Zohdi

Computational modeling
and design of new random

microheterogeneous materials
CISM Course Notes
T. I. Zohdi
COPYRIGHT WARNING
This document is under copyright. No part can be copied, electronically stored,
transmitted, reproduced or translated into another language without written permission from T. I. Zohdi. Please report any mistakes in this version to T. I. Zohdi.
Contents
Introduction
I.1 Basic concepts in macro-micro modeling
I.2 Historical overview
I.3 Objectives of this monograph
1 Some basics in the mechanics of solid continua
1.1 Kinematics of deformations
1.1.1 Deformation of line elements
1.1.2 Infinitesimal strain measures
1.1.3 The Jacobian of the Deformation Gradient
1.2 Equilibrium/kinetics of solid continua
1.2.1 Volumetric and surfacic postulates
1.2.2 Balance law formulations
1.3 Referential descriptions of balance laws
1.4 The First Law of Thermodynamics/An energy balance
1.5 The Second Law of Thermodynamics/A restriction
1.6 Infinitesimal linearly elastic constitutive laws
1.6.1 The infinitesimal strain case
1.6.2 Material symmetry
1.6.3 Material constant interpretation
xiii
xviii
xviii
xx
1
2
3
4
5
6
8
8
9
11
12
13
15
16
21
vii
viii
CONTENTS
1.7
1.8
1.6.4 Consequences of positive-definiteness

Hyperelastic finite strain material laws
1.7.1 Basic requirements for large strain laws
1.7.2 Determination of material constants
Moderate strain constitutive relations
22
24
24
26
27
2 Fundamental weak formulations

2.1 Direct weak formulations
2.1.1 An example
2.1.2 Some restrictions
2.1.3 The Principle of Minimum Potential Energy
2.1.4 Complementary weak forms
31
31
33
34
35
35
3 Fundamental micro-macro concepts

3.1 Testing procedures
3.1.1 The Average Strain Theorem
3.1.2 The Average Stress Theorem
3.1.3 Satisfaction of Hills energy condition
3.2 The Hill-Reuss-Voigt bounds
3.3 Observations
3.4 Classical micro-macro mechanical approximations
3.4.1 The asymptotic Hashin-Shtrikman bounds
3.4.2 The concentration tensor: microfield behavior
3.4.3 The Eshelby result
3.4.4 Dilute methods
3.4.5 The Mori-Tanaka method
3.5 Micro-geometrical manufacturing idealizations
3.5.1 Upper and lower variational bounds
3.5.2 Proof of energetic ordering
3.5.3 Uses to approximate the effective property
39
40
41
42
42
43
44
45
45
46
47
48
49
50
50
53
55
4 A basic finite element implementation

4.1 Finite element method implementation
4.2 FEM approximation
4.3 Global/local transformations
4.4 Differential properties of shape functions
4.5 Differentiation in the referential coordinates
4.6 Surface integrals
57
57
58
60
60
63
66
CONTENTS
4.7
4.8
4.9
ix
Post processing
Accuracy of the finite element method
Minimum principles and Krylov methods
4.9.1 Krylov searches and minimum principles
4.10 Exterior and interior augmentations
4.10.1 Properties: exterior/penalty functions
4.10.2 Properties: interior/barrier functions
67
68
70
71
74
75
76
5 Computational/statistical testing methods

5.1 A boundary value formulation
5.2 Numerical discretization
5.2.1 Topological resolution
5.2.2 A simple Galerkin error monitor
5.3 Elemental averaging and increased accuracy
5.4 Iterative Krylov solvers/microstructural correctors
5.5 Overall testing process: numerical examples
5.5.1 Successive sample enlargement
5.5.2 Multiple sample tests
5.6 A minimum principle interpretation
5.6.1 A proof based on partitioning results
5.6.2 Relation to the material tests
5.7 Partitioning and traction test cases
5.8 Dependency on volume fraction
5.9 Increasing the number of samples
5.10 Increasing sample size
5.11 An ergodic interpretation of the results
79
79
80
81
83
86
87
89
89
91
94
94
95
96
99
100
100
102
6 Higher order statistics

6.1 Statistical shifting theorems
6.2 Domain decomposition and ensemble averaging
6.2.1 Primal partitioning
6.2.2 Complementary partitioning
6.2.3 Homogenized material orderings
6.2.4 Embedded orthogonal monotonicities
6.3 Moment bounds on population responses
6.3.1 First order (average) bounds
6.3.2 Second order (standard deviation) bounds
6.3.3 Third order (skewness) bounds
115
115
116
116
118
119
119
120
121
121
122
CONTENTS
6.4
Remarks
123
7 Domain decomposition analogies and extensions

7.1 Boundary value problem formulations
7.2 Error in the broken problems
7.2.1 Multiscale proximity bounds
7.2.2 Uses of the bounds for domain decomposition
7.3 The connection to material testing
7.4 A total orthogonal sum
7.5 Iterative extensions
7.5.1 Method I: global/local CG iterations
7.5.2 Method II: iterative equilibration
7.5.3 Remarks
125
125
127
130
133
134
138
140
140
141
144
8 Nonconvex-nonderivative genetic material design

8.1 Computational material design
8.2 Minimization methods
8.2.1 Difficulties due nonconvexity
8.2.2 Local field design constraints
8.2.3 Difficulties due to nondifferentiability
8.3 Nonconvex-nonderivative genetic search
8.4 A semi-analytical example
8.5 Direct computational approaches
8.6 Primal and dual objective function envelopes

8.6.1 The implication of

8.6.2 The implication of
8.7 Numerical implementation
149
150
152
152
153
154
155
156
160
162
162
164
165
9 Inelastic responses
9.1 Introduction
9.2 A model problem
9.2.1 Constitutive assumptions: infinitesimal strain
case
9.2.2 A geometrically-nonlinear extension
9.3 A recursive staggering algorithm
9.4 Issues of convergence and staggering error control
9.4.1 Explicit time dependency
9.4.2 Contraction-mapping time stepping control
169
170
171
172
173
174
176
177
179
CONTENTS
9.5
9.6
Numerical Experiments
Multifield extensions
9.6.1 The infinitesimal strain case
9.6.2 A geometrically-nonlinear extension
9.6.3 Multifield staggering
xi
181
185
186
187
187
10 Closing comments
191
11 References
193
Introduction
BASE MATERIAL
ADDITIVES
MACROSCOPIC
STRUCTURE
NEW "MATERIAL"
MIXING
PROPELLOR
Fig. I.1 LEFT: Doping a base material with particulate additives. RIGHT: Typical loose
particulate additives.
A key to the success of many modern structural components is the tailored behavior of the material. A relatively inexpensive way to obtain macroscopically desired
responses is to enhance a base matrix properties by the addition of microscopic matter,
i.e. to manipulate the microstructure. Accordingly, in many modern engineering designs, materials with highly complex microstructures are now in use. The macroscopic
xiii
xiv
INTRODUCTION
characteristics of modified base materials are the aggregate response of an assemblage

of different "pure" components, for example several particles or fibers suspended in a
binding matrix material (Figure I.1). Thus, microscale inhomogeneities are encountered in metal matrix composites, concrete, etc. In the construction of such materials,
the basic philosophy is to select material combinations to produce desired aggregate
responses. For example, in structural engineering applications, the classical choice
is a harder particulate phase that serves as a stiffening agent for a ductile, easy to
form, base matrix material. A variety of materials are characterized by particulate
inhomogeneities embedded in an otherwise homogeneous base matrix, as shown in
Figures I.2, I.3 and I.4.
Fig. I.2 Left to right and top to bottom: (1) Ductile iron ( 100), (2) Aluminum matrix with
silicon-carbide particles ( 250), (3) Titanium-aluminum alloy ( 450), (4) Silver matrix with
copper inclusions ( 500), Source: The American Society for Metals [122].
If one were to attempt to perform a direct numerical simulation, for example of

the mechanical response of a macroscopic engineering structure composed of a microheterogeneous material, incorporating all of the microscale details, an extremely
INTRODUCTION
xv
Fig. I.3 Left to right and top to bottom: (1) Aluminum alloy reinforced with boron( 250),
(2) Aluminum alloy ( 250), (3) Zinc alloy ( 1000), (4) A cross-section of concrete ( 1.5).
Micrograph (1) is reprinted with permission of The Metalurgical Society, Warrendale, PA,
USA. For micrographs (2) and (3) Source: The American Society for Metals [122]. For
micrograph (4) Source: Van Vlack [181].
fine spatial discretization mesh, for example that of a finite element mesh, would be
needed to capture the effects of the microscale heterogeneities. The resulting system
of equations would contain literally billions of numerical unknowns. Such problems
are beyond the capacity of computing machines for the forseeable future. Furthermore, the exact subsurface geometry is virtually impossible to ascertain throughout
the structure. In addition, even if one could solve such a system, the amount of
information to process would be of such complexity that it would be difficult to extract any useful information on the desired macroscopic behavior. It is important to
realize that solutions to partial differential equations, of even linear material models, at infinitesimal strains, describing the response of small bodies containing a few
heterogeneities are still open problems. In short, complete solutions are virtually
xvi
INTRODUCTION
Fig. I.4 LEFT: Pearlite ductile iron with matrix etched away to show secondary graphite,
and bulls eye ferrite, around primary graphite nodules ( 475). RIGHT: Ductile iron, graphite
nodules (spherulites) encased in envelopes of free ferrite, all in a matrix of pearlite ( 100).
Source: The American Society for Metals [122].

ACTUAL STRUCTURE

STRUCTURE WITH
EFFECTIVE PROPERTIES
EFFECTIVE
PROPERTIES
DETERMINED
< > =EI :*< >
EI
(1)
EI
(2)
Fig. I.5 The use of effective properties in structural engineering.
impossible. Because of these facts, the use of regularized or homogenized material

models (resulting in smooth coefficients in the partial differential equations) are commonplace in practically all branches of the physical sciences. The usual approach
is the compute a constitutive relation between averages, relating volume averaged
field variables. Thereafter, the regularized properties can be used in a macroscopic
analysis (Figure I.5). The volume averaging takes place over a statistically repre-
INTRODUCTION
xvii
L1
L2
L3
L1 >> L2>> L 3
Fig. I.6 The size requirements of a representative volume element.
sentative sample of material, referred to in the literature as a representative volume

element (RVE). The internal fields to be volumetrically averaged must be computed
by solving a series boundary value problems with test loadings. Such regularization
processes are referred to as homogenization, mean field theories, theories of
effective properties, etc. For overviews, we refer the interested reader to Jikov et.
al. [94] for mathematical aspects or to Aboudi [1], Mura [130] and Nemat-Nasser
and Hori [131] for mechanically inclined accounts of the subject.
For a sample to be statistically representative it must usually contain a great deal
of heterogeneities (Figure I.6), and therefore the computations over the RVE are still
extremely large, but are of reduced computational effort in comparison with a direct attack on the real problem. Historically, most classical analytical methods for
estimating the macroscopic response of such engineering materials have a strongly
phenomenological basis, and are in reality non-predictive of material responses that
are unknown a-priori. This is true even in the linearly elastic, infinitesimal strain,
range. In plain words, such models require extensive experimental data to tune parameters that have little or no physical significance. Criticisms, such as the one stated,
have led to computational approaches which require relatively simple descriptions on
xviii
INTRODUCTION
the microscale, containing parameters that are physically meaningful. In other words,
the phenomenological aspects of the material modeling are reduced, with the burden
of the work being shifted to high performance computational methods. Stated clearly,
the "mission" of computational micro-macro mechanics is to determine relationships
between the microstructure and the macroscopic response or "structural property" of
a material, using models on the microscale that are as simple as possible. 1
I.1 BASIC CONCEPTS IN MACRO-MICRO MODELING

Initially, for illustration purposes, we consider the simple case of linear elasticity. In
this context, the mechanical properties of microheterogeneous materials are character
ized by a spatially variable elasticity tensor . Typically, in order to characterize the
(homogenized) effective macroscopic response of such materials, a relation between

averages
is sought, where
, and where and

are the stress and strain tensor fields within a statistically representative volume

element (RVE) of volume . The quantity,
, is known as the effective property,
and is the elasticity tensor used in usual structural scale analyses. Similarly, one can
describe other effective quantities such as conductivity or diffusivity, in virtually the
same manner, relating other volumetrically averaged field variables. However, for the
sake of brevity, we initially describe purely mechanical linear elastostatics problems.

2
It is emphasized that effective quantities such as
are not material properties,
but relations between averages. More appropriate terms might be apparent properties which is discussed indepth in Huet [79], [80]. However, to be consistent with

the literature, we continue to refer to
by the somewhat inaccurate term effective
property.
I.2 HISTORICAL OVERVIEW

The analysis of microheterogeneous materials is not a recent development. Within
the last 150 years, estimates on effective responses have been made under a variety of assumptions on the internal fields within the microstructure. Works dating
back at least to Maxwell [120] (1867), [121] (1873) and Lord Rayleigh [150] (1892)
have dealt with determining overall macroscopic transport phenomena of materials
consisting of a matrix and distributions of spherical particles. Voigt [183] (1889) is
usually credited with the first analysis of the effective mechanical properties of the
microheterogeneous solids, with a complementary contribution given later by Reuss
1 Appropriately,
the following quote, attributed to Albert Einstein: "A model should be as simple as possible,
but not simpler", is a guiding principle.
2 By direct extension, we are interested in "!$#&% vs "'(#% in the materially nonlinear/geometrically linear
range, or the appropriate conjugate pairing in the geometrically nonlinear range, such as *)+#,%.- vs 0/1#%2or 031#%.- vs *45#%.- , etc ..., where 3 and ) are the First and Second Piola Kirchhoff stresses, / is the
Right Cauchy-Green strain, 4 is the deformation gradient and where 6+7 is the initial reference volume.
HISTORICAL OVERVIEW
xix
[151] (1929). Voigt assumed that the strain field within

an aggregate sample of het

erogeneous material was uniform, leading to
as an expression of the effective
property, while Reuss approximated the stress fields within the aggregate of polycrystalline material as uniform. If the Reuss field
is assumed within the RVE, then an

. A fundamental result (Hill [70],
expression for the
effective
property
is

1952) is
.
These
inequalities
mean that the eigenvalues

of the tensors
and
are non-negative. Therefore, one
can interpret the Voigt and Reuss fields as providing two microfield extremes, since
the Voigt stress field is one where the tractions at the phase boundaries cannot be in
equilibrium, i.e. statically inadmissible, while the implied Reuss strains are such that
the heterogeneities and the matrix could not be perfectly bonded, i.e. kinematically
inadmissible. Typically, the bounds are quite wide and provide only rough qualitative information. Within the last 50 years improved estimates have been pursued.
For example, the Dilute family methods assume that there is no particle interaction.
With this assumption one requires only the solution to a single ellipsoidal particle in
an unbounded domain. This is the primary use of the elegant Eshelby [29] (1957)
formalism, based on eigenstrain concepts, which is used to determine the solution to
the problem of a single inclusion embedded in an infinite matrix of material under
uniform exterior loading. By itself, this result is of little practical interest, however
the solution is relatively compact and easy to use, and thus has been a basis for the
development of many approximation methods based on non-interacting and weakly
interacting (particle) assumptions. Non-interaction of particulates is an unrealistic
assumption for materials with randomly dispersed particulate microstructure, at even
a few percent volume fraction. One classical attempt to improve the Dilute approximations is through ideas of self-consistency (Budiansky [15], Hill [71]). For example,
in the standard Self-Consistent method, the idea is simply that the particle sees the
effective medium instead of the matrix in the calculations. In other words the "matrix material" in the Eshelby analysis is simply the effective medium. Unfortunately,
the Self-Consistent method can produce negative effective bulk and shear responses,
for voids, for volume fractions of 50 and higher. For rigid inclusions it produces
infinite effective bulk responses for any volume fraction and infinite effective shear
responses above 40 . For details, see Aboudi [1]. Attempts have also been made
to improve these approaches. For example the Generalized Self-Consistent method
encases the particles in a shell of matrix material surrounded by the effective medium
(see Christensen [23]). However, such methods also exhibit problems, primarily due
to mixing scales of information in a phenomenological manner, which are critically
discussed at length in Hashin [67]. For a relatively recent and thorough analysis of a
variety of classical approaches, such as the ones briefly mentioned here, see Torquato
[175]-Torquato [178]. In addition to ad-hoc assumptions and estimates on the interaction between microscale constituents, many classical methods of analysis treat
the microstructure as being a regular, periodic, infinite array of identical cells. Such
an assumption is not justifiable for virtually any real material. In this monograph
we do not consider the highly specialized case of periodic microstructures. There
exists a large body of literature for periodic idealizations, and we refer the reader, for
example, to Aboudi [1] for references.

xx
INTRODUCTION
I.3 OBJECTIVES OF THIS MONOGRAPH

It is now commonly accepted that some type of numerical simulation is necessary in
order to determine more accurate micro-macro structural responses. Ideally, in an
attempt to reduce laboratory expense, one would like to make predictions of a new materials behavior by numerical simulations, with the primary goal being to accelerate
the trial and error laboratory development of new high performance materials. Within
the last 20 years, many numerical analyses have been performed two-dimensionally,
which, in contrast to many structural problems, is relatively meaningless in micromacro mechanics. For a statistically representative sample of microstructural material, three-dimensional numerical simulations are unavoidable for reliable results.
As we have indicated, the analysis of even linearly-elastic materials, at infinitesimal
strains, composed of a homogeneous matrix filled with microscale heterogeneities,
is still extremely difficult. However, the recent dramatic increase in computational
power available for mathematical modeling and simulation raises the possibility that
modern numerical methods can play a significant role in the analysis of heterogeneous
structures, even in the nonlinear range. This fact, among others, has motivated the
work that will be presented in this monograph. The monograph is designed for a
self-contained one semester graduate course for students in the applied sciences and
mathematics with an interest in the computational micromechanical analysis of new
materials.
1
Some basics in the
mechanics of solid continua
Throughout this work, boldface symbols imply vectors or tensors. For the inner
product
of two vectors (first order tensors) and we have in three dimensions,

, where Cartesian bases and Einstein index
summation notation is used. At the risk of over simplification, we will ignore the
difference between second order tensors and matrices. Furthermore, we exclusively
employ a Cartesian bases. Readers that feel uncomfortable with this approach should
consult the wide range of texts which point out the subtle differences, for example the
texts of Malvern [118], Marsden and Hughes [119] or others listed in the references.
Accordingly, if we consider

(1.1)

then a first order
of two matrices
tensors)
"!$(two
contraction
102(inner
3 , product)
#%!& second
( ,') *& order
is defined by
with components of
, where
it is

clear that the range of/
the
inner
index
+
must
be
the
same
for
and
.
For
.
three
dimensions
+
.
The
second
order
inner
product
of
two
matrices
4! # 4! 658729: ,B*;< D =C ?> . The divergence of a vector (a contraction to isa
scalar) is defined
a second order tensor (a contraction to a
@A of, whereas
"!C ! . Thefor gradient
E hasbycomponents
vector) @A
of a vector (a dilation to a second
DC ! , whereas
order tensor) is @FAGIHKJGLNM OQPRKOQSTU SWVLXOGY
for a 4second
tensor (a
!C :C & . Theorder
dilation to a third order tensor) @A
gradient of a
Z
H
Q
J
N
L

M
Q
O
F
P

R
G
O
Z
S
U
T
W
S

V
N
L
G
O
Y
scalar (a dilation to a vector): @AW[HZJQL%M OQPRKOQSTUSWVLXOGY
[ . The scalar product of

two second order tensors, for example the gradients of first order vectors, is defined
102:3

D

C
!

C
!

!
/
.

, where ! are
as - +
, @

@

A

A

! are the Cartesian components of
partial derivatives
of and , and where
D

C

!

and and @FAG
has components of
in Cartesian bases.
2
SOME BASICS IN THE MECHANICS OF SOLID CONTINUA

1.1 KINEMATICS OF DEFORMATIONS
X3 x3
u+du
dx
dX
X+dX
P
X
O
X2 x
X1 x1
Fig. 1.1 Different descriptions of a deforming body.

9 5 >

9 5 >

!9 5 >
9! 5 > 9 9 5 > 5 > 5 5#"

The term deformation refers to a change in the shape of the continuum between
a reference configuration and current configuration. In the reference configuration,
a representative particle
of the continuum occupies a point in space and has the

position vector
(Figure 1.1), where
is a Cartesian reference triad, and
(with center ) can be thought of as labels for a

point. Sometimes the coordinates or labels
are called the referential
coordinates. In the current configuration the particle originally located at point is
located at point
can be expressed also in terms of another position vector , with
the coordinates
. These are called the current
coordinates. It is obvious

with this arrangement, that the displacement is
for a point originally
at
and with final coordinates . When a continuum undergoes deformation (or
flow), its points move
along various paths in space. This motion may be expressed by

, which gives
the present

location of a point that occupied the position
at time
, written
in terms of the labels
. The previous position vector may be interpreted
as a mapping of the initial configuration onto the current configuration. In classical approaches, it is assumed that such a mapping is a one-to-one and continuous,

KINEMATICS OF DEFORMATIONS

9 5>

9 5> 9 5>
with continuous partial derivatives to whatever order is required. The description

of motion or deformation expressed previously is known as the Lagrangian formulation. Alternatively,
if the independent variables are the coordinates and t, then

, and the formulation is denoted
as Eulerian (Figure 1.1).
1.1.1
Deformation of line elements
@
Partial differentiation of the displacement vector

produces the following displacement gradients:
where
@

@FA

A A A
A A A
A A A

A A A

A A A

A A A
and
with respect to

,and
@FA
(1.2)
(1.3)
where
is known as the material deformation gradient, and
is known as the
spatial deformation gradient.
Now,
consider
the
length
of
a
differential
element
in the

reference
configuration
and
in
the
current
configuration,

. Taking the difference in the magnitudes of these elements yields

Alternatively, we have with

9 @

Q9 ;

@ A
> Q9 @

>

> 0

(1.4)
and
9 @FA > G9 @ 0 A
Q9 ; >

>

(1.5)
Therefore we have the so-called Lagrangian strain tensor
9 ;
> 9 @ 9 @ > ; 9 @ > ; @

> (1.6)
Frequently, the Lagrangian strain tensor is defined in terms of the so-called right
9 > , ; . The Eulerian strain tensor is
Cauchy-Green strain,

defined as


9

; , > 9 @ A 9 @ A > ; 9 @ A > ; @ A >

(1.7)
9 >
In a similar manner as for the Lagrangian strain tensor, the Eulerian strain tensor can

be defined in terms of the so-called left Cauchy-Green strain,
,

.

F-1
REFERENCE
CURRENT
Fig. 1.2 Mapping back and forth with the deformation gradients.

can be reinterpreted as the result of a mapping

It should be clear that
Remark:
, or a change in configuration (reference to current),while ,

maps the current to the reference
and
9 system.
> Forandthe deformations
9 > to be .invertible,
physically realizable,
We note that
9 $5 ,> 9 $5 > . and the following obvious relation . It

.
should be clear that

1.1.2

Infinitesimal strain measures
9 @FA9 @ 9 @F> ; A @> ;

9 @ AG 9 > @F; AG@ >A; > @ AG
9 @ 9 @ > ; 9 @ > ; @ > , leading to
9 @ A 9 @ A > ; >
9 @ 9 @ > ; > . If the displacement gradients are small
and

and coincide closely to and , respectively. If we
compared with unity,

assume that,
, we may use
or interchangeably. Usually is the
In infinitesimal deformation theory, the displacement

gradient components
being

"small" implies that higher order terms
like
and
can

1

be neglected in the strain measure
and

symbol used for infinitesimal strains. Furthermore, to avoid confusion, when using
models employing the geometrically linear infinitesimal strain assumption we use the
or subscript.
symbol of with no
KINEMATICS OF DEFORMATIONS
(3)
dX
dX
(2)
dX (1)
(3)
dx
dx
(2)
(1)
dx
Fig. 1.3 A differential volume.
1.1.3
The Jacobian of the Deformation Gradient
The Jacobian of the deformation gradient,
$5

, is defined as
AA
A

AA
A

AA
A

(1.8)
&

To interpret the Jacobian in

a physical
way, consider
differential
a reference
volume

,
,
where
and
(Figure
1.3)
is
given
by

,
. The current differential element is described by

and

, where
is a unit vector, and

Q9

&
>

"

&

AA
AA
A
A

AA
A

(1.9)

Therefore,
. Thus, the Jacobian of the deformation gradient must remain
positive definite, otherwise we obtain physically impossible negative volumes.
Fig. 1.4 Newtons laws applied to a continua.
1.2 EQUILIBRIUM/KINETICS OF SOLID CONTINUA

We
start with the following postulated balance law for an arbitrary part of a body ,
, around a point P, with boundary
(Figure 1.4)

,
, (
,
G5

(1.10)

,

where is the material density, is the body force per unit mass (

) and
is the time derivative of the displacement. When the actual molecular structure is
considered on a sub-microscopic scale, the force densities, , which we commonly

refer to as surface forces, are taken to involve short-range intermolecular forces.
Tacitly we assume that the effects of radiative forces, and others which do not require
momentum transfer through a continuum, are negligible. This is a so-called local
action postulate. As long as the volume element is large, our resultant body and
surface forces may be interpreted as sums of these intermolecular forces. When we
pass to larger scales, we can justifiably use the continuum concept.
EQUILIBRIUM/KINETICS OF SOLID CONTINUA
X2
(3)
(1)
SIDE VIEW
X3
(2 )
X1
Fig. 1.5 Cauchy tetrahedron: a sectioned material point.
1.2.1
Volumetric and surfacic postulates

Consider
a tetrahedron
1.5.
From
Newtons laws,

in equilibrium,
as shown
in Figure

, where

is the surface area of the face of the tetrahedron with normal , and
is the
tetrahedron volume. Clearly, as the distance between the tetrahedron base (located
at

(0,0,0)) and
the surface center, denoted , goes to zero,
we
have

, and
. Geometrically, we have

"!

therefore
. It is clear

that forces on the surface areas could be decomposed into three linearly independent
components. It is convenient to express the concept of stress at a point, representing
the surfacic forces there, pictorially represented by a cube surrounding a point. The
fundamental issue that must be resolved is the characterization of these surface forces.
We can represent the force density vector, the so-called "traction", on a surface by the
$#
#
#

component representation:
, where the first index represents

the normal to corresponding coordinate plane, and the second index represents the
direction of the component. From this point forth, we will drop the superscript

notation of
, where it is implicit that
or explicitly
W9
Q9 >

> W9 > W9 >

9 > ;
N;
=?>
where
1.2.2
-+0
% &213(/.
% &4($.
10

%&')(+*,%&-$(/.
>
-3>
*,57698;:<*
>
>
4B-
>
-$1
>
1@-

131
=
.
1A4
>
41
-/4
>
434DC
(1.11)
4DCE
is the so-called Cauchy stress tensor.

Balance law formulations
Substitution of Equation 1.11 into Equation 1.10 yields

,
,

(
,
Q5

9
GF

,
"
(1.12)
A relationship can be
determined
between
thedensities
in the current and reference

configurations,
.
Therefore,
the Jacobian can also

densities
at a point. Since the
be interpreted as the ratio
of
material
volume is arbitrary,

we can assume
that
holds
at
every
point
in
the
body.
Therefore,
we may write

,
when
the
system
is
mass
conservative
over
time.
This leads to

H
B
H
B

IKJ

writing the last term in Equation 1.12 as B H
H
B

and an implicit assumption

that is

. From Gausss Divergence
theorem,

L!

differentiable, we have
. If the volume is argued as

being arbitrary, then the relation in the integral must hold pointwise, yielding
9 $" >
9 >
"
9 @FA
@A

>
"

(1.13)
REFERENTIAL DESCRIPTIONS OF BALANCE LAWS
33
22 21
23
12
32
11
31
13
X2
X1
X3
Fig. 1.6 Stress at a point.
1.2.2.1 Symmetry of the stress tensor Starting with an angular momentum

balance, under the assumptions that no infinitesimal micro-moments or so-called
couple-stresses exist, then it can
be shown
the stress tensor
must
be symmetric
that

(Malvern
[118]),
i.e.
which
implies

B
H

. It is somewhat easier to consider a differential element, such as in Figure
1.6
and
to simply
sum moments
about the center. Doing this one immediately obtains
#
#
#
#
#
#

. Therefore
and
N;

:

=?>
% &'(+*
% & -$($.
-B0
%&213(/.
1 0
%&4(/.
4
*,5
8 :
>
-3>
*
>
1@4B-
>
-$1
>
>
131
41
-/4
>
>
1A4
434DC
=
.
*,5
8 :
(1.14)
1.3 REFERENTIAL DESCRIPTIONS OF BALANCE LAWS

In many cases it is quite difficult to perform a stress analysis, for finite deformation solid mechanics problems, in the current configuration, primarily because it
is unknown a priori. Therefore all quantities are usually transformed (pulled)
back to the original coordinates, the referential frame. Therefore, it is preferable
to think of a formulation in terms of the referential fixed coordinated , a so
called Lagrangian formulation. With this in mind there are two commonly used
referential measures of stresses. We start by a purely mathematical result, leading
10
n0
da 0
da
-1
UNDEFORMED
DEFORMED
Fig. 1.7 A current and reference surface element.
> 9 >

9 > 9 > W9 >
9 5 > ; . This leads to 9 5 > ; . Applying the
defined by

9 > and ; $9

>

result we have

9 $5 ,> 9 > . This leads to ; 9 $5 ,> " " . This is

to the so-called Nanson formula for transformation of surface elements. Consider

of two
line
differential
elements in a current configuration,
the
product
cross

.
. An important vector identity (see
Chandriashakiah
and
Debnath
[20])
for
a
tensor
and two first order vectors and

is
, where the
is the transpose of the adjoint
9

the so-called Nanson formula. Knowing this, we now formulate the equations of
equilibrium in the current or a reference configuration.
Consider two surface elements, one on the current configuration, and one on a
reference configuration. If we form a new
kind of
stress
tensor,
call it , such thatthe

amount of force is the same
we
have

which implies
. The tensor is called the first Piola-Kirchhoff
stress, and gives the actual force on the current area, but calculated per unit area of
reference area. However, it is is not symmetric, and this
sometimes
causes
difficulties

in an analysis.
Therefore,
we
symmetrize
it
by

. The tensor
is called the second Piola-Kirchhoff stress. By definition

we have , and thus
; 9 $5 >
"
"
;

) * (

'
,+

! # "%$&
- /. 0 ' ,* + (

1 ! 2"%$3 1
-
7 0
7 0
; 9 $5 >

and therefore
" "
; 9 $5 ,>
" "
(1.15)
(1.16)

@ G9 %> "

Since
, control volume is arbitrary, we have
- . 3 ' , * + (
8
11
THE FIRST LAW OF THERMODYNAMICS/AN ENERGY BALANCE
-.

"
1

-

"

' , * + (
*
" . Since the

H
B
(1.17)
1.4 THE FIRST LAW OF THERMODYNAMICS/AN ENERGY BALANCE

The interconversions of mechanical, thermal and chemical energy in a system are
governed by the first law of thermodynamics, where the time rate of change of the
, is equal to the work rate, and the net heat supplied, ,
total energy,

. Here the kinetic energy of a subvolume of material
i.e. B H

or power of
contained in , denoted , is
, the rate of work

external forces acting on is given by

, the

heat flow into the volume by conduction is
,

the heat generated due to sources, such as chemical reactions, is
and

the stored energy is
. If we make the assumption that the mass in the
system is constant, one has,
9 >

!"$#%'&((
'
*
'
*
"
which implies

. Therefore,
energy balance leads to

,+ ' * ( &* (
2
"

' * ( * (
,

+
,+ ' * ( * (
' *( *
2
'
7*)

. Using this and the

' , + * ( * (

(1.19)
We also have

,+ ' ,+ '
76
D*
'
,

+

By using the divergence theorem, we obtain
(1.18)
E
54
@A
!+-,.+-"/&
01%'&((
*
7 0
7 0
' (
6
(1.20)
12

5
8K:
* ( -
$
-
*(

D*
8K5
*(
8
- * (
5
(1.21)
Combining the results, and enforcing a balance of momentum, leads to
' ( * ( ' * - ' - * ( -

'
' ( - * ( -
'
6

6
Since the volume
8;5
(1.22)
is arbitrary, the integrand must hold locally and we have
' (
- *( -
'
8
(1.23)
In later chapters dealing with multifield problems, this equation will be dealt with
more extensively.
Remark: Through similar arguments as above, one can reformulate the First Law
of Thermodynamics in the reference configuration as
where
"

(
' (
;.
7

-.
'
7
8

*
(1.24)
E
1.5 THE SECOND LAW OF THERMODYNAMICS/A RESTRICTION

,

Consider the quantity commonly known as the entropy of a volume

where is the entropy per unit mass, or specific entropy1. The rate of entropy input

is
, where
is the entropy source and is the entropy

influx. A basic postulate, that of Clausius (1854)-Duhem (1901), is
!

Q5

(1.25)
In other words, the time rate of change of entropy in the body is no less than the entropy
input into the body.
The last expression
be recast
forms. We

can
in various
useful

have B H
. Furthermore,
H B
- B H

by the divergence theorem we have
. Inserting these

"
1 Loosely
9 ">
"
@A
speaking, the entropy is a measure of the disorder in a system.
INFINITESIMAL LINEARLY ELASTIC CONSTITUTIVE LAWS
relations, and since the volume

have
13
is arbitrary, the integrand must hold locally and we
' ( ' -

(1.26)

@FA 9 @FA >5 , we obtain 9 @ A

@
A

9
>
>

we obtain
@F A . Employing
the
First
Law
of
Thermodynamics,

>

9

@ A / @FA . An unchallenged physical observation is that,
in the absence
of external input, heat flows from a region of higher temperature to
that of a lower
temperature,
hot to cold, i.e. heat never flows

against
the temperature

. Thus if Fouriers law
is
used
that
gradient, @A

@
A
F
@
A

must be positive definite. Since @A , we have implies
F
@
A
@ A 9 @ A >
9 @ A >
When
@ A then we
or
.

@ A or / @ A , either of which is referred
have

to as the Clausius (1854)-Planck (1887) inequality. Such inequalities can be used
By expanding

"!
to invalidate or construct (complicated) constitutive laws, although throughout this

monograph somewhat standard models, where the Second Law of Thermodynamics
is satisfied automatically.
1.6 INFINITESIMAL LINEARLY ELASTIC CONSTITUTIVE LAWS

The fundamental mechanism that produces forces in elastic deformation is the stretching of atomic bonds. Depicted in Figure 1.8 is a force/atom-separation curve. If
the deformations are small, then one can argue that we are dealing with only the
linear portion of the curve. Ultimately, this allows us to usually use linear relationships for models relating forces to deformations. Atoms in materials are held
together by electronic forces, or bonds, whose behavior resembles that of little
springs. The stiffness of an atomic bond force between two atoms is shown in Figure 1.8. If we
compute simple expressions for the force versus strain, we have

. Some materials, mainly ceramics, truly
exactly exhibit this theoretical stiffness. Most materials do not exhibit exactly this
stiffness, however, the relationships are approximately linear between the atomic
force and atom separation for small deformations. Therefore, to determine the material properties of structural materials, engineers usually resort to a macroscopic
tension test for material data.
The usual procedure to determine tensile properties of materials is to place samples
of material in testing machines, apply the loads, and then to measure the resulting
deformations, such as lengths and changes in diameter in a portion of the specimen,
of circular cross-section, called the gage length. The location of the gage length is
away from the attachments to the testing machine (Figure 1.9). The ends where the
samples are attached to the machine are larger so that failure will not occur there first,
which would ruin the experimental measurements. Slow rates of deformation are
applied, and the response is usually measured with strain gauges or extensometers.

14
FORCE

K

SEPARATION
0.25r
1.25r
Fig. 1.8 The atomic bond force as a function of atom separation.
MUP
MUP
M P
For a metal, samples are usually 1.25

cm in diameter and 5 cm in length. Compression

tests are usually on cubes (
) or cylinders (2.5 cm in diameter and
2.5 cm in length). Here both the load applied by the machine, and the shortening of
the specimen, should be measured over the gage length. For concrete, a material that
is expected to carry compressive loads, the specimens are usually on the order of 15
cm in diameter and 30 cm long. 2
The immediate result of a tension test is the axial force (F) divided by the original

area
(
), denoted, loosely speaking, the stress, and change in length, (

), per unit length, ( ), or engineering strain, measured by strain gauges.
As a first approximation we define the tensile stiffness of the material, known as

#

Youngs
Modulus,
denoted

,
by

. As we know, the
#
terms
and are, strictly speaking, not the true stress and strain, but simply serve
our presentation purposes. Clearly, what we have presented is somewhat ad hoc,
therefore, next we present the classical theory of elastic material responses for three
dimensional states of stress and strain.
"
"
"
2 Concrete
"
"
"
is aged usually 28 days before testing.
"
15
o
GAGE
LENGTH
Fig. 1.9 A standard metals testing machine.
1.6.1
The infinitesimal strain case
Initially we discuss relationships between the stress and strain, so-called material laws or constitutive relations for geometrically linear problems (infinitesimal deformations). Accordingly, if we neglect transfer of heat and production of

, which in the infinitesimal strain linheat, Equation 1.23 implies

early elastic
case
is
.
From
the
chain rule of differentiation we have

.
The
starting point to develop a constitu
H

tive theory is to assume a stored elastic energy function exists, a function denoted

on the mechanical deformation. The simplest func
only

, which depends
tion that fulfills
is
. Such a function satisfies the intuitive

physical requirement that, for any small strain from an undeformed state, energy must
be
stored
in the material.
Alternatively,
a small
strain material law
can be derived
from

and
which
implies
.

We are free to set
(it is arbitrary) to have zero strain energy at zero strain,
"!
and furthermore we assume that no stresses exist in the reference state (
), we
obtain the familiar relation
"
"
@A

16
(1.27)
This is a linear (tensorial) relation between stresses and strains. The existence of
a strictly positive stored energy function at the reference configuration implies that
the linear elasticity tensor must have positive eigenvalues at every point in the body.
Typically, different materials are classified according to the number of independent

constants in . A general material has 81 independent constants, since it is a fourth
order tensor relating 9 components of stress to strain. However, the number of constants can be reduced to 36 since the stress and strain tensors are symmetric. This is

easily seen from the matrix representation of :

Z

#

#
#

#
#
#

::
:

::: :: :::

:: : ::
: :

:
:
:

00 ::
0 :

(1.28)

The symbol is used to indicate standard matrix notation equivalent to a tensor
form, while
to
indicate vector representation. The existence of a scalar energy

function
forces
to be
symmetric
since
the
strains
are
symmetric,
in other
words

which

implies
. Consequently,
has only 21 free constants. The nonnegativity

of
imposes the restriction that
remain positive definite. At this point, based
on many factors that depend on the material microstructure, it can be shown that the

components of may be written in terms of anywhere between 21 and 2 independent
parameters. We explore such concepts further via the ideas of elastic symmetry.
1.6.2
> ; ;
; ;
Material symmetry

Transformation matrices are used in determining the symmetries. Consider a plane

of symmetry, the
plane (Figure 1.10). A plane of symmetry implies that the
material has the same properties with respect to that plane. Therefore we should be
able to flip the axes with respect to that plane, and have no change in the constitutive
law. By definition, a plane of elastic symmetry exists at a point where the elastic
constants have the same value for a pair of coordinate systems. The axes are referred
to as equivalent elastic directions. Also by definition,
an axis of symmetry of order

exists at a point when there are sets of equivalent elastic directions which can be
superposed by a rotation through an
angle
axis. Theway

about
an
to determine
elastic symmetry is as follows,
implying

which implies
, where is a

8

) )8

,
17
X3
X3
X1
^
X2
X2
REFLECTION
X1
X3
X3
X2
^
X1
X2
ROTATION
X1
Fig. 1.10 Top: reflection with respect to the

to the 4 axis.

1
plane. Bottom: rotation with respect

rotational transformation matrix. Imposing elastic symmetry means the components

are invariant with respect to the transformation, therefore
.

Therefore, all components that are not identical must be zero if the material has the
assumed elastic symmetry. In this fashion one can carve away components from a
general anisotropic material
tensor. The central point of such symmetries is that in a

new transformed state,
should not change.
1.6.2.1 Examples of elastic symmetry To make things clear consider the
following steps for one plane of symmetry starting with an originally Triclinic material
(21 free constants):
::

:

STEP 1: Reflect the
:: :: :: ::: ::

:: :: : :: :
: : :

(1.29)
axis

9 > . . .

(1.30)
18
STEP 2: Transform the stress and strain tensors with the same transformation,
but for second order tensor rules:
.
6

. . . . .
#

(1.31)
and thus

#
#
#
#

#

#
(1.32)
.
6

. . :: . . .

and also

::

Also,

::

::

(1.33)

(1.34)
STEP 3: Form the constitutive law in the primed frame:

>

>

and

>

>

-3>
131
434
>
>
>
-$1
1A4
4B-

*

131
434
>

-3>
>
-$1
>
1A4
4B-

13131A4
4341A4
434B-/4
-$1@-$1
-$131A4
-$1@-/4
1A43434
1A4B-$1
-/43434
131@-/4
-3-3-/4
-/4131
(1.35)
E
434B-$1
-$1A434
1A4131
-3-$1A4
1A4
4B-
4343434
-$13131
1A4B-3-
-$1
>
131@-$1
>
434131
-$1@-3-
434
131A434
434B-3

131
1313131

>

-3>
-3-3-$1
131@-3-
>
>
-3-/434
-/4B-3-
-3-$131
-3-3-3-

1A41A4
-/4B-$1
1A4B-/4
-/41A4
-/4B-/4

131
434
-$1
1A4

-3
4B-

(1.36)
19
STEP 4: Put everything in terms of the original variables, which implies that
the constitutive law must be the same as before,
if the plane was a plane of

symmetry, and thus the tensor relating and is :

-3-$131
1313131
-/4B-3-
-$1A434
1A4131
-/4131
1A43434
131@-$1
13131A4
-$131A4
1A4B-$1
-/4B-$1
434B-/4
-$1@-$1
131@-/4
4341A4
-/43434
434B-$1
1A41A4

-3-3-/4
-$13131
1A4B-3
4343434
-$1@-3
434131
-3-$1A4
131A434
-3-3-$1
131@-3434B-3-
-3-/434
-3-3-3
-$1@-/4
1A4B-/4
(1.37)
-/41A4
-/4B-/4
STEP 5: All components that are not equal before and after the reflection of
axes are zero:

#

#

:: ::

: :

:::::

::

0
00

(1.38)
The end result is a Monoclinic material, i.e, one plane of elastic symmetry (
plane)(13 free constants)

::: :: :::

: : :

or (

plane)(13 free constants)

: :
:: ::::

: :

::

:

(1.39)

::

(1.40)
The basic procedure is the same, for reflectional symmetry, rotational symmetries, etc.
What follows is a catalog of commonly referred to materials of various symmetries.

20
Two mutually perpendicular planes of symmetry reduce the material symmetry to nine
free constants, known as an orthotropic material. It also has a third mutually perpendicular plane of symmetry as a consequence, without changing the number of elastic
constants. In other words, if one were to reflect a reference frame located at a material
point (rotate by 180 degrees) the properties are the same. Accordingly, for an orthotropic
material, there are two planes of symmetry (nine free constants)
"
1313131
131A434
434B-3-
-3-/434
131@-3
-3-$131
434131
4343434
-$1@-$1

-3-3-3
1A41A4
(1.41)
-/4B-/4
In addition, if there is one plane in which the material properties are equal in all directions, there are only five free constants and the material is termed transversely isotropic.
Accordingly, for transversely isotropic material: two planes of symmetry and one plane
of directional independence (five free constants)
"

1313131
-/4B-/4
1A41A4
1A41A4

-$1@-$1
131A434
-3-/434
4343434
434131
1313131
131A434
434B-3-
-3-/434
131@-3
-3-3-3-
-3-$131
-3-3-3
and
(1.42)
-/4B-/4
-$1@-$1
*
1
-3-3-3-
-3-$131
Cubic materials have two planes of symmetry and two planes of directional independence, and can be shown to have three free constants. Accordingly, for a cubic material:
two planes of symmetry and two planes of directional independence (three free constants)
and
"
-3-3-3-
-3-$131
1313131
4343434
1313131
434131
131A434
434B-3-
131@-3
-3-/434
4343434
-3-$131
-3-/434
1A41A4
*
131A434

-$1@-$1
-3-3-3
(1.43)
E
-/4B-/4
-/4B-/4
*
4343434
*
-$1@-$1
Finally, if there are an infinite number of planes where the material properties are equal
in all directions, there are two free constants, the Lame parameters, and the material
is of the familiar isotropic variety. An isotropic body has material properties that are
the same in every direction at a point in the body, i.e., the properties are not a function
of orientation at a point in a body. Accordingly, for isotropic materials: two planes
of symmetry and an infinite number of planes of directional independence (two free
constants),
21

"
*

4

1
4

1

4
4

(1.44)
In this case we have

+

*
3
0
+

+

31
3
0
(1.45)
E
*
*
- .
where
and
. The eigenvalues of an isotropic elasticity

3 3 4

. Therefore, we must have
and
to retain
tensor are
E
E
positive definiteness
ofE E . E
Remark: Obviously, the elasticity tensor, regardless of the degree of anisotropy,

must have positive eigenvalues. It is symmetric,
and therefore, writing the similarity

mutually
orthonormal eigenvectors
of the
transform, the
collection of

elasticity as we have
, where
is
diagonalized:

which implies

. Clearly
the constitutive law in the new frame has the form:
; 8 ; .
8 ;

ZZ

#

#
#

#
#

< )

8 ; /
0
00

"!

(1.46)
The strain energy is simply
0. 9 > )G9 > <Q9 : > 9 $> G9 > 9 >

(1.47)
One interpretation is that of a generalized
decomposition of the energy into pieces

associated with the principle
elasticity tensor in matrix form.
- directions
. ,ofmustthe bethepositive.
Clearly, each eigenvalue,
$#

1.6.3
%
&%
&%
'
Material constant interpretation
There are a variety of ways to write isotropic constitutive laws each time with a
physically meaningful pair of material constants.
22
9 >

5 9 @ >
1.6.3.1 Splitting the strain It is sometimes

important
to split
infinitesimal

H
H
strains into two physically meaningful parts
. The Jaco
, and can be written as
bian, , of the deformation gradient is
AA
A

AA
A

AA
A

$5
$5 9 @ % >
9.

.

.

. 587 G58>7 @ "
9 @
587 >
(1.48)
. 85 7

and expanded as

, and therefore
with infinitesimal strains
- .

with
the
volumetric
part
of
the
deformation.
Furthermore,
Hence, is associated

H
, the so-called strain deviator can only affect the shape of a
since
differential element. In other words, it describes distortion in the material.
5872589 7 >

1.6.3.2
Associated infinitesimal strain linearly elastic material
The
laws

H
stress
can
be
split
into
two
parts
(dilatational
and
a
deviatoric):
>

, where we call the symbol the hydrostatic pressure, and

the stress deviator. This is one form of the so-called Hookes
resistance
9 Law.
> The, which
to change in the volume is measured by . We note that
indicates that this part of the stress
produces no distortion. Another fundamental form

587
, which implies

587 .

of Hookes law is
H

: :
0
587
:
:

To interpret
the constants,
consider
a uniaxial test where
G

#

#

#

,
. Under these conditions we have

. Therefore, , the so-called Youngs modulus, is the ratio
and
of the uniaxial stress to the corresponding strain component. The Poisson ratio, , is
the ratio of the transverse strains to the uniaxial strain.

Another
commonly
used set of stress-strain
forms are the Lame relations,

or
. To interpret the constants, consider a
#

#

#

#

#

#

pressure test where
, and where
. Under

,
and
. We
these conditions we have

implies
observe that
which implies that
, and
.
Therefore, from the fact that both and must be positive and finite, this implies

and

. For example, some polymeric foams exhibit
,
steels

, and some forms of rubber have

. However, no restrictions
arise on , i.e. it could be positive or negative.

. 3

1.6.4
587 :

.

Consequences of positive-definiteness
With a positive-definite linear elastic material law, at infinitesimal strains, the solution
is unique, in other words, there exists only one solution. To see this, let us first suppose
that there were two solutions that satisfy equilibrium

8

-*
& -$(
'*
*
& -$(
&"
$
8

-*
&213(
'*
*
&213(
(1.49)
E
23
and the specified traction boundary conditions

on

boundary conditions on , H
.
and specified displacement

H
>

, inte1.6.4.1 Elastostatic case Multiplying each equation by
grating over the volume, using the divergence theorem, and subtracting each equation
from one another we obtain
- *

*
& -$(
- *

&13(
& -$(
*

&213(

*

5
& -$(
.
&213( +8
& -$(
*
&213( +8;:

(1.50)
The integral on the right vanishes since the tractions from both solution must be the
same value on the traction specified boundary, and zero on the displacement part of the
boundary since the solutions are the same on the displacement part
of the boundary.

Since the domain is arbitrary, the integrand must vanish. Since is
definite
"
!
positive
the integrand can only vanish if the solutions are the same, i.e.
. For
completeness, we discuss the elastodynamic case.
>
1.6.4.2
case As in the static case, multiplying each equation
Elastodynamic

by
, integrating over the volume, using the divergence theorem, and

subtracting each equation from one another we obtain
- *

*
&-$(
&213(
- *(

& -$(
* (
&13(
'*

4
& -$(
*
&213(
$
8
*(
& -$(
* (
&213(

*
(1.51)
Let us define a function

5

>

>
9
9

@
@
0
9 > W9 >

(1.52)
We have that
>
9 >
@
(1.53)
9 > G9 > '
'
' , and thus the integral is equal to zero
where is a constant. The constant

@9

because the initial and boundary conditions are the same for both solutions. Since

the domain is arbitrary, the integrand must vanish. Since
is positive definite,

is positive, and the fact that both are independent of one another,
the
implies
that
!
integrand
vanish if the solutions are the same, i.e.
and

can
"only
!
.

24
1.7 HYPERELASTIC FINITE STRAIN MATERIAL LAWS

A commonly used model for purely elastic materials, such as some polymers, which
are expected to experience finite deformations, is hyperelasticity. Analogous to the
infinitesimal strain case, an elastic material is called hyperelastic if there exists a

stored energy
, that

function,

is only a function of the mechanical deformation, and

that
or
. These relations stem from the fact that the work
in a closed process (cycle) must be greater than or equal to zero, and in fact zero for
a

hyperelastic body. Various definitions for a nonnegative stored energy function, ,
in the initial configuration, follow from
H
H

G5 O
"

0.

" "

" G5

-
"

"

-
"

Q5
G5

(1.54)
Since the integrals are arbitrary, we have

,
and
.
Alternatively, these relations could have been
directly
obtained
from
the
First
Law
of

Thermodynamics, with no dissipation
(Box 1.24). Analagously

to the infinitesimal case, the results are
"

0

1.7.1

H

H
3Z

0

(1.55)
Basic requirements for large strain laws
.$>

9

>
In addition to being
material
frame
stored energy
function

invariant,
any

must
obey
4

criteria: (1)
, where
"!

are the principal invariants of , (2) , (3)
for
and (4) the material
constants in a finite deformation material law must
give
responses
with
known material
&
constants, for example in the isotropic case,
and , when perturbed
around
the undeformed configuration (see Ciarlet [24] for more details). Condition

implies that the material constants in a finite deformation material law must be
adjusted so that they give hyperelastic responses with known Lame constants, and
9 >

>
G5

25
HYPERELASTIC FINITE STRAIN MATERIAL LAWS
, when perturbed around the undeformed configuration, thus matching the simplest
(moderate strain) hyperelastic law, the Kirchhoff-St. Venant3
0
58729 >

9 >

(1.56)
For illustration purposes, consider a somewhat general standard form for a stored

energy function, , function as follows:

> 9
3 0
> 9

.$>

(1.57)
This is one possible form of a compressible material

response
functions. In this

model, the first and second invariants of the ,
and
, have been scaled by

the square-root of the third invariant, i.e.
and
to the compressible part

, to insure that they contribute

nothing

9
=

>
, where

of the response.
The
factors
stem from

, and , and where
,
. . We further note
,
which imply

and
that, defining an incompressible deformation gradient,

. , the first
;
. Note that because of the fact that

terms form a so-called incompressible function. In other words the corresponding

scaled third invariant is always unity. We may write
0
0
9
9

>
9

=>

>

(1.58)
where the following useful relations hold:

))
3 One

)
*,+
1
2
*
*

)
)

E
!!#!%$
&
.-
! & !!#!%$
&
*"!!#!%$
&
*!#!#!%$
&

1

E

('
)
*'
E
*!
&
))
*'
!! & !#!!%$0/

&
E
198

'
E
*"!!#!
& '
$ E
(1.59)
can alternatively write the equation in terms of the bulk 1325476 4 and shear moduli : . In general
a constitutive law of the form )"2<; =?>%= is known as a Kirchhoff-St. Venant material law. It is the
simplest possible finite strain law which is hyperelastic and frame indifferent.
26
where

B587
9

>

(1.60)
For proofs see Ciarlet [24].

1.7.2
Determination of material constants
The first three conditions for an admissible energy function are satisfied by construction leading to
> 9 0 9 >
0 9 =
3 3 > >

0 0
0 (1.61)

9 > .
yields
which when evaluated at

,) is also automatically satisfied. What
Therefore, condition ( ) (
for
9

remains is to satisfy condition ( ). We have a general tangent

9 9
9 ,89 =>
9 9 => >8
9 9 . > => 9 9 =>

9 9 => >
9 >

/

= >

(1.62)
MODERATE STRAIN CONSTITUTIVE RELATIONS
9 587 0 =
>
9 $>

0 0
9 9 > U >

27
Allowing
, and equating coefficients of the
tangent relations implied in Equa

tion
1.56
and
1.62,
we
have
from
,
which implies

,
and
from
,
.
Therefore,
the
coefficients
must

obey
, and we have in the general case
*
-
! &
.

0

-
!#! &

0

!#!#! &

2 31
(1.63)
, the material is called a Compressible

and
We remark that when
Neo-Hookean material, with a strain energy function of

3
*
! &
.

3
!#!! &
2 31
(1.64)
Remark: The
Cauchy stress can be split in the following manner,

H
where
, and thus
2 4
$ - !
6

4
$ 6
$ -
2 4
!

4
$ 6

$ -
6
/
(1.65)
)
We also have, by definition,

)
2

Since

2
6
/

2!

2
/
6

; , Equation 1.66 implies

6"#
/
/
$ -
(1.66)
)$
32

&%
2
('*)+
(1.67)
1.8 MODERATE STRAIN CONSTITUTIVE RELATIONS

From the point of view of an observer attached to a rigid body in motion, certain
quantities will remain unchanged. Quantities such as distance or stress are said to
be objective. For
example,
consider a rigid motion, a rotation
and translation, of
,

. An original differential element,
, will be oriented
the form

differently
after
the
motion,
,
however,
its
length
will
remain
unchanged, i.e.

"
" " . A quantity such as distance is said to be objective under rigid body
.,

motions. This definition holds for any vector that transforms
according
to
.
/,

,

Analogously, for second order tensors, we require
. The velocity
vector is an0
example
of a non-objective vector, because differentiating the rotated
,

mapping
, with respect to time yields

;
28
G5

/,

(1.68)
thus,
. The concept of objectivity under rigid body motions is central to
the construction of reliable constitutive relations. In this context, objectivity can be
used by studying the effect of rigid motions superposed on existing deformations. An
important example of a nonobjective tensor is the temporal derivative of the Cauchy
stress,

9
,
;>
G5
&,
"!

,

,
,
;

; (1.69)
,
Thus, unless
, it is not objective. The application of such concepts to constitutive relations is straightforward. Our goal is always to employ constitutive relations
that transform consistently, i.e. their character does not change under superposed
rigid

body motions. Consider the definition of the deformation gradient,
.
Superposed rigid motions leads to

Q9
.,

>

(1.70)
As a consequence, we have for the Green-Lagrange referential strain measure

since

0. 9 ;
;

&,
> 0. 9 9 > ;

,

> 0. 9 ;

>

(1.71)
. Also, for the second Piola-Kirchhoff strain measure,

9
; > *9 ; ; >
> *9
,
( ,
,

( ,

;

(1.72)

Therefore, a constitutive law such as the Kirchhoff-St. Venant law
, is
said to be
frame-indifferent,
since
it
employs
a
objective
stress
and
strain
measures,

i.e.
We note that simply because a strain of stress
measure employs quantities such as the deformation gradient in its definition, does
not mean that it will remain unaltered under rigid motions, for example, take the
Almansi (Eulerian) strain

0. 9

;

>

0. 9

>
+,
0,
+,
; (1.73)
However, such a measure will nonetheless

prove
to be useful. For example, we

could construct the following law
. However, it is important to note
that such a constitutive law relating the Cauchy stress and the Almansi strain tensor must be isotropic in order to remain
frame
indifferent
under all
superposed
rigid
$
,

,

implying
body rotations. To see this consider,
) ,8 )8 ,?;

8

29
MODERATE STRAIN CONSTITUTIVE RELATIONS

8 8
,

, where is a
which implies

transformation matrix, and
where
is
used
to
indicate
matrix
notation equivalent

to indicate a vector representation. Frame indifference
to a tensor form, while

,

requires that
,
and thus
. This relation holds only if

is isotropic.4 However, such a relation is only elastic, i.e. the deformation is
only a function of the deformed state, and not hyperelastic, i. e. it is not derivable
from a differentiation of a potential energy function. It is said to be Cauchy-elastic.
Hyperelastic constitutive relations, those derived from scalar energy
functions em
ploying objective strain measures, such as Kirchhoff St. Venant,
or Compressible Mooney Rivlin materials, which employ the principle invariants
of , are automatically frame indifferent. The simple Kirchhoff-St. Venant and
Almansi/Eulerian laws are extremely useful in applications where there are small or
moderate elastic strains, and large inelastic strains. For finite deformations with mod
erate elastic strains (
), the constitutive laws discussed yield virtually identical
responses. Consider
8

9 ; > 9 > ;

Kirchhoff-St. Venant
,0
9 ; > 3
3

9
>
9

>

Compressible Mooney-Rivlin

0 9 . >
9 > , with a special case of a Neo-Hookean material being .
Eulerian
,

We consider the following loadings on a homogeneous block

.
6
.
. .

.

(1.74)
3
The responses are shown in Figure 1.11. The relative
proximity
of
these
relation
9 > ; 9 @ > and 9 @ A > ;
should
a surprise for under
strains since @
9 @ A > notare besmall.

4 The
Kirchhoff St. Venant relation can be anistropic.

APPLIED TRIAXIAL DISPLACEMENT
SECOND PIOLA-KIRCHHOFF STRESS (GPa)
30000
20000
EUL.
K-ST.V.
COMP. N.-HK.
C. MN.-RIV.
10000
-10000
-20000
-30000
-0.1
-0.08
-0.06
-0.04
-0.02
0
0.02
0.04
GREEN-LAGRANGE STRAIN
0.06
0.08
0.1
0.12
0.08
0.1
APPLIED SHEAR DISPLACEMENT

6000
SECOND PIOLA-KIRCHHOFF STRESS (GPa)
30
4000
EUL.
K-ST.V.
COMP. N.-HK.
C. MN.-RIV.
2000
-2000
-4000
-6000
-0.1
-0.08
-0.06
-0.04
-0.02
0
0.02
GREEN-LAGRANGE STRAIN
0.04
0.06
Fig. 1.11 A comparison of various finite deformation laws.
2
Fundamental weak
formulations
In most problems of mathematical physics the true solutions are nonsmooth, i.e. the
strains and the stresses are not differentiable
in
the classical sense. For example in
,

!
the equation of static equilibrium
, there is an implicit requirement that

the stress was differentiable. 1 In many applications, this is too strong a requirement.
Therefore, when solving such problems we have two options: (1) enforcement of
jump conditions at every interface or (2) weak formulations (weakening the regularity
requirements). Weak forms, which are designed to accommodate irregular data and
solutions, are usually preferred. Numerical techniques employing weak forms, such
as the Finite Element Method, have been developed with the essential property that
whenever a smooth classical solution exists, it is also a solution to the weak form
problem. Therefore, we lose nothing by reformulating a problem in a weaker way.
However, an important feature of such formulations is the ability to allow natural
and easy approximations to solutions in an energetic sense, which is desirable in
framework of mechanics.
2.1 DIRECT WEAK FORMULATIONS

To derive a direct weak form for a body, we take the equilibrium equations (denoted
the strong form) and form a scalar product with
an arbitrary
smooth
vector
valued
"

function , and integrate over the body,
, where
9@ >

1 Throughout this chapter, we consider only static linear elasticity, at infinitesimal strains, and specialize
approaches later for nonlinear and time dependent problems.
31
32
FUNDAMENTAL WEAK FORMULATIONS
ONE TEST FUNCTION

r=RESIDUAL
ANOTHER TEST FUNCTION
Fig. 2.1 The idea of a test function.
9@
>
is called the residual. We

call a test function. If we were to
add a condition

that we
do
this
for
all
(
)
possible
test
functions
then

!

,
, implies
. Therefore if every possible test function was

!
considered, then
on any finite region in . Consequently,

the weak and strong statements would be equivalent provided the true solution is
smooth enough to have a strong solution. Clearly, can never be zero over any
finite region in the body, because the test function
will find them
(Figure 2.1).

1

Using the product rule
of
differentiation,

leads to, ,
, where we choose

the from an admissible set, to bediscussed
momentarily.
Using
the
divergence

theorem, leads to,
,
, which

leads to
.
If
we
decide
to
restrict
our
!

choices of s to those such that
, we have, where is the applied boundary
displacement on , for infinitesimal strain linear elasticity
@
9 > 9 @ %> @
@
9

9
>
%
>
@

@

@

@

* *
+-"

%

&

- *
E
*

(
#
/&
#

E
0

&

(2.1)
This is
called a weak form because it does not require the differentiability of the

stress . In other words, the differentiability requirements have been weakened. It is
clear that we are able to consider problems with quite irregular solutions. We observe
DIRECT WEAK FORMULATIONS
33
that if we test the solution with all possible test functions of sufficient smoothness,
then the weak solution is equivalent to the strong solution. We emphasize that provided
the true solution is smooth enough, the weak and strong forms are equivalent, which
can be seen by the above constructive derivation. To see this a bit more clearly, we
consider a simple example.
2.1.1
7' " 9 N> ' " 9 N>

7
An example
9 >
7
7 ' " 9 N>

7 7
7$ "

"
A B' 9 N>
A ' 9 N> ' 9 N>
' 9 N>
=' 9 N>

Let us define
a one-dimensional
continuous function

5

, on& a one-dimensional

domain,

.
Our
claim
is
that
,
, which implies

. This can be easily proven by contradiction. Suppose
at some point
. Since , there must exist a subdomain (subinterval), , defined

through ,
such that has the same sign as at point . Since is arbitrary,
we may choose to be zero outside
of
this interval,
and positive inside (Figure 2.2).

This would imply that
which is a contradiction. Now

select

. Therefore,

for example in one-dimensional infinitesimal
strain
linear
elasticity
the
equivalence

of weak and strong forms occurs if
.
Remark: For the general case of geometrical
and &material
nonlinearities, the
#

equivalence of weak and strong forms occurs if
.
r
v
0
Fig. 2.2 A residual function and a test function.
34
2.1.2
Some restrictions
A key question is the selection of the sets of functions in the weak form. Somewhat
naively, the answer is simple, the integrals must remain
finite. Therefore the following

restrictions hold ( ),
,
and

, and govern the selection of the approximation spaces. These relations simply
mean that the functions must be square integrable. In order to make precise statements
one must have a method of book keeping. Such a system is to employ so-called
Hilbertian Sobolev spaces. We recall that a norm has three main characteristics
for

1"
" "
" 1"
,
"
if
and
such
that
and
are
(1)
any vectors
!

and only if
, (2) "
"
"
" and (3) "

" " , where
is a
scalar. Certain types of norms, so-called Hilbert space norms,
are
frequently
used
in

solid mechanics. Following standard notation, we denote
as the& usual space of

scalar functions with generalized partial derivatives of order
in
, i.e. square

1

integrable, in other words
if " "

@
9 N. > 9 N>
A
A

9 N>

. We define 9 N> 9 N> 9 N>8 as the space of vector-valued functions whose
components are in
, i.e.

9
N

>

Y ! !

(2.2)
9 N> 9 N>8 . Using these definitions, a complete weak boundand we denote
ary value problem
can be written
follows. The data (loads) are assumed to be such
9 N> and
9 as> , but
9 N> without
that
less smooth9 Ndata
> andcan be considered

complications. Implicitly we require that
without con

&
&
&

"

&
&
&
tinually making such references. Therefore in summary we assume that our solutions
obey these restrictions, leading to the following infinitesimal strain linear elasticity
weak statement:
*
+-"

- *
*

(
#
/&
#

E

-

(2.3)
We note that if the data in (2.3) are smooth and if (2.3) possesses a solution that is
sufficiently regular, then is the solution of the classical linear elastostatics problem
in strong form:
@ Q9 @ >

9 @ >

"!
(2.4)
2.1.3
35
The Principle of Minimum Potential Energy
A useful concept in mechanics is that of minimum principles. By direct manipulation

we have

9

> 0

9 > 9 > 0 9 > 0
9 > 9 > 0 9 > 9 >
9 > 9 > 0 9 >
9 > 0 9 > 9 > 0
0 9 > 0 9 > 9 9 > 9 > >
9 > 9 >
(2.5)

9 > 9 > 9 > @
where we define the elastic potential

@
. This implies
0

9 9 > 9 >:> 7 9 > 9 >
(2.6)
"
"

"

"
D
where Equation 2.6 is known as the Principle of Minimum Potential Energy (PMPE).
In other words, the true solution possesses the minimum potential.
The minimum property of the exact solution can be proven by an alternative
technique. Let us construct a potential function, for a deviation away from the exact
solution , denoted
, where
is a scalar and is any admissible
variation

(test function)

. If we differentiate with respect to ,

, and set
, because

, we have

we know
that the exact
solution
is for

. This is sometimes referred to as

the Euler-Lagrange equation of the potential. Clearly, the minimizer of the potential
is the solution to the field equations, since it produces the weak formulation as a

result. This is a minimum since

. It is
important to note that the weak form, derived earlier, requires no such potential, and
thus is a more general approach than a minimum principle. Thus, in the hyperelastic
case, the weak formulation can be considered as a minimization of an potential energy
function. This is sometimes referred to as the Rayleigh-Ritz method.
>
@
2.1.4
9 > @ 9 > @ 9 > Q9

29 >

9@ >

"

@

"

@ @

Complementary weak forms

$

9
@
>
There exists another set of weak formulations

and minimum
principles called com
L!

L!
plementary principles. Starting
with
, multiplying
by the

solution leads to
. Using

the divergence theorem yields
36
+"

8K5

%

&
)5

5

$ -

*

8;:
E
*
8;:

8

(
#
/& #
*,%
&

8

8K:

*
(2.7)
This is called the complementary form of Equation 2.1. Similar restrictions are placed
on the trial and test fields to force the integrals to make sense, i.e. to be finite. Similar
boundedness restrictions control the choice of admissible complementary functions.
In other words we assume that the solutions produce finite energy. Despite the apparent simplicity of such principles they are rarely used in practical computations

because of the fact"that
it is very hard to find even approximate functions, , that
!
satisfy
.
2.1.4.1 The Principle of Minimum Complementary Potential Energy As

in the primal case, a similar process is repeated using the complementary weak formulation. We define a complementary norm

5
&
"

% (

5
$ -

5

'
*

5
(2.8)

,
9
> 0

,9 %> ,9 > 0 ,9 > 0

,9 > ,9 %> 0 ,9 > ,9 %>
,9 > ,9 %> 0 ,9 %>
,9 > ,
0 9 %> 9 %>0
,
90 > 0 9 > 9,9 % > 9 %> >
9 > 9 %>
(2.9)

9 > ,9 > 9 >

0 9 9 > 9 %> > 7 9 > 9 >

(2.10)
Again, by direct manipulation, we have

"
where
we define

. Therefore,
"

which is the Principle of Minimum Complementary Potential Energy (PMCPE). By

directly adding together the potential energy and the complementary energy we obtain
an equation of balance:
37
9 > 9
>

0.
@
.0
These concepts will prove pivotal later.
@

(2.11)
3
Fundamental micro-macro
concepts
As stated in the introduction, it is clear that for the relation between averages to
be useful it must be computed over a sample containing a statistically representative amount of material. This requirement can be formulated in a very precise
mathematical way, which has a clear physical meaning. A commonly accepted
macro/micro
criterion
used
in effective
property calculations is the so-called Hills

condition,
. Hills condition (Hill [70], 1952) dictates the
size requirements on the RVE. The classical argument is as follows. For any perfectly bonded heterogeneous body, in the absence of body forces, two physically
important loading states
satisfy
Hills condition.
They are pure linear displacements

2

of the form:
(1)
1

or (2) pure tractions in the form

1

; where and are constant strain and stress tensors,

respectively.
Clearly, for Hills conditions to be satisfied within a macroscopic body
under nonuniform external loading, the sample must be large enough to possess small
boundary field fluctuations relative to its size. Therefore applying (1)- or (2)-type
boundary conditions to a large sample is a way of reproducing approximately what
may be occurring in a statistically representative microscopic sample of material in
a macroscopic body. Thus, there is a clear interpretation to these test boundary conditions. Our requirement that the sample must be large enough to have relatively
small boundary field fluctuations relative to its size and small enough relative to the
macroscopic engineering structure, forces us to choose boundary conditions that are

uniform. This is not optional. Explicitly,
to determine
six

linearly

, one specifies

independent
loading
of
the
form,
1

or

where
and
are symmetric second order strain and stress tensors,
with spatially constant components. Each independent loading state provides six

9 . >

0
9>

39
40
FUNDAMENTAL MICRO-MACRO CONCEPTS
equations, for a total of 36, which are used to determine the tensor relation between

average stress and strain,
. If the effective response is assumed isotropic then
only
H
and
one
test
loading
(instead
of
usually
six),
containing
non-zero
dilatational
(

H

) and deviatoric components (

and
to determine the effective bulk and shear moduli:

0
JS

), is necessary

(3.1)

We note that even if the aggregate response is not purely isotropic, one can interpret
the above expressions as generalizations of isotropic responses.
3.1 TESTING PROCEDURES

Explicitly, in order to determine structural scale material properties of microheteroge

neous
material one computes the following relation between averages,

,

> -3-
> 131
> 434

1A4
>
-/4
>

>
%
%
%
%
-$1
%
%

-3 -$131
1313131

434131

-$ 13131
1A4131

-3 -3-$1
131@-$1

434B-$1

-$ 1@-$1
1A4B-$1

-3 -$1A4
13131A4

4341A4

-$ 131A4
1A41A4

-/4B-3-
-/4131
-/43434

-/4B-$1
-/41A4

-3 -3-/4
131@-/4

434B-/4

-$ 1@-/4
1A4B-/4

-3 -/434
131A434

4343434

-$ 1A434
1A43434

*

-3 -3-3131@-3
434B-3
-$ 1@-31A4B-3
-/4B-/4
-3- %
131 %
434 %
3
-$1 %
3
1A4
%
3
-/4
%

(3.2)

where
, and where and are the stress and strain tensor fields

within a microscopic sample of material, with volume . In mechanics literature,
such a sample is called a statistically representative volume element (RVE). In order
for such an analysis to be valid, i.e. to make the material data reliable, the sample
must be small enough that it can be considered as a material point with respect to the
size of the domain under analysis, but large enough to be a statistically representative

sample of the microstructure.
is exactly what appears in engineering books as
the properties of materials. A wide variety of microheterogeneous materials, such
as those shown in Figure I.2, can be treated by the general approach to be presented.
The usual choices for the six independent load cases are

=
*
CE

CE
=
CE

(3.3)
CE
41
TESTING PROCEDURES
Fig. 3.1 Nomenclature for the averaging theorems.
where is a load parameter. Each independent loading state provides six equations,
for a total of 36, which are used to determine the tensor relation between average

stress and strain,
. For completeness we record a few related fundamental results,
which are useful in micro-macro mechanical analysis.
3.1.1
The Average Strain Theorem
If a composite body has the following uniform loading on its surface:

then

2
3
2
*
%
*
- * - *
- * - *

*
*

3
3
2
3
3

%
2
3
A6

8
=
8

A6

8

A6
8

*
- * - *

*
*

*
*

*

*
%

A61

*

(3.4)
42
9
!! >

where
is a tensor product of the vector
describes the displacement jumps in the interfaces between
only if the material is perfectly bonded, then
2

and vector .
and . Therefore,

(3.5)
Note that the presence of finite body forces does not effect this result. Also note that
the third line in Equation 3.4 is not an outcome of the divergence theorem, but of a
generalization that can be found in a variety of places, for example Chandrasekharaiah
and Debnath [20] or Malvern [118].
3.1.2
The Average Stress Theorem

@ Q9
> 9 @ %>

Again
if we consider a body with
, then

. Therefore, substituting this yields

)
"! #%

"!
If there are no body forces

2

6

(3.6)
(3.7)
Note that debonding (interface separation) does not change this result.
3.1.3
Satisfaction of Hills energy condition
9 % >
@
, then

Consider
a body
with
a perfectly
bonded
microstructure
and

%>

@9 >

!
and

2
If

@

1
and

!

, then

.

. Therefore since

Noting that

2

, then

Q9
@
, we have

. If
(3.8)
9 X>
@
, as before we have

43
THE HILL-REUSS-VOIGT BOUNDS
3.2 THE HILL-REUSS-VOIGT BOUNDS

Until recently, the direct computation of micromaterial responses was very difficult.
Accordingly, classical approaches have sought to approximate or bound effective
responses. Many classical approaches start by splitting the
stress
field
within a sample

into a volume average, and a purely fluctuating part
and we directly
obtain

2

09

>
>

5
$ -
%
5
1 $

5

$ -
$

5

!
>

1
%
2
Similarly for the complementary case, with

sumption

2
>

(3.9)

, and the following as-

$ -
5
(3.10)
we have

5

*
*
5
5

$ 3 % $

% $ % 3 % %

$

$
%
%
%

$ -
5
5
5
0
5

$ %
$ %
-

%
(3.11)
Invoking Hills condition, which is loading independent in this form, we have

$ - $ %

%

(3.12)
where we emphasize
that
this
inequality means that the eigenvalues of the tensors

and
are non-negative. The practical outcome of the analysis is that bounds on effective properties are obtained. These bounds are commonly
known as the Hill-Reuss-Voigt bounds, for historical reasons. Voigt [183], in 1889,
assumed that the strain field within a sample of aggregate of polycrystalline material,
was uniform (constant), under uniform strain
exterior
loading. If the
constant strain

2

Voigt field is assumed
within
the
RVE,
,
then
,

which implies
. The dual assumption was made by Reuss [151], in 1929,
who approximated the stress
fields within the
aggregate
of
polycrystalline
material

.

, leading to
,
as uniform (constant),
"
"
"
"
44
(REUSS FIELD)
SPECIFIED
TRACTION
SPECIFIED
TRACTION
(VOIGT FIELD)
SPECIFIED
DISPLACEMENT
SPECIFIED
DISPLACEMENT
Fig. 3.2 Intuitive interpretations of the Reuss and Voigt fields with purely one dimensional
effects assumed.

. Equality is attained in the above bounds if the Reuss or

and thus
Voigt assumptions hold, respectively.
3.3 OBSERVATIONS
These results have a clear interpretation. To see this, consider a 1-dimensional bar of
length composed of random particles. There are a total of bands,
dark bands

( particles) and white bands ( particles), each of length
(upper
structure in Figure 3.2). We apply uniform strains on the structure and obtain the
following two point boundary value problem

9 9 > Q 9 > >
9> 9>

where is a constant. Simple calculations reveal that

(3.13)
9. >
(3.14)
. , and where is the volume fraction which equals . This
where
is exactly the Reuss bound, which is not surprising since the state of stress is constant
throughout the rod. Alternatively consider
3.2. The state
/9 . the lower
> structure
in Figure
of strain is uniform and we have
, which is the

Voigt approximation. The interpretation is clear, the response is bounded from below
(softer) by springs (moduli) in series (constant stress) and above (harder) by springs
(moduli) in parallel (constant strain). The bounds provide a rough and quick way of
determining approximate aggregate responses of microheterogeneous materials. The
wideness of the bounds grows with volume fraction, for example illustrated for an
45
CLASSICAL MICRO-MACRO MECHANICAL APPROXIMATIONS
aluminum/boron material combination (Figure 3.5). Material combinations such as

aluminum/boron are becoming more and more widely used in engineering because
they are lightweight and stiff. The boron particles are used as a stiffener for the easy
to form aluminum. Figure 3.5 illustrates the behavior of the bounds over the range
of all admissible volume fractions.
3.4 CLASSICAL MICRO-MACRO MECHANICAL APPROXIMATIONS

For completeness we review a few more classical approximations in micro-macro
mechanics.
3.4.1
The asymptotic Hashin-Shtrikman bounds
Improved bounds were developed in 1963 by Hashin and Shtrikman ([65], [66])
based on variational principles using the concept of polarization or filtering of
micro-macro mechanical fields. Based on these formulations they developed better
asymptotic bounds on effective properties. These bounds are sensitive to sample
size and are strictly valid only asymptotically useful when the sample size is infinite
relative to the microconstituent length scale. The Hashin-Shtrikman bounds are the
tightest possible bounds on isotropic effective responses, generated from isotropic
microstructure, where the volumetric data and phase contrasts of the constituents are
the only data known. The Hashin-Shrtikman principle represents a classical example
of the filtering of scales for materials with microstructure. It essentially involves
the Principle of Minimum Potential Energy (PMPE) written in terms of a filtered
variable that admits to a straightforward approximation of the internal fields. With
such an approximation
one can bound the macroscopic response.
One begins by

writing
, where
is spatially constant, and
, where
is
spatially nonconstant (heterogeneous). One defines polarization (filtered) variables,

1

,
and
, where

and
are
unknowns
in
a
filtered
variational
principle.
The
following
potential
is

constructed
"
"
9 >
where
"
"1
"
"
0.

!
"
">
"
0. 9 9 >
"
"
9
, @

"

"
>

!
and
">
1

"
(3.15)
"

. One can make the

following
observations: (1) The potential in Box

3.15
is
stationary
when
,
(2)
The stationary value is a maximum if
!
!

and a minimum if
, (3) In the isotropic case this becomes:

is an absolute minimum if
, and
is an absolute maximum if

, (4) When is a minimum
it is equal to the strain energy, (5) When

is vanishingly small compared to , then the variational principle collapses to
"
"

"
9 >
"
9 >
9 >
46
"
"
>

0 "

9 9 >
the PMCPE and (6) When

is infinitely large compared to , then the variational
principle collapses to the PMPE. Similar properties hold for a pure traction loaded
and
.
problem. The relations
are then
extremized by optimizing
with respect to

H

Noting that when

the strain energy is
,
one can arrive at, for a two-phase microstructure:

(

(

9 . >

.

are the bulk and shear moduli for the phases, while(3.16)
W is
and

!

(

(

where
the phase 2 volume fraction. The original proofs, which are algebraically complicated, can be found in Hashin and Shtrikman [65] and [66]. We emphasize that in
the derivation of the bounds, the body is assumed to be infinite, the microstructure
isotropic, and that the effective responses are isotropic. Also, a further assumption
is that
and . We remark that the bounds are the tightest possible,
under the previous assumptions, when no geometric (micro-topological) information
is included.
3.4.2

Consider the following identities: (I)

2

and (II)
direct manipulation we obtain
0!+# %
- "! # %
2
;=
;=

6 1 0!$# %
> 0' #%

N >

;=
'
;=
$ -

>
;=
'
;=
-

> "'(#%
Thereafter, we may write, for the variation in the stress
, which reduces to
> 0' #%

1

2
By
(3.17)
2"' #%
; = - > "' #% ' 1 0' # % 6 1 ; = 1 > "' #%

; = - 6 1 ; = 1 ' ; = - > /(> "'(#%
2
where
2
2

N> B

The concentration tensor: microfield behavior
(3.18)

.
is
known as the stress concentration tensor. Therefore, once either or

the other can be determined. In the case of isotropy we may write
+
"%
2
)
1 1
1 1 1
1 1 '
'
1 -

8
1 -
"
2
)
1 :
: 1 :
'
'
: 1 '
: -

: -
'
47
are known,
(3.19)
.
Clearly, the microstress fields are minimally distorted when
Remark: There has been no approximation yet. The burden in the computations
has shifted to the determination of . Classical methods approximate . For example,

the simplest approximation
is
, which is the Voigt approximation,

or for the Reuss
approximation

, where
.
'
'
9 9:9 > 9 9 > > > > > '
3.4.3
9 >
'
9 >
The Eshelby result
UNIFORM FAR FIELD IMPOSED
Fig. 3.3 The notation for the Eshelby formalism.
In 1957, Eshelby [29] developed an elegant formalism to determine the solution to

the elasticity solution of a single inclusion embedded in an infinite matrix of material
with uniform exterior loading. While alone this result is of little practical interest, this
solution, which is relatively compact, has been the basis of approximation methods for
effective properties for non-interacting and weakly interacting particulate solutions.
The essentials of the Eshelby result are as follows. Consider a single linearly elastic
particle, at infinitesimal strains, embedded in an infinite matrix: If the shape of the
particle is ellipsoidal, under uniform far field stress or strain, then the stress, and
hence the strain, in the particle is constant. As we have mentioned, this result is the
foundation of most methods of approximation for effective properties and is used to
approximate for ellipsoidal geometries. Ellipsoidal shapes are qualitatively useful
since the geometry can mimic a variety of microstructures (1) platelet behavior when
the ellipsoid is oblate with a high aspect ratio (2) needlelike microstructure when it is
prolate with high aspect ratio, (3) cracklike behavior as the material in the particle is
modeled as extremely soft with very high aspect ratios for the particle geometry and
'
48
(4) pores when the particle shapes are spheres with

very soft material inside. For a

single particle in three dimensions we have,
,
>
>
9 >
(3.20)

where is the far field uniform strain,
is the strain in the particle,

is the transformation strain and is the fourth-order Eshelby tensor which satisfies

. The Eshelby result, stated in a compact way, is:
and
are
constants in an isolated ellipsoidal particle under uniform far field loading.
3.4.4
Dilute methods
The use of the Eshelby result is straightforward as a method of determining approximate effective properties for regularly shaped particles. When the particle is a
sphere, is isotropic and the relations are relatively easy to write down. Breaking
the tensor relations into dilatational and deviatoric pieces yields,
one has

H

H
H

,
,
and
.

0
0

The constitutive relations are,
9 > , where

X , 587 587 ,
yields,

3 3

% , 9 > and

and
,
.

The approximations allow the determination of the strains in the particles as a func

. Combining the previous expressions

tion of the loading and geometry. Notice that the deviatoric and dilatational strain
components in the particle are the far field corresponding strain components multiplied by a factor composed by the corresponding components of the matrix material
divided by a convex combination of the matrix and particle materials. The transformation relations are tabulated for various simple particulate shapes in Mura [130].
It is important to realize that the stress fields in the matrix are not constant, and are
extremely complicated. Fortunately, in developing approximate expressions for the
effective response it unnecessary to know the fields in the matrix. Consequently, the
concentration tensor, under the non-interacting particle assumption, is isotropic and
can be written as

I
'
I
I

'

(3.21)
Therefore, in the case of noninteracting spheres, is isotropic with deviatoric and

dilatational components given by and , which are furnished by the Eshelby solution.

This is the framework for the Dilute method, which postulates the complete non

interaction
between
microstructural components. Therefore we have

implies
G9
> '
49
G9 >
JS

W9 >

(3.22)
Dilute approximations are usually only accurate at extremely low volume fractions.
3.4.5
The Mori-Tanaka method
Non-interaction of particulates is an unrealistic assumption for materials with randomly dispersed particulate microstructure at even a few percent volume fraction.
Weak interaction between particles, via a sensitivity to increased volume fraction,
involves slight modification of the
dilute method.
This involves assuming that
there

.

,
thus
implying
exists another tensor, such
that

be approximated.

. The tensor
must

There are two requirements: (1)
and (2)
. The
most widely used approach is that of Mori-Tanaka [128].
The
Mori-Tanaka
method

.
2
, thus implying
proceeds
by
constructing
a2tensor
such that

2

that
2

. The Mori-Tanaka method approximates by . The physical implications are that the particle sees a matrix material with an average stress
state provided by the average in the matrix. One can consider the Mori-Tanaka
as

one
possible
curve
fit
that
satisfies
the
two
mentioned
conditions,
i.e.
when
1

. The result is
and
when
9D >

P " '
P

9

>

'

. '

9
>

9
> (3.23)

JS

X
where
,
X , 587 587 , )

, where
,
and
. The interaction, via a
9 9

> >

volume fraction dependency, is explicitly clear in the expressions for and . It is

noted that for spherical particles the Mori-Tanaka matches the well known HashinShtrikman lower bound [66] on effective properties.
Remark: Another approach to incorporate particulate interaction is the Self Consistent method. The idea is simply that the particles see the effective medium instead
of the matrix in the computation of the transformation strain. In other words the matrix material in the Eshelby tensor is simply the effective medium. Therefore one
simply replaces the matrix material in the concentration formulae (only in ) with
the effective property to obtain
'

9 . > JGS
and where

9. >
I

(3.24)

. The approach is iden

where

tical to the dilute approximation, with the result being somewhat more complicated
50
A
I
v =1
2
v =0
2
Fig. 3.4 The Mori-Tanaka extrapolation.
expressions for the approximate effective property, however, which can be computed
in closed form in many cases. We refer the reader to the books of Mura [130], Aboudi
[1] and Nemat-Nasser and Hori [131]. The self-consistent method produces negative
effective bulk and shear responses for voids above volume fractions of 50 . For
rigid inclusions it produces infinite effective bulk responses for any volume fraction
and infinite effective shear responses above 40 . For proofs and discussions on
variations of this method, see Aboudi[1]. With these facts in mind, one can safely
employ such methods only at very low volume fraction levels.
Remark: For more indepth analyses, with extensions into nonlinear behavior,
using blending analytical, semi-analytical and numerical techniques, we refer the
reader to the extensive works of Llorca and co-workers: Segurado and LLorca [153],
Gonzlez and LLorca [56], Segurado et. al [154], Gonzlez and LLorca [57], LLorca
[111], Gonzlez, C. and LLorca [58], Poza and LLorca [148], LLorca and Gonzlez
[113], LLorca [114], LLorca [115].
3.5 MICRO-GEOMETRICAL MANUFACTURING IDEALIZATIONS

Obviously it is virtually impossible to manufacture particles with perfectly smooth geometries. However, if one were to attempt to simulate a samples response with highly
irregular shapes, a problem of such complexity would occur that virtually no analytical, semi-analytical or standard numerical technique would suffice. Fortunately, in
the computation of effective properties, the shapes can be safely approximated by
much smoother geometrical idealizations, which is the subject of this section.
3.5.1
Upper and lower variational bounds
Variational inequalities are now derived that govern the growth or decrease in system
energy due to the local stiffness reduction. With these inequalities, bounds can be
developed for the effective response for irregularly shaped particles by creating outer
MICRO-GEOMETRICAL MANUFACTURING IDEALIZATIONS
51
240
220
EFFECTIVE BULK MODULUS K*
200
180
REUSS
VOIGT
HASH/SHTRHASH/SHTR+
160
DILUTE
MORI-TANAKA
140
120
100
80
60
0
0.2
0.4
0.6
VOLUME FRACTION
0.8
1.2
0.4
0.6
VOLUME FRACTION
0.8
1.2
180
REUSS
VOIGT
HASH/SHTRHASH/SHTR+
DILUTE
140
MORI-TANAKA
EFFECTIVE SHEAR MODULUS MU*
160
120
100
80
60
40
20
0
0.2

Fig. 3.5 The predicted effective bulk ( ) and shear moduli ( ) for an Aluminum matrix
3
3
2 3
*

*
*
*

(
,
)/Boron particle (
,
).

52
MANUFACTURING
REALITY
LOWER
IDEALIZATION
UPPER
IDEALIZATION
Fig. 3.6 Upper and lower idealizations of shapes that can be manufactured.
and inner smooth envelopes, as shown in the Figure 3.6. In order to do this the
principle of virtual work and the complementary principle are employed.
Consider two symmetric positive definite material property (elasticity tensor) dis

tributions, and used in same boundary value problem.
The corresponding

stress and strain states using these materials are denoted
and
respectively. Consider the respective strain energies of both materials:
"
&

(
&
&

(

&
(
$

&
"
"%
&

0

&
Similarly, defining

,
(
&

&
(
0
&

(
$

, if

.

,
(
&

(
$
1 $

, one obtains, assuming, 9 >

Denoting

9 > 9 >
&
(
0
&
(

.

(3.25)
(3.26)
53
"
*

(
&
$ - 5
& (
% 1 $

(
&
&"
"

9 >
and thus, assuming

(
&
%1 $

%

$ - 5
& (

(
&

(
&

&
(

5
0
1 $
%

$
-
&
(
(
&

5
0

E
(3.27)

5
0
$
-
&
(
$
/
&
5
(

%

5
0
(3.28)
The general proofs are constructive, and we believe add extra insight the analysis to
follow later.
3.5.2
Proof of energetic ordering

9 > 9 >

Consider two symmetric

positive definite material property (elasticity tensor) distri
The corresponding
butions, and used in same boundary value problem.

stress and strain states using these materials are denoted
and
respectively. Consider their respective strain energies:
"
&

(
We have, denoting

*

(
$

&
&
&
(

&

&
/
(
&

*
&
/
(

&

/
(
&

&
(

(
(
&

(
&
1 $

%

&
(

(
&

(
(
&
&
/
&
(

(
&

(

,
,
(

&
&
(
&
(

&

&
(

0

(3.29)

0

.
(
&
(

&

(

(
"%

,

&
"
%

(
&

(
%

54

*

(
&
"
&
&

(
&

,
(
&
& (
& (

(

0
&

%

(

7
9
(3.30)
>
The vanishing terms in the preceding analysis terms are

due to a direct application

! implies
of the principle of virtual work. The end result is that
&

(
(
&
%1 $

%

&

(
Similarly, defining
and

(

, if

%1 $

H

&

(

&

(3.31)

(
&
(3.32)

9
> 9 %>

(3.33)
and we have

&

(

5
0
$ - )5
& (
$
-
&
(
&

(
&

(

5

&
(

"

$ & (

&
$ & (

&

&
5
&
(
&
&

(
&

(
&
(
5
$ -
5*
0

(
5
$ -
&
(
$ -
&
(

$ -
&
(

(
5

5

$ -
&
(
$ & (

$ & (
$
-
&
(
(
&

/

(
&

&
(
$
-
&
(

5

$ -
&
(
5

5

5*
0
$
-
&
(
5

(3.34)
55
9 >
&
5
Therefore

(
$

$ - )5
& (
&

(
implies

&
(
0

5*
$
&
5
(

%

$ -
&
(

5
0

(3.35)
Similar results, however for more restrictive (uniform) loading cases, and not in the
context of a microfailure analysis, can be found in Huet, Navi, and Roelfstra [79].
3.5.3
Uses to approximate the effective property
Under the special case of uniform test boundary loadings, one has

YN >
Y>

G
J ,S
JS

9

9

is assigned the real microstructure composed of irregular
Therefore, if

shapes,
, and represents a microstructure with a smoother particu
late geometry that envelopes the real particles
(Figure 3.6), denoted
, with corre

!

sponding effective property
, then
and
or " .
is assigned the real microstructure composed of irregular
Alternatively, if

represents a microstructure with a smoother par
shapes,
, and

ticulate geometry that is enveloped by the real particles (Figure
3.6), denoted ! ,

or
with corresponding effective property
, then
and

"
. Therefore, in total,

(3.36)
These relations, for uniform boundary conditions,appeare to have been first developed
in Huet et. al. [85], based on the results of Hill [74]. The result has been coined
by Huet et. al. [85] as the Hill Modification Theorem. Of critical importance is
the fact that one can safely use smooth idealizations to approximate rougher more
geometrically complicated shapes that really occur during manufacturing.
4
A basic finite element
implementation
Classical approximation techniques construct approximations from globally kinematically admissible functions. Two main obstacles arise (1) it may be very difficult
to find a kinematically admissible function over the entire domain and (2) if such
functions are found they lead to large, strongly coupled, and complicated systems
of equations. These problems have been overcome by the fact that local approximations, i.e. over a very small portion of the domain are possible that deliver adequate
solutions, and simultaneously lead to systems of equations which have an advantageous structure amenable to large scale computation by high speed computers. This
piece-wise or element-wise approximation technique has been recognized at least
60 years ago by Courant [25]. There have been a variety of such approximation methods to solve equations of elasticity. The most popular is the Finite Element Method
(FEM). The central feature of the method is to reduce the continuum field equations,
in a physically sound and systematic manner to an assembly of discrete subdomains
or elements. The process is designed to keep the resulting algebraic systems as
computationally manageable, and memory efficient as possible. The implementation, theory and application of FEM is a subject of immense literature. We in no way
will attempt to review this huge subject. For general references on the subject see the
standard books of Bathe [10], Becker, Carey and Oden[11], Hughes[92] and Szabo
and Babuska[172].
4.1 FINITE ELEMENT METHOD IMPLEMENTATION

Consider the following general form:
57
58
A BASIC FINITE ELEMENT IMPLEMENTATION
+-"
*

-

(
#
/&
#

- *

<*

-

0

*

+8
(4.1)
E
where the last term is to be thought of as a penalty term to enforce the

applied
is a
displacement (kinematic) boundary conditions. The (penalty) parameter
large number, whose selection will be discussed later in the presentation.
Remark: The ability to change the boundary data quickly will be very important
later in the finite element computations, and is the hallmark of the penalty method
in the context of a virtual work formulation. This is done by relaxing kinematic
assumptions on the members of the space of admissible functions and adding a term
to account for the violation on the exterior boundary. This is a widely used practice, and therefore to keep the formulation as general as possible we include penalty
terms, although this implementation is not mandatory. Obviously one could simply
condense out the known (imposed) values of boundary displacements. Nevertheless
we consider the (exterior) penalty method formulation for generality, although one
does not necessarily need to use it. For the sake of completeness, more discussion on
penalty methods is given at the end of this chapter.
4.2 FEM APPROXIMATION

It is convenient to write the bilinear form in the following (matrix) manner

6

*

, the deformation tensor, is

" * "

*
%

(4.2)
E
where

DE
"
DE
%
" ++
+
(4.3)
It is clear that in an implementation of the finite element method, the sparsity of

should be taken into account. It is also convenient to write
59
FEM APPROXIMATION

1

or

-

-

-
1
E

1
E
(4.4)
4
E
[ , where, for example, for trilinear shape functions
-

"

=
(4.5)
It is advantageous to write

"

1

A6

1

"

A6
! #"% $ "

0
Since
(4.6)
and

combination

"

we choose with the same basis, but a different linear
[ [ . ,Ifthen
we may write
4

"

"

"%
*

A6
%
A6

This is the system of equations that is to be solved.

%

(4.7)
, therefore

"

! "$ $ 1

is arbitrary, i. e. the weak statement implies

(4.8)
60
4.3 GLOBAL/LOCAL TRANSFORMATIONS

One strength of the finite element method is that most of the computations can be
done in an element by element manner. We define the entries of
,
"!

9: [ > ; 89 [ ! >

;
[ [ ! +
(4.9)
and

;
[ +

;
[ +
Breaking the calculations into elements,
4!

(4.10)
4! "! , where

9 [ > ; 9: [ ! >

;
[ +

9 >

[ ; [ ! +
(4.11)
In order to make the calculations systematic we wish to use the generic or master
. Accordingly, we need the
element defined in a local coordinate system
following mapping functions,
coordinates to the real space coordi from
the(formaster

example trilinear bricks):
nates,

9 > 9 >
"%
"%
"%

*
-
-

(4.12)

:

9

>
9

are true spatial
coordinates
of the ith node, and where [ >
where
9 9 $> W9 > W9 $>:> . These types of mappings are usu[
ally termed parametric maps. If the polynomial order of the shape functions is as high
4
-

as the element, it is an isoparametric map, lower, then subparametric map, higher,

then superparametric.
4.4 DIFFERENTIAL PROPERTIES OF SHAPE FUNCTIONS

The master element shape functions form a nodal bases of trilinear approximation
given by:
DIFFERENTIAL PROPERTIES OF SHAPE FUNCTIONS
MAPPED
ELEMENT
X2
1
MASTER
ELEMENT
X1
Fig. 4.1 A two-dimensional finite element mapping.
3
8
1
2
Fig. 4.2 A trilinear hexahedron or "brick".
61
62
2
*
2
2
2
*
2
-
2
2
2
2
2
0
2
2
-
2
2
2
E
2
2
2
E
2
1
2
2
2
2
1
2
(4.13)
E
For trilinear elements: Here we have a nodal basis consisting of 8 nodes, and
since it is vector valued, 24 total degrees of freedom, or 3 shape functions for
each node.
For triquadratic elements: Here we have a nodal basis consisting of 27 nodes,

and since it is vector valued, 81 total degrees of freedom, or 3 shape functions for
each node. The nodal shape functions can be derived quite easily, by realizing
that it is a nodal basis, i.e. they are unity at the corresponding node, and zero
at all other nodes., etc...
[
We note that the s are never really computed, we actually start with the s.
Therefore in the stiffness matrix and righthand side element calculations, all terms
must be defined in terms of the local coordinates. With this in mind we lay down
some fundamental relations, which are directly related to the concepts of deformation presented in our discussion in continuum mechanics. It is not surprising that a
deformation gradient reappears in the following form:

"

- E

4

"

!

(4.14)
The corresponding determinant is

*

-

4
The differential relations

-

4

4

4
(4.15)
, are

The inverse differential relations

, are

E
(4.16)
63
DIFFERENTIATION IN THE REFERENTIAL COORDINATES

and
and the inverse form

0

0
(4.17)
E

-
(4.18)

*

(4.19)
Noting the following relationship

where

-3*

*

*

*
-/4
131
4B-
434
$ - *

"
!

1A4

E
41

-3-
1@-
-$1

=
1@4B-

-$1

131
41
-/4
(4.20)
1A4
434

With these relations, one can then solve for the components of
and

(4.21)

, via
We now need to express
in terms

9 [ 9 > > 9 [ 9 A 9 > A 9 > A 9 >:>
(4.22)
4.5 DIFFERENTIATION IN THE REFERENTIAL COORDINATES
64

Therefore we write for the first column of

(4.23)

0

for the second column

0
and for the last column

(4.24)
E

(4.25)
[ ) has the following form for

0

For an element, our shape function matrix (

linear shape functions, for the first eight columns
-

(4.26)

for the second eight columns

for the last eight columns

-
(4.27)

(4.28)
E
1 This is for illustration purposes only. For computational efficiency, one should not program such operations in this way. Clearly, the needless multiplication of zeros is to be avoided.
DIFFERENTIATION IN THE REFERENTIAL COORDINATES
[ is a

which in total is a 3 24 matrix. Therefore the product
the form, for the first eight columns

0

0

*

0
6

% $

E

""

E

(4.31)

(4.30)
E
(4.29)

% matrix of
6 6
6
Finally with quadrature for each element

and for the last eight columns

0

and for the second eight columns

0

65

%
(4.32)
E
and

*
6 6
*
% $

""

-
6

6
%

%
(4.33)
E
66
where , etc, are Gauss weights.

Remarks: In other words, it is permitted to have material discontinuities within
the finite elements.
1 On, the implementation level, the system of equations
to be

,
where
the
stiffness
matrix
is
represented
by
solved
are

, where I,J are the global entries, and i,j are the local entries. The
amount of memory required with this
relatively
simple storage system is, for trilin

%
%
ear
hexahedra,
the
number
of finite elements, where the
are the individual element stiffness matrices. If matrix
symmetry
is taken into

%
%
account, the memory requirements are
the number of

finite elements. This simple approach is so-called element by element storage. The
element by element storage is critical in this regard to reduce the memory requirements. 2 Here a global/local index relation must be made to connect the local entry
to the global entry when solution time begins. This is a relatively simple and efficient
storage system to encode. The element by element strategy has other advantages with
regard to element by element system solvers. This is discussed later.

- + $>
0 10
> .
0 0

9 >
>
4.6 SURFACE INTEGRALS

The surface integrals also are computed of the master

element, and require delicate

=position
transformations. The following relation is critical (
vector)

9

9

$>

> W9

>

where, depending on the surface on which it is to be evaluated upon, one of the

plane,
components will be +1 or -1. Generally, for the

*

whose absolute value is evaluated at

"

4
-/4
. or . . For the plane
2 If a direct storage of the finite element storage of the stiffness matrix were attempted,
stations.
(4.34)
(4.35)

2
memory would be required, which, for large problems is an impossibility for most work-
POST PROCESSING
67
X3
n
dA
X2
X1
Fig. 4.3 A surface element.

-

1



1
. or . . For the

"

"

(4.36)
1A4
E
plane
. or . .
(4.37)
-$1
E
4.7 POST PROCESSING

Post processing for the stress, strain and energy from the existing displacement solution, i.e., the values of the nodal displacements,the shape functions, is straightforward.
Essentially
is the same as the formation of the virtual energy in the system
the process

. Therefore, for each element
$
68

-/4

-3
131

434
3
-$1
3
1A4
3

%

(4.38)
E
where the global coordinates must be transformed to the master system, in both the
deformation tensor, and the displacement representation. At each Gauss point, we
add up all eight contributions for each of the six components, then multiply by the
corresponding nodal displacements that have previously been calculated. The following expressions must be evaluated at the Gauss points, multiplied by the appropriate
weights and added together:

-
-
-

-

E

-
-

-
-

-

E

-
-
-

(4.39)
E
denotes the component of the displacement of the ith node. Combining

0

0
0
the numerical derivatives
to form the strains
we obtain
:
A and
A
A , :
A A
A , , and A ,

A
A .
where

4.8 ACCURACY OF THE FINITE ELEMENT METHOD

As we have seen, the essential idea in the finite element method is to select a finite
dimensional subspatial approximation of the true solution and form the following
weak boundary problem:
+-"

%

&
*

- *

-
E
*

*
8
*

-

E
*

(
#
/& #
E

&

(

(4.40)
69
ACCURACY OF THE FINITE ELEMENT METHOD
9 N>

9 N>
&
9 N>

9F

The critical point is that
. This inner approximation
allows
the
development
of
straightforward
subspatial
error
estimates. In most cases
&
&

and
coincide. We have for any kinematically admissible function
a definition of the so-called energy norm
9 N>

9 N>

1
"% - *
% (
&

- * -

5 5

9 >

(4.41)
Note that in the

event
that
displacements are specified
on the boundary,

nonconstant

!
then
is unobtainable unless
, and the semi-norm
in Equation (4.41) is a norm in the strict mathematical sense. Under standard assumptions the fundamental a-priori error
estimate for the finite element method is

"
"

, where is the (complete) polynomial order of the finite element method used, is the regularity of the exact solution, is a
global constant dependent on the exact solution and the polynomial approximation.
is independent of h, the maximum element diameter. For details see, for example,
Hughes [92]. Related forms of this estimate holds locally, but
with
constants
that

F

,
thereare element dependent.
To
prove
this
select
any

, where
fore "
, is a

"

generalization
have
,
and

of the
result with . We
F
,
, implying a Galerkin (Figure 4.4)

orthogonality property of inner approximations:
C
7

9 >

9 >
**

E

'

> 9 >
9 >

where the error is defined

by
can be represented
by
we have

9
9

sense:
With this we finally have

"

9 N>
7 >
9 N>
9 N>

- 9 .
9 > 9 >
9 N>
9 N>

-
(4.42)
. 0 Therefore
any member
the subspace

this ofrepresentation

9 > 9 > . Using
9 > , which implies
>
. We recall that
2 is bounded
; in the following

E

. 9 > . 9 >

'

"

"
(4.43)
This result holds locally as well, however with the constant and exponent being
different for each element.
70
TRUE
SOLUTION
RESIDUAL
ERROR
APPR.
SOL.
APPROXIMATION SPACE
Fig. 4.4 Orthogonality of the approximation error.
4.9 MINIMUM PRINCIPLES AND KRYLOV METHODS

There are two main approaches to solving systems of equations resulting from numerical discretization of solid mechanics problems, direct and iterative. There are a large
number of variants of each. Direct solvers are usually employed when the number of
unknowns is not very large, and there are multiple load vectors. Basically one can operate on the multiple righthand sides simultaneously via
Gaussian
elimination.
For a

back substitution the cost is
.

Therefore the total cost of solving such a system is the cost to reduce the system to
.
upper triangular form plus the cost of back substitution, i.e.

However since the operation counts to factor and solve an
system are
,
iterative solvers are preferred when the systems are very large. 3 Gaussian
elimina
tion is approximately

. In general, most modern solvers, for large symmetric systems,
like the ones of interest here, employ Conjugate Gradient (CG) type iterative techniques which can deliver solutions in
operations. It is inescapable, for almost
all variants of Gaussian elimination, unless they involve complicated sparsity tracking to eliminate unneeded operations on zero entries, that the operation costs are
. However band solvers, which exploit the bandedness of the stiffness matrix, can reduce the number of operation counts to , where is the bandwidth.
Many schemes exist for the optimal ordering of nodes in order to make the bandwidth
0
"
9 >
0 3
0
9 . . > & 9 .$> 9 . >
9 .$>
9 >

0
0 3 0

9 9 .$> 9 > > . > &

9 >
3 An operation such as the addition of two numbers, or multiplication of two numbers, is one operation
count.
MINIMUM PRINCIPLES AND KRYLOV METHODS
71
as small as possible. Skyline solvers locate the uppermost nonzero elements starting from the diagonal and concentrate only on elements below the skyline. Frontal
methods, which are analogous to a moving front in the finite element mesh, perform
Gaussian elimination element by element, before the element is incorporated into
the global stiffness matrix. In this procedure memory is reduced, and the elimination process can be done for all elements simultaneously, at least in theory. Upon
assembly of the stiffness matrix and righthand side, back substitution can be started
immediately. If the operations are performed in an optimal order, it can be shown
that the number of operations behaves proportionally to
. Such a process, is, of
course, nontrivial. We note that with direct methods, zeros within the band, below
the skyline, and in the front are generally filled and must be carried in the operations.
In very large problems the storage requirements and the number of operation counts
can become so large that solution by direct methods is not feasible. The data structures, and I/O are also non-trivial concerns. However, we notice that a matrix/vector
multiplication involves
operation counts, and that a method based on repeated
vector multiplication, if convergent in less than iterations, could be very attractive.
This usually the premise in using iterative methods, such as the Conjugate Gradient
Method (CG). A very important feature of iterative methods, is that the memory requirements remain constant during the solution process. It is important to note that
modern computer architectures, are based on (1) registers, which can perform very
fast operations on computer words, but have virtually no memory capabilities (2)
cache, which slightly larger memory capabilities, with a slight reduction in speed, but
are thermally very hot, and are thus limited for physical as well as manufacturing
reasons (3) main memory, which is slower since I/O is required, but still within a
workstation and (4) disk and tape or magnetic drums, which are out of core, and thus
require a huge I/O component, and thus are very slow. Therefore, one point that we
emphasize is that advantage is taken of the element by element structure inherent
in the finite element method for data storage and matrix vector multiplication in the
CG method. The element by element data structure is also critical for the ability to
fit matrix/vector multiplications into the computer cache, which essentially is a low
memory/high floating point operation per second portion of the computer hardware.
Remark: One singularly distinguishing feature of iterative solvers is the fact that
since they are based on successive updates of a starting guess solution vector, they
can be given a tremendous head start by a good solution guess, for example, provided
by an analytical or semi-analytical solution. We now discuss minimum principles,
which play a key role in the construction of a certain class of iterative solvers valuable
in solving micromechanics problems.
0
4.9.1
Krylov searches and minimum principles
By itself, the PMPE is a powerful theoretical result. However it can be used to develop
methods to solve systems of equations arising from a finite element discretization of
a infinitesimal strain linearly elastic structure. This result is the essence of the socalled Krylov
family of searches. Suppose we wish to solve
the discrete system

, .
is a symmetric positive definite
matrix,
is the
$

72
is the . righthand side. We define a potential

. solution
vector, and
; ; . Correspondingly, from basic calculus we have
. Therefore the minimizer
@

;
,

"!
of the potential is also the solution to the discrete system. A family of iterative
solving techniques for symmetric systems based upon minimizing by successively
updating a starting vector are the Krylov class. The minimization takes place over
vector spaces called the Krylov spaces. These methods are based on the hope that a
solution to a tolerable accuracy can be achieved in much less than
operations,
as required with most Gaussian-type techniques. The simplest of this family is the
method of steepest descent, which is a precursor to the widely used Conjugate Gradient
Method.
9 >
4.9.1.1 The Method of Steepest Descent The Method of Steepest Descent

is based upon the following simple idea: if the gradient of the potential is not zero
at a possible solution vector, then the greatest increase of the scalar function is in
. The
the direction of the gradient, therefore we move in the opposite direction

,
, and
ingredients in the methods are the residual,
@

C @ C

$.; We seek
the successive iterates,

a relation
$; such

that

C

C

C

$

$

is a global
minimum.
Directly
we
have

$; $ $; $; , where
it was
assumed
that

+

"!

. Therefore the method of steepest descent consists of the following:

2

*

3 !#"$%

& & * * (
(

*

'&
!

)&
&
&

(4.44)

!#"$%
( " ( * *
'&

was symmetric. Forcing
,and solving for
+* (
+* (
1,
1,
)&
)&
-!
.-!
-!
#!/-!0.

3

$ A $
9.
>
A $
The rate of convergence

is related
to21 the condition
number of the
of the
method

, where
stiffness matrix "

'9 >

43

. The rate of convergence of the method is typically

quite slow, however, a variant, the Conjugate Gradient Method, is guaranteed to
converge in N iterations at most, provided the algebra is performed exactly.

73
MINIMUM PRINCIPLES AND KRYLOV METHODS
9 >

$

@
such that

. We seek
and the successive iterates, $
a

0
:

3

minimum.
is a global
search vector, $
is computed
iteration
. , awhere
The , first
0:3
$

by
the update vector is defined as
is

W5 , for - .
;. C The
coefficient

+
chosen
so that is
to
, i.e.

. Therefore, the value of which minimizes

0. 9 > ; 89 > 9 > ;
$

$

(4.45)
1023

.
is ( ),
. The solution steps
are:
4.9.1.2 The Conjugate Gradient Method In the Conjugate Gradient Method

At each iteration the computational cost is
. We refer the reader to Axels

,
,
son [8] for details. We define the residual,

+
+
+

! +
!#"$%

3
(

!

!#"$%

-
&
*
1

-
&

"%

2 3

0

$ -

'&
*
-
*

&

&

*

*
&
(

!#"$%

*
*
* * * *

!#"$%
"
(

!
*
'&
2 !#"$%
+
)&
)&

*
*
-

2

*
-!
-!
#!-!0

(4.46)
A T JQM V%L O V OGS
4.9.1.3 Accelerating computations

rate of convergence
of the CG method
The

)
is related to the condition number (
74

*

*

-

2

2

*
-

-
(4.47)
Proofs of the various characteristics of the method can be found in Axelsson[8]. As

is standard, in an attempt to reduce the condition number and hence increase the rate
of convergence, typically preconditioning
of
is done by forming the following
1
transformation of variables,
, which produces a preconditioned system,
with stiffness matrix
. Ideally we would like

where
,
and
where
is
a lower triangular matrix, thus forcing

. However, the reduction of the stiffness
matrix into a lower triangular matrix and its transpose is comparable in the number
of operations to solving
the system by Gaussian elimination. Therefore only an

approximation to
is computed. Therefore inexpensive preconditioners are
usually used. For example diagonal preconditioning, which is essentially
the least

expensive involves defining
as a diagonal matrix with entries

, where the are the diagonal entries of . In this case the resulting
terms in the preconditioned stiffness
matrix are unity on the diagonal. The off diagonal

terms,
are divided by
. There are a variety of other preconditioning

$ ?;)
*;) % % ?; % N8 %?; N ; N
%

.

4!
!
N ;

- +
E
techniques, of widely ranging expense to compute. For more details see Axelsson
[8]. It is strongly suggested to precondition the system. For example with the simple
diagonal preconditioner we obtain the following stiffness matrix

-

-

-
/
/ /
-

-

2
-

2

-

-
/
-
-
/ /

2

2

(4.48)
4.10
EXTERIOR AND INTERIOR AUGMENTATIONS
As we have mentioned earlier, the exterior point (penalty) method was used to enforce
displacement boundary conditions. in the case where a potential exists We can now

motivate this by an augmented potential

,
, whose variation is
>
9 N>
>
9 >
> 9
+-"
*

)5

(
#
/&
#

<*

-

75

8

0

+8

(4.49)
Therefore, the penalty term can be thought of as a quadratic augmentation of the

potential energy. When no potential exists, the exterior point method can only be
considered as an enforcement of a constraint.
Remark: Throughout this entire document, the penalty method has been used to
enforce boundary conditions. The penalty parameters have been varied repeatedly in
simulations and the solutions have been checked to be independent of such parameters
beyond threshold values. This is primarily due to the fact that they produce isolated
increases in the spectrum of eigenvalues of the stiffness matrix, and thus affect the
solver minimally. It is a distributed increase in the spectrum which is a problem. This
is not the case with the penalty method. At any rate such augmentations have minimal
effects on iterative solver performance due to the fact that preconditioning is used.
In short, the penalty parameters are not an issue of importance in the simulations
presented later. For completeness, augmentation methods are discussed further next.
4.10.1
9
Properties: exterior/penalty functions
9 > ' "9 >
9>
& > 9 > 9 > 9 > ,

, if and only

Let us define the

following unconstrained
function
where (1)
, (2)
and (3)

if
, the feasible region.
is theminimizer
of
for weight
problem

9 2> 9
9 > 9 >
9 U> 9 >
9 > 9 >
9 > P
with the constraints:
/
+
1
We have
following properties, where
is the
and
the true minimizer for the exact

9 >
9 >
9 >
9 > 9 >
9 >
9 >

+
+
>
+
(4.50)
The important point is that for increasing penalty weights, we are assured to converge
to a feasible minimum. In general one uses a series of weights, and stops at finitely
large, penalty values. There are a variety of minor variations to such approaches since
their introduction in the late 1960s, but the essential idea has remained unchanged.
For more details we refer the interested reader to reviews in Fiacco and McCormick
76
[31] or Luenberger [117]. For the sake of completeness, we derive some fundamental
results.
4.10.1.1
Proofs for the penalty method Consider the following:
1. By definition:

-

-

-

0

0

-
"%

2. By definition:

- -
- -

-

-
(4.51)

0
and
. Adding the two previous results yields

.

3. By definition:

"

0

"
*
4. For each K,

"%

0

0
0

-

0

-

(4.52)
"

7

5. Utilizing the previous results, one can show that the method converges. Let - ,

, be a convergent subsequence of - , having limit . Therefore, by continuity of

* 5

we have 0-+-% . From the previous results

*

E
0 +-%
. There - - which
5implies
* 0-+-%
- *
0-+-%
E
which
implies
fore, E

*
. Since
, and therefore is feasible.
must be
- is continuous,
5

* 0-+-%
optimal since,
and
.
4.10.2

-
"

-

Properties: interior/barrier functions
9>9
&>
The dual to exterior/penalty formulations, so-called barrier formulations are now

discussed.
Let us
the following
unconstrained
function,

define

(1)
, (2)
and (3)

, where
, if
. We have
the
following
properties,
where
is the

with weight and is the true minimizer for the exact problem
minimizer of
with constraints:
9 9 > > 9 >

9 > 6' " 9 >
77

9 2> 9 > 9 >
9 > 9 > 9 >
9 U> 9 > 9 >
9 > 9 > 9 > 9 >
9 > P " 9 > 9 >
/
+
+
+
+
(4.53)
These results are proven in a straightforward manner.

4.10.2.1
Proofs for the barrier method Consider the following:
1. By definition:

0
0

(4.54)
"

-

-

and
- - - - -. Adding the
two previous results yields
- - .
-

- -
- -

2. By definition:

- -

-
3. By definition:

-

-
"%

4. For each K:

-

-

-

-

- -
- - "

-
"
*

-

-
(4.55)

.
5. Utilizing the previous results, one can show that the method converges. Let - ,

, be a convergent subsequence of - , having limit . Then by continuity

*

of we have 0-+-% . From the previous results

5

0-+-%
*

E
. Therefore
5- E* - 0 +-%

0-+-%
*
.

E
. Since
5 is continuous
is finite, and therefore is feasible.
must be optimal since,

5
*
* 0 +-%
and by continuity
.
-

"

+

Remark : The same comments go for the numerical conditioning of Barrier

methods as did for Penalty methods, i.e. the barrier parameters must be changed
gradually. There are a variety of minor variations to these types of approaches, since
their introduction in the late 1960s, but the essential idea has remained unchanged. A
78
primary work concerning this approach is found in the classical monograph Nonlinear Programming: Sequential Unconstrained Minimization Techniques (SUMT), by
Fiacco and McCormick [31]. This work summarizes and extends a series of papers
by Fiacco and McCormick on the subject, and serves as a primary source of results
related to penalty and barrier methods. In addition it contains a detailed historical
survey of these methods.
5
Computational/statistical
testing methods
The amount of uncertainty in the effective property bounds, as well the magnitude of
the internal solution difference bounds, leads one to consider direct numerical simulation, whereby the effective responses can be obtained by volumetrically averaging
(post processing) numerical solutions of boundary value problems representing the
response of aggregate assemblages of heterogeneous material. Accordingly, in this
chapter we investigate topics related to the numerical simulation of the testing of
mechanical responses of samples of microheterogeneous solid materials formed by
aggregates of particulates suspended in a binding matrix. Consistent with what is
produced in associated manufacturing methods, the microstructures considered are
generated by randomly distributing aggregates of particulate material throughout an
otherwise homogeneous matrix material. Therefore, the resulting microstructures are
irregular and nonperiodic. A primary issue in testing such materials is the fact that
only finite sized samples can be tested, leading to no single response, but a distribution of responses. A technique employing potential energy principles is developed
to interpret the results of testing groups of samples. Three dimensional numerical
examples employing the finite element method are given to illustrate the overall analysis and computational testing process. The basis for this chapter follows from work
found in Zohdi and Wriggers [200].
5.1 A BOUNDARY VALUE FORMULATION

We consider a sample of heterogeneous material (Figure 5.1), with domain , under

a given set of specified boundary loadings. The weak boundary value problem is
79
80
COMPUTATIONAL/STATISTICAL TESTING METHODS
X2
X1
X3
Fig. 5.1 A cubical sample of microheterogeneous material.

S=
@
9 N>

L M1HVHZJ V

9 N>

!
(5.1)
When we perform material tests satisfying Hills condition,
we
have,
in
the
case
of

displacement controlled tests (loading case (1))
and
or
for
1

traction controlled
tests (case (2)) H
and
. In either case we

!

consider
and that the material is perfectly
bonded.
However,
we note that
!

(1) Hills condition is satisfied with
(and no
debonding)
and
in
case (2) it is

!

satisfied even with debonding, however only if
. The boundary value problem

in Box 5.1 must be solved for each new sample, each possessing a different random

microstructure ( ). The solution is then post processed for the effective quantities.
It is convenient to consider the RVE domain as a cube, and we shall do so for the
remainder of the work.

5.2 NUMERICAL DISCRETIZATION

In order to computationally simulate effective responses, our choice for spatial discretization is the finite element method. There are essentially two choices to mesh
NUMERICAL DISCRETIZATION
81
the microstructure with the finite element method, an unaligned or an aligned approach. We refer to an unaligned approach as one which does not require the finite
element boundaries to coincide with material interfaces when meshing the internal
geometry (Figure 5.2). This leads to material discontinuities within the finite elements. An aligned approach would impose that the element boundaries coincide
with material interfaces and therefore the elements have no material discontinuities
within them. There are advantages and disadvantages to both approaches. Unaligned
meshing has the advantages of rapid generation of structured internal meshes and
consequently no finite element distortion arising from the microstructure. This is
critical to computational performance if iterative solvers are to be used. The aligned
meshing usually will require less finite elements than the unaligned approach for the
same pointwise accuracy. However, the disadvantages are the mesh generation for
irregular microstructures in three dimensions. Even if such microstructures can be
meshed in an aligned manner, the finite element distortion leads to stiffness matrix ill
conditioning and possible element instability (shape nonconvexity). For numerical
studies comparing the meshing approaches, see Zohdi et. al [195]. Here our emphasis
is on admitting general irregular microstructures, and rapidly evaluating them during
the testing process, thus we have taken the unaligned approach, which we discuss in
more detail next.
5.2.1
Topological resolution
Inherent in the unaligned approach is the integration of discontinuous integrands

(Figure 5.3). The topology is not embedded into the finite element a-priori, as it would
be in an aligned approach, via isoparametric maps onto material interfaces. To some
extent if the elements are much smaller than the particle length scales, the topology
will be approximately captured. However, one can improve this representation (Figure
5.2). Since the finite element method is an integral-based method, the quadrature rules
can be increased in an element by element fashion to better capture the geometry in
elements with material discontinuities. A primary question is: If there is an integrand
discontinuity in a finite element, how high should the quadrature rules be to perform
accurate integration?
5.2.1.1 Topological uncertainty for unaligned meshing Throughout the
simulations, we shall employ the finite element method, along with a technique
of Gauss point oversampling to accurately resolve the topological features of the
microstructure. We first consider a one dimensional reference finite element do with a jump discontinuity at (Figure 5.3),
main, and associated integrands,
which admit to a decomposition of the function into continuous and discontinu

ous parts,
, where is the local coordinate, and

9 > ' 9 > 9 > 9 >

9 >U 9 "> 9 > is the jump operator. We assume that the elements are small

9>

compared to the length scales of the particulate matter and as a consequence that
there is at most one discontinuity within the element. However, this assumption
makes no difference at the implementation level. Integrating over a reference element,
82
Fig. 5.2 Unaligned meshing with material discontinuities within an element.
9 >
9'9 > 9 > 9 >:>

we have
. We perform a straightforward Gauss-Legendre quadrature, with quadrature points, where points lie
before the discontinuity at . We assume that the continuous function, possibly a
polynomial,
can be integrated exactly, or nearly exactly, with standard quadrature

, where are the Gauss weights and where are

'9 > '9 >
0
. The remainder is
the Gauss point locations. We note the property

9

>

9
>
9 > 9 . > . We may write

the discontinuous part

9 > 9 > 9 >U 9 > . . As a

+
consequence, the maximum amount of variation in the computed integration is

! #"

in the interval
the dependence of the largest quadrature
weight,
3
9. In .order
. > , towithbound
" % (Gauss rules of
rule, we
least
squares
curve-fit
. )thewithGauss-Legendre
!

$

.

,
the following, PJ
with
2
'
(5.2)
83
=1
=1
M GAUSS POINTS
M+1 GAUSS POINTS
REFERENCE ELEMENT LENGTH=2
Fig. 5.3 A material discontinuity in a reference finite element.
/.

9 . $. > 9 . . > 9 . .$>
0 0 00
0 0 0
#
3 3 3
# # . 3W0
# # . 3

0 0 0
where
indicates a perfect regression value of the curve fit. Simple three
dimensional estimates can be made by applying this procedure in all three directions
on a reference element. For example, consider a 3-D step function discontinuity over
a reference finite element (
). Denoting
- the reference

/
element

volumetric normalization of the bound by
, since
=volume of the
element, one
has for (I) a
Gauss
rule
reference

% , (III) a %
%
% Gauss
Gauss
rule

, (II) a

rule
and (IV) a
Gauss rule
. Consequently,
the amount of variation in the integral is at most approximately 2 . Therefore, for
efficient implementation, a 2/5 rule should be used, whereby a
Gauss rule
if there is no material discontinuity in the element, and a
rule if there is a
material discontinuity. We emphasize that this procedure is used simply to integrate
elemental quantities with discontinuities accurately. For a variety of numerical tests
in Zohdi [207], the typical mesh density to deliver mesh
results, for the
insensitive

trilinear finite element
quantities of interest in the upcoming simulations, was
hexahedra (approximately 2200-3000 degrees of freedom (DOF)) per particle. A
disk-type

and a diamond-type microstructure, as seen by the meshing algorithm with
%
% trilinear hexahedra mesh density for a total of 46875 degrees of freedom
a %

(approximately
hexahedra or 2344 degrees of freedom per element), are
shown in Figure 5.4.

0 0 0
5.2.2
A simple Galerkin error monitor
In order to judge the quality of the numerical results, we need an indication of

the Galerkin "subspatial" approximation error. As stated earlier, under standard
assumptions the classical a-priori error estimate for the finite element method is
84
Fig. 5.4 A random microstructure consisting of 20 non-intersecting particles. LEFT a

2
*
*
*
1
4
). RIGHT an oblate disk-type microstructure
diamond-type microstructure ( (aspect ratio of 3:1). Both microstructures contain particles which occupy approximately 7
of the volume.
C
9

> 7, where is the (complete) polynomial

order of the finite element method used, is the regularity of the exact solution
"

and is a global constant dependent on the exact solution and the polynomial approximation, but independent of h, the maximum element diameter. Related forms
of this estimate hold locally, but with constants that are element dependent. The
smoothness of the solution is no better than

for elements containing material discontinuities. However, the solution can be quite smooth, i.e.

, for elements with no material discontinuities. For simplicity we employ the h-version finite element
method, i.e. successive
mesh subdi

vision using trilinear hexahedra (
), and therefore
in elements with
no discontinuities.
We recall
the well known property (PMPE), introduced earlier,

. By solving the boundary value problem as"

sociated for a given sample for
three
successively
finer
, with

meshes,

the following property:
,
we
can
set
up

the following system of equations for unknown constants , , and ,
7 . 9
7 . 9
>
9 9 > 9 >>
9 > 9 >

9 >

9 >

9 >

9 >

>
">
(5.3)
85
J(u 1)
MESH=h 1
J(u )
h
J(u 2)
MESH=h 2
MESH=h 3
J(u 3 )
1/h
h1
J(u )
J(u 2)
J(u 3)
Fig. 5.5 Successively refined (embedded) meshes used to estimate the error.
where the exponents and are selected to reflect the estimated regularity of the
solution in the elements. In the
case of using trilinear hexahedra we are led to the

logical choice of exponents:
, representing elements with no material

discontinuities,
and
,
representing
elements with material discontinuities.

It is clear that
, and we can thus define the following quantity to monitor
the Galerkin error
9
">

9 9 > >

(5.4)
In summary, to monitor the discretization error, we apply the following algorithm:
86
! .

%"$

%"$

!#"$%

"$

*
*<
1
*<
4

-

1

&

*
*
1
&
% (

3

"% *

1
% (
&
% (
"
"

&

(5.5)
These operations are performed for each sample, since they each have a different
random microstructure.
5.3 ELEMENTAL AVERAGING AND INCREASED ACCURACY
'
7 .

, and thus the error in the energy

For elements with material discontinuities,
norm is bounded from above by . However, if we ask only to compare the average
of the solutions over each element we have

%

%
>

>
>

%
>

&

>

>

%

&

4
E
(5.6)
where
. In short, the rates of convergence between the averaged quantities
are much faster than between the unaveraged quantities. Consider a one-dimensional
example. Suppose the true
solution contained
in an element domain
containing ma

and

,
if
terial discontinuities is

if
, as shown in Figure 5.6. Now consider a field produced by a

linear finite element
approximation The field will be a purely linear function, which
we denote
. The difference between the true and the approximate solution
is
A #

A
# 9 F# >
87
ITERATIVE KRYLOV SOLVERS/MICROSTRUCTURAL CORRECTORS

.

"

Q

"

G

A

9.

"
Q

"

9 #

G >

For example, take the special case when

> '

(5.7)

0 #

(5.8)
Therefore, we see that the rate of convergence is at least linear. However, this can
be increased further, for example via more sophisticated post processing extraction
techniques (see Szabo and Babuska [172]). Essentially the error in the element
averaged stress and strain fields scale at least linearly with element size.
5.4 ITERATIVE KRYLOV SOLVERS/MICROSTRUCTURAL

CORRECTORS
For micro-macro mechanics problems, one important feature of iterative solvers, such
as the CG method, is the fact that since they are based on successive updates of a
starting guess solution vector. Consequently, they can be given a head start by a good
solution guess, for example, provided by an inexpensive analytical approximation
solution. For example, in a micro-macro mechanical context, the Conjugate Gradient
(CG) method can be interpreted as making fine scale (high frequency) corrections
with each iteration. Importantly, one should use inexpensive approximate solutions
as initial guesses for the iterative solver. This further enhances the effectiveness of
the CG searches for this class of problems. A well known fact, with regard to iterative
solvers is the fact that they are quite adept at capturing local high-frequency responses.
In other words, localized effects are resolved quickly during CG iterations. However,
Krylov methods typically may require many iterations to capture long-wave modes.
For a discussion see Briggs [14]. In other words, a starting vector that captures apriori the long wave components of the solution is advantageous. For example, if
we assume that either of the
the two classical
boundary conditions
that satisfy
Hills

,
and (2)

conditions hold: (1)
.

The classical assumptions of Voigt [183] (uniform strain) and Reuss [151] (uniform
stress) produce the following:
1

, which implies
(1)VOIGT:

88
h
u =Dx
u=Bx+(A-B)x*
u=Ax
x
x*
MATERIAL 1
MATERIAL 2
Fig. 5.6 An intuitive characterization of heterogeneous element averaged quantities.

(2)REUSS:

, which implies
(1)VOIGT:
(2)REUSS:

The Reuss solution, while statically admissible, is highly oscillatory in the displacement field variable and kinematically inadmissible. The Voigt assumption, which is
statically inadmissible and kinematically admissible, is not oscillatory in the displacement field variable. Therefore, from an iterative solver point of view, the Reuss field
assumption is not of much value as a starting vector for a displacement based method.
Furthermore, such guess cannot be projected onto the initial nodal values because it is
double valued at material discontinuities (not kinematically admissible). However, it
is important to note that if a complementary based numerical method were to be used,
where the primary variable is the stress, the Reuss guess would be advantageous due
to the smooth nature of stress field assumed. The Voigt (constant strain) assumption
produces a low frequency initial guess. Therefore, from an iterative solver point of
OVERALL TESTING PROCESS: NUMERICAL EXAMPLES
89
view, it is of value when using displacement based weak form such as the standard
FEM. Furthermore, such a guess can easily be projected onto initial nodal values, because it is single valued. The central point is that each Conjugate Gradient iteration
can be viewed as a microstructural corrector
to an initially statically inadmissible

should be used for the displacement
initial guess (Figure 5.7), and that
tests.
u=
.x
CORRECTIONS TO u=
.x
Fig. 5.7 Krylov corrections to a smooth Voigt starting vector.
5.5 OVERALL TESTING PROCESS: NUMERICAL EXAMPLES

As considered before, a typical example of a composite material combination is that
of an aluminum matrix (77.9, 24.9 GPa) embedded with (stiffening) boron particles
(230, 172 GPa). We chose AL/B as a material combination which exhibits significant
enough mismatch in the mechanical properties to be representative of a wide range
of cases. All tests were run on a single IBM RISC 6000 workstation. Comparable
hardware is available in most academic and industrial work places, therefore such
simulations are easily reproducible elsewhere for other parameter selections.
5.5.1
Successive sample enlargement
In a first set of tests, the number of particles contained in a sample were increased
holding the volume fraction constant. During the tests, we repeatedly refined the
mesh to obtain invariant macroscopic responses. A sample/particle size ratio was
used as a microstructural control parameter. This was done by defining a subvolume

, where
is the number of particles in the entire sample and
size

90
X2
X1
X
Fig. 5.8 A microheterogeneous "tailored" solid.

) sample. A generalized diameter was
where is the length of the (cubical
defined, , which was the diameter of the smallest sphere that can enclose a single
particle, of possibly non-spherical shape if desired. The ratio between the generalized
. For a variety of numerical tests,
diameter and the subvolume was defined by
discussed momentarily, the typical
mesh density to deliver invariant volumetrically

averaged responses was
trilinear finite element
(approximately

hexahedra

2200-3000 degrees of freedom) per particle. We used
, which resulted in
a (fixed) volume fraction of approximately 22 . The following particle per sample
sequence (Figure 5.10) was used to study the dependence of the effective responses on
the sample size: 2 (5184 DOF), 4 (10125 DOF), 8 (20577 DOF), 16 (41720 DOF), 32
(81000 DOF) and 64 (151959 DOF) particles. In order to get more reliable response
data for each particle number set, the tests were performed five times (each time
with a different particulate distribution) and the responses averaged. Throughout the
tests,
we considered
a single combined
boundary loading satisfying Hills condition,

,
. We tracked the strain energy, as well as

and , as defined in Equation (3.1). Table 5.1 and Figure 8.5 depict the dependency
of the responses with growth in particle number. Justified by the somewhat ad-hoc
fact that for three successive enlargements of the number of particles, i.e. 16, 32 and

4!
0:3
.-+ .

91
Fig. 5.9 6061 aluminum alloy reinforced with pitch-55 boron additives ( 250), reprinted
with permission of The Metalurgical Society, Warrendale, PA, USA.
16
32
64
SAME VOLUME FRACTION

INCREASING NUMBER
OF PARTICLES
Fig. 5.10 A series of test samples with increasingly more particles, but with the volume
fraction fixed.
0 0 0
64 particle samples, the responses differed from one another, on average, by less than
for further tests. We remark that we
1 , we selected the 20-particle
microstructures

applied a 2/5 rule,
i.e.
a
Gauss
rule
if there is no discontinuity in the

element, and a
rule if there is a discontinuity, which is consistent with the
earlier derivation in the work (Table ??). The microstructure, as seen by this mesh
density, is shown in Figure 5.12.

5.5.2
Multiple sample tests
0 0 0
For further tests, we simulated 100 different samples, each time with a different

random distribution of 20 nonintersecting particles occupying 22
( ).
%
%
%
Consistent with
the previous tests mesh densities per particle, we used a
mesh (
trilinear hexahedra or 2344 DOF per particle, 46875 DOF per test
sample). The numerical error monitor, , which measures microscopic differences
92
PART
DOF
W (GPa)
2
4
8
16
32
64
0.595
0.472
0.375
0.298
0.236
0.188
5184
10125
20577
41720
81000
151959
0.001444
0.001408
0.001386
0.001375
0.001365
0.001358

(GPa)
(GPa)
98.2
97.3
96.5
96.2
95.9
95.7
46.7
44.3
43.2
42.5
41.6
41.4
Table 5.1 Results of successive sample enlargements. Five tests, each with a different random distribution, were performed at each sample/particulate size ratio level to obtain somewhat
representative data.
5184 DOF
d/L=0.595
(GPa)
48
10125 DOF
d/L= 0.472
(2)
46
20577 DOF
d/L= 0.375
41270 DOF
d/L= 0.298
< >:< > 1/2

( < >:< > )
2 =
(4)
44
(8)
81000 DOF
d/L= 0.236
(16)
(32)
42
(64)
151959 DOF
d/L= 0.188
40
50 60
10
20 30 40
NUMBER OF PARTICLES
Fig. 5.11 The values of the effective shear responses for samples containing increasingly
larger numbers of particles. One hundred tests were performed per particle/sample combination
and the results averaged.
0 0 0 0 0 0
. . .
0 0 0
0 0 0

%
%
(described in Box 5.5), was constructed for the each sample with a %
mesh

,
density
by
using
a
mesh
sequence
of
embedded
finite
element
meshes:

%
% (Figure 5.5). The %
%
% mesh was used in all
and %
final computations for each sample, with an accompanying computed each time.
The plots of the behavior of the various quantities of interest are shown in Figures 5.13-5.22. The averages, standard deviations and maximum/minimum of these
quantities are tabulated in Table 5.2. For the 100 sample tests, with 20 particles per
sample, the results for the effective responses were
Z.3

%
0 .3

.G.Q.

(5.9)
QUAN.
'

(GPa)
(GPa)
(GPa)
0.001373
96.171
42.350
0.0376
49.079

.
3

!

0.2025
0.4798
0.0045
1.5790
93
0.950
2.250
0.024
9
Table 5.2 Results of 100 material tests for randomly distributed particulate microstructures
(20 spheres). We note that for a zero starting guess for the iterative solver, the average number
8
*
of CG-iterations was 55, therefore using the initial Voigt (
) guess saved approximately
2 3
of the computational solution effort.

where and are the averaged effective responses from the 100 tests, and where the
lower and upper bounds are, respectively, the classical Reuss [151] and Voigt [183]
bounds. We also compared the computed results to the well-known Hashin-Shtrikman
bounds ([65], [66]) which are, strictly speaking, only applicable to asymptotic cases of
an infinite (sample length)/(particulate length) ratio and purely isotropic macroscopic
responses. The bounds were as follows:
3W0
33
%

%

0 .3

03

%
%

(

.

.

9
.

>

(

(

(

(5.10)
94

and
where
are the bulk and shear moduli for the matrix and particle
phases. Despite the fact that the bounds are technically inapplicable for finite sized
samples, the computed results did fall within them. The time to preprocess, solve
and postprocess each 20 particle finite element test took no more than one minute
on a single RISC 6000 workstation. Therefore, as before, 100 of such tests lasted
approximately 1.5 hours.
Remarks: A primary question is: "What other information can we extract from
the average of the responses of many samples? This is discussed next.
5.6 A MINIMUM PRINCIPLE INTERPRETATION

Consider the following process for a sample of material with

1. Step 1: Take the sample, and cut it into N pieces,
. The pieces
do not have to be the same size or shape, although for illustration purposes it
is convenient to take a uniform (regular) partitioning (Figure 5.23)
2. Step
2: Test each piece (solve the subdomain BVP) with the loading:

. The function
is the solution to the BVP posed over subsample

3. Step 3: One is guaranteed the following properties:
5
"
* *

$ -
* " *
*

%
&

% (

%

%

E

$ -
*
"%
*

%

- -

-
%

%

E
$ -

%
$ -

E

%
E
(5.11)
The same process can be done for traction test loading cases:
. The

effective material ordering, line three in Box 5.11, has been derived by Huet [83]. The
second line of Box 5.11, and generalizations to nonuniform loading, were developed
in Zohdi and Wriggers [203]. The proofs, which are constructive, and illustrative of
the main ideas, are provided next. They result from a direct manipulation of classical
energy minimization principles.

5.6.1

A proof based on partitioning results
Consider a large sample whose domain consists of the union of many smaller domains,

, as depicted in Figure 5.23. The corresponding solution , denoted
95
A MINIMUM PRINCIPLE INTERPRETATION
the globally exact solution, is characterized by the following weak boundary value
problem
*

+"

- *
<*
(
#
&
#

*

-

*
(5.12)

Specific choices for , and will be given
Now consider the boundary

, as , momentarily.
of an individual subdomain
( is the total number of
subdomains)
S

@

9 >

9 >

9
N
>

L M1H V:HZJV

(5.13)
!
A specific choice for will also be given momentarily. The individual subdomain
solutions form an approximate solution to the globally exact problem, . This approximate solution is constructed by a direct assembly process

> 9 > 9 >

(5.14)
The approximate displacement field is continuous, however, the approximate traction

field is usually discontinuous. Clearly, if there are no jumps in the tractions, the
solution is exact.
5.6.2
Relation to the material tests

>
We have for any kinematically

admissible function

5

notation,
"
"
0 , using the0 (bilinear operator)

9 > 9 > 9 >
9 > , where 9 > 9 > 9 > . In other words, the true solution has
a minimum
By applying the PMPE, for the test loading
0
potential (PMPE).
, we
, , with the9 specific
> 9 > choice
9 > (kinematically
9 >:> 9 admissible)

.
obtain

The critical observation is that
when choosing 0
in Box 5.13, then ,
0

as defined in Box 5.14,
is also9 kinematically
admissible.
Therefore,
we have by direct

9 > 9 >:> 9 9 > 9 > >
.
expansion

"
1
Since

"

"

is a solution to a subdomain boundary value problem posed over
"
, it
96
9 ">
9

>
9
>
9
>

0

9 9 > 9 > >

(5.15)

90 >

By direct expansion we have
> . We have
09
0 by definition, 9 >
,

9

>
9 >
0

, and
, which implies
9 >
. By direct substitution this completes the first of the
9 >

minimizes the corresponding subdomain potential energy function

. Therefore,

. Consequently,

"
"

5

"

5
assertions in Box 5.11. We can directly interpret the testing of many smaller samples
as simply the partitioning of a very large one which we cannot easily test (Figure
5.23). Therefore, for displacement tests, the averaged effective responses generated
will always bound the response of the very large sample from above. Therefore, the
average of the 100 sample tests provide us with tighter upper bounds on the response
of a very large sample. Therefore, by the previously proven results in Box 5.11, the
Reuss-Voigt bounds in Box 5.9 are tightened by the following factors
3
. . 3
.Q.G. .

03 3
JS

%
(5.16)
5.7 PARTITIONING AND TRACTION TEST CASES

Suppose we repeat the partitioning process for an applied internal traction set of tests.
As will be shown traction tests will bound the response of the very large sample from
below. Employing the Principle
of Complementary
Potential Energy, we obtain for

the test loading
, with
(a statically admissible trial field),
"

"

9 9
> 9 %> > 9 9 %> 9 > >

9
>

9
0
9 "> 9 %>

"
"

(5.17)

where
, being the stress

field produced by solving each subsample problem with the uniform stress boundary

conditions and forming

. By direct expansion we have

"
"

.

The complementary forms collapse to

,

and

7

is the subdomain complementary potential energy function for

, where
subdomain . By direct substitution this yields
0 >

9 %>
9
>
0 >
9 >
PARTITIONING AND TRACTION TEST CASES
2

"

9

>

>

97
(5.18)
As for the primal (displacement) case, the second line of Box 5.18, and generalizations
to nonuniform loading, were developed in Zohdi and Wriggers [196]. The fourth line
has been derived in Huet [83] by other means.
5.7.0.1 Consequences/difficulties
with traction tests If the sample were

an RVE, we have
, then the preceding analysis yields the following
two sided ordering of approximate effective material responses,

(5.19)
>
We emphasize that
is an assumption, which may not be true for a
finite sized sample. Therefore, in theory, under the RVE assumption, traction tests
form lower bounds on the effective responses. However, traction tests pose great
difficulties, which are as follows:
1. Numerically pure traction boundary data cause rigid motions (singular FEM
stiffness matrices),
2. The
FEM is a method based upon generating kinematically admissible solutions

. The traction tests result is based upon the assumption that statically admissible trial field are generated. Statically admissible fields cannot be achieved
by a standard FEM approach,
3. True laboratory tests specifying the force on a sample are far more difficult
than specifying the displacements.
5.7.0.2 Isolating the subsampling error/numerical error orthogonality
According to the results in Box 5.11, we have the following normalized estimate:

* *
&
% (

/

$ %$

%

(5.20)
where, since we have used displacement controlled tests, we have used the Voigt
materials energy, which corresponds to assuming a constant strain throughout the
material equal to , to normalize the results.
98
It is important to note that the numerical error, which is implicitly included in

the estimate in Equation 5.20, is orthogonal to the partitioning or subsampling
error. In other words, while the numerical error is known, it can be directly filtered out of the sampling error estimates. To see this, consider that the boundary
value formulations
in Boxes
5.12 and 5.13 directly
imply,
for
any type
of loading

!

,

,
.

This directly implies that
. Therefore,
@9
@

>
* *
1

9
N
>

@9 > @9
>
* * * *
- * * - * *
- * * - * *

! $ !
- * * - * *
* *
* *

% (
&
023

% (
&
7

% ( 0
&
% (
&
(5.21)
Under the special case that

, we have the following orthogonal decompo
sition of sampling and numerical error:
/
* *

*
C
/

* *

C

0

-
-
&

-
C

1
-
&
% (
(5.22)
%

(

and

. Therefore, the
where

expression in Box 5.22 indicates that the estimates made on the effective response
is an orthogonal sum of both the numerical and sampling error. Therefore, one can
isolate either the numerical or sampling error, if the other is known, or estimated, by
simply subtracting it from the the total expression (the lefthand side of Box 5.22).
Therefore, for example, approximating the numerical error by the error monitor, ,
we have

#

.

>

(5.23)

This quantity represents the isolated error in the internal fields induced by the partitioning or subsampling only.
DEPENDENCY ON VOLUME FRACTION
99
5.8 DEPENDENCY ON VOLUME FRACTION
We repeated the 100 sample tests procedure for samples containing significantly
higher volume fraction, approximately 32 . Unlike fiber-reinforced composites,
which can contain well over 50 volume fraction of fibers, particulate composites
usually contain no more than approximately 25 volume fraction of particulates.
Particulate volume fractions over 15 are already high in three dimensions. For the
relevant examples, 22 particulate volume fraction in three dimensions corresponds
roughly to 44
in two dimensions, while 32
in three dimension corresponds
roughly to 57
in two dimensions. Cross-sections of these volume fractions are
depicted in Figure 5.24. The results for the 32 case are tabulated in Table 5.3. The
averages for the effective properties of the samples were
33

. 0 3

0
. 0 3

(5.24)
where and are the averaged quantities from the 100 tests, and where the lower
and upper bounds are, respectively, the classical Reuss [151] and Voigt [183] bounds.
The Reuss-Voigt bounds in Box 5.24 are tightened by the following factors
3
.0 .
.

0 3 3 3
3 3
0 3

JS

%
.
(5.25)
The error estimate in this case is

#

.

03
>

.
Z3
(5.26)
which is approximately 2 higher error estimate than the case of 22 particulate

volume fraction. Intuitively one would expect such a result, since the particles interact
more at such a high volume fraction (Figure 5.24). In other words, they "feel" the
presence of the other particles more at higher volume fractions than at lower volume
fractions. As before, we also compared the computed results to the Hashin-Shtrikman
bounds [65] and [66] which are, strictly speaking, only applicable to asymptotic
cases of an infinite (sample length)/(particulate length) ratio, and purely isotropic
macroscopic responses
33
.

3
. 0 3 .

.G. 3 .
%
%
(5.27)
100
QUAN.

'
(GPa)
(GPa)
(GPa)

0.001593
106.831
52.635
0.0523
48.950

.G. .

.

!

0.2505
0.5876
0.0044
1.716
1.350
3.095
0.0255
8
Table 5.3 Results of 100 material tests for randomly distributed particulate microstructures
(20 spheres, =0.85, or approximately 32 volume fraction).
5.9 INCREASING THE NUMBER OF SAMPLES

Suppose we increase the number of samples tested. For purposes illustration purposes,
let us test 512 samples of material. The number 512 is not accidental, since it is a
common number of independent processors in modern parallel processing machines.
Table 5.4 illustrates that the averaged results are virtually identical to the 100 sample
tests for all the quantities. Testing more and more samples will not help obtain
better average results. However, despite practically the same average values, one can
observe from the Figures 5.25-5.32 that the 512 sample tests have a more Gaussian
distribution, relative to the 100 sample tests, for the responses. However, for even
more accurate average responses, we must test larger samples of material. This is
explored further in the next section.
QUAN.
'

(GPa)
(GPa)
(GPa)

0.001373
96.169
42.353
48.8984
0
0.1967
0.4647
1.7737

1.203
3.207
11
Table 5.4 Results of 512 material subdomains, each containing 20 randomly distributed
spheres for a total of 10240.
5.10
INCREASING SAMPLE SIZE
Beyond a certain threshold, it is simply impossible to obtain any more information

by testing samples of a certain size. As we have shown the reason for the inability for
obtaining further information is the fact that the testing conditions are uniform on the
"subsamples". This idealization is valid only for an infinitely large sample. However,
101
INCREASING SAMPLE SIZE
suppose we increase the number of particles per subsample even further, from 20 to
say 40 then 60, each time performing the 100 tests procedure. With this information
one could then possibly extrapolate to a (giant) sample limit. The results for the 40
and 60 particle cases are shown in Table 5.5 for 22 boron volume fraction. Using
these results, along with the 20 particle per sample tests, we have the following curve
fits

.

3 3

.
0
3 3
3Q0

where is the sample size, . is the diameter of the particles. Thus as

obtain estimates of
GPa,

GPa and

3Q0
0.

(5.28)
, we
GPa
as the asymptotic energy, effective bulk modulus, and effective shear modulus,
re
spectively. Indeed, judging from the degree of accuracy of the curve-fit, (

is perfect) the curve fit appears to be perfect for
. For
and , the slightly

less accurate reliability (regression values of
) is attributed to the fact that

absolute perfect isotropy is impossible to achieve with finite sized samples. In other
words the extrapolations using various samples exhibit slight isotropic inconsisten
cies. However, the energy , has no built-in assumptions whatsoever, thus leading to
the nearly perfect curve-fit. The monotonicity of the testing curves is to be expected,
and is explained further in the next chapter.
PART
DOF
40
40
40
40
0.2193
0.2193
0.2193
0.2193
98304
98304
98304
98304
60
60
60
60
0.1916
0.1916
0.1916
0.1916
139968
139968
139968
139968
QUAN.
'

(GPa)
(GPa)
(GPa)

'

(GPa)
(GPa)
(GPa)

.

0.0013617
95.7900
41.6407
60.3200
0.0013586
95.6820
41.4621
66.500
0.1413
0.3245
1.6302

3 3

0.1197
0.2801
1.9974

!
.

%
0 3
.

00
0.6600
1.590
7

0.6214
1.503
10
Table 5.5 Results of material tests for randomly distributed particulate microstructures for

*
100 (
, approximately 22 ) samples of 40 and 60 particles per sample.
102
5.11
AN ERGODIC INTERPRETATION OF THE RESULTS
It is possible to interpret the results in ergodic manner, which we now briefly discuss. In statistical mechanics a frequently invoked assumption is that of statistical
ergodicity. Importantly, in the context of material testing, Hills condition can be
interpreted as such an assumption. To see this, consider several bodies with the same
external geometry composed of the same material volume fractions but with different random microstructure. Let us focus on the same spot on each body (indexed
$

by
) and compute
, where N

is the number of macroscopic structures. The tensorial quantity,
, is called an
ensemble average (Figure 5.33). Now we perform another experiment. We have a
single body, and focus
on one spot,
but repeatedly
and enlarge the sample
compute

),
. This is our familiar
size (indexed by
(volumetric averaged)
effective
property.
A
classical
ergodicity
assumption is that as

that
. Therefore, volumetric averaging and ensemble averaging
yield the same result. In other words, an infinitely large samples volumetric average
must equal the ensemble average of infinitely many finite samples at a point, from
different bodies.

If we split the stress
and
strain fields into a purely fluctuating (zero mean) ( )

and average parts (
) we see by direct expansion:
- .
-+
0

+ 6.
02

9 >

! !

%> 9

>

"

"

. In other words, the product of two
The ergodicity assumption is that
purely fluctuating random fields is again purely fluctuating (Figure
5.34). The impli
cation is that as the sample becomes infinitely large that
is purely fluctuating.
This is exactly the implication of the Hill condition. This motivates the use of exterior
uniform loading for an infinitely large sample. Clearly, the partitioning, and the error
estimates, can be interpreted testing an ergodic assumption. However, the loadings
are still not rough enough, since it is clear that uniform loading is an idealization
and thus will only be present within the microstructure of a finite sized engineering
(macro)structure to an error. For more on ergodic hypotheses see the classical work
of Kroner [100].
103
Fig. 5.12 TOP: a random microstructure consisting of 20 non-intersecting boron spheres,

occupying approximately 22 of the volume in an aluminum matrix, as seen by the algorithm
3
3
3
with a
trilinear hexahedra mesh density for a total of 46875 degrees of freedom

(approximately
hexahedra or 2344 degrees of freedom per element). A 2/5 rule,

3
3
3
Gauss rule if there is no discontinuity in the element, and a
rule if
i.e. a
there is a discontinuity, was used. BOTTOM: a zoom on one particle.

104

0.001395
ENERGY
0.00139
SYSTEM ENERGY (GPa)
0.001385
0.00138
0.001375
0.00137
0.001365
0.00136
0.001355
0
10
20
30
40
50
SAMPLE
60
70
-*
80
90

- *
100
Fig. 5.13 100 SAMPLES: The energy responses,

(GPa), of
% %
a block with 20 randomly distributed Boron spheres embedded in an Aluminum matrix. Each
point represents the results of one test.
25
SAMPLES
20
15
10
0.00136
0.00137
0.00138
0.00139
ENERGY (GPa)
-*

-*
Fig. 5.14 100 SAMPLES: The corresponding histogram for the variations in energy,
(GPa).
% %
105
96.8
KSTAR
96.7
EFFECTIVE BULK MODULI (GPa)
96.6
96.5
96.4
96.3
96.2
96.1
96
95.9
95.8
95.7
0
10
20
30
40
50
SAMPLE
60
70
80
90
100
Fig. 5.15 100 SAMPLES: The bulk responses, , of a block with 20 randomly distributed
Boron spheres embedded in an Aluminum matrix. Each point represents the results of one test.
SAMPLES
20
15
10
95.75
96
96.25
96.5
EFFECTIVE BULK MODULUS (GPa)
Fig. 5.16 100 SAMPLES: The corresponding histogram for the variations in the effective

bulk responses, .
106
43.5
MUSTAR
EFFECTIVE SHEAR MODULI (GPa)
43
42.5
42
41.5
41
0
10
20
30
40
50
SAMPLE
60
70
80
90
100

Fig. 5.17 100 SAMPLES: The shear responses, , of a block with 20 randomly distributed
Boron spheres embedded in an Aluminum matrix. Each point represents the results of one test.
20
SAMPLES
15
10
0
41
41.5
42
42.5
43
EFFECTIVE SHEAR MODULUS (GPa)

shear responses, .
107
0.05
NUMERICAL ERROR
ERROR IN ENERGY (GPa)
0.045
0.04
0.035
0.03
0.025
0
10
20
30
40
50
SAMPLE
60
70
80
90
100
Fig. 5.19 100 SAMPLES: The numerical error monitor of a block with 20 randomly
distributed Boron spheres embedded in an Aluminum matrix. Each point represents the results
of one test.
SAMPLES
20
15
10
0.025
0.03
0.035
0.04
0.045
NORMALIZED NUMERICAL ERROR
Fig. 5.20 100 SAMPLES: The corresponding histogram for the variations in the numerical

error monitor .
108

54
CG SOLVES
53
CONJUGATE GRADIENT ITERATIONS
52
51
50
49
48
47
46
45
0
10
20
30
40
50
SAMPLE
60
70
80
90
100
Fig. 5.21 100 SAMPLES: The Conjugate Gradient iterations of a block with 20 randomly
distributed Boron spheres embedded in an Aluminum matrix. Each point represents the results
of one test.
25
SAMPLES
20
15
10
46
48
50
52
Fig. 5.22 100 SAMPLES: The corresponding histogram for the variations in Conjugate
Gradient iterations needed for a solution.
109
X2
LARGE PROBLEM
STATISTICALLY
REPRESENTATIVE
SAMPLE
X1
k
PARTITIONED PROBLEM
"SUBSAMPLES"
Fig. 5.23 The idea of partitioning a sample into smaller samples or equivalently combining
smaller samples into a larger sample.
22 % VOLUME FRACTION SIZE

32 % VOLUME FRACTION SIZE

Fig. 5.24 A comparison of relative sizes of particles occupying approximately 22

*
) and 32 (
) volume fraction.
110
0.0014
ENERGY
0.001395
0.00139
SYSTEM ENERGY (GPa)
0.001385
0.00138
0.001375
0.00137
0.001365
0.00136
0.001355
0.00135
0
100
200
300
SAMPLE
"%
400
-*
500
600

- *
Fig. 5.25 512 SAMPLES: The energy responses,

(GPa),
% %
of a block with 20 randomly distributed boron spheres embedded in an aluminum matrix.
140
130
120
110
SAMPLES
100
90
80
70
60
50
40
30
20
10
0.00135
0.00136
0.00137
0.00138
0.00139
ENERGY (GPa)
-*

-*
Fig. 5.26 512 SAMPLES: The corresponding histogram for the variations in energy,

(GPa).
% %
"
*
111
96.8
KSTAR
EFFECTIVE BULK MODULI (GPa)
96.6
96.4
96.2
96
95.8
95.6
0
100
200
300
SAMPLE
Fig. 5.27 512 SAMPLES: The bulk responses,

boron spheres embedded in an aluminum matrix.
400
500
600
, of a block with 20 randomly distributed
120
110
100
SAMPLES
90
80
70
60
50
40
30
20
10
96
96.5
EFFECTIVE BULK MODULUS (GPa)

bulk responses, .
112
44.5
MUSTAR
44
43.5
43
42.5
42
41.5
41
40.5
0
100
200
300
SAMPLE
Fig. 5.29 512 SAMPLES: The shear responses,

boron spheres embedded in an aluminum matrix.
400
500
600
, of a block with 20 randomly distributed
160
140
120
SAMPLES
100
80
60
40
20
41
42
43
EFFECTIVE SHEAR MODULUS (GPa)

shear responses, .
113
54
CG SOLVES
52
50
48
46
44
42
0
100
200
300
SAMPLE
400
500
600
Fig. 5.31 512 SAMPLES: The Conjugate Gradient iterations of a block with 20 randomly
distributed boron spheres embedded in an aluminum matrix.
110
100
90
SAMPLES
80
70
60
50
40
30
20
10
44
46
48
50
52
Fig. 5.32 512 SAMPLES: The corresponding histogram for the variations in Conjugate
Gradient iterations needed for a solution.
114
ENSEMBLE AVERAGING
MORE
BODIES
VOLUMETRIC AVERAGING
LARGER
SAMPLES
Fig. 5.33 Ensemble and volumetric averaging processes.
SAMPLE LENGTH
SAMPLE LENGTH
SAMPLE LENGTH
Fig. 5.34 An illustration of an ergodicity assumption. Essentially it states that the product
of two random purely fluctuating fields is also purely fluctuating and random.
6
Higher order statistics
6.1 STATISTICAL SHIFTING THEOREMS
, i=1, 2, ...N=samples)
Consider any tested quantity, , with a distribution of values (

about an arbitrary reference point, denoted , as follows:

"

$
-
"
*
(6.1)
. The various moments characterize the dis " measures

the first deviation from the average,
"

9 > , (3)

which equals zero, (2)

is the standard deviation and (4)
is the skewness. The skewness measures
the bias, or asymmetry of the distribution of data. The moments of the data can be

9>

where
and
tribution, for example: (1)

expressed about the average and related to any other reference point, denoted here as
, using parallel axis type theorems. Using the notation in Box
6.1, let us define

. There are
useful properties associated with these relations, in particular a shifting property for
the standard deviation

115
116
HIGHER ORDER STATISTICS
"%

1
3
1

0

1
*
3

0
1
E
(6.2)
E

*

31

as well as for the skewness
"

4
*

1

$
3
0

*

4
*
1
1
0

4
E
4

E

3
0
(6.3)
A4
From Box 6.2 we may write

-

3

*
3
0

A4

3
1
3
3

$
31

therefore,

$

-

1

0
3
1

*
4

$

(6.5)
We now proceed to bound the average, standard deviation and skewness associated
with the ensemble averaging of a population of samples.
6.2 DOMAIN DECOMPOSITION AND ENSEMBLE AVERAGING

6.2.1
(6.4)
E
Primal partitioning
Consider a (large) sample of material with the following general boundary value
representation
117
DOMAIN DECOMPOSITION AND ENSEMBLE AVERAGING
Fig. 6.1 A regular partitioning of a large sample of material.
+"
*

-

-
*

*
(
#
&
#

8

*

-
(6.6)

Now partition the domain into S subdomains,
. The pieces do not have
to be the same size or shape, although for illustration purposes it is convenient to take
a uniform (regular)
partitioning
(Figure
6.1). Consider a kinematically admissible

function,
and
,
which
is projected onto the internal boundaries

(
) of the subdomains. Any subdomain boundaries coinciding with the exterior
surface retain their original boundary conditions (Figure 6.1). Accordingly,
we have

the following virtual work formulation, for each subdomain,
9 N>
+"

%

-

-

%
&

-
*
- *

5

E
<*

%
&
%
%
)(
*

*
(

8

0

(
#
/& #

-

(6.7)

The individual
subdomain solutions, , are zero outside of the corresponding sub
domain
. In this case the approximate solution is constructed by a direct assembly

process,
. The approximate

&
displacement field is in
, however,
the
approximate
traction
field is possibly

, will be given momentarily. It
discontinuous. Logical choices of , i.e.
9 > 9 N > 9 > 9
>
118
should be clear that if

on the internal partition boundaries, then the approximate solution is exact. Since we employ energy type principles to generate approximate solutions, we use an induced energy norm to measure the solution differences

"
1"
. It is convenient to cast

@9 >
9 > @

the error in terms of
the potential energy,
@

0
9 , where
is any kinematically admissible function. This leads to

9 > 9 > > or 9 > 9 > , which0 is a form of the Principle
solution possesses a minimum
of Minimum Potential Energy. In other words, the true

9 9 > 9 > > . In the
potential. By direct substitution we have
, which is equivalent
with
special case that
to testing
9 each subsample
> , where

and
, then

"
6.2.2

>
"

@9
2

and
1"

Complementary partitioning
We can repeat the partitioning process for an applied internal traction set of tests. The
equivalent complementary form for the exact (undecomposed) problem is
+-"
8K5

*

)5
$ -

8K:
E
*
8;:

(
#
/& #
*,%

8

*

8;:
(6.8)

For the complementary problem, similar restrictions are placed on the solution and
test fields to force the integrals to make sense. In other words, we assume that solutions produce finite global energy. When employing the applied internal traction
approach, in order to construct approximate
solutions, a statically admissible func
, is projected onto the internal boundaries
tion, , with the property that

of the subdomain partitions. As in the applied displacement case, any subdomain
boundaries coinciding with the exterior surface retain their original boundary conditions. Accordingly, we have the
complementary virtual work formulation,
following

for each subdomain,

+-"
8 5

$ -

0

5

8;:

%
8;:

%
8;:

&
%

&

(

*
8K:

%
&
%

(
(
#
&
#

8

*
%
E
<*
(6.9)
The individual
subdomain solutions, , are zero outside of the corresponding

subdomain
. In this case the approximate solution is constructed by a direct
DOMAIN DECOMPOSITION AND ENSEMBLE AVERAGING
> 9
> 9
5 5
>
119

assembly process

The stress field is statically admissible, however, the approximate
displacement
field

, will be given
is possibly discontinuous. Logical choices of ,
momentarily. It should be clear that if
on the internal partition boundaries,

then the approximate solution is exact. We define the complementary norm

. As in the primal case, it is

"

9 %>
9 %>
9 >
convenient to cast the error in terms of the potential complementary energy for the

,

case of linear elasticity, where
where is any statically admissible function. The well known
relationship,
for a

statically admissible function , is "

or
. This is a form of the Principle of Minimum Complementary Potential Energy.
In other words,
the true solution
possesses a minimum
complementary
potential.

Choosing
,
we
have
.
In
the
special
case

that
, which is equivalent
to
testing
each
subsample
with
,

"

,
where

9 >
and

6.2.3

9 9 %> 9 > >
9 9 %> 9 > >
>
9
>
Homogenized material orderings
If the sample is statistically representative, we have

, then the previous results imply, under the assumption that the uniform loadings are arbitrary, the
following two sided ordering of approximate effective material responses,

(6.10)
9 >
9 >
where the
tensor inequality notation means, for example, that the difference ten

sor (
) is positive definite, etc. Since
, we also have

9 >

9 X>

. Alternatively, since
, then
. To

the knowledge of the author, the result in Box 6.10 was first derived in Huet [83],
however by other analysis techniques.
Remark: For isotropic responses, we have

6.2.4
Embedded orthogonal monotonicities
(6.11)

9 9 > 9 >>

Since is kinematically admissible, we have "
. If
1
we
repartition
the
existing
subdomains
into
more
subdomains
(Figure
6.2),
and
use

for the local boundary conditions on the finer partition, upon solving the local
120
9 9 > 9 >>
boundary value problems and assembling the local solutions together

(just as before

1"
for ), we have, denoting the solution by , "
.
Adding the two previous relations together yields an orthogonal decomposition
"

"

1

"

(6.12)
This implies that the error monotonically grows for successively finer embedded partitions. Intuitively one expects this type of growth in the error, since one is projecting
more inaccurate data onto the interfaces. Simply stated, more embedded subdomains,
more error. Furthermore, the relationship is monotone. As in the displacement
controlled tests, for traction controlled tests we have

"
> 9
> 9
"
>

"
In terms
of effective properties, Equation 6.12 implies

>

>

>
(6.13)

, while Equation 6.13 implies
"

. To streamline
to denote embedded partitions, etc. The

the notation we employ the notation
higher the number the more embedded partitions exist (Figure 6.2). Therefore, for

, we have
and
.

$ -
$ -

(6.14)
The results of this section generalize and extend relations found in Huet [83],Hazanov
and Huet [68], Huet [91], Zohdi and Wriggers [196], Zohdi [201] and Zohdi et. al
[206].
Remark: For isotropic material responses we have
$ -
$ -

$ -
$ -

(6.15)

%
6.3 MOMENT BOUNDS ON POPULATION RESPONSES
-5
Let us take any effective property, averaged over the subsample, denoted by ,

and correspondingly then is the average of all of the samples. Let
denote the

true effective property. Employing the notation
,
where is any arbitrary second order (strain) tensor.
9
>
MOMENT BOUNDS ON POPULATION RESPONSES
121
Fig. 6.2 Successively embedded partitions.
6.3.1
First order (average) bounds
Box 6.2.4 implies

$

$
6.3.2

$

$

$

$

$

(6.16)
E
Second order (standard deviation) bounds
For the second order moments, the shifting theorems imply

1
$

$

0

0
$

(6.17)

$

$

$

1
122
and

1

1

$

$

(6.18)

$

The results in Boxes 6.16-6.18 imply

$

$

$

$

$

$

$

(6.19)
Remark: In the case of isotropy, the moment results hold for the effective shear
and bulk moduli individually.
6.3.3
Third order (skewness) bounds
From Box 6.5, we may write the skewnesses as

$

$

$

$

$

$

$

$
*
4

$

$

$

$
0
-

$

$

$

$

$

$

$

$

$

$

(6.20)

E

E

E

123
REMARKS
and as

$

$

*

*

$

4

$

4

$

$

0
-

-
$
0

$

3
3
$

-
(6.21)
The above expressions imply

$

$

$

$

$

$

$

$

(6.22)
The derived results allow one to bound, above and below, the unknown SRVE
response in terms of the ensembles averages. Related forms of the first order bounds
have been derived in various forms dating back to Huet [79], [80], [83], Hazanov and
Huet [68], Hazanov and Amieur [69] and Huet [90], [91], Zohdi et. al [194], Oden
and Zohdi [135], Zohdi and Wriggers [196],[203],[205], Zohdi [201] and Zohdi et.
al [206]. The second and third order bounds appear to have been unknown. Using
similar techniques, bounds on even higher order moments, such as the kurtosis (fourth
moment), which measures the peakness of the distribution, are possible.
6.4 REMARKS
The results derived here can be used in conjunction with a variety of methods to
perform large-scale micro-macro simulations. Noteworthy are the multiscale methods: Fish and Wagiman [32], Fish et. al [33], Fish and Belsky [34], Fish and Belsky
[35], Fish and Belsky [36], Fish et. al [37], Fish et. al [38], Fish and Shek [39],
Fish and Ghouli [40], Fish and Yu [41], Fish and Chen [42], Chen and Fish [21] and
Wentorf et. al [184], Voronoi cell methods: Ghosh and Mukhopadhyay [45], Ghosh
and Moorthy [51], Ghosh et. al [47], Ghosh and Moorthy [48], Ghosh et. al [49],
Ghosh et. al [50], Lee et. al [102], Li et. al [109], Moorthy and Ghosh [126] and
Raghavan et. al [149], transformation methods: Moulinec et. al [127] and Michel
et al. [124], partitioning methods: Huet [79], [80], [83], Hazanov and Huet [68],

E

$
$

$

$

3

$

3

$

$

$

4

$

124
Hazanov and Amieur [69] and Huet [90], [91] and the adaptive hierarchical modeling methods: Zohdi et. al [194], Oden and Zohdi [135], Moes et. al [125], Oden
and Vemaganti [136], Oden et. al [137] and Vemaganti and Oden [182] and finally
multipole methods adapted to such problems by Fu et. al [43]. Particularly, attractive are iterative domain decomposition type strategies, whereby a global domain is
divided into nonoverlapping subdomains. On the interior subdomain partitions an
approximate globally kinematically admissible displacement is projected. This allows the subdomains to be mutually decoupled, and therefore separately solvable.
The subdomain boundary value problems are solved with the exact microstructural
representation contained within their respective boundaries, but with approximate
displacement boundary data. The resulting microstructural solution is the assembly
of the subdomain solutions, each restricted to its corresponding subdomain. As in the
ensemble testing, the approximate solution is far more inexpensive to compute than
the direct problem. Numerical and theoretical studies of such approaches have been
studied in Huet [79], Hazanov and Huet [69], Zohdi et. al [194], Oden and Zohdi
[135], Zohdi and Wriggers [196],[203],[205], Zohdi [201] and Zohdi et. al [206].
Clearly, when decomposing the structure by a projection of a kinematically admissible function onto the partitioning interfaces, regardless of the constitutive law, the
error is due to the jumps in tractions at the interfaces (statical inadmissibility). If the
interfaces would be in equilibrium, then there would be no traction jumps. Therefore,
if the resulting approximate solution is deemed not accurate enough, via a-posteriori
error estimation techniques, the decoupling function on the boundaries of the subdomain is updated using information from the previously computed solution, and the
subdomains are resolved. Methods for updating subdomain boundaries can be found
in Zohdi et. al [206]. They bear a strong relation to alternating Schwarz methods (see
Le Tallec [105] for reviews) and methods of equilibration (see Ainsworth and Oden
[2]).
7
Domain decomposition
analogies and extensions
One can consider the multiple sample testing introduced in the last chapter as a type
of domain decomposition. In order to see this, we consider a partitioning approach
whereby a global domain, under arbitrary loading,is divided into nonoverlapping subdomains. On the interior subdomain partitions an approximate globally kinematically
admissible solution is projected. This allows the subdomains to be mutually decoupled, and therefore separately solvable. The subdomain boundary value problems are
solved with the exact microstructural representation contained within their respective boundaries, but with approximate displacement boundary data. The resulting
microstructural solution is the assembly of the subdomain solutions, each restricted
to its corresponding subdomain. The approximate solution is far more inexpensive
to compute than the direct problem. The work follows from results found in Zohdi
et al. [204].
7.1 BOUNDARY VALUE PROBLEM FORMULATIONS

The globally exact solution, , is characterized by the following virtual work formulation:
+-"

)5

E
*

(
#
/& #
E
0

-

E
(7.1)

E
125
126
DOMAIN DECOMPOSITION ANALOGIES AND EXTENSIONS
PROJECT A
KINEMATICALLY
ADMISSIBLE FUNCTION
Fig. 7.1 Construction of the approximate solution.
where

is the Cauchy stress. In the infinitesimal strain, linearly elastic, case

. In order to construct approximate solutions, next we consider the
subdomain boundary value problems, which have the exact microstructural representation contained within the domain, but approximate displacement boundary data on
the interior subdomain boundaries.
To construct the approximate microstructural solutions we first partition the do

. We define the
main, , into N nonintersecting open subdomains

boundary of an individual subdomain
, as
. When employing

the applied
internal
displacement
approach,
a
kinematically
admissible
function,

and
, is projected onto the internal boundaries of the subdomain partitions.
Any subdomain boundaries coinciding with the exterior surface retain their original
boundary conditions (Figure 7.1). Accordingly,
we have the following virtual work

formulation, for each subdomain,

+"

%

-

-

E
*

%
&

8
%
9 N>
*

(
%
&

)(
*

-
(
#
/& #

-

(7.2)
127
ERROR IN THE BROKEN PROBLEMS
The constitutive law and microstructure are identical to that

of
the globally
exact

problem. In the infinitesimal strain, linearly elastic, case
. The
individual
subdomain solutions, , are zero outside of the corresponding subdomain

. In this case the approximate solution is constructed by a direct assembly process
* "
*

0
*
-

%
*
1

%
9 N>
*

%
(7.3)

The approximate displacement field is in

, however,
the approximate traction

field is possibly discontinuous. Logical
choices
of
will
be given later in the

on the internal partition boundaries, then the
work. It should be clear that if
approximate solution is exact.
Remark I: In theory, an approach of projecting the tractions could be developed.
However, such an approach has a number of difficulties, in particular the pure traction
subdomain problems must have equilibrated tractions to be well posed. To impose
this on the numerical implementation level is not a trivial task. Primarily for this
reason the projected displacement approach is preferable. However, the traction case
is discussed later in a chapter on optimization.
Remark II: For example, if one considers a cubical domain, whose response
is simulated using numerical unknowns, for example employing a finite element
discretization, approximately
floating point operations are needed to solve

the system. The value of , typically
, depends on the type of algebraic
system solver used. If the cube was divided into
equal subdomains (subcubes),
'
the# number
of floating point
operations
needed
would
be approximately

the order

on

of
, and thus
.
For example, if one had 1000 subdomains, the broken solution costs between
and times less to compute than the globally exact solution. The advantages
are not limited to the possible reduction of operation counts, since the subdomain
problems can be solved separately, and trivially in parallel. If parallel processing is
used for the decomposed problem, which by construction has a perfect speedup, then

the ratio of costs becomes
, where denotes the number of processors.

9 > 0
' '
'
'

7.2 ERROR IN THE BROKEN PROBLEMS

Since we employ energy type principles to generate approximate solutions, we use
energy type measures for the error. Directly by use of the divergence theorem one
has
128
DISCONTINUOUS TRACTIONS
Fig. 7.2 The consequences of the projection approach.
- * *

5

5

*

- -
- -

-
/
-

5

8
+8.
-

5

* -

*
- * *

- * *

%
- -
- -

*

8B:

-
8
2% 1

*
# & #!+-" %

8 + ( 0 &%

02
"7# !

-

(7.4)
where is an interior subdomain interface, and
=number of interior
subdomain interfaces. For the applied internal displacement case, the tractions may
suffer discontinuities at the interior subdomain boundaries (see Figure 7.2). In the
case of infinitesimal strain, linear elasticity, we have
129
* *

- * * *
#

" - * * - * *
- -
* *

% (
&
1
&
% (

8;:
(7.5)
9 > @
It is convenient to cast the error in terms of the potential energy for the case of

,
linear elasticity,

where is any kinematically admissible function. The well known relationship for
any kinematically admissible function is
"
"

9 9 > 9 >:> 7 9 >

9 >
(7.6)
which is the Principle of Minimum Potential Energy (PMPE). In other words, the
true solution possesses a minimum potential. Therefore,
since is kinematically

admissible, we immediately have "
. The critical
"
observation is that if we can bound
from below, then we can bound the error

from above. In other words, what we seek is
.
To bound
in terms of easily accessible quantities, we first calculate a computationally inexpensive, kinematically admissible, regularized test solution. The
regularized test solution, , is characterized by a virtual work formulation:
& *
+-"
5
9 >

9

*

9 9 > 9 >>
9 >
(
#
/& #
>

-

(7.7)

*
Here the constitutive law for

should be taken to be as simple as a possible, i.e. a
constant (regularized)
fourth
order
positive definite linear elasticity
tensor

, defined via symmetric
)
. In general, the regularized solution (
does not coincide
with
the
field
associated
with
the
decoupling
solution
.
If
we

choose
in Equation 7.12, which is an admissible
function,
we
obtain

"

"
" .
, which implies

Our objective is to form an upper bound on "
in terms of , and
, to obtain a lower bound on

. For the bound to be useful, it should contain
no unknown constants, and should be solely in terms of the regularized
solution and

the material data. In other words we seek
, which will lead

"

to
, thus supplying an upper bound for the

quantities in Box 7.5.
@
0
9 9 > 9 :> >
9 > 9 C
>
9 >

9 >
9 >
9 >
C
130
7.2.1
Multiscale proximity bounds
The solution corresponding to a material with microstructure is , and is characterized

by the following virtual work formulation:
+-"

5

&

*

-
(
#
/&
#

E
*

&

8

8;5

%

5

5
&

$ -

*

8;:
8;:

8
&
(7.8)
. The equivalent
(
#
&
#
*,%
E
*

*
In the infinitesimal strain linearly elastic case, then

complementary form is
+-"

-

*
8;:

*
(7.9)
The solution to the constant coefficient problem, denoted the regular solution,
, is characterized by a virtual work formulation:
+-"

%

&

5

%

$ -

&

@

0

)5

8

-

*
(7.10)
. The equivalent complementary form is

5
8K:

8;:

&

E
*

&
(
#
/&
#
8B5
*

+-"

5

where
*
&
(
#
&
#
8B:

*

(7.11)
For the complementary problem, similar restrictions are placed on the solution and
test fields to force the integrals to make sense. In other words, we assume that solutions
produce finite global energy.
131
We have for any kinematically admissible function

energy norm

1
&
" - *
% (
-

&
* -

* *
, a definition of the primal

3

3

(7.12)
E
where we define the elastic potential by

"
*
2
3
*

*
3
-

8
(7.13)
Equation (7.12) is a form of the Principle of Minimum Potential Energy. In other

words, the true solution possesses a minimum potential. Similarly, for the complementary formulation, for any statically admissible function ( ), one has

5
% (
&
"

5

&

9 > 0. 9 > 9 > 0.
$ -

5

5

3
3
*
(7.14)
E
where

(7.15)
This is a form of the Principle of Minimum Complementary Potential Energy. By

directly adding together the potential energy and the complementary energy we obtain
an equation of balance:
9 > 9
>

0.
@
.0

9
>

(7.16)
kinematically admissible
function,
we obtain
9 9
> , which
9
>
9
>
9 is>:> awhich

.

implies
0

is statically admissible, we have

9 9 > 9 %> > which, which
9
>
9
>

. Combining
implies
If we choose

"
Also, choosing

"
"

the two previous results yields

3
*

*

* *
1
&
% ( 0
5

5

1
&
% ( 0
3
0
7
*
5
(7.17)
132
This estimate is an exact measure of this particular norm, and requires no computation of the exact microfield boundary value problem. If one wishes to isolate either the
primal or complementary norms, one must resort to two-sided bounding techniques.
Clearly, the primal norm is desirable in order to determine the starting vectors proxis kinematically admissible, from
imity to the exact fine-scale solution. Since
Equation 7.12, we immediately have
9 >

9 > 0.

9 > 0. 9 C
>

(7.18)
we
.
@
From Boxes (7.8) and (7.10),

@ , issoughtafter.

9
@

@
@

@
from both> sides yields
>
9

@

@

@
. Since is an arbitrary virtual

to yield
displacement, we may set
@
where the potential

lower
bound,

have

Subtracting

* *
- * * - * - * &
- * * - *
- * * - * *

* *

- *
- *

% (
&

*

$ -

(7.19)

$ -

Therefore, a primal solution difference upper bound is

* *
1
&
% (

-*

C

&
- *
/
$ -

(7.20)
To derive this result we used the fact that a symmetric positive-definite matrix has

a unique square
Cauchy-Schwarz inequality. Note
root,
and
, the

that the
-bound depends only on
and . The process is repeated in
a similar manner to form a complementary upper bound:
5

5

1
&
% (

%
$ -

$ -
/
)5

&
K
$ -

$ -

)5

(7.21)
C
133

,
Analogous to the
-bound, the
. From Equation (7.17), lower primal and complementary bounds follow,
* *

5

5
&
*
3
*
% (
&
3
3
% (

0
*
*
3
*

5

5

5

5

0

3
*
* *

&

0
"

and

& (
$
% (
&
3
*
1

5
"

&)(
$
(7.22)
3

*

0

0

*

*
& (
$
& (

1

* *
1
5
&

% (

% (
&

&
& (
3
*
*
1
3
*

5

(7.23)
These two sided bounds hold for any loading, external geometry and pointwise
positive-definite microstructure. Related forms of such estimates can be found in
Zohdi [199].
7.2.2
Uses of the bounds for domain decomposition
From the previous results one immediately has

* *
1
&
% (

-*

C
/
$ -

- * "

*
&
( 1
(7.24)
To derive this result we used the fact that a symmetric positive definite matrix has

a unique positive square
root,
Cauchy-Schwarz inequality.

and, theand
Note that the
. Therefore, using

Box 7.20 we have

, and thus

9 > 9 C
>
* *
* *

1
1@& &
&
% (
$ &
( (
*
3
9 > @

1
Therefore, in summary
3
1
E
*
&)(
% (
&
(7.25)
Remark I: Bounds of such type have been studied in Zohdi et. al [194], Oden
and Zohdi [135] and
Zohdi and Wriggers [196], however, they required that the decoupling solution be identical to
. Clearly this is unnecessary. In the preceding
bound, the tensor
did not have to be constant, and this has been exploited to
generate hierarchical micromechanical models in Zohdi et. al [194]. More general
134
theoretical aspects of such bounds, including their complementary (dual) formulation counterparts, can be found in Zohdi and Wriggers [196] and Zohdi [199]. Under
certain conditions, such bounds coincide with results found in Huet [78], [79].
Remark II: It is important to realize that the only approximation
in this entire

is an
estimate is "
. Thus the ratio

extremely good indicator of the quality of the estimate. Numerical and theoretical
studies of the proximity of to unity has been studied in Zohdi et. al [194], Oden
and Zohdi [135], Zohdi and Wriggers [196][200] and Zohdi [199]. It is clear that
is free to choose, provided it is symmetric and positive
the tensorial parameter
!
, in other words
,
definite. Thus, ideally, one would want to minimize

!

. Unfortunately, minimization of
is usually impossible,

since
" is unknown. However, another avenue tooptimize
bound
, is tothe
is possible. Formally, the process to obtain
the! optimal
choice,
maximize
"! . For example,
the lower bound
, in other words
if we consider a one-dimensional displacement controlled structure

C

*

7

(7.26)
E
it is easy to show that

&

*
&
% (

also maximizes
where the estimate is exact (
9>
A9 >

-

-

(. ) for

"
(7.27)

E
"

. This
. We remark that for traction controlled loading

1
,
, for a given , the error estimate is exact for any . For
, which is problem dependent, is
three dimensional problems, the optimal tensor
easily computable, since it requires no microscale computations.
value of

7.3 THE CONNECTION TO MATERIAL TESTING

As an example, we apply the decomposition approach to a sample of material containing a significant amount of microstructure (Figure 7.3). We simulated a rectangular
block composed of an Aluminum matrix containing 10240 Boron particles. The
partitioning consisted of 512 subblocks (subdomains), each containing a random dis22 of volume.
tribution of nonintersecting particles occupying approximately

We considered external boundary loading of the form
:

"!
102:3
. - + .
135
THE CONNECTION TO MATERIAL TESTING
512 SUBDOMAINS
20 PARTICLES
PER SUBDOMAIN
MESH DENSITY
VARIABLE GAUSS RULES
Fig. 7.3 A large sample of Aluminum with 10240 embedded Boron particles. In reality the
2
2
2
, while the particles are approximately
sample is approximately
0.0035 mm in diameter.

::

(7.28)
which is relatively standard in micromechanical analyses. The usual motivation for

such a loading is that the sample is considered so large that it experiences nearly uniform strain (linear displacements) on its
a straightforward
boundary. For this
loading,

kinematically admissible choice for
(in Box 7.2) is
. Before we ran
the full tests, we repeatedly refined the finite element mesh for several subdomains
(subblocks), each containing the same number of spheres, but with different random
distributions, until we obtained invariant
total strain energy responses. The result

%
% mesh (46875 dof per test) per subdomain,
was that beyond typically a %
roughly 2344 dof per particle, delivered macroscopically (volumetrically averaged)
mesh invariant responses. The resulting number of total degrees of freedom (n), if the
entire domain was directly simulated with this mesh density, would be n=21,567,171.
Important: Under these special loading conditions, this decomposition process
can be considered as equivalent to the statistical testing of 512 samples of heterogeneous materials in the previous chapter, since the loading on each subdomain will
be uniform. It is important
to realize that since all subdomains are the same size,

on the interfacial boundaries would produce the
that the projection of
0 0 0
136
same strain energy in each subdomain if the material had been uniform. 1 Therefore,
it is logical to (statistically) compare the solution behavior between the subdomains.
Throughout the tests a standard preconditioned Conjugate Gradient (CG) solver was
used. The primary premise for its (wide) use is that a solution, to tolerable accuracy,
operations, which is required with most
can be achieved in much less than
Gaussian-type techniques. For overviews of a variety of such methods see Axelsson
[8]. The CG method, is guaranteed to converge in iterations, provided the algebra is
performed exactly. At each iteration the computational cost is
operations, and
if the number of CG iterations is , then the total cost is
operations,
where typically
. Use of an iterative solver allowed the use of the vector as a starting
vector to speed up the solution process. In other words, a starting vector that captures
a-priori the long wave components (low frequency eigenmodes) of the solution is
advantageous. This increases the effectiveness of the CG searches for this class of
problems. A well known fact with regard to iterative solvers is the fact that they are
quite adept at capturing high-frequency responses. However, they typically may require many iterations to capture long-wave modes. For a discussion see Briggs [14].
For this decomposition, since the number of CG iterations to solve each subdomain
% (Table 5.4). Therefore, the cost to solve all of the subdomain
problem was

%
, as opposed to the direct cost of approximately
problems was

, where
. In this case the approximate cost savings are approxi

mately
. The

time to preprocess, solve and postprocess each 20 particle subdomain took no more
than one minute on a single RISC 6000 workstation. Clearly, the overall solution
process is trivially parallelizable.
We computed the following upper bound, which
coincided with the previous

chapters results, normalized by the energy of
, on the error

9 >
'
'

>
9

'
9 >

'
' '

* *
1
1
&
'
3
% (
&
1
% (
0.

2

stems from computing
where

!

5

(7.29)

% (
&
where the maximized lower bound on the potential is
(7.30)

, has been used. The optimal

.
Remarks: In the general case, in order to determine the optimal regularization for
which
a given boundary value problem, one must search for the components of
minimize any one of the estimates derived earlier, which, for the sake of generality,
we denote by . This entails that boundary value problems must be solved repeatedly

to form, for example, a multivariate Newton-type search for the components of
1 Clearly,
it is a homogeneous deformation in this case.

137
THE CONNECTION TO MATERIAL TESTING
.

"!

Explicitly written, a Newton type system is

,

where
is the Hessian matrix (
), with components
,

is the gradient (
), with components
and where
is the

increment (
), with components
, needed to update the
components
(which are initially guessed). To construct the system of ,
and must be determined numerically, for example using the finite element
method, to determine and its derivatives. As with most complicated systems, finite
difference approximations of the gradient and Hessian components are constructed
with respect to the search parameters by perturbing the search variables around a base
point. The reader is referred to Gill et. al [52] for details. The number of objective
function evaluations to construct the derivatives of numerically with respect to the
is triclinic, then there are up to
design variables can be large. For example, if
21 free constants, while for an isotropic material ther are only two free constants.
The construction of the discrete Hessian is the main expense, since the number of
objective functions to form it scales with the number of search variables as
.
There exist a variety of quasi-Newton methods, which approximate the Hessian in
order to reduce computational effort (see Gill et. al [52]). It is important to realize
, because
that these problems may possess multiple local minima with respect to

of the possibly nonconvex dependence of on
. To deal with this one should
restart the search from different starting guesses, to locate other potential minima, if
they exist. Alternatively, so-called GO (Global Optimization) techniques based upon
rapid, non-derivative stochastic function evaluation, may be useful alternatives. We
refer the reader to Horst and Tuy [76], Zhigljavsky [188] or Davis [27] for reviews.
we note that regardless of the search method, a restriction is that the tensor
Finally,
remain
positive definite. Clearly, for the class of problems under consideration,
solutions must be generated numerically, for example by the finite element method.
We now consider the effect of the use of the finite element method to generate a
subspatial approximation to
, denoted
denoted

. .

9 >
+"
*

- *
9 N>
*

-

*

*

&
9 N>

(
#
/& #

(7.31)
C
9 N>

, and thus
is kinematically
A critical point is that
admissible with respect to the true fine-scale problem. For the regularized and discrete
regularized solutions we have

* *
1
&
% (
3
*

(7.32)
E
and
138

* *
1
3
*
% (
&
*
*

(7.33)
Subtracting the two results yields

* *
1
% (
&
* * * *
" 1 1 "
*

&
3
% ( 0
(7.34)
Clearly, the total difference is composed of two mutually orthogonal parts: (1) a
difference in scales and (2) a numerical error, introduced by spatial discretization
of the coarse scale problem. Formally speaking, estimates can be derived for the
complementary norm in a similar manner. However, there is a fundamental difficulty
with such an approach
space of admissible trial stress functions must satisfy
since the

!
equilibrium,
, and the test functions must be divergence free,

! . However, formally
proceeding we have
5

3
*
% (
&

5
*
5
(7.35)
E
and
5
1

5
% (
&
3
*

5

5
(7.36)
E
leading to
5
1

&
% (
5

3

" 1 & 1 " 1

&
% ( 0
(7.37)
Finally, by adding Equations 7.33 and 7.36 we obtain an exact mixed estimate for a
solution generated by a standard finite element method ( ) and a solution generated
by a method which generates a statically admissible solution ( )

* *
1
&
% ( 0
5

5
1

&
% (
3
*

0

(7.38)
7.4 A TOTAL ORTHOGONAL SUM

Clearly, for the class of problems under consideration, solutions must be generated
numerically, for example, as in the last section, by the finite element method. We now
consider the effect of the use of
the finite element method to generate a subspatial
approximation to , denoted , governed per subdomain by
139
A TOTAL ORTHOGONAL SUM
+"

%

-

-

*

&

-

-

9 N>

9 N>
&
&
(
#
/& #
(7.39)
9 N>
&

. This inner approximation
A critical point is that
allows the development of a straightforward decomposition. Using a subspatial approximation, by direct expansion we have
*

- *
- *
*

- -
.
- * *

Subtracting

- *
*

*

-
- *
* *
* *
- * * - * *

- * * - * *
02
.
* * * *
0

$
(7.40)
- -
- *

- * *
2
0
% 8
9 > from both sides of Equation (7.40) leads to

- -

where, by Box 7.39, for

*

- -

%

*
- -

%
- -
*

- -
*

<*
(7.41)
- * *

(7.42)
E
which immediately implies

* *
1
&
% (

* *
1
&
% ( 0

* *
1
&
% (
(7.43)
Therefore the estimate of the error, for example in Boxes 7.25 and 5.20 implicitly
contains the numerical discretization error. Clearly, the total error is composed of
140
two mutually orthogonal parts: (1) a partitioning error introduced by subdividing

the body and applying inexact boundary data on the interfaces and (2) a numerical
error, introduced by spatial discretization. Therefore, if one has a numerical error
estimator, one can isolate the pure decomposition error. Such filtering processes
have been performed in Zohdi and Wriggers [200]. For details on elementary a priori
numerical error estimates, see the text of Hughes [92], while for a posteriori estimates
see Ainsworth and Oden [2].
7.5 ITERATIVE EXTENSIONS

In order to achieve further reduction of the global/local error, the subdomain interfacial boundary conditions must be improved. Clearly, the error bound in Box 7.25,
and the orthogonality expression in Box 7.43 still hold, provided that the modified,
hopefully improved, local boundary data deliver a globally kinematically admissible
solution. Thus, we now discuss a few methods which consist of updating the numerical degrees of freedom on the interface of a finite element discretization. Without loss
of generality, and for simplicity of presentation, we consider that we have meshed
each subdomain in the substructuring process so that a global nodally-conforming
trilinear hexahedral finite element mesh results.
7.5.1
Method I: global/local CG iterations
9 >
9 >
One can consider a process of updating interfacial nodal values as

having as its goal

the minimization of the discrete analog of
, namely
, which is convex
in terms of the finite element nodal displacements. There exist a variety of classical
techniques which could drive a process to minimize potential energies, the most
applicable in the linearly elastic case being the CG method, which, for the sake of
completeness, we briefly discuss in the context of a global/local
strategy.
1

Suppose we wish to solve the discrete
system
.
is a symmetric

positive definite
matrix,
is the
solution vector,
is the
righthand side, and is the number of discrete unknowns. We define a discrete

potential
. Correspondingly, from basic calculus

;

.
;
$

$;

=
we have

. Therefore

the minimizer of the potential is also the solution to the discrete system. The
CG method is based upon minimizing by successively updating a starting vector.
Therefore, the substructuring method described in this
work can be considered as a
construction of an advanced initial starting guess, , to the solution of the global
problem formed by directly discretizing the entire body with no decomposition. If

we are to directly apply the CG method, one will reduce
, to the minimum
value for the corresponding global mesh discretization. Since a CG approach updates
all values in the entire domain simultaneously, via a global matrix-vector multiply,
each iteration can be thought of as providing global
transfer of information, i.e.

microstructural corrections to the initial solution , and thus the interfacial boundary

9 >
141
ITERATIVE EXTENSIONS
conditions. The minimum of the discrete potential is the global solution for the
given finite element discretization. However, instead of a direct application of the

, then
CG method, one can perform a set number of global CG iterations,
reevaluate the local boundary value problems with the new local boundary data. Per
subdomain we have
operation counts, where is the number of local
CG iterations for subdomain solution convergence. For simplicity, we have assumed
that one would use the CG method to solve each subdomain problem, and that
is approximately the same for each subdomain. Therefore, for total subdomains
we have
operation counts. Let us denote the number of times that
we globally iterate and resolve
the local problems
as, . Thus, in total, we have

. Therefore, a direct CG approach
and a projecting-global/local resolve approach compare as follows
9 >

9 >
9 9 > > 9 >
"

"

1 3 %" 1
' '
Clearly the magnitudes of and determine the relative performance of the two

&
"
(7.44)
methods.
7.5.2
Method II: iterative equilibration
Clearly, when decomposing the structure by a projection of a kinematically admissible

function onto the partitioning interfaces, regardless of the constitutive law, the error
is due to the jumps in tractions at the interfaces (Box 7.4). If the interfaces would be
in equilibrium, then there would be no traction jumps. Therefore, an approach which
attempts to eliminate the jumps in tractions is sought next.
GLOBAL INTERFACE
PROBLEM IS SOLVED
LOCAL INTERFACE (NODAL)

PROBLEMS ARE SOLVED
Fig. 7.4 Global equilibration and local equilibration.
142
7.5.2.1 Global equilibration For global

F equilibration one treats the entire interface as a single subdomain, denoted
, and solves the following (Figure
7.4)

+-"
*

-
5

-
&
(
DE *
&
(
#
/& #
(7.45)
)(
The new interfacial displacements are then computed, and the subdomain problems
are re-solved. The process is repeated until convergence. The entire interface problem
can be quite large. A less expensive approach, amenable to trivial parallel processing,
follows.
7.5.2.2 Local (nodal) equilibration
For local equilibration, one prescribes

(Figure
7.4), surrounding each interfacial
small non-intersecting
"nodal
domains",

( = number of interfacial nodes)
node

+-"

-
-
5

&

&
)(

(

(
#
/&
#
E

(7.46)
The new interfacial displacements are then computed, and the subdomain problems
are resolved. The process is repeated until convergence. The advantage of such an
approach is that the "nodal domains" can be processed independently of one another.
The disadvantage is that global equilibrium is not satisfied throughout the entire
interface. Either of the introduced approaches can be considered as a fixed-point
iteration, which we discuss next.
,9 >
7.5.2.3 General fixed-point iterations

Consider a general system of coupled

partial differential equations given by
, where is a solution, and where it
is assumed that
the
operator,
,
is
invertible.
One
that the sequence of iterated

desires

solutions,
,
,
converge
to
as
. If
is a function of

one
says
that
is
the
order
of
iteration.
It
is
assumed

that the

iterate can be represented by some arbitrary function
.

One makes the following split
. For this method
to
be
useful

the exact
should
be reproduced. In other words, when
, then
solution
1

.
Therefore,
one
has
the
following
consistency

condition
, and as a consequence,

. Convergence of the iteration can be studied by defining the
error vector:
102
/.

5
9 > 99 > 9
9 > 9 9 >>
9 >

9 >

> > 9 9 >:>

9 >
9 >
9 >
143

*
*

*
$ -

$ -
$ -
$ -
$ -

$ -
$ -
9

One
sees
that,
if
is
linear
and
invertible,
the
above
reduces
to

9 >>
9 > . Therefore, if the spectral radius of , i.e. the magnitude
of its largest eigenvalue, is less
" asthan unity .for each iteration , then for any
arbitrary starting solution

(7.47)
E1

E2
U(0)
U(L)
Fig. 7.5 A one-dimensional equilibration example.
A 9 9 >
A >
9 >
9 > " 9 >

9 > 9 >
102
7.5.2.4 An example Again

consider

the one dimensional structure

,

,
in Figure 7.5. Consider two subdomains,

and one interfacial
problem,
which
are
solved
in an iterative manner
,

until

2

3

1

- .

*
*

*

7
*
1
0

1

*

0
1

-
(7.48)
Since there is only a single one-dimensional interface, the global and local equilibration methods are
solution,
same. A simple starting value, that of the regularized
the

, where
is chosen. After some algebra we have,

, yielding
"
9 >

"
144

*
1

2

$ -

9.

1
0
$ -

E
(7.49)

1

2
> .
1

2
*

2
*
1

0

7

One sees that convergence occurs if
, which implies G . The rate of
convergence is linear and is controlled by the size of .
To illustrate the approach we compared it to classical a classical overlapping
Schwarz method (see Le Tallec [105]), which for this simple structure could be
written as

2

3

1

0

E
*

7

*

-
0
1

$
1
. .
0
1
E
*

(7.50)
Depicted in Figures 7.6-7.8 are the relative performances of both approaches
(equili

bration
and
overlapping
Schwarz).
The
parameters
were
,
GPa,

GPa. Each of the approaches iteratively improved the interfacial

data by balancing the jumps in the tractions via information exchange between subdomains. Clearly the overlapping Schwarz method attained a faster rate of convergence
since more information is provided via the fact that the entire extended subdomain
problem is solved during the iterative process. Generally, the local equilibration
method is less expensive to compute for three dimensional problems, however, as
illustrated by the preceding one dimensional example, it will probably exhibit slower
rates of convergence.
7.5.3
Remarks
Depending on the problem, one method may exhibit superiority over another in terms
of the overall cost for a desired solution accuracy. It is important to remark that, in
the general three dimensional case, for either the global/local CG approach or the
iterative equilibration appraoch, one can determine whether the solution is improving

by monitoring the potential of each iterate,
. A decreasing potential indicates

that the solution is improving since
and
"
is fixed. Clearly the error bound in Box 7.25 holds for the iteratively generated
kinematically admissible solutions, and is

* *

1
&
% (
3
*

9 >
3
9 9 > 9 >>
*

9 >
(7.51)
E
where

9
> 9:9

> @ >
9 9

> @ >

145

, and retain it for any subsequent
iterations, since it is iteration independent. As we have
indicated
earlier, the only

approximation in this entire estimate is "

, which is exact

for the previous one dimensional example, using the regularized parameter

Thus one should maximize

over
, via
"

. Application of the global/local CG approach,iterative equilibration

or classical overlapping methods to three dimensional problems is under current
investigation.
Remark: The rate of convergence of the CG method is

* *
&
% (
* *

&
(7.52)
% (
E
where a functional of the condition number of the discretized stiffness matrix.

Proofs of the various characteristics of the
method can be found in Axelsson [8]. It is
controls a great deal of the error during the
clear that the magnitude of "
"
initial stages of the CG process. Therefore, on the discrete level, a good initial choice
for , which captures the long range features of the fine-scale solution is desired.
As mentioned, the CG method will then successively endow the long range solution
with local (high frequency) corrections. Therefore, it is clear that is
is available

it should be projected onto the fine-grid to produce a starting vector

.
146
0.5
EQUILIBRATION
OVERLAPPING SCHWARZ
0.45
0.4
NORMALIZED ERROR
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0
10
15
20
ITERATIONS
25
30
35
40
0.5
EQUILIBRATION
OVERLAPPING SCHWARZ
0.45
0.4
NORMALIZED ERROR
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0
Fig. 7.6 NORMALIZED ERROR

*
3
BOTTOM for
.
"%
*
8
10
ITERATIONS

&

(
$
&
&
(
12
14
16
. TOP for an overlap of
18
2
and
147
9e+08
EQUILIBRATION
OVERLAPPING SCHWARZ
8e+08
JUMP IN TRACTIONS
7e+08
6e+08
5e+08
4e+08
3e+08
2e+08
1e+08
0
0
10
15
20
ITERATIONS
25
30
35
40
9e+08
EQUILIBRATION
OVERLAPPING SCHWARZ
8e+08
JUMP IN TRACTIONS
7e+08
6e+08
5e+08
4e+08
3e+08
2e+08
1e+08
0
0
Fig. 7.7 JUMP IN TRACTIONS

*
*
2
3
and BOTTOM for
.
"
*
8
10
ITERATIONS

&
12

14

&
16
18
. TOP for an overlap of
148
4e+06
EQUILIBRATION
OVERLAPPING SCHWARZ
3.5e+06
ERROR WORK
3e+06
2.5e+06
2e+06
1.5e+06
1e+06
500000
0
0
10
15
20
ITERATIONS
25
30
35
40
4e+06
EQUILIBRATION
OVERLAPPING SCHWARZ
3.5e+06
ERROR WORK
3e+06
2.5e+06
2e+06
1.5e+06
1e+06
500000
0
0
"
8
10
ITERATIONS

*
-
Fig. 7.8 "ERROR WORK"

*
*
2
overlap of
and BOTTOM for

&

12

&
14

16
18

. TOP for an
8
Nonconvex-nonderivative
genetic material design
Until this point in the analysis, we have considered the microstructure of the material as given data. It is important to realize that microstructural selections, such as
combinations of volume fraction, stiffness mismatches and particulate topology, are
usually the result of tedious trial and error laboratory experiments. In this chapter,
a computational strategy is developed to simulate and accelerate the associated trial
and error development of tailored dispersed-type materials. The objective is to develop a procedure to determine multiple possible microstructures that can deliver the
same prespecified effective linearly elastic response. The determination of multiple
possible microstructural combinations gives an engineer a wide variety of design
and manufacturing alternatives. However, there exist a variety of difficulties in the
computational design of macroscopic solid material properties formed by doping a
base matrix material with randomly distributed particles of different phases. Three
primary problems are (1) the wide array of free microdesign variables, such as particle topology, volume fraction and mechanical property phase contrasts, which force
associated objective functions to be highly nonconvex, (2) the associated objective
functions are not continuously differentiable with respect to design space, primarily
due to microscale design constraints, such as limits on the desired local stress field
intensities, and (3) the effective responses of various finite sized samples, of equal
volume but of different random distributions of the particulate matter, exhibit mutual
fluctuations, leading to amplified noise in optimization strategies where objective
function sensitivities or comparisons are needed. The effects in (3) become amplified
when computing design sensitivities or comparisons needed in optimization strategies, thus extremely large samples must be used to eliminate detrimental size effects.
The presented work concentrates on the parametrization of microscale parameters
149
150
NONCONVEX-NONDERIVATIVE GENETIC MATERIAL DESIGN
inherent in such materials, with the goal of computational optimization of material

microstructure. The outline of the presentation is as follows. The fundamentals of
computational material design formulations are given. Attention is drawn to some of
the difficulties encountered, i.e. nonconvexity and nondifferentiability of objective
functions, as well as size effects. Solution strategies based on genetic algorithms are
then presented.
8.1 COMPUTATIONAL MATERIAL DESIGN
S=10
S=2
S=1
S<1
Fig. 8.1 Parametrization of a generalized ellipsoid, with various parametrized cross-sections

shown.
We are primarily concerned with the construction of an inverse problem, where

combinations of particulate and matrix materials are sought which minimize the
following normalized objective function
2
;=
'
;=
;=

(8.1)
COMPUTATIONAL MATERIAL DESIGN

151
where
is a prespecified desired effective response,
is the effective response

produced by a trial microstructure, and where " " is an appropriate admissible norm
that is selected for the application at hand. A microstructural design can be defined

through an -tuple design vector, denoted
, for example
consisting of the following components: (1) the mechanical properties of the foreign
particulate matter (2) the volume fraction of the foreign particulate matter and (3) the
topology of the foreign particulate matter. Ellipsoidal shapes are qualitatively useful
since the geometry can closely represent a variety of particulate types, for example
platelets when the ellipsoids are oblate or needles (discontinuous fibers) when the
ellipsoids are prolate. Such shapes can be generalized by considering the following
(Figure 8.1):
0
where the

>

(8.2)

. produce nonconvex shapes, while

s are exponents. Values of
'
'
'
values produce blocklike shapes. The following are free variables:
Particulate mechanical properties: for example, assuming local isotropy of the particles, the
bulk and shear moduli, 1 1 and : 1 (two variables).
Particulate topology: for example, the polynomial order of generalized ellipsoids,

ables).

"
Particulate aspect ratio: for example, defined by

, where
2
2
for prolate geometries and
) for oblate shapes (one variable).

"

(three vari-
,
4
Particulate volume fraction: for example, 1 2

% , where is the volume occupied by the
particulates, and 6 is the total volume of the material (one variable).
Particulate orientation: for example, within the last decade there have viable been processing
methods developed to control the orientation of particulate matter by coating them with a conducting
liquid material and introducing them into the molten matrix material (three free variables, i.e. Euler
angles). Thereafter an electrical current is applied, forcing the particles to align themselves along
the field lines. This can produce globally anisotropic properties (see Michaud [124] for details).
Matrix mechanical properties: for example, if the matrix material is variable, assuming local
isotropy of the matrix material, 1 - and : - (two variables).
Therefore, in this simple linear elasticity formulation, we have 12 microdesign

variables. It is remarked that for both manufacturing and physical reasons, generally,
each design variable will have a constrained search space. For example none of
the design variables in this formulation can be negative, with the exception of the
orientation angles, furthermore the volume fraction must be less than one. Further
discussions of inverse formulations of this kind can be found in Hyun and Torquato
[93].
Remark: There exist a variety of difficulties whenever one attempts to design
the effective mechanical properties of particulate materials, which on a practical
level involves finite sized samples possessing irregular heterogeneous microstructure.
These are discussed next.
152
8.2 MINIMIZATION METHODS

>

"!
Following
a standard
Newton-type multivariate search, a new design increment,

, for a microstructural
design vector, , is achieved by solving

, where
the following system,
),
is the Hessian matrix (

, is the gradient (
), with components
with components

and where
is the design increment (
), with components

. After the design
increment
has
been
solved
for,
one
then
forms an updated

design vector,
, and the process is repeated until "
.
Explicitly, the incremental system is
G

&

&

&

&

&

8.2.1

&

&

(
&

&

&

&

&

& (

&

(

&

& (

& (

& (

& (

(

(

(

(

& (

..

-

2
'

& (
&
&

&

&

(

(8.3)
Difficulties due nonconvexity
A severe difficulty is that the system in Equation 8.3 becomes noninvertible throughout
the design space, due to the nonconvexity of the objective function . We illustrate
the problem of nonconvexity via a simple
one dimensional example. Consider a

simple one dimensional bar of length composed of random particles, i.e. strips in

one dimension. There are a total of
strips,
dark strips, each of thickness ,
and the white strips representing the rest of the material. The Youngs modulus
corresponds to the "particles", while corresponds the "matrix". Suppose
one
/#

,
of
such
a
structure
defined
by
wishes
to
design
an
effective
response,

5

two point boundary value problem,
. Consider
,
the following

, where is a constant. One finds that the effective response

,

is the harmonic average, which can be written as

9> 9 >
A 9
A >

where
of
and

(8.4)
, . and . Clearly, there is no unique combination

to produce the same desired effective response. Consider

(8.5)
153
MINIMIZATION METHODS
where
is a desired effective response. Following a standard Newton-type
multivariate search, a new design increment,
one would obtain the following Hessian

and
system for the two design variables,
=
& (
& (

& (

*

& (
& (

(8.6)
& (

1
However, this system becomes noninvertible throughout the design space, due to the
nonconvexity of the objective function . As an example see Figure 8.2.
COST FUNCTION
2.5
2
1.5
1
0.5
0
0.7
10
15
MISMATCH
20
25
30 0
0.1
Fig. 8.2 The behavior of the objective function for
8.2.2
0.2
0.8
0.6
0.5
0.4
VOLUME
FRACTION
0.3
2
*

and

2
Local field design constraints
Consider the deviation of the stress field from its volumetric average,
which leads to
0.9

(8.7)
Therefore

(8.8)
154
In order to incorporate the deviation in a cost function, we introduce a tolerance

where ideally,

(8.9)
then the level of exceedance is incorIf the normalized deviation exceeds the
porated as a multilateral constraint to the macroscopic objectives by

"
8.2.3
"
C

"

"

if
where

be given shortly.

.

and

(8.10)

otherwise. Detailed examples will
Difficulties due to nondifferentiability
Independent of the problem with nonconvexity, due to the presence of design constraints, the objective may be nondifferentiable. For most complicated systems, usually finite difference approximations of the gradient and Hessian components are
constructed with respect to the design parameters. The finite difference size for the
approximate numerical derivatives, which is different for each component, is denoted

. In a practical setting, for example, for each variable, the numerical derivative
step sizes are scaled
current value of that variable, by a small num to the size of. the
ber
,
Numerically, the components of the gradient and
Hessian can be approximated by the following second order central finite difference
stencils such as

&
"! #! #! #!
"! "!

! ! ! !
#! #!
&
&
&
&
(8.11)

0
The number of objective
evaluations
to form the Hessian and
0 ofnecessary
. , wherefunction
gradient is
is the0 number
microstructural
design variables.
9 > objective evaluations
This stems from the fact that one needs
for the
off-diagonal terms of the Hessian, for the diagonal terms, and one evaluation
for the base point (the current design). Therefore even with a small number
of
0
design
variables
the
number
of
objective
function
evaluations
can
be
quite
large.
For
0
0 0 microstructural
0 variables introduced earlier, . , and one

example with
the 9 12
.
.
>
.

has
objective function evaluations per search step. It
'
is clear that the construction of the discrete Hessian is the main expense. There exist
NONCONVEX-NONDERIVATIVE GENETIC SEARCH
155
a variety of quasi-Newton methods, which in some manner attempt to approximate

the Hessian in an inexpensive way (see Gill et. al [52]). However, regardless of
the reduction of the number costly objective function evaluations, such procedures
require an immense amount of effort which grows dramatically with sample size.
Implicitly, such objective functions require a sufficient degree of regularity for the
derivatives to make sense. For example, due to the fact that design constraints are
present, there are many points throughout the domain
where the objective function

is not twice differentiable, implying
, thus rendering the Hessian
matrix nonsymmetric. This implies that the second order finite difference stencils
are insufficient, since they inherently construct a symmetric approximation to the
Hessian. Therefore, their use to construct the numerical Hessian can only be justified
near smooth convex optima. To put it succinctly, such methods are not robust.

8.3 NONCONVEX-NONDERIVATIVE GENETIC SEARCH

The lack of robustness of classical gradient based deterministic optimization processes
can be rectified by application of a family of methods, usually termed genetic
algorithms. Genetic algorithms are search methods based on the principles of natural
selection and, as such, they are highly probabilistic. There are a variety of such
methods, which employ concepts of species evolution, such as reproduction, mutation
and crossover. Such methods stem from the work of John Holland and his colleagues
in the late 1960s and early 1970s at the University of Michigan (Holland [75]). For
reviews of such methods, the interested reader is referred to Goldberg [54], Davis
[27] and Onwubiko [139]. A recent overview of the state of the art of the field can
be found in a collection of recent articles, edited by Goldberg and Deb [55]. In
this presentation, we concentrate on adapting genetic-type algorithms to nonconvex
inverse problems of micro-macro materials design. The key conceptual feature is
that the microstructural parameters form a genetic string. Thereafter, evolutionary
concepts are applied to a population of such strings. One possible implementation of
such an approach is as follows:

% ' )( * # &
%+ . '
# &
! ! ! ! !/
, -

0/1 324 ( 5 " )6 #
( 87 !

!9: 7 ! #
!;: 8 4 2 8 (<(

= ? > #

> #

=
& > #

> #

> . @
(<( ( 4 87 4
-A@ >

0B1 ( 4 8 8" " + @ + @ +
,

@ , @ ,

!C: (3(ED GF 4 3 !

0H1 3 JIK57 2 3I 4 8 # =

!" $#
2

156
We remark that the definition of fitness of a genetic string in this algorithm

indicates the value of the objective function. In other words, the most fit genetic
string is simply the one with the smallest objective function. STEPS 1-7, which are
associated with the genetic part of the overall algorithm, attempt to collect multiple
local minima.1 At first glance, it seems somewhat superfluous to retain even the top
parents in such an algorithm. However, as we will illustrate shortly, implementing the
algorithm in Box 9.31, it is found that the retention of the top old fit genetic strings
is critical. By observing Figure 8.3 one sees that if the objective functions is highly
nonconvex, there exists a strong possibility that the inferior offspring will replace
superior parents. Therefore, retaining the top parents is not only less computationally expensive, since these designs do not have to be reevaluated, it is theoretically
superior. With parent retention, the minimization of the cost function is guaranteed
to be monotone, if the top parents are retained.
FOR ALL PAIRS:
TWO ADJACENT PARENTS
10
11
12
RANDOM CONVEX COMBINATIONS
GLOBAL
MINIMUM
TWO OFFSPRING
GENE POOL
Fig. 8.3 LEFT: The class of objective functions of interest. MIDDLE: Genetic string
combinations. RIGHT: A loss of superior older genetic strings if the top parents are not
retained.
8.4 A SEMI-ANALYTICAL EXAMPLE

As an example, consider the widely used Hashin and Shtrikman bounds [65], [66],
for isotropic materials with isotropic effective responses; for the bulk moduli,
1
$
"
2
1 - 6
+
$ +

4B& -
6
4
1 1 6
+
and for the shear moduli

:
$
"%
2
: - 6
8
1 It

-
$ 8
& -
1
$ /( & +
8 &4 +
198
$ +
"

)$' 1
(8.12)
"

(8.13)

: 1 6
8
$ 8
) ' 1
&+
6
8 &4 +
198
is remarked that if the function is allowed to be greater than unity, one can consider the resulting
convex combination (offspring) as a mutation.

157
A SEMI-ANALYTICAL EXAMPLE
where
and are the bulk moduli and
and
are the shear moduli of the respective phases (( ) and (
)), and where is the second phase volume
fraction. Such bounds are the tightest possible on isotropic effective responses, with
isotropic two phase microstructures, where only the volume fractions and phase contrasts of the constituents are known. As we have mentioned earlier, during effective
material design development, when selecting particulate micro-additives for a base
matrix, information about the changes in the otherwise (relatively) smooth internal
fields, corresponding to the matrix material alone, is valuable to characterize a new
tailored materials performance. One way to characterize the smoothness of the microscopic field behavior is via concentration tensors, which provide a measure of the
deviation away from
the
mean fields
throughout the material. Consider the follow

! and (II)
ing identities: (I)
"# $
&% "'( ")*
"#!*
"+ ,

. By direct manipulation
we obtain
-/.102
-5.02
-/.802
43
6 76

-/:702
-/:;02
43; =930>
6 76 ; =6>

-/:702
-5:702
-5:702
;= 3 >
' 6
6 6 ; = 6 >
-5:?0 2

; =3 6 76 ; =6' ; =<3+ >>=(>
2

2
2
2
where
6' ; =<3+ $ 3
A 76
B?C
E& F/G
=
)
;=
>
;=

'
3
;=
>
-5:?02
-5:702
2
(8.14)
(8.15)
we may write, for the variation in the stress HJILKNM)O7P

Thereafter,
"#
, which reduces to
6 7> =.>
;=
;=
$ 3 -/.802RQ;SUT
>
2
=
>
-5.102
-/.802
2
"+
K5M'P
(8.16)
H is known as the stress concentration tensor. Therefore, once either H or KVMWO are
known, the other can be determined. In the case of isotropy we may write
Q;SUT ) :6 : ' :X3
L6 1 ' 13

8 2
(8.17)

:6 ' :X3
76 1
1L6 ' 13
?6 :
7ZU[ "
]7ZU[ "
"^_
]"^/

\
where Y
and where Y
Clearly, the mi \ are minimally
)
a` . For. the
Y
Y
crostress fields
distorted
when
matrix, since

"# #
" "
Pb
, therefore
b
5
1
.
02
-5.802
-5.802

-5.102
-/.102
' 6 =.>
' 6 =(>
(8.18)
2
2
2 =.>
43
43
Q;ST
and therefore, in the case of isotropy,

+
Q;2 SUT
43

76
) '
+

8
Q;2 ST
43

?6
)'
8
Therefore for the deviation in the particulate stress fields away from the mean
(8.19)
158

L
Y

!
`

(8.20)
Y
and
% " ^ !," ^ 1 I % " ^ !" ^ !1

Y `
" ^ , " ^ ,
I
Y

(8.21)
and for the matrix material

L
Y

!
`

(8.22)
and

`
% " ^ !" ^ 1 I % " ^ !" ^ !1
Y
" ^ " ^
I
Y

In order to incorporate the deviation in a cost function, we introduce a tolerance

where ideally,
`

`

Y

(8.24)
Y
(8.23)
and
`

Y
Y

Y

`

Y

(8.25)
If the normalized deviation exceeds the corresponding

, then the level of exceedance is incorporated as a multilateral constraint to the macroscopic objectives
by
As an example, our immediate goal is to computationally design the macroscale
effective bulk and shear moduli O and O , using convex combinations of the`& Hashin% " O7P
Shtrikman bounds
for
the effective
moduli O!#" O%$
% as`( approximations

`
and O '" O%$
. The micro-macro objective function
" O?P where )
"
is
A SEMI-ANALYTICAL EXAMPLE

O

`&
O

O
`

159

`

P `

P `
P `

, (II) if

P

\ ,
where (I) if
P
, then \
, then \
'
, (IV) if

P
'
,
, then
, then
(III) if
P

, (VI) if

P

,
(V) if
P
, then
, then

, (VIII) if
`

(VII) if
P
, then
, then
. Here the design variables are
, and their constrained ranges
are P
,
, P
and ! P " # . There are

\

!

P+

&

two characteristics of such a formulation which make the application of standard

gradient type minimization schemes, such as Newtons method, inapplicable: (I) the
incorporation of limits on the microfield behavior, as well as design search space
restrictions, renders the objective function not continuously differentiable in design
space and (II) the objective function is nonconvex, i.e. the system Hessian is not
positive definite (invertible) throughout design space.

We consider
matrix material (aluminum) of fixed material
a base

` values, 8
GPa and
GPa. The desired values are O
)) GPa, O
)

=0.5 and
GPa,
=0.5. The (constrained) design variables ranges are
5`

`
N`

`
P
,)
and
)

P

$
)

$
)

<
\
P
$
)

.
The
weights
were
set
to
<
`

)
. We used "
for the Hashin-Shtrikman bound combination. The number
of genetic strings was set to 1000, for ten generations, keeping the offspring of the
top 100 parents
after
` each generation. Three cases were considered: (I) Additionally
keeping top
)) parents after each generation, thus with 800 new genetic strings
`
infused and (II) Not keeping top
)) parents after each generation, thus with
900 new genetic strings infused.
Figure 8.2 depicts the results for several generations. After ten generations, a
dominant genetic has appeared for each approach ((I) and (II)). The minimization of
the cost function is guaranteed
to be monotone, if the top parents are retained, i.e.
% O
% O $ , where O $ and O are the best genetic strings from
%$'& (
*

.
/ 0 21 / 0
-&

)
&
+ #, +
0
0
160
SEARCH
K.P.C.
N.K.P.C.
K.P.N.C.
N.K.P.N.C.
1.972
2.172
2.377
2.507

4.755
5.730
6.753
7.408
0.3179
0.2771
0.2466
0.2311
0.0000000098
0.0000004874
0.0000000057
0.0000056610
Table 8.1 Top designs after 10 generations. K.P.C.=keeping parents with constraints,
N.K.P.C.=not keeping parents with constraints, K.P.N.C.=keeping parents with no constraints,
N.K.P.C.=not keeping parents with no constraints.
`

and respectively. There is no such guarantee if the top parents
generations
are not retained. While the nonretention of parents allows more newer genetic strings
to be evaluated in the next generation, numerical studies conducted thus far imply,
for sufficiently large populations, that the benefits of parent retention outweigh this
advantage, as well as any disadvantages of inbreeding, i. e. a stagnant population
with insufficient new genetic material. As the results illustrate, not retaining the
parents is suboptimal due to the possibility that inferior offspring will replace superior
parents. Furthermore, parent retention is computationally less expensive, since these
designs do not have to be reevaluated, although this was not a concern for the types
of simulations in this semi-analytical example.
8.5 DIRECT COMPUTATIONAL APPROACHES

In order to perform meaningful computational optimization of the material microstructure, one needs reliable effective responses. It is clear that for the relation between
averages to be useful, i.e. statistically representative, the sample must be so large
that, for further enlargements, K5M O changes minimally. For illustration purposes,
we used
(particle) combination. Typically, the boron
an aluminum
(matrix)/boron
`

(
) ,
)
is
used as a stiffener for the aluminum matrix

, 1
). We used a moderate particulate volume fraction
(
of approximately 22 . The following particle per sample sequence (Figure 8.5) was
used to study the dependence of the effective responses on the sample size: 2 (5184
DOF), 4 (10125 DOF), 8 (20577 DOF), 16 (41720 DOF), 32 (81000 DOF) and 64
(151959 DOF) particles. A meaningful
parameter to track was the ratio of the diam

eter of the individual particles ( ) to the length of the sample ( ). Throughout the
tests, we considered
a single
satisfying Hills condition
4

combined
`boundary

` loading

(Hill [70]),
,
. We tracked the effective bulk
) ))
and shear moduli, O and O , respectively. At each sample size, size effects, i.e.
scatter in the effective responses, occur. To eliminate the effects of scatter, the tests
were performed 100 times for each sample size (each time with a different random
$
$'& (
DIRECT COMPUTATIONAL APPROACHES
161
0.018
KPC
NKPC
KPNC
NKPNC
0.016
0.014
FITNESS
0.012
0.01
0.008
0.006
0.004
0.002
0
1
5
6
GENERATION
10
0.12
KPC
NKPC
KPNC
NKPNC
0.1
FITNESS
0.08
0.06
0.04
0.02
0
1
5
6
GENERATION
10
Fig. 8.4 Generational values of (TOP) the best designs objective function and (BOTTOM)
the average of the best 10 designs objective functions.
particulate distribution) and the responses averaged. 2 Since the effective bulk and
shear responses behave in a quantitatively similar manner, for brevity, we show only
the effective shear responses. Figure 8.5 depicts the dependency of the responses
with growth in particle number per sample, keeping the volume fraction constant.
For example, the time to preprocess, solve and postprocess a ) particle sample finite
element test is roughly 10 seconds on a single workstation, using a research code
written by the author.3 For three successive enlargements of the number of particles,
i.e. 16, 32 and 64 particle samples, the responses differed from one another, on average, by less than 1 . Typically one will select reasonably large samples (for a
2 This
procedure can be done trivially in parallel.

course, these tests could have been performed trivially in parallel, and thus the entire 100 sample
testing run would have taken 10 seconds.
3 Of
162
given ) with a fixed number of particles, for example 20. However, the samples will
exhibit mutual fluctuations in their respective effective responses due to the various
random particulate realizations. A stabilization procedure is then applied, by which
a series of the samples are tested until the progressive ensemble average converges.
For example, consider a certain microstructural design specification ( ), and a process where a sample of finite size, with a random microstructure, is tested, and the
effective response recorded. Consider a repetition of the test for another equal sized
sample, with the same microstructural design vector, however with another random
microstructural realization. The tests are repeated, for more and more samples, until
the sequential
in the ensemble average falls below a given tolerance for further

` change

tests (
)
!$
/

$
$

$

(8.26)
However, one must also interatively enlarge the sample sizes. Therefore, one must
(1) for a given sample size, ensemble test the samples until stabilization and (2)
enlarge the sample (increase the number of particles with the volume fraction fixed)
and ensemble test the samples until stabilization (3) compare the results of (1) and
(2) and (4) if the difference is below a set tolerance then one may stop, however, if
the difference is above a tolerance the process (1)-(4) is repeated with larger samples.
From a practical point of view, adaptively selecting the number of samples needed
to stabilize the results is important since (1) it is impossible to know a-priori the
size for approximate stabilization of the samples, (2) it is impossible to know apriori how many samples to test for the ensemble average to stabilize and (3) it more
cost effective than simply taking a large (overkill) preset number of samples of a
large size. Therefore in Box 9.31, STEP 1.2, one must only replace COMPUTE
with ITERATIVELY ENLARGE AND ENSEMBLE-COMPUTE. This procedure
is quite effective in eliminating detrimental size effect noise.
8.6 PRIMAL AND DUAL OBJECTIVE FUNCTION ENVELOPES
Let us consider the implications of the

bounds on the responses with regard to objective
functions such as
8.6.1
KUM K M
KUM P .
The implication of

For illustration purposes only, consider the following scalar analogies
% O

O

O

O

O

% O

O

O

O
(8.27)
PRIMAL AND DUAL OBJECTIVE FUNCTION ENVELOPES
5184 DOF
d/L=0.595
48
10125 DOF
d/L= 0.472
163
(GPa)
(2)
46
20577 DOF
d/L= 0.375
41270 DOF
d/L= 0.298
44
< >:< > 1/2

( < >:< > )
2 =
(4)
(8)
81000 DOF
d/L= 0.236
(16)
(32)
42
(64)
151959 DOF
d/L= 0.188
40
50 60
10
20 30 40
NUMBER OF PARTICLES
43.5
MUSTAR
43
42.5
42
41.5
41
0
10
20
30
40
50
SAMPLE
60
70
80
90
100
Fig. 8.5 TOP: The values of the effective shear responses for samples containing increasingly
larger numbers of particles. One hundred tests were performed per particle/sample combination
and the results averaged. BOTTOM: 100 SAMPLES: The shear responses, , of a block with
20 randomly distributed boron spheres embedded in an aluminum matrix. Each point represents
the results of one test.
164
TRACTION TESTS
WITH LARGER SAMPLES
DISPLACEMENT TESTS
WITH LARGER SAMPLES
E*
D
D
E* E *
E*
E* +
D
E*
Fig. 8.6 Two sided envelopes for the infinitely sized samples objective function.

1
/
O
O, O
O
) ,
) and
where
the function
O
O

) . Generally,

% O
with
.
Referring
to
Figure
8.6,
we
see
that
if
,
then
O
O
O
O varies

% O , while if O
O , then % O
% O . If O

O
O

%
% O
O , then there is an indefinite relationship between
O and
O . For the
K5M O
KVM O and
tensorial analogue, the above results directly imply that, with KVM O
O
KVM
,

1

!"#$&%!')(! +*

, .-0/ !1" , 1 2'3(546#7&%!')(! +8
(8.28)
O
O

O

9
If KVM O 9
, then there is an indefinite relationship between
5
K
M
N
K
M
V
K
M
/ /

% KNM O and % K5M O .
8.6.2
The implication of
1
As in the previous subsection, following similar arguments, we have, defining KNM O

O O
KVM O
KVM , KNM
,
<;/< = !1 = , 1 2'3(546#7&%!')(! +*

,
= !"#$&%!')(! +8
(8.29)
O
O
O 9
V
K
M
V
K
M
N
K
M
If KVM O 9
, then there is an indefinite relationship between
/ /

% KNM O and % K5M O .
Remark I: It is clear that, in the case of displacement controlled tests, since the
ensemble-generated effective responses are stiffer than the true effective response,
they achieve the desired response K M O with microstructural designs that are, casually
9
SEARCH TYPE
K. P.
5.540
6.349
K. P. I. P.
4.728
N. K. P.
3.720
N. K. P. I. P.
3.730

AR
0.358
0.183
1.373
4.759
0.341
0.213
7.670
0.385
0.206
5.144
0.344
0.205
165
NUMERICAL IMPLEMENTATION
5.155
7.055
7.901
.00032
1.178
6.210
6.210
6.210
.00130
1.379
3.528
5.653
4.606
.00073
1.208
5.950
5.950
5.950
.00263
Table 8.2 Top designs and average of the top 6, after 10 generations. From top to bottom:
(1) keeping parents and with an nonisotropic polynomial order (2) not keeping parents and
with an nonisotropic polynomial order (3) keeping parents and with an isotropic polynomial
order (4) not keeping parents and with an isotropic polynomial order.
speaking, softer than they should be. For the case of traction controlled tests,
since the ensemble-generated effective responses are weaker than the true effective
response, they achieve the desired response KVM O
with microstructural designs that
are stronger than they should be. Mathematically speaking, the preceeding two
statement can be represented by
9
;/

= = = - / = =

- A/ B? C D *

= ;A/ = B? C = D 8

(8.30)
Remark II: The preceding results imply a series of envelopes as depicted in

Figure 8.6, which converge to the true objective function as the sample size tends
towards infinity. The implication to objective functions such as the ones considered
in this work involving isotropic responses, when the constraints are inactive,

B?C
7

!"#%$%&('#)&(*,+-#/.%0214365
7
B7C

(8.31)
!")#)$8&9':#)&(*8+;#<'=2>?"@$2A
are rather obvious.
8.7 NUMERICAL IMPLEMENTATION
( ( (
CB
CB
A fixed mesh of approximately

trilinear hexahedra or 2344 numerical degrees of freedom per particle, for a total 46875 degrees of freedom, was used. This
mesh density delivered mesh independent results over the course of the numerical
166

0.08
KEEP PARENTS/NONISOTROPIC S
KEEP PARENTS/ISOTROPIC S
NOT KEEP PARENTS/NONISOTROPIC S
NOT KEEP PARENTS/ISOTROPIC S
0.07
0.06
FITNESS
0.05
0.04
0.03
0.02
0.01
0
1
5
6
GENERATION
10
0.18
KEEP PARENTS/NONISOTROPIC S
KEEP PARENTS/ISOTROPIC S
NOT KEEP PARENTS/NONISOTROPIC S
NOT KEEP PARENTS/ISOTROPIC S
0.16
0.14
0.12
FITNESS
0.1
0.08
0.06
0.04
0.02
0
1
5
6
GENERATION
10
Fig. 8.7 Top design (TOP) and top six average designs (BOTTOM) after 10 generations.
experiments. In other words, the same final designs occurred using finer meshes.
To illustrate the search process, continuing with 20 particle subsamples, the effective response produced by a sample containing approximately 22 boron spheres
in an aluminum matrix ( =25.9 GPa =77.9
The effec GPa) was firstcomputed.

and O

. Our objective
tive response was approximately O
was to find alternative
microstructures

which could deliver the same effective response ( O
and O
), while obeying the stress perturbation
constraints, to a specified tolerance.4 We considered an objective function of the
following form:
+9
the tests, we considered a single combined boundary loading satisfying Hills condition,
" # $! % & .
2
, !
4 Throughout
A A
167
NUMERICAL IMPLEMENTATION

B?C
!"#%$%&('#%&9*,+;#
7

.%02143 5
7
B7C
! "#)$8&('#)&(*,+-#

'=2>?"@$ 5

. .

.
.

!
A
B?C
D
+-#%$%& +;>,1
'! &(&)=2>8':'
With this in mind, the algorithm was started away from the original spherical aluminum/boron
design location,

however, still prescribing the same objective: O
and O
. The matrix material was fixed to be aluminum,

however, all other design parameters, with the exception of the particle orientations, since isotropic objectives were
sought,
` were allowed to vary. The following
weights were used:

`
. The following exponents were used

in Equation 8.32: !
. The constraints on the design
space were: (1)
5`

`
N`

`

,
(2)
)

P

$
)

)

P

$

` ) , (3)
5`

`
`

"%$
)
#"%$ P
"&$ $ ) , (4) (' P
'
' $ ) and (5)

.
The
volume
fraction
was
controlled
via a par)
)
*)P
)
) $

ticle/sample size ratio (one variable), defined by a subvolume size +(,.2
-0/ 1436 14351 ,

where 7 is the number of particles
in
the
entire
sample
and
where
is
the
length
B B

of the (cubical) sample,
. A generalized diameter is defined, , which
is the diameter of the smallest sphere that can enclose a single particle of possibly
non-spherical shape. The ratio between the generalized diameter and the subvolume
[
is one design parameter defined by ) ,. -0/ 8 9 . The number of genetic strings was set
to 20, for ten generations keeping the
offspring of the top 6 parents. Two cases
were
considered: (1) Keeping top
parents
and
(2)
Not
keeping
top

parents
and also keeping the polynomial order the same for all three axes and allowing all axes
to vary. It appears that even for isotropic overall responses, allowing the polynomial
order of the axes to vary appears to minimize the objective even further. Table 8.2 and
Figure 8.7 depict the results. As mentioned earlier, retaining the parents generates a
better set of designs at the end
of the multiple generation cycle. The total number of
% ` (B % : =< , where is the number of generations,
global evaluations is ;:
<
: is the total number of genetic strings in the population, and if the number of
parents kept after each generation. The minimization of the cost function is guaranteed to be monotone, if the top parents are retained. For the best design, generated
when retaining the parents and with nonisotropic topological exponents, 146 genetic
strings were tested after 10 generations. A total 1664 samples (each containing 20
particles) were tested to stabilize the ensemble averaging process, for an average of
11.8 with samples per string. The cuboid type microstructures that occur are similar
to microstructures generated by Vigdergauz ([192], [193]) as well as by Torquato and
Hyun [179], however for somewhat different objective functions in two dimensions
for periodic media. Clearly, the rounding of the corners of the particles of the cuboid
microstructure (Figure 8.8) is a product of the stress field constraints.
+9
-$
)

+9
(8.32)
168

8
8 8

8

Fig. 8.8 LEFT: The original spherical microstructure:
,
,

(volume fraction: 6
),
,
. RIGHT: A general alternative microstructure
(to spheres) produced by minimizing the perturbation in the stresses within the material while
maintaining the same effective response.
9
Inelastic responses
A drawback in the use of solid structures composed of heterogeneous media is the

existence of highly distorted and intense microfields, which may cause a significant
amount of microscale failure in the form of subcontinuum scale microfissures, interface separation, void nucleation, etc. As the loading progresses, these local failures
increase in size and number and eventually merge to produce observed macroscopic
failure. Therefore, the desirable aggregate response of the heterogeneous material
may be lost. In particular, of interest to a structural engineer is knowledge of the
deterioration of KNM O with progressive loading, or overloading. The typical approach
to describe the deterioration of the macroscopic response is to employ phenomenologically based macroscopic damage parameters, and corresponding evolution laws.
These are applied directly on the macroscopic level in order to mimic the observed
weakening (softening) responses of material samples with progressive loading. There
are a variety of such approaches, loosely referred to as Continuum Damage Mechanics (CDM) models. For an overview of the field see Krajcinovic [99]. Common to
such approaches is that the mentioned evolution law is based on, or operates solely
upon, macroscopic quantities such as KNM O . Typically, the evolution law is a curve fit of
laboratory observed macroscopic behavior relating to . Consequently, it is
fair to state that such phenomenological models are not really predictive of responses
that are unknown a-priori. It is for this reason that numerical simulations can play a
pivotal role in helping to determine inelastic responses of microheterogeneous solids,
in particular those which have been designed solely based on linear analyses.
169
170
INELASTIC RESPONSES
9.1 INTRODUCTION
The focus of the presentation is primarily on the numerical and algorithmic aspects
of the solution to a class of time-dependent inelastic finite strain processes involving
simultaneous damage and plasticity in solids possessing randomly distributed particulate microstructure. The model problem considered is closely related to models found
in Huet [81]-[86], Huet et. al [87] and Huet [88]-[91]. Following the usual approach,
the spatial discretization of the heterogeneous microstructure is achieved by the finite
element method, while finite differences are used for the temporal discretization. For
this class of problems, in order to accurately capture the microstructure, one must use
extremely fine finite element meshes. Schemes based on attempts to solve the entire
system simultaneuosly (monolithically), then applying Newton-type algorithms, are
not robust due to the possiblity of a zero, or near zero, tangents that occur when
describing a weakening material. For example, consider a material law of the form

and is` a continuity
KNM 'I , where KVM is the virgin undamaged material,

(isotropic damage) parameter
(Kachanov
[95]),
where
)
indicates an

`

initial state, )
and
takes on different values
) . The scalar function
throughout the body, dictated by the evolution law. This constitutive law possesses
KNM I
KVM I , -0/ KNM
I . In this case,
a tangent of the form
KVM
can lose positive definiteness when the change in is sufficiently large, which
in turn will lead to an indefinite-type system of algebraic equations. This one primary
reason why explicit-type schemes enjoy wide usuage for complicated phenomena
involving damage.
One can interpret solution techniques of so-called "staggering-type" used in multifield simulation processes, whereby, within a time step, each field equation or evolution law is solved individually, allowing only the primary field variable to be active,
as a type of explicit approach, if the process is not recursive. In explict staggering
schemes, after the solution of each field equation, the primary field variable is updated, and the next field equation is addressed in a similar manner. In the standard
approach, after this process has been applied, only once, to all of the field equations,
the time step is incremented and the procedure is repeated. The classical solution
process is nonrecursive, and is highly sensitive to the order in which the field variables are staggeredly solved. The order of staggering is usually selected based upon
somewhat ad hoc arguments pertaining to which field drives the other. For example for a lightly coupled thermo-mechanical problem, within a time step, one might
first solve for the thermal field, using the displacement field from the previous time
step as given data, thereafter using the solved for thermal field as a given load in the
balance of momentum to solve for an updated value of the displacement field. Thereafter, the time step would be incremented, and the process repeated. For accurate
numerical solutions, such approaches require small time steps, primarily because the
staggering error accumulates with each passing increment. For details, see Park and
Felippa [145], Zienkiewicz [190], Schrefler [152], Lewis et al. [107] and Doltsinis
[28]. A particularly lucid review can be found in Lewis and Schrefler [108]. A further primary problem is when there are multiple fields with complicated coupling,

171
A MODEL PROBLEM
it is extremely difficult to ascertain in what order to stagger the solution process.

Furthermore, as time progresses, for complicated systems, the coupling can change,
becoming stronger, weaker or oscillatory.
In this presentation an implicit, temporally adaptive, recursive staggering scheme
is developed to rapidly simulate a class of time-dependent multiphase solid materials
which are inelastically deforming at finite strains. The problem is cast as a coupled pseudo-multifield system consisting of (1) a dynamic equation of momentum
balance, where the primary field variable is the displacement, (2) the evolution of material changes where the primary field variable is a damage state function and (3) the
evolution of inelastic strains in the solid where the primary field variable is a plastic
strain function. Clearly, the damage and plasticity variables are implicitly functions
of the displacement, however, for the staggering scheme strategy, it is convenient
to formulate them as individual fields. The key concept for the strategy to operate
efficiently is to estimate and control the so-called staggering error, i .e. the error due
to incompletely resolving the coupling between the field equations in a staggering
process. This error is a function of the time step size. However, because the coupling
is temporally variable, possibly becoming stronger, weaker, or oscillatory, it is extremely difficult to ascertain a-priori the time step size needed for prespecified error
control. Presently, a solution strategy is developed, whereby the time step size is manipulated, enlarged or reduced, to control the intrinsic contraction mapping constant
of system staggering operator in order to induce desired staggering rates of convergence within each time step, thus controlling the staggering error. The overall goal is
to deliver accurate solutions where temporal discretization error control dictates upper limits on the time step size, while the iterative staggering strategy refines the step
size further to control the staggering error. Numerical experiments are performed to
illustrate the solution strategy in practical three dimensional settings involving solids
possessing irregular heterogeneous microstructure.
9.2 A MODEL PROBLEM
\
A structure which occupies an open bounded domain in
$ , with boundary
, is considered. The boundary consists of
on which the displacements
are
prescribed and a part Z on which tractions are prescribed. The mechanical properties
of the heterogeneous
material are characterized by a spatially varying elasticity tensor
\ \
KVM
$ 3 and various evolution (rate) parameters for describing inelastic strains
and damage in the material. KVM is assumed to be symmetric bounded positive definite
tensor-valued functions. For reasons of clarity, strong forms are used to derive the
governing equations, possibly assuming more regularity than warranted. Afterwards,
only the weak forms, which produce solutions which coincide with strong forms,
when the solutions are smooth enough, are employed.

172
INELASTIC RESPONSES
9.2.1
Constitutive assumptions: infinitesimal strain case

We first discuss infinitesimal strains and then a finite strain analog. To a first approximation the constitutive relationship can be written as KVM I
, where
is the Cauchy stress. The symbol is used for the inelastic strains. Specifically, we
describe the material changes (damage) and plastic strains via evolution laws of the
following form1:
A B7C D -
+*
*

- ; -

! , +*

A
B7C
D

" # $ - " Q ST "' & )( " &+*-,
& Q;ST + *

A
B7C
D
%

(9.1)
with the following activation conditions:
.0/1
.0/1
.0/1
.0/1
^
, 3
-0/ 2
^
, 3
-0/ 2
4^ [ Z6587:9
^
I

\
I
1 4^ [ Z6587:9
'
')
,

\ \O

)
4 [ Z 5;7:9 '< and

ZU[ 1 ZU[

\
\ 4 [ Z 5;7:9 < <O .
ZU[
^
-0/ = ABA @?? ABA , @\ C D$EF , G
-0/ = < , = and = < are spatially variable
Here O , <O and \O , 4 [ Z ,.>
ZU[
^
\IH and where KJ represents plastic strains. As
material parameters,

ZU[
\
ZU[
\
')
indicated in the introduction, here KVM

K5M , where KVM is the virgin undamaged
material, and is the classical
continuity
(isotropic
damage) parameter (Kachanov
`
[95]), where
) indicates an initial state. The scalar function takes on
different
values throughout the body, dictated by the evolution law. For any
later
time

`
, one can consider a deteriorating (weakening) material where )
( and
\ being negative), and where as
) the material becomes completely damaged.
For further details on these types of phenomenological (damage) formulations, the
interested reader is referred to Kachanov [95], Krajcinovic [99] or Lemaitre and
Chaboche [104].
ML
1 The
field
could, in general, include other effects, such as thermal strains.
173
A MODEL PROBLEM
9.2.2
A geometrically-nonlinear extension
Now we construct a geometrically nonlinear formulation, as an extension to the infinitesimal strain

` case. We consider the case of finite deformations with small elas ) and moderate inelastic strains. Later in the work, an updated
tic strains (
Lagrangian staggering type scheme, formulated directly in the current configuration, will be used, and thus current configuration-based material laws are advantageous. A relatively straightforward extension to the infinitesimal deformation con
P
stitutive laws is to employ
the Eulerian/Almansi
strain tensor2 ( , -0/

):
< H
<
A B?C D
-

+*

- ; -

+ , +*
A
B?C
D

" $ - "
Q ST " &
( " & *-,
& Q;SUT +*
B?C
D
A
%
(9.2)

where
, ^ ZU,[ ^4 [ Z , -0^ / = ABA @?? ABA , ZU[ \ C ^ D$E F ,. -0/ = < , = ^ and = < being
\IH , P
material constants,

P , 4 [ Z P
^
4 [ Z P with flow activation conditions specified as specified previously where
,
are referential coordinates and are current coordinates. The

indicates
symbol
differentiation with respect to the current configuration. Such

#
a law is frame indifferent under rigid body rotations and translations. Clearly,
is objective as are the scalar-based evolution laws for and . By construction,
the geometrically nonlinear formulation matches the geometrically
linear case
for

M
infinitesimal deformations. If one adheres to the split
, then M

, where M
. The Eulerian spilt is closely

"$#

H
H

< # #
<
related to the well-known split of Green and Naghdi ([59], [60]), M # ,. -0/ M
M ,

where M , M # and M
are the total, elastic and plastic Green-Lagrange strains, has

been used. Casey [19] justifies the Green-Naghdi split as an approximation valid for
small elastic and moderate plastic strains. Computational aspects of such strain splits,
in the context of pure elastoplasticity, are discussed in Simo and Ortiz [167] and Simo
[168]. We remark that M
is
not a true elastic strain, except in very special cases.

One can simply consider M
M as an internal parameter used in a constitutive law,
which is not intended to have any kinematical meaning at finite strains. The Eulerian
2 As
with
in the infinitesimal strain case, the field
could include other effects, such as thermal strains.
174
INELASTIC RESPONSES

M . We remark that the coefficients

type of split is merely a push forward of M
through \ could be temperature dependent, although in ths work we do not consider
this case. Also, we remark that for the plastic strain, the manipulations involving
the Second Piola-Kirchhoff stress, , are necessary to achieve a frame-indifferent
constitutive formulation. For the damage this is unnecessary due to the fact that it is
a purely scalar quantity.
9.3 A RECURSIVE STAGGERING ALGORITHM

As indicated in the introduction, a popular class of solution techniques for coupled
systems are so-called "staggering schemes", whereby, within a time step, each field
equation is solved individually, allowing only the primary field variable to be active.
After the solution of each field equation, the primary field variable is updated, and the
next field equation is addressed in a similar manner. In an explicit variant of such an
approach, after this process has been applied, only once, to all of the field equations,
the time step is incremented and the procedure is repeated. An explicit type of this
process is nonrecursive, and for accurate numerical solutions, the approach requires
small time steps, primarily because the staggering error accumulates with each passing
increment. For details, see Park and Felippa [145], Zienkiewicz [190], Schrefler [152],
Lewis et al. [107], Doltsinis [28], Piperno [147], Lewis and Schrefler [108] and Le
Tallec and Mouro [106]. In Zohdi [208], a recursive staggering strategy which allowed
the adaptive control of time step sizes, was developed. In that approach, in order to
reduce the error within a time step, the staggering methodology was formulated as
a recursive fixed-point iteration, whereby the system was repeatedly re-solved until
fixed-point type convergence was achieved. Extending the approach in Zohdi [208]
to the class of problems at hand, we consider a system consisting of four fields:
,
and
. Clearly, the pseudo-field variables, and
, are implicitly function
" ,
of the displacement, i. e.
and
, however, for the staggering

scheme strategy, it is convenient to formulate them as individual fields. In some
cases, inertial terms can play an important role in the model under consideration. A
relatively straightforward approach for the inertial terms, which is amenable to time

step adaptivity, is ( being the time step counter)

F

-

-

-
<

(9.3)

which collapses to a central difference-like stencil of
1 $
,
P Z $
when the time step size is uniform. For the evolution equations, consistent with the
fully staggered solution approach we freeze the and , using the most current
state variable values, and integrate analytically. Additionally, the geometric configuration of the system is frozen during each equation solve, and is updated only at the
end of each individual system iteration. This is essentially an updated Lagrangian
formulation, where all variables are referred to the last calculated configuration as opposed to a total Lagrangian formulation where all variables are referred to the initial
A RECURSIVE STAGGERING ALGORITHM
175
configuration. In this particular case, the last calculated configuration is the previously computed one within the staggering scheme. Employing weak formulations,
algorithmically the staggering scheme is as follows:

!#"%$
&('*)+-,/.0 21 #
!3435!1.!% "
&('*)+67 1 8 9(.! #:./;

<
= ">
==
=
?
= " *@ >
>
= =F

% <
= "> %
%
= ? %
= " % @ % = F

F
ED
=
?BA
= " A % @ > %
C
C
F F D ?
= F F

?
KML >O P
=
GH >
=JI "N

&('*)+
9-MOQ9 2RS;3# 54J
.T
5U V I P
= JWXZY
= P
= F [ Y

= "\
^]7_O`MaJb/adceb*fhg8WjiJY
= g F [ Y
=

(
= lnm g
=joqp r $ s lnm P
= ltm P
= /uwv lnm P
= 2x
= !yz

= P
= v g z (

=}| P <P v g z
(

F
F F~

=t{

=t{
= vg z
= v g zO"$

(

( [
=

=t
F

GH x
= I " KML A C
=

F F P
= P
= F F
=

&('*)+ -M3 H 3#539 1 ;3# 2
F F P
= F F
= u;

&('*)+754( 3# 4;;
H #" 1-
&('*)+-754( 3# 4
H 3#
n %7-"% .0 2(9R( #h

"k
(9.4)
Here
is the space of admissible trial functions, while
is the space of admissible
test functions. For most loading cases and data these spaces will correspond to
.
In an abstract setting, one can consider the following active set solution strategy
where the underlined variable is momentarily active,the corresponding field equations
solved, the active variable updated, and the process repeated for the next field equation:
S7
7

A

M
M4-}
-.8

s

7
M }
j} -

d

7
M
^

A
A

A
(9.5)
A
A in the form presented
A
A
A to algebraic
A
Writing the system
leads
systems which are symmetric and positive definite. Therefore, somewhat standard iterative solvers such as
the preconditioned Conjugate Gradient Method, can be used. Such solvers are highly
advantageous since any previous solution, from a previous time step or staggered
iteration can be used as the first guess in the solution procedure, thus providing a
head start in the solution process.
176
INELASTIC RESPONSES
9.4 ISSUES OF CONVERGENCE AND STAGGERING ERROR

CONTROL
The algorithm outlined in Boxes 9.31 and 9.32 can be considered as fixed-point
scheme, whose convergence within each time step is dependent on the time step size
itself. The step size can be manipulated, enlarged or reduced, to induce the desired
rates of convergence within a time step, in order to achieve an error tolerance within
a prespecified number of iterations. Consider
the general equation:

It is advantageous to write
this in the
form

(9.6)

A straightforward fixed point iterative

scheme
is

(9.7)

(9.8)
The convergence of such a scheme is dependent

on the behavior of . Namely, a
5
sufficient
condition
for
convergence
is
that
is
a
contraction mapping for all
,

Convergence of the iteration can be studied by defining the error vector

. A necessary condition for convergence is iterative self consistency, i.

e. the exact solution must be represented by the scheme
. Enforcing
this condition, a sufficient condition for convergence is the existence of a contraction
mapping
$

0 0

where,
if

as

L

(9.9)

for each iteration , then 0 L
for any arbitrary starting
solution
A
. If
is differentiable,
we
may
write

! = " #$
(9.10)

and thus, provided that the functional

behavior is locally convex
(Perron,
[146], Os5
5
A
trowski [143], [144], Ortega and Rockoff [141] and Kitchen [98]) in the neighborhood
of the solution,
= "

= 5
0 13 2 = 4
5=

Therefore, unconditional convergence is attained if for 0 ,

!
!
!
& #$
% %
'()( ) '+* -, " /.
(9.11)
67
$ 8 , 0 .
General overviews of fixed-point algorithms can be found in Ames [4].

Remark: An alternative approach is to attempt to solve the entire system simultaneuosly (monolithically). This would involve the use of a Newton-type scheme,
which can also be considered as a type of fixed-point iteration. Newtons method is
covered as a special case of this general analysis. To see this, consider
9 0
177
ISSUES OF CONVERGENCE AND STAGGERING ERROR CONTROL
9 0
9 = 0 0

<

dating can be written as
6 0
0 0
. Therefore, Newton up-
9 0
(9.12)
is the residual, and 6 0

9 is the 6 tangent. Therefore, in the fixed-point
6
form one has6 the< opera

tor

. The gradient is

9

6
6

, -0/
where $
. Therefore, the convergence criteria is,

where
9

,. -0/

, following the approach for the first fixed-point scheme,

0 1 2

5=
|
Equivalently, one can write
9 < 9 9 ~

4

= . As
(9.13)
mentioned at the outset of this work, one immediately sees a fundamental difficulty,
due to the possiblity of a zero, or near zero, tangent when employing a Newtons
can lose positive
method to a system describing a weakening material, since
definiteness when the change in is sufficiently large, which in turn will lead to an
indefinite-type system of algebraic equations. Therefore, while Newtons method
usually converges at a faster rate than a direct fixed point iteration, quadratic as
opposed to superlinear, its convergence criteria is less robust than the presented fixedpoint algorithm, due to its dependence on the gradients of the solution. In this work,
since we consider materials which experience damage and plasticity, under dynamics
conditions, it is unlikely that the gradients of remain positive definite and thus we opt
for the more robust gradient-less staggering scheme. A further issue complicating
the use of Newton-type schemes for simulation microheterogeneous solids is the
low regularity of the the operator with respect to , thus taking gradients highly
unattractive and in many cases impossible.

9.4.1
Explicit time dependency
For the class of coupled systems considered in this work the coupled
operators

spectral radius is directly dependent on the time step discretization . We consider
a simple example which illustrates the essential concepts. Consider coupling of a
second-order system and two first order systems

;
*

;
*

(9.14)
9 ,
178
INELASTIC RESPONSES

,
< 1

< 9

When discretized in time, for example with a backward Euler scheme,

and 1
9

following coupled system:

4
)

)

<1

$
9

$
, one obtains the

$ 1 <1
1
(9.15)
where is a time increment counter. For a recursive staggering scheme of Jacobitype, where the updates are made only after one complete iteration, considered here
only for algebraic simplicity, one has3

)
<1 0

31 2 4
0
$
$
0

$

$
$
Rewriting this in terms of the

)
)
0 4

132
)

The eigenvalues of

132
1
<1
1

0
0

$
$
0

$

-form yields
0

<1 0
$

1 $

<1 $
4 1
(9.16)

1

0
$
$
9
are 4 9

1
$ 1 <1
0
31 2

1

4
0
<1 0
1
0

$
$

$
(9.17)
. One sees that the spectral radius of the staggering
operator
grows quasilinearly with the time step size, specifically, superlinearly as
9
. Following Zohdi [208], a somewhat less algebraically complicated example
illustrates a further characteristic of such solution processes. Consider the following
example of reduced dimensionality, namely a coupled first-order system
)
and
) . When discretized in time with a backward Euler scheme and
repeating the preceding procedure one obtains the following -form
<

<

)
132

)

4
3 A Gauss-Seidel-type
1
0
<1 0

!
0 132 <1 4

"
1 #
<1 # 0

(9.18)
approach would involve using the most current iterate. Typically, under very general
conditions, if the Jacobi method converges, the Gauss-Seidel method converges at a faster rate, while if
the Jacobi method diverges, the Gauss-Seidel method diverges at a faster rate. The Jacobi method is
easier to address theoretically, thus it is used for proof of convergence, and the Gauss-Seidel method at the
implementation level.
179
ISSUES OF CONVERGENCE AND STAGGERING ERROR CONTROL

The eigenvalues of are

. One sees that the convergence of the staggering
scheme is directly related (linearly in this case) to the size of the time step. The
solution to the example is

; +,

- +,

+,
0 132 4

+,

0 - 132 4 ;

; 888

(9.19)
and

; ,

- +,

+,

0 - 31 2 4

+,
0 12 4

; 888

(9.20)
As pointed out in Zohdi [208], the time step induced restriction for convergence
matches the radius of analyticity of a Taylor series expansion of the solution, which
converges in a ball of radius from the point of expansion to the nearest singularity,
around the previous time increment solution. In other
words, the limiting step size

is given by setting the denominator to zero,
! , which is in agreement
with the condition derived from the analysis of the eigenvalues of .
Generally speaking, if the recursive process is not employed, i. e. an explicit
scheme, the staggering error can accumulate relatively rapidly. However, an overkill
approach involving very small time steps, smaller than needed to control the discretization error, simply to suppress a nonrecursive staggering process error, is computationally inefficient. Therefore, the objective of the next section is to develop
a strategy to adaptively adjust, in fact maximize, the choice of the time step size
to control the staggering error, while simultaneously staying below a critical time
step size needed to control the discretization error. An important related issue is to
simultaneously minimize the computational effort involved. The number of times
the multifield system is solved, as opposed to time steps, is taken as the measure of
computational effort, since within a time step, many multifield system re-solves can
take place.
9.4.2
<
Contraction-mapping time stepping control
The type of contraction condition discussed is sufficient, but not necessary, for convergence. In order to construct an adaptive staggering strategy, we modify an approach
in Zohdi [208] originally
developed for thermochemical multifield problems, one ap

proximates #"
$ , where one
estimates the error within an iteration to behave

$ &%
according to
, where % is the initial staggering error
'% ,
and is a function intrinsic to the system. Our target or ideal condition is to meet an
error tolerance in a given number of iterations,
more, and not less. One
writes
)(*,not
+
54
z
this in the following approximate form,
-$
,
where
is the
%
/.1032
number of desired iterations. Therefore, if the error tolerance is not met in a desired

0
7 $

0
180
INELASTIC RESPONSES
number of iterations, the contraction constant is too large. Accordingly, one can
solve for a new smaller step size, under the assumption that is constant,
#
##

(*,+

C =z
C =

(9.21)
The assumption that is constant is not overly severe, since the time steps are to be
recursively refined and unrefined. Clearly, the expression in Equation 9.21 is used
4
for time step enlargement,
if convergence is met in less than iterations. One sees
(* +

4
that if % z %
and
, then the expression in Equation 9.21 collapses
to a
4
ratio of the error tolerance to the achieved level of iterative error after
iterations,
(*,+

=
= z C z , and thus the step size will be scaled by the ratio of the error

to the tolerance. We first define the normalized
error,
within each time step (
)
#

,

and its corresponding violation ratio

(9.22)
0
, -0/
(9.23)
:

Here, a general formulation has been written for a more general

set of variables, which
and a scaling factor
will be introduced later. One then determines the violation

u@ $ C = z
@ =$ C =
@
,. -0/
(9.24)
Thereafter, the following criteria for temporal adaptivity is adopted
"!#

(
=<O
%$
! &N('*)(,+-/.
01-32/-5467!#80159". #
015>3?!#7-7!@2!A. #
;:
$ #
) *C C
B
&
# H$ ,&E 0FGG/5;-A 7!#;019

JI
"8!#5
! ,& K)J-+,.

=<O

=P
0159D7!#D.
019DL,M.
0F,32/,N4O7!#D019". #
015>3?!#7-7!@2!A. #
301GQ/N87!#;019
# ) HB $*C C #
!
5
;: z
(9.25)
NUMERICAL EXPERIMENTS
181
9.5 NUMERICAL EXPERIMENTS

8
8

8
Fig. 9.1 LEFT: The numerical resolution of

particles, with an (oblate) aspect ratio

(20 particles shown).
, resulting in a volume fraction of
of
RIGHT: A single particle.
Material Property

(GPa)
(GPa)
#4

< # 4

#4
=

= <

Matrix
Particles
78
26
-0.1, -0.05, -0.01
-0.1, -0.05, -0.01
0.0001, 0.00005, 0.00001
100
100
2700
230
172
-0.2, -0.1, -0.02
-0.2, -0.1, -0.02
0.00005, 0.000025, 0.00005
300
300
2330
Table 9.1 Material properties used in the computational examples.
To illustrate the B algorithm,

we considered a cube of matrix material, with normalB
ized dimensions
, containing randomly distributed inhomogeneities. The
various physical properties of the two materials are shown in Table 9.1. We consider
a set of topological microstructural variables which can be conveniently parametrized
by a generalized ellipsoid
182
INELASTIC RESPONSES

<

7
(9.26)

where the ' s are exponents. Values of '

produce nonconvex shapes, while '
values produce blocklike shapes. The types of suspensions to be introduced
in
the matrix binder can be controlled by (1) the polynomial order, ' , ' and ' , (2)

the aspect ratios defined by "%$ , -0/ 9 , where
, "%$
for prolate

geometries and "&$
for oblate shapes and (3) the volume fractions.
We
considered

the following
boundary conditions on

,

<B
the exterior of the cube:

,
=0.01,
,
where
is
a
referential
position

vector to the boundary of the cube. We note that while the displacement loading
is relatively small, the internal strains may be significantly larger. The material
parameters, selected only for the purposes of numerical experiment, are shown in
Table 9.1. From a practical engineering point of view, macroscopic quantities, which
are volumetrically averaged outcomes of the simulated microstructural events, are
of interest. Such quantities include, (1) the mechanical response, and (2) the
average change in the material, for example damage, . For any microstructural
combination, volume fractions, phase contrasts, etc..., the samples must be tested
and enlarged, holding the volume fraction constant, but increasing the number of
particles, for example from 2, 4, ... etc ...until the macroscopic results stabilize.
Over the course of such tests the finiteB element
meshes were repeatedly refined,
B
and a mesh density of approximately
trilinear hexahedra (approximately
between 2200-3000 DOF for the vector-valued balance of momentum) per particle
was found to deliver mesh independent results. Therefore, for example for the 10
particle test, 24000 DOF for the balance of momentum equation, for 20 particles,
46875 DOF, etc. During the computations, a "2/5" Gauss rule was used, B whereby
B
elements containing material discontinuities had increased Gauss rules (
)
to enhance the resolution of the internal
geometry,
while
elements
with
no
material
B
B
discontinuities had the nominal
rule. The numerical resolution of the
microstructure is shown in Figure 9.1, for a topological exponent of ' 5 . Samples
containing approximately 20 particles were used, although strictly speaking they
were far too small to be statistically representative. However, they gave reasonably
stable macroscopic results for the purposes of these numerical experiments. A more
detailed and rigorous analysis of size effects for such systems is beyond the scope
of this presentation. The reader is referred to the following series of works Huet
[78]-[84], Huet et. al [85], Guidoum and Navi [62], Amieur et. al [5], Guidoum
[63], Amieur [6] Hazanov and Huet [68], Hazanov and Amieur [69], Amieur et.
al. [7] and Huet [89]-[91], as well as some recent work of the author (Zohdi [208]
and Zohdi [220]). In particular, for related work in the optimization of the material
microstructure, controlling parameters such as ' , see Zohdi [220].
A time step size of one second was the starting value. The algorithmic staggering
tolerance was set to %
for the normalized/global control. The designated
4
maximum number of internal iterations, , was set to five. However, during the
simulations, this level of coarse time step discretization was not encountered. The

8
<
<

NUMERICAL EXPERIMENTS

183
starting time step size was

second. In order to smoothly refine and unrefine
the timesteps, the time step contractions and enlargements were bounded in the range
. The total simulation time was set to 60 seconds. The time step
limit size was set to 1 second. This was done in order to illustrate that the approach
can be used when there is a maximum step size limit, for example set for reasons
of time discretization (truncation) error control. No lower time step size was set.
The results are depicted in Figures 9.2-9.7 for strong, moderate and weak damage
and plasticity rates (Table 9.1). As the figures indicate, initially the time step size
needed for the coupling error control was larger than that needed for the discretization
error control. For the weakest damage and plasticity rates (Figures 9.6-9.7), the time
discretization error step size limit was adequate to control the staggering error as
well. However, when significant inelastic effects become stronger (Figures 9.2-9.5),
the time steps required for staggering error control are smaller than those needed for
time step discretization error control.
Remark: An interesting outcome of the simulations depicted, was the development of anisotropic responses due to slightly imperfect random distribution of the
particles. Clearly, for more statitically representative, isotropic, results, even larger
samples would be needed.

450
0.9
400
0.8
350
0.7
300
(MPa)
0.6
<>
<>

0.5
0.4
250
200
150
0.3
100
0.2
50
0.1
11-COMP
22-COMP
33-COMP
0
0
10
20
30
40
TIME (SEC)
50
60
70
10
20
30
40
TIME (SEC)
50
Fig. 9.2 For the strongest damage and plasticity rates (Table 9.1). LEFT: The volumetric
average of the degradation versus time. RIGHT: The volumetric average of the normal stresses
versus time.
60
70
184
INELASTIC RESPONSES
100
90
0.95
0.9
12-COMP
23-COMP
32-COMP
70
60
0.85
t (SEC)
<>
(MPa)
80
50
40
0.8
0.75
0.7
0.65
30
0.6
20
0.55
10
0.5
0
10
20
30
40
TIME (SEC)
50
60
70
10
20
30
40
TIME (SEC)
50
60
70
60
70
450
0.9
400
0.8
350
0.7
300
(MPa)
0.6
<>
<>
Fig. 9.3 For the strongest damage and plasticity rates (Table 9.1). LEFT: The volumetric
average of the shear stresses versus time. RIGHT: The adapted time step sizes versus time.
0.5
0.4
250
200
150
0.3
100
0.2
50
0.1
11-COMP
22-COMP
33-COMP
0
0
10
20
30
40
TIME (SEC)
50
60
70
10
20
30
40
TIME (SEC)
50
Fig. 9.4 For the middle damage and plasticity rates (Table 9.1). LEFT: The volumetric
versus time.
185
MULTIFIELD EXTENSIONS
110
100
0.95
0.9
12-COMP
23-COMP
32-COMP
80
70
t (SEC)
<>
(MPa)
90
60
50
40
0.85
0.8
0.75
30
0.7
20
10
0.65
0
10
20
30
40
TIME (SEC)
50
60
70
10
20
30
40
TIME (SEC)
50
60
70
60
70
Fig. 9.5 For the middle damage and plasticity rates (Table 9.1). LEFT: The volumetric
average of the shear stresses versus time. RIGHT: The adapted time step sizes versus time.
600
0.9
500
0.6
0.5
(MPa)
<>
0.7
400
<>
0.8
300
11-COMP
22-COMP
33-COMP
200
0.4
100
0.3
0.2
0
0
10
20
30
40
TIME (SEC)
50
60
70
10
20
30
40
TIME (SEC)
50
Fig. 9.6 For the weakest damage and plasticity rates (Table 9.1). LEFT: The volumetric
versus time.
9.6 MULTIFIELD EXTENSIONS

Finally, we illustrate how more field equations can be included in a staggering process.
As a model problem, we choose the energy equation,

- ;
-

(9.27)
186
INELASTIC RESPONSES
140
1.015
120
1.01
12-COMP
23-COMP
32-COMP
80
1.005
t (SEC)
<>
(MPa)
100
60
1
0.995
40
0.99
20
0
0.985
0
10
20
30
40
TIME (SEC)
50
60
70
10
20
30
40
TIME (SEC)
50
Fig. 9.7 For the weakest damage and plasticity rates (Table 9.1). LEFT: The volumetric
average of the shear stresses versus time. RIGHT: The adapted time step sizes versus time (no
adaptation was necessary).
which was introduced earlier. We now procede to explicitly identify each of the
terms in Equation 9.27.
9.6.1
The infinitesimal strain case
<
We first discuss infinitesimal strains and then a finite strain analog. A simple classical

choice for the stored energy is
,. -0/

, where
denotes the infinitesimal strain,
the inelastic (eigen-) strains, such as thermal or
plastic strains, and denotes the heat capacity per unit mass. To a first approximation
,
the thermo-mechanical constitutive relationship can be written as

, where is the Cauchy stress,
is the coefficient (matrix in the
anisotropic case) of thermal expansion, and is the reference temperature at which
no thermal strains occur. The symbol is used for the sum of all inelastic strains.
With
, the following balance of energy can be written

<

(9.28)

where the approximation that "

, partially justified since the strains are assumed
to be infinitesimal.4 We describe the effects of multiple eigenstrains by a slight

modification of the previous constitutive law, namely
, where

4 This
J

and
H .
0 1 2 4

assumption will be removed in the ensuing fully finite deformation analysis.
60
70
9.6.2
187
A geometrically-nonlinear extension

As before, a relatively straightforward extension to the infinitesimal deformation constitutive laws is to employ the Eulerian (or Almansi) strain tensor

and
$ , and
. In light of the previously mentioned deformation ranges under consideration, namely small elastic deofrmations and moderate
inelastic strains, we make the following simple constitutive assumption
H

Therefore, since

-
-$

-

;
-

- ; 8
and thus
(9.29)
, we have

-$ ; ; ;

0
1 2
4

- 0 |
- 31 2
- ; ~ 4 8
<
{
{
(9.30)
Finally, the production of energy at a point is phenomenologically modeled as being

, -0/ ) , where )
related to the change in the materials properties, )
is a spatially variable material constant. The parameter ) is negative for exothermic
reactions and positive for endothermic reactions.
9.6.3
Multifield staggering
Employing weak formulations, algorithmically the staggering scheme is as follows:
188
INELASTIC RESPONSES

!#"%$
&('*)+-,/.0 21 #
!3435!1.!% "
&('*)+67 1 8 9(.! #:./;

<
= ">
==
= ?
= " @ >
>
= =F

% <
= "> %
%
= ? %
= " % @ % = F

?BA
= " A % @ > %
F
D ?
==
C
C
F F D
F F

?
KML 8v
= $ /, A
=
x
= I "

C

KML >O
= P
=
>
= I "N

=

" {
= <
=

]/_d`7ab/aOcQb*f-W
= F [
= "$

(
= l m v
= v l m
= z F (
= {
=
=
= z "

= =
= = =
F (
= {

z (
= {
<

z

v }zO

= z (
(

[

=

= {
=

|
=
(
<A
x <
;
= s A
= 4x
= ~ z

=
| A = 4x
= < = A = 2x
=

(

d

~ z

=
|

l m <P
= s
= d A
= 8x
= ! z
(

~

=
=

<

=

|
=
=
8x
A = 8x =
A

(

= {{
=

;
s

; ~ z
&('*)+MOQ9 2RS;3# 54J
.T
WXZY
= P
= F [ Y
= "\
^ ]7_O`MaJb/adceb*fhg8WjiJY
= g F [ Y
= "k
5U V I P
= J

(
= l m g
=jo $ s l m P
= l m P
= u v l m P
= 2x
= y z

= P
= v g z (

=}| P <P v g z
(
F
F F~
t
=
{

=t{

= vg z
= v g zO"$

(

( [
=

=
F

KML vO
= $ /, A
=
GH x
= I "
C

&('*)+-M3 H 3#539 1 ;3# 2
F F P
= P
= F F
=
F F
=
= F F
=
Y

0
F F P
= F F
= u;
F F
= F F
= u

&('*)+754( 3# 4;;
H #" 1-
&('*)+!-754( 3# 4
H 3#
n %7-"% .0 2(9R( #h

KML >O
= P
= 0
GH >
= I "N

&('*)+
9-M3n. H 1 / ;

{
&('*)
+ ;-MOQ9 2 1 #:.!

5U V I
= W

=
= F [
= "
(9.31)
189
Here and
are spaces of admissible trial functions, while + and
are
spaces of admissible test functions. For most loading cases and data these spaces
and . In an abstract setting, one can consider the following
will correspond to
active set solution strategy where the underlined variable is momentarily active, the
corresponding field equations solved, the active variable updated, and the process
repeated for the next field equation:
S7
7

7

7

A A

A A
A A
7
7
A
A
4-j
d
s
^

A A A s
7
A A A A A

A

d
A
- 8
-} -
4j
M

As before, we first define the normalized errors, within each time step (

,

7
}

(9.32)
(9.33)
and their corresponding violation ratios
, -0/
:

, -0/ :
%
(9.34)
Here, a general formulation has been written for a more general set of variables,
, -0/
which will
the
violation $
be
introduced later. One then determines
maximum

"
and a minimum scaling factor 7
from

where

, -0/
u $ C = z
@
@ = =
C
@ $

(9.35)
. Thereafter, the following criteria for temporal adaptivity is adopted
190
INELASTIC RESPONSES
M %$
,>-"!#

(
%<O
%<O

%P
! &N 'T*) -+-/.
01-32/,N4O7!#DA 019". #
015>3?!#7,7!@2,!A. #
019 7!#D.
,>-"?!#5 5

0F-59"L3ME.
: :
#
) *C C
HB
# ,&EH$ 01G$ G /58 A 7!#;019

01-3238,5467!#80F-59D. #
013/!#7,7!@2,!A. #
-01GG/N87!#80F-59
The process proceeds as before.
! &N
#
) HB *C C
) -+,5 ?.
z
z
: :
A
&
(9.36)
10
Closing comments
It is deceptive to call a single uniaxial test of a heterogeneous material sample an

evaluation of material properties. Such criticisms are not a recent development, as
illustrated by this somewhat classical quote:
The tensile test is very easily and quickly performed but it is not possible to do
much with the results, because one does not know what they really mean. They are
the outcome of a number of very complicated physical processes... The extension of a
piece of metal is in a sense more complicated that the working of a pocket watch and
hope to derive information about its mechanism from two or three data derived from
measurement during the tensile test is perhaps as optimistic as would be an attempt
to learn about the working of a pocket watch by determining its compressive strength.
E. Orowan (1944) [140].
It is frequently asked of a theoretical or computational analyst, "Have the theoretical and computational results been compared to experiments?" Computational and
theoretical micro-macro mechanics open the reverse question: "Have the experimental results been compared to computations and theory?" At the risk of closing on a
controversial note, it is not unfair to think that computational methods have perhaps
become more accurate than experiments. Minimally, no one can argue that computations cannot now play a strong complementary role to laboratory experiments.
Furthermore, the continual rise in computational power will only push research more
towards reliance on numerical simulations employing non-phenomenological micromacro models, which are less expensive than laboratory tests. For successful modern
191
192
CLOSING COMMENTS
research, computations, modeling and experiments should be symbiotic and interactive.

In closing it must be remarked that the general areas of micro-macro material
theory and computational mechanics are two that are changing with such rapidity,
that it is questionable whether it is sensible to write any summaries, monographs or
books on these subjects, for they will certainly be outdated, at best, within no more
than five years. On a humorous note, perhaps they need an expiration date, like a
bottle of milk. Stated in a far more elegant way: Science does not last longer than
fish (Alfred North Whitehead, Logician and Philosopher (1861-1947)).
11
References
1. Aboudi, J. (1992). Mechanics of composite materials-a unified micromechanical approach. Elsevier, 29.
2. Ainsworth, M. and Oden, J. T., (2000). A posterori error estimation in finite
element analysis. John-Wiley.
3. Ames, W.F.(1965). Nonlinear Partial differential equations in engineering.
Academic Press.
4. Ames, W.F.(1977). Numerical methods for partial differential equations. 2nd
edition. Academic Press.
5. Amieur, M., Hazanov, S. and Huet, C. (1993). Numerical and experimental
study of size and boundary conditions effects on the apparent properties of specimens not having the representative volume. In Micromechanics of Concrete
and Cementitious Composite, C. Huet, editor.
6. Amieur, M. (1994). Etude numerique et experimentale des effets dechelle et
de conditions aux limites sur des e prouvettes de beton nayant pas le volume
representatif. Doctoral dissertation No 1256, Ecole Polytechnique Federale de
Lausanne, Lausanne, Switzerland.
7. Amieur, M., Hazanov, S. and Huet, C. (1995). Numerical and experimental assessment of the size and boundary conditions effects for the overall properties of
granular composite bodies smaller than the representative volume. In IUTAM
193
194
REFERENCES
Symposium on Anisotropy, Inhomogeneity and Nonlinearity in Solid Mechanics, D.F. Parker and A.H. England, editors. Kluwer Academic Publishers, The
Netherlands. 149-154.
8. Axelsson, O.(1994). Iterative solution methods. Cambridge University Press.
9. Babuska, I., Andersson, B., Smith, P. J., Levin, K. (1999) Hierarchical modeling of heterogeneous solids. Damage analysis of fiber composites Part I:
Statistical analysis on fiber scale. Computer Methods in Applied Mechanics
and Engineering. 172. 27-77.
10. Bathe, K. J. (1996). Finite element procedures. Prentice-Hall.
11. Becker, E.B., Carey, G.F., and Oden J. T. (1980) Finite elements: an introduction Prentice-Hall.
12. V.Belsky, M.W. Beall, J. Fish, M.S. Shephard, S. Gomaa, "Computer-Aided
Multiscale Modeling Tools for Composite Materials and Structures,"International
Journal of Computing Systems in Engineering, Vol. 6, No. 3, pp. 213-223,
(1995).
13. Bonet, J. and Wood, R. D. (1997) Nonlinear continuum mechanics for finite
element analysis. Cambridge University Press.
14. Briggs, W. (1987). A multigrid tutorial. Siam Issue.
15. Budiansky, B. (1965). On the elastic moduli of some heterogeneous materials.
Journal of the Mechanics and Physics of Solids. 13, 223-227.
16. Carol, I., Rizzi, E. and Willam, K. (1994) A unified theory of elastic degradation
and damage based on a loading surface. The International Journal of Solids
and Structures. 31, No. 20, 2835-2865.
17. Courant, R. (1943) Variational methods for the solution of problems of equilibrium and vibrations. Bull. Amer. Math Soc. 49, 1-23.
18. Carey, G.F., and Oden J. T. (1983) Finite elements: a second course. Prenticehall.
19. Casey, J. (1985). Approximate kinematical relations in plasticity. The International Journal of Solids and Structures. 21, 671-682.
20. Chandrasekharaiah, D. S. and Debnath, L. (1994) Continuum mechanics. Academic press.
21. Chen, W and Fish, J (2001) A Dispersive Model For Wave Propagation in Periodic Heterogeneous Media Based on Homogenization With Multiple Spatial
and Temporal Scales. Journal of Applied Mechanics, Vol. 68, No. 2, 153-161.
195
22. Cherkaev, A. V. and Gibiansky, L. V. (1992). The exact coupled bounds for
effective tensors of electrical and magnetic properties of two-component twodimensional composites. Proc. of Royal Soc. of Edinburgh. 122A, 93-125.
23. Christensen, R. (1990). A critical evaluation for a class of micromechanics
models Journal of the Mechanics and Physics of Solids. 38, No. 3, 379-404.
24. Ciarlet, P. G. (1993). Mathematical Elasticity. Elsevier.
25. Courant, R. (1943). Variational methods for the solution of problems of equilibrium and vibration. Bull. Am. Math. Soc. 49, 1-23.
26. Davidon, W. C.. (1959). Variable metric method for minimization. Research
and development report. ANL-5990 (Ref.). U. S. Atomic Energy Commision,
Argonne National Laboratories.
27. Davis, L. (1991). Handbook of Genetic Algorithms. Thompson Computer
Press.
28. Doltsinis, I. St. (1993). Coupled field problems-solution techniques for sequential and parallel processing. In Solving large-scale problems in mechanics. M
. Papadrakakis, editor.
29. Eshelby, J. D. (1957). The elastic field of an ellipsoidal inclusion, and related
problems. Proc. Roy. Soc. A241, 376-396.
30. Eustathopoulos, N. and Mortensen, A. (1993). Capillary phenomena, interfacial bonding and reactivity. In Fundamentals of metal matrix composites.
Suresh, S., Mortensen, A. and Needleman, A. editors. Butterworth-Heineman
publishers.
31. Fiacco, A. V. and McCormick G. P, (1990). Nonlinear Programming: Sequential Unconstrained Minimization Techniques. Siam reissue. John Wiley and
Sons, New York and Toronto.
32. Fish, J., and Wagiman, A. (1993) Multiscale finite element method for a locally
nonperiodic heterogeneous medium. Comput. Mech. 12. 164-180.
33. Fish, J., Pandheeradi, M. and Belsky, V. (1995), An efficient multilevel solution
scheme for large scale nonlinear systems. International Journal for Numerical
Methods in Engineering. Vol. 38, 1597-1610.
34. Fish, J., and Belsky, V., (1995) Multigrid method for periodic heterogeneous
media Part I: Convergence studies for one dimensional case,Computer Methods
in Applied Mechanics and Engineering. 126 pp. 1-16
35. Fish, J., and Belsky, V., (1995) Multigrid method for periodic heterogeneous
media Part II: Multiscale modeling and quality control in multidimensional
case, Computer Methods in Applied Mechanics and Engineering. 126 pp. 1738
196
REFERENCES
36. Fish, J. and Belsky, V. (1997) Generalized Aggregation Multilevel Solver. International Journal for Numerical Methods in Engineering. Vol. 40, 43414361.
37. Fish, J., Shek, K., Pandheeradi, M., Shephard, M. S. (1997) Computational
plasticity for composite structures based on mathematical homogenization:
Theory and practice Computer Methods in Applied Mechanics and Engineering. 148. (1-2), 53-73.
38. Fish, J., Yu, Q. and Shek, K. L. (1999) Computational Damage Mechanics for
Composite Materials Based on Mathematical Homogenization. International
Journal for Numerical Methods in Engineering. Vol. 45, pp. 1657-1679.
39. Fish, J. and Shek, K. (1999). Finite deformation plasticity for composite structures: Computational models and adaptive strategies Computer Methods in
Applied Mechanics and Engineering. 172. 145-174.
40. Fish, J. and Ghouli, A. (2001) Multiscale Analytical Sensitivity Analysis for
Composite Materials. International Journal for Numerical Methods in Engineering. Vol. 50, 1501-1520.
41. Fish, J. and Yu, Q. (2001) Multiscale Damage Modeling for Composite Materials: Theory and Computational Framework. International Journal for Numerical Methods in Engineering. Vol. 52, No 1-2, 161-192.
42. Fish, J. and Chen, W. (2001) Uniformly Valid Multiple Spatial-Temporal Scale
Modeling for Wave Propagation in Heterogeneous Media. Mechanics of Composite Materials and Structures Vol. 8, 81-99.
43. Fu, Y., Klimkowski, K., Rodin, G. J., Berger, E., Browne, J. C., Singer, J. K.,
Van de Geijn, R. A. and Vemaganti, K. (1998). Fast solution method for threedimensional many-particle problems of linear elasticity. The International
Journal of Numerical Methods in Engineering. 42, 1215-1229.
44. Fletcher, R. and Powell, M. J. D. , (1963). A rapidly convergent descent method
for minimization. Computer Journal 7, 149-154
45. Ghosh, S., and Mukhopadhyay, S. N., (1993) A material based finite element
analysis of heterogeneous media involving Dirichlet tessellations, Computer
Methods in Applied Mechanics and Engineering, Vol. 104., 211-247.
46. Ghosh, S. and Moorthy, S. (1995) Elastic-plastic analysis of arbitrary heterogeneous materials with the Voronoi Cell finite element method. Computer
Methods in Applied Mechanics and Engineering. 121 (1-4), 373-409.
47. Ghosh, S., Kyunghoon, L., Moorthy, S. (1996) Two scale analysis of heterogeneous elastic-plastic materials with asymptotic homogenization and Voronoi
cell finite element model. Computer Methods in Applied Mechanics and Engineering. 132. (1-2), 63-116
197
48. Ghosh, S. and Moorthy, S. (1998) Particle fracture simulation in non-uniform

microstructures of metal-matrix composites. Acta mater.46, No. 3, 965-982.
49. Ghosh, S., Lee, K., Raghavan, P. (2001) A multi-level computational model for
multi-scale damage analysis in composite and porous materials. International
Journal of Solids and Structures. 38, 2335-2385.
50. Ghosh, S., Ling, Y., Majumdar, B. and Kim, R. (2001) Interfacial debonding
analysis in multiple fiber reinforced composites. Mechanics of Materials.32,
562-591.
51. Ghosh, S., Kyunghoon, L., Moorthy, S. (1996) Two scale analysis of heterogeneous elastic-plastic materials with asymptotic homogenization and Voronoi
cell finite element model. Computer Methods in Applied Mechanics and Engineering. 132. (1-2), 63-116
52. Gill, P. Murray, W. and Wright, M. (1995). Practical optimization. Academic
Press.
53. Golanski, D., Terada, K., and Kikuchi, N. (1997) Macro and Micro Scale
Modeling of Thermal Residual Stresses in Metal Matrix Composite Surface
Layers by the Homogenization Method, Computational Mechanics. Vol. 19No. 3, pp.188-202.
54. Goldberg, D. E. (1989). Genetic algorithms in search, optimization and machine learning. Addison-Wesley.
55. Goldberg, D. E. and Deb, K. (2000) Special issue on Genetic Algorithms.
Computer Methods in Applied Mechanics and Engineering. 186 (2-4) 121124.
56. Gonzlez, C. and LLorca, J. (2002). An Analysis of the Effect of Hydrostatic Pressure on the Tensile Deformation of Aluminum-Matrix Composites.
Materials Science and Engineering A. (2002).
57. Gonzlez, C. and LLorca, J. (2001). Micromechanical Modelling of Deformation and Failure in Ti-6Al-4V / SiC Composites. Acta Materialia, 49, 35053519.
58. Gonzlez, C. and LLorca, J. (2000). A Self-Consistent Approach to the ElastoPlastic Behaviour of Two-Phase Materials Including Damage. Journal of the
Mechanics and Physics of Solids, 48, 675-692.
59. Green, A. E. and Naghdi, P. M. (1965) A general theory of an elasto-plastic
continuum. Arch. Rat. Mech. Anal. 18, 251-281.
60. Green, A. E. and Naghdi, P. M. (1971) Some remarks on elastoplastic deformation at finite strain. Int. J. Engrg. Sci. 9, 1219-1229.
198
REFERENCES
61. Guedes, J. M., and Kikuchi, N. (1990) Preprocessing and postprocessing for
materials based on the homogenization method with adaptive finite element
methods. Computer Methods in Applied Mechanics and Engineering. 83,
143-198.
62. Guidoum, A. and Navi, P. (1993). Numerical simulation of thermo-mechanical
behaviour of concrete through a 3-D granular cohesive model. In C. Huet
(Ed.) Micromechanics of Concrete and Cementitious Composites. Presses
Polytechniques et Universitaires Romandes, Lausanne. 213-228.
63. Guidoum, A. (1994). Simulation numrique 3D des comportements des btons
en tant que composites granulaires. Doctoral Dissertation No 1310. Ecole
Polytechnique de Lausanne, Switzerland.
64. Gurtin, M. (1981). An introduction to continuum mechanics. Academic Press.
65. Hashin, Z. and Shtrikman, S. (1962). On some variational principles in anisotropic
and nonhomogeneous elasticity. Journal of the Mechanics and Physics of
Solids. 10, 335-342.
66. Hashin, Z. and Shtrikman, S. (1963). A variational approach to the theory of
the elastic behaviour of multiphase materials. Journal of the Mechanics and
Physics of Solids. 11, 127-140.
67. Hashin, Z. (1983). Analysis of composite materials: a survey ASME Journal
of Applied Mechanics. 50, 481-505.
68. Hazanov, S. and Huet, C. (1994). Order relationships for boundary conditions
effect in heterogeneous bodies smaller than the representative volume. Journal
of the Mechanics and Physics of Solids. 42, 1995-2011.
69. Hazanov, S. and Amieur, M. (1995). On overall properties of elastic heterogeneous bodies smaller than the representative volume. Int. J. Eng. Science. 33,
9, 1289-1301.
70. Hill, R. (1952). The Elastic Behaviour of a Crystalline Aggregate. Proc. Phys.
Soc. (Lond.) A65, 349-354.
71. Hill, R. (1965). A self consistent mechanics of composite materials, Journal
of the Mechanics and Physics of Solids. 13, 213-222
72. Hill, R. and Rice, J. R.(1973). Elastic potentials and the structure of inelastic
constitutive laws. SIAM J. Appl. Math. 25, 448-461.
73. Hill, R. (1963). New derivations of some elastic extremum principles. In
Progress in Applied Mechanics, Prager Anniversary Volume New York. 91106.
74. Hill, R. (1963). Elastic Properties of reinforced solids: some theoretical principles Journal of the Mechanics and Physics of Solids. 11. 357-372.
199
75. Holland, J. H. (1975). Adaptation in natural and artificial systems. Ann Arbor,
Mich. University of Michigan Press.
76. Horst, R. and Tuy, H. (1996). Global optimization-deterministic approaches.
Springer Verlag.
77. Hsiao, S.W., and Kikuchi, N., Numerical analysis of deep drawing process
for thermoplastic composite laminates, Journal of Engineering Materials and
Technology, ASME Transactions, Vol. 119, pp.314-318, July 1997.
78. Huet, C. (1981). Remarques sur lassimilation dun materiau heterogene a
un milieu continu e quivalent. In C. Huet and A. Zaoui, editors.Rheological
behaviour and Structure of Materials. Presses ENPC, Paris. 231-245.
79. Huet, C. (1982). Universal conditions for assimilation of a heterogeneous
material to an effective medium. Mechanics Research Communications. 9 (3),
165-170.
80. Huet, C. (1984). On the definition and experimental determination of effective constitutive equations for heterogeneous materials. Mechanics Research
Communications. 11 (3), 195-200.
81. Huet, C. (1987) Thermo-hygro-mechanical couplings in wood technology and
rheological behaviours. In Bui, H. D. and Nguyen, Q. S. (Eds.), Thermomechanical couplings in solids, IUTAM symposium, Paris, 1986. North-Holland.
pp 163-182.
82. Huet, C. (1988) Modelizing the kinetics of the thermo-hygro-viscoelastic behavior of wood under constant climatic conditions. In Itani, R. (Ed.). Proceedings of the 1988 International Conference on Timber Engineering, Seattle. pp.
395-401
83. Huet, C. (1990). Application of variational concepts to size effects in elastic
heterogeneous bodies. Journal of the Mechanics and Physics of Solids. 38,
813-841.
84. Huet, C. (1991). Hierarchies and bounds for size effects in heterogeneous
bodies. In Continuum Models and discrete systems, G. A. Maugin, editor. Vol.
2, 127-134.
85. Huet, C., P. Navi, P. E. Roelfstra (1991). A homogenization technique based
on Hills modification theorem. In Continuum models and discrete systems.
86. Huet, C. (1995) Hybrid continuum thermodynamics framework and numerical
simulations examples for the delayed micromechanical behavior of heterogeneous materials with chemical, climatic and defect sensitivity. In Nguyen, Q.
S. and Nguyen, V. D. (Eds.). Engineering mechanics today. Proc. Int. Conf.
Hanoi, University of Hanoi, pp. 170-184.
200
REFERENCES
87. Huet, C., Guidoum, A. and Navi, P. (1995) A 3D micromechanical model for
numerical analysis and prediction of long term deterioration in concrete. In.
Sakai, K. Banthia, N. Gjorv, O. E. (Eds.), Concrete under severe conditions:
environment and loading. Vol 2, Spon, London, pp. 1458-1467.
88. Huet, C. (1996) Recent advances in the long term deformation and deterioration behavior of structural materials and components through the integrated micromechanics and thermodynamics of solids approach. In Advances in building
materials. Ed. A. Gerdes. Aedificatio, Freiburg. pp. 161-196.
89. Huet, C. (1997). Activities 1989-1996. Laboratory for Building Materials. P.
Navi and A. Tolou editors. Department of Materials Report.
90. Huet C. (1997). An integrated micromechanics and statistical continuum thermodynamics approach for studying the fracture behaviour of microcracked heterogeneous materials with delayed response. Engineering Fracture Mechanics,
Special Issue. 58, 5-6, 459-556.
91. Huet, C. (1999). Coupled size and boundary condition effects in viscoelastic
heterogeneous bodies. Mechanics of Materials. 31, 12, 787-829.
92. Hughes, T. J. R. (1989). The finite element method. Prentice Hall.
93. Hyun, S. and Torquato, S. (2001). Designing material microstructures with
target properties. Journal of Materials Research. 16, 280.
94. Jikov, V. V., Kozlov, S. M., Olenik, O. A.(1994). Homogenization of differential
operators and integral functionals. Springer-Verlag.
95. Kachanov, L. M. (1986). Introduction to continuum damage mechanics. Martinus Nijoff, Dordricht.
96. Kee, A., Matic, P. and Everett, R. K. (1998). A mesoscale computer simulation
of multiaxial yield in gasar porous copper. Mater. Sci. Eng. A249, 30-39.
97. Kikuchi, N., Nishiwaki, S., Fonseca, J.O., and Silva, E.C.N., Design Optimization Method for Compliant Mechanisms and Material Microstructure,
Computer Methods in Applied Mechanics and Engineering, Vol. 151, 1998,
pp.401-417.
98. Kitchen, J. (1966). Concerning the convergence of iterates to fixed points.
Studia Math. 27, 247-249.
99. Krajcinovic, D. (1996). Damage Mechanics. North Holland.
100. Kroner, E. (1972). Statistical Continuum Mechanics. CISM Lecture Notes.
92, Springer-Verlag.
101. Ladeveze, P. and Leguillon, D. (1983). Error Estimate Procedure in the Finite
Element Method and Applications. SIAM J. Numerical Anal. 20, 485-509.
201
102. Lee, K., Moorthy, S., Ghosh, S. (1999). Multiple scale computational model
for damage in composite materials, Effective properties of composite materials
with periodic microstructure: a computational approach Computer Methods in
Applied Mechanics and Engineering. 172. 175-201.
103. Lemaitre, J.(1985). Coupled elasto-plasticity and damage constitutive equations. Computer Methods in Applied Mechanics and Engineering. 51, 31-49.
104. Lemaitre, J. and Chaboche, J.-L.(1990). Mechanics of solid materials. Cambridge Univ. Press, Cambridge.
105. Le Tallec, P. (1994). Domain Decomposition Methods in Computational Mechanics. Computational Mechanics Advances. 1, 121-220.
106. Le Tallec, P. and Mouro, J. (2001) Fluid structure interaction with large structural displacements. Computer Methods in Applied Mechanics and Engineering 190 (24-25) 3039-3067.
107. Lewis, R. W., Schrefler, B. A. and Simoni, L (1992). Coupling versus uncoupling in soil consolidation. Int. J. Num. Anal. Metho. Geomech. 15,
533-548.
108. Lewis, R. W. and Schrefler, B. A. (1998). The finite element method in the static
and dynamic deformation and consolidation of porous media. 2nd edition.
Wiley press.
109. Li, M., Ghosh, S. and Richmond, O. (1999) An experimental-computational
approach to the investigation of damage evolution in discontinuously reinforced
aluminum matrix composite. Acta mater.47, No. 12, 3515-3532.
110. Liu, H. W. (1970). Stress corrosion cracking and the interaction between cracktip stress-field and solute atoms. J. Basic Engr. ASME. 633-638.
111. LLorca, J. (2000) Void Formation in Metal Matrix Composites. Comprehensive
Composite Materials. Vol. 3, Metal Matrix Composites. (T. W. Clyne, Ed.),
Pergamon, Amsterdam, pp. 91-115, 2000.
112. LLorca, J. (2000) J. LLorca, M. Elices, and Y. Termonia. Elastic Properties
of Sphere-Reinforced Composites with a Mesophase. Acta Materialia, 48,
4589-4597.
113. LLorca, J. and Gonzlez, C. (1998). Microstructural Factors Controlling the
Strength and Ductility of Particle-Reinforced Metal-Matrix Composites. Journal of the Mechanics and Physics of Solids, 46, 1-28.
114. LLorca, J. (1994). A Numerical Stydy of the Mechanisms of Cyclic Strain
Hardening in Metal-Ceramic Composites. Acta Metallurgica et Materialia, 42,
pp. 151-162.
202
REFERENCES
115. LLorca, J. (1991). J. LLorca, A. Needleman and S. Suresh. An Analysis of the

Effects of Matrix Void Growth on Deformation and Ductility in Metal-Ceramic
Composites. Acta Metallurgica et Materialia, 39, 2317-2335.
116. Lubliner, J. (1990). Plasticity theory. Macmillan Publishers.
117. Luenberger, D. (1974). Introduction to Linear and Nonlinear Programming
Addison-Wesley, Menlo Park.
118. Malvern, L. (1968). Introduction to the mechanics of a continuous medium.
Prentice Hall.
119. Marsden, J. E. and Hughes, T. J. R. (1983). Mathematical foundations of
elasticity. Prentice Hall.
120. Maxwell, J. C.(1867). On the dynamical theory of gases. Philos. Trans. Soc.
London. 157, 49.
121. Maxwell, J. C. (1873). A treatise on electricity and magnetism. 3rd. Ed.
Clarendon Press, Oxford.
122. Metals Handbook (1975). Failure Analysis and Prevention. 8, The American
Society for Metals.
123. Michaud, V. (1992). Liquid state processing. In Fundamentals of metal matrix
composites. Suresh, S., Mortensen, A. and Needleman, A., editors.
124. Michel, J. C., Moulinec, H., Suquet, P. (1999) Effective properties of composite
materials with periodic microstructure: a computational approach Computer
Methods in Applied Mechanics and Engineering. 172. 109-143.
125. Moes, N., Oden, J. T., and Zohdi, T. I.(1998). Investigation of the interaction of
numerical error and modeling error in the Homogenized Dirichlet Projection
Method. Computer Methods in Applied Mechanics and Engineering. 159,
79-101.
126. Moorthy, S. and Ghosh, S. (2000) Adaptivity and convergence in the Voronoi
cell finite element model for analyzing heterogeneous materials. Computer
Methods in Applied Mechanics and Engineering. 185, 37-74.
127. Moulinec, H. and Suquet, P. (1998). A numerical method for computing the
overall response of nonlinear composites with complex microstructure. Computer Methods in Applied Mechanics and Engineering. 157. 69-94.
128. Mori, T. and Tanaka, K.(1973) Average stress in matrix and average energy of
materials with misfitting inclusions. Acta. Metall. 21, 571-574.
129. Moorthy, S. and Ghosh, S. (2000) Adaptivity and convergence in the Voronoi
cell finite element model for analyzing heterogeneous materials. Computer
Methods in Applied Mechanics and Engineering. 185, 37-74.
203
130. Mura, T. (1993). Micromechanics of defects in solids, 2nd edition. Kluwer

Academic Publishers.
131. Nemat-Nasser, S. and Hori, M. (1999). Micromechanics: overall properties of
heterogeneous solids. 2nd edition. Elsevier, Amsterdam.
132. Nesterenko, V. F., Meyers, M. A., Chen, H. C., LaSalvia, J. C. (1994). Controlled high rate localized shear in porous reactive media. Applied physics
letters. 65 (24), 3069-3071.
133. Nye, J. (1984). Physical properties of crystals. Oxford science publications.
134. Oden J. T. and Carey, G.F. (1984). Finite elements: mathematical aspects.
Prentice-hall.
135. Oden, J. T. and Zohdi, T. I. (1997). Analysis and adaptive modeling of highly
heterogeneous elastic structures. Computer Methods in Applied Mechanics and
Engineering. 148, 367-391.
136. Oden, J. T. and Vemaganti, K. (1999) Adaptive Hierarchical Modeling of Heterogeneous Structures. Physica D, volume 133, pp. 404-415.
137. Oden, J. T., Vemaganti, K., Moes, N. (1999). Hierarchical modeling of heterogeneous solids. Computer Methods in Applied Mechanics and Engineering.
172. (1-27).
138. Ogden, R. W. (1999). Nonlinear Elasticity. Dover reissue.
139. Onwubiko, C. (2000). Introduction to engineering design optimization. Prentice Hall.
140. Orowan, E. (1944). Proc. Instn. Mech. Engrs. 151, 133 (discussion of paper
by H. ONeill).
141. Ortega, J. and Rockoff, M. (1966). Nonlinear difference equations and GaussSeidel type iterative methods. SIAM J. Numer. Anal. 3, 497-513.
142. Ortiz, M. and Pandolfi. A. (1999). Finite deformation irreversible cohesive
elements for three-dimensional crack-propagation analysis. The International
143. Ostrowski, A. (1957). Les points d attraction et de repulsion pour literation
dans lespace a n dimensions. C. R. Acad. Sci. Paris. 244, 288-289.
144. Ostrowski, A. (1966). Solution of equations and systems of equations. Academic Press, New York.
145. Park, K. C. and Felippa, C. A. (1983). Partitioned analysis of coupled systems.
In Computational methods for transient analysis. T. Belytschko and T. J. R.
Hughes, editors.
204
REFERENCES
146. Perron, O. (1929). Uber

Stabilitat und asyptotisches Verhalten der Lo sungen
eines Systems endlicher Differenzengleichungen. J. Reine Angew. Math. 161,
41-64.
147. Piperno, S. (1997) Explicit/implicit fluid/structure staggered procedures with a
structural predictor and fluid subcycling for 2D inviscid aeroelastic simulations.
Int. J. Num. Meth. Fluids. 25, 1207-1226.
148. Poza, P. and LLorca, J. (1999) Mechanical Behaviour of Al-Li/SiC Composites.
Part III: Micromechanical Modeling. Metallurgical and Materials Transactions
30A, 869-878.
149. Raghavan, P., Moorthy, S., Ghosh, S. and Pagano, N. J. (2001) Revisiting the
composite laminate problem with an adaptive multi-level computational model.
Composites Science and Technology.61, 1017-1040.
150. Rayleigh, J. W. (1892). On the influence of obstacles arranged in rectangular
order upon properties of a medium. Phil. Mag. 32, 481-491.
151. Reuss, A. (1929) Berechnung der Fliessgrenze von Mischkristallen auf Grund
der Plastizitatsbedingung fur Einkristalle. Z. angew. Math. Mech. 9, 49-58.
152. Schrefler, B. A, (1985). A partitioned solution procedure for geothermal reservoir analysis. Comm. Appl. Num. Meth. 1, 53-56.
153. Segurado, J. and LLorca, J. (2002). A Numerical Approximation to the Elastic
Properties of Sphere-Reinforced Composites. Journal of the Mechanics and
Physics of Solids, 50.
154. Segurado, J., LLorca, J., and Gonzlez, C. (2002). On the Accuracy of Meanfield Approaches to Simulate the Plastic Deformation of Composites. Scripta
Materialia, 46, 525-529.
155. Silva, E.C.N., Fonseca, J.S.O., and Kikuchi, N. (1997) Optimal design of piezoelectric microstructures. Computational Mechanics. Vol 19, 397-410.
156. Silva, E.C.N., Fonseca, J.S.O., and Kikuchi, N. (1998) Optimal design of periodic piezocomposites. Computer Methods in Applied Mechanics and Engineering, 159, 49-77.
157. Sigmund, O. (2000). Topology Optimization: A Tool for the Tailoring of
Structures and Materials. A special issue of the Philosophical Transactions of
the Royal Society: Science into the next Millennium (Issue III, Mathematics,
Physics and Engineering). 358(1765), pp. 211-228.
158. Sigmund, O. (2000). A new Class of Extremal Composites. Journal of the
Mechanics and Physics of Solids. 48(2), pp. 397-428.
159. Sigmund, O. and Torquato, S. (1999). Design of smart composite materials
using topology optimization. Smart Materials and Structures. 8, pp. 365-379.
205
160. Sigmund, O. and Torquato, S. (1996) Composites with extremal thermal expansion coefficients. Applied Physics Letters. 69(21):3203-3205.
161. Sigmund, O. (1995) Tailoring materials with prescribed elastic properties. Mechanics of Materials. 20:351-368.
162. Sigmund, O. and Torquato, S. (1999) Design of Smart Composite Materials
Using Topology Optimization. Smart Materials and Structures. 8, 365.
163. Sigmund, O, Torquato, S. and Aksay, I. A. (1998). On the optimization of 1-3
piezo-composites using topology optimization. Journal of Materials Research.
14, 1038.
164. Sigmund, O. (2001) Design of multiphysics actuators using topology optimization - Part I: One-material structures. Computer Methods in Applied Mechanics
and Engineering. 190 (49-50), pp. 6577-6604.
165. Sigmund, O. (2001) Design of multiphysics actuators using topology optimization - Part II: Two-material structures. Computer Methods in Applied
Mechanics and Engineering 190 (49-50), pp. 6605-6627.
166. Simo, J. C., Hughes, T. J. R(1998). Computational inelasticity. Springer
Verlag.
167. Simo, J. C. and Ortiz, M. (1985). A unified approach to finite deformation
elastoplastic analysis based on the use of hyperelastic constitutive laws. Computer Methods in Applied Mechanics and Engineering. 49, 221-245.
168. Simo, J. C. (1988). A framework for finite strain elastoplasticity based on maximum plastic dissipation and the multiplicative decomposition: Part 1. Continuum formulation. Computer Methods in Applied Mechanics and Engineering.
66, 199-219.
169. Southwell, R. V. (1940). Relaxation methods in engineering science. Oxford
University Press.
170. Southwell, R. V. (1946). Relaxation methods in theoretical physics. Oxford
University Press.
171. Suresh, S., Mortensen, A. and Needleman, A.(1993). Fundamentals of metal
matrix composites. Butterworth-Heineman publishers.
172. Szabo, B. and Babuska, I.(1991). Finite element analysis. Wiley Interscience.
173. Terada, K., and Kikuchi, N. (1996) Microstructural Design of Composites
Using the Homogenization Method and Digital Images. Materials Science
Research International. Vol.2-No. 2, 65-72.
206
REFERENCES
174. Terada, K., and Kikuchi, N. (1996) Global-Local Constitutive Modeling of

Composite Materials by the Homogenization Method. Materials Science Research International, Vol.2-No. 2, 73-80.
175. Torquato, S. (1991). Random heterogeneous media: microstructure and improved bounds on effective properties. Appl. Mech. Rev. 44, 37-76.
176. Torquato, S. (1997). Effective stiffness tensor of composite media I. Exact
series expansions. Journal of the Mechanics and Physics of Solids. 45, 14211448.
177. Torquato, S. (1998). Effective stiffness tensor of composite media II. Applications to isotropic dispersions. Journal of the Mechanics and Physics of Solids.
46, 1411-1440.
178. Torquato, S. (2002). Random Heterogeneous Materials: Microstructure and
Macroscopic Properties Springer-Verlag, New York.
179. Torquato, S. and Hyun, S. (2001). Effective-medium approximation for composite media: Realizable single-scale dispersions. J. Appl. Phys., 89, 17251729.
180. Truesdell, C. and Noll, W.(1992). The nonlinear field theories. Springer Verlag.
181. Van Vlack, L. H. (1966). Elements of materials science. Addison-Wesley.
182. Vemaganti, K. S. and Oden, J. T. (2001) Estimation of Local Modeling Error
and Goal-Oriented Adaptive Modeling of Heterogeneous Materials. Computer
Methods in Applied Mechanics and Engineering. 190, 46-47, 6089 6124.
183. Voigt, W. (1889). Uber

die Beziehung zwischen den beiden Elastizitatskonstanten
isotroper Korper. Wied. Ann. 38, 573-587.
184. Wentorf, R., Collar, R., Shephard, M. S. and Fish, J. (1999) Automated Modeling for Complex Woven Mesostructures. Comp. Meth. Appl. Mech. Engng.,
Vol. 172, 273-291.
185. Wriggers, P. (2001). Nichtlineare Finite-Element-Methoden. Springer-Verlag.
186. Wriggers, P., Zavarise, G. and Zohdi, T. I. (1998). A computational study of
interfacial debonding damage in fibrous composite materials. Computational
Materials Science. 12 , 39-56.
187. Young. D. M. (1950). Iterative methods for solving partial difference equations
of elliptic type. Doctoral thesis. Harvard University.
188. Zhigljavsky, A. A. (1991). Theory of global random search. Kluwer
189. Zienkiewicz, O. C. and Taylor R. L. (1991). The finite element method. Vols.
I and II. McGraw-Hill.
207
190. Zienkiewicz, O. C. (1984). Coupled problems and their numerical solution, in

R. W. Lewis, P. Bettess and E. Hinton, editors. Numerical methods in coupled
systems. Wiley, Chichester. 35-58.
191. Zhang, J., Lin, Z., Wong, A., Kikuchi, N., Li, V.C., Yee, A.F., and Nusholtz,
G.S., Constitutive modeling and material characterization of polymetric foams,
Journal of Engineering Materials and Technology, ASME Transactions, Vol.
119, pp.279-283, July 1997.
192. Vigdergauz, S. (2001) Genetic algorithm perpective to identify energy optimizing inclusions in an elastic plate. The International Journal of Solids and
Structures. 38, 6851-6867.
193. Vigdergauz, S. (2001) The effective properties of a perforated elastic plate:
numerical optimization by genetic algorithm. The International Journal of
Solids and Structures. 38, 8593-8616.
194. Zohdi, T. I., Oden, J. T. and Rodin, G. J. (1996). Hierarchical modeling of
heterogeneous bodies. Computer Methods in Applied Mechanics and Engineering. 138, 273-298.
195. Zohdi, T. I., Feucht, M., Gross, D. and Wriggers, P.(1998). A description of
macroscopic damage via microstructural relaxation. The International Journal
of Numerical Methods in Engineering. 43, 493-507.
196. Zohdi, T. I. and Wriggers, P. (1999). A domain decomposition method for bodies with microstructure based upon material regularization. The International
Journal of Solids and Structures. 36, No. 17, 2507-2526.
197. Zohdi, T. I., Hutter, K., and Wriggers, P. (1999). A technique to describe
the macroscopic pressure dependence of diffusive properties of solid materials
containing heterogeneities. Computational Materials Science. 15, 69-88.
198. Zohdi, T. I. and Wriggers, P. (1999). On the effects of microstress on macroscopic diffusion processes. Acta Mechanica. 136, No 1-2, 91-107.
199. Zohdi, T. I. and Wriggers, P. (2000). A computational model for interfacial
damage through microstructural cohesive zone relaxation. The International
Journal of Fracture. 101 No. 3, L9-L14.
200. Zohdi, T. I. and Wriggers, P. (2000). On the sensitivity of homogenized material
responses at infinitesimal and finite strains. Communications in Numerical
Methods in Engineering. 16. 657-670.
201. Zohdi, T. I. (2000). Overall solution-difference bounds on the effects of material
inhomogeneities. The Journal of Elasticity. 58 (3), 249-255.
202. Zohdi, T. I. and Wriggers, P. (2001). A model for simulating the deterioration
of structural-scale material responses of microheterogeneous solids. Computer
Methods in Applied Mechanics and Engineering. 190, 22-23, 2803 - 2823.
208
REFERENCES
203. Zohdi, T. I. and Wriggers, P. (2001). Aspects of the computational testing

of the mechanical properties of microheterogeneous material samples. The
International Journal of Numerical Methods in Engineering. 50, 2573-2599.
204. Zohdi, T. I. and Wriggers, P. (2001). Modeling and simulation of the decohesion
of particulate aggregates in a binding matrix. Engineering Computations. 18,
1/2, 79-95.
205. Zohdi, T. I. and Wriggers, P. (2001). Computational micro-macro material
testing. Archives of Computational Methods in Engineering. Vol 8, 2, 131228.
206. Zohdi, T. I., Wriggers, P. and Huet, C. (2001). A method of substructuring
large-scale computational micromechanical problems. Computer Methods in
Applied Mechanics and Engineering. 190. 43-44, 5639-5656.
207. Zohdi, T. I. (2001). Computational optimization of vortex manufacturing of advanced materials. Computer Methods in Applied Mechanics and Engineering.
190. 46-47, 6231-6256.
208. Zohdi, T. I. (2002). An adaptive-recursive staggering strategy for simulating
multifield coupled processes in microheterogeneous solids. The International
209. Zohdi, T. I. (accepted). On the propagation of microscale material uncertainty
in a class of hyperelastic finite deformation stored energy functions. The International Journal of Fracture.
210. Zohdi, T. I. (accepted). Modeling and simulation of variably coupled timetransient thermo-chemo-mechanical processes in multiphase solids. Journal
of Applied Mathematics and Physics. (Zeitschrift fur Angewandte Mathematik
und Physik)
211. Zohdi, T. I. (accepted). Incorporation of microfield distortion into rapid effective property design. Mathematics and Mechanics of Solids.
212. Zohdi, T. I. (accepted). On the tailoring of microstructures for prescribed
effective properties. The International Journal of Fracture.
213. Zohdi, T. I. (accepted). Statistical ensemble error bounds for homogenized
microheterogeneous solids. Journal of Applied Mathematics and Physics.
(Zeitschrift fur Angewandte Mathematik und Physik)
214. Zohdi, T. I., Kachanov, M. and Sevostianov, I. (accepted). On perfectly-plastic
flow in porous material. The International Journal of Plasticity.
215. Zohdi, T. I. and Steigmann, D. J. (accepted). The toughening effect of microscopic filament misalignment on macroscopic fabric response. The International Journal of Fracture.
209
216. Zohdi, T. I. (accepted). Modeling and simulation of progressive penetration of

multilayered ballistic fabric shielding. Computational Mechanics.
217. Zohdi, T. I. (accepted). Review article: Computational micro-macro material
testing. Annual publication of the German Association for Computational
Mechanics.
218. Zohdi, T. I., Monteiro, P. J. M. and Lamour, V. (accepted). Extraction of the
mechanical properties of dense ettringite from porous compacts. The International Journal of Fracture.
219. Zohdi, T. I. (accepted). Bounding envelopes in multiphase material design.
The Journal of Elasticity.
220. Zohdi, T. I. (accepted). Large-scale statistical inverse computation of inelastic
accretion in transient granular flows. The International Journal of Nonlinear
Mechanics.
221. Zohdi, T. I. (under review). Genetic optimization of statistically uncertain
microheterogeneous solids.
222. Zohdi, T. I. (under review). Compaction of cohesive hyperelastic granules at
finite strains.
223. Zohdi, T. I. (under review). Staggering-error controlled simulation of inelastic
processes in random microheterogeneous solids.
224. Zohdi, T. I. (under review). Modeling and rapid computational design of ballistic fabric.
225. Zohdi, T. I. (under review). Computational design of fatigue-resistant microheterogeneous materials.

Zohdi

Diunggah oleh

Informasi Dokumen

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Zohdi

Diunggah oleh

Hak Cipta:

Format Tersedia

Computational modeling

and design of new random

CISM Course Notes

1.6.4 Consequences of positive-definiteness

2 Fundamental weak formulations

3 Fundamental micro-macro concepts

4 A basic finite element implementation

5 Computational/statistical testing methods

6 Higher order statistics

7 Domain decomposition analogies and extensions

8 Nonconvex-nonderivative genetic material design

characteristics of modified base materials are the aggregate response of an assemblage

If one were to attempt to perform a direct numerical simulation, for example of

< > =EI :*< >

Fig. I.5 The use of effective properties in structural engineering.

impossible. Because of these facts, the use of regularized or homogenized material

Fig. I.6 The size requirements of a representative volume element.

sentative sample of material, referred to in the literature as a representative volume

I.1 BASIC CONCEPTS IN MACRO-MICRO MODELING

I.2 HISTORICAL OVERVIEW

[151] (1929). Voigt assumed that the strain field within

I.3 OBJECTIVES OF THIS MONOGRAPH

SOME BASICS IN THE MECHANICS OF SOLID CONTINUA

1.1 KINEMATICS OF DEFORMATIONS

with continuous partial derivatives to whatever order is required. The description

Deformation of line elements

Partial differentiation of the displacement vector

Alternatively, we have with

Therefore we have the so-called Lagrangian strain tensor

> 9 @ 9 @ > ; 9 @ > ; @

SOME BASICS IN THE MECHANICS OF SOLID CONTINUA

; , > 9 @ A 9 @ A > ; 9 @ A > ; @ A > 

can be reinterpreted as the result of a mapping

Infinitesimal strain measures

9 @FA9 @ 9 @F> ; A @> ;

In infinitesimal deformation theory, the displacement

Fig. 1.3 A differential volume.

The Jacobian of the Deformation Gradient

The Jacobian of the deformation gradient,

To interpret the Jacobian in

SOME BASICS IN THE MECHANICS OF SOLID CONTINUA

Fig. 1.4 Newtons laws applied to a continua.

1.2 EQUILIBRIUM/KINETICS OF SOLID CONTINUA

EQUILIBRIUM/KINETICS OF SOLID CONTINUA

Fig. 1.5 Cauchy tetrahedron: a sectioned material point.

SOME BASICS IN THE MECHANICS OF SOLID CONTINUA

Volumetric and surfacic postulates

is the so-called Cauchy stress tensor.

Substitution of Equation 1.11 into Equation 1.10 yields

REFERENTIAL DESCRIPTIONS OF BALANCE LAWS

Fig. 1.6 Stress at a point.

1.2.2.1 Symmetry of the stress tensor Starting with an angular momentum

1.3 REFERENTIAL DESCRIPTIONS OF BALANCE LAWS

SOME BASICS IN THE MECHANICS OF SOLID CONTINUA

Fig. 1.7 A current and reference surface element.

> 9  >

to the so-called Nanson formula for transformation of surface elements. Consider

@ G9 %> "

THE FIRST LAW OF THERMODYNAMICS/AN ENERGY BALANCE

" . Since the

1.4 THE FIRST LAW OF THERMODYNAMICS/AN ENERGY BALANCE

. Using this and the

By using the divergence theorem, we obtain

SOME BASICS IN THE MECHANICS OF SOLID CONTINUA

> 9 @ 9 @ > ; 9 @ > ; @

; , > 9 @ A 9 @ A > ; 9 @ A > ; @ A >

9 @FA9 @ 9 @F> ; A @> ;

> 9 >

@ G9 %> "

' ( * ( ' - ' - ( -

::: :: :::

rotational transformation matrix. Imposing elastic symmetry means the components

:: :: :: ::: ::

:: ::

:::::

::: :: :::

::

0. 9 > )G9 > <Q9 : > 9 $> G9 > 9 >

, where we call the symbol the hydrostatic pressure, and

and the specified traction boundary conditions

Since the integrals are arbitrary, we have

58729 >

to the compressible part