tension of the spring, cf. [86]. Hook’s law for linear elastic materials states that the stress
on a solid medium is directly proportional to the strain produced, as long as the limit
of elasticity is not exceeded. As a consequence, in the absence of stresses, a solid elas-
tic body remains in its undeformed reference state. Also, no permanent deformations
occur. After loading, when the stresses are again zero, the body returns to its original
configuration.
1.1 Motivation 3

Griffith used a thermodynamic approach to analyse the centrally cracked glass plate
present in an earlier work of Inglis [41]. His theory was strictly restricted to elastic brittle
materials like glass, in which virtually no plastic deformation near the tip of the crack
occurs. However, extensions that account for such a deformation and further extend
this theory were later suggested, for example in [42–45, 65].

Up to the present day fracture mechanics has remained a hot topic, with many open and
intriguing questions. In particular there is great interest in predicting crack propagation
on composite materials. Composites consist of two or more chemically or physically
dissimilar constituents that are bonded together along interior material interfaces and
do not dissolve or blend into each other, cf. [48, 52]. The idea is that by putting together
the right ingredients a material with a better performance can be obtained. Composites
are heterogeneous at the microscale, by which we mean the scale of the constituents,
but they can be considered homogeneous at the macroscale. When the constituents are
finely mixed, the materials are said to be highly heterogeneous.

Figure 1.2: Manufacturing of bricks in Egypt; tomb of Rekhmire.

Nowadays composites such as fiber reinforced polymers or plywood are very common
and can be found in many daily life products. But composites have actually been used
by humans for thousands of years. An early example are the bricks made of mud and
straw used in Ancient Egypt and mentioned in the Bible in the book Exodus. The man-
ufacturing process of these bricks is represented in Figure 1.2.
4 Introduction

1.2 Problem setting

To discuss the problem of crack propagation further, we begin by considering a simple

example related to the cracked plate represented in Figure 1.3, which is loaded along its
upper and lower edges so that the crack surfaces are pulled apart in the x2 -direction. We
want to know under which conditions the initially stationary crack will start to grow,
assuming the plate to be isotropic, homogeneous, brittle and linearly elastic.

x2

x1

Figure 1.3: Finite plate with a centre through crack under tension.

The analysis of the fracture phenomenon focuses on the behaviour of the stress field in
the crack tip region. Each stress component displays a singular behaviour characterised
by local parameters called stress intensity factors (SIFs). These parameters are related
to the geometry, the constitution of the material and the external boundary conditions,
cf. for example [15, 94]. In particular, for the plate we have considered the horizontal
tensile stress ahead of the crack tip and along the x1 -axis is characterised by the mode
I SIF denoted by KI . A fracture criterion tells us that the crack will propagate when KI
reaches a critical value that is material dependent, see [15, 37, 94].

In this simple example the crack may only grow in the direction of the x1 -axis. In more
general situations the direction of propagation is not known a priori. The previous frac-
ture criterion can then be naturally extended and the crack growth assumed to occur in
a direction given as a function of the SIFs, cf. for example [30, 90]. In turn, the SIFs can
be computed from the displacement and stress fields so that solving the elasticity prob-
lem is the first step in predicting whether and how a crack present on a homogeneous
material will propagate. In general, the solutions of elasticity problems are obtained
in terms of the displacements. There are many situations in which these solutions are
either known [49] or can be easily approximated using classical numerical techniques
1.2 Problem setting 5

such as finite element methods or boundary element methods, cf. [7, 10, 11].

Given a cracked plate, an incremental approach to predict the future crack path can then
be implemented, [30,37,55,67]. The crack is modelled as a very thin and long hole in the
geometry, along which stress-free boundary conditions are prescribed. Assuming the
crack to be static, the elasticity problem is solved and the direction for crack propagation
determined. The crack is then incremented by updating the geometry. The procedure is
repeated to determine the path further.

Figure 1.4: Cracked thin plate subject to vertical loading (left). Zoom in of the crack tip
region displaying the various constituents of the composite (right).

Having considered crack propagation on homogeneous materials only is a rather large

simplification. It is a much more complicated problem to know whether and how a pre-
existent crack on a composite material may grow, see Figure 1.4. For now let us consider
the situation in which the crack is propagating inside a homogeneous component of
the composite and far from any internal material interfaces. As before, to model the
crack propagation, we would have to first solve the elasticity problem. When dealing
with composites, the coefficients of the underlying PDEs are material dependent and
it is often the case that the constituents of the composite are finely mixed so that these
coefficients jump between different values along the spatial coordinates very rapidly.
This gives rise to very complicated problems with PDEs that have highly oscillatory
coefficients. These are usually very hard to solve also numerically. When an accurate
representation of the evolution of cracks is necessary, alternatives have to be found.

Many strategies that extend and adapt the method of finite elements for problems in-
volving PDEs with coefficients that oscillate very rapidly have been developed, cf. [5, 6,
28, 38, 39]. A different approach is given by the theory of homogenisation. It can be effi-
ciently applied to PDEs with periodic coefficients. It allows for the approximation of the
solutions of the original heterogeneous problem in terms of the solutions of the homo-
geneous problem comprised by the same boundary conditions but simplified PDEs, the
homogenised or effective equations. The idea is to establish the macroscopic behaviour of a
system that has heterogeneities on the microscopic level. The starting point is then a PDE
with -periodic coefficients, with   1. The crucial step of the homogenisation process
6 Introduction

consists of taking  to zero. This is the same as assuming that the heterogeneities in the
composite are so small that we may replace it by a fictitious homogeneous material. This
asymptotic process, together with solving the homogenised problem, is in general compu-
tationally much easier than solving the original problem. The concept of homogenisa-
tion has been associated to other techniques and looked at from different points of view,
giving rise to very efficient algorithms, cf. for example [2, 46, 71]. We like to cite [21] in
particular, which updates the earlier writings of [8, 12, 78].

For periodic composite materials, this upscaling procedure has permitted significant
progress. In particular it allows for an enormous simplification in the study of layered
materials, cf. [68]. This type of materials is singled out due to their simplicity and be-
cause they are very commonly employed in the construction of structures with high
performance and are also used for a wide range of applications, from sensor devices
to magnetic or optical imaging. As an example, one may think of multilayer capaci-
tors [51, 93] or laminar composites [25, 54].

A distinct approach to tackle elasticity problems related to highly heterogeneous com-

posites consists of employing domain decomposition methods, see for example [26, 72, 79].
Unlike homogenisation, the heterogeneities are now resolved, which leads to more ac-
curacy but also more computational complexity. Domain decomposition allows the
division of the computational domain into smaller subdomains where the differential
equation is to be solved. The original problem is then split into problems set on these
subdomains, coupled together by matching conditions. In this way, instead of solv-
ing one very complicated problem, several less complicated problems will be solved.
Domain decomposition is a very versatile technique. It allows for a parallel implemen-
tation as well as the usage of different numerical schemes within distinct subdomains. It
is also advantageous for dealing with irregular geometries, discontinuous coefficients,
local grid refinements, boundary layers and coupling between equations of different
types, cf. for example [72, 80].

Both homogenisation and domain decomposition were introduced to deal with the com-
plexity of elasticity problems for composite materials. A hybrid approach that borrows
concepts from these two techniques is further considered. The advantage of this alter-
native approach is that it allows for a microscopic analysis to be restricted to where it
is relevant, thus reducing the computational complexity of the problems dramatically
when compared to domain decomposition methods. On the other hand, it provides
more accuracy than homogenisation, cf. [69, 70]. This hybrid approach is particularly
useful if on one or more localised regions of the computational domain a phenomenon
such as crack propagation is to be analysed, for which the microstructure plays an im-
portant role. The study of these critical regions may be performed separately from the
remainder of the domain, where a macroscopic approximation is sufficient. Such a com-
promise between the techniques of domain decomposition and homogenisation proves
very effective if one intends to describe the behaviour of a pre-existent crack on a com-
posite. Indeed, this is a local phenomenon that is primarily affected by the microstruc-
ture in the vicinity of the crack tip.

Let us now go back to the problem of fracture on a composite. When the crack is prop-
1.3 Thesis layout 7

agating inside one the the material components, the incremental approach for homo-
geneous materials can be used effectively by employing domain decomposition or the
hybrid approach to solve the elasticity problems. However, the crack will eventually
propagate on further to other materials and interact with internal boundaries. Unless
we look at this problem at a macroscale and deal with a homogeneous material, fac-
tors such as the nature and distribution of the constituents or the effect of the interfaces
have to be taken into account, cf. [19]. This yields an extremely complex problem, and
actually other aspects such as the existence of defects, the possibility of branching, cre-
ation of new cracks or delamination may also have to be incorporated into the analysis,
cf. [9, 24, 63].

To provide a concrete framework that captures the fundamental aspects of the problem
of fracture on a composite we consider a crack propagating through a periodically lay-
ered material. The aforementioned incremental approach can be adapted and extended
for this problem by accounting for the interaction with the interfaces and making use of
the hybrid technique to solve the complex elasticity problems.

1.3 Thesis layout

The present thesis is constituted by 7 chapters. In Chapter 2 the basic concepts of the
classical theory of linear elasticity are introduced. The governing equations are estab-
lished in Section 2.1. The problem of linear elasticity is then presented in a general
setting in Section 2.2. In particular this can be formulated to model the behaviour of
composite materials. The situation of thin plates under plane stress is also addressed.
We conclude this chapter by introducing a weak formulation of the problem of linear
elasticity and indicating how numerical approximations for its solution can be found
using finite element methods.

Being able to determine the stress and displacement fields of a given cracked plate is
not sufficient to model the growth of a propagating crack. One further needs to decide
on criteria to determine under which conditions cracks will propagate as well as the
direction of propagation. This is the main topic of Chapter 3, which begins with a short
introduction in Section 3.1. Subsequently, in Section 3.2 several fracture parameters are
introduced. We distinguish between global parameters and local parameters. Fracture
criteria given in terms of the fracture parameters are then introduced in Section 3.3.
The chapter is concluded with a discussion of various numerical aspects in Section 3.4.
Fracture parameters are computed using several numerical methods and it is shown
that with the J-integral method one obtains accurate reliable approximations. Finally,
an algorithm is set up to predict the path of a growing pre-existent crack on an isotropic
homogeneous linear elastic plate loaded in a mixed mode situation. This chapter may
be regarded as an overview of some aspects of linear elastic fracture mechanics and is
an enlarged version of [67].

One of the main issues arising in modelling crack propagation on composite elastic ma-
8 Introduction

terials is solving the related elasticity problems, which are often very complex. When
these materials are periodically distributed, one may apply the homogenisation the-
ory, which is the object of study of Chapter 4. We start in Section 4.1 by presenting
a one-dimensional example of an elliptic differential equation with periodic oscillating
coefficients. This is meant to show how homogenisation techniques provide an accurate
approximation for the solution of this example problem without having to resolve the
microscale, hence avoiding prohibitively large computational costs. It is also shown that
if more accuracy is sought correctors can be employed to recover the heterogeneities.
Next, in Section 4.2, the 2D elasticity problem for periodically distributed composite
materials is considered. The asymptotic behaviour of the solution of the underlying
equations is analysed and the necessary fundamental concepts of the homogenisation
theory literature are introduced. The homogenised equations are presented, as well as
the main convergence result. The elastic behaviour of layered composites, a prime ex-
ample of periodically distributed composites due to their widespread applications, is in-
vestigated. Explicit formulae for the respective effective coefficients are found, allowing
us to establish several useful properties. In particular it is shown that the homogenised
material that corresponds to an isotropic elastic layered material is orthotropic, with en-
gineering constants given as a function of those characterising the constituents of the
composite. The chapter is concluded with a discussion of some numerical aspects. An
efficient procedure to determine the effective solutions for elasticity problems related
to periodic structures numerically is proposed and illustrated with several examples.
We note that the novel results of this chapter, in particular sections 4.2.2 and 4.2.3, are
contained in [68].

In Chapter 5 domain decomposition methods are studied. Like the homogenisation

techniques these methods are considered to handle the computational complexity of
elasticity problems related to highly heterogeneous materials. The key idea is to di-
vide the computational domain into smaller subdomains where the original equation
is to be solved. This permits the solution to be obtained with microscopic resolution
throughout, eventually employing parallelisation. Following the same pattern as in the
previous chapter, we begin by illustrating the basic principles of the domain decom-
position methods with a one-dimensional example in Section 5.1. It is discussed how
approximations for the solution of the problem can be found using either overlapping
or non-overlapping methods. We conclude the section by establishing a convergence re-
sult. In Section 5.2 the extension to the planar elasticity problem is considered. Domain
decomposition methods are related to an interface equation expressed in terms of the
Steklov-Poincaré operator. The Dirichlet-Neumann method is presented and it is shown
how it reduces to the preconditioned Richardson method for the Steklov-Poincaré equa-
tion, which allows us to establish the convergence. An algorithm expressing an over-
lapping scheme is presented. We conclude by studying the Dirichlet-Neumann method
at the discrete level in Section 5.3 and it is discussed how the acceleration parameter for
this method can be optimised.

In many problems where different length scales are involved the microstructure is only
really relevant on a localised subdomain where the phenomena one wants to model
occurs. Everywhere else a macroscopic analysis is deemed sufficient. A hybrid ap-
proach to deal with this class of problems is proposed in Chapter 6, combining ho-
1.3 Thesis layout 9

mogenisation and domain decomposition to form a new method, cf. [69, 70]. The idea
is to resolve the microstructure where necessary and homogenise elsewhere to obtain
an accurate solution with reasonably small computational effort. We start by analysing
a one-dimensional boundary value problem in Section 6.1. An iterative scheme to find
numerical approximations for the solution of this problem is proposed and the associ-
ated error is studied. In Section 6.2 the hybrid approach is extended for the problem
of linear elasticity. Finally examples of application for the two problems considered are
presented in Section 6.3.

The numerical techniques presented in the three previous chapters allow us to extend
the analysis of Chapter 3 to the study of cracked composite materials. We begin Chap-
ter 7 by investigating how stress intensity factors are affected by the local microscopic
structure. The advantages of employing the hybrid approach to solve the related elas-
ticity problem are highlighted in Section 7.1. Finally, in Section 7.2 the influence of
internal material boundaries on a propagating crack is analysed. An algorithm to pre-
dict the path of pre-existent cracks on highly heterogeneous materials, making use of
the aforementioned hybrid approach, is proposed. The technique is illustrated for a pe-
riodic composite and shown to produce very satisfactory results when compared to a
reference solution.
Chapter 2

Mathematical modelling of linear

elastic materials

The behaviour of linear elastic materials such as steel or copper can be modelled by
second order elliptic PDEs. These differential equations take into account the specific
properties of the medium. This sometimes leads to very complicated problems, partic-
ularly for highly heterogeneous structures.

In this chapter we introduce the basic concepts and equations that allow the mathe-
matical analysis of mechanics problems involving linear elastic materials. The general
elasticity problem is presented. Special attention is devoted to thin plates under plane
stress.

2.1 Governing equations

The deformation of a body is expressed in terms of the strain and stress tensors, as well
as the displacement vector. These fields satisfy the kinematic relations, conservation
laws and constitutive equations that we will introduce in this section.

2.1.1 Kinematics

To study the motion of a deformable 3-dimensional continuous body B, we start by

setting up a suitable kinematical framework that does not depend on the constitution
of the body or the forces that act on it, cf. [33, 36, 87]. As a result of the action of several
forces, the body B will move and deform. We refer to the configuration of B at time t,
12 Mathematical modelling of linear elastic materials

i.e., the region Ωt that B occupies at time t, as the current configuration. We assume
that for t = 0, the body is both undeformed and unstressed. It is then said to be at its
reference configuration Ω.

We identify each point P of B with the position x = (x1 , x2 , x3 ) ∈ Ω it occupies in

the reference configuration, with respect to a Cartesian coordinate system with origin O
and unit basis vectors {e1 , e2 , e3 }. We refer to x as a material point. As the body moves,
the position of this point will also vary and we denote by y = y(x, t) the position that it
occupies at time t. The vector function y is called the motion.

W u(x,t)
x Wt
y(x,t)
e2

O
e1
e3

Figure 2.1: Motion and displacement of the material point x.

For each point x ∈ Ω, clearly y(x, 0) = x. Rather than working with y as a primary
variable, it is useful to introduce the displacement vector field u(x, t) = y(x, t) − x, see
Figure 2.1. The displacement field characterises the motion of the body. It accounts for
both the rigid body motions, i.e., translations and rotations and also for the deforma-
tions in which there are relative movements and distortions within the body. Not being
interested in rigid body motions, we introduce a quantity that allows us to measure only
deformation, the strain tensor η = (ηij )1≤i,j≤3 given by

1
η= (∇u + (∇u)T + (∇u)T ∇u). (2.1.1)
2

Here,
2.1 Governing equations 13

 ∂u1 ∂u1 ∂u1 

∂x1 ∂x2 ∂x3
 
∂u2 ∂u2 ∂u2
∇u = 
 
∂x1 ∂x2 ∂x3 
 
∂u3 ∂u3 ∂u3
∂x1 ∂x2 ∂x3

is the displacement gradient matrix. To interpret the physical meaning of η, we start

by considering the points x + δxi in the reference configuration Ω, for i = 1, 2. In
the current configuration, x and the two vectors δxi will be mapped to y and δyi =
y(x + δxi , t) − y(x, t) respectively, see Figure 2.2.

W Wt
dy1
dy2
dx1 y
dx2
x

Figure 2.2: Configurations of B.

To analyse the changes in both lengths and relative angles of the vectors δx1 and δx2
after deformation, we take the difference δy1 · δy2 − δx1 · δx2 . We assume that it is
possible to expand the functions y(x + δx1 , t) and y(x + δx2 , t) in Taylor series about
x, so that we can write

δy1 · δy2 − δx1 · δx2 = (∇uδx1 ) · δx2 + (∇uδx2 ) · δx1

+(∇uδx1 ) · (∇uδx2 ) + . . . (2.1.2)

Taking the limit in (2.1.2) as δ := max{kδx1 k, kδx2 k} goes to zero, where k · k denotes the
Cartesian norm, it follows that

δy1 · δy2 − δx1 · δx2

lim = 2m · η(u)n, (2.1.3)
δ→0 δ2

where m := δx1 /δ and n := δx2 /δ are fixed vectors, independent of h.

14 Mathematical modelling of linear elastic materials

It is clear that if the body moves as a rigid body, then η(u) = 0, because the term on
the left hand side of (2.1.2) must be zero. The converse is also true, as can be seen
by interpreting the components of η. To do so, we start by looking at the diagonal
components of the strain tensor. These measure the relative elongation of vectors that
are parallel to the coordinate axes. Indeed, assume for example that δx1 = δx2 = e1 .
Then from (2.1.3) we see that

1 kδy1 k2 − kδx1 k2
η11 = lim .
2 δ→0 δ2

As for the off-diagonal components of η, they give a measure for the change in the
angle between two vectors each parallel to a different coordinate axes. To illustrate this,
assume that δx1 and δx2 have the same length δ and that they are parallel to e1 and e2 ,
respectively. Then

1 δy1 · δy2
η12 = lim .
2 δ→0 δ2

Throughout this thesis we will assume that the components of the displacement gradi-
ent are small. The material body B is then said to undergo infinitesimal deformation.
The nonlinear term in (2.1.1) can be neglected and the strain tensor η may be replaced
by the symmetric infinitesimal strain tensor  given by

1
(u) = (∇u + (∇u)T ). (2.1.4)
2

Henceforth we will simply refer to  as the strain tensor. The relationships expressed by
(2.1.4), relating the strain tensor to the displacement vector, are known as the kinematic
equations.

2.1.2 Conservation laws

Classical continuum mechanics is based upon a system of fundamental laws that ap-
ply for all material bodies, both solid and fluid. These express the balances of mass,
momentum, moment of momentum and energy by postulating that in the absence of a
source, these quantities remain unchanged. Here we present the laws of balance in their
global form and introduce the notion of the stress tensor. The equations of motion for a
continuum are deduced.

The global balance laws are formulated for an arbitrary sub-body B 0 of a body B. At all
times during the motion, B 0 contains the same set of points. We denote by Ω 0 ⊂ Ω and
Ωt0 ⊂ Ωt the regions occupied by B 0 in the undeformed state and at the time instant
2.1 Governing equations 15

t, respectively. Amongst the forces acting on B 0 , we distinguish between the body force
f = ρb(x, t) and the surface traction sn (x, t), where ρ is the mass density and b is the
specific external force field, or specific volume force field. The body force is due to
external forces, e.g. the gravity. It represents the force per unit of volume acting at the
time instant t on each particle x of B 0 . As for the surface traction, also called the stress
vector, it is the force per unit of area exerted by the part of B outside of B 0 across the
border of B 0 , see Figure 2.3. It depends on the point of action x ∈ ∂Ωt0 , on the time t and
also on the outward pointing normal n in x on ∂Ωt0 .

sn(x,t)
n

Wt
Wt

Figure 2.3: Stress vector field defined over the border of Ωt0 .

We are now ready to state the mechanical global balance laws.

Conservation of mass: in the absence of mass sources, the mass of the material con-
tained in Ω 0 is constant,

Z
d
ρdV = 0. (2.1.5)
dt Ωt0

Conservation of linear momentum: the rate of change of momentum of the material

contained in B 0 , due to the movement of the material with velocity u̇, is equal to the
sum of the body forces and surface forces acting on the material,

Z Z I
d
ρu̇dV = ρbdV + sn dS (2.1.6)
dt Ωt0 Ωt0 ∂Ωt0

Conservation of angular momentum: the rate of change of angular momentum in Ω 0 ,

when the control volume moves with the continuum, is equal to the sum of the moment
16 Mathematical modelling of linear elastic materials

of the body forces and the moment of the surface forces acting on it,

Z Z I
d
x × ρu̇dV = x × ρbdV + x × sn dS. (2.1.7)
dt Ωt0 Ωt0 ∂Ωt0

Conservation of energy: the rate of change of kinetic energy and internal energy is
equal to the mechanical power of the stresses and body forces acting on the material
added to the heat supplied by internal sources and exchanged around the border,

Z I Z I Z
d
ρEdV = sn · u̇ds + ρb · u̇dV − q · ndS + ρrdV. (2.1.8)
dt Ωt0 Ωt0 Ωt0 Ωt0 Ωt0

In (2.1.8), q is the heat flux vector, r the specific heat supply and E is the specific energy,
cf. [62].

Since we assume the deformations to be infinitesimal, the distinction between the refer-
ence and current configurations may be ignored, cf. [36]. One would like to extract local
relationships, valid for each of the material points in the body, from the global laws. Fol-
lowing [62, 87], we first introduce the second order tensor field σ = (σij )1≤i,j≤3 called
the Cauchy stress, which verifies the Cauchy law

σ · n = sn , (2.1.9)

for each unit vector n. The Cauchy stress characterises the state of stress for each point
of B. Now, using (2.1.9) and Gauss’ theorem, we can rewrite (2.1.6) in the form

Z
ρü − ρb − ∇ · σdV = 0, (2.1.10)
Ωt0

attending to the fact that by the transport theorem

Z Z
d
ρu̇dV = ρüdV. (2.1.11)
dt Ωt0 Ωt0

Since (2.1.10) holds for any subdomain Ωt0 of Ωt , the integrand must vanish and

∇ · σ + ρb = ρü. (2.1.12)
2.1 Governing equations 17

This equation is known as Cauchy’s equation of motion. In the static case, when ü = 0,
it reduces to the equation of equilibrium

∇ · σ + ρb = 0. (2.1.13)
We note that (2.1.7) allows us to show that the tensor σ is symmetric, cf. [62].

2.1.3 Constitutive equations

The kinematic equations (2.1.4) and the Cauchy’s equation of motion (2.1.12) hold for
any continuous medium. However, the mechanical properties of the material must also
be taken into account. These are expressed by the constitutive equations or material laws.
For linear elastic materials, the stress σ depends linearly on the strain . They are related
by the generalised Hook’s law

σ = A, (2.1.14)

where A = A = (aijkh )1≤i,j,k,h≤2 is a fourth-order tensor called the elasticity tensor,

see [36, 87]. It is convenient to write (2.1.14) in the matrix form

a1111 a1122 a1133 a1112 a1123 a1131

 
 
   a2211 a2222 a2233 a2212 a2223 a2231   
σ11 


 11

 σ22  
  a3311 a3322 a3333 a3312 a3323 a3331  
 22 

 σ33    33 
 =  . (2.1.15)

 σ12  
  a1211 a1222 a1233 a1212 a1223 a1231 

 12 

 σ23    23 
σ31

 a2311 a2322 a2333 a2312 a2323 a2331 
 31
 
 
a3111 a3122 a3133 a3112 a3123 a3131

The elasticity tensor is a function of position in the body, but it is independent of time
as long as the material remains elastic, A = A(x). Due to the symmetry of the stress and
strain tensors, the components of A satisfy

aijkh = ajikh = akhij , for i, j, k, h = 1, 2, 3.

The body B with elastic behaviour characterised by A is called homogeneous if both

ρ and A are independent of spacial coordinates. Another classification arises from the
18 Mathematical modelling of linear elastic materials

fact that some materials have certain properties of symmetry which allow the number
of independent components of the elasticity tensor to be reduced, cf. [33]. A medium
is then called isotropic if at each material point its mechanical properties are identical in
all directions and anisotropic otherwise. It is called orthotropic if it has different materials
properties along different orthogonal directions. In particular, for isotropic materials,
(2.1.14) can be rewritten as

σ = λ(tr)I + 2µ, (2.1.16)

where λ and µ are the so called Lamé moduli. These are related to the relevant material
parameters Young’s modulus E, the Poisson’s ratio ν and the shear modulus G by

µ(2µ + 3λ) λ E
E= , ν= and G = .
µ+λ 2(µ + λ) 2(1 + ν)

2.2 The problem of linear elasticity

The governing equations that have been introduced allow the mechanical behaviour of
linear elastic materials to be translated into a mathematical formulation. In particular,
the elasticity problem for composite materials may be considered.

2.2.1 Problem formulation

In the dynamic case, the linear elastic behaviour of a material body B can be expressed
in terms of the displacement vector u and the strain and stress tensors  and σ by the
following equations, cf. for example [58, 62, 94].

∇ · σ + f = ρü, (2.2.1)
1
(u) = (∇u + (∇u)T ), (2.2.2)
2
σ = A. (2.2.3)

We note that the static case is obtained replacing ρü by 0 in (2.2.1). In order for the elas-
ticity problem to be complete and consistent, suitable boundary conditions are needed
at each point of ∂Ω, cf. [61, 87]. For this the equations (2.2.1)-(2.2.3) are supplemented
by boundary conditions
2.2 The problem of linear elasticity 19

u = ϕD , x ∈ ΓD , (2.2.4)
σ · n = ϕN , x ∈ ΓN , (2.2.5)

related to the undeformed boundary Γ = ∂Ω. Here, ΓD and ΓN are the subsets of Γ where
the displacements and the stress vectors are prescribed and the indices in ϕD and ϕN
refer to Dirichlet and Neumann, respectively. We note that for the dynamic case initial
boundary conditions must be added for the displacements and their time derivatives.
The existence and uniqueness of the solution for this problem is ensured, possibly apart
from rigid body motions, [87].

The equations of the problem of elasticity can be reduced to a system of three equations
in which the displacement vector field is the primary unknown. We eliminate the stress
and strain from the governing equations by substitution to obtain the Navier equations
for the displacements. These read

∇ · (A(u)) + f = ρü.

As for the boundary condition expressed by (2.2.5) it is given by

A(u) · n = ϕN , x ∈ ΓN . (2.2.6)

When dealing with composite materials, the components of the elasticity tensor vary
with the spacial coordinates, giving rise to highly oscillatory PDEs. At the internal ma-
terial interfaces, the coefficients of the differential equation display a discontinuity. The
mathematical problem for composites is increasingly complex for more heterogeneous
materials. It is still formulated as (2.2.1)-(2.2.5), but at the internal boundary no-jump
conditions are prescribed, i.e., it is assumed that the displacements and the tractions are
continuous.

2.2.2 Plane stress

The elasticity problem for the three-dimensional case is not easy to solve. However,
under certain conditions, the problem may be simplified. This is the case for situations
of plane stress, where thin plate-like structures are considered. For these, the normal and
shear stresses in the x3 -direction can be assumed to be zero.

In this thesis we will consider the static elasticity problem for thin plates in a plane stress
situation that reads
20 Mathematical modelling of linear elastic materials



 −∇ · (A(x)(u)) = f, x ∈ Ω,


u = 0, x ∈ ΓD , (2.2.7)


 σ(u) · n = ϕN ,
 x ∈ ΓN .

The primary variable is the displacement vector u = (ui )1≤i≤2 . The elasticity tensor A
which relates σ and  can be represented by a 3 × 3 matrix. For isotropic materials we
have

 
E/(1 − ν2 ) (Eν)/(1 − ν2 ) 0
 
A =  (Eν)/(1 − ν2 ) E/(1 − ν2 ) 0 , (2.2.8)
 
 
0 0 E/(2(1 + ν))

where the Young’s modulus E and the Poisson’s ratio ν are the material parameters
introduced earlier. For orthotropic materials A is given in terms of the Young’s moduli
Exy , Eyx , the Poisson’s ratios νxy , νyx and the shear modulus Gxy , so that

 
Ex /(1 − νxy νyx ) (Ex νxy )/(1 − νxy νyx ) 0
 
A =  (Ey νyx )/(1 − νxy νyx ) Ey /(1 − νxy νyx ) 0 . (2.2.9)
 
 
0 0 Gxy

where νxy Ey = νyx Ex .

For future convenience we introduce the weak formulation for (2.2.7). Let Ω be a con-
nected bounded open set in Rn . Moreover, let ∂Ω = ΓN ∪ ΓD be Lipschitz continuous
such that ΓD is of measure greater than zero. From a Green’s formula, we obtain the
following variational formulation

a(u, v) = hl, vi, ∀ v ∈ V, (2.2.10)

where we define

Z X
2
a(u, v) := σij (u)ij (v)dV, (2.2.11)
Ω i,j=1
Z X
2 I X
2
hl, vi := fi vi dV + ϕN,i vi dS. (2.2.12)
Ω i=1 ΓD i=1
2.2 The problem of linear elasticity 21

Here, ϕN = (ϕN,i )i=1,2 and v = (vi )i=1,2 . As for V it is such that

V := {v | v ∈ (H1 (Ω))2 , γ(v) = 0 on ΓD }, (2.2.13)

where γ(v) := v|∂Ω is the trace of v on ∂Ω. Furthermore, we assume that f ∈ (L2 (Ω))2
and ϕN ∈ (L2 (ΓN ))2 . The bilinear form a is V-elliptic and continuous, whereas the
linear functional l is bounded and linear, cf. [21]. The Lax-Milgram lemma assures that
(2.2.10)-(2.2.12) has a unique solution.

A numerical approximation uh for the solution u ∈ V of (2.2.10) can be obtained by

using finite elements. The domain Ω is triangulated into non-overlapping elements, typ-
ically triangles, rectangles or tetrahedra. We consider a finite dimensional subspace of
V of dimension N, which we denote by Vh . The parameter h in subscript expresses the
maximum width of the side of the elements. A nodal basis {φj }j=1,...,N is defined for
Vh such that at the finite element nodes denoted by aj we have

φi (aj ) = δij ,

where δij is the Kronecker symbol. A finite element approximation uh to u consists of a

linear combination of elements of the nodal basis,

X
N
uh = yj φj .
j=1

Inserting this into the variational problem yields

XN
a( yj φj , φk ) = hl, φk i, ∀ k = 1, . . . , N. (2.2.14)
j=1

This produces a linear system for the vector of unknowns y = (yi )i=1,...,N that reads

Dy = b,

where we define the stiffness matrix D = [dij ]i,j=1,...,N and the load b = (bi )i=1,...,N by

dij := a(φi , φj ), bi := hl, φi i.

For a more thorough analysis of the finite element method, cf. for example [7, 10].
Chapter 3

The mechanics of crack

propagation

Many tragic examples have shown that the slightest flaws in engineering structures
might lead to major catastrophes. There is then great need of understanding the me-
chanics of fracture. This branch of science plays a central role in providing useful
tools that allow for an analysis of materials exhibiting cracks. A main goal is to pre-
dict whether and in which manner failure might occur.

In this chapter we focus on the brittle fracture of thin plates in a plane stress situation,
subject to in-plane loading. We start by introducing the relevant fracture parameters,
the energy release rate, the J-integral and the stress intensity factors. Different fracture
criteria are presented and related. We conclude by looking into numerical techniques for
determining the fracture parameters and by predicting the path for a crack propagating
on a homogeneous material.

3.1 Elastic failure

Let us consider a plate displaying a pre-existent crack. We can distinguish several ways
in which a force may be applied to the plate causing the crack to propagate. In [45]
a classification is proposed corresponding to the three situations represented in Figure
3.1. We will refer to these as mode I, mode II and mode III, respectively.

In the mode I, or opening mode, the body is loaded by tensile forces, such that the crack
surfaces are pulled apart in the x2 -direction. The deformations are symmetric with re-
spect to planes perpendicular to the x2 -axis and the x3 -axis.
24 The mechanics of crack propagation

In the mode II, or sliding mode, the body is loaded by shear forces parallel to the crack
surfaces, which slide over each other in the x1 -direction. The deformations are sym-
metric with respect to a plane perpendicular to the x3 -axis and skew-symmetric with
respect to a plane perpendicular to the x2 -axis.

Finally, in the mode III, or tearing mode, the body is loaded by shear forces parallel to
the crack surfaces. These slide over each other in the x3 -direction. The deformations are
skew-symmetric with respect to planes perpendicular to the x3 -axis and the x2 -axis.

x2
x3
x1

Figure 3.1 - a) Opening. b) Sliding. c) Tearing.

For each of these modes crack extension may only take place in the direction of the x1 -
axis, the original orientation of the crack. More generally, we typically find a mixed mode
situation, where there is a superposition of the modes and the crack may propagate in
any direction. For such a problem the principle of stress superposition states that the
individual contributions to a given stress component are additive, so that if σIij , σII
ij and
III
σij are the stress components associated to the modes I, II and III respectively, then
the stress component σij of a plate in a mixed mode situation is given by

σij = σIij + σII III

ij + σij , for i, j = 1, 2. (3.1.1)

Within the scope of the theory of linear elasticity, a crack introduces a discontinuity in
the elastic body such that the stresses tend to infinity as one approaches the crack tip.
Using the semi-inverse method as in [91], Irwin [43,44] related the singular behaviour of
the stress components to the distance to the crack tip r. In particular, when we consider
a Cartesian coordinate system centred at the crack tip, in the opening mode and in polar
coordinates we have

KI I
σij ' √ fij (θ), (3.1.2)
2πr
3.2 Fracture parameters 25

where the angular variation function fIij , which will be introduced later, depends only
on θ. The parameter KI , the stress intensity factor (SIF), plays a fundamental role in
fracture mechanics, as it characterises the stress field, reflecting the geometry of the
structure and the loading it is subject to. This approximation for the stresses is assumed
to be valid in the vicinity of the crack tip for linearly elastic materials. Actually, the
materials yield or deform inelastically very near the crack tip and so there is a region of
validity of the approximation, cf. [73, 95].

Henceforth we will consider problems of linear elastic, isotropic, homogeneous and

brittle cracked plates loaded under plane stress conditions. Both pure and mixed mode
III situations will not be considered.

3.2 Fracture parameters

We begin by introducing the concepts of the strain release rate and the J-integral. These
are global fracture parameters, based on an energetic approach. Next we focus on the
stress intensity factors, which arise from the local distribution of the stresses in the vicin-
ity of the crack tip.

3.2.1 Global approach

The failure of brittle materials, characterised by having virtually no plastic deforma-

tion near the tip of the crack, was first explained by Griffith using a thermodynamic
approach, cf. [34, 92]. He used the mathematical work of Inglis [41], who had earlier
analysed the case of an elastic infinite plate under uniform tensile stress. This had a
central elliptic flaw of semi-major axis a = c cosh ξ and semi-minor axis b = c sinh ξ,
for given values of c, see Figure 3.5 a). The key idea is that when ξ → 0, the elliptical
flaw degenerates into a line of length 2a which can be considered to represent a crack.
Under the assumption that the faces of the crack are free of stress, it can be assumed
that crack propagation will occur if the energy required to form a new crack can be de-
livered by the system, see [15, 94]. Mathematically, this can be translated in terms of the
quantities

σ2 πa2
U= , and W = 2aγ. (3.2.1)
2E

Here, U is the strain energy released in the formation of a crack of length a and W is
the corresponding surface energy increase, both per unit thickness of plate. The Young’s
modulus is denoted by E and γ is the material surface energy density or surface tension.
The criterion for crack growth is then given by the equality
26 The mechanics of crack propagation

∂U ∂W
G = R, where G = − and R = . (3.2.2)
∂a ∂a

This holds if the strain energy release rate G during crack growth is large enough to exceed
the rate of increase in surface energy R associated with the formation of new crack
surfaces.

The energy release rate can also be given in terms of the path-independent J-integral,
cf. [74]. This is defined as a line integral along a counterclockwise contour C which
surrounds the crack tip,
P see Figure 3.2. Its components are given in terms of the strain
energy density We = i,j=1,2 σij ij and the stresses by

I X ∂ui
Jk = (We nk − σij nj )dS, (3.2.3)
C ∂xk
i,j=1,2

for k = 1, 2 and where n = (n1 , n2 ) is the outward-pointing normal vector defined over
C, cf. [37, 90, 95]. The vector J = (Jk )k=1,2 can be regarded as the energy flux per unit
length into the crack tip. It is called path-independent because it does not depend on
the choice of the curve C, cf. [37].

Figure 3.2: Integration path (gray) around the crack tip.

In the local coordinate system with the crack tip as origin and the crack located along
the negative x1 -axis represented in Figure 3.3 the J-integral and the strain release rate
are related by

J1 = G, (3.2.4)

cf. for example [50].

3.2 Fracture parameters 27

3.2.2 Local approach

Let us consider a static crack in a plate which is in a plane stress situation. Assume that
the crack surfaces are free of stress and that the axes are positioned as in Figure 3.3.

x2

q
x1
Figure 3.3: Crack tip coordinates.

For linearly elastic materials, the stress field in the vicinity of the crack tip is given in
polar coordinates by

KI I KII II
σ(r, θ) = √ f (θ) + √ f (θ), (3.2.5)
2πr 2πr

√
up to an error of order of r for each component, cf. for example [15, 20, 31, 45]. In
(3.2.5), the angular variation functions fI = (fIij )i,j=1,2 for mode I are given by

 
1 1 3
fI11 (θ) = cos( θ) 1 − sin( θ) sin( θ) , (3.2.6)
2 2 2

 
1 1 3
fI22 (θ) = cos( θ) 1 + sin( θ) sin( θ) , (3.2.7)
2 2 2

1 1 3
fI12 (θ) = fI21 (θ) = cos( θ) sin( θ) cos( θ), (3.2.8)
2 2 2

while the equivalent functions fII = (fII

ij )i,j=1,2 for mode II are

 
1 1 3
fII
11 (θ) = − sin( θ) 2 + cos( θ) cos θ) , (3.2.9)
2 2 2
28 The mechanics of crack propagation

1 1 3
fII
22 (θ) = cos( θ) sin( θ) cos( θ), (3.2.10)
2 2 2

 
1 1 3
fII
12 (θ) = fII
21 (θ) = cos( θ) 1 − sin( θ) sin θ) . (3.2.11)
2 2 2

As for KI and KII , they represent the SIFs for the modes I and II respectively and are
defined by

√
KI := lim K∗I (r), where K∗I (r) = 2πrσ22 (r, 0), (3.2.12)
r→0
√
KII := lim K∗II (r), where K∗II (r) = 2πrσ12 (r, 0). (3.2.13)
r→0

It is also possible to obtain equations for the corresponding√ displacement field near the
crack tip, cf. [20, 45, 66]. Again up to an error of order of r, this is given by

r r
KI r I KII r II
u(r, θ) = f̄ (θ) + f̄ (θ). (3.2.14)
G 2π G 2π

I II
The angular variation functions f̄ = (f̄Ii )i=1,2 and f̄ = (f̄II
i )i=1,2 in (3.2.14) are

 
1 1−ν 2 1
f̄I1 (θ) = cos( θ) + sin ( θ) , (3.2.15)
2 1+ν 2

 
1 2 1
f̄I2 (θ) = sin( θ) − cos2 ( θ) , (3.2.16)
2 1+ν 2

 
1 2 1
f̄II
1 (θ) = sin( θ) + cos2 ( θ) , (3.2.17)
2 1+ν 2

 
1 1−ν 2 1
f̄II
2 (θ) = cos( θ) − + sin ( θ) . (3.2.18)
2 1+ν 2

These formulae allow us to have a characterisation of the stresses and the displacements
in the vicinity of a crack tip.
3.3 Fracture criteria 29

3.3 Fracture criteria

Two different types of fracture criteria must be distinguished. Global criteria are based
on an energy balance and are related to the energy release strain and the J-integral. Local
criteria, on the other hand, make use of the key role of the SIFs in the stress state near
the tip of the crack, see for example [37, 90, 94].

We proceed to present global and local criteria for a stationary semi-infinite line crack,
loaded in a mode I situation. The symmetry of the deformations then implies that the
crack may only propagate in a direction perpendicular to the loading. All that is re-
quired is a condition for crack growth.

A possible criterion, so-called global criterion, states that the crack will propagate when
the energy stored is sufficient to break the material. This is expressed by the following
equation for the energy release rate

G = Gc . (3.3.1)

Here, Gc is the critical energy release rate. It is a material parameter that may be deter-
mined experimentally. Given that the first component of J equals the energy release
rate, (3.3.1) can also be expressed in terms of the J-integral.

It is also possible to establish a different criterion making use of the fact that the singular
stresses are characterised by KI in the region surrounding the tip of the crack. This local
criterion reads

KI = KIc . (3.3.2)

Here KIc , which behaves as a threshold value for KI , is called the critical stress intensity
factor or the mode I fracture toughness. We include in Table 3.1 examples of experimen-
tal data for the fracture toughness of some materials, as taken from [4].

The criteria (3.3.1) and (3.3.2) are actually equivalent, as the energy release rate and the
stress intensity factor are related by

K2I
G= . (3.3.3)
E

Naturally the critical energy release rate and the critical stress intensity factor are also
similarly related.

We now turn our attention to the more general case when the loading is a combination
of modes I and II, assuming that KI > 0. Now there is a fundamental difference to the
30 The mechanics of crack propagation

Material KIc (MN/m3/2 )

Mild steel 140
Titanium alloys 55 − 120
High carbon steel 30
Nickel, copper > 100
Concrete (steel reinforced) 10 − 15
Concrete (unreinforced) 0.2
Glasses, rocks 1
Ceramics (Alumina, SiC) 3−5
Nylon 3
Polyester 0.5

Table 3.1: Examples of fracture toughness.

mode I loading situation, where the direction of the crack growth is trivially determined.
In a mixed mode situation, criteria on whether the crack will propagate but also on
which direction it will do so must be decided upon.

A global criterion for propagation, expressed in terms of the J-integrals defined earlier, is
given by

kJk2 = Gc , (3.3.4)

which can be seen as a generalisation of (3.3.1) when written in terms of J1 . The direction
(J)
θp of the crack extension is the same as the direction of the vector J

J2
θ(J)
p = arctan ( ). (3.3.5)
J1

Moreover, a local criterion based on the circumferential tensile stress σθθ given by

σ11 + σ22 σ11 − σ22

σθθ = − cos(2θ) − σ12 sin(2θ), (3.3.6)
2 2

also known as the tangential stress, can be obtained by rewriting the stresses
√ in (3.2.5)
into local polar coordinates, cf. [15, 20]. We thus obtain, up to an order of r,
3.3 Fracture criteria 31

Kθθ (θ)
σθθ (r, θ) = √ , (3.3.7)
2πr

where

1 1 1
Kθθ (θ) = KI cos3 ( θ) − 3KII sin( θ) cos2 ( θ), (3.3.8)
2 2 2

is the circumferential stress intensity factor. By the maximum circumferential tensile stress
criterion, crack growth will occur when

max Kθθ (θ) = KIc , (3.3.9)

θ

which can be seen as a generalisation of (3.3.2). This criterion states that the direction of
(K)
propagation is given by the angle θp which maximises Kθθ (θ),

 q 
KI − K2I + 8K2II
θ(K)
p = 2 arctan  , (3.3.10)
4KII

cf. [30, 55]. This formula can be used to rewrite the condition for crack extension (3.3.9).
Indeed, maxθ Kθθ (θ) can be calculated by inserting the value of the angle given by
(3.3.10) in the circumferential tensile stress expressed in (3.3.8). We thus obtain

√
 q 
4 2K3II KI + 3 K2I + 8K2II
  32 = KIc . (3.3.11)
q
K2I + 12K2II − KI K2I + 8K2II

In a mixed mode situation the components of the J-integral are related to the stress
intensity factors

k+1 2 k+1
J1 = (KI + K2II ); J2 = − KI KII , (3.3.12)
8G 4G

where k = (3 − ν)/(1 + ν), ν is the material’s Poisson’s ratio and G its shear modulus,
cf. [37, 90]. These relations allow the comparison between the global fracture criterion
(3.3.4) - (3.3.5) and the local fracture criterion (3.3.10) - (3.3.11) for a mixed mode sit-
uation. When the fracture is dominated by mode I, i.e., when |KII /KI | is small, these
32 The mechanics of crack propagation

criteria are nearly equivalent, cf. [90]. This is illustrated in Figure 3.4, where the propa-
gation angles are displayed.

0
θ(J)
P
(K)
θP
−10
Propagation angle in degrees

−20

−30

−40

−50

−60
0 0.2 0.4 0.6 0.8 1
KII/KI

Figure 3.4: Crack propagation angles.

Henceforth we adopt the local criterion (3.3.10) - (3.3.11) and denote the propagation
(K)
angle θP simply by θP .

3.4 Numerical aspects

The fracture parameters which were introduced earlier are usually computed applying
numerical techniques. Once they have been determined, the fracture criteria allow the
prediction of the consecutive positions of the crack tip of a propagating crack.

3.4.1 Stress intensity factors

As we have seen the criteria for crack propagation can be given in terms of the SIFs. It
is then important to be able to approximate these accurately. There are classical exam-
ples of cracked geometries for which the stress intensity factors have been computed or
approximated explicitly, cf. [83].
3.4 Numerical aspects 33

s s

2a a

Figure 3.5 - a) Infinite plate with a centre through crack b) Semi-Infinite plate with a centre through crack
under tension. under tension.

s s

2b b

2a a

c) Infinite stripe with a centre through crack under tension. d) Infinite stripe with an edge through crack under tension.

As an example we include the values of the SIFs for the various geometries represented
in Figure 3.5, each subject to a uniform tensile stress σ. These values are as follows,
where the letters a) - d) used to identify the formulae correspond to those of the figure.

√
a) KI = σ πa;

√
b) KI = 1.1215σ πa;

√
c) KI = σ πa 1 − 0.025ρ2 + 0.06ρ4 sec πρ
p

2 ;
34 The mechanics of crack propagation

√ q 2 πρ 0.752+2.02ρ+0.37(1−sin(πρ/2))3

d) KI = σ πa πρ tan 2 cos(πρ/2) .

Here, ρ = a/b. The previous examples involve geometries of infinite dimensions. In

this section, we will instead focus on a finite geometry. Consider a square plate, of side
length 2b, with a central crack of length 2a, see Figure 3.6. Let the plate be loaded from
its upper and lower edges again by a uniform tensile stress σ.
s

2b

2b

2a

Figure 3.6: Finite plate with a centre through crack under tension.

For this cracked plate, the SIF has been estimated in [1] to be given by

√
KI = σ πa(1 + 0.043ρ + 0.491ρ2 + 7.125ρ3 − 28.403ρ4
+59.583ρ5 − 65.278ρ6 + 29.762ρ7 ). (3.4.1)

From this formula we compute reference solutions for the stress √ intensity factors for
several values of ρ = a/b. The values of KI /K0 , where K0 := σ πa, are displayed in the
second line of Table 3.2. Note that KI /K0 does not depend on the loading applied to the
plate, but only on the geometrical ratio ρ. In what follows we discuss several methods to
compute the SIFs numerically, namely the stress correlation method, the displacement
correlation method and the J-integral method, cf. for example [95].

Earlier we presented formulae that allow us to characterise the stress field in the vicinity
of a crack tip. In turn, if the stress field is known, this allows determining the SIFs using
(3.2.12) and (3.2.13). This is the key idea of the stress correlation method. The technique
is applied as follows: one simply finds the functions K∗I (r) and K∗II (r) for values of the
stress computed ahead of the crack tip, using finite elements for instance, and extrapo-
lates back to r = 0 to find approximations for KI and KII respectively. This method is
quite simple, but its accuracy depends on the mesh refinement and also on how well
3.4 Numerical aspects 35

the solution of the elasticity problem captures the crack singularity.

An alternative technique is given by the displacement correlation method. It makes use of

the equations for the displacement field near the crack tip, from which it can be shown
that

r
E 2π
KI = lim K∗∗
I (r), where K∗∗
I (r)= u2 (r, π) , (3.4.2)
r→0 4 r
r
∗∗ ∗∗ E 2π
KII = lim KII (r), where KII (r) = u1 (r, π) . (3.4.3)
r→0 4 r

The determination of the SIFs then follows a similar procedure to the one described for
the stress correlation method.

Finally, the J-integral method makes use of the J-integral. This can be computed along a
path surrounding the crack tip by means of (3.2.3). In turn the SIFs can be determined
using (3.3.12). The main advantage of the J-integral method is that it is very accurate in
general without requiring the usage of very fine meshes. This is because capturing the
crack tip singular stress field is no longer necessary. It suffices to be able to find a good
approximation for the stress and strain fields over the curve of the J-integral. It is in that
sense a global approach.

To illustrate the usage of these techniques we consider once again the problem of the
plate illustrated in Figure 3.6. For each value of ρ we start by finding an approximation
for the SIF by applying the displacement correlation method. We first solve the elasticity
problem by a finite element method. A mesh with quadratic triangular elements with
maximum width of the element side h = 0.01 is generated. Next, we approximate the
function K∗∗I by a linear polynomial within an interval contained in [0, a]. The choice of
the interval is not arbitrary. Indeed, since the limit when r → 0 is sought, we want to
consider the function in a domain where the values of r are small. On the other hand
the crack tip singularity affects the accuracy of the results near r = 0. This means that
the values of K∗∗
I when r is too small must be disregarded.

To illustrate this procedure, we take ρ = 0.4. Without loss of generality, let b = 1. The
values of K∗∗
I /K0 can be plotted versus the values of r, see the full line in Figure 3.7.

We use a linear interpolator, represented by a dashed line, to fit the function in the
interval [0.03, 0.2]. The data over the interval [0, 0.03] are neglected due to the crack tip
singularity and the consequent non-linear behaviour of K∗∗ I . To proceed, we extrapolate
the value of the interpolator back to r = 0 and obtain the desired approximation. The
SIFs computed using this method yield the results displayed in the third line of Table
3.2 for the various values of ρ contained in the first line of the table.

The stress correlation method works similarly. The results are shown in the fourth line
of Table 3.2. They are not as accurate as those obtained using the displacement cor-
36 The mechanics of crack propagation

1.3

1.225

1.1 KI∗∗ /K0

0.9

0.7

0.5
0 0.05 0.1 0.15 0.2
r

Figure 3.7: Extrapolation using the displacement correlation method, ρ = 0.4.

SIF/K0 \ ρ 0.1 0.2 0.4 0.6

KI /K0 1.014 1.055 1.216 1.481
KI (u)/K0 1.004 1.058 1.225 1.525
KI (σ)/K0 1.226 1.145 1.211 1.802
KI (J)/K0 1.004 1.051 1.212 1.477

Table 3.2: Stress intensity factors.

relation method. This is to be expected, because now we are using the stresses to ap-
proximate the SIFs, whereas the primary result of the finite element computation is the
displacement field.

It should also be pointed out that the two previous methods require a choice to be made
for the interval where the linear interpolator is to be applied. This is not completely ob-
jective and requires intuition, which does affect the results. The inherent subjectivity of
the choice of an interval is no longer a problem for the method employing the J-integral,
as this fracture parameter is path-independent. In this sense this is a much more reliable
method than the previous two as it allows for quite accurate approximations of the SIFs.
The results obtained using this method can be seen in the last line of Table 3.2. They are
the average of a number of J-integrals, each computed over a small circle centred at the
crack tip.
3.4 Numerical aspects 37

3.4.2 Simulation of crack growth

Earlier a static fracture analysis was performed, where the goal was merely to compute
the stress intensity factors. With the fracture criteria introduced in section 3.3, we can
now look into the actual propagation of the crack.

x2

a 2b
x1

Figure 3.8: Finite plate with a crack subject to mixed mode loading.

Consider the cracked rectangular linear elastic plate of width b and height 2b, repre-
sented in Figure 3.8. Let the ratio between the length of the crack measured along the
x1 -axis and the width of the plate be such that a/b = 1/5. In order to consider a mixed
mode situation the inclined initial crack is considered to be at a α = 45◦ angle with the
horizontal axis. Without loss of generality, we take b = 1.

If the uniform tensile stress σ, which is applied at the lower and upper horizontal
boundaries of the plate, is large enough so that the fracture criterion (3.3.11) holds, we
do know that the initial crack will propagate. We assume it will do so in the direction
given by (3.3.10).

We are now ready to set up a procedure for the prediction of the crack path. Denote
(0) (0) (0)
the initial crack tip coordinates by xtip = (xtip , ytip ) and the initial crack angle by θ(0) .
Following an incremental approach to keep track of the position of the crack tip as the
crack grows as suggested in [37, 55, 67], the crack path is determined step by step. At
each step the direction of propagation is determined and the crack is incremented with
a line segment of length ∆a. This procedure is structured as follows.
38 The mechanics of crack propagation

Algorithm 3.1

(0)
Set n = 0. Given: ∆a, xtip , θ(0) .

1- Solve the elasticity problem.

2- Compute the values of the stress intensity factors, KI and KII .

(K)
3- Determine the angle of propagation θp = θp using equation (3.3.10).

4- Update the crack angle and the crack tip coordinates

θ(n+1) := θ(n) + θP , (3.4.4)

(n+1) (n)
xtip = xtip + ∆a(cos(θ(n+1) ), sin(θ(n+1) )). (3.4.5)

5- Increment n, n → n + 1.

6- If required to determine the path further, repeat steps 1 to 6.

This algorithm can be employed to determine the set of consecutive crack tip coordi-
(n) (n) (n)
nates xtip = (xtip , ytip ) for the propagating crack of the plate illustrated in Figure 3.8,
with ∆a = 0.025. We use the numerical package Abaqus to solve the elasticity problem
in the first step of the algorithm. Quadratic rectangular elements are employed every-
where except near the crack tip, where quadratic triangular elements are used. The SIFs
are determined using the J-integral method. The results obtained using the aforemen-
tioned algorithm are displayed in Table 3.3.

(n) (n) (n) (n) (n) (n)

n xtip ytip n xtip ytip n xtip ytip
0 0.100 0.100 5 0.225 0.101 10 0.350 0.101
1 0.125 0.102 6 0.250 0.101 11 0.375 0.101
2 0.150 0.101 7 0.225 0.101 12 0.400 0.102
3 0.175 0.101 8 0.300 0.101 13 0.425 0.102
4 0.200 0.101 9 0.325 0.101 14 0.475 0.102

Table 3.3: Crack tip positions.

The discretisation elements of the crack tip region are illustrated in Figure 3.9, where
the crack is represented by the darker region.
3.4 Numerical aspects 39

Figure 3.9: Crack tip region.

The deformed state of the plate before propagation and the deformed plate geometry
obtained after 15 steps of the algorithm are displayed in Figure 3.10 a) and b), respec-
tively.

Figure 3.10 - a) Deformed initial plate. b) Deformed plate for n = 15.

Finally, the crack path can be seen in Figure 3.11. In a) we plot the initial crack segment
 40 The mechanics of crack propagation

and in b) the crack path as it was determined after 15 steps.

0.1 0.1
y1

1
y
0 0
0 0.25 0.5 0 0.25 0.5
x1 x1

Figure 3.11 - a) Initial crack position. b) Crack path for n = 15.

There appears to be a strong tendency for the crack to propagate mainly in mode I,
during continued fracture. This agrees with the predictions of [20, 37]. We also observe
that this simulation of crack growth is based on the static stress intensity factors. Dy-
namic effects such as wave propagation are not taken into account. A dynamic analysis
is presented in [90].
Chapter 4

Homogenisation for periodic

structures

Earlier the elliptic PDEs that model the behaviour of composite linear elastic materials
were introduced. When these materials are very heterogeneous the coefficients of the
differential equations become highly oscillatory. Classical numerical techniques such as
finite elements or boundary elements are then computationally too expensive to obtain
accurate approximations for the solutions of the PDEs. An alternative is provided by
the theory of homogenisation when the coefficients are periodic. This converts the orig-
inal problems into simpler homogenised problems, which provide good macroscopic
approximations for the solutions of the heterogeneous problems.

In this chapter the basic ideas of homogenisation are introduced. Analytical expressions
for the coefficients of the homogenised problem are obtained for layered materials and
some properties of the homogenised materials are analysed. An efficient algorithm for
the numerical determination of the effective coefficients is proposed.

4.1 One-dimensional model

Let us consider the one-dimensional Dirichlet boundary value problem with periodi-
cally oscillating coefficients

x
a (x) := ã( ), (4.1.1)


which is given by
42 Homogenisation for periodic structures

 d
  d 


 − dx a (x) dx u (x) = f(x), x ∈ (0, 1),

u (0) = 0, (4.1.2)

 
u (1) = 0.

Here, a is an -periodic function and   1. We assume that there exist two real
constants α and β such that

0 < α < ã(x/) < β.

The analysis of this problem will give us insight into many of the relevant issues present
in the sophisticated problems dealt with by the theory of homogenisation. To illustrate
this approach let

1
a (x) := . (4.1.3)
2 + sin( 2πx
 )

We observe that a is an oscillating function, see Figure 4.1. For smaller values of , it
oscillates more rapidly.

1.1

0.9

0.8

0.7

0.6

0.5

0.4

0 0.2 0.4 0.6 0.8 1

x

Figure 4.1: Oscillating behaviour of a ,  = 0.1.

The exact solution u of (4.1.2) with f(x) = 1, represented in Figure 4.2 taking a of the
form (4.1.3), is given by

Zx  R1
[ã(y/)]−1 y dy

−y + c0
u (x) = dy, c0 = R01 . (4.1.4)
0 ã(y/) [ã(y/)]−1 dy
0

In general it is hard and often even impossible to find the analytical solution for prob-
lems with oscillating coefficients like (4.1.2). We must then resort to numerical methods,
such as finite element methods.
4.1 One-dimensional model 43

Let {xi }i=0, ..., N+1 , with xi = i/(N + 1), be a discretisation of the interval [0, 1] and
denote by h the distance between two adjacent points. We represent by V0h the subspace
of H1 (0, 1) given by the functions that are linear in every interval [xi−1 , xi ], for i =
1, . . . , N + 1, and assume that f ∈ L2 (0, 1). Then the finite element approximation
uh ∈ V 1 (0, 1) to u ∈ H10 (0, 1) satisfies

Z1 Z1
duh dvh
ã(x/) (x) (x)dx = f(x)vh (x)dx, ∀vh ∈ V0h , (4.1.5)
0 dx dx 0

and its error is estimated by

β2 h
ku − uh kH1 (0,1) ≤ c kfkL2 (0,1) , (4.1.6)
α3 

with c ∈ IR, cf. [60]. For a fixed value of , it is clear that uh converges to u when
h → 0. However, the convergence of this method is not uniform in , see [76]. This
implies that when finite elements are used the grid size h must be taken sufficiently
smaller than , cf. [39].

0.25 exact
h=0.05
h=0.01

0.2

0.15

0.1

0.05

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

Figure 4.2: Solutions for the heterogeneous problem.

This behaviour is illustrated in Figure 4.2, again taking a of the form (4.1.3). Here we
show numerical approximations of (4.1.2) with f(x) = 1, for some values of h. Note that
an accurate approximation for the exact solution is only obtained for h = 0.01 <  = 0.1.

Observe that taking h <  to obtain an accurate approximation is computationally very

expensive when  takes small values, particularly for problems in higher dimensions.
44 Homogenisation for periodic structures

In that case, even with heavy computer efforts it is impossible to obtain the desired
accuracy.

4.1.1 Multiple scales method

One alternative approach to obtain an accurate numerical approximation for (4.1.2) in a

cheap manner, using a mesh size h > , is given by the homogenisation theory. It allows
for the formulation of a second problem, still of the form of (4.1.2) but with constant co-
efficients, called homogenised or effective coefficients, so we are rid of the oscillations. The
homogenised or effective solution u of the new problem should capture the macroscopic
behaviour of u , but not the oscillations on the microscopic scale.

In order to look for the homogenised approximation for the solution of (4.1.2), we use
the multiple scale method. Following [2], consider the ansatz

u (x, y) = u0 (x, y) + u1 (x, y) + 2 u2 (x, y) + . . . , (4.1.7)

where we assume the functions ui to be 1-periodic in y and sufficiently smooth. They

depend explicitly on x and y := x/. This is a mathematical translation of the idea that
u depends on two length scales, macroscopic and microscopic. We will treat x and
y as independent variables. By restricting ourselves to the first three terms of the last
expansion and substituting the ansatz in the differential equation in (4.1.2) we obtain

1 1 1
A0 u0 + (A1 u0 + A0 u1 ) + 0 (A2 u0 + A1 u1 + A0 u2 ) + . . . = f. (4.1.8)
2  

Here we define the operators

∂ ∂
A0 := − (ã(y)) , (4.1.9)
∂y ∂y
∂ ∂ ∂ ∂
A1 := − (ã(y)) − (ã(y)) , (4.1.10)
∂y ∂x ∂x ∂y
∂ ∂
A2 := − (ã(y)) . (4.1.11)
∂x ∂x

Equating coefficients of equal power in  in (4.1.8), we obtain the equalities

A0 u0 = 0, (4.1.12)
A0 u1 = −A1 u0 , (4.1.13)
4.1 One-dimensional model 45

A0 u2 = −A1 u1 − A2 u0 + f, (4.1.14)

The existence of ui , i = 0, 1, 2, up to constants in y is ensured by the Fredholm alterna-

tive, cf. [71]. Now, it can be shown that the solution of (4.1.12) actually does not depend
on y. Indeed, multiplying A0 u0 by u0 and integrating over [0, 1] yields

Z1
0 = u0 A0 u0 dy
0
Z1 Z1
∂ ∂u0 ∂u0 2
= − u0 (ã(y) )dy = − ã(y)( ) dy, (4.1.15)
0 ∂y ∂y 0 ∂y

where the last equality was obtained using integration by parts. Since ã(y) > α, then
∂u0 /∂y must be zero. We conclude that the first term of the ansatz does not actually
depend on the microscopic variable. It is the homogenised solution we are looking for.
We proceed to finding a way to compute u0 , which we now denote as u := u0 . To do so,
we start by using separation of variables and seeking a solution of (4.1.13) of the form

∂u
u1 (x, y) = χ(y) . (4.1.16)
∂x

Inserting (4.1.16) in (4.1.13), we obtain the cell problem



 − d (e dχ
a(y) dy ) = − de
a(y)
dy , y ∈ (0, 1),
 dy
(4.1.17)

 χ is periodic,
 R1
0
χ(y)dy = 0,

where the last condition is added to ensure the problem is well-posed. Finally we use
(4.1.14) to look for the homogenised solution. Because the function u2 is 1-periodic, it
R1
can be shown by using the fundamental theorem of calculus that 0 A0 u2 dy = 0, and
so from (4.1.14) we have

Z1
−A1 u1 − A2 u0 + fdy = 0. (4.1.18)
0

Using the fact that the function u depends only on x and that u1 is given by (4.1.16), the
previous equation yields the homogenised problem
46 Homogenisation for periodic structures

 d du

 −a dx ( dx ) = f(x), x ∈ (0, 1),
(4.1.19)


u(0) = 0,
u(1) = 0,

where the homogenised coefficient a is given by

Z1 Z1
dχ
a= a
e(y)dy − a
e(y) (y)dy. (4.1.20)
0 0 dy

In summary, we obtain the homogenised problem (4.1.19) which corresponds to (4.1.2).

This allows us to find an approximation u for the solution u of the original problem
without having to solve the microscale. Instead we are left to deal with the simpler
homogenised problem, once the effective coefficient a in (4.1.20) has been determined.

We may now relate u and u and estimate the approximation error, cf. [21, 60]. The
following theorem justifies the validity of approximating the solution of a problem with
highly oscillating periodic coefficients by the solution of the corresponding homoge-
neous problem, cf. [82].

Theorem 4.1.1 Let a e=a e(x) be a function in L∞ (0, l1 ) which we extend to IR such that it is
l1 -periodic, and f ∈ L (d1 , d2 ) a given function. Define a (x) := ã(x/). Assume that there
2

exist two real constants α and β such that

0 < α < ã(y) < β.

Moreover, let u ∈ H10 (d1 , d2 ) be the unique solution of the problem

 d
  d 


 − dx a (x) dx u (x) = f(x), x ∈ (d1 , d2 ),

u (d ) = 0, (4.1.21)

  1
u (d2 ) = 0,

and u ∈ H10 (d1 , d2 ) be the unique solution of the homogenised problem

 d2

 −a dx2 u(x) = f(x), x ∈ (d1 , d2 ),
(4.1.22)

 u(d1 ) = 0,
u(d2 ) = 0,
4.1 One-dimensional model 47

where the effective coefficient a is given by

1
a = R l1 . (4.1.23)
0
a(y)]−1 dy
[e

Then the solution of problem (4.1.21) converges to the solution of problem (4.1.22) in H10 (d1 , d2 ),
i.e.

u → u weakly in H10 (d1 , d2 ). (4.1.24)

Finally, let a ∈ W 1,p (IR), for some p > 2. Under the conditions of the previous theorem
it can be shown that there exists a constant c ∈ IR such that

ku − ukH1 (d1 ,d2 ) ≤ ckukH2 (d1 ,d2 ) . (4.1.25)

4.1.2 Recovering the heterogeneities

The homogenised solution u captures the macroscopic behaviour of u and disregards the
oscillations. We would like to be able to recover these without having to employ finite
elements with a very fine mesh. In order to do so we consider one more term in the
asymptotic expansion (4.1.7), the first order corrector u1 .

Though we might expect u + u1 to be a better approximation than u, it does not neces-
sarily verify the same boundary conditions as the heterogeneous solution u . A function
C which we call the boundary corrector should then be introduced so that u+u1 +C is the
sought approximation that recovers the oscillations and at the boundaries u+u1 +C =
u . Let C satisfy the original heterogeneous equation with source term equal to zero

 d
  d


 − dx a (x) dx C(x) = 0, x ∈ (0, 1),
(4.1.26)


C(0) = −u1 (0),
C(1) = −u1 (1).

In practice C may be approximated by a function that satisfies (4.1.26) with a replaced

by a. It can be shown that there exists a constant d such that we have

ku − (u + u1 + C)k0 ≤ d2 kuk3 , (4.1.27)

and
ku − (u + u1 + C)k1 ≤ dkuk2 , (4.1.28)
48 Homogenisation for periodic structures

cf. [88]. The heterogeneities neglected by u can be recovered up to a point by the cor-
rected homogenised solution. This is an advantageous process as it allows for a cheap so-
lution to be obtained for a problem that would otherwise require a large computational
effort.

4.1.3 Numerical example

Let us consider the boundary value problem (4.1.2) with (4.1.3) and f(x) = 1, as well as
the corresponding homogenised problem (4.1.19). The solutions of these problems are
given by (4.1.4) and

  Z1
1 2 1
u(x) = − (x − x) dy, (4.1.29)
2 0 a
e(y)

respectively. We can then determine the error E = u − u, which is displayed in the

second line of the Table 4.1, in the norms k · k∞ and k · kL2 (0,1) , for  = 0.01 and  = 0.1.
We obtain better approximations when the period of the heterogeneities is small. This
was to be expected, as the homogenisation procedure works better for materials that are
finely mixed.

Better approximations for u can be achieved using the first order corrector u1 given by
(4.1.16) and the boundary corrector C which solves (4.1.26). The latter can be approxi-
mated by the solution of the homogenised problem

 d
d 

 −a dx dx C(x) = 0, x ∈ (0, 1),
(4.1.30)


C(0) = −u1 (0),
C(1) = −u1 (1),

where a is given by (4.1.20). We display the norms of the error EC = u − (u + u1 + C)

in the last line of Table 4.1. Note that in this case it was possible to determine both u1
and C analytically.

 = 0.01  = 0.1
k · k∞ k · kL2 (0,1) k · k∞ k · kL2 (0,1)
E 1.6E − 3 5.6E − 4 1.5E − 2 5.6E − 3
EC 4.1E − 6 2.4E − 6 4.1E − 4 2.4E − 4

Table 4.1: Error of the approximations.

4.2 Homogenisation for elasticity 49

As expected, the corrected homogenised solution is a better approximation for u than

u. This is better illustrated in Figure 4.3. The exact solution u is depicted as dots. The
homogenised solution u and its corrected counterpart u + u1 + C are represented as
the darker and lighter solid lines, respectively.

0.25 Exact solution

Homogenised
Homogenised
with correction
0.2

0.15

0.1

0.05

0
0 0.2 0.4 0.6 0.8 1
x

Figure 4.3: Solution of the heterogeneous problem and corresponding homogenised

approximations,  = 0.1.

4.2 Homogenisation for elasticity

In this section we briefly analyse the asymptotic behaviour of the solution of elasticity
problems associated to periodic composite materials using the theory of homogenisation,
cf. [21, 71]. These results are presented for a 2D plane stress situation, though they can
be extended to higher dimensions in a straightforward way.

Consider a composite plate with constituents periodically distributed over Ω ⊆ IR2 . We

can then define a reference cell Y = [0, l1 ] × [0, l2 ] such that Ω is covered by a mosaic of
cells of the form Y = [0, l1 ] × [0, l2 ], see Figure 4.4.
50 Homogenisation for periodic structures

l2
el2

el1 Y
l1
W

Figure 4.4: Domain Ω and reference cell Y.

The behaviour of the material is characterised by the fourth-order elasticity tensor à =

aijkh )1≤i,j,k,h≤2 defined over the reference cell Y. Assume that there exist α,
Ã(y) = (e
β ∈ IR such that à ∈ Me (α, β, Y), i.e.,

eijkh ∈ L∞ (Y), for any i, j, k, h ∈ {1, 2};

• a
• a
eijkh = a ekhij , for any i, j, k, h ∈ {1, 2};
ejikh = a
• αkMk2 ≤ ÃMM, for any symmetric 2 × 2 matrix M;
• kÃ(y)Mk ≤ βkMk, for any 2 × 2 matrix M, almost everywhere on Y,

where for any fourth-order tensor B = (bijkh )1≤i,j,k,h≤2 and for any matrices M =
(mij )1≤i,j≤2 and N = (nij )1≤i,j≤2 , we denote

X
2
kMk := ( m2ij )1/2 ,
i,j=1
X
BM := (( bijkh mkh )ij )1≤i,j≤2 ,
k,h=1,2
X
BMN := bijkh mij nkh .
i,j,k,h=1,2

The functions a eijkh are extended to IR2 periodically, which allows us to define an -
periodic tensor A = A (x) = (aijkh )1≤i,j,k,h≤N such that for x = (x1 , x2 ) ∈ IR2 ,

x
aijkh (x) := a
eijkh (y) = a
eijkh ( ), (4.2.1)


where we denote y := x/, for y = (y1 , y2 ) ∈ IR2 . It can be shown that if à ∈

Me (α, β, Y), then also A ∈ Me (α, β, Y).
4.2 Homogenisation for elasticity 51

We now formulate the linear elasticity problem for the periodically distributed compos-
ite material, that is stated as



 −∇ · (A (x)(u )) = f, x ∈ Ω,


u = 0, x ∈ ΓD , (4.2.2)


 
 σ(u ) · n = ϕN , x ∈ ΓN ,

where the functions ϕN and f are given and , σ and u denote the strain tensor, the
stress tensor and the displacement vector respectively. We recall from Chapter 2 that
these are such that for every function w,

1
(w) = (∇w + (∇w)T ), σ(w) = A (w). (4.2.3)
2

We look for weak solutions of (4.2.2) in the space V defined by

V = {v | v ∈ (H1 (Ω))2 , γ(v) = 0 on ΓD }. (4.2.4)

Moreover, we define the following norm

!1/2
X
2
kvkV = k∇vi k2L2 (Ω) , (4.2.5)
i=1

for v = (v1 , v2 ) ∈ V, cf. [21].

4.2.1 Main result

Our objective is to determine the homogenised version for the problem (4.2.2), for which
it is useful to introduce some auxiliary functions. Consider the family of vector-valued
functions Plm (y) = (Pklm (y))k=1,2 defined by

Pklm (y) := ym δkl , k = 1, 2, (4.2.6)

where δkl is the Kronecker symbol and l, m ∈ {1, 2}. Let χlm (y) = (χlm
k (y))k=1,2 ∈
(Wper (Y))2 be the unique solution of the cell problem
52 Homogenisation for periodic structures


 −∇ · (Ã(y)∇(χlm − Plm )) = 0, y ∈ Y,
χlm is Y-periodic , (4.2.7)

MY (χlm
k ) = 0,

where

Z
1
MY (f) := f(y)dy and Wper (Y) := {v ∈ H1per (Y)|MY (v) = 0}. (4.2.8)
|Y| Y

Here, H1per (Y) is the closure for the H1 -norm of C∞ ∞ 2

per (Y), the subset of C (IR ) of peri-
odic functions over Y. In these conditions, it can be shown that the solution of (4.2.7) is
unique.

We now aim at characterizing the homogeneous solution of the elasticity problem. The
following result can be thought of as a generalisation of the theorem presented earlier
on section 4.1 to the context of planar elasticity, cf. [21] for a proof.

Theorem 4.2.1 Let us consider the heterogeneous elasticity problem



 −∇ · (A (x)(u )) = f, x ∈ Ω,


u = 0, x ∈ ΓD , (4.2.9)


 
 σ(u ) · n = ϕN , x ∈ ΓN ,

where f ∈ V 0 , ϕN ∈ (H−1/2 (ΓN ))2 and u ∈ V is its unique solution. The corresponding
homogeneous problem is given by



 −∇ · (A(u)) = f, x ∈ Ω,


u = 0, x ∈ ΓD , (4.2.10)


 u · n = ϕN ,
 x ∈ ΓN ,

and it admits a unique solution u ∈ V. In these conditions,

u → u weakly in V, (4.2.11)
A (u ) → A(u) weakly in (L2 (Ω))2×2 . (4.2.12)

Moreover, the homogenised tensor A = (aijkh ) is symmetric and

4.2 Homogenisation for elasticity 53

Z Z X
1 1 ∂χkh
l
aijkh = a
eijkh (y)dy − a
eijlm (y) (y)dy. (4.2.13)
|Y| Y |Y| Y l,m=1,2 ∂ym

This theorem is the fundamental result which allows us to analyse the behaviour of lay-
ered materials in the next subsection. In practical terms, one may avoid computing the
solution of (4.2.9) directly. Instead we solve the cell problem (4.2.7) to find the effective
coefficients aijkh and finally compute the approximation u for u by means of (4.2.10).

As in the one dimensional case, correctors may be introduced to recover the hetero-
geneities that are disregarded by the homogenised solution. That will be done in Chap-
ter 6.

4.2.2 Layered composites

We will now focus on layered periodic composite materials, which are such that their
properties vary only along one spatial coordinate. We show how the effective coeffi-
cients can be determined analytically and use this result to characterise the homogenised
material. Let us again consider the elasticity problem (4.2.2) where now the tensor com-
ponents read

x x
aijkh (x) = a eijkh ( 1 ) = aijkh (x1 ).
eijkh ( ) = a (4.2.14)
 

For the sake of simplicity and without loss of generality, we further assume that the
reference cell is such that |Y| = 1. Then under the conditions of theorem 4.2.1, we make
use of (4.2.14) to solve the cell problem analytically and find explicit expressions for the
effective coefficients.

Theorem 4.2.2 Consider the problem (4.2.2) with (4.2.14), where the components of the tensor
A are such that

det(y1 ) := a
e1111 (y1 )e e22111 (y1 ) 6= 0, ∀y1 ∈ [0, l1 ].
a2121 (y1 ) − a (4.2.15)

Furthermore, denote

a
e2111 2 a
e1111 a
e2121
Det =: [MY ( )] − MY ( )MY ( ), (4.2.16)
det det det
a
eij11 a
e2121 a
eij21 a
e1121
Aij := −MY ( ) + MY ( ), (4.2.17)
det det
54 Homogenisation for periodic structures

a
eij11 a
e1121 a
eij21 a
e1111
Bij := MY ( ) − MY ( ). (4.2.18)
det det

Then the effective solution u for this problem satisfies



 −∇ · (A(u)) = f, x ∈ Ω,


u = 0, x ∈ ΓD , (4.2.19)



 σ(u) · n = ϕN , x ∈ ΓN ,

where the components of A are symmetric and given by

Aij a
e11lm Bij a
e21lm
aijlm = MY ( )+ MY ( ), (4.2.20)
Det det Det det

for all i, j, l, m ∈ {1, 2} with i + j ≥ l + m, except for i = j = l = m = 2. In the later case,

a2222 = MY (e
a2222 )
2e
a2211 a
e2221 a e22211 a
e2111 − a e22221 a
e2121 − a e1111
+MY ( )
det
A22 a
e1111 B22 a
e1121
−[ MY ( )+ MY ( )]A22
Det det Det det
B22 a
e2121 A22 a
e2111
−[ MY ( )+ MY ( )]B22 . (4.2.21)
Det det Det det

Proof:

Since a eijkh (y1 ), the homogenised coefficients given in (4.2.13) read

eijkh (y) = a

Z Z X
1 1 ∂χkh
l
aijkh = a
eijkh (y1 )dy1 − a
eijlm (y1 ) (y1 )dy1 . (4.2.22)
|Y| Y |Y| Y l,m=1,2 ∂ym

The family of functions ∂χkh

l /∂ym is determined by integrating the first equation in
(4.2.7), which yields

∂χlm ∂χlm
a
e1111 1
+a a11lm + clm
e1121 2 = −e 1 , (4.2.23)
∂y1 ∂y1
4.2 Homogenisation for elasticity 55

∂χlm ∂χlm
a
e2111 1
+a a21lm + clm
e2121 2 = −e 2 . (4.2.24)
∂y1 ∂y1

The integration constants clm lm

k can be determined using the fact that the functions χk
are Y-periodic, for k = 1, 2. The result then follows from (4.2.22).

This theorem allows us to compute the elasticity tensor A explicitly for all media that are
such that there exists a direction along which its properties are constant. In particular
we can apply the previous result to a linear elastic isotropic non-homogeneous material
with Poisson’s ratio ν and with Young’s modulus E. We recall from Chapter 2 that the
symmetrical coefficients aeijkh of the tensor A
e that describe the material behaviour are

E Eν
a
e2222 = a
e1111 = ; a
e2211 = ; (4.2.25)
1 − ν2 1 − ν2
E
a
e2121 = ; a
e2111 = a
e2221 = 0. (4.2.26)
2(1 + ν)

We examine the situation when the material parameters are of the form

ν = ν(y1 ) and E = E(y1 ). (4.2.27)

We can then establish the following property.

Property 4.2.3 The homogeneous problem related to (4.2.2) with (4.2.25)-(4.2.27) is



 −∇ · (A(u)) = f, x ∈ Ω,


u = 0, x ∈ ΓD , (4.2.28)



 σ(u) · n = ϕN , x ∈ ΓN ,

where the effective coefficients aijlm are symmetric and given by

1
a1111 = ; a2111 = a2221 = 0; (4.2.29)
MY (1/ea1111 )
MY (e
a2211 /e
a1111 ) 1
a2211 = ; a2121 = ; (4.2.30)
MY (1/ea1111 ) MY (1/e
a2121 )
56 Homogenisation for periodic structures

2
e22211
a [MY (e
a2211 /e
a1111 )]
a2222 = MY (e
a2222 ) − MY ( )+ . (4.2.31)
a
e1111 MY (1/e a1111 )

Proof: It is a consequence of Theorem 4.2.2.

It is common to think of situations where the Poisson ratio is constant throughout the
material, and only the Young’s modulus varies as a function of x1 . We will show that
in this situation the homogenised material is orthotropic. As was seen in Chapter 2 this
means that the coefficients of the effective elasticity tensor A are of the form

Ex Ex νyx
a1111 = ; a2211 = , (4.2.32)
1 − νxy νyx 1 − νxy νyx
Ey
a2222 = ; a2121 = Gxy , (4.2.33)
1 − νxy νyx
a2111 = a2221 = 0, (4.2.34)

where νxy Ey = νyx Ex . Before we present this result in the form of a property, we state
the following lemma.

Lemma 4.2.4 In the conditions of Theorem 4.2.2 assume that there exist constants cijlm and a
function E = E(y1 ) such that the coefficients a
eijlm read

a
eijlm = E(y1 )cijlm . (4.2.35)

Then the effective coefficients are


1
aijlm = c11lm (−cij11 c2121 + cij21 c2111 )
[c22111 − c1111 c2121 ]MY (1/E)

+c21lm (cij11 c2111 − cij21 c1111 ) , (4.2.36)

for all indices i, j, l, m ∈ {1, 2} not simultaneously 2 and i + j ≥ l + m. In the later case,
4.2 Homogenisation for elasticity 57


a2222 = MY (E) c2222 +

2c2211 c2221 c2111 − c22211 c2121 − c22221 c1111


−
c22111 − c1111 c2121
2c2211 c22111 (−c2211 c2121 + c2221 c2111 )

+
(c22111 − c1111 c2121 )2
c2221 c1111 c2111 [c2211 c2121 − c2221 c1111 ]
+
(c22111 − c1111 c2121 )2
c1111 (−c2211 c2121 + c2221 c2111 )2
+
(c22111 − c1111 c2121 )2
c2121 (c2211 c2111 − c2221 c1111 )2

1
. (4.2.37)
(c22111 − c1111 c2121 )2 MY (1/E)

Proof: It is a consequence of Theorem 4.2.2.

We are now ready to show that a finely mixed linear elastic material with constant Pois-
son’s ratio ν and Young’s modulus E = E(y1 ) can be replaced effectively by an or-
thotropic material, in the sense of homogenisation.

Property 4.2.5 Let us consider the problem (4.2.2) with (4.2.25)-(4.2.26), where ν is a constant
function and E = E(y1 ).

The effective coefficients are such that the homogenised material is orthotropic. The correspond-
ing material constants read

AB
Ex = , Ey = A, (4.2.38)
A + ν2 (B − A)
Bν
νxy = 2
, νyx = ν, (4.2.39)
A + ν (B − A)

B
Gxy = , (4.2.40)
2(1 + ν)

where

1
A = MY (E), B = .
MY (1/E)
58 Homogenisation for periodic structures

Proof: It is a consequence of Lemma 4.2.4.

In the conditions of the previous property, the upper and lower bounds for the material
constants given in (4.2.38)-(4.2.39) can be determined in terms of the Young’s modulus
and Poisson’s ratio of the original heterogeneous material.

Property 4.2.6 The material constants of the homogenised material satisfy

1
≤ Ex ≤ MY (1/E), (4.2.41)
MY (1/E)
1
ν ≤ νxy ≤ ν. (4.2.42)
MY (E)MY (1/E)

Proof: Follows from using the Cauchy-Schwarz inequality.

Finally we can think of a mixture of two isotropic homogeneous materials. Consider a

composite layered material on Ω, which is such that a reference cell Y = [0, 1] × [0, 1] can
be decomposed into two subdomains Y1 = [0, 21 ] × [0, 1], Y2 = [ 12 , 1] × [0, 1]. Suppose
that for i = 1, 2, Yi is occupied by a linear elastic material with Young’s modulus Ei and
Poisson’s ratio νi .

Then, for a given value of  > 0 the domain Ω can be decomposed into two subdomains

x
Ω1 = {x ∈ Ω |χ1 ( ) = 1},

x
Ω2 = {x ∈ Ω |χ2 ( ) = 1}.


Here, χ1 and χ2 are the characteristic functions of the sets Y1 and Y2 given respectively by

1, y ∈ Yi ,
χi (y) = (4.2.43)
0, y ∈ Y − Yi ,

that are extended by periodicity over Y. In this conditions, we state the following

Property 4.2.7 Let us consider the problem (4.2.2) with (4.2.25) - (4.2.26), where
4.2 Homogenisation for elasticity 59

E(y) = E1 χ1 (y1 ) + E2 χ2 (y1 ), (4.2.44)

ν(y) = ν1 χ1 (y1 ) + ν2 χ2 (y1 ). (4.2.45)

Then the corresponding homogenised material is orthotropic with material constants given by

E
Ex = 2
, Ey = E, (4.2.46)
ν + AE

ν
νxy = 2
, νyx = ν, (4.2.47)
ν + AE

1
G= , (4.2.48)
(1 + ν1 )/E1 + (1 + ν2 )/E2

where

1 1 1 1 − ν21 1 − ν22
E= (E1 + E2 ), ν = (ν1 + ν2 ), A = ( + ). (4.2.49)
2 2 2 E1 E2

Proof: It is a consequence of Property 4.2.3.

We conclude this subsection with an example to illustrate the previous property. We

consider an elastic plate occupying the region Ω = [0, 1] × [0, 1]. Let ν1 = 0.1, ν2 = 0.3,
E1 = 1, E2 = 10 and the boundary conditions be such that:


 0, x2 = 1,
ΓD = {0} × [0, 1], ϕN (x) = 0, x2 = 0, (4.2.50)

(1, 0), x1 = 1.

We use a triangular grid with quadratic finite elements to solve the heterogeneous prob-
lem (4.2.2) with (4.2.25) - (4.2.26) for  = 0.1 and f = 0. The vertical component of the
displacements is plotted in Figure 4.5 a). We also plot the value of the horizontal dis-
placement along the vertical line x2 = 0.5 in Figure 4.5 b).
60 Homogenisation for periodic structures

b) Plot of the horizontal component of u along

Figure 4.5 a) Plot of the vertical component of u . The
x2 = 0.5.
contour lines highlight the oscillating behaviour.

Finally we represent a plot of the solution of the corresponding homogenised problem

in Figures 4.6 a) and b), obtained after applying Property 4.2.7. Again we used finite
elements in the same manner as before.

Figure 4.6 a) Plot of the vertical component of u, b) Plot of the horizontal component of u along x2 = 0.5.
which captures the essential behaviour of u .

As expected, the effective solution captures the essence of the behaviour of the original
heterogeneous solution, disregarding only the oscillations.

4.2.3 Numerical aspects

Thus far our attention was devoted to finding explicit analytical expressions for the ho-
mogenised coefficients for layered materials. In this subsection we present a simple
4.2 Homogenisation for elasticity 61

procedure to find the homogenised solution related to (4.2.2) for any material with pe-
riodically distributed coefficients. This is applied to three different problems. The first
two deal with layered materials. We start by considering an example where the coeffi-
cients of the underlying PDE are continuous functions of space. In the second example
the coefficients are only piecewise continuous, as one would expect when dealing with
composite materials. Finally we include another example where a more complicated
structure is considered.

Let us again look at the elasticity problem (4.2.2). In order to determine the homogenised
solution for this problem, it is necessary to first solve the cell problem and determine
the effective coefficients. This procedure can be structured as the following algorithm.

Algorithm 4.1

1 - Solve the cell problem (4.2.7);

2 - Compute the coefficients aijkh by means of (4.2.13);

3 - Determine the solution of the homogenised problem (4.2.10).

We note that numerical techniques may have to be used at any of these steps. It is then
important to estimate the error committed on each step, as the accuracy of the estimate
u allowed by Theorem 4.2.1 depends also on these errors.

Example 4.1 Continuous coefficients

Consider the problem (4.2.2) with (4.2.25)-(4.2.26), where the reference cell is now given
by Y = [−0.5, 0.5] × [−0.5, 0.5] and the material properties characterised by

E(y) = E(y1 ) = 1 + y21 , ν = 0.3. (4.2.51)

The non-trivial effective coefficients for this problem are easily determined using Prop-
erty 4.2.3. The exact values of the coefficients are displayed in the second column of
Table 4.2 with 6 significant digits.

To determine a numerical approximation for the coefficients, we follow steps 1 and 2 of

Algorithm 4.1.

We use finite elements on a quadrangular mesh with quadratic elements with side size
h = 0.01 in order to solve the cell problem. The solution functions χlm
k are either the
62 Homogenisation for periodic structures

constant zero function or are similarly shaped to the function χ11

1 , which we represent
over the reference cell in Figure 4.7.

Figure 4.7: The periodic cell function χ11

1 .

For the second step two distinct strategies were followed in order to determine an ap-
proximation for the effective coefficients. Both aim at solving (4.2.13), but they differ in
how the integrals

Z Z X ∂χkh
l
a
eijlm (y1 )dy1 , a
eijlm (y1 ) dy1 , (4.2.52)
Y Y l,m=1,2 ∂ym

are dealt with.

Coefficients Exact Ap3 Ap9

a1111 1.18506 1.18350 1.18506
a2121 0.41477 0.41423 0.41477
a2211 0.35552 0.35505 0.35552
a2222 1.18999 1.18985 1.18999

Table 4.2: Effective coefficients.

The first approach employs the expressions (4.2.25)-(4.2.26) of the heterogeneous coeffi-
cients a
eijkh . Using these it is possible to compute the values of the integrals in (4.2.52)
analytically when we use for example a least square polynomial approximation for χkh l .
In the third and fourth columns of Table 4.2 we display the approximations Ap3 and
Ap9 for the relevant effective coefficients, obtained by using the mentioned polynomi-
als with degree 3 and 9, respectively.
4.2 Homogenisation for elasticity 63

The values of the residuals of the approximations of the relevant non-zero functions
χkh
l , denoted by Res3 and Res9 , are given in Table 4.3.

Function Res3 Res9

χ11
1 2.1E − 3 1.7E − 7
χ21
2 2.1E − 3 1.7E − 7
χ22
1 6.2E − 4 5.1E − 8

Table 4.3: Residual of the approximation of χkh

l .

The second strategy that we adopt consists of computing the right hand side of (4.2.13)
avoiding fitting. Instead, numerical differentiation and integration techniques are used.
Firstly, for the approximation of ∂χkh l /∂ym , central differences are employed for all
points except for the boundaries. Forward and backward differences are used for x =
−0.5 and x = 0.5, respectively. To compute the integrals, the classical composite Trape-
zoidal and composite Simpson rules with 25 and 101 equidistant points are used. The
values of the estimates for the coefficients are given in Table 4.4.

Trapezoidal Simpson
Coefficients 25 points 101 points 25 points 101 points
a1111 1.18539 1.18508 1.18515 1.18507
a2121 0.41489 0.41478 0.41480 0.41477
a2211 0.35562 0.35552 0.35554 0.35552
a2222 1.19029 1.19001 1.19000 1.18998

Table 4.4: Effective coefficients.

We conclude that by using finite differences followed by the Simpson rule with 101
points one obtains results which are quite accurate. These are indeed nearly as accurate
as the ones obtained by employing fitting with a polynomial of degree 9. Now, once
the effective coefficients have been determined, the homogenised solution can be easily
computed applying the third step of Algorithm 4.1.

Example 4.2 Piecewise coefficients

We consider the problem described in the previous example once again, but where now
instead of (4.2.51) we have
64 Homogenisation for periodic structures

E(y) = E(y1 ) = χ1 (y1 ) + 3χ2 (y1 ); ν = 0.3, (4.2.53)

where χ1 and χ2 are the characteristic functions of the sets Y1 = [−0.5, 0] × [−0.5, 0.5]
and Y2 = [0, 0.5] × [−0.5, 0.5], respectively, as defined in (4.2.43). In order to deal with
the discontinuous behaviour of E(y), we approximate it by the continuous function

2
Eλ (y1 ) = 2 + arctan(λy1 ). (4.2.54)
π

λ=50
λ=100
2.5 λ=1000
E(y1)

1.5

−0.5 0 0.5
y1

Figure 4.8: The functions Eλ .

Note that the approximation is more accurate as the value of λ grows, as illustrated in
Figure 4.8.

Coefficients Exact Approximation

λ = 50 λ = 100 λ = 1000
a1111 1.64835 1.74169 1.70408 1.67335
a2121 0.57692 0.60959 0.59643 0.58907
a2211 0.49451 0.52251 0.51125 0.50200
a2222 2.14835 2.15675 2.15337 2.15192

Table 4.5: Effective coefficients.

We now use the second strategy described earlier. Table 4.5 contains the values of the
approximations obtained for the non-zero coefficients of the homogenised PDE. These
4.2 Homogenisation for elasticity 65

can be compared to the exact solution given by Property 4.2.3, displayed in the second
column of Table 4.5, with 6 significant digits.

As one would expect, the results are more accurate for larger values of λ. This parameter
may be taken arbitrarily large, but at the cost of having to work with a fine mesh, locally
around y1 = 0, the material interface.

Example 4.3 A complicated structure

Let us again consider the elasticity problem (4.2.2) with (4.2.25)-(4.2.26) where over the
reference cell Y = [0, 1] × [0, 1] the material properties now vary along both spatial
coordinates. They are characterised by

E(y) = 1.1 + cos(2πy1 ) sin(2πy2 ), ν = 0.3. (4.2.55)

Let Ω = [0, 1] × [0, 1], f = 0 and the boundary conditions be such that


 (0, 1), x2 = 1,
ΓD = {0} × [0, 1], ϕN (x) = 0, x2 = 0, (4.2.56)

(1, 0), x1 = 1.

As before, we compute the solution functions χlm

k of the cell problem using finite ele-
ments. A fine triangular grid with 42778 quadratic elements is employed. For illustra-
tion, χ12 12
1 and χ2 are plotted in Figure 4.9.

Figure 4.9 a) The periodic cell function χ12

1 . b) The periodic cell function χ12
2 .

We follow Algorithm 4.1 to approximate the value of the effective coefficients of the
homogenised PDE. In Figure 4.10 a) we compare the horizontal components of the ef-
fective and the heterogeneous solutions along the line x2 = 0.5 and in Figure 4.10 b) the
66 Homogenisation for periodic structures

respective vertical components along x1 = 0.5, for  = 1/4. The homogeneous solution
is represented by the dashed lines.

Figure 4.10 a) Horizontal components of u and b) Vertical components of u and the approximation u
the approximation u along x2 = 0.5. along x1 = 0.5.

Finally, we compute the norms of the horizontal and vertical components of u − u for
different values of , see Table 4.6.

 = 1/4  = 1/8  = 1/12

ku1 − u1 kL2 (Ω) 1.4E − 2 7.0E − 3 4.6E − 3
ku2 − u2 kL2 (Ω) 9.5E − 3 4.4E − 3 3.0E − 3

Table 4.6: Error in k · kL2 (Ω) .

As one would expect, for smaller values of , the heterogeneous solution is closer to the
homogenised solution.

We have thus set up a procedure to compute a macroscopic approximation for the solu-
tion of complicated boundary value problems related to periodically distributed struc-
tures. In particular, this is useful to deal with the composite materials that we will
consider in following chapters.
Chapter 5

Domain decomposition

Finding the numerical solutions of problems of practical interest such as elasticity prob-
lems related to highly heterogeneous composites often involves solving large-scale al-
gebraic systems. Domain decomposition methods deal with this by partitioning the
computational domain into a finite number of subdomains. The original PDEs are then
solved on each of these subdomains. Instead of solving one complicated problem, one
solves a number of simpler problems that require less computational effort. By patch-
ing together the subdomains, a sequence of numerical approximations converging to
the sought solution is obtained.

In this chapter we analyse domain decomposition iterative schemes that employ either
overlapping or non-overlapping subdomains. Several algorithms are presented and
their convergence is established.

5.1 One dimensional example

The basic principles of the domain decomposition methods can be illustrated in a one
dimensional setting. Let us consider the boundary value problem

 d
  d 


 − dx a (x) dx u (x) = f(x), x ∈ (0, 1),

u (0) = 0, (5.1.1)

 
u (1) = 0.

Suppose that
68 Domain decomposition

1
a (x) := , f(x) = 1. (5.1.2)
2 + sin( 2πx
 )

The problem (5.1.1) with (5.1.2) was already introduced in the previous chapter. There
we found that the oscillating behaviour of the coefficient a leads to difficulties in solv-
ing (5.1.1) when classical numerical techniques are employed, particularly for   1.
The theory of homogenisation was presented as a valid alternative which simplified the
problem by replacing a by an adequate constant function. The solution u can then be
approximated by the solution u of the homogenised problem

 d2

 −ā dx2 u(x) = 1, x ∈ (0, 1),
(5.1.3)

 u(0) = 0,
u(1) = 0.

Recall that the constant function a is given by (4.1.23). Homogenisation can be ap-
plied effectively to media with constituents that are finely mixed in such a way that the
properties of any material sample are about the same. However, this excludes a great
number of materials. One other drawback of this technique is that even using correc-
tors, homogenisation may yield insufficiently accurate results if the local microscopic
behaviour is sought. In those situations alternative ways of reducing the computational
complexity of the original problem have to be used.

Domain decomposition methods are designed to take advantage of the architectural

structure of modern parallel computers, cf. [47,80]. Though the distinction is not always
clear, the algorithms are usually divided into Schwarz methods and substructing or
Schur complement methods, cf. [3, 13, 79, 81]. The latter are associated with the usage
of non-overlapping subdomains, while the former employ overlap. More details about
the relationships between these different schemes are given in [14, 18, 40].

5.1.1 Substructuring iterative methods

To illustrate how (5.1.1) can be solved using domain decomposition methods, we start
by splitting the computational domain Ω, with Ω = (0, 1), into two non-overlapping
subdomains Ω1 and Ω2 , where

Ω1 = (0, 1/2), Ω2 = (1/2, 1). (5.1.4)

Denote by ui the restriction of the solution u of (5.1.1) to Ωi , i = 1, 2. We suppress

the superscript  and write ui instead of ui for simplicity reasons. Assuming that u ∈
C2 (Ω), (5.1.1) is equivalent to
5.1 One dimensional example 69


 d d
  
 − dx a (x) dx u1 (x) = f(x), x ∈ Ω1 ,





 u1 (0) = 0,





u1 (1/2) = u2 (1/2),
(5.1.5)
 ∂u1 − ∂u2
 ∂x (1/2 ) = ∂x (1/2 ),
+






 d d
  
 − dx a (x) dx u2 (x) = f(x), x ∈ Ω2 ,



u2 (1) = 0.

In fact, due to the continuity of u and of its derivative in x = 1/2, the matching conditions

u1 (1/2) = u2 (1/2),
∂u1 − ∂u2 + (5.1.6)
∂x (1/2 ) = ∂x (1/2 ),

(k)
are redundant. Splitting Ω into Ω1 and Ω2 , sequences of functions {ui }k are gener-
ated, each defined on the subdomain Ωi , i = 1, 2. The elements of each sequence satisfy
a boundary value problem comprised by the original PDEs and the boundary condition
(k)
ui = 0 at ∂Ω ∩ ∂Ωi . As for the other boundary condition, there are different ways
(k)
to prescribe it to ensure that ui converges to ui as k increases. The common idea is
(k) (k)
to couple u1 and u2 by imposing one of the matching conditions and to iterate to
achieve the other. If for example we impose continuity of derivatives by

(k) (k)
∂u2 ∂u1
(1/2+ ) = (1/2− ),
∂x ∂x

(k) (k)
the iterative process must be such that as k increases the quantity |u1 (1/2) − u2 (1/2)|
tends to zero. Accordingly we present the following algorithm to solve (5.1.1), where
we represent by θ, α(0) and Tol the acceleration parameter, the initial approximation for
u (1/2) and the tolerance for the stopping condition, respectively.

Algorithm 5.1

Set k = 0. Given: θ, α(0) , Tol.

1 - Solve
70 Domain decomposition



h i
d d (k+1)
 −
 dx a
(x) u
dx 1 (x) = f(x), x ∈ Ω1 ,
(k+1) (5.1.7)


u1 (0) = 0,
 (k+1)
u1 (1/2) = α(k) ,

and

 h i


d
− dx d (k+1)
a (x) dx u2 (x) = f(x), x ∈ Ω2 ,


(k+1) (5.1.8)

 u2 (1) = 0,

 ∂u(k+1) ∂u1
(k+1)
2
∂x (1/2+ ) = ∂x (1/2− ).

2 - Update α(k) and increment k

(k+1)
α(k+1) = θu2 (1/2) + (1 − θ)α(k) , (5.1.9)
k → k + 1. (5.1.10)

(k) (k)
3 - Return to step 1 until |u1 (1/2) − u2 (1/2)| < Tol.

The so-called Dirichlet-Neumann procedure, described in Algorithm 5.1, will converge

as long as θ is conveniently chosen. As we will see later, θ is an acceleration parameter
that can be optimised for faster convergence. We note that in order to ensure faster
convergence it is important to choose a value for α(0) that is as good an approximation
as possible to u (1/2). An educated guess is given by the homogenised solution, α(0) =
u(1/2). This has the advantage of being easy to compute.

We will now consider an example of application of the previous algorithm for which
α(0) will not be given by the homogenised solution. We do this because this educated
guess would be an unrealistically good initial guess, much better than what is usually
available in problems where homogenisation is not applicable. Moreover, we simply
take θ = 0.4.

Tol ku(k) − u kL2 (Ω) k

1E − 3 2.31E − 4 5
1E − 4 9.24E − 6 7
1E − 5 1.85E − 6 8

Table 5.1: Error and number of iterations, Eα = 0.25.

5.1 One dimensional example 71

Using Algorithm 5.1 we approximate the solution of (5.1.1) with (5.1.2) and  = 0.1
(k)
by a sequence of functions {u(k) }k defined by u(k) |(Ωi \{1/2}) := ui and u(k) (1/2) :=
(k)
u1 (1/2). For each value of k, we employ finite elements to solve (5.1.7)-(5.1.8). Uni-
form meshes with width h = 1E − 4 are used for each boundary value problem. The
L2 norms of the errors u(k) − u are displayed in the second column of Table 5.1 for the
smallest number of iterations k that ensures the stopping condition is achieved, taking
various values for Tol. The initial guess is taken α(0) = 0.

Tol ku(k) − u kL2 (Ω) k

1E − 3 1.39E − 4 6
1E − 4 2.77E − 5 7
1E − 5 1.11E − 6 9

Table 5.2: Error and number of iterations, Eα = 0.75.

Similar calculations are presented in Tables 5.2 and 5.3 for α(0) = 1 and α(0) = 10,
respectively. The error Eα := |α(0) − u (1/2)| of the initial guesses, as indicated in the
legends of the tables, is increasing. A larger number of iterations is then necessary to
achieve the stopping condition.

Tol ku(k) − u kL2 (Ω) k

1E − 3 7.21E − 5 8
1E − 4 1.44E − 5 9
1E − 5 2.88E − 6 10

Table 5.3: Error and number of iterations, Eα = 9.75.

Finally, we note that the value of the tolerance Tol should be chosen carefully taking the
expected behaviour of the error function into account. For instance, in the case of the
problem we have considered, since the sign of a is constant, it can easily be shown
that when the desired tolerance is attained the maximum norm of the error u(k) − u is
smaller than Tol. The reasoning is similar to the proof of the Theorem 5.1.1 that will be
presented later.

5.1.2 Alternating Schwarz methods

Let us once again consider the problem (5.1.1) where we now take the following subdo-
mains
72 Domain decomposition

Ω1 = (0, γ2 ), Ω2 = (γ1 , 1), (5.1.11)

which are such that 0 < γ1 < γ2 < 1. We denote the length of the overlap by δ := γ2 −γ1
and assume that there exist constants α and β such that 0 < α < a (x) < β. It is easily
concluded that


 d d
  
 − dx a (x) dx u1 (x) = f(x), x ∈ Ω1 ,





 u1 (0) = 0,





u1 (γ1 ) = u2 (γ1 ),
(5.1.12)

 u1 (γ2 ) = u2 (γ2 ),





 d d

  
 − dx a (x) dx u2 (x) = f(x), x ∈ Ω2 ,



u2 (1) = 0,

is equivalent to (5.1.1) where ui = u |Ωi as before, for i = 1, 2. We observe that the

matching conditions

u1 (γ1 ) = u2 (γ1 ),
(5.1.13)
u1 (γ2 ) = u2 (γ2 ),

are now both of the same type, but hold at different points in Ω. In order to obtain
approximations for the solution of the problem we set up an iterative process that gen-
(k)
erates a sequence of approximation functions {ui }k defined on the subdomains Ωi ,
i = 1, 2. One of the matching conditions will be imposed for all the iteration steps and
we will iterate in order to obtain the other condition. The iterative process, so-called
the alternating Schwarz method, can be described in terms of the following algorithm. We
denote by α(0) and Tol the initial approximation for u (γ2 ) and the tolerance for the
stopping condition, respectively.

Algorithm 5.2

Set k = 0. Given: α(0) , Tol.

1 - Solve



h i
d d (k+1)
 −
 dx a
(x) u
dx 1 (x) = f(x), x ∈ Ω1 ,
(k+1) (5.1.14)


u1 (0) = 0,
 (k+1)
u1 (γ2 ) = α(k) ,
5.1 One dimensional example 73

and



h i
d d (k+1)

 − a 
(x) u (x) = f(x), x ∈ Ω2 ,
 dx dx 2

(5.1.15)


(k+1)
u2 (1) = 0,

 (k+1) (k+1)
u2 (γ1 ) = u1 (γ1 ).

2 - Update α(k) and increment k

(k+1)
α(k+1) = u2 (γ2 ), (5.1.16)
k → k + 1. (5.1.17)

3 - Return to step 1 until

(k) (k)
|u1 (γ2 ) − u2 (γ2 )| < Tol. (5.1.18)

(k)
To show the convergence of this iterative process, let us define the error functions Ei :=
(k) (0)
ui − u |Ωi , for k ≥ 1. Moreover, let E2 (x) := α(0) − u2 (x), x ∈ Ω2 . Then we have



h i
d d (k)

 − a 
(x) E (x) = 0, x ∈ Ω1 ,
 dx dx 1

(5.1.19)


(k)
E1 (0) = 0,

 (k) (k−1)
E1 (γ2 ) = E2 (γ2 ),

and



h i
d d (k)

 − dx a (x) dx E2 (x) = 0, x ∈ Ω2 ,

(5.1.20)


(k) (k)
E2 (γ1 ) = E1 (γ1 ),

 (k)
E2 (1) = 0.

(k)
Observe that if the sign of a is constant, the functions Ei are strictly monotonous. It
is then easy to see that the modulus of the error function defined by


(k)
(k) E1 (x), x ∈ [0, γ1 ],
E (x) := (k) (5.1.21)
E2 (x), x ∈ [γ1 , 1],
74 Domain decomposition

decreases as k increases, i.e., |E(k) (x)| < |E(k−1) (x)|, for x ∈ Ω. The error behaves as
(k) (k) (k) (k−1)
illustrated in Figure 5.1. Note that E1 (γ1 ) = E2 (γ1 ) and E1 (γ2 ) = E2 (γ2 ).

(1) E(1)
E1 2

E(2)
1
(2)
E2
E(3)
1

0 γ1 γ 1
2
x

Figure 5.1: Error reduction.

We can now state the following theorem.

Theorem 5.1.1 Let p be the smallest natural number for which (5.1.18) holds and assume that
there exist real constants α and β such that 0 < α < a (x) < β. Then the approximation
function


(p)
(p) u1 (x), x ∈ [0, γ1 ],
u (x) := (p) (5.1.22)
u2 (x), x ∈ [γ1 , 1],

for the exact solution u of (5.1.1) is such that

Rγ1
1/a (u)du
ku (p)
− u k∞ ≤ R0γ2

Tol. (5.1.23)
γ1
1/a (u)du

(p)
Proof: The maximum of |E(p) | is attained at x = γ1 and so ku(p) − u k∞ ≤ |E2 (γ1 )|. In
(p)
order to estimate an upper bound for the value of |E2 (γ1 )|, we note that from (5.1.19)
and (5.1.20) it can easily be seen that there exist constants ci and di such that
5.1 One dimensional example 75

Zx Zx
(p) 1 (p) 1
E1 (x) = c1 du + d1 and E2 (x) = c2 du + d2 .
0 a (u) γ1 a (u)

These constants can be determined in terms of integrals of the coefficient function a .

(p) (p)
For that we use the boundary conditions for E1 and E2 on x = 0 and x = 1 respec-
tively, and also the conditions

(p) (p) (p) (p)

E1 (γ1 ) = E2 (γ1 ) and |E1 (γ2 )| = |E2 (γ2 )| + Tol.

The result then follows.

The Schwarz algorithm can be employed to find an approximation to the solution of

the problem (5.1.1) with (5.1.2). This yields the results displayed in Tables 5.4 to 5.6 for
α(0) = 0,  = 0.1 and for various overlapping intervals, centred at x = 1/2 and with
length δ. The minimum number of iterations k needed to obtain the desired tolerance
for each case is displayed in the third columns of the tables. Both (5.1.14) and (5.1.15)
are solved employing finite elements. For each of these problems a uniform mesh with
width h = 1E − 4 is used.

Tol ku(k) − u kL2 (Ω) k

1E − 3 1.40E − 3 12
1E − 4 1.26E − 4 18
1E − 5 1.13E − 5 24

Table 5.4: Error and number of iterations, δ = 0.1.

The L2 norms of the errors u(k) − u are displayed in the second column of the tables.

Tol ku(k) − u kL2 (Ω) k

1E − 3 3.23E − 4 8
1E − 4 6.38E − 5 10
1E − 5 5.60E − 6 13

Table 5.5: Error and number of iterations, δ = 0.2.

From these results it appears that larger values of δ yield the stopping condition to
be achieved after fewer iterations. The drawback is that the computational effort per
76 Domain decomposition

Tol ku(k) − u kL2 (Ω) k

1E − 3 1.51E − 4 6
1E − 4 4.39E − 5 7
1E − 5 3.68E − 6 9

Table 5.6: Error and number of iterations, δ = 0.3.

iteration increases when the problems have to be solved on a larger region. Also, as
it would be expected, smaller values of Tol imply more accuracy but require a larger
number of iterations.

This agrees with the result established in the theorem. In fact, using the upper and lower
bounds of a , it follows from (5.1.23) that

β γ1
ku(k+1) − u k∞ < Tol. (5.1.24)
α δ

5.2 The elasticity problem

In the previous chapter, elasticity problems concerning periodically distributed mate-

rials were considered. We now look into the more general problem of linear elastic
materials that do not necessarily display a periodic behaviour, for which the domain
decomposition methods provide suitable numerical techniques.

5.2.1 The Steklov-Poincaré equation

Let us consider the elasticity problem given by



 −∇ · (A(x)(u)) = f, x ∈ Ω,


u = 0, x ∈ ΓD , (5.2.1)


 σ(u) · n = ϕN ,
 x ∈ ΓN .

To partition the computational domain into smaller non-overlapping subdomains Ω1

and Ω2 , we note that the physics of the problem suggests the continuity of both the
displacements and stresses over the interface Γ = ∂Ω1 ∩ ∂Ω2 . Bearing this in mind and
5.2 The elasticity problem 77

assuming u to be sufficiently smooth, we may reformulate our original problem in the

equivalent form


 −∇ · (A(x)(u1 )) = f, x ∈ Ω1 ,





 u1 = 0,
T
x ∈ ΓD T ∂Ω1 ,



 σ(u 1 ) · n = ϕN , x ∈ ΓN ∂Ω1 ,







u1 = u2 , x ∈ Γ,
(5.2.2)

 σ(u 1 ) · n = σ(u2 ) · n, x ∈ Γ,







 −∇ · (A(x)(u2 )) = f, x ∈ Ω2 ,






T

 u2 = 0, x ∈ ΓD T ∂Ω2 ,
σ(u2 ) · n = ϕN , x ∈ ΓN ∂Ω2 .

Here ui is the restriction of the solution u of (5.2.1) to Ωi , for i = 1, 2. We would like to

split the problem (5.2.2) into two separate problems, one defined on Ω1 and the other
defined on Ω2 . We then start by considering the boundary value problems



 −∇ · (A(x)(wi )) = f, x ∈ Ωi ,



 T
wi = 0, x ∈ ΓD T ∂Ωi ,
(5.2.3)

 σ(wi ) · n = ϕN , x ∈ ΓN ∂Ωi ,



 w = λ,
i x ∈ Γ,

where λ is a given vector function. Clearly, if the value of u|Γ were known, it would
suffice to simply set λ := u|Γ and (5.2.3) would be equivalent to (5.2.2), in the sense that
wi = ui . However, since the function u|Γ is unknown, to ensure this equivalence λ must
be prescribed so that the solutions of (5.2.3) satisfy the following matching conditions on
Γ:

w1 = w2 , (5.2.4)
σ(w1 ) · n = σ(w2 ) · n. (5.2.5)

Note that (5.2.4) is trivially verified. As for (5.2.5), this requirement on λ can be written
in the form of the Steklov-Poincaré equation

Sλ = χ, (5.2.6)

cf. [72]. The Steklov-Poincaré operator S plays a central role in the convergence analysis
of the domain decomposition methods we will look into. To properly define both S and
78 Domain decomposition

χ, let us denote by Hi λ the harmonic extension of λ into Ωi , that is, the solution of the
problem



 −∇ · (A(x)(Hi λ)) = 0, x ∈ Ωi ,





T
 Hi λ = 0, x ∈ ΓD T ∂Ωi ,
σ(Hi λ) · n = 0, x ∈ ΓN ∂Ωi , (5.2.7)





 Hi λ = λ, x ∈ Γ.



Furthermore, let us denote by Gi f the solution of



 −∇ · (A(x)(Gi f)) = f, x ∈ Ωi ,





T
 Gi f = 0, x ∈ ΓD T ∂Ωi ,
σ(Gi f)f · n = ϕN , x ∈ ΓN ∂Ωi , (5.2.8)





 Gi f = 0, x ∈ Γ.



Then the Steklov-Poincaré operator S is defined by S := S1 + S2 , where

S1 λ = σ(H1 λ).n, S2 λ = −σ(H2 λ).n, (5.2.9)

and χ is given by

χ = σ(G2 f).n − σ(G1 f).n.

It can then easily be seen that if λ satisfies the Steklov-Poincaré equation the solutions of
(5.2.3) satisfy (5.2.5) and so ui = wi . We must then focus on determining the solution of
(5.2.6). Several iterative schemes can be employed for that. In particular, the Dirichlet-
(k) (k)
Neumann method imposes continuity of stresses over Γ for all terms u1 and u2 of
the sequences of problems that it generates on Ω1 and Ω2 respectively. The quantity
(k) (k)
u1 − u2 should then converge to zero as k tends to infinity. As we will see, this
is equivalent to generating a sequence of functions λ(k) that converges to the solution
of the Steklov-Poincaré interface equation on Γ . The Dirichlet-Neumann method can be
written as the following algorithm, where θ is a positive acceleration parameter, an
initial guess function λ(0) must be provided and Tol = (Tol1 , Tol2 ) is the tolerance for
the stopping condition.
5.2 The elasticity problem 79

Algorithm 5.3

Set k = 0. Given: θ, λ(0) , Tol.

1 - Solve



 −∇ · (A(u1
(k+1)
)) = f, x ∈ Ω1 ,





 u(k+1) = 0, T
1 x ∈ ΓD ∂Ω1 ,
(k+1) (5.2.10)

T
 σ(u ) · n = ϕN , x ∈ ΓN ∂Ω1 ,


1



 (k+1)
u1 = λ(k) , x ∈ Γ,

and

 (k+1)

 −∇ · (A(u2 )) = f, x ∈ Ω2 ,





 (k+1) T
 u2 = 0, x ∈ ΓD ∂Ω2 ,
(k+1)
(5.2.11)
T
σ(u2 ) · n = ϕN , x ∈ ΓN ∂Ω2 ,




 σ(u(k+1) ) · n = σ(u(k+1) ) · n,


 x ∈ Γ.
 2 1

2 - Update λ(k) and increment k

(k+1)
λ(k+1) = θu2 |Γ + (1 − θ)λ(k) , (5.2.12)

k → k + 1. (5.2.13)

3 - Return to step 1 until

(k) (k)
k(u1 |Γ − u2 |Γ )i k∞ < Toli . (5.2.14)

(k) (k) (k) (k)

Here, (u1 |Γ − u2 |Γ )i is the i-th component of the vector function (u1 |Γ − u2 |Γ ),
for i = 1, 2. Now, the iterative scheme (5.2.10)-(5.2.12) can be reduced to an interface
equation given in terms of the components of the Steklov-Poincaré operator. In fact we
have
80 Domain decomposition

(k+1) (k+1)
u2 |Γ = (u2 − G2 f)|Γ
(k+1)
= S−1
2 S2 (u2 − G2 f)|Γ
h i
−1 (k+1)
= S2 −σ(u1 ) · n + σ(G2 f) ·n
h i
= S−1
2 −σ(H 1 λ (k)
+ G 1 f) · n + σ(G 2 f) · n ⇒
(k+1)
|Γ S−1
 
u2 = 2 −S1 λ(k) + χ . (5.2.15)

Inserting (5.2.15) into (5.2.12) we verify that the iterative scheme on λ(k) described in
the previous algorithm reduces to the preconditioned Richardson method

S2 (λ(k+1) − λ(k) ) = θ(χ − Sλ(k) ), k ≥ 0, (5.2.16)

for the Steklov-Poincaré equation with the preconditioner S2 .

Several alternatives of domain decomposition methods exist that approximate the solu-
tion of the elasticity problem (5.2.1). Among these we distinguish substructuring methods
such as the Neumann-Neumann, the Robin method, or the Dirichlet-Neumann scheme
we have described earlier, cf. [72]. All of these partition the computational domain in
non-overlapping subdomains. Sometimes however it is more appropriate to have over-
lap. In particular, the Schwarz scheme for (5.2.1) with overlapping subdomains Ω1 and
Ω2 is given by Algorithm 5.4.

W1

G2
G1
W2

Figure 5.2: Overlapping subdomains.

Let once again θ be a positive acceleration parameter, λ(0) an initial guess function and
Tol = (Tol1 , Tol2 ) the tolerance for the stopping condition. Moreover, let Γ1 and Γ2 be the
5.2 The elasticity problem 81

portions of ∂Ω1 and ∂Ω2 that are not in ∂Ω, respectively, see Figure 5.2. The algorithm
reads

Algorithm 5.4

Set k = 0. Given: λ(0) , Tol.

1 - Solve



 −∇ · (A(u1
(k+1)
)) = f, x ∈ Ω1 ,





 u(k+1) = 0, T
1 x ∈ ΓD ∂Ω1 ,
(k+1) (5.2.17)

T
 σ(u ) · n = ϕN , x ∈ ΓN ∂Ω1 ,


1



 (k+1)
u1 = λ(k) , x ∈ Γ1 ,

and



 −∇ · (A(u2
(k+1)
)) = f, x ∈ Ω2 ,




 (k+1) T
u2 = 0, x ∈ ΓD ∂Ω2 ,
(5.2.18)


(k+1) T
 σ(u2 ) · n = ϕN , x ∈ ΓN ∂Ω2 ,



 (k+1) (k+1)
u2 = u1 , x ∈ Γ2 .

2 - Update λ(k) and increment k

(k+1)
λ(k+1) = u2 | Γ1 , (5.2.19)
k → k + 1. (5.2.20)

3 - Return to step 1 until

(k) (k)
k(u1 |Γ1 − u2 |Γ1 )i k∞ < Toli . (5.2.21)

We note that both iterative schemes included in Algorithms 5.3 and 5.4 can be gener-
alised to the case of having more subdomains. The idea is to perform each iteration
step on all subdomains, with exchange of information occurring only over the inter-
faces of neighbouring subdomains. This can be implemented using parallel computing,
cf. [35, 80].
82 Domain decomposition

5.2.2 Convergence analysis

Earlier the Dirichlet-Neumann iterative scheme expressed by (5.2.10)-(5.2.12) was pro-

posed to solve (5.2.1). Following [72] we will establish the convergence of this scheme.
We begin by rewriting (5.2.1) in a variational formulation

a(u, v) = hl, vi, ∀ v ∈ V, (5.2.22)

where, as we have already introduced in Chapter 2,

Z X
2
a(u, v) = σij (u)ij (v)dV, (5.2.23)
Ω i,j=1
Z X
2 I X
2
hl, vi = fi vi dV + ϕN,i vi dS. (5.2.24)
Ω i=1 ΓD i=1

We assume that f ∈ (L2 (Ω))2 , ϕN ∈ (L2 (ΓN ))2 and

V = {v | v ∈ (H1 (Ω))2 , γ(v) = 0 on ΓD }. (5.2.25)

Recall that (5.2.1) can be rewritten as (5.2.2). This can in turn be written in a variational
formulation as follows



 a1 (u1 , v1 ) = hl1 , v1 i, ∀ v1 ∈ V10 ,

u1 = u2 , x ∈ Γ,
(5.2.26)

 a2 (u2 , v2 ) = hl2 , v2 i, ∀ v2 ∈ V20 ,

a2 (u2 , R2 µ) = hl2 , R2 µi + hl1 , R1 µi − a1 (u1 , R1 µ), ∀ µ ∈ Λ,

where ui and vi denote the restrictions of u = (u1 , u2 ) and v = (v1 , v2 ) to Ωi , for

i = 1, 2. Moreover, Ri denotes any possible extension operator from Λ to Vi and

Vi := {vi ∈ (H1 (Ωi ))2 | vi |(∂Ω∩∂Ωi ) = 0},

Vi0 := (H10 (Ωi ))2 ,
Λ := {η ∈ (H1/2 (Γ ))2 | ∃ v ∈ V : η = v|Γ },
R P2
ai (ui , vi ) := Ωi j,k=1 σjk (ui )jk (vi )dV,
R P2 H P2
hli , vi i := Ωi j=1 [fj vj ]|Ωi dV + ΓD ∩∂Ωi j=1 [ϕN,j vj ]|Ωi dS.
5.2 The elasticity problem 83

Next we want to write the Steklov-Poincaré equation in a variational formulation. For that
we first introduce for each η ∈ Λ the solution Hi η ∈ Vi of the Dirichlet boundary value
problem

ai (Hi η, vi ) = 0, ∀ vi ∈ Vi0 ,
(5.2.27)
Hi η = η, x ∈ Γ,

and the solution G i f ∈ Vi0 of

ai (G i f, vi ) = hli , vi i, ∀ vi ∈ Vi0 . (5.2.28)

We are now able to define the symmetric and coercive Steklov-Poincaré operators. For
each η, µ ∈ Λ, let

hSi η, µi := ai (Hi η, Hi µ).

Finally, let λ = u|Γ , S := S1 + S2 and χ := χ1 + χ2 , with

hχi , µi := hli , Hi µi − ai (G i f, Hi µ), ∀ µ ∈ Λ.

The stress continuity condition (5.2.5) then translates into the following interface equa-
tion on Γ :

Sλ = χ. (5.2.29)

We now turn our attention to the Dirichlet-Neumann scheme (5.2.10)-(5.2.12). In a weak

form, this is given by the following algorithm, where θ is an acceleration parameter,
λ(0) is the initial approximation for u|Γ and Tol = (Tol1 , Tol2 ) is the tolerance for the
stopping condition.

Algorithm 5.5

Set k = 0. Given: λ(0) , Tol.

(k+1) (k+1)
1 - Find u1 ∈ V1 and u2 ∈ V2 such that


(k+1)
a1 (u1 , v1 ) = hl1 , v1 i, ∀ v1 ∈ V10 ,
(k+1) (5.2.30)
u1 = λ(k) , x ∈ Γ,
84 Domain decomposition

and


 (k+1)
 a2 (u2 , v2 ) = hl2 , v2 i, ∀ v2 ∈ V20 ,
(k+1)
a2 (u2 , R2 µ) = hl2 , R2 µi + hl1 , R1 µi (5.2.31)

 (k+1)
−a1 (u1 , R1 µ), ∀ µ ∈ Λ.

2 - Update λ(k) and increment k

(k+1)
λ(k+1) = θu2 |Γ + (1 − θ)λ(k) , (5.2.32)
k → k + 1. (5.2.33)

3 - Return to step 1 until

(k) (k)
k(u1 |Γ − u2 |Γ )i kH1/2 (Γ ) < Toli . (5.2.34)

We observe that this iterative scheme reduces to the preconditioned Richardson method

S2 (λ(k+1) − λ(k) ) = θ(χ − Sλ(k) ), k ≥ 0, (5.2.35)

for (5.2.29), where S2 is employed as a preconditioner for S. The convergence of {λ(k) }k

to the unique solution λ of the Steklov-Poincaré equation is established in the following
general theorem, cf. [72].

Theorem: Let X be a real Hilbert space, X 0 its dual space, Q : X → X 0 a linear invertible
continuous operator and G ∈ X 0 . Assume that Q = Q1 + Q2 , where Q1 and Q2 are
continuous linear operators, Q2 is also coercive and there exist positive constants α, k∗
and βi such that

hQi η, µi ≤ βi kηkX kµkX , ∀ η, µ ∈ X,

hQ2 η, ηi ≥ αkηk2X , ∀ η ∈ X,
∗
hQ2 η, Q−1 2
2 Qηi + hQη, ηi ≥ k kηkX , ∀ η ∈ X.

for i = 1, 2. Then, for any given λ(0) ∈ X and for any θ satisfying 0 < θ < θmax , with

k∗ α2
θmax := ,
β2 (β1 + β2 )2
5.3 Discrete solution 85

the sequence {λ(k) }k given by

Q2 (λ(k+1) − λ(k) ) = θ(G − Qλ(k) ), (5.2.36)

converges in X to the solution λ ∈ X of Qλ = G.

5.3 Discrete solution

Domain decomposition algorithms are often perceived to be quite complex. They are
more easily understood when presented in the discrete form, using matrix notation. In
this section the Dirichlet-Neumann method is studied at a discrete level for the solution
of the elasticity problem that in the variational formulation is given by (5.2.22). A con-
vergence analysis is carried out and an optimal value for the acceleration parameter θ
is discussed.

As we have seen in Chapter 2, a finite element approximation uh to the solution u of

(5.2.22) consists of a linear combination of the form

X
N
uh = yj φj . (5.3.1)
j=1

Here, {φj }j=1,...,N is a base of a finite dimensional subspace Vh of V, with dim Vh = N.

We denote the finite element nodes by aj so that we have

φi (aj ) = δij . (5.3.2)

To determine the vector y = (yi )i=1,...,N , we insert (5.3.1) into the variational problem
(5.2.22), which yields

XN
a( yj φj , φk ) = hl, φk i, k = 1, . . . , N. (5.3.3)
j=1

This leads to the linear system

Dy = b, (5.3.4)
86 Domain decomposition

where the components of b = (bi )i=1,...,N and the components of the finite element
stiffness matrix D = [dij ]i,j=1, ..., N are given by

bi := hl, φi i and dij := a(φi , φj ), (5.3.5)

respectively. Before we proceed, we would like to draw a parallel to section 5.2. Like
there we will rewrite the problem at hand, expressed by (5.3.4), in a form that suggests
it may be split into smaller problems. For that we decompose the vector of unknowns y
into three vectors y1 , y2 and yΓ of length N1 , N2 and NΓ respectively. The components
of these vectors are the components of y corresponding to the nodes belonging to Ω1 \Γ ,
Ω2 \ Γ and Γ , respectively. Likewise, b is decomposed into b1 , b2 and bΓ . The equation
(5.3.4) can then be rewritten in the form

    
D11 0 D1Γ y1 b1
 0 D22 D2Γ   y2  =  b2  . (5.3.6)
DΓ1 DΓ2 DΓΓ yΓ bΓ

Here the jth line and kth column of the matrices Dii and DiΓ are given by

(i) (i)
ai (φj , φk ), for j, k = 1, . . . , Ni and
(Γ ) (i)
ai (φj , φk ), for j = 1, . . . , NΓ , k = 1, . . . , Ni ,

(i)
respectively. The basis functions denoted by φj correspond to the nodes lying in Ωi \Γ ,
(Γ )
while φj are associated with the nodes lying on Γ . As for DΓΓ , its components read

(Γ ) (Γ ) (Γ ) (Γ )
a1 (φj , φk ) + a2 (φj , φk ), for j, k = 1, . . . , NΓ .

(i)
We denote the contribution to DΓΓ coming from ai by DΓΓ . Eliminating y1 and y2 from
(5.3.6) we obtain

Σh yΓ = χh , (5.3.7)

with

Σh := DΓΓ − DΓ1 D−1 −1

11 D1Γ − DΓ2 D22 D2Γ , (5.3.8)

χh := bΓ − DΓ1 D−1 −1
11 b1 − DΓ2 D22 b2 . (5.3.9)
5.3 Discrete solution 87

The matrix Σh is the Schur complement matrix, the algebraic counterpart of the Steklov-
Poincaré operator on the discrete level. The solution of (5.3.7) can be determined easily
because Σh is symmetric and positive definite. However, it is also ill-conditioned, cf.
[72]. To deal with that, we note that like DΓΓ , Σh can also be split to two components

Σh = Σ1,h + Σ2,h , (5.3.10)

where

(i)
Σi,h = DΓΓ − DΓi D−1
ii DiΓ . (5.3.11)

The idea is to employ any of the matrices Σ1,h and Σ2,h as a preconditioner for Σh .
We can then determine yΓ without having to solve (5.3.7) directly. As we will see, the
discrete version of the Dirichlet-Neumann method uses Σ2,h as a preconditioner for Σh .
Before we introduce the corresponding algorithm, let us define

Vi,h := {vh |Ωi | vh ∈ Vh },

0
Vi,h := {wh ∈ Vi,h | wh |Γ = 0},
Λh := {vh |Γ | vh ∈ Vh },

(0)
and denote by Ri,h any extension operator from Λh to Vi,h , i=1,2. Let λh be an initial
approximation for uh |Γ and Tol the tolerance for the stopping condition. The algorithm
then reads

Algorithm 5.6

(0)
Set k = 0. Given: λh , Tol.

(k+1) (k+1)
1 - Find u1,h ∈ V1,h and u2,h ∈ V2,h satisfying


 a1 (u(k+1) , v1,h ) = hl1 , v1,h i, 0
∀ v1,h ∈ V1,h ,
1,h
(5.3.12)
 u(k+1) = λ(k) , x ∈ Γ,
1,h h

and
88 Domain decomposition




(k+1) 0
 a2 (u2,h , v2,h ) = hl2 , v2,h i, ∀ v2,h ∈ V2,h ,
a2 (u2,h
(k+1)
, R2,h µh ) = hl2 , R2,h µh i + hl1 , R1,h µh i (5.3.13)


 (k+1)
−a1 (u1,h , R1,h µh ), ∀ µh ∈ Λh .

(k)
2 - Update λh and increment k

(k+1) (k+1) (k)

λh = θu2,h |Γ + (1 − θ)λh , (5.3.14)
k → k + 1. (5.3.15)

3 - Return to step 1 until

(k) (k)
k(u1,h |Γ − u2,h |Γ )i kH1/2 (Γ ) < Toli . (5.3.16)

It is convenient to interpret the scheme described in the previous algorithm in an alge-

(k) (k) (k) (k)
braic form. For that we introduce the vectors y1 = (y1,j )j=1,...,N1 , y2 = (y2,j )j=1,...,N2
(k) (k)
and yΓ = (yΓ,j )j=1,...,NΓ , which are such that

(k)
X
N1
(k) (1)
u1,h = y1,j φj ,
j=1

(k)
X
N2
(k) (2)
u2,h = y2,j φj ,
j=1

(k)
X
NΓ
(k) (Γ )
λh = yΓ,j φj .
j=1

Now (5.3.12) corresponds to the Dirichlet problem

(k+1) (k)
D11 y1 = b1 − D1Γ yΓ . (5.3.17)

(k+1)
Solving (5.3.17) for y1 yields

(k+1) (k)
y1 = D−1 −1
11 b1 − D11 D1Γ yΓ . (5.3.18)

We can rewrite (5.3.13) as

5.3 Discrete solution 89

 " (k+1)
#  
D22 D2Γ y2 b2
(2) (k+1/2) = (k+1) (1) (k) . (5.3.19)
DΓ2 DΓΓ yΓ bΓ − DΓ1 y1 − DΓΓ yΓ

From (5.3.18) and (5.3.19) we conclude

(2) (k+1/2) (1) (k)

(DΓΓ − DΓ2 D−1
22 D2Γ )yΓ = (−DΓΓ + DΓ1 D−1
11 D1Γ )yΓ
+bΓ − DΓ1 D11 b1 − DΓ2 D−1
−1
22 b2 , (5.3.20)

which is equivalent to

(k+1/2) (k)
Σ2,h yΓ = χΓ − Σ1,h yΓ . (5.3.21)

Finally, (5.3.14) leads to

(k+1) (k) (k)

yΓ = θΣ−1
2,h (χh − Σ1,h yΓ ) + (1 − θ)yΓ , (5.3.22)

which can be recognised as a Richardson iteration

(k+1) (k) (k)

Ph (yΓ − yΓ ) = θ(χh − Σh yΓ ), for k ≥ 0 (5.3.23)

for (5.3.7), where Ph = Σ2,h is being used as a preconditioner for Σh .

The convergence of the Dirichlet-Neumann scheme is stated in the following theorem,

cf. [72].

(k)
Theorem 5.3.1 : The sequence (yΓ )k generated by the preconditioned Richardson iterations
(5.3.23) converges to the unique solution yΓ of (5.3.7) if and only if θ satisfies

2
0<θ< , (5.3.24)
νmax

where we denote by νmax and νmin the maximum and minimum eigenvalues of the matrix
−1
Σ2,h Σh . Moreover, the optimal value θopt which minimises the spectral radius of the iteration
matrix is

2
θ = θopt := . (5.3.25)
νmin + νmax
90 Domain decomposition

This result allows us to find the optimal value for the acceleration parameter θ in terms
of the two components of the Schur complement matrix.
Chapter 6

Hybrid approach

Up to now, two distinct techniques were analysed for solving boundary value problems
with oscillating coefficients that express the behaviour of materials with complex mi-
crostructures. These techniques share the feature of reducing the computational com-
plexity of the problem at hand. In general terms, homogenisation was presented for
periodic structures, whilst domain decomposition is applicable to more complex struc-
tures but requires the resolution of the heterogeneous scale.

In this chapter boundary value problems related to periodic materials with localised
imperfections are considered. Algorithms that borrow concepts from the theory of ho-
mogenisation and also from the domain decomposition methods, making use of the
periodic structure of the material as much of possible, are introduced and studied.

6.1 One-dimensional problem

Let us begin by considering a modification of the boundary value problem (4.1.2) stud-
ied earlier given by

 d
 d


 − dx a(x) dx u(x) = f(x), x ∈ (0, 1),
(6.1.1)


u(0) = 0,
u(1) = 0.

The coefficient of the differential equation is now the more general piecewise function
92 Hybrid approach

a1 (x), x ∈ [0, γ1 ],
a(x) = (6.1.2)
a2 (x), x ∈ (γ1 , 1],

where a2 is an -periodic function,   1 and

0 < α < a(x) < β, for x ∈ [0, 1]. (6.1.3)

We like to think of γ1 < 1/2, so that a = a(x) is periodic everywhere except on the
smaller interval [0, γ1 ]. In a broader setting, there are many examples of problems for
which the material’s microstructure is only really relevant on a localised subdomain.
This may happen either because the composition of the material differs in this local
region or the phenomenon that is being modelled is occurring there. In the remainder of
the computational domain, the material may be replaced by an equivalent homogenised
medium. The one-dimensional example we now consider should serve as illustration
for this type of problems.

6.1.1 The algorithm

As we have seen in the previous chapter, approximations for (6.1.1) can be found em-
ploying domain decomposition, either using substructuring or overlapping algorithms.
The latter produce some advantages, namely they are more generally applicable and
robust [17, 40]. Furthermore, they involve computing u and not its derivative du/dx
at the interfaces. Note that when the function a = a(x) oscillates rapidly, expressing
a complex microstructure, the derivative of u will change more abruptly than u itself.
Moreover, the function u is the primary unknown that we approximate using finite el-
ements and there is a natural loss of accuracy when du/dx is determined numerically.
Finally, there is another advantage in employing Schwarz methods to deal with prob-
lems with a localised relevant microstructure if in the remainder of the computational
domain a macroscopic approach is to be adopted. In that case we expect the overlap-
ping region to smooth out the effect of prescribing macroscopic boundary conditions to
the smaller microscopic regions.

In order to consider an overlapping scheme for (6.1.1) we introduce γ2 ∈ (0, 1) that is

such that γ1 < γ2 . As we have seen, the Schwarz procedure described in Algorithm 5.2
(k) (k)
iterates over [γ1 , γ2 ] such that sequences of functions {u1 }k and {u2 }k are generated.
These converge to u|[0,γ2 ] and u|[γ1 ,1] , respectively.

Naturally, the periodic behaviour of the coefficients of the differential equation over
[γ1 , 1] suggest that homogenisation can and should be employed over this interval. Do-
ing so prevents resolving the microscale unnecessarily, thus reducing the complexity
of the problem. The idea now is to decouple the problem into two smaller problems
and using homogenisation on the subproblem that allows it. This results in the hybrid
6.1 One-dimensional problem 93

approach expressed in the following algorithm, where we denote by bλ(0) the initial ap-
proximation for u(γ2 ) and Tol is the tolerance for the stopping condition. Moreover, a
is the effective coefficient associated to a2 and χ is the solution of the respective cell
problem, as described in Chapter 4.

Algorithm 6.1

Set k = 0. Given: bλ(0) , Tol.

1 - Solve



h i
d d (k+1)

 − a(x) u (x) = f(x), x ∈ (0, γ2 ),
 dx dx 1
b
(6.1.4)

 b (k+1)
u (0) = 0,
 1
 (k+1)
u1
b (γ2 ) = bλ(k) ,

and



h i

 − d
a(x) d
w (k+1)
(x) = f(x), x ∈ (γ1 , 1),
 dx dx

(6.1.5)

 b (k+1)
w(k+1) (γ1 ) = u (γ1 ),

 (k+1) 1
w (1) = 0.

2- Set

(k+1) dw(k+1)
w1 (x) := χ(x/) (x). (6.1.6)
dx

3- Solve

 d d

 
 − dx a(x) dx C(k+1) (x) = 0, x ∈ (γ1 , 1),

C(k+1) (γ1 ) = −w1
(k+1) (6.1.7)


(γ1 ),
 (k+1) (k+1)
C (1) = −w1 (1).

(k+1) (k+1)
4- Set u
b2 := w(k+1) + w1 + C(k+1) .

5 - Update bλ(k) and increment k

94 Hybrid approach

b (k+1)
bλ(k+1) = u
2 (γ2 ). (6.1.8)
k → k + 1. (6.1.9)

(k) (k)
6 - Return to step 1 until |b b 2 (γ2 )| < Tol.
u1 (γ2 ) − u

(k) (k)
This algorithm generates sequences of approximation functions {b u1 }k and {bu2 }k to
(k)
u|[0,γ2 ] and u|[γ1 ,1] , respectively. Note that u
b 2 is the homogenised corrected solution
corresponding to the problem that reads


 d d
  
 − dx a2 (x) dx w(k) (x) = f(x), x ∈ (γ1 , 1),

b (k) (6.1.10)

 w(k) (γ1 ) = u 1 (γ1 ),
 (k)
w (1) = 0.

It is possible to skip steps 2 and 3 of the algorithm and avoid computing the correctors.
In that case, step 4 should be replaced by setting u b (k+1)
2 := w(k+1) . This approach will
then become less complex, but also less accurate.

6.1.2 Behaviour of the error

We can now analyse the behaviour of the error associated to the previous iterative
scheme. Let us denote the error of the approximations u b (k)
1 and ub (k) b(k) := u
2 by E1 b (k)
1 −
(k) (k) b(k)
u|[0,γ2 ] and E2 := u2 − u|[γ1 ,1] , respectively. Clearly for k ≥ 1 the function E
b b 1
satisfies



h i
d d b (k)

 − a(x) E (x) = 0, x ∈ (0, γ2 ),
 dx dx 1

(6.1.11)

 b(k) (0) = 0,
E
 1
 b(k)
E (γ2 ) = E b(k−1) (γ2 ),
1 2

where we define Eb(0) (γ2 ) as the error of the initial guess at x = γ2 . As for the function
2
b , it can be written as the sum of E(k) with e(k) , where
E
(k)
2 2 H



h i
d d (k)

 − a
(x) E (x) = 0, x ∈ (γ1 , 1),
 dx 2 dx 2

(6.1.12)


(k) b(k) (γ1 ),
E2 (γ1 ) = E

 (k)
1
E2 (1) = 0,
6.1 One-dimensional problem 95

(k)
and eH := u b (k)
2 − w
(k)
is the homogenisation error at step k. Recall that w(k) is the
solution of (6.1.10). Now take a fixed value for k. Assume for the sake of simplicity that
b(k−1) (γ2 ), the error arising of the previous iteration, is positive. The more general case
E 2
follows a similar proof. From the differential equations in (6.1.11) and (6.1.12) it can be
seen that both E b(k) and E(k) are strictly monotonous functions. They are illustrated in
1 2
Figure 6.1. Note that E b(k) (γ1 ) = E(k) (γ1 ) and E
b(k) (γ2 ) = E
b(k−1) (γ2 ).
1 2 1 2

Êk1
Ek2

0
γ1 γ 1
2 x

Figure 6.1: Errors.

It is now easy to see that

(k) (k)
E2 (γ2 ) < E
b (γ2 ).
1

(k) (k) (k)

b (γ2 ) − E (γ2 ) > |e (γ2 )|, we have
Also, as long as E1 2 H

b(k+1) (γ2 ) = E
E b(k) (γ2 ) = E(k) (γ2 ) + e(k) (γ2 ) < E
b(k) (γ2 ).
1 2 2 H 1

In other words, at the end of the iteration step, the maximum of the approximation error
coming from the previous iteration, E b(k) (γ2 ), will be reduced to E
b(k+1) (γ2 ).
1 1

A natural stopping criterion for the hybrid approach algorithm would then be given by

(k) (k)
|E
b (γ2 ) − E (γ2 )| ≤ eH ,
1 2 (6.1.13)
96 Hybrid approach

where eH is the maximum of the error, in modulus, of the approximation yielded by the
homogenisation procedure. Indeed, until this condition is satisfied the error decreases at
each iteration step. Note, however, that Eb(k) (γ2 ) and E(k) (γ2 ) are not known. Therefore,
1 2
instead of (6.1.13) we will adopt the following stopping criterion

|E b(k) (γ2 )| ≤ Tol,

b(k) (γ2 ) − E (6.1.14)
1 2

which can also be written as

(k) (k)
|b b 2 (γ2 )| ≤ Tol.
u1 (γ2 ) − u (6.1.15)

We can now establish the following theorem that gives the error of the approximation
b(k) (γ2 ) − E
obtained when the quantities |E b(k) (γ2 )| decrease with k until (6.1.15) holds,
1 2
for a given Tol.

Theorem 6.1.1 Let p be the smallest natural number such that (6.1.15) holds. Then the ap-
proximation


(p) b (p)
u 1 (x), x ∈ [0, γ1 ],
u (x) := (6.1.16)
b (p)
b
u 2 (x), x ∈ (γ1 , 1],

for the exact solution u of (6.1.1) with (6.1.2) is such that

Rγ1
1/a(u)du
kb
u (p)
− uk∞ ≤ R0γ2 (eH + Tol) + eH . (6.1.17)
γ1
1/a(u)du

Proof: For each iteration, the maximum of the function


b(p) (x), x ∈ [0, γ1 ],
E
E(p) (x) := 1
(p) (6.1.18)
E2 (x), x ∈ [γ1 , 1],

is attained for x = γ1 . Now, since

b(p) (x) = E(p) (x) + e(p) (x) ≤ E(p) (x) + eH ,

E2 2 H 2

the maximum of E(p) added to eH gives an upper bound for the error function
6.1 One-dimensional problem 97


b(p) (x), x ∈ [0, γ1 ],
E
b(p)
E (x) := 1 (6.1.19)
b(p) (x), x ∈ (γ1 , 1].
E2

Consequently,

(p)
u(p) − uk∞ ≤ E2 (γ1 ) + eH .
kb (6.1.20)

(p)
In order to estimate the value of E2 (γ1 ), we proceed like in Theorem 5.1.1. From
(6.1.11) and (6.1.12) it can be easily seen that there exist constants ci and di such that we
have

Zx Zx
b(p) (x) = c1 1 (p) 1
E1 du + d1 and E2 (x) = c2 du + d2 .
0 a(u) γ1 a(u)

b(p) and E(p) on x = 0

To determine these constants we use the boundary conditions for E1 2
and x = 1 respectively, and also the conditions

b(p) (γ1 ) = E(p) (γ1 ) and E

E b(p) (γ2 ) = E(p) (γ2 ) + eH + Tol.
1 2 1 2

(p)
The latter equality ensures that we compute an upper bound for E2 (γ1 ). The result
then follows from (6.1.20).

An estimate for the error in which the size of the overlapping region plays an explicit
role is given in the following result.

Corollary 6.1.2 Under the conditions of the previous theorem, an upper bound for the error is

Rγ1
β 1/a(u)du
u(k) − uk∞ ≤
kb 0
(eH + Tol) + eH .
γ2 − γ1

Proof: It follows from (6.1.17) and (6.1.3).

The latter inequality tells us that the size δ := γ2 − γ1 of the overlapping region should
be chosen as large as possible in order to minimise the error at the end of the iterative
process. The drawback is that choosing a bigger value for δ implies at each iteration
step a bigger computational effort.
98 Hybrid approach

On the other hand producing a good choice for the value of Tol is not trivial. The quan-
(k)
tity |b b (k)
u1 (γ2 ) − u 2 (γ2 )| will initially decrease as k increases, but at a certain point
the homogenisation error may prevent it from decreasing further and it might even
(k)
increase. The iterative process can then be stopped after |b u1 (γ2 ) − ub (k)
2 (γ2 )| starts
increasing and the approximation we seek taken from the previous iteration. Alterna-
tively, a value for Tol may be prescribed, which might be useful to avoid having too
many iteration steps. In any case, the result from the previous theorem will then give
an upper bound for the error. In Section 6.3 we will present a couple of examples where
we estimate eH and simply take Tol = eH /2.

Finally, we note that the error of this iterative process depends a great deal on the ac-
curacy of the homogenisation procedure. It may then be wise to improve the quality of
the homogeneous solution by employing correctors, though this is likely unnecessary
when  is very small.

6.2 The elasticity problem

Let us consider a composite plate with linear elastic constituents. When these are finely
mixed, computing the displacement vector field for this plate is rather complex, requir-
ing the adoption of appropriate numerical techniques. In particular, if the material is
periodically distributed, this problem is expressed by (4.2.2) and so the theory of Chap-
ter 4 may be employed. For the more general case, one may recur to domain decom-
position methods that cut up the original problem into smaller and more manageable
problems. In what follows we adapt and extend the results of the previous section to
the problem of linear elasticity given by



 −∇ · (A(x)(u)) = f, x ∈ Ω,


u = 0, x ∈ ΓD , (6.2.1)



 σ(u) · n = ϕN , x ∈ ΓN ,

where the domain Ω is partitioned into non-overlapping subdomains Ω1 and Ω2 such

that Ω = Ω1 ∪ Ω2 and

A1 (x), x ∈ Ω1 ,
A(x) := (6.2.2)
A2 (x), x ∈ Ω2 .

Here A2 is an -periodic tensor and it is assumed that |Ω1 | < |Ω2 |. Note that the ho-
mogenisation method described earlier may not be employed to solve (6.2.1) with (6.2.2)
as A is not periodic everywhere. On the other hand, using domain decomposition tech-
niques has the drawback of not making use of the periodicity of the elasticity tensor over
6.2 The elasticity problem 99

Ω2 . We then proceed as in Section 6.1 and establish a hybrid approach for this problem
where again we combine homogenisation and domain decomposition techniques.

We will introduce a sequence of problems defined on the two overlapping subdomains

Ωb 1 and Ω2 , where Ω1 ⊂ Ω
b 1 ⊂ Ω. Let Γ1 = ∂Ω1 ∩ ∂Ω2 and Γ2 = ∂Ω
b 1 ∩ Ω2 , as illustrated
in Figure 6.2.

Overlapping
region
G2
G2 G1
W2
G1 W^ 1

Figure 6.2: Computational domain Ω, the union of the overlapping subdomains Ω

b 1 and
Ω2 .

(0)
The algorithm expressing the hybrid approach for elasticity is as follows. Let b λ be the
initial guess for u|Γ2 and Tol = (Tol1 , Tol2 ) the tolerance for the stopping condition.

Algorithm 6.2

(0)
Set k = 0. Given: b
λ , Tol.

1 - Solve

 (k+1)

 −∇ · (A(b u1 )) = f, x ∈ Ω
b 1,





 u
b(k+1) = 0,
T b
x ∈ ΓD ∂Ω
1 1,
(k+1)
) · n = ϕN ,
T b
x ∈ ΓN ∂Ω (6.2.3)


σ( u 1 1,

b




 (k+1) b (k)
u
b1 =λ , x ∈ Γ2 ,

and
100 Hybrid approach



 −∇ · (A(w(k+1) )) = f, x ∈ Ω2 ,




 w(k+1) = 0, T
x ∈ ΓD ∂Ω2 ,
(6.2.4)
 (k+1) T
 σ(w ) · n = ϕN , x ∈ ΓN ∂Ω2 ,




 (k+1) b(k+1)
w = u1 , x ∈ Γ1 .

2- Set

1 X
 
∂wi ∂wj x
w1 (x) = + χij ( ). (6.2.5)
2 ∂xj ∂xi 
i,j=1,2

3- Solve



 −∇ · (A(C(k+1) )) = 0, x ∈ Ω2 ,


C(k+1) = −w1 , x ∈ ΓD ∩ ∂Ω2 , (6.2.6)



 σ(C
(k+1)
) · n = 0, x ∈ ΓN ∩ ∂Ω2 .

(k+1) (k+1)
4- Set u
b2 := w(k+1) + w1 + C(k+1) .

(k)
5 - Update b
λ and increment k

(k+1)
λ
b b(k+1)
=u2 | Γ2 , (6.2.7)
k → k + 1. (6.2.8)

6 - Return to step 1 until

(k) (k)
u1 |Γ2 − u
k(b b2 |Γ2 )i k∞ < Toli , for i = 1, 2. (6.2.9)

(k) (k) (k)

u1 |Γ2 − u
Here, (b b2 |Γ2 )i represents the ith component of the vector function (b
u1 |Γ2 −
(k)
b2 |Γ2 ).
u

We note that this algorithm may be significantly simplified if the accuracy requirements
are not too demanding. The iteration steps 2 to 4, where correctors are employed to
improve the accuracy of the overall procedure, may be replaced by simply having
6.3 Numerical results 101

b2(k+1) := w(k+1) . In any case the error originating from this procedure will depend
u
on the accuracy of the homogenised solution, which improves for smaller values of .

6.3 Numerical results

We conclude this chapter with a discussion of the implementation of the algorithms

presented earlier. A one-dimensional example related to Algorithm 6.1 is included. The
underlying differential equation has periodic coefficients everywhere except for a small
portion of the computational domain. A layered material with localised inclusions is
also analysed, for which Algorithm 6.2 is employed.

6.3.1 One-dimensional example

Let us consider the problem (6.1.1) with


2 + 1.9 cos(2πx/0.03), x ∈ [0, γ1 )
a(x) = −1 , f(x) = 1,  = 0.1, γ1 = 0.1. (6.3.1)
2 + sin( 2πx

 ) , x ∈ [γ1 , 1]

For future convenience we start by computing a reference solution u for this problem,
using quadratic finite elements on a very fine mesh.

An approximation for u can be determined using the hybrid approach described in

Algorithm 6.1 that takes the -periodic behaviour of a|[γ1 ,1] into account. This will
generate the sequence of approximations {b u(k) }k defined in (6.1.16), as we iterate over
the overlapping region [γ1 , γ2 ], for some given γ2 that satisfies 0 < γ1 < γ2 < 1.
In order to start the iterative process, an initial approximation bλ(0) for u(γ2 ) must be
provided. For this we consider the auxiliary problem (6.1.1) with

1
a(x) := a (x) = , for x ∈ [0, 1], f(x) = 1,  = 0.1. (6.3.2)
2 + sin( 2πx
 )

This is a modification of the original problem (6.1.1) with (6.3.1) obtained by disregard-
ing the behaviour of the coefficients a = a(x) in (6.3.1) over the smaller interval [0, γ1 ].
Instead, -periodic coefficients are considered throughout Ω. This periodic problem
had already been considered in Chapter 4. There we used the homogenised solution u
and the homogenised corrected solution u + u1 + C as approximations for u , cf. Table
4.1. In this setting, we may use either u or u + u1 + C as cheap initial guesses for the
solution u of (6.1.1) with (6.3.1). The L2 and maximum norms of the errors E := u − u
and EC := (u + u1 + C) − u are displayed in Table 6.1.
102 Hybrid approach

k · k∞ k · kL2 (0,1)
E 2.2E − 2 1.2E − 2
EC 2.7E − 2 1.3E − 2

Table 6.1: Error of the approximations.

We now turn our attention again to (6.1.1) with (6.3.1) and employ Algorithm 6.1, tak-
ing bλ(0) = u(γ2 ). Approximations for the solutions of the problems (6.1.4), (6.1.5) and
(6.1.7), as well as the cell problem, are computed using quadratic finite elements. A finer
uniform grid with grid size h = 1E − 4 is required to solve the problem on [0, γ1 ]. As
for the other problems, the grid size is taken to be h = 1E − 1. Finally, note that w1
depends on components of the cell function and on derivatives of components of the
homogenised function w. The latter are determined by numerical differentiation.

k · k∞ k · kL2 (0,1) k Upper bound

E 1.5E − 2 5.7E − 3 2 5.0E − 2
EC 7.9E − 4 3.9E − 4 6 1.4E − 3

Table 6.2: Error arising from the hybrid procedure for γ2 = 0.15.

k · k∞ k · kL2 (0,1) k Upper bound

E 1.4E − 2 5.2E − 3 2 2.9E − 2
EC 5.6E − 4 2.8E − 4 4 9.7E − 4

Table 6.3: Error arising from the hybrid procedure for γ2 = 0.25.

k · k∞ k · kL2 (0,1) k Upper bound

E 1.4E − 2 4.7E − 3 2 2.2E − 2
EC 4.8E − 4 2.5E − 4 4 8.0E − 4

Table 6.4: Error arising from the hybrid procedure for γ2 = 0.35.

Tables 6.2 to 6.4 display the errors obtained by employing the procedure above for var-
ious sizes of the overlapping region, with the parameter γ2 taking the values 0.15, 0.25
and 0.35, respectively. On the second lines of the tables, E gives the error when simple
homogenisation is applied on the domain [γ1 , 1], i.e., when steps 2 and 3 of the algo-
rithm are skipped. As for EC , it corresponds to the error obtained by improving the
6.3 Numerical results 103

homogenised solution with the first order corrector and the boundary corrector. The
fourth column of the tables displays the number of iterations k required to satisfy the
stopping condition (6.1.15) with Tol = eH /2, and the last column the theoretical upper
bound given by (6.1.17). Estimates for the maximum error eH of the homogenisation
procedure were taken from Table 4.1. In Figure 6.3 we show the exact solution of (6.1.1)
with (6.3.1) in black dots, the initial approximation u as a dotted line and the hybrid so-
lution given by Algorithm 6.1, employing the correctors, as the full gray line that nearly
coincides with the exact solution.

Exact solution
0.25 Initial guess
Homogenised
with correction

0.2

0.15

0.1

0.05

0
0 0.2 0.4 0.6 0.8 1
x

Figure 6.3: Exact and approximated solutions.

As one can see, the accuracy of the results depends on that of the homogenisation pro-
cedure. In practice,  should be much smaller than 0.1, the value that we took for this
example. This will lower the homogenisation error and improve the overall results fur-
ther.

6.3.2 Layered elastic materials

In what follows we present an example related to the elasticity problems introduced in

Section 6.2. This demonstrates the capability of the hybrid approach to deal with mate-
rials with localised imperfections. We will consider a problem in which the material is
periodic everywhere except for a small portion of the domain, where some inclusions
are added. As a first step, we disregard the inclusions.
104 Hybrid approach

Let us then begin by considering the problem (6.2.1) with Ω = [0, 1] × [0, 1] and as-
sume that the underlying material is isotropic. This means that the components of the
elasticity tensor A
e can be written as

E(y) E(y)ν(y)
a
e2222 (y) = a
e1111 (y) = ; a
e2211 (y) = ; (6.3.3)
1 − ν2 (y) 1 − ν2 (y)
E(y)
a
e2121 (y) = ; a
e2111 (y) = a
e2221 (y) = 0, (6.3.4)
2(1 + ν(y))

see Chapter 4. We further assume that the material is also layered and that its reference
cell Y = [0, 1] × [0, 1] can be decomposed into two subdomains Y1 = [0, 12 ] × [0, 1],
Y2 = [ 21 , 1] × [0, 1], see Figure 6.4. These are such that Yi is occupied by a linear elastic
material with Young’s modulus Ei = Ei (y1 ) and Poisson’s ratio νi = νi (y1 ), where
y = (y1 , y2 ) and i = 1, 2. We can then write

E(y) = E1 χ1 (y1 ) + E2 χ2 (y1 ), (6.3.5)

ν(y) = ν1 χ1 (y1 ) + ν2 χ2 (y1 ), (6.3.6)

where χ1 and χ2 are the characteristic functions of the sets Y1 and Y2 .

1

E1 E2
Y1 n1 n2 Y2
y2
y1
0 0.5 1

Figure 6.4: Reference cell Y, composed of two different materials.

Finally let

ν1 = ν2 = 0.3, E1 = 1, E2 = 3,  = 0.02, f = 0, (6.3.7)

and the boundary conditions be such that


 (0, 1), x2 = 1,
ΓD = {0} × [0, 1], ϕN (x) = (0, 0), x2 = 0, (6.3.8)

(1, 0), x1 = 1.
6.3 Numerical results 105

A reference solution u for (6.2.1) with (6.3.3)-(6.3.8) can be determined using finite
elements with a very fine mesh. What we want to do now is to approximate the solution
of this problem using homogenisation to obtain a cheap initial guess for the problem we
will consider next, where inclusions are present. Recall that the homogenised medium
corresponding to the isotropic layered material we have mentioned is orthotropic with
material constants given by

E
Ex = 2
, Ey = E, (6.3.9)
ν + AE
ν
νxy = 2
, νyx = ν, (6.3.10)
ν + AE
1
G= , (6.3.11)
(1 + ν1 )/E1 + (1 + ν2 )/E2

where

1 1 1 1 − ν21 1 − ν22
E= (E1 + E2 ), ν = (ν1 + ν2 ), A = ( + ). (6.3.12)
2 2 2 E1 E2

We use quadratic finite elements to compute the homogenised solution u. A uniform

square grid with 50 × 50 elements is employed. The norms of the horizontal and vertical
components of the error E := u − u of this approximation are displayed in the third
line of Table 6.5.

Horizontal component Vertical component

k · k∞ k · kL2 (0,1) k · k∞ k · kL2 (0,1)
E 1.2E − 1 2.2E − 2 1.9E − 1 6.2E − 2
EC 3.3E − 2 3.5E − 3 3.0E − 2 3.6E − 3

Table 6.5: Error arising from the homogenisation procedure.

In order to obtain a better approximation for u , we will make use of the first order
corrector u1 and the boundary corrector C. The former is given by

1 X
 
∂ui ∂uj x
u1 (x) = + χij ( ), (6.3.13)
2 ∂xj ∂xi 
i,j=1,2

where χij is the solution of the cell problem (4.2.7). The boundary corrector C can be
approximated by the solution of the following problem
106 Hybrid approach



 −∇ · (A(C)) = 0, x ∈ Ω,


C = −u1 , x ∈ ΓD , (6.3.14)



 σ(C) · n = 0, x ∈ ΓN .

Now, to compute u1 we first have to solve the cell problem. We do so by approximating

the piecewise constant function E by a continuous smooth function as in Chapter 4.
As before, we employ square quadratic finite elements. The derivatives of components
of the homogenised function u in (6.3.13) are determined by numerical differentiation.
As for the boundary corrector, it can be computed from (6.3.14) again by using finite
elements.

Exact
Homogenised
0.3
Homogenised
with correction

0.2

0.1

0
0 0.5 1
x1

Figure 6.5: Exact and approximated solutions for the horizontal component of the dis-
placement along x2 = 0.7.

The norms of the horizontal and vertical components of the error function EC := (u +
u1 +C)−u are displayed in the last line of Table 6.5. As expected, the use of correctors
gives better results than using the homogenised solution u as an approximation. In
Figure 6.5 we plot the horizontal components of the exact solution of the problem u ,
the homogenised function u and the homogenised corrected approximation u+u1 +C,
along x2 = 0.7.

Up to this point we considered a problem concerning an -periodic material plate.

We will now look into a modification of this problem where inclusions are present,
as originally intended. Consider (6.2.1) with (6.2.2), where the computational domain
Ω = [0, 1]×[0, 1] is split into the non-overlapping subdomains Ω1 = [0, 0.1]×[0, 0.1] and
Ω2 = Ω − Ω1 . The components of the elasticity tensor are assumed to satisfy (6.3.3)-
(6.3.6) with (6.3.7) throughout Ω except for some circular inclusions on Ω1 , see Figure
6.6. Furthermore, we assume that (6.3.8) holds.
6.3 Numerical results 107

0.1

P8
0.08 P7
P5
0.06
P6
P4
0.04
P2 x2
0.02 P1
P3
0

0 0.02 0.04 0.06 0.08 0.1

x1

Figure 6.6: Inclusions present on the subdomain Ω1 .

The inclusions consist of a linear elastic material characterised by Young’s modulus

E3 = 4 and Poisson’s ratio ν3 = 0.1. The radius of each inclusion is r = 0.01. They are
centred at the points

P1 = (0.02, 0.02), P2 = (0.05, 0.03), P3 = (0.08, 0.015),

P4 = (0.015, 0.05), P5 = (0.04, 0.07), P6 = (0.08, 0.055),
P7 = (0.015, 0.08), P8 = (0.07, 0.085).

We thus deal with a problem related to a periodic structure with localised imperfections.
A reference solution u for this problem can be determined using finite elements with a
very fine mesh.

To find an approximation for u, we employ the hybrid approach for elasticity expressed
by Algorithm 6.2. Let Ωb 1 = [0, 0.15] × [0, 0.15], cf. Figure 6.2. The algorithm iterates
over the overlapping region [0.10, 0.15] × [0.10, 0.15], generating a sequence of approx-
imations


(k) b(k)
u1 (x), x ∈ Ω1 ,
u (x) := (6.3.15)
b(k)
b
u2 (x), x ∈ Ω2 .

(0)
We take bλ := u|Γ2 , where u is the homogenised solution computed previously. At
each iteration step, finite elements and numerical differentiation are used, adapting the
procedure described for the one-dimensional example.

The third line of Table 6.6 displays the norms of the error E0 := u − u of the approxima-
tion which we give as an initial guess. On the fifth line of the table we show the norms of
108 Hybrid approach

Horizontal component Vertical component Iterations

k · k∞ k · kL2 (0,1) k · k∞ k · kL2 (0,1) k
E0 1.7E − 1 2.8E − 2 2.3E − 1 7.5E − 2 −
E 8.2E − 2 8.5E − 3 1.0E − 1 2.7E − 2 5
EC 3.8E − 2 5.4E − 3 3.9E − 2 5.8E − 3 7

Table 6.6: Error arising from the hybrid procedure.

the approximations obtained by using the hybrid approach as described. Alternatively,

steps 2 to 3 of Algorithm 6.2 may be skipped, avoiding having to compute correctors.
We then obtain the results shown in the fourth line of Table 6.6.
Chapter 7

Fracture of composites

In a linear elastic, homogeneous and isotropic material, the stress and displacement
fields around a crack tip are characterised by the SIFs, as was seen in Chapter 3. Once
these parameters have been determined, it is possible to predict the future behaviour of
a crack, namely whether it will propagate and if so in what direction.

For composite materials finding the path of a pre-existent crack is a very challenging
problem. Not only the interaction of the crack with the heterogeneities has to be ac-
counted for, but already solving the underlying elasticity problems is computationally
very complex. It requires appropriate numerical techniques such as those introduced in
Chapters 4 to 6.

In this chapter the effect of the local structure on the process of crack propagation oc-
curring on a highly heterogeneous composite plate is analysed. The evaluation of stress
intensity factors using the hybrid approach, which leads to accurate and computation-
ally feasible solutions of the elasticity problems, is discussed. Finally, an algorithm to
determine the crack path is proposed.

7.1 Behaviour of the SIFs

On the microscopic level the propagation of a given crack on a composite linear elastic
material is influenced by aspects like the orientation of the material constituents and the
existence of internal defects such as microcracks, voids and inclusions. Mathematically
this can be seen in the way the SIFs vary according to the constitution of the vicinity of
the crack tip.
110 Fracture of composites

7.1.1 Effects of the local structure

In what follows we investigate how the local structure of a plate influences the SIFs of
a pre-existent static crack. To do this consider a cracked plate Ω = ΩP − ΩC , where
ΩP = [−0.5, 0.5] × [−0.5, 0.5] and ΩC is the crack line [−0.5, −0.3] × {0} of length a = 0.2,
see Figure 7.1.

0.5

x2
a
0
O x1

-0.5

-0.5 -0.3 0.5

Figure 7.1: Cracked plate.

Assume that the plate is composed of a periodically layered elastic material with isotropic
homogeneous components, its behaviour being modelled by (4.2.2). The components of
the related elasticity tensor A
e then read

E(y) E(y)ν(y)
a
e2222 (y) = a
e1111 (y) = , a
e2211 (y) = , (7.1.1)
1 − ν2 (y) 1 − ν2 (y)
E(y)
a
e2121 (y) = , a
e2111 (y) = a
e2221 (y) = 0. (7.1.2)
2(1 + ν(y))

Let the plate be pulled at its upper and lower edges. The remaining boundaries, includ-
ing the crack edges, are considered to be stress-free, so that the following conditions
hold


 (0, 1), x2 = 0.5,
ϕN (x) = (0, −1), x2 = −0.5, (7.1.3)

(0, 0), otherwise.

We assume that the reference cell for this material, given by Y = [0, 1] × [0, 1], is the
union of the two subdomains Y1 = [0, 0.5] × [0, 1], Y2 = [0.5, 1] × [0, 1], composed
respectively by materials A and B, see Figure 7.2. For i = 1, 2, Yi is occupied by a
material with Young’s modulus Ei and Poisson’s ratio νi and we have
 7.1 Behaviour of the SIFs 111

E(y) = E1 χ1 (y1 ) + E2 χ2 (y1 ), (7.1.4)

ν(y) = ν1 χ1 (y1 ) + ν2 χ2 (y1 ), (7.1.5)

where χ1 and χ2 are the characteristic functions of the sets Y1 and Y2 .

1
Material A

Material B

Y1 Y2
0 0.5 1

Figure 7.2: Reference cell Y, composed by materials A and B.

Finally let

ν1 = 0.3, ν2 = 0.1, E1 = 10, E2 = 1,  = 0.02, f = 0. (7.1.6)

The problem we have described up to now has been set up so that the material layers
are orthogonal to the crack line, see Figure 7.3 e). Other material orientations can also be
considered. In particular, we look into the five situations depicted in Figure 7.3, repre-
senting zoom-ins of the vicinity of the crack tip. The figures labelled a) to e) correspond
to various material orientations characterised by the angle θ0 that the layer boundaries
form with the horizontal axis. We assume that the crack tip lays inside material A.
5

0.05
05
0.0

0.

-0.05
1
0.
0.1

q0 q0

-0.3 -0.3 -0.3

Figure 7.3 - a) θ0 = 0o b) θ0 = 30o . c) θ0 = 45o .

112 Fracture of composites

Material A
5
0.0 Material B

0.1
q0 q0 0.1

-0.3 -0.3

o
d) θ0 = 60 . e) θ0 = 90o .

We now want to compute the mode I and mode II stress intensity factors. Once this has
been done, we may also obtain the propagation angle θP for each of these configurations
using the maximum circumferential tensile stress criterion. We recall that this is the an-
gle under which the crack will propagate if the loading is large enough for crack growth
to occur. Before we can compute the SIFs, we have to solve the elasticity problems (4.2.2)
with (7.1.1)-(7.1.6) for the various values of θ0 . For that we employ the finite element
method. A fine mesh with quadratic triangular elements in the vicinity of the crack tip
and quadratic rectangular elements everywhere else is generated. This allows for the
computation of the displacement and stress fields for the different material orientations.
The SIFs KI and KII are then determined using the J-integral method and their values
are displayed in Table 7.1. As for the propagation angle θP also included in the table, it
is zero when θ0 = 0o or θ0 = 90o . The other cases are computed using (3.3.10).

θ0 KI KII θP
0o 2.20 0 0o
30o 2.13 −2.23E − 2 1.2o
45o 2.10 −5.22E − 2 2.8o
60o 2.11 −8.20E − 2 4.4o
90o 2.13 0 0o

Table 7.1: SIFs and propagation angles. The material surrounding the crack tip is char-
acterised by E1 and ν1 .

Next we interchange the roles of materials A and B. We assume they are now charac-
terised by E2 and ν2 and by E1 and ν1 respectively. We solve the previous problem with
this new assumption. The values we find are shown in Table 7.2.

By comparing the results in the previous two tables, it can be seen that the local structure
can influence the value of the SIFs quite dramatically. Indeed, if we take for instance the
values of KI and KII along the second and third columns of Table 7.1 or of Table 7.2 we
7.1 Behaviour of the SIFs 113

θ0 KI KII θP
0o 8.16E − 1 0 0o
30o 7.75E − 1 7.18E − 2 −10.4o
45o 7.10E − 1 1.07E − 1 −16.4o
60o 6.24E − 1 1.21E − 1 −20.5o
90o 5.39E − 1 0 0o

Table 7.2: SIFs and propagation angles. The material surrounding the crack tip is char-
acterised by E2 and ν2 .

see that the SIFs vary with the material orientation. However, there is a greater differ-
ence when we compare the situations in which the roles of the materials A and B are
interchanged. The values of the KI and KII change dramatically when we compare the
two tables for the same values of θ0 . This sensitivity of the SIFs to the material structure
is particularly relevant for composites or other materials that are very heterogeneous.

We note that we dealt with a heterogeneous periodic plate where the period  of the
heterogeneities was rather large. For that accurate numerical approximations of the
solutions of the elasticity problems and consequently of the SIFs can be obtained by
the finite element method and taken as reference solutions. When  is much smaller, the
typical mesh width h must satisfy h   to yield reasonable results and so prohibitively
thin meshes have to be employed. Alternatively, we may adopt a macroscopic approach
and replace the heterogeneous medium by a fictitious equivalent homogeneous mate-
rial. In particular, instead of the layered material with vertically disposed layers for
which (7.1.4) and (7.1.5) hold, one may consider the orthotropic material characterised
by the elasticity tensor A with components

Ex Ex νyx
a1111 = , a2211 = , (7.1.7)
1 − νxy νyx 1 − νxy νyx
Ey
a2222 = , a2121 = Gxy , (7.1.8)
1 − νxy νyx
a2111 = a2221 = 0, (7.1.9)

where νxy Ey = νyx Ex , cf. Chapter 4. Homogenisation can also be applied for the re-
maining material orientations other than for θ0 = 90o , by rotating the axes of orthotropy.

The advantage of the homogenisation procedure lays in the very significative simplifica-
tion of the underlying elasticity equations, as the -dependent coefficients are replaced
by constants. However this comes at the price of losing the effects of locality. Indeed,
homogenisation intends to provide the average behaviour of a material in some sense,
114 Fracture of composites

not its local behaviour, so it is not as accurate. Moreover, we note that if we are inter-
ested in computing the SIFs and propagation angles for the homogenised material it
should be taken into account that the homogenised material is orthotropic. In particular,
the maximum circumferential tensile stress criterion should be reformulated, cf. [64].

r
P

Figure 7.4: Inclusion near the crack tip.

Besides the material orientation, the behaviour of the SIFs is also affected by the pres-
ence of defects, as we will illustrate in the following example. Consider again the
cracked plate represented in Figure 7.1, and let it now be composed of a linear elas-
tic isotropic homogeneous material with E = 1 and ν = 0.1. We assume that it is subject
to the boundary conditions (7.1.3), so that a mode I situation arises. Using the finite
element method to compute the SIF for this configuration, we find that KI = 1.18.

Now consider the same plate but containing one linear elastic circular inclusion of ra-
dius 0.05, centred at the point P and characterised by E2 = 10 and ν2 = 0.3, cf. Figure
7.4. We take various coordinates for P as listed in Table 7.3, so that the SIF can be mea-
sured when the inclusion is closer or further from the crack tip, located at (−0.3, 0).

P d KI
(−0.2, 0) 0.05 1.11
(−0.1, 0) 0.15 1.16
(−0.2, 0.2) 0.17 1.14
(−0.1, 0.2) 0.23 1.15
(0, 0) 0.25 1.17
(0.2, 0) 0.45 1.18

Table 7.3: SIF affected by the proximity of an inclusion.

When the inclusion is at a larger distance d from the crack tip, the SIF is about the
same as the one computed for the plate without any inclusion. Conversely, this fracture
parameter is affected by a closer proximity of the inclusion, see Table 7.3.
7.1 Behaviour of the SIFs 115

This example illustrates the sensitivity of the stress intensity factors to the local char-
acteristics of the surroundings of the crack tip. Many other authors have also studied
this phenomenon, see for example [24, 84, 85]. If one looks for accurate approximations
for the SIFs, the local structure in the vicinity of the crack tip may not be disregarded.
As we move away from this region, the heterogeneities have much less influence on the
crack behaviour.

7.1.2 Computational aspects

The first step in predicting the behaviour of a pre-existing crack in a composite plate
lays in solving the related elasticity problem. Doing so using the finite element method
without disregarding the heterogeneities is usually too demanding in terms of the com-
putational complexity.

To deal with this problem we would like to have the best of both worlds and design
a method that is both accurate - especially in the vicinity of the crack tip - and com-
putationally feasible. This is the goal of the hybrid approach presented in the previous
chapter. It is a flexible method that allows the region around the crack tip to be treated
separately. With this procedure the computational complexity of the original problem
is significantly reduced, as the heterogeneities are only resolved where it is relevant to
do so.

In what follows we will apply the hybrid approach to a cracked elastic plate Ω = ΩP −
ΩC , where ΩP = [−1, 1] × [−1, 1] and ΩC is the closed crack line [−a, a] × {0}.

x2=1
E1
n1
E2
Symmetry
n2
line x1=0

Symmetry x1=1 2a
line x2=0

Figure 7.5: Cracked plate (left) and zoom-in around crack region (right).

We assume that the plate is composed of a layered material with isotropic components
such that (4.2.2) and (7.1.4)-(7.1.5) hold. The elasticity tensor A
e = A(y)
e depends only on
o
y1 so that the layers are disposed at a 90 angle with the horizontal axis. Besides this,
we take
116 Fracture of composites

ν1 = ν2 = 0.3, E1 = 3, E2 = 1,  = 2.5E − 2, f = 0. (7.1.10)

Let the plate be pulled at its upper and lower edges while it remains free of stress along
the other boundaries so that we have the following boundary conditions


 (0, 1), x2 = 1,
ϕN (x) = (0, −1), x2 = −1. (7.1.11)

(0, 0), otherwise.

Due to the underlying symmetry of the elasticity problem, it suffices to consider the
domain [0, 1] × [0, 1] − ([0, a] × {0}) represented in Figure 7.5.

Consider the values for a, half of the crack length, included in the first line of Table
7.4. For each of these values a solution is found for the elasticity problem employing
finite elements, with very fine meshes. We take these as reference solutions. Using the
J-integral method the mode I stress intensity factor KI is computed. It is included in the
second line of the table.

SIF \ a 2.5E − 3 5.0E − 3 7.5E − 3 1.0E − 2

KI 1.33E − 1 1.93E − 1 2.46E − 1 3.24E − 1
Kha
I 1.31E − 1 1.90E − 1 2.47E − 1 3.18E − 1

Table 7.4: SIF for various crack lengths.

Next, we look for an approximation of the SIF applying the hybrid approach. The
computational domain ΩCp = [0, 1] × [0, 1] is split into the overlapping subdomains
b 1 = [0, 0.15] × [0, 0.15] and Ω2 = ΩCp − ([0, 0.1] × [0, 0.1]). We compute the elasticity
Ω
problem by Algorithm 6.2, as in the previous chapter, and determine the stress intensity
factors Kha
I with the J-integral method. These values are displayed in the last line of
Table 7.4.

This example illustrates the computational capabilities and the accuracy of the hybrid
approach. We note that similarly accurate results could have been obtained without
resorting to domain decomposition as in Algorithm 6.2. Indeed, whilst in the vicinity of
the crack tip the original differential equations must be considered, the heterogeneous
material can be replaced by an equivalent homogenised material in the remainder of
the plate. This gives rise to a problem that may be solved by the finite element method
employing a mesh that is finer near the crack tip. In the case of the problem we have
just considered such an approach is computationally feasible. It is then a good idea to
use it because it is easier to implement than Algorithm 6.2.

Note however that if we had more than one crack or if we wanted the subdomain where
7.2 Crack paths in layered materials 117

the microscale structure must be resolved to be handled separately then Algorithm 6.2
would have to be used without suppressing domain decomposition.

7.2 Crack paths in layered materials

The problem of the growth of a pre-existent crack on a homogeneous linear elastic plate
under plane stress conditions was addressed earlier in Chapter 3. The crack path was
determined adopting the incremental approach proposed in Algorithm 3.1, with the di-
rection of propagation being computed at each iteration for the updated plate geometry.

In this section we want to extend the previous analysis to deal with the problem of a
crack propagating on a linear elastic plate under plane stress conditions, but now con-
stituted of a highly heterogeneous composite material. This implies that a growing crack
will likely interact with more than one material and also with the respective interfaces.
At a given moment, we can then distinguish between the situations when the crack tip
region is contained inside one of the components of the composite and when the tip is
at an interface.

Naturally, these two situations are not likely to be static. A crack that has its tip at
the interface may be directed away from it or even continue propagating between the
materials, i.e., delaminating. Also, any crack propagating through a material component
may encounter an interface. When this happens, several situations are possible. The
crack may penetrate into the adjacent material, eventually suffering a change in angle,
reflect back to the material it was originally propagating in or be deflected along one or
both sides of the interface so that delamination occurs, see Figure 7.6 and cf. [29, 59, 75].

Material 1

(c)
Material 2

q0

(b)
(a)

Figure 7.6: The crack (full line) may penetrate (a), be reflected back (b), or deviate along
the interface (c).

For reasons of simplicity we will assume that the components of the composite are well
bonded so that no delamination may occur in what follows. We will also disregard the
118 Fracture of composites

possibility of branching, creation of new cracks or the existence of flaws in the plate as
well as of interfaces between more than two materials. One may e.g. consult [9, 24, 63]
for this.

Under these assumptions, for a crack present in a linear elastic plate with isotropic ho-
mogeneous constituents, the stress and displacement fields are characterised by (3.2.5)
and (3.2.14) when the crack tip is laying inside one of the constituents. This is no longer
the case when the crack tip is positioned at an interface. Let us consider the situation
depicted in Figure 7.6 where the crack penetrates into the adjacent material. The order
of the stress singularity is then characterised by the solution λ of the equation

α + β2 − (1 − β)(α − β)(1 − cos(2θ0 ))λ2


2
+(1 − β2 )(π − 2θ0 )cos(λπ)cos(λ)
+(1 − β2 )(π − 2θ0 ) sin2 (λ) (1 − β2 ) cos2 (λπ) + β2 − α2 = 0,
 
(7.2.1)

where α and β are the so-called Dundurs bimaterial parameters, cf. [22, 27, 59]. Let us
denote by σm m
θθ and σrθ the tangential and radial stress fields defined over the material
m, m = 1, 2, where the axes are positioned as represented in Figure 3.3. Then when λ is
a real number these fields satisfy

σm m
θθ + iσrθ = r
λ−1 m
F (θ, α, β, KI , KII ). (7.2.2)

When λ = Re(λ) + i Im(λ) is complex, they read

σm m
θθ + iσrθ = r
Re(λ)−1 m
G (θ, α, β, KI , KII ). (7.2.3)

The complex functions Fm and Gm can be found for example in [19], where the SIFs KI
and KII are also defined for a crack terminating at the interface.

We are now ready to predict the propagation of a crack in a composite material. When
the crack growth occurs far from any interface, the incremental approach of Algorithm
3.1 is applicable. This implies solving an elasticity problem at each iteration using the
finite element method, finding approximations for the SIFs and computing the propa-
gation angle from these by using the maximum circumferential tensile stress criterion.
The crack is then incremented, i.e., the geometry is updated by subtracting the portion
of new crack created.

When the crack tip is at the interface, the stress fields display the more complex be-
haviour described by (7.2.2) or (7.2.3). We assume that the maximum circumferential
tensile stress criterion can again be used to predict the direction of propagation. We
approximate this direction by averaging the amplitudes of a number of angles, each
7.2 Crack paths in layered materials 119

maximising σθθ on a small circle centred at the crack tip. The stress fields are again
computed by the finite element method.

It is important to note that in the region near an interface the accuracy of the computed
stress fields is less than elsewhere, cf. [84]. To deal with this, our strategy is to adjust
the size of the new crack increment when the crack tip approaches this critical region.
The propagation angle is kept unchanged and the updated crack tip position lays at the
interface. This seems to be a good idea for situations when the angle θ0 represented
in Figure 7.6 is not expected to be close to 0o or 180o . If the crack may grow nearly
parallel to the interface line, a different approach should be used. Not only small errors
in the propagation angle may lead to very different updated crack tip positions, but
also cracks parallel to layers behave differently from those that propagate through the
layers, see for example [84] and references within.

The procedures we have described can be structured in the following algorithm for a
composite plate with a pre-existent crack. We denote by ∆a and δR the length for the
crack increments and the minimum distance between the crack tip and the interface for
(n)
which we assume the stress fields are accurately computed. Moreover, let xtip denote
the crack tip coordinates with respect to a fixed coordinate system at the iteration step
n. Finally, θ(n) is the angle between the crack segment containing the crack tip and the
x1 -axis of the coordinate system.

Algorithm 7.1

(0)
Set n = 0. Given: ∆a, δR , xtip , θ(0) .

1- Solve the elasticity problem.

(n)
2- If xtip is at an interface, compute σθθ using (3.3.6). Otherwise skip to step 4.

3- Compute the value of the propagation angle θP using the maximum circumferential
tensile stress criterion. Skip to step 6.

4- Compute the values of the stress intensity factors KI and KII .

(K)
5- Determine the angle of propagation θP = θP using equation (3.3.10).

6- Update the crack angle and the crack tip coordinates

θ(n+1) := θ(n) + θP , (7.2.4)

(n+1) (n)
xtip := xtip + ∆a(cos(θ(n+1) ), sin(θ(n+1) )). (7.2.5)
120 Fracture of composites

7- Find the coordinates y(n+1) of the point in the interface that is collinear with the
(n) (n+1) (n+1) (n+1)
points of coordinates xtip and xtip . If ky(n+1) − xtip k < δR replace xtip by
(n+1)
y ,

(n+1)
xtip → y(n+1) . (7.2.6)

8- Increment n, n → n + 1.

9- If required to determine the path further, repeat steps 1 to 9.

We will now consider a periodically layered plate Ω = ([−0.5, 0.5] × [−0.5, 0.5]) \ ΩC ,
displaying a pre-existent crack, the crack line being given by ΩC = [−0.50, −0.49] × {0},
cf. Figure 7.7. The layers are at a 45o angle with the horizontal axis and the two compo-
nents of the plate are isotropic, homogeneous and linear elastic, being characterised by
E1 = 1, ν1 = 0.1 and E2 = 10, ν2 = 0.3.

E1, n1

E2, n2

0.05

o
45

-0.5
-0.49

Figure 7.7: Zoom in of the crack tip region.

To complete the problem formulation, we consider the following boundary conditions


 (0, 1), x2 = 0.5,
ϕN (x) = (0, −1), x2 = −0.5, (7.2.7)

(0, 0), otherwise,

and that body forces are absent. Our goal is to predict the path of the crack assuming
that the loading is sufficiently large to ensure crack growth. We will employ the pre-
vious algorithm with ∆a = 0.01 and δR = 0.0125 in two ways that differ in how the
first step of the algorithm is done. The first of these consists of computing a reference
7.2 Crack paths in layered materials 121

solution for the crack path. After this a computationally cheaper approximation will
be determined by using the hybrid approach described earlier to solve the elasticity
problem in step 1 of the algorithm.

The reference solution is obtained first by solving the elasticity problem by the finite
element method using fine meshes, composed of quadratic triangular elements near
the crack tip and quadratic rectangular elements elsewhere. The J-integral method is
employed to compute the SIFs in the fourth step of the algorithm.

n (∆a)x1 (∆a)x2 θP n (∆a)x1 (∆a)x2 θP

0 9, 89E − 3 1.48E − 3 8.52o 4 9.85E − 3 1.74E − 3 30.65o
1 9.95E − 3 −9.90E − 4 −14.20o 5 9.99E − 3 4.85E − 4 −12.78o
2 9.82E − 3 −1.88E − 3 −5.16o 6 1.00E − 2 1.13E − 4 3.43o
3 1.82E − 2 −7.09E − 3 −9.81o 7 1.65E − 2 2.29E − 3 7.27

Table 7.5: Crack tip increments.

This procedure enables us to predict the amplitude of the propagation angles θP of the
new crack segments, as well as the corresponding horizontal and vertical components
(∆a)x1 and (∆a)x2 of the increment vector. These values are displayed in Table 7.5 for
the first 8 increments.

Figure 7.8: Crack propagating through layers.

We note that the lengths of the increments when n = 3 and n = 7 are greater. This is due
to the proximity of the crack tip to the interface. The updated crack tip positions were
computed with step 7 of the algorithm. Since the plate is pulled in the vertical direction,
the path propagates essentially in a horizontal direction, as can be seen in Figure 7.8.
122 Fracture of composites

A second approximation for the crack path is obtained by employing the hybrid ap-
proach to solve the first step of Algorithm 7.1. The idea is to reduce the computational
complexity of the problem. As we have seen, working with composites often leads
to problems that may not be solved efficiently by applying the finite element method
in a straightforward manner. Since we considered one crack only, it is not necessary
to employ domain decomposition. The hybrid approach procedure then consists of
homogenising the material of the plate outside of the rectangle Ω1 = ([−0.5, 0.25] ×
[−0.25, 0.25]) \ ΩC . Next we solve the elasticity problem for a plate consisting of three
distinct materials, the original material components over Ω1 and the homogenised ma-
terial over Ω \ Ω1 . The crack path computed using Algorithm 7.1 with the hybrid ap-
proach is depicted in dots in Figure 7.9. There we also show the path obtained by using
the reference solution, represented as a full line. We can see that the plots almost coin-
cide.

0.01

0.008
Reference
0.006 Hybrid approach

0.004

0.002
2

0
x

−0.002

−0.004

−0.006

−0.008

−0.01
−0.5 −0.48 −0.46 −0.44 −0.42 −0.4 −0.38
x1

Figure 7.9: Paths determined by application of Algorithm 7.1.

It is then a good strategy to employ the hybrid approach to solve the elasticity prob-
lems when one wants to predict crack propagation for highly composite materials. This
will actually work better for materials with smaller heterogeneities due to the decrease
in the homogenisation error. In that case, the size of the region where we perform a
microscopic analysis can be smaller.

Finally, we note that a macroscopic approach can be adopted to find the path for the
homogenised plate. Naturally the effects of locality will then be lost, but the procedure
is much more straightforward. After employing homogenisation, one can simply ap-
ply a modification of Algorithm 3.1. Since the homogenised material is orthotropic, the
7.2 Crack paths in layered materials 123

maximum circumferential tensile stress criterion needs to be reformulated, see for ex-
ample [56,64]. One other aspect to consider is that the homogenised model does not take
into account the effect of the internal boundaries, so this would have to be incorporated
somehow.
Bibliography

[1] M. H. Aliabadi and M. H. López. Database of stress intensity factors. Computational

Mechanics Publications, 1996.

[2] G. Allaire. Shape optimization by the homogenization method. Springer, New York,
2002.

[3] A. Alonso, R. L. Trotta, and A. Valli. Coercive domain decomposition algorithms

for advection-diffusion equations and systems. J. Comput. Appl. Math., 96(1):51–76,
1998.

[4] M. F. Ashby and D. R. Jones. Engineering materials I, an introduction to their properties

and applications. Butterworth Heinemann, 1996.

[5] I. Babuska and J. E. Osborn. Generalized finite element methods: their performance
and their relation to mixed methods. SIAM J. Numer. Anal., 20(3):510–536, 1983.

[6] I. Babuska and J. E. Osborn. Finite element methods for the solution of problems
with rough data. In Singularities and constructive methods and their treatment, pages
1–18. Spring-Verlag, Berlin-New York, 1985.

[7] I. Babuska and T. Strouboulis. The finite element method and its reliability. Clarendon
Press, Oxford, 2001.

[8] N. Bakhvalov and G. Panasenko. Homogenization: averaging process in periodic media.

Kluwer Academic, Dordrecht, Boston, London, 1989.

[9] L. Banks-Sills and D. Ashkenazi. A note on fracture criteria for interface fracture.
Int. J. Fract., 103(2):177–188, 2000.

[10] E. J. Barbero. Finite element analysis of composite materials. CRC Press, 2008.

[11] A. A. Becker. The boundary element method in engineering. McGraw-Hill, Maiden-

head, 1992.

[12] A. Bensoussan, J. L. Lions, and G. Papanicolaou. Asymptotic analysis of periodic

structures. North-Holland, Amesterdam, 1978.
126 Bibliography

[13] P. E. Bjørstad and O. B. Widlund. Iterative methods for the solution of elliptic prob-
lems on regions partitioned into substructures. SIAM J. Numer. Anal., 23(6):1097–
1120, 1986.
[14] P. E. Bjørstad and O. B. Widlund. To overlap or not to overlap: a note on a domain
decomposition method for elliptic problems. SIAM J. Sci. Stat. Comput., 10(5):1053–
1061, 1989.
[15] D. Broek. Elementary engineering fracture mechanics. Kluwer Academic Publishers,
1986.
[16] L. Ceriolo and A. Tommaso. Fracture mechanics of brittle materials: a historical
point of view. 2nd Int. PhD Sympodium in Civil Engineering, 1998.
[17] T. F. Chan. Domain decomposition algorithms and computational fluid dynamics.
Int. J. High Perform. Comput. Appl., 2(4):72–83, 1988.
[18] T. F. Chan and D. Goovaerts. Schwarz = Schur: overlapping versus nonoverlapping
domain decomposition. CAM 88-21, Department of Mathematics, UCLA, 1988.
[19] J. Chang and J. Xu. The singular stress field and stress intensity factors of a crack
terminating at a bimaterial interface. Int. J. Mech. Sci., 49, 2007.
[20] G. P. Cherepanov. Mechanics of brittle fracture. MacGraw-Hill, 1979.
[21] D. Cioranescu and P. Donato. An introduction to homogenization. Oxford University
Press, 1999.
[22] A. Cirello and B. Zuccarello. On the effects of a crack propagating toward the
interface of a bimaterial system. Eng. Fract. Mech., 73(9):1264–1277, 2006.
[23] B. Cotterell. The past, present, and future of fracture mechanics. Eng. Fract. Mech.,
69:533–553, 2002.
[24] I. Demir, H. M. Zbib, and M. Khaleel. Microscopic analysis of crack propagation for
multiple cracks, inclusions and voids. Theor. Appl. Fract. Mech., 36:147–164, 2001.
[25] L. Dickinson. Using fiber to fight delamination: trans-laminar reinforcement out
of the lab and into production. Compos. Manuf., 2005.
[26] O. Dubois. Optimized Schwarz methods for the advection-diffusion equation. PhD thesis,
McGill University, Montréal, 2003.
[27] J. Dundurs. Discussion of edge-bonded dissimilar orthogonal elastic wedges under
normal and shear loading. J. Appl. Mech., 36, 1969.
[28] W. E, B. Engquist, L. Li, W. Ren, and E. Vanden-Eijnden. The heterogeneous multi-
scale method: a review. http://www.math.princeton.edu/multiscale, 2005.
[29] F. Erdogan and T. S. Cook. Antiplane shear crack terminating at and going through
a bimaterial interface. Int. J. Fract., 10:227–240, 1974.
[30] F. Erdogan and G. C. Sih. On the crack extension in plates under plane loading and
transverse shear. J. Basic Eng., 85:519–527, 1963.
Bibliography 127

[31] L. B. Freund. Dynamic fracture mechanics. Cambridge University Press, 1990.

[32] G. Galilei. Discorsi e dimonstrazioni matematiche intorno duo nuove scienze, 1638.

[33] P. Gould. Introduction to linear elasticity. Springer-Verlag, 1983.

[34] A. A. Griffith. The phenomena of rupture and flows in solids. Phil. Trans. Roy. Soc.
London, A221:163–198, 1921.

[35] P. de Haas. Numerical simulation of nonlinear water waves using a panel method; domain
decomposition and applications. PhD thesis, University of Twente, 1997.

[36] W. Han and B. D. Reddy. Plasticity mathematical theory and numerical analysis.
Springer, Berlin, 1999.

[37] D. Hegen. An element-free Galerkin method for crack propagation in brittle materials.
PhD thesis, Technical University of Eindhoven, 1997.

[38] T. Y. Hou and X. H. Wu. A multiscale finite element method for elliptic problems
in composite materials and porous media. J. Comput. Phys., 134:169–189, 1997.

[39] T. Y. Hou, X. H. Wu, and Z. Cai. Convergence of a multiscale finite element method
for elliptic problems with rapidly oscillating coefficients. Math. Comp., 227(68):913–
943, 1999.

[40] G. Houzeaux and R. Codina. An iteration-by-subdomain overlapping Dirichlet /

Robin domain decomposition method for advection-diffusion problems. J. Comput.
Appl. Math., 158(2):243–276, 2003.

[41] C. E. Inglis. Stresses in a plate due to the presence of cracks and sharp corners.
Proc. Inst. Naval Architects, 60:219–41, 1913.

[42] G. R. Irwin. Fracture dynamics. Fracturing of Metals, Proceedings of the ASM Sympo-
sium on Fracturing of Metals, pages 147–166, 1948.

[43] G. R. Irwin. Analysis of stresses and strains near the end of a crack transversing a
plate. Trans. ASME, J. Appl. Mech., 24:361–364, 1957.

[44] G. R. Irwin. Encyclopedia of physics (Handbuch der physic), volume IV. Springer,
Berlin, 1958.

[45] G. R. Irwin. Encyclopedia of physics (Handbuch der physic), volume VI. Springer,
Berlin, 1958.

[46] V. V. Jikov, S. M. Kozlov, and O. A. Oleinik. Homogenization of differential operators

and integral functionals. Springer-Verlag, Berlin, New York, 1994.

[47] P. K. Jimack. Engineering computational technology, chapter Domain decomposition

preconditioning for parallel PDE software, pages 193–219. Civil-Comp Press, 2002.

[48] L. Mishnaevsky Jr. Computational mesomechanics of composites: numerical analysis of

the effect of microstructures of composites on their strength and damage resistance. John
Wiley, 2007.
128 Bibliography

[49] M. Kachanov, B. Shafiro, and I. Trukrov. Handbook of elasticity solutions. Kluwer

Academic, 2003.

[50] M. F. Kanninen and C. H. Popelar. Advanced fracture mechanics. Clarendon Press,

1985.

[51] H. Kishi, Y. Mizuno, and H. Chazono. Base-metal electrode-multilayer ceramic

capacitors: past, present and future perspectives. AAPPS Bulletin, 14(2), 2004.

[52] A. K. Kulshreshtha and C. Vasile. Handbook of polymer blends and composites. Rapra
Technology Ltd., 2002.

[53] J. Lancaster. Engineering catastrophes: causes and effects of major accidents. Abington
Publishing, 2000.

[54] F. F. Lange. Fiber and laminar composites with strong interfaces. Silic. Indus., 63,
1998.

[55] R. W. Lewis and B. Koosha. A mixed mode rock fracture model for the prediction
of crack path. Int. J. Numer. Anal. Meth. Geomech., 23:281 – 294, 1999.

[56] W. Lim, S. Choi, and B. Sankar. Biaxial load effects on crack extension in anisotropic
solids. Eng. Fract. Mech., 68:403–416, 2001.

[57] J. R. Lund and J. P. Byrne. Leonardo da Vinci’s tensile strength tests: Implications
for the discovery of engineering mechanics. Civ. Eng. Environ. Syst., 18(3):243–250,
2001.

[58] A. I. Lurie and A. Belyaev. Theory of elasticity. Springer, Berlin, 2005.

[59] K. Madani, M. Belhouari, B. Bachir Bouiadjra, B. Serier, and M. Benguediab. Crack

deflection at an interface of alumina/metal joint: A numerical analysis. Comput.
Mater. Sci., 38:625–630, 2006.

[60] A. L. Madureira. Métodos numéricos para equações diferenciais parciais com múltiplas
escalas. Notas de aula de Minicurso no Seminário brasileiro de Análise, Vol: 31, No.
2, 2005.

[61] L. E. Malvern. Introduction to the mechanics of a continuous medium. Prentice-Hall,

1969.

[62] R. Mattheij, S. Rienstra, and J. ten Thije Boonkamp. Partial differential equations:
modelling, analysis, computation. SIAM, 2005.

[63] M. A. Meggiolaro, A. C. O. Miranda, J. T. P. Castroa, and L. F. Martha. Stress

intensity factor equations for branched crack growth. Eng. Fract. Mech., 72:2647–
2671, 2005.

[64] L. Nobile and C. Carloni. Fracture analysis for orthotropic cracked plates. Compos.
Struct., 68:285–293, 2005.

[65] E. Orowan. Energy criteria of fracture. Weld J. Res. Suppl., 30:157–160, 1955.
Bibliography 129

[66] D. R. J. Owen and A. J. Fawkes. Engineering fracture mechanics: numerical methods

and applications. Pineridge Press, 1983.
[67] M. Patrı́cio and R. Mattheij. Crack propagation analysis. CASA report 07-03, 2007.
[68] M. Patrı́cio, R. Mattheij, and G. de With. Homogenisation with application to lay-
ered materials. To appear in Math. Comput. Simulat., 2008.
[69] M. Patrı́cio, R. Mattheij, and G. de With. Solutions for periodically distributed
materials with localised imperfections. ICCES’08: International Conference on Com-
putational & Experimental Engineering and Sciences, 2008.
[70] M. Patrı́cio, R. Mattheij, and G. de With. Solutions for periodically distributed
materials with localised imperfections. To appear in Comput. Model. Eng. Sci., 2008.
[71] G. A. Pavliotis and A. M. Stuart. Multiscale methods: averaging and homogenization.
Springer, 2008.
[72] A. Quarteroni and A. Valli. Domain decomposition methods for partial differential equa-
tions. Clarendon Press, Oxford, 1999.
[73] J. R. Rice. Fracture: an advanced treatise, volume 2, chapter Mathematical analysis in
the mechanics of fracture. Academic Press, 1968.
[74] J. R. Rice. A path independent integral and the approximate analysis of strain
concentration by notches and cracks. Transactions of ASME, J. Appl. Mech., 35:79–
386, 1968.
[75] S. Roham, K. Hardikar, and P. Woytowitz. Crack penetration and deflection at a
bimaterial interface in a four-point bend test. J. Mater. Res., 19:3019–3027, 2004.
[76] H. Roos, M. Stynes, and L. Tobiska. Numerical methods for singularly perturbed differ-
ential equations. Spring-Verlag, 1991.
[77] H. P. Rossmanith. Fracture mechanics and materials testing: forgotten pioneers of
the early 20th century. Fatigue Fract. Eng. Mater. Struct., 22:781–797, 1999.
[78] Sanchez-Palencia. Non-homogeneous media and vibration theory, volume 127.
Springer Verlag, Berlin, 1980.
[79] H. A. Schwarz. Über einen grenzübergang durch alternierendes verfahren. Viertel-
jahrsschrift der Naturforschenden Gesellschaft in Zürich, (15):272–286, 1870.
[80] B. Smith, P. Bjørstad, and W. Gropp. Domain decomposition: parallel multilevel methods
for elliptic partial differential equations. Cambridge University Press, 1996.
[81] B. F. Smith. Domain decomposition algorithms for the partial differential equations of
linear elasticity. PhD thesis, New York, 1990.
[82] S. Spagnolo. Sul limite delle soluzioni di problemi di Cauchy relativi all’equazione
del calore. Ann. Sc. Norm. Sup. Pisa, (21):657–699, 1967.
[83] H. Tada, P. C. Paris, and G. R. Irwin. The stress analysis of cracks handbook. ASME
Press, 2000.
130 Bibliography

[84] M. T. Tilbrook, R. Moon, and M. Hoffman. Crack propagation in graded compos-

ites. Compos. Sci. Technol., 65:201–220, 2005.

[85] M. T. Tilbrook, K. Rozenburg, E. Steffler, L. Rutgers, and M. Hoffman. Crack propa-

gation paths in layered, graded composites. Composites Part B, 37(6):490–498, 2006.
[86] S. P. Timoshenko. History of the strength of materials. MacGraw-Hill, 1953.
[87] A. A. F. van de Ven. Field theory for continuous media. Lecture notes (Technical
University of Eindhoven).
[88] H. M. Versieux and M. Sarkis. Numerical boundary corrector for elliptic equations
with rapidly oscillating periodic coefficients. Commun. Numer. Meth. Eng., 22, 2006.
[89] L. da Vinci. Codice Atlantico. Folio 82 recto-B, date unknown.
[90] J. C. W. van Vroonhoven. Dynamic crack propagation in brittle materials: analyses based
on fracture and damage mechanics. PhD thesis, Technical University of Eindhoven,
1996.
[91] H. M. Westerngaard. Bearing pressures and cracks. Trans. ASME, J. Appl. Mech.,
1939.
[92] T. Wilshaw. Fracture of brittle solids. Cambridge University Press, 1975.
[93] G. de With. Structural integrity of ceramic multilayer capacitor materials and ce-
ramic multilayer capacitors. J. Eur. Ceram. Soc., 12, 1993.
[94] G. de With. Structure, deformation and integrity of materials. Wiley-VCH Verlag, 2006.
[95] A. Zehnder. Lecture notes on fracture mechanics. Cornell University, 2007.
Index

-periodic tensor, 50 hybrid approach, 98–101

linear, 2
anisotropic, 18 problem, 18, 49, 76, 98
tensor, 17
body force, 15 energy release rate, 26, 29
brittle, 3
finite elements, 5, 21, 42, 85
cell problem, 45, 51 fracture
characteristic functions, 58 composites, 3, 117–123
composite, 3 criteria, 29–32
conservation laws, 14–17 parameters, 25–28
constitutive equation, 17 simulation, see crack growth simulation
corrector, 47, 48
boundary, 47, 105 heterogeneous material, 3
first order, 47, 105 homogeneous material, 23–40
crack, see fracture homogenisation, 5, 41–66
growth simulation, 37–40, 117–123 approximation with correction, 48, 105
path, 5 elasticity, 49–66
criterion layered materials, 53–60
circumferential stress, 31, 114 homogenised
global, 29, 30 coefficient, 44, 46, 53
local, 29, 30 equation, 5
problem, 6, 45
defects, 109 solution, 44, 47
delamination, 117 hybrid approach, 6, 91–108, 115
Dirichlet-Neumann, 70, 78, 83, 87
displacement isotropic, 18
correlation method, 35
field, 28 J-integral, 26
vector, 11–14 J-integral method, 35
domain decomposition, 6, 67–90
elasticity, 76–90 kinematic equations, 11–14
Dundurs parameters, 118
Lamé moduli, 18
effective, see homogenised layered material, 6, 53–60
elasticity, 18–21 load, 21
domain decomposition, 76–90
homogenisation, 49–66 macroscopic, 5
132 Index

matching conditions, 69, 72, 77

microscopic, 5
modes of fracture
mixed mode, 24
mode I, 23
mode II, 24
mode III, 24
multiple scales method, 44–47

orthotropic, 18

path, see crack

plane stress, 19
Poisson’s ratio, 18

Richardson method, 80, 84

Schur
complement matrix, 87
complement methods, 68–71
Schwarz methods, 71–76, 80
shear modulus, 18
SIFs, 4, 25, 27–28, 32–36, 109–117
circumferential SIF, 31
critical SIF, 29
orthotropic material, 114
Steklov-Poincaré
convergence, 82–85
equation, 76–81, 83
operator, 77, 83
stiffness matrix, 21, 86
strain tensor, 11–14
stress
circumferential tensile, 30
correlation method, 34
field, 27
intensity factor, see SIFs
tangential, 30
tensor, 14–17
vector, 15
substructuring methods, 68–71, 80

Young’s modulus, 18
Summary

In this thesis we study the problem of crack growth on highly heterogeneous periodi-
cally distributed composites. The material components are assumed to be linearly elas-
tic, isotropic and homogeneous. Depending on the length scale at which the problem
is viewed, several points of view can be adopted. The simplest approach is given by
looking at the macroscale. It is assumed that the propagation occurs on a homogeneous
material and we are within the scope of linear elastic fracture mechanics. Predicting
the behaviour of a growing crack then involves solving the elasticity problems for the
cracked structure as the propagation occurs and having a fracture criterion available.

Instead, when the scale of the constituents is considered, the composites are seen to
have complex microstructures. This microscopic point of view gives rise to complicated
elasticity problems that can be simplified by employing homogenisation. The heteroge-
neous material is replaced by a homogenised material that has an equivalent behaviour
on the macroscopic scale. Mathematically this means that the highly oscillatory coef-
ficients that characterise the behaviour of the heterogeneous material are replaced by
constant coefficients.

Homogenisation allows us to look at a heterogeneous composite at the macroscale as

if it were homogeneous. This means that we have less computational complexity but
also less accuracy. As the purpose of this thesis is to look at the problem of fracture
from the scale of the constituents, homogenisation per se is not an option. However,
the prohibitive complexity of elasticity problems for composites has to be addressed
somehow, and so we look for other techniques. Domain decomposition methods can
be considered as a valid alternative. They solve the original problem by dividing the
computational domain in smaller subdomains. This divide and conquer logic certainly
looks promising, even allowing parallel implementations, but it requires the microscale
to be resolved throughout which seems unnecessary. Indeed, when a crack propagates,
we focus on the crack tip and its vicinity. Elsewhere the heterogeneous nature of the
material does matter, but it is not as relevant.

This leads us to consider a hybrid approach combining homogenisation and domain

decomposition techniques. The idea is to look at the problem of crack propagation at a
microscopic length scale only in the vicinity of the crack tip and at the macroscale every-
where else. This approach makes the problems of elasticity of periodically distributed
134 Summary

composites feasible and we show that it provides the desired microscopic accuracy. In
particular, with this technique we are able to set up a procedure to predict the path that
a crack propagating on a layered composite will follow. Naturally, the interaction of the
crack with the material interfaces also has to be addressed.
Samenvatting

In dit proefschrift bestuderen we het probleem van scheurpropagatie (scheurgroei) in

extreem heterogene, periodiek verdeelde, composietmaterialen. De componenten van
het materiaal worden lineair elastisch, isotroop en homogeen verondersteld. Afhanke-
lijk van de schaal waarop het probleem wordt bekeken kunnen verschillende aanpak-
ken worden toegepast. De eenvoudigste aanpak wordt verkregen door naar de macro-
schaal te kijken. Hierbij gaat men uit van een scheur die propageert in een homogeen
materiaal en die beschreven kan worden met de theorie van lineair elastische breuk-
mechanica. Het gedrag van een groeiende scheur kan dan voorspeld worden door de
elasticiteitsproblemen op te lossen in het domein waarin de scheur propageert en door
een breukcriterium te gebruiken.

Als daarentegen op de schaal van de materiaalcomponenten wordt gekeken, dan wordt

duidelijk dat de composietmaterialen een ingewikkelde microstructuur hebben. Deze
microschaal introduceert lastige elasticiteitsproblemen die vereenvoudigd kunnen wor-
den door gebruik te maken van homogenisatietechnieken. Het heterogene materiaal
wordt dan vervangen door een homogeen materiaal dat hetzelfde gedrag vertoont op
de macroschaal. Wiskundig gezien betekent dit dat de sterk fluctuerende coëfficiënten
die het gedrag van het heterogene materiaal beschrijven, worden vervangen door con-
stante coëfficiënten.

Via homogenisatie kunnen we naar heterogene composietmaterialen kijken als ware

deze homogeen. Dit betekent dat de complexiteit van de berekening omlaag gaat maar
tegelijkertijd ook minder nauwkeurig wordt. Omdat het doel van dit proefschrift is
om te kijken naar breuken op de schaal van de materiaalcomponenten, is simpelweg
toepassen van homogenisatie geen optie. Daarentegen is de complexiteit van het elas-
ticiteitsprobleem voor de composietmaterialen te hoog en zal dip op de één of ander
manier aangepakt moeten worden. Dit is de reden dat we naar andere technieken kij-
ken en domein decompositie biedt hier een mogelijk alternatief. Deze techniek lost het
oorspronkelijke probleem op door het rekenkundig domein op de splitsen in kleinere
subdomeinen. Deze verdeel en heers strategie ziet er veelbelovend uit, met de mogelijk-
heid tot parallellisatie in de implementatie, maar vereist wel dat de microschaal overal
opgelost moet worden wat gevoelsmatig niet erg zinvol lijkt te zijn. Wanneer name-
lijk een scheur propageert focussen we op het uiteinde van de scheur en diens directe
omgeving. In de rest van het domein is het heterogene aspect van het materiaal wel
136 Samenvatting

aanwezig, maar minder belangrijk.

Dit in gedachte hebbende zet ons ertoe om een hybride methode te bekijken dat een
combinatie is van homogenisatie en domein decompositie technieken. Het idee is nu
om het probleem van de scheurpropagatie op de microschaal alleen bij het uiteinde van
de scheur en diens directe omgeving te bekijken en op de macroschaal in de rest van het
domein. Deze aanpak maakt het mogelijk om de elasticiteitsproblemen van periodiek
verdeelde composietmaterialen op te lossen. Bovendien laten we zien dat ook aan de
gewenste nauwkeurigheid op de macroschaal wordt voldaan. In het bijzonder is het
mogelijk om een procedure op te stellen dat het pad van een scheur voorspeld wanneer
deze propageert in een gelaagd composietmateriaal. Uiteraard wordt ook de nodige
aandacht besteed aan de interactie tussen de scheur en de raakvlakken van de gelaagde
materialen.
Curriculum vitae

Miguel Patrı́cio was born in Coimbra (Por-

tugal) on October 24, 1980. He finished his
pre-university education at the Escola Se-
cundária de Avelar Brotero in his home ci-
ty in 1998. After that he enrolled at the De-
partment of Mathematics of the University
of Coimbra where he graduated in 2003 in
(Pure) Mathematics.

Subsequently, Miguel moved to The Nether-

lands where he joined the Centre for Analysis, Scientific Computing and Applications
(CASA) for a period of eight months. He was then granted a four years scholarship by
the Fundação para a Ciência e a Tecnologia to work on a different project still at CASA,
leading to this thesis.