Interdisciplinary Applied Mathematics

Volume 23

Editors
S.S. Antman J.E. Marsden
L. Sirovich S. Wiggins

Geophysics and Planetary Sciences

Mathematical Biology
L. Glass, J.D. Murray
Mechanics and Materials
R.V. Kohn
Systems and Control
S.S. Sastry, P.S. Krishnaprasad

Problems in engineering, computational science, and the physical and biological

sciences are using increasingly sophisticated mathematical techniques. Thus, the
bridge between the mathematical sciences and other disciplines is heavily trav-
eled. The correspondingly increased dialog between the disciplines has led to the
establishment of the series: Interdisciplinary Applied Mathematics.

The purpose of this series is to meet the current and future needs for the interac-
tion between various science and technology areas on the one hand and mathe-
matics on the other. This is done, firstly, by encouraging the ways that mathe-
matics may be applied in traditional areas, and well as point towards new and
innovative areas of applications; and, secondly, by encouraging other scientific
disciplines to engage in a dialog with mathematicians outlining their problems to
both access new methods and suggest innovative developments within mathe-
matics itself.

The series will consist of monographs and high-level texts from researchers
working on the interplay between mathematics and other fields of science and
technology.
Interdisciplinary Applied Mathematics
Volumes published are listed at the end of this book.

Springer
New York
Berlin
Heidelberg
Hong Kong
London
Milan
Paris
Tokyo
Muhammad Sahimi

Heterogeneous Materials
Nonlinear and Breakdown Properties
and Atomistic Modeling

With 119 Illustrations

Muhammad Sahimi
Department of Chemical Engineering
University of Southern California
Los Angeles, CA 90089-1211
USA
[email protected]
Editors
J.E. Marsden L. Sirovich
Control and Dynamical Systems Division of Applied Mathematics
Mail Code 108-81 Brown University
California Institute of Technology Providence, RI 02912
Pasadena, CA 91125 USA
USA [email protected]
[email protected]
S. Wiggins S.S. Antman
School of Mathematics Department of Mathematics
University of Bristol and
Bristol, BS8 1TW Institute of Physical Science and Technology
United Kingdom University of Maryland
[email protected] College Park, MD 20742
USA
[email protected]
Cover illustration:

Mathematics Subject Classification (2000): 82-02, 65M

Library of Congress Cataloging-in-Publication Data
Sahimi, Muhammad.
Heterogeneous materials / Muhammad Sahimi.
p. cm. — (Interdisciplinary applied mathematics ; 22-23)
Includes bibliographical references and indexes.
Contents: [1] Linear transport and optical properties — [2] Nonlinear and breakdown
properties and atomistic modeling.
ISBN 0-387-00167-0 (v. 1 : alk. paper) — ISBN 0-387-00166-2 (v. 2 : alk. paper)
1. Inhomogenesou materials. 2. Composite materials. I. Title. II. Interdisciplinary
applied mathematiccs ; v. 22-23.
TA418.9.I53 S24 2003
620.1⬘1—dc21 2002042744
ISBN 0-387-95541-0 Printed on acid-free paper.
© 2003 Springer-Verlag New York, Inc.
All rights reserved. This work may not be translated or copied in whole or in part without the writ-
ten permission of the publisher (Springer-Verlag New York, Inc., 175 Fifth Avenue, New York, NY
10010, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in
connection with any form of information storage and retrieval, electronic adaptation, computer
software, or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if
they are not identified as such, is not to be taken as an expression of opinion as to whether or not
they are subject to proprietary rights.
Printed in the United States of America.
9 8 7 6 5 4 3 2 1 SPIN 10885680
www.springer-ny.com
Springer-Verlag New York Berlin Heidelberg
A member of BertelsmannSpringer Science+Business Media GmbH
 To children of the third world
who have the talent but not the means to succeed
and to
the memory of my father, Habibollah Sahimi,
who instilled in me, a third world child, the love of reading
Preface

Disorder plays a fundamental role in many natural and man-made systems that
are of industrial and scientific importance. Of all the disordered systems, hetero-
geneous materials are perhaps the most heavily utilized in all aspects of our daily
lives, and hence have been studied for a long time. With the advent of new ex-
perimental techniques, it is now possible to study the morphology of disordered
materials and gain a much deeper understanding of their properties. Novel tech-
niques have also allowed us to design materials of morphologies with the properties
that are suitable for intended applications.
With the development of a class of powerful theoretical methods, we now have
the ability for interpreting the experimental data and predicting many properties
of disordered materials at many length scales. Included in this class are renor-
malization group theory, various versions of effective-medium approximation,
percolation theory, variational principles that lead to rigorous bounds to the ef-
fective properties, and Green function formulations and perturbation expansions.
The theoretical developments have been accompanied by a tremendous increase in
the computational power and the emergence of massively parallel computational
strategies. Hence, we are now able to model many materials at molecular scales
and predict many of their properties based on first-principle computations.
In this two-volume book we describe and discuss various theoretical and com-
putational approaches for understanding and predicting the effective macroscopic
properties of heterogeneous materials. Most of the book is devoted to comparing
and contrasting the two main classes of, and approaches to, disordered materials,
namely, the continuum models and the discrete models. Predicting the effective
properties of composite materials based on the continuum models, which are based
on solving the classical continuum equations of transport, has a long history and
goes back to at least the middle of the nineteenth century. Even a glance at the liter-
ature on the subject of heterogeneous materials will reveal the tremendous amount
of work that has been carried out in the area of continuum modeling. Rarely, how-
ever, can such continuum models provide accurate predictions of the effective
macroscopic properties of strongly disordered multiphase materials. In particular,
if the contrast between the properties of a material’s phases is large, and the phases
form large clusters, most continuum models break down. At the same time, due to
their very nature, the discrete models, which are based on a lattice representation
of a material’s morphology, have the ability for providing accurate predictions for
the effective properties of heterogeneous materials, even when the heterogeneities
are strong, while another class of discrete models, that represent a material as a
collection of its constituent atoms and molecules, provides accurate predictions of
viii Preface

the material’s properties at mesoscopic scales, and thus, in this sense, the discrete
models are complementary to the continuum models. The last three decades of
the twentieth century witnessed great advances in discrete modeling of materials
and predicting their macroscopic properties, and one main goal of this book is to
describe these advances and compare their predictions with those of the continuum
models. In Volume I we consider characterization and modeling of the morphology
of disordered materials, and describe theoretical and computational approaches for
predicting their linear transport and optical properties, while Volume II focuses
on nonlinear properties, and fracture and breakdown of disordered materials, in
addition to describing their atomistic modeling. Some of the theoretical and com-
putational approaches are rather old, while others are very new, and therefore we
attempt to take the reader through a journey to see the history of the development
of the subjects that are discussed in this book. Most importantly, we always com-
pare the predictions with the relevant experimental data in order to gain a better
understanding of the strengths and/or shortcomings of the two classes of models.
A large number of people have helped me gain deeper understanding of the
topics discussed in this book, and hence have helped me to write about them.
Not being able to name them all, I limit myself to a few of them who, directly
or indirectly, influenced the style and contents of this book. Dietrich Stauffer has
greatly contributed to my understanding of percolation theory, disordered media,
and critical phenomena, some of the main themes of this book; I am deeply grateful
to him. For their tireless help in the preparation of various portions of this book, I
would like to thank two of my graduate students, Sushma Dhulipala and Alberto
Schroth. Although they may not be aware of it, Professors Pedro Ponte Castañeda
of the University of Pennsylvania and Salvatore Torquato of Princeton University
provided great help by guiding me through their excellent work, which is described
in this book; I would like to thank them both. Some of my own work described in
this book has been carried out in collaboration with many people; I am pleased to
acknowledge their great contributions, especially those of Dr. Sepehr Arbabi, my
former doctoral student. The constant encouragement and support offered by many
of my colleagues, a list of whom is too long to be given here, are also gratefully
acknowledged. I would like particularly to express my deep gratitude to my former
doctoral student Dr. Jaleh Ghassemzadeh, who provided me with critical help at
all stages of preparation of this book. Several chapters of this book have been used,
in their preliminary versions, in some of the courses that I teach, and I would like
to acknowledge the comments that I received from my students.
My wife, Mahnoush, and son, Ali, put up with the countless hours, days, weeks,
and months that I spent in preparing this book and my almost complete absence
during the time that I was writing, but never denied me their love and support
without which this book would have never been completed; I love and cherish
them both.

Muhammad Sahimi
Los Angeles, California, USA
May 2002
Contents

Preface vii

Abbreviated Contents for Volume I x

Introduction to Volume II 1

1 Characterization of Surface Morphology 6

1.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.1 Self-Similar Fractal Structures . . . . . . . . . . . . . . . . . 7
1.2 The Correlation Function . . . . . . . . . . . . . . . . . . . . 9
1.3 Rough Surfaces: Self-affine Fractals . . . . . . . . . . . . . . 10
1.4 Generation of Rough Surfaces: Fractional
Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4.1 The Power-Spectrum Method . . . . . . . . . . . . . . 12
1.4.2 Successive Random Additions . . . . . . . . . . . . . 15
1.4.3 The Weierstrass–Mandelbrot Algorithm . . . . . . . . 15
1.5 Scaling Properties of Rough Surfaces . . . . . . . . . . . . . . 16
1.6 Modeling of Growth of Thin Films with Rough Surface . . . . 19
1.7 Measurement of Roughness Exponent . . . . . . . . . . . . . 22
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

I Effective Properties of Heterogeneous Materials with

Constitutive Nonlinearities 25
2 Nonlinear Conductivity and Dielectric Constant:
The Continuum Approach 27
2.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.1 Variational Principles . . . . . . . . . . . . . . . . . . . . . . 29
2.2 Bounds on the Effective Energy Function . . . . . . . . . . . . 34
2.2.1 Lower Bounds . . . . . . . . . . . . . . . . . . . . . . 35
2.2.1.1 One-Point Bounds . . . . . . . . . . . . . . 35
2.2.1.2 Two-Point Bounds . . . . . . . . . . . . . . 36
2.2.1.3 Three-Point Bounds . . . . . . . . . . . . . 38
2.2.2 Approximate Estimates of the Effective Energy . . . . 39
2.2.2.1 Conductor–
Superconductor Composites . . . . . . . . . 39
2.2.2.2 Conductor–Insulator Composites . . . . . . 40
x Contents

2.2.3 Upper Bounds and Estimates . . . . . . . . . . . . . . 40

2.3 Exact Results for Laminates . . . . . . . . . . . . . . . . . . . 42
2.4 Effective Dielectric Constant of Strongly
Nonlinear Materials . . . . . . . . . . . . . . . . . . . . . . . 45
2.4.1 Inclusions with Infinite Dielectric Constant . . . . . . 46
2.4.2 Inclusions with Zero Dielectric Constant . . . . . . . . 47
2.5 Effective Conductivity of Nonlinear Materials . . . . . . . . . 47
2.5.1 Materials with Nonlinear Isotropic Phases . . . . . . . 48
2.5.2 Strongly Nonlinear Materials with
Isotropic Phases . . . . . . . . . . . . . . . . . . . . . 50
2.6 Second-Order Exact Results . . . . . . . . . . . . . . . . . . . 53
2.6.1 Strongly Nonlinear Isotropic Materials . . . . . . . . . 55
2.6.1.1 The Maxwell–Garnett Estimates . . . . . . . 56
2.6.1.2 Effective-Medium
Approximation Estimates . . . . . . . . . . 57
2.6.2 Conductor–Superconductor Composites . . . . . . . . 58
2.6.3 Conductor–Insulator Composites . . . . . . . . . . . . 58
2.6.4 General Two-Phase Materials . . . . . . . . . . . . . . 60
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

3 Nonlinear Conductivity, Dielectric Constant, and

Optical Properties: The Discrete Approach 64
3.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.1 Strongly Nonlinear Composites . . . . . . . . . . . . . . . . . 64
3.1.1 Exact Solution for Bethe Lattices . . . . . . . . . . . . 66
3.1.1.1 Microscopic Versus
Macroscopic Conductivity . . . . . . . . . . 68
3.1.1.2 Effective-Medium Approximation for
Bethe Lattices . . . . . . . . . . . . . . . . 71
3.1.2 Effective-Medium Approximation for
Three-Dimensional Materials . . . . . . . . . . . . . . 71
3.1.3 The Decoupling Approximation . . . . . . . . . . . . 75
3.1.4 Perturbation Expansion . . . . . . . . . . . . . . . . . 76
3.1.5 Variational Approach . . . . . . . . . . . . . . . . . . 76
3.1.6 Exact Duality Relations . . . . . . . . . . . . . . . . . 77
3.1.7 Scaling Properties . . . . . . . . . . . . . . . . . . . . 79
3.1.7.1 Series Expansion Analysis . . . . . . . . . . 81
3.1.7.2 Field-Theoretic Approach . . . . . . . . . . 82
3.1.8 Resistance Noise, Moments of Current Distribution,
and Scaling Properties . . . . . . . . . . . . . . . . . . 83
3.2 Nonlinear Transport Caused by a Large External Field . . . . . 85
3.3 Weakly Nonlinear Composites . . . . . . . . . . . . . . . . . 89
3.3.1 Effective-Medium Approximation . . . . . . . . . . . 90
 Contents xi

3.3.2 Resistance Noise, Moments of Current Distribution,

and Scaling Properties . . . . . . . . . . . . . . . . . . 93
3.3.3 Crossover from Linear to Weakly
Nonlinear Conductivity . . . . . . . . . . . . . . . . . 97
3.3.4 Exact Duality Relations . . . . . . . . . . . . . . . . . 99
3.3.5 Comparison with the Experimental Data . . . . . . . . 101
3.4 Dielectric Constant of Weakly Nonlinear Composites . . . . . 103
3.4.1 Exact Results . . . . . . . . . . . . . . . . . . . . . . 104
3.4.2 Effective-Medium Approximation . . . . . . . . . . . 105
3.4.3 The Maxwell–Garnett Approximation . . . . . . . . . 105
3.5 Electromagnetic Field Fluctuations and
Optical Nonlinearities . . . . . . . . . . . . . . . . . . . . . . 106
3.5.1 Scaling Properties of Moments of the
Electric Field . . . . . . . . . . . . . . . . . . . . . . 109
3.5.1.1 Distribution of Electric Fields in
Strongly Disordered Composites . . . . . . . 110
3.5.1.2 Moments of the Electric Field . . . . . . . . 116
3.5.1.3 Field Fluctuations at Frequencies
Below the Resonance . . . . . . . . . . . . 118
3.5.1.4 Computer Simulation . . . . . . . . . . . . . 122
3.5.1.5 Comparison with the
Experimental Data . . . . . . . . . . . . . . 124
3.5.2 Anomalous Light Scattering from Semicontinuous
Metal Films . . . . . . . . . . . . . . . . . . . . . . . 125
3.5.2.1 Rayleigh Scattering . . . . . . . . . . . . . 126
3.5.2.2 Scaling Properties of the
Correlation Function . . . . . . . . . . . . . 128
3.5.3 Surface-Enhanced Raman Scattering . . . . . . . . . . 130
3.5.3.1 General Formulation . . . . . . . . . . . . . 131
3.5.3.2 Raman and Hyper-Raman Scattering in
Metal–Dielectric Composites . . . . . . . . 133
3.5.3.3 Comparison with the
Experimental Data . . . . . . . . . . . . . . 135
3.5.4 Enhancement of Optical Nonlinearities in
Metal–Dielectric Composites . . . . . . . . . . . . . . 135
3.5.4.1 Kerr Optical Nonlinearities . . . . . . . . . 135
3.5.4.2 Enhancement of Nonlinear Scattering
from Strongly Disordered Films . . . . . . . 139
3.5.4.3 Comparison with the
Experimental Data . . . . . . . . . . . . . . 143
3.6 Electromagnetic Properties of Solid Composites . . . . . . . . 143
3.6.1 Effective-Medium Approximation . . . . . . . . . . . 144
3.7 Beyond the Quasi-static Approximation: Generalized Ohm’s Law 149
3.8 Piecewise Linear Transport Processes . . . . . . . . . . . . . . 157
3.8.1 Computer Simulation . . . . . . . . . . . . . . . . . . 159
xii Contents

3.8.2 Scaling Properties . . . . . . . . . . . . . . . . . . . . 160

3.8.3 Effective-Medium Approximation . . . . . . . . . . . 160
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

4 Nonlinear Rigidity and Elastic Moduli:

The Continuum Approach 164
4.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
4.1 Constitutive Relations and Potentials . . . . . . . . . . . . . . 165
4.2 Formulation of the Problem . . . . . . . . . . . . . . . . . . . 169
4.3 The Classical Variational Principles . . . . . . . . . . . . . . . 170
4.3.1 One-Point Bounds . . . . . . . . . . . . . . . . . . . . 172
4.3.2 Two-Point Bounds: The Talbot–Willis Method . . . . . 172
4.4 Variational Principles Based on a Linear
Comparison Material . . . . . . . . . . . . . . . . . . . . . . 175
4.4.1 Materials with Isotropic Phases . . . . . . . . . . . . . 176
4.4.2 Strongly Nonlinear Materials . . . . . . . . . . . . . . 178
4.4.3 Materials with Anisotropic Phases . . . . . . . . . . . 178
4.4.3.1 Polycrystalline Materials . . . . . . . . . . . 179
4.4.3.2 Strongly Nonlinear Materials . . . . . . . . 180
4.4.3.3 Materials with Isotropic and Strongly
Nonlinear Phases . . . . . . . . . . . . . . . 181
4.4.3.4 Strongly Nonlinear
Polycrystalline Materials . . . . . . . . . . . 181
4.4.3.5 Ideally Plastic Materials . . . . . . . . . . . 182
4.5 Bounds with Piecewise Constant Elastic Moduli . . . . . . . . 182
4.5.1 Materials with Isotropic Phases . . . . . . . . . . . . . 183
4.5.2 Polycrystalline Materials . . . . . . . . . . . . . . . . 184
4.6 Second-Order Exact Results . . . . . . . . . . . . . . . . . . . 186
4.6.1 Weak-Contrast Expansion . . . . . . . . . . . . . . . . 186
4.6.2 Strong-Contrast Expansion . . . . . . . . . . . . . . . 188
4.7 Applications of Second-Order Exact Results . . . . . . . . . . 192
4.7.1 Porous Materials . . . . . . . . . . . . . . . . . . . . 192
4.7.1.1 Two-Point Bounds . . . . . . . . . . . . . . 193
4.7.1.2 Three-Point Bounds . . . . . . . . . . . . . 194
4.7.2 Rigidly Reinforced Materials . . . . . . . . . . . . . . 195
4.7.2.1 Two-Point Bounds . . . . . . . . . . . . . . 196
4.7.2.2 Three-Point Bounds and Estimates . . . . . . 197
4.7.3 Completely Plastic Materials . . . . . . . . . . . . . . 198
4.8 Other Theoretical Methods . . . . . . . . . . . . . . . . . . . 202
4.9 Critique of the Variational Procedure . . . . . . . . . . . . . . 203
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
 Contents xiii

II Fracture and Breakdown of

Heterogeneous Materials 207
5 Electrical and Dielectric Breakdown: The Discrete Approach 209
5.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
5.1 Continuum Models of Dielectric Breakdown . . . . . . . . . . 211
5.1.1 Griffith-like Criterion and the Analogy with
Brittle Fracture . . . . . . . . . . . . . . . . . . . . . 212
5.1.2 Computer Simulation . . . . . . . . . . . . . . . . . . 215
5.2 Discrete Models of Electrical Breakdown . . . . . . . . . . . . 215
5.2.1 The Dilute Limit . . . . . . . . . . . . . . . . . . . . 216
5.2.2 The Effect of Sample Size . . . . . . . . . . . . . . . 217
5.2.3 Electrical Failure in Strongly
Disordered Materials . . . . . . . . . . . . . . . . . . 218
5.2.4 Computer Simulation . . . . . . . . . . . . . . . . . . 220
5.2.5 Distribution of the Failure Currents . . . . . . . . . . . 222
5.2.6 The Effect of Failure Thresholds . . . . . . . . . . . . 224
5.2.7 Dynamical and Thermal Aspects of Electrical
Breakdown . . . . . . . . . . . . . . . . . . . . . . . 226
5.2.7.1 Discrete Dynamical Models . . . . . . . . . 227
5.2.7.2 Breakdown in an AC Field:
Thermal Effects . . . . . . . . . . . . . . . 230
5.2.7.3 Comparison with the
Experimental Data . . . . . . . . . . . . . . 232
5.3 Electromigration Phenomena and the Minimum Gap . . . . . . 234
5.4 Dielectric Breakdown . . . . . . . . . . . . . . . . . . . . . . 237
5.4.1 Exact Duality Relation . . . . . . . . . . . . . . . . . 237
5.4.2 Stochastic Models . . . . . . . . . . . . . . . . . . . . 238
5.4.3 Deterministic Models . . . . . . . . . . . . . . . . . . 241
5.4.3.1 Scaling Properties of
Dielectric Breakdown . . . . . . . . . . . . 243
5.4.3.2 Distribution of Breakdown Fields . . . . . . 245
5.4.4 Comparison with the Experimental Data . . . . . . . . 247
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248

6 Fracture: Basic Concepts and Experimental Techniques 249

6.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
6.1 Historical Background . . . . . . . . . . . . . . . . . . . . . . 250
6.2 Fracture of a Homogeneous Solid . . . . . . . . . . . . . . . . 252
6.3 Introduction of Heterogeneity . . . . . . . . . . . . . . . . . . 253
6.4 Brittle Versus Ductile Materials . . . . . . . . . . . . . . . . . 254
6.5 Mechanisms of Fracture . . . . . . . . . . . . . . . . . . . . . 255
6.5.1 Elastic Incompatibility . . . . . . . . . . . . . . . . . 255
6.5.2 Plastic Deformation . . . . . . . . . . . . . . . . . . . 255
6.5.3 Coalescence of Plastic Cavities . . . . . . . . . . . . . 256
xiv Contents

6.5.4 Cracks Initiated by Thin Brittle Films . . . . . . . . . 256

6.5.5 Crazing . . . . . . . . . . . . . . . . . . . . . . . . . 257
6.5.6 Boundary Sliding . . . . . . . . . . . . . . . . . . . . 257
6.6 Conventional Fracture Modes . . . . . . . . . . . . . . . . . . 257
6.7 Stress Concentration and Griffith’s Criterion . . . . . . . . . . 258
6.8 The Stress Intensity Factor and Fracture Toughness . . . . . . 261
6.9 Classification of the Regions Around the Crack Tip . . . . . . 263
6.10 Dynamic Fracture . . . . . . . . . . . . . . . . . . . . . . . . 265
6.11 Experimental Methods in Dynamic Fracture . . . . . . . . . . 266
6.11.1 Application of External Stress . . . . . . . . . . . . . 266
6.11.1.1 Static Stress . . . . . . . . . . . . . . . . . . 266
6.11.1.2 Initiation of Fractures . . . . . . . . . . . . 268
6.11.1.3 Dynamic Stress . . . . . . . . . . . . . . . . 268
6.11.2 Direct Measurement of the Stress
Intensity Factor . . . . . . . . . . . . . . . . . . . . . 269
6.11.2.1 The Method of Caustics . . . . . . . . . . . 269
6.11.2.2 Photoelasticity . . . . . . . . . . . . . . . . 269
6.11.3 Direct Measurement of Energy . . . . . . . . . . . . . 270
6.11.4 Measurement of Fracture Velocity . . . . . . . . . . . 270
6.11.4.1 High-Speed Photography . . . . . . . . . . . 270
6.11.4.2 Measurement of Resistivity . . . . . . . . . 271
6.11.4.3 Ultrasonic Measurements . . . . . . . . . . 271
6.11.5 Measurement of the Thermal Effects . . . . . . . . . . 272
6.11.6 Measurement of Acoustic Emissions
of Fractures . . . . . . . . . . . . . . . . . . . . . . . 272
6.12 Oscillatory Fracture Patterns . . . . . . . . . . . . . . . . . . 273
6.13 Mirror, Mist, and Hackle Pattern on a Fracture Surface . . . . . 275
6.14 Roughness of Fracture Surfaces . . . . . . . . . . . . . . . . . 277
6.14.1 Measurement of Roughness of Fracture Surface . . . . 279
6.14.2 Mechanisms of Surface Roughness Generation . . . . 283
6.14.2.1 Growth of Microcracks . . . . . . . . . . . . 283
6.14.2.2 Plastic Deformation . . . . . . . . . . . . . 284
6.14.2.3 Macroscopic Branching
and Bifurcation . . . . . . . . . . . . . . . . 284
6.15 Cleavage of Crystalline Materials . . . . . . . . . . . . . . . . 284
6.16 Fracture Properties of Materials . . . . . . . . . . . . . . . . . 286
6.16.1 Polymeric Materials . . . . . . . . . . . . . . . . . . . 287
6.16.2 Ceramics . . . . . . . . . . . . . . . . . . . . . . . . 288
6.16.3 Metals . . . . . . . . . . . . . . . . . . . . . . . . . . 289
6.16.4 Fiber-Reinforced Composites . . . . . . . . . . . . . . 290
6.16.5 Metal-Matrix Composites . . . . . . . . . . . . . . . . 290
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291

7 Brittle Fracture: The Continuum Approach 292

7.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
 Contents xv

7.1 Scaling Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 294

7.1.1 Scaling Analysis of Materials Strength . . . . . . . . . 294
7.1.2 Scaling Analysis of Dynamic Fracture . . . . . . . . . 295
7.2 Continuum Formulation of Fracture Mechanics . . . . . . . . . 298
7.2.1 Dissipation and the Cohesive Zone . . . . . . . . . . . 298
7.2.2 Universal Singularities near the Fracture Tip . . . . . . 299
7.3 Linear Continuum Theory of Elasticity . . . . . . . . . . . . . 300
7.3.1 Static Fractures in Mode III . . . . . . . . . . . . . . . 303
7.3.2 Dynamic Fractures in Mode I . . . . . . . . . . . . . . 304
7.4 The Onset of Fracture Propagation: Griffith’s Criterion . . . . . 308
7.5 The Equation of Motion for a Fracture in an
Infinite Plate . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
7.5.1 Mode III . . . . . . . . . . . . . . . . . . . . . . . . . 314
7.5.2 Mode I . . . . . . . . . . . . . . . . . . . . . . . . . . 316
7.6 The Path of a Fracture . . . . . . . . . . . . . . . . . . . . . . 318
7.6.1 Planar Quasi-static Fractures: Principle of
Local Symmetry . . . . . . . . . . . . . . . . . . . . . 318
7.6.2 Three-Dimensional Quasi-static Fractures . . . . . . . 319
7.6.3 Dynamic Fractures: Yoffe’s Criterion . . . . . . . . . . 320
7.7 Comparison with the Experimental Data . . . . . . . . . . . . 321
7.7.1 The Limiting Velocity of a Fracture . . . . . . . . . . . 323
7.8 Beyond Linear Continuum Fracture Mechanics . . . . . . . . . 325
7.8.1 The Dissipated Heat . . . . . . . . . . . . . . . . . . . 325
7.8.2 The Structure of Fracture Surface . . . . . . . . . . . . 327
7.8.3 Topography of Fracture Surface . . . . . . . . . . . . 327
7.8.4 Properties of Fracture Surface . . . . . . . . . . . . . 327
7.8.5 Conic Markings on Fracture Surface . . . . . . . . . . 328
7.8.6 Riblike Patterns on Fracture Surface . . . . . . . . . . 329
7.8.7 Roughness of Fracture Surface . . . . . . . . . . . . . 329
7.8.8 Modeling Rough Fracture Surfaces . . . . . . . . . . . 332
7.8.9 Fracture Branching at Microscopic Scales . . . . . . . 334
7.8.10 Multiple Fractures Due to Formation and
Coalescence of Microscopic Voids . . . . . . . . . . . 334
7.8.11 Microscopic Versus Macroscopic
Fracture Branching . . . . . . . . . . . . . . . . . . . 335
7.8.12 Nonuniqueness of the Stress Intensity Factor . . . . . . 336
7.8.13 Dependence of the Fracture Energy on
Crack Velocity . . . . . . . . . . . . . . . . . . . . . . 336
7.8.14 Generalized Griffith Criterion for Fractures with
Self-Affine Surfaces . . . . . . . . . . . . . . . . . . . 337
7.8.15 Crack Propagation Faster Than the Rayleigh
Wave Speed . . . . . . . . . . . . . . . . . . . . . . . 340
7.9 Shortcomings of Linear Continuum Fracture Mechanics . . . . 342
xvi Contents

7.10 Instability in Dynamic Fracture of Isotropic

Amorphous Materials . . . . . . . . . . . . . . . . . . . . . . 342
7.10.1 The Onset of Velocity Oscillations . . . . . . . . . . . 343
7.10.2 Relation Between Surface Structure and
Dynamical Instability . . . . . . . . . . . . . . . . . . 344
7.10.3 Mechanism of the Dynamical Instability . . . . . . . . 345
7.10.4 Universality of Microbranch Profiles . . . . . . . . . . 347
7.10.5 Crossover from Three-Dimensional to
Two-Dimensional Behavior . . . . . . . . . . . . . . . 347
7.10.6 Energy Dissipation . . . . . . . . . . . . . . . . . . . 348
7.10.7 Universality of the Dynamical Instability . . . . . . . . 349
7.11 Models of the Cohesive Zone . . . . . . . . . . . . . . . . . . 349
7.11.1 The Barenblatt–Dugdale Model . . . . . . . . . . . . . 350
7.11.2 Two-Field Continuum Models . . . . . . . . . . . . . 351
7.11.3 Finite-Element Simulation . . . . . . . . . . . . . . . 354
7.11.4 Fracture Propagation in Three Dimensions . . . . . . . 357
7.11.5 Failure of Dynamic Models of Cohesive Zone . . . . . 362
7.12 Brittle-to-Ductile Transition . . . . . . . . . . . . . . . . . . . 363
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366

8 Brittle Fracture: The Discrete Approach 367

8.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
8.1 Quasi-static Fracture of Fibrous Materials . . . . . . . . . . . 371
8.1.1 Equal-Load-Sharing (Democratic) Models . . . . . . . 372
8.1.2 Local-Load-Sharing Models . . . . . . . . . . . . . . 375
8.1.3 Computer Simulation . . . . . . . . . . . . . . . . . . 381
8.1.4 Mean-Field and Effective-
Medium Approximations . . . . . . . . . . . . . . . . 384
8.2 Quasi-static Fracture of Heterogeneous Materials . . . . . . . 390
8.2.1 Lattice Models . . . . . . . . . . . . . . . . . . . . . 391
8.2.1.1 Shape of the Macroscopic Fracture . . . . . 397
8.2.1.2 Dependence of the Elastic Moduli
on the Extent of Cracking . . . . . . . . . . 400
8.2.1.3 Fracture Strength of Materials with
Strong Disorder . . . . . . . . . . . . . . . . 402
8.2.1.4 Distribution of Fracture Strength . . . . . . . 405
8.2.1.5 Size-Dependence of
Fracture Properties . . . . . . . . . . . . . . 408
8.2.2 Comparison with the Experimental Data . . . . . . . . 412
8.2.3 Percolation Versus Quasi-static Brittle Fracture . . . . 413
8.2.4 Universal Fixed Points in Quasi-static
Brittle Fracture . . . . . . . . . . . . . . . . . . . . . 416
8.3 Dynamic Brittle Fracture . . . . . . . . . . . . . . . . . . . . 421
8.3.1 Dynamic Fracture in Mode I . . . . . . . . . . . . . . 424
8.3.2 Dynamic Fracture in Mode III . . . . . . . . . . . . . 426
 Contents xvii

8.3.2.1 Phonon Emission . . . . . . . . . . . . . . . 434

8.3.2.2 Forbidden Fracture Velocities . . . . . . . . 436
8.3.2.3 Nonlinear Instabilities . . . . . . . . . . . . 436
8.3.2.4 The Connection to the
Yoffe’s Criterion . . . . . . . . . . . . . . . 437
8.3.3 The Effect of Quenched Disorder . . . . . . . . . . . . 438
8.3.4 Comparison with the Experimental Data . . . . . . . . 442
8.4 Fracture of a Brittle Material by an Impact . . . . . . . . . . . 443
8.5 Dynamic Fracture of Materials with Annealed Disorder . . . . 446
8.6 Fracture of Polymeric Materials . . . . . . . . . . . . . . . . . 448
8.7 Fracture of Thin Solid Films . . . . . . . . . . . . . . . . . . . 451
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453

III Atomistic and Multiscale Modeling of Materials 455

9 Atomistic Modeling of Materials 457
9.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
9.1 Density-Functional Theory . . . . . . . . . . . . . . . . . . . 461
9.1.1 Local-Density Approximation . . . . . . . . . . . . . 464
9.1.2 Generalized Gradient Approximation . . . . . . . . . . 466
9.1.3 Nonperiodic Systems . . . . . . . . . . . . . . . . . . 467
9.1.4 Pseudopotential Approximation . . . . . . . . . . . . . 467
9.2 Classical Molecular Dynamics Simulation . . . . . . . . . . . 471
9.2.1 Basic Principles . . . . . . . . . . . . . . . . . . . . . 472
9.2.2 Evaluation of Molecular Forces in a
Periodic System . . . . . . . . . . . . . . . . . . . . . 476
9.2.3 The Verlet and Leapfrog Algorithms . . . . . . . . . . 477
9.2.4 Constant-Energy Ensembles . . . . . . . . . . . . . . 479
9.2.5 Constant-Temperature Ensembles . . . . . . . . . . . 479
9.2.6 Constant-Pressure and Temperature Ensembles . . . . 481
9.2.7 Simulation of Rigid and Semirigid Molecules . . . . . 481
9.2.8 Ion–Ion Interactions . . . . . . . . . . . . . . . . . . . 486
9.3 Nonequilibrium Molecular Dynamics Simulation . . . . . . . . 490
9.4 Quantum Molecular Dynamics Simulation:
The Car–Parrinello Method . . . . . . . . . . . . . . . . . . . 494
9.4.1 The Equations of Motion . . . . . . . . . . . . . . . . 495
9.4.2 The Verlet Algorithm . . . . . . . . . . . . . . . . . . 496
9.4.3 The Kohn–Sham Eigenstates and
Orthogonalization of the Wave Functions . . . . . . . 497
9.4.4 Dynamics of the Ions and the Unit Cell . . . . . . . . . 498
9.4.4.1 The Hellmann–Feynman Theorem . . . . . . 499
9.4.4.2 Pulay Forces and Stresses . . . . . . . . . . 500
9.4.4.3 The Structure Factor and Total
Ionic Potential . . . . . . . . . . . . . . . . 501
xviii Contents

9.4.5Computational Procedure for Quantum

Molecular Dynamics . . . . . . . . . . . . . . . . . . 502
9.4.6 Linear System-Size Scaling . . . . . . . . . . . . . . . 506
9.4.7 Extensions of the Car–Parrinello Quantum
Molecular Dynamics Method . . . . . . . . . . . . . . 506
9.4.8 Tight-Binding Methods . . . . . . . . . . . . . . . . . 507
9.5 Direct Minimization of Total Energy . . . . . . . . . . . . . . 507
9.5.1 The Steepest-Descent Method . . . . . . . . . . . . . 508
9.5.2 The Conjugate-Gradient Method . . . . . . . . . . . . 508
9.5.3 Minimizing the Total Energy by the
Conjugate-Gradient Method . . . . . . . . . . . . . . 509
9.6 Vectorized and Massively-Parallel Molecular
Dynamics Simulation . . . . . . . . . . . . . . . . . . . . . . 513
9.6.1 Vectorized Molecular Dynamics Algorithms . . . . . . 513
9.6.2 Massively-Parallel Molecular
Dynamics Algorithms . . . . . . . . . . . . . . . . . . 514
9.6.2.1 Atom-Decomposition Algorithms . . . . . . 515
9.6.2.2 Force-Decomposition Algorithms . . . . . . 517
9.6.2.3 Spatial-Decomposition Algorithms . . . . . 519
9.6.2.4 Load Balance in Massively-Parallel
Molecular Dynamics Simulation . . . . . . . 521
9.6.2.5 Selecting a Massively-Parallel
Molecular Dynamics Algorithm . . . . . . . 522
9.7 Interatomic Interaction Potentials . . . . . . . . . . . . . . . . 523
9.7.1 The Embedded-Atom Model . . . . . . . . . . . . . . 524
9.7.2 The Stillinger–Weber Potential . . . . . . . . . . . . . 527
9.7.3 The Tersoff Potentials . . . . . . . . . . . . . . . . . . 529
9.7.4 The Brenner Potentials . . . . . . . . . . . . . . . . . 533
9.7.5 Other Interaction Potentials . . . . . . . . . . . . . . . 537
9.8 Molecular Dynamics Simulation of
Fracture Propagation . . . . . . . . . . . . . . . . . . . . . . . 538
9.8.1 Early Simulations . . . . . . . . . . . . . . . . . . . . 540
9.8.2 Large Size and Scalable Molecular Dynamics
Simulation of Fracture . . . . . . . . . . . . . . . . . 544
9.8.3 Comparison with the Experimental Observations . . . 546
9.8.3.1 Fracture Instabilities . . . . . . . . . . . . . 546
9.8.3.2 Morphology of Fracture Surface . . . . . . . 548
9.8.3.3 Fracture Propagation Faster Than the
Rayleigh Wave Speed . . . . . . . . . . . . 549
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 550

10 Multiscale Modeling of Materials: Joining Atomistic

Models with Continuum Mechanics 551
10.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 551
10.1 Multiscale Modeling . . . . . . . . . . . . . . . . . . . . . . . 554
 Contents xix

10.1.1 Sequential Multiscale Approach:

Atomistically-Informed Continuum Models . . . . . . 554
10.1.2 Parallel Multiscale Approach . . . . . . . . . . . . . . 556
10.2 Defects in Solids: Joining Finite-Element and
Atomistic Computations . . . . . . . . . . . . . . . . . . . . . 557
10.2.1 The Quasi-continuum Formulation . . . . . . . . . . . 559
10.2.2 Constitutive Models . . . . . . . . . . . . . . . . . . . 563
10.2.3 The Atomistic Model . . . . . . . . . . . . . . . . . . 563
10.2.4 Field Equations and Their Spatial Discretization . . . . 563
10.2.5 Local Quasi-continuum Formulation . . . . . . . . . . 565
10.2.6 Nonlocal Quasi-continuum Formulation . . . . . . . . 567
10.2.7 The Criterion for Nonlocality of Elements . . . . . . . 568
10.2.8 Application to Stacking Faults in FCC Crystals . . . . 570
10.2.9 Application to Nanoindentation . . . . . . . . . . . . . 573
10.3 Fracture Dynamics: Joining Tight-Binding, Molecular
Dynamics, and Finite-Element Computations . . . . . . . . . . 576
10.3.1 The Overall Hamiltonian . . . . . . . . . . . . . . . . 576
10.3.2 The Tight-Binding Region . . . . . . . . . . . . . . . 577
10.3.3 Molecular Dynamics Simulation . . . . . . . . . . . . 578
10.3.4 Finite-Element Simulation . . . . . . . . . . . . . . . 579
10.3.5 Interfacing Finite-Element and Molecular
Dynamics Regions . . . . . . . . . . . . . . . . . . . 581
10.3.6 Interfacing Molecular Dynamics and
Tight-Binding Regions . . . . . . . . . . . . . . . . . 584
10.3.7 Seamless Simulation . . . . . . . . . . . . . . . . . . 587
10.3.8 Multiscale Simulation of Fracture Propagation
in Silicon . . . . . . . . . . . . . . . . . . . . . . . . 587
10.4 Other Applications of Multiscale Modeling . . . . . . . . . . . 588
10.4.1 Atomistically Induced Stress Distributions in
Composite Materials . . . . . . . . . . . . . . . . . . 588
10.4.2 Chemical Vapor Deposition . . . . . . . . . . . . . . . 589
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 590

References 593

Index 633
Abbreviated Contents for Volume I

Preface

Abbreviated Contents for Volume II

1 Introduction
I Characterization and Modeling of the Morphology
2 Characterization of Connectivity and Clustering
3 Characterization and Modeling of the Morphology
II Linear Transport and Optical Properties
4 Effective Conductivity, Dielectric Constant, and
Optical Properties: The Continuum Approach
5 Effective Conductivity and Dielectric Constant:
The Discrete Approach
6 Frequency-Dependent Properties: The Discrete Approach
7 Rigidity and Elastic Properties: The Continuum Approach
8 Rigidity and Elastic Properties: The Discrete Approach
9 Rigidity and Elastic Properties of Network Glasses, Polymers,
and Composite Solids: The Discrete Approach

References

Index
Introduction to Volume II

In Volume I of this book, we presented a self-contained analysis of the morphol-

ogy of heterogeneous materials and their effective linear properties. Some of the
properties of heterogeneous materials that were studied in Volume I were,
(1) the effective (electrical, thermal, hopping, and Hall) conductivity;
(2) the effective dielectric constant and optical properties, and
(3) the effective elastic moduli.
In addition, we also considered some aspects of the classical (as opposed to
quantum-mechanical) superconductivity of composite materials. Both steady-state
and time- and frequency-dependent properties were considered, and the most sig-
nificant theoretical developments for modelling the morphology of heterogeneous
materials and predicting their effective linear properties were described in detail.
In addition, we also described the techniques for computer simulations of trans-
port processes in disordered materials, and compared their predictions with the
theoretical ones and also the relevant experimental data.
In the present Volume, we continue our study of transport processes in heteroge-
neous materials, except that we consider their effective nonlinear properties. After
the introductory Chapter 1 in which we study characterization of surface structure
of materials when the surface is rough, we embark on studying various nonlinear
processes in heterogeneous materials. To do this, we divide nonlinear transport
processes into two groups, which are as follows.

A. Constitutive Nonlinearity
Materials of this type always behave nonlinearly. For example, if in a composite
material the relation between the current I and voltage V is given by

I = gV n

where g is a generalized conductance of the material, then, as far as the electri-

cal conductivity is concerned, for n
= 1 the material always behaves nonlinearly.
We will study such nonlinear phenomena in Chapters 2–4, and describe vari-
ous approaches for predicting and estimating the effective nonlinear conductivity,
dielectric constant, optical properties, and elastic moduli and rigidity.
2 Introduction to Volume II

B. Threshold Nonlinearity
In this class of materials are those for which the nonlinearity arises as a result
of imposing on them an external field of sufficient intensity. Brittle fracture and
dielectric breakdown of composite solids are two important examples of such
nonlinear transport processes. In brittle fracture, for example, the elastic response
of a solid material is governed by the equations of linear elasticity until the external
stress or strain that has been imposed on the material exceeds a critical value, at
which time the material breaks down and microcracks begin to emerge. A list
of all possible nonlinear transport processes of this type is very long. This type
of nonlinearity will be studied in Chapters 5–8, and will include electrical and
dielectrical breakdown, brittle fracture, and the transition between brittle fracture
and ductile behavior.
One important point to remember is that, the interplay between a nonlinear
transport process and the disordered morphology of a composite material gives
rise to a rich variety of phenomena that are usually far more complex than what one
usually must deal with in linear processes. Over the past 15 years, an increasing
number of investigations have been devoted to such nonlinear transport processes,
and deeper insight into their properties has been acquired. A major goal of Volume
II is to describe this progress and compare various properties of nonlinear transport
processes in heterogeneous materials with their linear counterparts.

C. Theoretical Approaches
Although the analysis of transport processes in composite materials has a long
history, it is only in the past three decades that this analysis has been extended
to include detailed structural properties of the materials, and in particular the
distribution of their heterogeneities. Deriving exact results for the effective prop-
erties of composite materials with anything but the simplest morphologies is
extremely difficult, if not impossible, and thus one must resort to various approx-
imate techniques. At the same time, however, the advent of powerful computers
and development of efficient computational algorithms have allowed us to esti-
mate various properties of heterogeneous materials to practically any desired or
affordable accuracy.
To describe the theoretical approaches for estimating the effective properties of
composite materials, we divide them into two classes. In the first class of models are
what we refer to as the continuum models, while the second class is made of the dis-
crete models. Both types of models are described and analyzed in this Volume, and
what follows is a brief description of the general features of each class of models.

C.1 The continuum models

The physical laws that govern the transport processes at the microscopic level are
well understood. Thus, one can, in principle, write down the differential equa-
tions that describe transport of energy, charge, or stress in a material and specify
the associated initial and boundary conditions. However, as the morphology of
most real composite materials is very irregular, practical and economically fea-
 Introduction to Volume II 3

sible computations for exact estimation of the effective properties are still very
difficult—even in the event that one knows the detailed morphology of the material.
Thus, it becomes essential to adopt a macroscopic description at a length scale much
larger than the dimension of the individual phases of a composite material. The
governing equations are then discretized and solved numerically, provided that the
effective properties that appear in the transport equations are either supplied as the
inputs (through, for example, experimental measurements), or else a model for the
morphology of the material is assumed so that the effective transport properties
can be somehow estimated, so that the numerical solution yields other quantities
of interest, such as the potential distribution in the material. We refer to various
models associated with this classical description as the continuum models. These
models have been widely used because of their convenience and familiarity to the
engineers and materials scientists. Their limitations will be described and discussed
in the subsequent chapters.
In addition to deriving the effective macroscopic equations and obtaining their
solution by numerical calculations, one may also derive exact results in terms of
rigorous upper and lower bounds to the properties of interest. Hence, powerful
tools have been developed for deriving accurate upper and lower bounds and
estimates. Finally, various approximations, such as the mean-field and effective-
medium approximations, have also been developed in the context of the continuum
models. We will describe most of these theoretical approaches throughout both this
book and Volume II.

C.2 The discrete models

The second class of models, the discrete models, are free of many limitations of
the continuum models. They themselves are divided into two groups.
(1) In the first class of discrete models, a material is represented by a discrete
set of atoms and molecules that interact with each other through interatomic
potentials. In a solid material, the distance between the atoms is fixed. One
then carries out atomistic simulations of the materials’ behavior under a va-
riety of conditions. Several types of such simulations have been developed
over the past few decades. With the advent of massively-parallel computa-
tional algorithms, atomistic simulations have increasingly become a viable
and quantitative method of predicting the effective properties of materials.
We will describe such approaches in Chapters 9 and 10.
(2) In the second class are lattice models of composite materials. The bonds of
the lattices represent microscopic elements of the material. For example, they
represent a conducting or insulating elements, or an elastic or a plastic region.
They do not represent molecular bonds, and therefore such lattice models are
appropriate for length scales that are much larger than molecular scales. These
models have been advanced to describe various phenomena at the microscopic
level and have been extended in the last several years to also describe them
at the macroscopic length scales. We will describe both classes of discrete
models in this volume.
4 Introduction to Volume II

The main shortcoming of both groups of the discrete models, from a practical
point of view, is the large computational effort required for a realistic discrete repre-
sentation of the material and simulating its behavior, although the ever-increasing
computational power is addressing this difficulty.

D. The Organization of the Book

What we intend to do in this Volume, similar to Volume I, is describing the most
important developments in predicting the effective nonlinear properties of com-
posite materials, and comparing the predictions with the relevant experimental
data. To accomplish our goal, for each effective property we describe and discuss
the continuum and discrete models separately. Then, in Chapter 10 we describe
recent advances in multiscale modelling of materials’ properties—a method that
combines a discrete approach with a continuum model. Similar to Volume I, the
structure of each chapter is as follows.
(1) The main problem(s) of interest is (are) introduced.
(2) The problem(s) is (are) then analyzed by several methods, each of which
provide valuable insight into the solution of the problem(s) and the physical
phenomena that it (they) represent. Typically, each chapter starts with exact
and rigorous results, then describes analytical approximations, and finally dis-
cusses the numerical and computer simulation methods. The weakness and
strengths of each method are also pointed out. In this way, the most important
progress in understanding the physical phenomena of interest is described and
discussed.
(3) When possible (which is almost always the case), we compare the theoret-
ical predictions with the experimental data and/or high-resolution computer
simulation results.
Characterization of surface morphology materials, which are directly relevant
to most of what is discussed in this Volume, will be described in Chapter 1. Aside
from this chapter, this Volume is divided into three parts. In Part I (Chapters 2–
4) we study transport processes in heterogeneous materials that are characterized
by constitutive nonlinearities. Part II (Chapters 5–8) contains the description and
discussion of transport processes with threshold nonlinearity, including electrical
and dielectric breakdown, and brittle fracture of disordered materials. Finally, in
Part III we will describe (in Chapters 9 and 10) advances in atomistic modelling of
materials, and how a powerful new approach that combines atomistic simulations
with the continuum description—in effect a combination of a discrete approach
with a continuum model—promises to provide much deeper understanding of
materials, and deliver quantitative predictions for their effective properties.
Let us emphasize that, as in Volume I, although every attempt has been made to
discuss and cite the relevant literature on every subject that we consider, what we
do cite and bring to the attention of the reader represents what was known to us at
the time of writing this book, and/or what we considered to be the most relevant.
As such, this two-volume book represents the author’s biased view of the subject
of composite materials.
1
Characterization of Surface Morphology

1.0 Introduction
Natural, as well as man-made, materials have enormous variations in their mor-
phology, which consists of materials’ geometry, topology and surface structure.
The geometry refers to sizes of the micro- and mesoscale elements of the mate-
rials, as well as their shapes which range anywhere from completely ordered to
complex and seemingly chaotic patterns. Generally speaking, regular Euclidean
shapes are formed under close-to-equilibrium conditions, although even in such
cases equilibrium thermodynamics is often incapable of describing the process
that gives rise to such shapes. The topology of materials describes how the micro-
and mesoscale elements are connected to one another. The structure of materials’
surface, especially those that are produced under far-from-equilibrium conditions,
is also very important because the surface is often very rough and possesses com-
plex features. In recent years it has become clear that characterizing the surface
roughness will go a long way toward giving us a much better understanding of
materials’ microstructure and hence many of their effective properties. However,
when we speak of surface roughness, we must specify the length scales over
which the roughness is measured. Even the most rugged mountains look perfectly
smooth when viewed from the outer space! Therefore, surface roughness (and,
more generally, all the morphological characteristics) depends on the length scale
of observations or measurements. The effect of topology of disordered materials
on their effective transport properties is quantified by percolation theory which,
together with the effect of the geometry, was described in Chapters 2 and 3 of
Volume I, and their significance was emphasized throughout Volume I where we
analyzed effective linear properties of disordered materials. For other applications
of percolation theory see Sahimi (1994a). Stauffer and Aharony (1992) present a
simple introduction to the concepts of percolation theory. In this chapter, we con-
sider the structure and characteristics of materials’ surface, and describe various
theoretical and experimental methods of studying rough surfaces, which are di-
rectly relevant to the nonlinear phenomena in heterogeneous materials considered
in this Volume, particularly to their brittle fracture and dielectric breakdown.
An important example of a material with a rough surface are the thin films
that produced by molecular beam epitaxy, and are utilized for manufacturing of
semiconductors and computer chips. These films are made of silicon and other
6 1. Characterization of Surface Morphology

elements, and are prepared by deposition of atoms on a very clean surface. Thin
films with rough surfaces are also made by sputtering in which an energized beam
of particles is sent toward the bulk of a material. Collision of the beam particles
with the material causes ejection of some particles from the material’s surface,
which then deposit on another surface and start to grow a thin film of the original
material.
Although the enormous variations in the morphology of natural, and even man-
made, materials, particularly in their surface, are such that, up until a few decades
ago, the problem of describing and quantifying such morphologies seemed hope-
less, many experimental and theoretical developments of the past two decades have
brightened the prospects for deeper understanding of materials’ microstructures,
and in particular the structure of their surface. Among them are the advent of pow-
erful computers and novel experimental techniques that allow highly sophisticated
computations of materials’ properties and their measurement. In addition, the real-
ization that the complex microstructure and behavior of a wide variety of materials
can be quantitatively characterized by utilizing the ideas of fractal distributions,
have advanced our understanding of materials’ surface structure. As we discuss in
this chapter, fractal concepts provide us with a powerful tool for characterizing the
structure of materials’ surface and its roughness, and the long-range correlations
that often exist in their morphology.
The purpose of this chapter is to describe and discuss the essential features of
surface morphology and its dynamics during the process in which it is formed, and
how fractal concepts can be utilized for characterizing it. We already described in
Chapter 2 of Volume I most of the main concepts of fractal geometry, and therefore
in this chapter we restrict ourselves to a brief discussion of such concepts, after
which we study and analyze rough surfaces.

1.1 Self-Similar Fractal Structures

An intuitive and informal definition of a self-similar fractal object is that, in such
objects the part is reminiscent of the whole, implying that the object possesses
scale-invariant properties, i.e., its morphology repeats itself at different length
scales. This means that above a certain length scale—the lower cutoff scale for
fractality—the structure of a piece of the object can be magnified to recover its
structure at larger length scales up to another length scale—the upper cutoff for
its fractality. Below the lower cutoff and above the upper cutoff scales the system
loses its self-similarity. While there are disordered media that are self-similar at
any length scale, natural materials and media that exhibit self-similarity typically
lose their fractal characteristics at sufficiently small or large length scales.
One of the simplest characteristics of a self-similar fractal system is its fractal
dimension Df , which is defined as follows. We cover the fractal system by non-
overlapping d-dimensional spheres of Euclidean radius r, or boxes of linear size
r, and count the number N (r) of such spheres that is required for complete coverage
 1.1. Self-Similar Fractal Structures 7

of the system. The fractal dimension Df of the system is then defined by

ln N
Df = lim . (1)
r→0 ln(1/r)
Estimating the fractal dimension through the use of Eq. (1) is called the box-
counting method. For non-fractal objects, Df = d, where d is the Euclidean
dimensionality of the space in which they are embedded. Note that, in order to
be able to write down Eq. (1), we have implicitly assumed the existence of a
lower and an upper cutoff length scale for the fractality of the system which are,
respectively, the radius r of the spheres and the linear size L of the system.
One can also define the fractal dimension Df through the relation between the
system’s mass M and its characteristic length scale L. If the system is composed
of particles of radius rand mass m, then
M = cm(L/r)Df , (2)
where c is a geometrical constant of order 1. Since we can fix the dependence of
M on m and r, we can write
M(L) ∼ LDf . (3)
Often, measuring M entails using an ensemble of samples with similar structures,
rather than a single sample. In this case
M = cm(L/r)Df , (4)
where · implies an average over the mass of a large number of samples with
linear sizes in the range L ± δL, centered on L.
Most natural fractals are what we call statistically self-similar because their
self-similarity is only in an average sense. One of the most important examples
of such fractals is one which is generated by the diffusion-limited aggregation
model (Witten and Sander, 1981). In this model the site at the center of a lattice is
occupied by a stationary particle. A new particle is then injected into the lattice,
far from the center, which diffuses on the lattice until it reaches a surface site,
i.e., an empty site which is a nearest neighbor of the stationary particle, at which
time the particle sticks to it and remains there permanently. Another diffusing
particle is then injected into the lattice to reach another surface (empty) site and
stick to it, and so on. If this process continues for a long time, a large aggregate
is formed. The most important property of diffusion-limited aggregates is that
they have a self-similar fractal structure (for a review see, for example, Meakin,
1998) with Df  1.7 and 2.45 for 2D and 3D aggregates, respectively. A two-
dimensional (2D) example of such aggregates is shown in Figure 1.1. Diffusion-
limited aggregates have found wide applications, ranging from colloidal systems,
to miscible displacement processes in porous media, to describing cellular patterns
in human bone marrow (Naeim et al., 1996). We will come back to this model in
Chapters 5 and 8, where we describe models of dielectric breakdown and fracture
of composite materials.
8 1. Characterization of Surface Morphology

Figure 1.1. A two-dimensional diffusion-limited aggregate.

1.2 The Correlation Function

A powerful method for testing self-similarity of disordered media is to construct a
correlation function Cn (rn ) defined by
Cn (rn ) = ρ(r0 )ρ(r0 + r1 ) · · · ρ(r0 + rn ), (5)
where ρ(r) is the density at position r, and the average is taken over all possible
values of r0 . Here rn denotes the set of points at r1 , · · · , rn . If an object is self-
similar, then its correlation function defined by Eq. (5) should remain the same, up
to a constant factor, if all the length scales of the system are rescaled by a constant
factor b. Thus, one must have
Cn (br1 , br2 , · · · , brn ) = b−nx Cn (r1 , · · · , rn ). (6)
It is not difficult to see that only a power-law correlation function can satisfy
Eq. (6). Moreover, it can be shown that one must have x = d − Df , where the
quantity x is called the co-dimensionality. However, in most cases only the two-
point, or the direct, correlation function can be computed or measured with high
precisions, and therefore we focus on this quantity. In practice, to construct the
direct correlation function for use in analyzing a self-similar fractal structure, one
typically employs a digitized image of the system. The correlation function is then
written as
1

C(r) = s(r )s(r + r ), (7)
 
r

where s(r) is a function such that s(r) = 1 if a point at r belongs to the system,
s(r) = 0 otherwise, and r = |r|. Because of self-similarity of the system, the direct
 1.3. Rough Surfaces: Self-affine Fractals 9

correlation function C(r) decays as

C(r) ∼ r Df −d . (8)
This power-law decay of C(r) not only provides a test of self-similarity of a dis-
ordered medium or material, it also gives us a means of estimating its fractal
dimension since, according to Eq. (8), if one prepares a logarithmic plot of C(r)
versus r, then for a fractal object one should obtain a straight line with a slope
Df − d. Estimating the fractal dimension based on the direct correlation function
has proven to be a very robust and reliable method. Equation (8) has an impor-
tant implication: There are long-range correlations in a self-similar fractal system,
because C(r) → 0 only when r → ∞. The existence of such correlations has
important implications for estimating the effective transport properties of disor-
dered materials (see, for example, Sahimi, 1994b, 1995a, and references therein).
Other experimental methods of estimating the fractal dimension were described
in Chapter 2 of Volume I, and therefore are not repeated here.

1.3 Rough Surfaces: Self-affine Fractals

The self-similarity of a fractal structure implies that its microstructure is invariant
under an isotropic rescaling of lengths, i.e., if all lengths in all directions are
rescaled by the same scale factor. However, there are many fractals that preserve
their scale-invariance only if lengths in different directions are rescaled by factors
that are direction dependent. In other words, the scale-invariance of such systems
is preserved only if lengths in x-, y-, and z-directions are scaled by scale factors bx ,
by , and bz , where in general these scale factors are not equal. This type of scale-
invariance implies that the fractal system is, in some sense, anisotropic. Such
fractal systems are called self-affine, a term that was first used by Mandelbrot
(1985). If a fractal structure is self-affine, it can no longer be described by a single
fractal dimension Df , and in fact if one utilizes any of the methods of estimating
a fractal dimension that were in Sections 1.1 and 1.2, then, the resulting fractal
dimension would depend on the length scales over which the method is utilized.
A well-known example of a process that gives rise to a self-affine fractal is
a marginally stable growth of an interface. For example, if water displaces oil
in a porous medium, the interface between water and oil is a self-affine fractal.
Well-known examples of man-made materials with rough and self-affine surfaces
include thin films that are formed by molecular beam epitaxy. Among naturally-
made surfaces that are rough and have self-affine properties are bacterial colonies,
and pores and fractures of rock and other types of porous media. Many properties
of such materials are described by a function f (x) that also possesses a self-affine
structure. For example, the surface height h(x, y) at a lateral position x of a rough
surface, e.g., the internal surface of a rock fracture, and the porosity distribution of
rock along a well at depths x, both have self-affine property. Self-affinity of many
natural systems that are associated with Earth, such as various properties of natural
rock, is quite understandable, since gravity plays a dominant role in one direction
10 1. Characterization of Surface Morphology

but has very little effect in the other directions, hence generating anisotropy in the
structure of rock. The interested reader is referred to Family and Vicsek (1991) for
an excellent collection of articles which describe a wide variety of rough surfaces
with self-affine properties.
Self-affine fractals that one encounters in practical situations are typically disor-
dered, and thus their self-affinity is only in a statistical sense. For the problems that
are of interest to us in this book, a disordered self-affine fractal can be thought of as
the fluctuations about a straight line or a flat surface. Such fluctuations can generate
rough self-affine curves or surfaces. If we consider the height difference between a
pair of points h(x1 ) and h(x2 ) on a self-affine surface h(x) that lie above or below
points separated by a distance x1 − x2 = x = |x| on a flat reference surface (or
line), then
|h(x1 ) − h(x2 )| ∼ x H , (9)
where H is called the Hurst exponent. One may generalize Eq. (9) to higher dimen-
sions, and generate rough surfaces that are encountered in a variety of contexts,
from surface of pores of a natural porous medium (see, for example, Sahimi, 1993b,
1995b, for comprehensive discussions) to fracture surface of heterogeneous ma-
terials (see Chapters 6 and 7), to thin films that are formed by a deposition process
(see below).

1.4 Generation of Rough Surfaces: Fractional Brownian

Motion
We now describe two fractal processes that are used for generating rough curves
(1D profiles) and surfaces. The properties that these self-affine fractal processes
possess may also be used as guides to better understanding of rough surfaces
that one encounters in practical applications. In addition, because these stochastic
processes generate fractal sets with long-range correlations, they have been widely
used for modeling of a variety of phenomena in engineering and materials science
in which the effect of long-range correlations is paramount. We first consider the
1D case, and define a stochastic process BH (t), called the fractional Brownian
motion (fBm), by (Mandelbrot and Van Ness, 1968)
 t 
1
BH (t) − BH (0) = K(t − s)dB(s) , (10)
(H + 1/2) −∞
where t can be a spatial or temporal variable. Here (x) is the gamma function,
H is the Hurst exponent defined above, and the kernel K(t − s) is given by

(t − s)H −1/2 0≤s≤t
K(t − s) = −1/2 −1/2 (11)
(t − s)H − (−s) H s < 0.
It is not difficult to show that
BH (bt) − BH (0) ≡ bH [BH (t) − BH (0)], (12)
 1.4. Generation of Rough Surfaces: Fractional Brownian Motion 11

where “≡” means “statistically equivalent to.”Aremarkable property of fBm is that

it generates correlations with infinite extent. To see this, consider the correlation
function C(t) of future increments BH (t) with past increments −BH (−t) which
is defined by
−BH (−t)BH (t)
C(t) = . (13)
BH (t)2 

It is straightforward to show that C(t) = 2(22H −1 − 1), independent of t. More-

over, the type of the correlations can be tuned by varying H . If H > 1/2, then
fBm displays persistence, i.e., a trend (for example, a high or a low value) at t is
likely to be followed by a similar trend at t + t, whereas if H < 1/2, then fBm
generates antipersistence, i.e., a trend at t is not likely to be followed by a similar
trend at t + t. For H = 1/2 the past and future are not correlated, and thus the
increments in BH (t) are completely random and uncorrelated. Thus, varying H
allows us to generate infinitely long-range correlations or anticorrelations.
We can generalize the above 1D fBm to 2D or 3D. Hence, if we consider two
arbitrary points x and x0 in 2D or 3D space, the fBm is defined by

[BH (x) − BH (x0 )]2  ∼ |x − x0 |2H . (14)

Figure 1.2 presents 1D and 2D rough profiles and surfaces generated by fBm.
The increments in fBm are stationary but not ergodic. The variance of a fBm
for a large enough array is divergent (i.e., the variance increases with the size of
the array without bounds). Its trace in d dimensions is a self-affine fractal with
a local fractal dimension Df = d + 1 − H . Fractional Brownian motion is not
differentiable at any point, but by smoothing it over an interval one can obtain its
approximate numerical derivative which is called fractional Gaussian noise (fGn),
a 1D example of which is shown in Figure 1.3, which should be compared with its
counterpart in Figure 1.2. We should point out that the correlation function C(r)
of a fBm is given by

C(r) − C(0) ∼ r 2H (15)

so that, as long as H > 0 (which are the only physically-acceptable values of H ),

the correlations increase as r does.
Efficient and accurate generation of a d-dimensional array that follows the statis-
tics of a fBm is not straightforward. Rambaldi and Pinazza (1994) describe a
numerical algorithm based on Eqs. (10) and (11). In addition to their method,
there are at least three other techniques for numerically generating a fBm array
with a given Hurst exponent H (Mehrabi et al., 1997) which we now describe.

1.4.1 The Power-Spectrum Method

A convenient way of representing a stochastic function is through its power spec-
trum S(ω), the Fourier transform of its covariance. The power spectrum of a
12 1. Characterization of Surface Morphology

Figure 1.2. Examples of one- and two-dimensional rough profiles and surfaces generated
by the fractional Brownian motion with various Hurst exponents H .

d-dimensional fBm is given by

ad
S(ω) = d . (16)
( i=1 ωi2 )H +d/2
where ω = (ω1 , · · · , ωd ) is the Fourier-transform variable, and ad is a d-dependent
constant. The spectral representation (16) also allows us to introduce a cutoff length
scale co = 1/ωco such that
ad
S(ω) =  . (17)
2 +
(ωco d 2 H +d/2
i=1 ωi )

The cutoff co allows us to control the length scale over which the spatial properties
of a system are correlated (or anticorrelated). Thus, for length scales L < co
the properties preserve their correlations (anticorrelations), but for L > co they
become random and uncorrelated. Note that the power spectrum of fGn in, for
 1.4. Generation of Rough Surfaces: Fractional Brownian Motion 13

Figure 1.3. An example of one-dimensional fractional Gaussian noise.

example, 1D is given by

bd
S(ω) = , (18)
ω2H −1
where bd is another d-dependent constant. The spectral representation of fBm (and
fGn) provides a convenient method of generating an array of numbers that follow
the fBm statistics, using a fast Fourier transformation (FFT) technique. In this
method, one first generates random numbers, distributed either uniformly in [0,1),
or according to a Gaussian distribution with random phases, and assigns them to
the sites of a d-dimensional lattice. In most cases the linear size L of the lattice is a
power of 2, but the only requirement is that L can be partitioned into small prime
numbers, so that a FFT algorithm can be used. One must also keep in mind that,
since the variance σ 2 of a fBm increases with the size L of the array, generating a
fBm array with a given variance requires selecting an appropriate L. In any case,
the Fourier transformation of the resulting d-dimensional array of the√numbers is
then calculated numerically, the resulting numbers are multiplied by S(ω), and
the results then inverse Fourier transformed back into the real space. The array so
obtained follows the statistics of a fBm with the desired long-range correlations
and the specified value of H . To avoid the problem associated with the periodicity
of the numbers arising as a result of their Fourier transforming, one must generate
14 1. Characterization of Surface Morphology

the array using a much larger lattice size than the actual size that is to be used in
the analysis, and use the central part of the array (or lattice).

1.4.2 Successive Random Additions

In the successive random addition method (Voss, 1985) one begins with the two
end points in the interval [0,1], and assigns a zero value to them. Then Gaussian
random numbers 0 with a zero mean and unit variance are added to these values.
In the next stage, new points are added at a fraction r of the previous stage by
interpolating between the old points (by either linear or spline interpolation), and
Gaussian random numbers 1 with a zero mean and variance r 2H are added to the
new points. Thus, given a sample of Ni points at stage i with resolution λ, stage
i + 1 with resolution rλ is determined by first interpolating the Ni+1 = Ni /r new
points from the old points, and then Gaussian random numbers i with a zero
mean and variance r 2(i−1)H are added to all of the new points. At stage i with
r < 1, the Gaussian random numbers have a variance
σi2 ∼ r 2iH . (19)
This process is continued until the desired length of the data array is reached.
Typically r = 1/2 is used to generate a fBm.
The problem with this method is that the points that are generated in earlier gen-
erations are not statistically equivalent to those generated later. To remedy this, one
can add, during the nth stage of the process, a random Gaussian displacement with
a variance r 2(n−1)H to all of the points. This of course increases the computation
time (it roughly doubles it). Moreover, if one is interested in generating a fBm array
with a very wide range, one may start the process by assigning a Gaussian random
number with a variance 22H to one end of the [0, 1] interval. The generalization
of this method to higher dimensions is straightforward.

1.4.3 The Weierstrass–Mandelbrot Algorithm

In the Weierstrass–Mandelbrot (WM) method (Voss, 1985) one first divides the
interval [0,1] into n − 1 equally-spaced subintervals, where n is the size of the
data array that one wishes to generate, and assigns zero value to all the points in
the interval. Then, to point i at a distance xi from the origin one adds a random
number generated by the Weierstrass function defined by
∞


W(xi ) = Cj r j H sin(2π r −j xi + φj ) (20)
j =−∞

where Cj and φj are random numbers distributed according to Gaussian and

uniform distributions, respectively, and r is a measure of the distance between the
frequencies, which is usually chosen to be small, e.g., r = 0.9. The variance of
Cj is proportional to r 2j H , and the random phases φj are distributed uniformly
on [0, 2π]. Usually, the infinite series in Eq. (20) is approximated by a finite
number of terms, but the number of terms included in the series must be large to
 1.5. Scaling Properties of Rough Surfaces 15

ensure accuracy. For example, in our own work we have used up to 140 terms in
−70 ≤ j ≤ 70 to obtain accurate results. The power spectrum of the data array
generated by the WM method is discrete and does not contain all the frequencies.
However, it is still proportional to ω−(2H +1) , in agreement with Eq. (16).

1.5 Scaling Properties of Rough Surfaces

How do we characterize a rough self-affine surface, either generated synthetically
(numerically) or by a physical process, such as fracturing of a material? We define
a height correlation function Cn (x) by
Cn (x) = |h(x0 + x) − h(x0 )|n 1/n , (21)
where h(x) is the height of the surface at a transverse position x above a reference
surface that can be a smooth, coarse-grained approximation to the rough surface,
and the averaging is over all the initial x0 . The choice of the reference surface can
be tricky. For example, if the rough surface has been grown from a planar substrate,
then a plane parallel to the substrate and in a coordinate system that moves with
the rough surface can be taken to be the reference plane. In any case, it has been
found for many rough surfaces that
Cn (x) = Cn (x)|x|=x ∼ x H (n) , (22)
where the averaging is taken with respect to all the origins x0 in the smooth
reference plane. In most cases, the exponents H (n) take on the same value H for
all n, but there are also some exceptions to this, as discussed by Barabási and Vicsek
(1990).Asurface with correlation function (22) is a self-affine fractal over the range
of length scales in which Cn (x) is computed. Typically, the height correlation
function C2 (x), denoted simply as C(x), has been utilized for estimating H , and
has proven to be a very robust and accurate method.
In practical applications, such as analyzing rough fracture surfaces, the self-
affinity of the surface is bounded by an upper correlation length ξ + and a lower
correlation length ξ − in both the horizontal ( ) and vertical (⊥) directions. That is,
self-affine behavior is restricted to the ranges, ξ − < δx < ξ + and ξ⊥− < δh < ξ⊥+ .
Because of the self-affinity property we must have
 + H
ξ⊥+ ξ
− = . (23)
ξ⊥ ξ−
The correlation function Cn (x) satisfies a general scaling equation given by
Cn (x) = x H Fn (x/ξ + , x/ξ − ). (24)
For x  ξ − , such that x/ξ − → ∞, scaling equation (24) simplifies to
Cn (x) = x H fn (x/ξ + ), (25)
where the scaling function f (y) has the properties that f (y) = c for y  1 and
f (y) ∼ y −H for y  1, where c is a constant of order unity. Hereafter, we delete
the superscripts and use ξ and ξ⊥ for the upper cutoff length scales.
16 1. Characterization of Surface Morphology

If the rough self-affine fractal surface is growing with the process time t as in,
for example, deposition on a flat surface, then one must define a more general
correlation Cn (x, t) in a manner similar to that used for Cn (x), namely,
Cn (x, t) = [|h(x0 + x, t + t  ) − h(x0 , t  )|]n
1/n
, (26)
t .
where the averaging is over all the initial position x0 and times Then, due to
self-affinity of the surface, the correlation function Cn (x, t) has the property that
Cn (bx, bz t) = bα Cn (x, t). (27)
Similar to C2 (x), one usually constructs C2 (x, t) and attempts to extract from it
information about the surface. Under the dynamic conditions in which a rough and
self-affine surface grows, there exists a time scale tc over which the time correla-
tions are important. For rough surfaces that begin growing from a smooth surface,
it has been found in most cases that ξ⊥ and ξ satisfy the following power laws,
ξ⊥ ∼ t β , t  tc (28)
ξ ∼t 1/z
, t  tc (29)
where t is either the time (for a growing rough surface) or the surface’s mean thick-
ness. For t  tc the magnitude of ξ saturates, ξ = L. The quantity z is called
the dynamical exponent of the surface, while β is called the growth exponent. The
quantities ξ⊥ and ξ are actually related to each other by
ξ⊥ ∼ ξ α . (30)
α is called the roughness exponent. Although we are not aware of an experimental
realization of a case for which α and the Hurst exponent H are different, we keep
both α and H to make our discussion as general as possible.
The roughness of a dynamic, growing surface is characterized by the width
w(L) defined as,
1/2
w(L) = [h(x) − hL ]2 , (31)
where h(x) is, as before, the height of the surface at position x, and hL is its
average over a horizontal segment of length L (normalized by the “volume” Ld−1 ).
According to the dynamic scaling theory of Family and Vicsek (1985) for growing
rough surfaces, one has the following dynamic scaling equation
h(x) − hL ∼ t β f (x/t β/α ), (32)
where α and β, the two exponents defined above, satisfy the following scaling
relation
α
α + = 2, (33)
β
and the scaling function f (u) has the properties that |f (u)| < c for u  1, and
f (u) ∼ Lα f (Lu) for u  1, where c is a constant. Note that the ratio α/β can be
replaced by the dynamical exponent z. It is then straightforward to show that
w(L, t) ∼ Lα g(t/Lα/β ), (34)
 1.5. Scaling Properties of Rough Surfaces 17

where g(u) is a universal scaling function. Note also that w(L, t) is a measure of
the correlation length ξ⊥ along the direction of growth. As the rough surface grows,
the wavelength of the spatial fluctuations and the length over which the fluctuations
are correlated both grow with time. However, the length L is the maximum spatial
extent to which the correlations can grow in the d − 1 dimensions along the surface.
When the correlations reach this scale, they cannot extend further, and therefore
the rough surface reaches a steady-state which is characterized by a constant width.
Then, the surface is scale invariant and the saturation value w(L, ∞) is expected
to have a power-law dependence on L:
w(L, ∞) ∼ Lα . (35)
The correlation time tc also scales with L as
tc ∼ Lα/β ∼ Lz , (36)
Equation (34) indicates that, if one plots w/Lα versus t/Lα/β , then, due to the
universality of g(u), all the results for various t and L should collapse onto a single
universal curve [representing the scaling function g(u)]. Figure 1.4 presents such
a data collapse for a rough surface grown by a ballistic deposition process (Vold,
1963). In the simplest version of ballistic deposition, one begins with a line of
L, selects at random a horizontal line above the line of particles, and places a
α

Figure 1.4. Data collapse for a rough surface grown by ballistic deposition (courtesy of
Ehsan Nedaaee Oskoee).
18 1. Characterization of Surface Morphology

Figure 1.5. An example of a rough surface grown by ballistic deposition (courtesy of Ehsan
Nedaaee Oskoee).

particle there. The particle is then allowed to fall along a straight line vertically
downward. When the particle touches the original particles, it sticks to them and
becomes part of the particle pile. A large deposit is then grown by repeating this
procedure. Extensive numerical simulations indicate that the deposit is compact
and non-fractal, but its surface is rough and self-affine. An example is shown in
Figure 1.5.

1.6 Modeling of Growth of Thin Films with Rough

Surface
How can we describe the growth of a rough surface? If the surface is characterized
by a single-valued height function h(x, t), then in general we can describe the
growth of the surface by the following equation
∂h
= R(x, t) + N (x, t), (37)
∂t
where R(x, t) represents all the various (deterministic) physical factors that con-
tribute to the rate of growth of h(x, t), and N represents the noise or randomness in
 1.6. Modeling of Growth of Thin Films with Rough Surface 19

the growth of the rough surface. However, because of various constraints that are
imposed by the physics of growing a rough surface, the set of acceptable functions
R(x, t) is limited. Some of these constraints are as follows (Barabási and Stanley,
1995).
(1) The growth of the surface should be independent of where h = 0 is defined,
i.e., it should be invariant under the transformation h → h + δh. Therefore,
R cannot depend explicitly on h, but should be built from such terms as ∇ n h
(with n = 1, 2, · · ·).
(2) The equation must have rotation and inversion symmetry with respect to the
direction of the growth, implying that it cannot contain odd-order derivatives
in the coordinates, such as ∇h and ∇(∇ 2 h).
(3) The equation must be invariant under time translation t → t + δt, which
means that R cannot depend explicitly on t. It should also be translationally
invariant in the direction perpendicular to the growth direction, and therefore
R cannot contain terms that are explicit in x.
(4) Since the fluctuations in the rough surface must be similar with respect to
the mean position of the surface—the so-called up-down symmetry (h → −h
invariance)—the equation cannot contain terms such as (∇h)n with n being an
even number. However, this symmetry can be broken if there exists a driving
force F, perpendicular to the rough surface, which selects a particular direction
for the growth of the surface. The existence of this driving force is a necessary
but not sufficient condition for breaking this symmetry.
Therefore, the most general form of the equation that describes the growth of a
rough surface is given by
∂h
= ∇ 2 h + ∇ 4 h + · · · + (∇ 2 h)(∇h)2 + · · · + (∇ 2k h)(∇h)2j + N (x, t).
∂t
(38)
To investigate the scaling properties of a growing surface, we consider the
hydrodynamic limit, t → ∞ and x → ∞. In this limit, the higher-order deriva-
tives of h are much smaller than the lowest-order one. Consider, as examples,
∇ 2 h and ∇ 4 h. Writing x → x ≡ bx, we must have h → h ≡ bα h, and thus,
∇ 2 h → ∇ 2 h ≡ bα−2 ∇ 2 h and ∇ 4 h → ∇ 4 h ≡ bα−4 ∇ 4 h. In the limit b → ∞,
∇ 4 h decays much faster than ∇ 2 h and can therefore be neglected.
Given such considerations, the simplest possible equation has the following
form
∂h
= D∇ 2 h + N (x, t), (39)
∂t
which was proposed by Edwards and Wilkinson (1982). In most cases, the noise
term has been assumed to be Gaussian:
N (x, t)N (x , t) = 2Aδ(x − x )δ(t − t  ), (40)
where A is the amplitude of the noise. Equation (40) implies that there is no
correlation in space or time, since the average N (x, t)N (x , t) vanishes (except,
20 1. Characterization of Surface Morphology

of course, at x = x and t = t  ). The Edwards–Wilkinson model, which satisfies the

four constraints described above, can be solved exactly (this is made possible by the
linearity of the equation). One obtains, α = 12 (2 − d), β = 12 α, and hence z = 2.
This model describes the growth of a surface by random deposition of particles
on a growing surface, starting from a flat surface, in which, upon landing on the
growing surface, the particles diffuse on the surface until they find a point with
the lowest height at which they stop. Note that the Edwards–Wilkinson equation
predicts that for d = 2 (growth on a 2D surface) α = 0, which should be interpreted
as implying a logarithmic dependence of the width w on L, i.e., w(L, ∞) ∼ ln L.
The growth of a variety of thin films with rough, self-affine surfaces, such as
those that are formed by ballistic deposition, and the dynamical scaling of the
height and width of such surfaces, are described by the stochastic differential
equation proposed by Kardar, Parisi, and Zhang (KPZ) (1986):
∂h 1
= D∇T2 h + v|∇h|2 + N (x, t), (41)
∂t 2
where v is the growth velocity perpendicular to the surface, and D is a diffusivity.
Equation (41) satisfies the first three constraints listed above, but violates the fourth
constraint since, for example, in ballistic deposition there is lateral growth of the
surface (i.e., the growth occurs in the direction of local normal to the growing
surface), and this is equivalent to having a net driving force F. The lateral growth
is represented by the nonlinear term 12 v|∇h|2 . To see how this term arises, suppose
that a new particle is added to the growing surface. If the surface grows in the
direction of local normal to the surface, then its growth δh is given by, δh =
[(vδt)2 + (vδt∇h)2 ]1/2 = vδt[1 + (∇h)2 ]1/2 . Thus, if |∇h|  1, one must add a
term 12 v(∇h)2 to the Edwards–Wilkinson equation. In the literature one often finds
that σ is used instead of the diffusivity D, and is referred to as a “surface tension,”
since ∇ 2 h tends to smoothen the surface, as does a surface tension. However, we
prefer to use D as the term D∇T2 h represents a diffusion process that arises when
the depositing particles land on the growing surface, diffuse on the surface, and
only stop when they find the point with the lowest height. This diffusion process
also helps smoothen the growing surface (and counter the effect of lateral growth,
represented by the nonlinear term 12 v|∇h|2 , which tends to roughen the surface).
Kardar et al. (1986) considered the case in which the noise was assumed to be
Gaussian with the correlation function (40). For their model, it has been proposed
(Kim and Kosterlitz, 1989; Hentschel and Family, 1991) that for a d-dimensional
surface,
2
α= , (42)
d +2
1
β= , (43)
d +1
and therefore the dynamical exponent z is given by, z = 2(d + 1)/(d + 2). Equa-
tions (42) and (43) are not exact, but provide accurate estimates of α and β (and
hence z). Note that the KPZ equation predicts that z = 2 only when d → ∞.
 1.7. Measurement of Roughness Exponent 21

Another stochastic equation was proposed by Koplik and Levine (1985)

∂h
= D∇T2 h + v + AN (r, h), (44)
∂t
a linear equation in which the term representing the noise is more complex than the
corresponding term in the KPZ equation. For this model, the numerical simulations
indicate that α(d = 2)  3/4, which should be compared with that of the KPZ
surfaces, α = 2/3. The growth of a rough surface can sometimes stop because
it is pinned. To see how the pinning occurs, consider Eq. (44) in zero transverse
dimension:
∂h
= v + AN (h). (45)
∂t
If v > ANmax , where Nmax is the maximum value of N , then ∂h/∂t > 0, and
the surface always moves with a velocity that fluctuates around v. If, however,
v< ANmax , the surface will eventually arrive at a point where v +AN = 0, and
will be pinned down. Therefore, for a fixed v there must be a pinning transition at
some finite value of A.

1.7 Measurement of Roughness Exponent

The numerical value of the Hurst exponent H or the roughness exponent α is not
enough for characterizing the roughness of a surface. It only indicates how the
roughness (or the variance in the height) varies as the transverse length scale, over
which it is measured, changes. A complete characterization of the rough surface
would require not only H or α, but also the amplitudes of the height fluctuations
as well as the transverse correlation lengths. One way of characterizing a rough
surface is by measuring the width w over a segment of size  from the surface.
Then for   ξ we must have
w() ∼ H . (46)
For   ξ we must of course have w = ξ⊥ .
Another method of characterizing a rough surface is by the so-called slit island
method (Mandelbrot et al., 1984). In this method, the rough surface is coated with
another material and then polished carefully parallel to the flat reference surface
(described above) to reveal a series of horizontal cuts. As the coating material is
removed, islands of the surface material appear in a sea of the coating material.
With further removal of the coating material, the islands will grow and merge. If
we consider a region of linear size  and height fluctuations h = h(x) − h ,
the distribution P (h) can be described by the following scaling law
P (h) = w()−1 f [h/w()], (47)
where w() is the width of the region. Since w() follows Eq. (46), the implication
is that the density ρ() in a cross-section of size  is given by
ρ() ∼ −H . (48)
22 1. Characterization of Surface Morphology

Equation (48) suggests that the interface between the two materials, i.e., between
the rough surface and the coating material, in the cross-sections parallel to the
reference plane is a self-similar fractal with a fractal dimension
Df = d − H, (49)
where d is the Euclidean dimensionality of the reference surface. Therefore, if the
fractal dimension Df can be estimated independently, then the Hurst exponent
H can also be evaluated. Typically, the islands that appear have a surface area
distribution nS such that
nS ∼ S −τ , (50)
where nS is the number of islands with areas S in the range [S − 1
2 S, S+ 1
2 S].
The exponent τ is related to the fractal dimension Df through the following
equation
1 
τ= Df + d , (51)
d
so that measurement of the islands’areas yields Df , from which the Hurst exponent
H can be estimated.
The third method of analyzing a rough, self-affine surface is through its power
spectrum which, in d dimensions, is given by Eq. (16). However, as Hough (1989)
pointed out, interpreting a power-law power spectrum is not without difficulties,
and thus one must be careful in using such an analysis. In particular, a power-law
power spectrum might also be the result of a non-stationary and non-fractal system.
We will come back to this issue in Chapters 6 and 7, where we describe fracture
surface of materials which are typically very rough.

Summary
An important characteristics of morphology of disordered multiphase materials
is the structure of their surface, and in particular their surface roughness. The
concepts of modern statistical physics of disordered media can now quantify the
roughness in terms of self-affine fractals, and the roughness or Hurst exponent.
The dynamics of growth of such surfaces can also be described by dynamical
scaling, discrete models of material growth, and suitable continuum differential
equations. Moreover, fractal geometry, and the associated power-law correlation
functions, point to the fundamental role of length scale and long-range correlations
in the macroscopic homogeneity of a heterogeneous material. If the largest relevant
length scale of the material, e.g., its linear size, is less than the length scale at
which it can be considered homogeneous, then the classical equations that describe
transport processes in the material must be fundamentally modified.
 Part I

Effective Properties of
Heterogeneous Materials with
Constitutive Nonlinearities
2
Nonlinear Conductivity and Dielectric
Constant: The Continuum Approach

2.0 Introduction
The main focus of Volume II is on nonlinear properties of heterogeneous mate-
rials. In general, there are two fundamental classes of nonlinearity that one may
encounter in disordered materials:
(1) One class of nonlinear materials is described by what we call constitutive non-
linearity, which is one in which the basic local constitutive law that expresses
the relation between the flux (of current, force, etc.) and the potential (volt-
age, stress, etc.) gradient is nonlinear. As a result, the macroscopic behavior
of such materials must also be described by nonlinear transport equations. In
particular, the effective transport properties of such materials are nonlinear
in the sense of being functions of the external potential gradient. Such mate-
rials are of great practical importance, since, for example, one may be able
to design new nonlinear optical materials by tuning their nonlinear response
which can be achieved by, for example, changing the volume fraction of their
constituents. For example, it has been suggested that strong local field effects,
such as the large local field at the surface plasmon resonance frequency of a
metallic inclusion, may lead to enhanced nonlinear response in a heteroge-
neous material. Constitutive nonlinearity is the subject of this and the next
two chapters. Even within this restricted class of nonlinear materials, one may
imagine a very large number of nonlinear constitutive equations (similar to
those that have been proposed, for example, for polymeric fluids). Therefore,
while we describe in this chapter results for general constitutive nonlinearity,
their application is restricted mostly to strongly nonlinear materials, i.e., those
that are described by a power-law relation between the flux and the current.
In the next two chapters we will also describe the macroscopic behavior of
nonlinear materials that can be described by a few other types of nonlinear
constitutive equations, for which considerable progress has been made, and a
comparison between the theoretical predictions and the experimental data is
possible.
(2) In the second class of nonlinearities, a material is characterized by thresholds
in the (local as well as macroscopic) potential gradient. Then, depending on
the physics of the phenomenon under study, one of the following two scenarios
may arise.
26 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

(i) The transport properties of the material vanish below the threshold, but
above the threshold the material behaves linearly (or, possibly, exhibits
constitutive nonlinearity) and possesses non-zero effective transport prop-
erties. For example, consider a resistor network in which each bond is
insulating if the voltage drop between its two ends is less than a threshold
value, but becomes conducting (either linearly or nonlinearly) if the volt-
age drop exceeds a threshold. An example of a material to which such a
model is directly relevant is foam. As described in Chapter 9 of Volume I,
foams behave both as solid materials (in the sense of exhibiting an elas-
tic response when exposed to an external stress or strain), and as a fluid
when the applied stress that they are exposed to reaches a threshold value.
Therefore, foams do not flow if the stress applied to them is less than the
threshold. As a result, if we consider, for example, flow of foams in a
porous medium (which is usually modeled as a network of tubes), there
would be no macroscopic flux of foams unless the pressure gradient ap-
plied to the porous medium exceeds a threshold. We must, however, point
out that this type of threshold behavior is not the same as that of a percola-
tion system below and above the percolation threshold, i.e., this threshold
behavior is not a geometrical effect, although, as we will show in Chapter
3, there are certain similarities between the two types of phenomena.
(ii) The second scenario arises when the material behaves linearly (or, pos-
sibly, exhibits constitutive nonlinearity) if the applied potential gradient
is less than a threshold, but exhibits highly nonlinear properties when the
threshold is exceeded. Well-known examples of this type of phenomenon
are brittle fracture and dielectric breakdown of solid materials, phenomena
that will be studied beginning with Chapter 5.

Compared to linear systems, the number of studies in which an attempt has been
made to obtain estimates of the effective nonlinear properties is small. This is par-
ticularly true in the context of continuum models of disordered materials. Discrete
models have received much more attention, and will be described and discussed
in Chapter 3. To our knowledge, Marcellini (1978) was perhaps the first to un-
dertake a systematic study of effective transport properties of nonlinear materials,
and attempted to estimate their effective dielectric constant. He considered a two-
phase composite in which one phase had a constant dielectric constant, while the
dielectric constant of the second phase, that consisted of spherical inclusions, was
a function of the local electric field. The particles were arranged either at random
or in a periodic manner, similar to the periodic models that were described and
analyzed in detail in Chapter 4 of Volume I. Miksis (1983) obtained slightly more
general results for the effective properties of periodic arrays, and random distribu-
tions of nonlinear spherical inclusions in a linear matrix. The methods of Marcellini
and Miskis were more or less straightforward generalization of those described
in Chapter 4 of Volume I, and hence need not be described again. Willis (1986)
applied the approach of Hill (1963) (see below; see also Chapter 7 of Volume I
for more details) to nonlinear dielectrics. In terms of deriving rigorous bounds for
 2.1. Variational Principles 27

the effective nonlinear electrical conductivity and dielectric constant, Talbot and
Willis (1985, 1987, 1994) and Willis (1986) proposed extensions of the Hashin–
Shtrikman variational principles (Hashin and Shtrikman, 1962a,b, 1963) (see also
Chapters 4 and 7 of Volume I) to nonlinear heterogeneous materials. In a series
of papers, Ponte Castañeda and co-workers (Ponte Castañeda, 1992b, 1998; Ponte
Castañeda and Kailasam, 1997) analyzed the effective nonlinear conductivity and
dielectric constant of two-phase heterogeneous materials using two different tech-
niques. One of the methods is exact to first-order in contrast between the properties
of the two phases, and is capable of delivering rigorous lower bounds and approx-
imate estimates for the upper bounds (not the upper bounds themselves), while
the second method is exact to second order in the contrast between the phases’
properties. To our knowledge, their work is the most advanced attempt in the area
of continuum description of the effective nonlinear conductivity and dielectric
constant of disordered materials, and is described in detail in this chapter.

2.1 Variational Principles

Volume I of this book should have made it abundantly clear that the effective
linear properties of heterogeneous materials are not characterized by simple aver-
ages of the properties of the constituent phases, weighted, for example, by their
respective volume fractions. In fact, in addition to the volume fractions, the effec-
tive properties depend in general on certain microstructural parameters which are
themselves functions of the volume fractions. The same is true about nonlinear
effective properties of disordered materials. Moreover, due to the nonlinearity, a
proper definition of the effective properties is even more important than the linear
case because, for example, nonlinear effective properties may exhibit sensitive
dependence on the boundary conditions.
Consider a heterogeneous dielectric material that occupies a region  in space.
The nonlinear constitutive behavior of the material may be characterized in terms
of an electric energy-density function w(x, E) which depends on the position x
and the electric field E(x), such that the electric displacement field D(x) is given
by
∂
D(x) = w(x, E). (1)
∂E
Furthermore, if one assumes local isotropy, then, w(x, E) = e(x, E), where e :
 × R → R is continuous, convex and coercive (in the sense that, w → ∞ as
E → ∞) which satisfies the conditions, e(x, E) ≥ 0 and e(x, 0) = 0. Here, R is
the set of the extended real numbers.
The effective constitutive behavior of the heterogeneous material is then defined
by
∂
D = He (E), (2)
∂E
28 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

where · · · denotes an spatial average. We should keep in mind that the effective
behavior of the heterogeneous material, as characterized by the energy functional
He (E), may in general be anisotropic, even if the material’s phases themselves
are isotropic. In principle, He (E) is determined by solving the usual electro-
static problem on , defined by, ∇ × E = 0, and ∇ · D = 0, subject to a uniform
boundary condition, ϕ = −E · x on the external surface of , where ϕ is the
electrostatic potential defined by, E = −∇ϕ(x) in . This boundary condition
ensures that the average of the electric field is in fact E, in the sense that

E = E(x) dx. (3)

Moreover, the average displacement field is defined by a similar relation:

D = D(x) dx, (4)

so that one obtains the effective energy He that evaluates the pertinent energy
functional for the heterogeneous material,

He (E) = w(x, E) dx, (5)

at the actual electric field solving the electrostatic problem for a given microstruc-
ture. Due to the complexity of the morphology of real materials, it is not practical
to solve the electrostatic problem. For this reason, variational formulations of
the problem based on the minimum energy and minimum complementary-energy
principles provide useful alternative routes for analyzing the problem. Thus, let
us state these principles here (which were also utilized in Volume I for obtaining
estimates of effective linear properties).
According to the minimum energy principle, expressed in terms of the energy
functional H, one can obtain the following expression for the effective energy He
of a heterogeneous material,
He (E) = min H(E), (6)
E∈S1

where
S1 = {E|E = −∇ϕ(x) in , and ϕ = −E · x on ∂}. (7)
Note that, to guarantee the existence of the minimizer (6), certain conditions on the
behavior of w (or e) as E → ∞ are required, which is why one assumes that w is
coercive. Moreover, strict convexity of He guarantees uniqueness of the solution,
convexity of w ensures that of He , and if the fields are smooth enough, Eq. (6)
will be equivalent to the original electrostatic problem defined above.
The second characteristic of the heterogeneous material is obtained from its
complementary-energy function Hec , defined in terms of the principle of minimum
complementary energy:
Hec (D) = min Hc (D), (8)
D∈S2
 2.1. Variational Principles 29

where

Hc (D) = w ∗ (x, D) dx (9)

is the complementary energy functional, expressed in terms of
w∗ (x, D) = sup{E · D − w(x, E)}, (10)
E
with
S2 = {D|∇ · D = 0 in , and D · n = D · n on ∂} (11)
being the set of admissible electric displacement fields. Note that, if Eq. (3) is
reinterpreted as a definition for the average electric field, then, one has
∂
E = Hc (D). (12)
∂D e
In general, it can be shown that
Hec (E) ≥ Hec∗ (E). (13)
The reason for the inequality (13) is related to the fact that definitions of H and
Hc correspond to different boundary conditions on the heterogeneous material
(Dirichlet versus Neumann conditions), hence leading to generally distinct effec-
tive energies. However, the strict equality holds in (13) if the composite can be
homogenized, in the sense that it can be considered as homogeneous on a large
enough scale. Finally, note that
w ∗ (x, D) = e∗ (x, D), (14)
where e∗ is the convex polar function (Legendre transform) of e and D is the
magnitude of D.
Ponte Castañeda (1992b) proposed new variational principles in order to obtain
upper and lower bounds and estimates for the effective energy functions of non-
linear materials. These variational principles are equivalent to the standard ones
described above, under appropriate hypothesis on the energy-density function.
The new variational principle is based on a change of variables r = h(E), with
h : R + → R + (R + is the set of non-negative reals) given by h(E) = E 2 . One
than obtains a function f :  × R + → R + , such that
f (x, r) = e(x, E) = w(x, E), (15)
has the same dependence on x as e and w, and that it is continuous and coercive (but
not necessarily convex) in r. Moreover, f is a non-negative function satisfying,
f (x, 0) = 0. Then, if we define the Legendre transform (convex polar) of f by
f ∗ (x, p) = sup{rp − f (x, r)}, (16)
r≥0

it follows that
f (x, r) ≥ sup{rp − f ∗ (x, p)}. (17)
p≥0
30 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

Note that x is fixed in (16) and (17), and that the suprema are evaluated over the
sets of non-negative r and p, respectively. In addition, the right-hand side of (17)
is the bipolar of f , which has the geometric interpretation of the convex envelope
of f , and hence the inequality. The equality in (17) is achieved if f is convex and
continuous in r. Therefore, assuming that the energy function w in (16) is such
that f is convex (note that convexity of f implies that of w), one obtains from (17)
the following representation for the local energy density function of the nonlinear
heterogeneous material,
w(x, E) = sup {w 0 (x, E) − v(x,  0 )}, (18)
 0 ≥0

where [from Eq. (16)]

v(x,  0 ) = sup{w 0 (x, E) − w(x, E)}, (19)
E

where p has been identified with 12  0 and r with E 2 , in such a fashion that
w0 (x, E) = 12  0 (x)E 2 and v(x, 12  0 ) = f ∗ (x,  0 ). Thus, w0 corresponds to the
local energy-density function of a linear, heterogeneous comparison material with
arbitrary (but not necessarily constant) non-negative dielectric constant  0 (x). The
minimum energy formulation of the variational principle follows by making use
of the representation (18) in the classical minimum energy principle, and inter-
changing the order of the infimum in (16) and the supremum in (18). The result
(Ponte Castañeda, 1992b) is the following theorem.
Theorem 1: Suppose that the local energy-density function w of a given nonlin-
ear heterogeneous material with isotropic phases satisfies condition (15) with
f a non-negative, continuous, coercive and convex function of r = E 2 , with
f (x, 0) = 0. Then, the effective energy function of the nonlinear heterogeneous
material He is determined by the variational principle,
He = sup {He0 (E) − V ( 0 )}, (20)
 0 (x)≥0

where

V ( ) =
0
v[x,  0 (x)] dx, (21)

and He0 is the effective energy function of a linear heterogeneous comparison
material with local energy function w 0 , such that

He0 = min w 0 (x, E) dx. (22)
E∈S1 

The complementary-energy formulation of the new variational principle follows

in a similar fashion from the change of variables s = h(D), where h is the same
function as before which induces a function g :  × R + → R + , such that
g(x, s) = e∗ (x, D) = w∗ (x, D), (23)
where g is continuous and coercive in s, and is a non-negative function such that,
 2.1. Variational Principles 31

g(x, 0) = e∗ (x, 0) = 0. Then, if one defines the concave polar of g by

g∗ (x, q) = inf {sq − g(x, s)}, (24)
s≥0

it follows that
g(x, s) ≤ inf {sq − g∗ (x, q)}, (25)
q≥0

with the equality holding true if g is concave. Assuming then that the complemen-
tary energy density function w ∗ of the nonlinear heterogeneous material is such
that g is concave, it follows from (25) that
w∗ (x, D) = inf {w 0∗ (x, D) + v(x,  0 )}, (26)
 0 ≥0

where q has been identified with (2 0 )−1 and s with D 2 , such that w 0∗ (x, D) =
[ 12  0 (x)]D 2 is the complementary-energy function of the linear, heterogeneous
comparison material with arbitrary non-negative dielectric coefficient  0 (x), and
v(x,  0 ) = g ∗ (x, 12  0 ). Given these, one can state the following theorem (Ponte
Castañeda, 1992b)
Theorem 2: Suppose that the (convex) local complementary-energy function
w∗ of a given nonlinear heterogeneous material with isotropic phases satisfies
condition (23) with g being a non-negative, continuous, coercive and concave
function of s = D 2 , and g(x, 0) = 0. Then, the effective complementary-energy
function Hec of the nonlinear heterogeneous material is given by
Hec (D) = inf {He0c (D) + V ( 0 )}, (27)
 0 (x)≥0

where

He0c (D) = min w 0∗ (x, D) dx (28)
D∈S2 

is the effective complementary-energy function of the linear comparison material.

Note that without the hypotheses of convexity of f and concavity of g the
equivalence between the classical minimum energy and the new variational prin-
ciples would not hold. It can be shown that concavity of g implies convexity of
f . Moreover, recall that, so far, it has only been assumed explicitly that w is
convex and coercive. Since concavity of g implies convexity of f , it implies in
turn that w ≥ αE 2 (α > 0) as E → ∞. Thus, a sensible condition may be that,
w(x, E) ∼ E 1+n (n ≥ 1) as E → ∞. Then, f is stronger than, or at least as strong
as, affine at infinity, consistent with its convexity. On the other hand, the above
assumption for w implies that w ∗ (x, D) ∼ D 1+1/n as D → ∞, and therefore g is
weaker than, or at least as weak as, affine at infinity, consistent with its concav-
ity. Other conditions are possible, but the bounds and estimates that are derived
below may require reinterpretation, if the conditions are different. For example, if
one lets n in the above conditions be such that 0 < n ≤ 1, then, the suprema and
infima in the above relations would have to be replaced by infima and suprema,
respectively.
32 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

2.2 Bounds on the Effective Energy Function

One can now determine bounds and estimates for the effective energy functions
of nonlinear heterogeneous materials that are characterized by some appropriate
statistical data on their morphology. The main idea of Ponte Castañeda (1992b)
is to make use of corresponding bounds and estimates for linear heterogeneous
comparison materials, which were described in detail in Chapters 4 and 7 of Volume
I, such as the Wiener one-point bounds, the Hashin–Shtrikman two-point bounds,
and the Beran three-point bounds, in order to derive the corresponding results for
the nonlinear materials. The linear comparison material has the same morphology
as the nonlinear composite. In particular, consider heterogeneous materials with
N homogeneous phases, characterized by the isotropic energy functions ei (i =
1, · · · , N), such that the local energy function w of the heterogeneous material is
given by


N
w(x, E) = mi (x)ei (E), (29)
i=1
where mi (x) is the exclusion indicator function of phase i defined by, mi (x) = 1 if
x is in phase i, and mi (x) = 0 otherwise. The volume fractions φi of the constituent
phases are assumed fixed and given by

φi = mi (x) dx. (30)

Before proceeding with the determination of the bounds and estimates, the follow-
ing useful corollaries to Theorems 1 and 2 must be stated. Their proofs (which are
simple) are given by Ponte Castañeda (1992b).
Corollary 1: Suppose that Eq. (29) characterizes the local energy-density func-
tion of a N -phase nonlinear composite, satisfying the hypotheses of Theorem 1.
Then, the effective energy function He of the composite satisfies the inequality
 

N
He (E) ≥ sup He0 (E) − φi vi (i0 ) , (31)
i0 >0 i=1

where He0 is the effective energy function of a linear comparison material with N
phases of dielectric constants i0 with volume fractions φi , such that the effective
dielectric constant e0 of the comparison composite is given by


N
e0 (x) = mi (x)i0 . (32)
i=1
The function vi is given by Eq. (19), written for the ith phase, and the supremum
in (31) is evaluated over the set of constants i0 (i = 1, · · · , N).
Corollary 2: Suppose that the appropriate complementary version of (29)
characterizes the local complementary energy function w∗ of a N-phase non-
 2.2. Bounds on the Effective Energy Function 33

linear composite, satisfying the hypotheses of Theorem 2. Then, the effective

complementary-energy function Hec satisfies
 

N
He (D) ≤ inf He (D) +
c 0c 0
φi vi (i ) , (33)
i0 i=1

where He0c is the effective complementary-energy function of a linear comparison

composite with N phases of dielectric constants i0 and volume fractions φi , such
that the effective dielectric constant of the comparison composite is given by


N
e0 (x) = mi (x)i0 . (34)
i=1

2.2.1 Lower Bounds

Similar to the effective linear conductivity and dielectric constant of disordered
materials described in Chapter 4 of Volume I, we can now derive one-, two- and
three-point bounds for the effective nonlinear conductivity and dielectric constant
of materials. What follows is a description of derivation of such bounds.

2.2.1.1 One-Point Bounds

Consider utilizing the one-point lower bound of Wiener (1912) for linear,
anisotropic materials described in Chapter 4 of Volume I for generating a corre-
sponding bound for nonlinear, anisotropic composites. Recall that the bounds are
given by, g(r)−1  ≤ ge ≤ g(r). Although these bounds are not very sharp, their
derivation for nonlinear materials provides a useful demonstration of utility of the
variational principles of Ponte Castañeda (1992b), described above, for deriving
rigorous bounds which will then be used in order to derive the Hashin–Shtrikman
and Beran bounds. The Wiener lower bound may be specified as a bound on the
effective energy functions of linear composites with dielectric constants i0 and
volume fractions φi (with i = 1, · · · , N) via the relation
N −1
1
φi
He ≥
0
0
E2 , (35)
2 
i=1 i

where He0 = 12 ( 0e E) · E is the effective energy function of the linear material
with effective dielectric tensor  0e . The nonlinear Wiener lower bound for the
effective energy functions He of the nonlinear materials is obtained by applying
Eq. (31) to the set of nonlinear composites with given phase volume fractions,
and combining the result with the lower bound (35) for the corresponding linear
comparison materials. The result is
⎧  −1 ⎫
⎨1
N
φi
N ⎬
He ≥ sup E 2
− φ v (
i i i
0
) , (36)
⎩
 0 >0 2 i0 ⎭
i i=1 i=1
34 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

with
 
1 0 2
vi (i0 ) = sup i s − ei (s) . (37)
s>0 2

Clearly, the number of optimizations implicit in (36) and (37) is 2N , but this
number may be significantly reduced by using the identity,
N −1 N 

φi

0
= inf φi i (1 − ωi ) ,
2
(38)
 ωi
i=1 i i=1

where the infimum is over the set of variables ωi (i = 1, · · · , N) which are sub-
ject to a zero-average constraint, i.e., ω = N i=1 φi ωi = 0. This identity, when
applied to the nonlinear lower bound for He in (36), yields
 N 

1
He ≥ sup inf φi i (1 − ωi ) E − vi (i )
2 2 0
, (39)
 0 >0
ωi 2
i i=1

which in turn leads to

⎧ ⎫
⎨
N  ⎬
1
He ≥ inf φi sup i (1 − ωi )2 E2 − vi (i0 ) . (40)
ωi ⎩ i0 >0 2 ⎭
i=1

In (40), the saddle point theorem and the fact that the argument of the nested
supremum and infimum is concave in i0 (since the functions vi are convex in
i0 ) and convex in ωi have been used in order to justify the interchange of the
supremum and infimum operations. Finally, application of Eq. (18), specialized to
each phase in the form
 
1 0 2
ei (s) = sup i s − vi (i0 ) , (41)
 0 >0 2 i

leads to
He ≥ inf {φi ei (|1 − ωi |E)} , (42)
ωi

which is much simpler than the bounds (36) and (37), as it involves only a
N-dimensional optimization, with one linear constraint, which can easily be em-
bedded in the optimization operation by suitable relabelling of the optimization
variables. For example, for a two-phase material, bound (42) becomes
He ≥ inf {φ1 e1 (|1 − φ2 ω|E) + φ2 e2 (|1 + φ1 ω|E)}, (43)
ω

where the optimization variable ω is now unconstrained.

2.2.1.2 Two-Point Bounds

One can now use the same technique to derive the Hashin–Shtrikman-type bound
for nonlinear isotropic materials. To do this, one should first note that the effective
 2.2. Bounds on the Effective Energy Function 35

dielectric tensor of a linear isotropic heterogeneous material is isotropic (i.e.,  0e =

e0 U, where U is the identity tensor). Then, the Hashin–Shtrikman lower bound
(l) (l)
e for the effective dielectric constant e0 , satisfying e0 ≥ e is given by the
expression (see also Chapters 4 and 7 of Volume I)
N −1

φi
e = inf
(l)
− (d − 1) (l) , (44)
ωi  + (d − 1)
i=1 i
0 (l)

where  (l) = inf s {s0 }. Equation (44), which is subject to the constraint that, ω =
0, may be rewritten as
N 

 
e = inf
(l)
φi i (1 − ωi ) + (d − 1) ωi
2 (l) 2
. (45)
ωi
i=1

Observe that the effective energy functions He of the macroscopically-isotropic,

nonlinear materials can be estimated from relation (31), where He0 now repre-
sents the effective energy function of the linear comparison materials with phases
of dielectric constants i0 and volume fractions φi . Note that, while not all mi-
crostructures that are isotropic for linear materials are also isotropic in the nonlinear
context, nonlinear isotropic microstructures must also be isotropic in the linear con-
text. Therefore, a lower bound for the effective energy function of linear, isotropic
comparison materials is also a lower bound for the subclass of linear comparison
composites with “nonlinearly isotropic” microstructure. Hence, replacing He0 in
(31) by the lower bound given by (45) generates a lower bound for the nonlinear
isotropic composites, with the result being
 N 

1
He ≥ sup inf φi (i (1 − ωi ) + (d − 1) ωi )E − vi (i )
2 (l) 2 2 0
i0 >0
ωi 2
i=1
⎧ N ⎫
⎨
1   ⎬
= inf sup φi i (1 − ωi )2 + (d − 1) (l) ωi2 E2 − vi (i0 ) ,
ωi ⎩ 0 2 ⎭
 >0
i i=1

(46)
where the saddle point theorem has been used to justify interchanging the
supremum and infimum operations. Then, using (41), one obtains
⎧ ⎧
⎨ ⎨
N
He ≥ min inf φi ei (|1 − ωi |E)
s ⎩ ωi ⎩
i=1,i
=s
 ⎫⎫
 ⎬⎬
 1

N
+ φs es (1 − ωs )2 + (d − 1) φj ωj2 E , (47)
φs ⎭⎭
j =1

which represents the Hashin–Shtrikman lower bound for nonlinear isotropic ma-
(l)
terials with isotropic phases of given volume fractions, and is denoted by HHS .
36 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

For a two-phase material, the nonlinear lower bound reduces to

 ! 
(l) inf ω {φ1 e1 (|1 − φ2 ω|E) + φ2 e2 (1 + φ1 ω)2 + (d − 1)φ1 ω2 E}
HHS (E) = min ! .
inf ω {φ1 e1 (1 − φ2 ω)2 + (d − 1)φ2 ω2 E + φ2 e2 (|1 + φ1 ω|E)}
(48)
Note that for a two-phase nonlinear material, the bounds given above involve only
one optimization. Moreover, the method described here has a distinct advantage in
that, it utilizes the linear heterogeneous comparison material in conjunction with
its linear bounds and estimates (other than, for example, the Hashin–Shtrikman
bounds) to yield the corresponding nonlinear bounds and estimates.

2.2.1.3 Three-Point Bounds

As another illustration of this feature of the method, the lower bounds of the
Beran-type for two-phase, nonlinear isotropic materials are derived which, as
discussed in Chapters 4 and 7 of Volume I, are generally tighter than the Hashin–
Shtrikman bounds except, of course, for those microstructures for which the
Hashin–Shtrikman bounds become exact estimates, such as the coated-spheres
model (see Sections 4.4 and 7.2.3 of Volume I). As discussed in Chapters 4 and
7 of Volume I, the Beran bound (Beran, 1965), simplified by Milton (1981a,b),
depends on the volume fraction of the phases and on one additional microstructural
parameter ζi , and is given by
 2 −1

φ i
e(l) = 0 + (d − 1) (l)
− (d − 1) (l) , (49)
i=1
i

which is identical in form to (44), except that  (l) is now given by

 2 −1

ζi
 (l) = , (50)
i=1 i
0
where the third-order microstructural parameters ζ1 and ζ2 = 1 − ζ1 are both in
the interval [0,1], and were described in detail in Chapters 4 (see Sections 4.5.2
and 4.5.3) and 7 of Volume I (see section 7.4.3). Substituting (49) into the lower-
bound approximation (31) and following a procedure very similar to that used for
the Hashin–Shtrikman bound, one arrives at the following lower bound for the
nonlinear energy function,
 "
(l)
HB (E) = inf φ1 e1 (1 − φ2 ω)2 + (d − 1)φ2 ζ1 ω2 (1 − ζ2 γ )2 E
ω,γ
" 
+ φ2 e2 (1 + φ1 ω)2 + (d − 1)φ1 ζ2 ω2 (1 + ζ1 γ )2 E . (51)

Note that the corresponding nonlinear Hashin–Shtrikman lower bound follows

immediately from (51) by choosing either ζ1 = 0 or ζ2 = 1, whichever yields
the lowest value (note also that the infimum problem over γ becomes trivial in
 2.2. Bounds on the Effective Energy Function 37

either case), which is completely analogous to the corresponding result for linear
two-phase materials.

2.2.2 Approximate Estimates of the Effective Energy

Although the above developments were for the effective dielectric constant of
nonlinear materials, they are equally applicable to the problem of estimating their
nonlinear conductivity. We will discuss this problem in detail later in this chapter,
but it is useful to note here the work of Gibiansky and Torquato (1998a). They
wrote Eq. (51) in a more general form
 "
He (E) = inf φ1 e1 (1 − φ2 ω)2 + (d − 1)φ2 ζ1 ω2 (1 − ζ2 γ )2 + Bφ2 ζ1 ζ2 ω2 γ 2 E
ω,γ
" 
+ φ2 e2 (1 + φ1 ω)2 + (d − 1)φ1 ζ2 ω2 (1 + ζ1 γ )2 E (52)

which must be optimized over the two scalar variables ω ∈ (−∞, ∞) and γ ∈
(−∞, ∞). The optimization can be carried out either analytically, if the energy
functions of the nonlinear phases are sufficiently simple, or numerically. Here, B
is a parameter which is given by (Torquato, 1985a,b)
(d − 1) − ζ2
B = (d − 1) . (53)
1 − (d − 1)ζ2
We can now consider two important limiting cases.

2.2.2.1 Conductor–Superconductor Composites

If we assume that the inclusion phase 2 is a superconducting material, i.e., if

0, if E = 0,
e2 (E) = (54)
∞, if E
= 0,
then, for such a composite, the right-hand side of Eq. (52) will be divergent unless
the argument of the function e2 is equal to zero, i.e., unless
"
(1 + φ1 ω)2 + (d − 1)φ1 ζ2 ω2 (1 + ζ1 γ )2 E = 0,
which is possible (for d
= 1) only if,
ω = −(φ1 )−1 , γ = −(ζ2 )−1 , (55)
which represent the optimal values of these parameters.An approximate expression
for the effective energy of the nonlinear material is then obtained:
# 
ζ1 + (d − 1)φ2 + Bφ2 ζ2
He (E) = φ1 e1 E . (56)
ζ1 φ12
Therefore, if, for example, the matrix is a strongly nonlinear material with the
(n) (n)
energy function, e1 = g1 E n /n, and if the effective nonlinear conductivity ge
38 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

(n)
is defined by, He (E) = ge En /n, one obtains
 n/2
(n)
ge ζ1 + (d − 1)φ2 + Bφ2 ζ2
= φ1 . (57)
(1)
g1 ζ1 φ12
Equations (56) and (57) can now be utilized for estimating the effective energy
of nonlinear composites with superconducting inclusions, provided that the ap-
propriate expressions for the microstructural parameters ζ1 and ζ2 = 1 − ζ1 are
available, a matter that was discussed in detail in Section 4.5.3 of Volume I. In
particular, it can be shown that Eq. (56) provides an estimate of the effective energy
which is always larger than the estimates provided by Eqs. (48) and (51), hence
satisfying these rigorous bounds.

2.2.2.2 Conductor–Insulator Composites

Consider now the opposite limit in which the inclusion phase is insulating, so
that e2 (E) = 0 for all E. Then, the optimal values of ω and γ are obtained by
minimizing
"
(1 − φ2 ω)2 + (d − 1)φ2 ζ1 ω2 (1 − ζ2 γ )2 + Bφ2 ζ1 ζ2 ω2 γ 2 E,
with respect to these parameters. It is straightforward to show that the optimal
values are given by
B + (d − 1)ζ2 d −1
ω= , γ = , (58)
Bφ2 + (d − 1)(Bζ1 + φ2 ζ2 ) B + (d − 1)ζ2
which then lead to
# 
(d − 1)Bζ1
He (E) = φ1 e1 E . (59)
Bφ2 + (d − 1)(Bζ1 + ζ2 φ2 )
For a strongly-nonlinear (power-law) matrix, the effective conductivity of the
composite is then given by
(n)  n/2
ge (d − 1)Bζ1
= φ1 . (60)
(n)
g1 Bφ2 + (d − 1)(Bζ1 + ζ2 φ2 )
Let us mention that Eqs. (56), (57), (59) and (60) are accurate only if the inclusion
phase does not form large clusters.

2.2.3 Upper Bounds and Estimates

The derivation of upper bounds for the effective energy functions of nonlinear
materials is intrinsically more difficult than the corresponding lower bounds. This
is because approximations such as (31) do not work in this case. While it is possible
to derive the Wiener upper bound, derivation of upper bounds of the Hashin–
Shtrikman- and Beran-type bounds has proven to be very difficult. Instead, one
may obtain upper estimates or, more precisely, lower estimates for the upper bound,
of the Hashin–Shtrikman- and Beran-types.
 2.2. Bounds on the Effective Energy Function 39

The derivation of the Wiener upper bound is accomplished by the corresponding

upper bound for linear materials with an arbitrary dielectric constant  0 (x), and is
given by
 
1
He0 (E) ≤  0 (x)dx E2 . (61)
2 
Then, application of (61) to (20) leads to
    
1
He (E) ≤ sup  0 (x)dx E2 − v[x,  0 (x)]dx
 0 (x)≥0 2  
   
1 0
= sup  E2 − v(x,  0 ) dx = e(x, E)dx, (62)
  0 ≥0 2 

which, via (29), leads to the nonlinear Wiener upper bound

(u)


N
HW ≤ φi ei (E). (63)
i=1
The determination of an estimate for the Hashin–Shtrikman upper bound, or
the upper estimate, is accomplished by application of approximation (31) to the
Hashin–Shtrikman upper bounds for the linear comparison material. The upper
bound for the effective energy function of the linear comparison material may be
given in terms of the upper bound for its effective dielectric constant:
N −1

φ
e+ = − (d − 1) + ,
i
0 + (d − 1) +
(64)
i=1
i

where  + = supi {i0 }. The procedure that utilizes the lower bound (44) for the
linear comparison material to obtain a lower bound for the nonlinear material may
now be repeated. To derive the upper estimates, one utilizes (64) instead of (44),
in which case the result would be the same as (47) and (48) for the N -phase and
two-phase nonlinear materials, respectively, with the difference that the outermost
minimum operations must now be replaced by maximum operations. The result,
+
denoted by HHS , is referred to as the Hashin–Shtrikman upper estimate. However,
+
as shown below, HHS is not, in general, an upper bound for He .
The same arguments and analyses also apply to the Beran upper bounds. Hence,
one can obtain upper estimate for the Beran-type bounds. If


2
+
 = ζi i0 , (65)
i=1
then, the corresponding result for the upper estimate (which, in general, is not an
upper bound) for nonlinear isotropic materials is given by
 "
+
HB (E) = inf φ1 e1 (1 − φ2 ω)2 + (d − 1)φ2 ζ1 ω2 E
ω
" 
+ φ2 e2 (1 + φ1 ω)2 + (d − 1)φ1 ζ2 ω2 E . (66)
40 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

2.3 Exact Results for Laminates

Having derived the rigorous lower bounds and also the lower estimates for the
upper bounds, two important issues must be now addressed.
(1) How accurate are the lower bounds given above for any type of materials’
morphology?
(2) Do the upper estimates represent rigorous bounds?
To address these issues, one can, for example, analyze the effective properties of
sequentially-laminated materials (Ponte Castañeda, 1992b) which have provided
useful insights into the properties of linear materials, even though they represent
highly ideal models. A sequentially-laminated material (or laminate, for short) is
an iterative construction obtained by layering one type of laminated material with
other types of laminated materials, or directly with the homogeneous phases that
make up the composite, in such a way as to produce hierarchical microstructures of
increasing complexity. The rank of the laminate is the number of layering opera-
tions required to reach the final iterated morphology. Figure 2.1 presents a first-rank
laminate, constructed by mixing layers of two homogeneous phases to obtain a
simple laminate with layering direction n1 . A second-rank laminate, also shown
in Figure 2.1, is obtained by layering the first-rank laminate with a third phase
or, alternatively, with one of the original phases (say 2), in a different layering
direction n2 . In general, n1 and n2 can take on any orientation. It is assumed that
the length scale of the embedded laminates is small compared with the length scale
of the embedding laminates. Under this assumption, the fields will be essentially
constant within each elemental layer, provided that the boundary conditions ap-
plied to the laminate are uniform. This feature greatly simplifies the computation of
effective properties, thereby making sequentially-laminated materials very useful
constructions. With such a microstructure, the effective energy function of a simple

Figure 2.1. Examples of first-rank (left) and second-rank laminates (right).

 2.3. Exact Results for Laminates 41

linear laminate lies within and attains (for specific orientations of the applied fields)
the Wiener bounds. Thus, at least in this case, the Wiener bounds on the effective
energy function of arbitrarily anisotropic-linear materials are sharp.
In the context of two-phase linear materials, it is known that only iterated lami-
nates of rank greater than or equal to the dimension of the underlying physical space
(d = 2 or 3) can have isotropic properties. The isotropy is obtained by choosing
the relative volume fractions and the layering directions of each of the embedded
laminates in such a way that the tensor representing the effective property of in-
terest is isotropic, while the absolute volume fractions of the constituent phases
remain fixed. One might criticize sequentially-laminated materials by noting that
the inclusions are flat, whereas in practice the inclusions are often equi-axed.
However, one must note that iterated laminates can be used to model arbitrarily
close the properties of any two-phase microstructure (Milton, 1986). For exam-
ple, the coated-spheres model of Hashin and Shtrikman (1962a,b, 1963) possesses
exactly the same effective properties as an isotropic iterated laminate with the
same volume fractions. In the coated-spheres model (see also Chapters 3, 4 and
7 of Volume I) the material consists of composite spheres that are composed of a
spherical core of conductivity g2 and radius a, surrounded by a concentric shell of
conductivity g1 with an outer radius b > a. The ratio a/b is fixed, and the volume
fraction φ2 of inclusions in d dimensions is given by φ2 = (a/b)d . The composite
spheres fill the space, implying that there is a sphere size distribution that extends
to infinitesimally-small spheres. In Chapters 4 and 7 of Volume I we derived exact
expression for the effective conductivity and elastic moduli of the coated-spheres
model and low-rank laminates.
As shown in Chapters 4 and 7 of Volume I, the Hashin–Shtrikman bounds for
the coated-spheres model, which represent isotropic microstructures, are exact es-
timates. Thus, it may seem that the coated-spheres model may be more realistic
than the iterated laminates. However, the laminates have a distinct advantage over
the coated-spheres model in that, they contain a finite number of length scales, in
contrast with the coated-spheres microstructure which involves an infinite number
of length scales because, as described above, the composite spheres must cover
all sizes to fill the space. Another advantage of sequentially-laminated materials
is that, when subjected to uniform boundary conditions, the fields are piecewise
constant within the material (regardless of whether the composite’s phases are
linear or nonlinear), except in small boundary layer regions at the interfaces sep-
arating laminates of different ranks, the effect of which is made negligible by the
hypothesis of separated length scales. This fact was used for deriving the exact
results for the effective linear properties of the laminates presented in Chapters 4
and 7 of Volume I.
To compute the effective energy function of nonlinear rank−d laminates (d = 2
and 3) with layering directions n1 , · · · , nd , we denote by φI the volume fraction of
phase 1 with energy function e1 in the first-rank laminate, and note that 1 − φI is the
corresponding volume fraction of phase 2 with energy e2 . Ponte Castañeda (1992b)
showed that the effective energy function of the nonlinear, first-rank laminate is
42 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

given by
HeI (E) = inf {φI e1 (s1 ) + (1 − φI )e2 (s2 )} , (67)
(1) (2)
ωI ,ωI

(1) (2)
subject to the constraints that ωI  = φI ωI + (1 − φI )ωI = 0, and where
"
(1)
s1 = E2 − E12 + [1 − ωI ]2 E12 ,
"
(2)
s2 = E2 − E12 + [1 − ωI ]2 E12 , (68)
where E1 = E · n1 .
Consider now the second-rank laminate obtained by mixing layers of the first-
rank laminate with layers of a third phase characterized by an energy function
e3 and relative (to the second-rank laminate) volume fractions φI I and 1 − φI I ,
respectively. The new lamination direction n2 is orthogonal to n1 . Then, the fol-
lowing energy function for the nonlinear second-rank laminate in dimension d ≥ 2
is obtained (Ponte Castañeda, 1992b):
HeI I (E) = inf {φI I φI e1 (s1 ) + φI I (1 − φI )e2 (s2 ) + (1 − φI I )e3 (s3 )} ,
(1) (2) (1) (2)
ωI ,ωI ,ωI I ,ωI I
(69)
subject to the constraints that ωI  = ωI I  = 0 (where ωI I  is defined in a
manner analogous to ωI ), and where
"
(1) (1)
s1 = E2 − E12 − E22 + [1 − ωI ]2 E12 + [1 − ωI I ]2 E22 ,
"
(2) (1)
s2 = E2 − E12 − E22 + [1 − ωI ]2 E12 + [1 − ωI I ]2 E22 , (70)
"
(2)
s3 = E2 − E22 + [1 − ωI I ]2 E22 ,
where Ei = E · ni .
Asimilar result can be obtained for a two-phase, nonlinear second-rank laminate.
In this case, the result for HeI I is generally anisotropic and direction-dependent,
but it may be used in two dimensions (2D) for deriving an isotropic result, for
each value of E, by an appropriate choice of φI I (but not the choice that makes
the corresponding linear second-rank laminate isotropic), which is obtained by
requiring that φI I (0 ≤ φI I ≤ 1) and E1 satisfy the following relations
∂HeI I ∂HeI I
= 0 and = 0, (71)
∂E1 ∂φI I
where the first relation is subject to the constraint that, E12 + E22 = E2 , while
in the second relation one assumes that E is fixed. These conditions follow by
performing a Taylor series expansion of (69) in φI I and E1 and requiring that
the expansion yield the same result for any choice of φI I and E1 . Physically, this
corresponds to selecting a microstructure (by choosing φI I )—with fixed over-
all volume fractions of the phases—for each value of E, ensuring that HeI I is
 2.4. Effective Dielectric Constant of Strongly Nonlinear Materials 43

independent of the direction of E, thus guaranteeing that the resulting energy
function is isotropic. However, the resulting energy function does not correspond
to a fixed microstructure, rather to a family of (anisotropic) microstructures, each
one of which is obtained from one value of the applied electric field.
The effective energy function of a nonlinear third-rank laminate is obtained by
analyzing the effective behavior of a simple laminate made up of layers of the
second-rank laminate and of layers of a fourth phase with energy function e4 and
volume fractions φI I I and 1 − φI I I , respectively. The new layering direction n3
is selected to be orthogonal to both n2 and n1 . Then, the effective energy function
of the nonlinear third-rank laminate is given by (Ponte Castañeda, 1992b)
HeI I I (E) = inf {φI I I φI I φI e1 (s1 )
(1) (2) (1) (2) (1) (2)
ωI ,ωI ,ωI I ,ωI I ,ωI I I ,ωI I I

+φI I I φI I (1 − φI )e2 (s2 ) + · · · + φI I I (1 − φI I )e3 (s3 ) + (1 − φI I I )e4 (s4 )} ,

(72)
subject to the constraints that ωI  = ωI I  = ωI I I  = 0, and where
"
(1) (1) (1)
s1 = E2 − E12 − E22 − E32 + [1 − ωI ]E12 + [1 − ωI I ]2 E22 + [1 − ωI I I ]2 E32 ,
"
(2) (1) (1)
s2 = E2 − E12 − E22 − E32 + [1 − ωI ]E12 + [1 − ωI I ]2 E22 + [1 − ωI I I ]2 E32 ,
"
(2) (1)
s3 = E2 − E22 − E32 + [1 − ωI I ]E22 + [1 − ωI I I ]2 E32 ,
"
(2)
s4 = E2 − E32 + [1 − ωI I I ]2 E32 , (73)
where, as before, Ei = E · ni .
The effective energy function of a two-phase, nonlinear third-rank laminate may
be obtained by letting e4 = e3 = e2 in (72). Then, for 3D third-rank laminates
Eq. (72) may be used to obtain an isotropic energy by choosing φI I and φI I I , and
E1 , E2 , and E3 with E12 + E22 + E32 = E2 , such that
∂HeI I I ∂HeI I I ∂HeI I I ∂HeI I I
= = 0, = = 0. (74)
∂E1 ∂E2 ∂φI I ∂φI I I

2.4 Effective Dielectric Constant of Strongly Nonlinear

Materials
To illustrate the application of the methods described above, we consider two
important examples that we have been considering throughout this book, both in
Volume I and the present Volume. Both limits involve a nonlinear matrix with
isotropic potential e2 = e (subject to the restrictions of Theorem 1), and an inclu-
sion phase that, similar to the case of nonlinear conductivity discussed above, has
either an infinite dielectric constant or, alternatively, a zero dielectric constant. In
the first case, e1 = 0 if E = 0, or e1 = ∞ otherwise, while in the second, e1 = 0
44 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

regardless of the value of E. Moreover, we specialize the results to the case in

which the nonlinearity of the matrix is of power-law type, which is usually re-
ferred to as strong nonlinearity. This type of nonlinearity is characterized by the
energy-density function
e(E) = (n + 1)−1  (n) E n+1 , (75)
where n ≥ 1, and  (n)
is the nonlinear dielectric constant. Equation (75) has the
advantage that it yields the same type of behavior for the isotropic composite
materials with perfectly conducting or insulating inclusions. Thus, for both types
of isotropic composites, we have
He (E) = (n − 1)−1 e(n) En+1 , (76)
(n)
where e is the effective nonlinear dielectric constant of the material. For the
anisotropic materials, the form of the effective energy will, in general, be different,
but the Wiener bounds will be of the same form. We can then characterize the
behavior of the Wiener, Hashin–Shtrikman and isotropic (in the sense defined
earlier) laminates for this class of materials in terms of the effective nonlinear
dielectric constant.

2.4.1 Inclusions with Infinite Dielectric Constant

The results for the bounds and estimates of 2D materials are not essentially different
from those for 3D composites, and therefore only the results for the 3D materials are
presented. Consider first the Wiener and the Hashin–Shtrikman lower bounds and
the isotropic laminate estimate for nonlinear materials with perfectly conducting
inclusions. These results, expressed in terms of the effective nonlinear dielectric
constant, are given by (Ponte Castañeda, 1992b)
(l)
W
= (1 − φ)−n , (77)
 (n)
(l)
HS (1 + 2φ)(n+1)/2
= , (78)
 (n) (1 − φ)n
 $ % 
eI I I 2 − y (n+1)/(n−1)
= sup (1 − y)2n/(n−1) + (xy − φ)2n/(n−1)
 x,y y
$ % (1−n)/2
2 − x (n+1)/(n−1)
× + [(1 − x)y]2n/(n−1) , (79)
x
where φ = φ2 is the volume fraction of the inclusions, and the optimization vari-
ables x and y are subject to the constraints, 0 ≤ x, y ≤ 1 and xy ≥ φ. Note that
as n → ∞,
eI I I
→ (1 − 8φ + 12φ 4/3 − 6φ 5/3 + φ 2 )−n/2 , (80)
 (n)
(l)
which is different from, but close to, HS / 1/n in the same limit. In general,
the Hashin–Shtrikman bound provides estimates that are very close to those for
 2.5. Effective Conductivity of Nonlinear Materials 45

the laminates, and both differ strongly from the Wiener bound, with the latter
yielding estimates that are larger than the former two.

2.4.2 Inclusions with Zero Dielectric Constant

Consider now the corresponding results for 3D nonlinear materials with perfectly
insulating inclusions. The results for the Wiener upper bound, the Hashin–
Shtrikman upper estimate, and the exact estimate for the isotropic laminate are
given by (Ponte Castañeda, 1992b)
(u)
W
= 1 − φ, (81)
 (n)
+
HS 1−φ
= , (82)
 (n) 1 + 12 φ (n+1)/2
 
eI I I [(1 − x)y + (1 − y)p](n+1)/2
= sup ,
x,y [(xy − φ)q
 (n) (n+1)/(n−1) + (1 − x)y + (1 − y)p (n+1)/(n−1) ](n−1)/2

(83)
where p is the root of the quadratic equation,
 
11−y 1
(2 − x)p 2 − xy + 2(1 − y) p + (1 − x)(1 − y) = 0,
21−x 2
and
xy 1 − x
q= .
xy − φ 2 − x
In this case, the Hashin–Shtrikman upper estimates for the isotropic composite
lie well below the Wiener bounds for arbitrarily anisotropic composites. On the
other hand, the exact estimates for the nonlinear isotropic laminates lie above the
Hashin–Shtrikman upper estimates, hence verifying that the Hashin–Shtrikman
upper estimates are not in general upper bounds. This is due to the fact that
the isotropic laminates correspond to specific microstructures within the class
of isotropic composite materials, and if the Hashin–Shtrikman upper estimates
were rigorous bounds for such materials, they would have to lie above all possible
isotropic microstructures, and, in particular, they must lie above the isotropic lam-
inates. Nevertheless, the effective dielectric constants of the isotropic laminates
are not far from the Hashin–Shtrikman upper estimates.

2.5 Effective Conductivity of Nonlinear Materials

The above methods of deriving bounds and estimates for the effective dielectric
constant of heterogeneous nonlinear materials can also be used for estimating their
effective conductivity (Ponte Castañeda, 1998). Equation (75) is now written as
ei (E) = (n + 1)−1 gi |E|n+1 ,
(n)
(84)
46 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

(n)
where gi is the generalized nonlinear conductivity of phase i. The linear
comparison materials are now defined by the quadratic energy-density function,
1 0
w 0 (x, E) = g (x)E 2 , (85)
2
where g 0 (x) is the conductivity of the fictitious linear material. Then, under the
hypothesis that the functions ei of the nonlinear material are convex on E 2 , the
analogues of Eqs. (37) and (41) for the conductivity problem are given by
 
1 0
ei (E) = max g (x)E 2 − vi (g 0 ) , (86)
g 0 ≥0 2
 
1 0 2
vi (g 0 ) = max g E − ei (E) . (87)
E 2
Note that if the functions ei are smooth, the maxima in Eqs. (86) and (87) are
attained at
1 2 ∂vi 1 ∂ei
E = 0 , g0 = , (88)
2 ∂g E ∂E
respectively, which are inverse of each other. Then, the analogue of Eq. (31) for
the conductivity problem is given by
 

N
He (E) = max He (E) −
0
φi vi [g (x)]i ,
0
(89)
g 0 (x)≥0
i=1

where He0 is the effective energy function of the linear comparison material, with
local energy function (85), such that
He0 (E) = min w 0 (x, E), (90)
E∈S2

as before, where S2 is the set defined by (11).

Equation (89), together with Eqs. (87) and (90), provide variational represen-
tation of the effective energy function of the nonlinear material in terms of the
effective energy function of a fictitious linear composite, the choice of which is
determined by Eq. (89). It should be emphasized that the conductivity g 0 (x) of
the comparison material is an arbitrary non-negative function of x, and that the
minimum principle (89) is valid only under the hypothesis that the functions ei
are convex on E 2 . If these functions are concave on E 2 (as when, for example,
0 ≤ n < 1), an analogous result would hold, but with the maximum in Eq. (89)
replaced by a minimum, and with the function vi redefined such that the maximum
in Eq. (87) is replaced by a minimum.

2.5.1 Materials with Nonlinear Isotropic Phases

Even if each of the nonlinear phases is homogeneous, the solutions for the com-
parison conductivities g 0 (x) in the variational principle (89) is not, in general,
 2.5. Effective Conductivity of Nonlinear Materials 47

constant over the individual phases, unless the actual fields are constant through-
out the phases. However, as discussed earlier in this chapter, a lower bound for
He0 can be obtained by restricting the class of trial comparison conductivity fields
to be constant within each phase such that


N
g 0 (x) = mi (x)gi0 , (91)
i=1

where gi0 is constant, and mi (x) is the indicator function of phase i, as before.
Equation (91) follows from the fact that the maximum over a set is, in general,
larger than the maximum over any subset of the original set. Therefore, from
Eqs. (89) and (91), it follows that
 

N
He (E) ≥ max He0 (E) − φi vi (gi0 ) , (92)
gi0 >0 i=1

where
 
1

N
1
He0 (E) = E · [ge(l) E] = min φi gi E i .
0 2
(93)
2 E∈S2 2
i=1
(l)
Here, ge is the effective conductivity tensor of a linear comparison material with
precisely the same morphology as the original nonlinear composite which, in
general, is anisotropic.
As discussed earlier in this chapter, the above estimates for N -phase nonlinear
(l)
materials represent lower bounds for He . Thus, lower bounds for ge may be
(l)
used to generate lower bounds for He , but upper bounds for ge cannot be used
for deriving upper bounds for He . In this case, one can ignore the inequality in
(92) and reinterpret it as an approximate equality in order to obtain estimates for
specific types of materials. Denoting by ĝi0 the optimal values of gi0 from Eq. (92),
it follows that the average current field I is given by
   


N
1
(l)
∂ge ∂vi 0 ∂ ĝi0
I = ge(l) (ĝ10 , · · · , ĝN
0
)E + E · ( ĝ 0
1 , · · · , ĝN
0
)E − φ i ( ĝ i ) ,
i=1
2 ∂gi0 ∂gi0 ∂E
(94)
so that, the maximum in (92) for the general bound is attained at
 
(l)
1 ∂ge ∂vi
E · 0
(ĝ10 , · · · , ĝN
0
)E = φi 0 (ĝi0 ) (i = 1, · · · , N). (95)
2 ∂gi ∂gi
The constitutive relation that defines the effective conductivity of the nonlinear
material reduces to
I = ge(l) (ĝ10 , · · · , ĝN
0
)E. (96)
Note that, Eq. (96) is fully nonlinear because the variables depend nonlinearly ĝi0
(l)
on E. Since the linear conductivity ge is a homogeneous function of degree
48 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

one in the conductivity constants gi0 of the linear comparison material (see also
Chapters 2, 4 and 6 of Volume I), then


N (l)
∂ge
gi0 = ge(l) . (97)
i=1
∂gi0

Therefore, Eq. (83) is rewritten as

   


N
1 gi0
(l)
∂ge
He (E) ≥ max φi E · E − vi (gi0 ) , (98)
gi0 >0 i=1 2 φi ∂gi0

and Eq. (86) implies that

N
He (E) ≥ φi ei (Êi ), (99)
i=1

where
  1/2
(l)
1 ∂ge
Êi = E · (ĝ1 , · · · , ĝi )E
0 0
(i = 1, · · · , N). (100)
φi ∂gi0
Finally, the constitutive relation describing the effective behavior of the nonlinear
material is written in the following form,
 
1 ∂e1 1 ∂eN
I = ge(l) (Ê1 ), · · · , (ÊN ) E, (101)
Ê1 ∂E Ên ∂E
where Êi are functions of the (average) applied field E, the nonlinear properties
of the constituent phases of the material, and the material’s microstructure.

2.5.2 Strongly Nonlinear Materials with Isotropic Phases

Consider now the class of materials that is defined by Eq. (84) for the phase
potentials ei , for which it is possible to simplify further the two equivalent forms
(92) and (99). Thus, since


N  
φi gi0 (Êi )2 = E · ge(l) E , (102)
i=1

then, for a power-law material,

N
1

N
He (E) ≥ φi ei (Êi ) = φi gi0 (Êi )2 , (103)
n+1
i=1 i=1

and therefore
   
1 1 ∂e1 1 ∂eN
He (E) ≥ E · ge
(l)
(Ê1 ), · · · , (ÊN ) E . (104)
n+1 Ê1 ∂E ÊN ∂E
 2.5. Effective Conductivity of Nonlinear Materials 49

If the material’s microstructure is statistically isotropic, then writing

1
He (E) = g (n) En+1 , (105)
n+1 e
(n)
and using (95), we obtain an equation for ge , the effective nonlinear conductivity
of the material (Wan et al., 1996; Ponte Castañeda, 1998).
For statistically-isotropic, nonlinear materials, we need to consider only
(l) (l)
isotropic linear comparison composites with He0 (E) = 12 ge E2 , where ge
is now a scalar function of the nonlinear conductivities gi0 , the volume fractions
φi , and the material’s microstructure. In particular, as discussed in Section 4.6.1.1
of Volume I, for a two-phase material, there are several closely related bounds
and estimates for linear materials which can be all characterized in terms of the
d-dimensional quantity,
φ1 φ2 (g10 − g20 )2
ge(l) = φ1 g10 + φ2 g20 − , (106)
φ2 g10 + φ1 g20 + (d − 1)g 0
where g 0 takes on different values for different types of estimates. For example,
assuming that g10 > g20 , then,
(1) the limits g 0 → ∞ and 0 correspond, respectively, to the Wiener (one-point)
upper and lower bounds.
(2) The choices g 0 = g10 and g20 yield the two Maxwell–Garnett approximations
for particulate microstructures with phases 1 and 2, respectively, in the ma-
trix phase (recall from Section 9.4.9 of Volume I that the Maxwell–Garnett
approximation is not symmetric with respect to the two phases).
(3) The same choices as in (2) also lead to the Hashin–Shtrikman upper and lower
bounds.
(4) If we choose, g 0 = ζ1 g10 + ζ2 g20 and (ζ1 /g10 + ζ2 /g20 )−1 , we obtain, respec-
tively, the upper and lower bounds of Beran, in terms of the microstructural
parameters ζ1 and ζ2 = 1 − ζ1 .
(l)
(5) Finally, the choice g 0 = ge yields the effective-medium approximation
(EMA).
For a two-phase material, one can obtain an expression for He in terms of
only one nonlinear equation for the ratio ĝ10 /ĝ20 . Computing the variables Ê1 and
(n)
Ê2 in terms of this ratio, the resulting effective nonlinear conductivity ge is
presented in Figure 2.2 for 2D, statistically isotropic, two-phase, power-law con-
(n) (n)
ductors with n = 3 and g2 /g1 = 1000, where W(l) and W(u) correspond to
the rigorous upper and lower Wiener bounds for heterogeneous materials with
arbitrary microstructures, MG(l) and MG(u) represent the Maxwell–Garnett esti-
mates for particulate microstructures with the less and more conducting materials
occupying the matrix phase, respectively. Because the MG(l) estimate for the con-
ductivity of a linear material coincides with the Hashin–Shtrikman lower bound
for the set of all statistically isotropic composites, the MG(l) results are identical to
the rigorous nonlinear Hashin–Shtrikman lower bound. The results of numerical
50 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

Figure 2.2. Comparison of various bounds and estimates for the effective nonlinear conduc-
(n)
tivity ge of 2D, isotropic, two-phase, power-law conductors with power exponent n = 3
(n) (n)
and g2 /g1 = 1000. Symbols show the results of numerical simulations with random
resistor networks (RRN). The lower bound MG(l) obtained from the Maxwell–Garnett ap-
proximation is identical with that obtained from the Hashin–Shtrikman lower bound. MG(u)
and B(u) denote, respectively, the estimates for the upper bound using the Maxwell–Garnett
approximation and the Beran upper bound. (after Ponte Castañeda, 1998).

simulation using a random resistor network (RRN) model, obtained by Wan et

al. (1996) using a square network, are also shown. The resistor network models
will be described and discussed in detail in Chapter 3. The label B(l) represents
the rigorous lower bound of Beran for statistically isotropic microstructures with
ζ1 = φ1 , which is presumably appropriate for symmetric cell microstructures that
are similar to the RRN models. The label B(u) denotes the estimate (not a rig-
orous bound; see above) which is obtained by using the Beran upper bound for
the linear comparison composite. As expected, the EMA estimates are in good
agreement with the RRN simulations. The Wiener, Hashin–Shtrikman, and Beran
bounds progressively narrow the range of possible behavior by introducing, as dis-
cussed in detail in Chapters 3, 4, and 7 of Volume I, first-, second- and third-order
statistical information about the microstructure of the material, respectively. Al-
though the Maxwell–Garnett approximation and the EMA are generally accurate
for particulate- and granular-type microstructures, respectively (see Chapters 4 and
 2.6. Second-Order Exact Results 51

7 of Volume I), the Beran bounds provide a way of estimating the effective proper-
ties of more general types of microstructures for which the Maxwell–Garnett and
EMA estimates may not be accurate.

2.6 Second-Order Exact Results

The method described and utilized so far is most suited for deriving lower bounds
and estimates for the upper bounds. These estimates are exact to first order in the
contrast between the properties of various phases of a multiphase material. In this
section, we describe and discuss another method, developed by Ponte Castañeda
and Kailasam (1997), which yields estimates that are exact to second order in the
contrasts. As such, they are more accurate than the predictions that are provided
by the method described above.
We should mention that Blumenfeld and Bergman (1991b) developed a general
method for reducing the solution of the scalar-potential field problems to the so-
lution of a set of linear Poisson-type equations in suitably rescaled coordinates.
In particular, for power-law type nonlinearities, they solved explicitly for the ef-
fective dielectric constant of a two-phase material to second order in the contrast
between the phases’ properties. Despite its elegance, their solution yields unphysi-
cal results for strong nonlinearity, even when the contrast is not very large, whereas
the method described below does not suffer from this shortcoming. We will come
back to this point at the end of this section.
The key idea of Ponte Castañeda and Kailasam (1997) is developing a Taylor
expansion for the phase energy functions wi , around appropriately defined phase-
average electric fields Ei , so that
1
wi (E) = wi (Ei ) + I(i) · (E − Ei ) + (E − Ei ) · [ĝ(i) (E − Ei )], (107)
2
where I(i) and ĝ(i) are reference current densities and conductivity tensors with
components
∂wi (i) ∂ 2 wi
I (i) = (Ei ), gj k = (E(i) ), (108)
∂Ei ∂Ei ∂Ej
where E(i) is a reference electric field given by, E(i) = λ(i) Ei  + [1 − λ(i) ]E, with
0 < λ(i) < 1. We now rewrite Eq. (107) in terms of the average E and fluctuating
E components, E = E + E :
1
wi (E + E ) = vi (E) + Pi · E + E · [ĝ(i) E ], (109)
2
where
1
vi (E) = wi (Ei ) + Pi · (E − Ei ) − (E − Ei ) · [ĝ(i) (E − Ei )],
2
Pi = I(i) + ĝ(i) (E − Ei ). (110)
52 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

Then, the effective energy He of the nonlinear composite material is given by

1
He (E) = min v + P · E + E · (ĝE ), (111)
E ∈S  2
where S  denotes the set of admissible fields E , such that, E = ∇ϕ  in the subspace
 and ϕ  = 0 on ∂, and


N

N

N
v(x) = mi (x)vi , P(x) = mi (x)Pi , ĝ(x) = mi (x)ĝ(i) . (112)
i=1 i=1 i=1

Equation (111) assumes that the reference fields E(i) are known in terms of the
λ(i) , which, in general, are functions of the actual electric field E, as well as of
the (unknown) Ei , and therefore the problem posed by (111) for He is nonlinear.
However, provided that the second derivatives of wi in Eq. (107) vary slowly
with the E(i) , Eq. (111) suggests, as an approximation, replacing E(i) by a (as-yet
unknown) constant, in which case ĝ(i) , Pi and vi will also be constant within each
phase, hence leading to the following expression for He ,


N
He (E) = φi vi (E) + P̃ (E), (113)
i=1

where
1
P̃ (E) = min P · E + E · (ĝE ). (114)
E ∈S  2
The interesting feature of Eq. (113) is that it requires only the solution of the
linear problem (114) for P̃ which is, physically, equivalent to a problem for a
linear conductor with N anisotropic constituents with conductivity tensors ĝ(i) and
prescribed polarizations Pi , a problem much simpler to analyze than the original
nonlinear problem for He . The question then arises as to what the best choices are
for these constants.
The optimal choice for each Ei  is Ei , the average of the actual field E over
phase i:
Ei  = Ei , (115)
where · · ·i denotes a volume average over phase i. Although Ei cannot be
obtained exactly, a consistent estimate for it may be obtained by noting that, Ei =
E + E i , where
1 ∂ P̃
E i = (116)
φi ∂Pi
with ĝ(i) held fixed. On the other hand, although the best choice for the E(i) is not
a priori clear, given the approximation that was made in deriving (113), the choice
E(i) = Ei , (117)
 2.6. Second-Order Exact Results 53

is simple and plausible. In particular, Eqs. (115) and (117) are exact for laminated
materials, where the fields are constant within each phase. Thus, any solution for
the problem posed by (114), together with the associated estimates (116), may
be utilized for obtaining corresponding estimates for He via Eq. (113), together
with the (self-consistent) equations (115) and (117). Note that E(i) = Ei , and,
for this reason, ĝ(i) is henceforth denoted by gi , the phase conductivity tensor. In
particular, for two-phase composite materials one can show that
1  (l) 
P̃ (E) = (ge − g)(g)−1 P · (g)−1 P, (118)
2
from which it follows, using (116), that
1
E1 = E + (g)−1 (ge(l) − g)(g)−1 P, (119)
φ1

1
E2 = E + (g−1 )(ge(l) − g)(g)−1 P, (120)
φ2
(l)
where g = g1 − g2 , P = P1 − P2 , g = φ1 g1 + φ2 g2 , and ge is the effective
conductivity tensor of a two-phase linear material with phase conductivity tensors
g1 and g2 , volume fractions φ1 and φ2 , and precisely the same microstructure as
the nonlinear composite. This means that any estimate that is available for the
(l)
effective conductivity tensor ge of a two-phase linear material, including, for
example, the Maxwell–Garnett and EMA estimates, can be used for generating
the corresponding estimates for He of a two-phase nonlinear material. Note that
the approximate estimate of He given by Eq. (113) is a convex function. Since
the exact expression for He is also known to be convex, it follows that deriva-
tives of the approximate expressions for He should provide a reasonably accurate
approximation to the exact constitutive relation.

2.6.1 Strongly Nonlinear Isotropic Materials

Consider now a class of two-phase materials for which Eq. (84) describes the
constitutive relation. We already saw that for statistically isotropic materials, the
macroscopic behavior is described by Eq. (105). For such materials, it is reasonable
to assume that the reference fields E(i) and I(i) are aligned with the corresponding
applied fields. If so, one can define scalar variables ωi and νi , such that
E(i) = (1 + ωi )E, I(i) = (1 + νi )I, (121)
from which it follows that ei = E/E and Ii = I/I , for all the phases i. This
implies that the conductivity tensor gi of all the phases in the linear comparison
material has exactly the same symmetry. Then, using (121), we find from Eq. (110)
that
νi (E) = (1 + n)−1 gi E1+n ,
(p) (v)
Pi = gi En−1 E, (122)
54 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

with
(p)
gi = gi [1 + (1 − n)ωi ], (123)
 
(v) 1
gi = gi 1 + (1 − n)ωi + n(n − 1)ωi2 . (124)
2
Then, from Eqs. (84), (113) and (118) we obtain
 
n + 1  (l)  g (p) − g (p) 2
ge(n) = g (v)
+ mge − ng 1 2
, (125)
2n2 g1 − g 2
(n) (n)
where here gi = gi |1 + ωi |n−1 , with gi being the nonlinear conductivity of
phase i, and g (v)  defined in a manner analogous to g. The variables ωi are
determined by Eqs. (115), (119), (120) and (121); they yield, ω1 = φ2 ω, ω2 = φ1 ω,
where
(l) (p) (p)
1 1 mge − ng g1 − g2
ω= . (126)
φ1 φ 2 n 2 g1 − g 2 g1 − g 2
(n)
Estimates of ge based on the Maxwell–Garnett approximation and the EMA
can now be obtained by using their corresponding estimates for the effective con-
(l)
ductivity tensor ge , which are in terms of the phase conductivity tensor gi . Since
the Maxwell–Garnett approximation is not symmetric in material’s phases, one
obtains two classes of Maxwell–Garnett estimates, corresponding to particulate
microstructures with the less and more conducting material designated as the ma-
trix phase. On the other hand, due to its symmetry, the estimates provided by the
EMA are unique. Moreover, it should be pointed out that the Maxwell–Garnett
(n)
and EMA estimates for the effective nonlinear conductivity ge are not exactly
equivalent. In fact, it can be shown that while for sufficiently weak nonlinearity
(i.e., for n  1) these estimates are in close agreement with each other, they can be
significantly different for stronger nonlinearities (i.e., as n → 0 or ∞). The reason
for the differences is associated with the nature of the approximations made in
going from the exact estimate (111) for He to the approximation (113), assuming
that the reference conductivity tensors ĝ(i) vary slowly with E(i) , so that the re-
placement of the E(i) by Ei does not introduce significant errors. In what follows,
we summarize the results obtained with the Maxwell–Garnett and EMA estimates.
The details of derivation of these results, which is straightforward, are given by
Ponte Castañeda and Kailasam (1997).

2.6.1.1 The Maxwell–Garnett Estimates

The Maxwell–Garnett estimates that correspond to designating the matrix as phase
2 are obtained from the following equations (see Section 4.9.4 of Volume I for the
corresponding Maxwell–Garnett equations for linear materials),
 −1
2 1
mge − ng = −nφ1 φ2 (g1 − g2 )
(l)
α(n)g2 + φ2 (g1 − g2 ) , (127)
n
 2.6. Second-Order Exact Results 55

where α(n) is a function of n, given by

√
α(n) = n n, 2D, (128)
 & −1
1 n−1
α(n) = (n − 1) 1 − √ arcsin , 3D. (129)
n−1 n
The 3D expression for α is valid for n ≥ 1, but the corresponding expressions for
n ≤ 1 may be easily obtained by analytic continuation. Then, Eq. (127), together
(n)
with Eqs. (125) and (126), provide one of the Maxwell–Garnett estimates for ge .
The other Maxwell–Garnett estimate, with the matrix designated as phase 1, is
obtained by simply interchanging the roles of 1 and 2.

2.6.1.2 Effective-Medium Approximation Estimates

In this case,
⎧
n ⎨ g − m(g1 + g2 )/α(m)
ge(l) =
m⎩ 2[1 − m/α(m)]
# ⎫
 2 ⎬
g − m(g1 + g2 )/α(m) g 1 g2
+ + (130)
2[1 − m/α(m)] α(m)/m − 1 ⎭

and
g − (g1 + g2 )/β(m)
ge(l) =
2[1 − 1/β(m)]
#
 
g − m(g1 + g2 )/β(m) 2 g 1 g2
+ + (131)
2[1 − m/β(m)] β(m)/m − 1
where
⎧ √
⎨ 1+ n, 2D,
$ " %−1
β(n) = (132)
⎩ 2(1 − n) 1 − √n arcsin n−1
n , 3D,
n−1

and g = φ1 g1 + φ2 g2 , as before. Equations (130) and (131), obtained from the
(l)
two independent components of the anisotropic tensor ge , depend on the func-
tions α and β which, in turn, are known functions of the unknown parameter m.
(l)
Therefore, m is obtained by equating (130) and (131). Once m is obtained, ge
(n)
and hence ge are computed.
We now consider the application of these results to estimating the effective
nonlinear conductivity of two important classes of heterogeneous materials that we
have been studying throughout this book, namely, those with superconducting or
insulating inclusions.As we emphasized in Volume I, because these two composites
represent two extreme limits of contrast between the properties of the two phases,
they provide stringent tests of any theory. In other words, if a theory is reasonably
accurate in these limits, it will be even more accurate in less extreme cases.
56 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

2.6.2 Conductor–Superconductor Composites

It is straightforward to show that, in this limit, Eq. (125) yields the following
(n)
Maxwell–Garnett estimates for ge :
(n)   
ge 1 1 1
= 1 + n(n + 1)φ 1 α(n) − 1 , (133)
g
(n) φ2n 2 n
2
(n)
where 0 < n < 1. The corresponding EMA estimates for ge are given by
Eq. (133) with the factor α(n)/n − 1 replaced by [α(m) − m]/[m − α(m)φ1 ],
where m is the solution of the equation,
m − α(m)φ1 = n[1 − β(m)φ1 ]. (134)
The EMA estimates are valid for φ1 < 1/β(m). The limit φ1 = 1/β(m) defines
(n) (n)
the percolation thresholds for ge at which ge → ∞. The Maxwell–Garnett
estimates, on the other hand, do not exhibit any percolation behavior, which is an
undesirable aspect of these approximations, as already pointed out in Chapter 4 of
Volume I.
Figure 2.3 presents the 3D Maxwell–Garnett, EMA, Hashin–Shtrikman and
(n) (n)
Wiener estimates for the effective resistivity Re /R2 of the composite material,
as functions of the volume fraction φ1 of the inclusions, for the power-law expo-
nent n = 3. As one might expect, both the Maxwell–Garnett and EMA estimates
lie below the Wiener and Hashin–Shtrikman upper bounds. Moreover, it can be
shown that the differences between the new Maxwell–Garnett estimates and the
old Hashin–Shtrikman bounds (derived earlier in this chapter) increase as n in-
creases, whereas they agree for n = 1. As usual, the EMA estimates exhibit sharply
the percolation limit at a finite value of φ1 , a distinct advantage of this method.

2.6.3 Conductor–Insulator Composites

(n)
In this limit [when g1 → 0], Eq. (125) yields the following Maxwell–Garnett
(n)
estimates for ge :
(n)  
ge 1
(n)
= φ 2 |1 − φ 1 ω| n
1 + φ 1 ω(n − 1) , (135)
g2 2

with ω = [φ1 + α(n) − n]−1 . The corresponding EMA estimates are obtained
from Eq. (135) with ω = [φ1 + nα(m)φ2 /m − 1]−1 , where m is the root of the
following equation
   
φ1 β(m) φ1 α(m)
m 1+ =n 1+ . (136)
1 − β(m) m − α(m)
The EMA estimates are valid for φ1 ≤ 1 − β(m)−1 . The limit φ1 = 1 − β(m)−1
(n)
defines the percolation threshold at which ge vanishes.
 2.6. Second-Order Exact Results 57

(n)
Figure 2.3. The effective resistivity Re of 3D, isotropic, two-phase, power-law materials,
as predicted by the various approximations, versus the volume fraction φ1 of the supercon-
ducting inclusions, with n = 3. Note that only the effective-medium approximation indi-
cates the existence of a percolation threshold (after Ponte Castañeda and Kailasam, 1997).

Figure 2.4 presents the 3D Maxwell–Garnett, EMA, Hashin–Shtrikman and

(n) (n) (n) (n) (n)
Wiener estimates for ge /g2 = [R2 /Re ]n , where Re is the effective resis-
tivity of the material, as functions of the volume fraction φ1 of the inclusions, for
n = 3. Both the Maxwell–Garnett and EMA estimates lie below the Wiener up-
(n)
per bound for ge [Eq. (81)], while the Maxwell–Garnett estimates lie above the
(n)
Hashin–Shtrikman lower bound for Re for particulate microstructures. The EMA
estimates that correspond to granular microstructures (which are different from
particulate microstructures) are not constrained to satisfy the Hashin–Shtrikman
bound and vanish at a finite value of φ1 , the percolation threshold. The difference
between the Maxwell–Garnett estimates and the Hashin–Shtrikman lower bound
increases with increasing n; recall that they are identical in the limit n = 1.
It can also be shown that all the nonlinear Maxwell–Garnett and EMA estimates
for the effective conductivity or resistivity agree to first order in the volume fraction
φ1 , with the result being
(n)
Re
(n)
= 1 + γ (n)φ1 + O(φ12 ), (137)
R2
58 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

05

Figure 2.4. Same as in Figure 2.3, but with insulating inclusions (after Ponte Castañeda
and Kailasam, 1997).

with
 
1 n+1
γ (n) = −1 . (138)
n 2[α(n) − n]
Analogous expressions can also be derived for the Wiener and Hashin–Shtrikman
bounds. Figure 2.5 presents a comparison of γ (n) for the MG/EMA estimates ver-
sus the Wiener and Hashin–Shtrikman bounds for the 2D materials, along with the
numerical results of Lee and Mear (1992), who reported their results for transverse
shear of fiber-reinforced power-law ductile composite conductors. As one might
expect, the MG/EMA estimates lie above the rigorous Wiener lower bound (for
the resistivity) for particulate microstructures. Moreover, the new estimates are in
excellent agreement with the numerical estimates of Lee and Mear (1992).

2.6.4 General Two-Phase Materials

In addition to the above limits, one may also consider general two-phase power-
law materials at arbitrary contrast between the phases. Since the behavior of the
(n)
general estimate (125) for ge for n < 1 is similar to that of the effective resistivity
(n) (n)
Re for n > 1, we discuss the various types of estimates for Re for n > 1. In
addition, recall that since two types of Maxwell–Garnett estimates are possible
 2.6. Second-Order Exact Results 59

Figure 2.5. Dependence of the coefficient γ on the power-law exponent n, for a 2D,
isotropic, two-phase materials with insulating inclusions. Symbols represent the numerical
results of Lee and Mear (LM) (1992) (after Ponte Castañeda and Kailasam, 1997).

(n) (n)
for a given value of the ratio R1 /R2 , depending on whether phase 1 or 2 is
designated as the matrix phase (and vice versa for the inclusion phase), we restrict
(n) (n)
our attention to R1 /R2 > 1, and denote by MG1 and MG2 the two estimates
corresponding to designating phases 1 and 2, respectively, as the matrix phase.
Ponte Castañeda and Kailasam (1997) showed that as the volume fraction φ1
(n)
of the inclusions increases, the MG2 estimates for Re also increase. However,
the rate of the increase decreases with increasing n. In particular, for sufficiently
(n) (n)
large n, there is hardly any increase in Re over the matrix resistivity R2 . The
reason for this effect is the fact the current density becomes concentrated in the
more conducting matrix phase as n increases, and therefore the effect of the inclu-
sions becomes insignificant. Moreover, as the volume fraction φ1 of the inclusions
(n)
increases, the MG1 estimates for Re decrease, with the rate of the decrease in-
creasing with increasing n. In addition, as is the case for the estimates of the linear
(n)
EMA (see Chapter 4 of Volume I), the nonlinear EMA estimates for Re agree
with the MG1 and MG2 estimates in the limits of small volume fractions of phases
2 and 1, respectively.
(n)
The 3D Maxwell–Garnett and EMA estimates for Re also agree with the corre-
sponding small-contrast asymptotic results of Blumenfeld and Bergman (1991b),
which are known to be exact to second order in the contrast, and with the Wiener
upper and lower bounds (see above). In fact, the Maxwell–Garnett and EMA es-
60 2. Nonlinear Conductivity and Dielectric Constant: The Continuum Approach

(n) (n)
timates for ge and Re reproduce the asymptotic estimates of Blumenfeld and
Bergman:
n + 1 g 2  − g2
ge(n) ∼ g − , (139)
2α(n) g
n + 1 R 2  − R2
Re(n) ∼ R − , (140)
β(n) R
where R = φ1 R1 + φ2 R2 , with R1 and R2 being the resistivities of phases 1
and 2, respectively. The agreement for small enough contrast (to second-order)
is a consequence of the fact that the effective behavior of weakly heterogeneous,
nonlinear materials with statistically isotropic microstructures is dependent only
upon the phase volume fractions (Blumenfeld and Bergman 1991). However, while
(n)
the small-contrast expansions of Blumenfeld and Bergman (1991) for ge and
(n)
Re diverge as n → 0 and ∞, respectively, and can therefore yield unphysical
results even at relatively small contrasts, the estimates provided by Eq. (125) do
not diverge and always yield physically meaningful results. The new Maxwell–
Garnett estimates presented in this section satisfy all the known rigorous bounds,
including the Wiener bounds and the Hashin–Shtrikman upper bounds of Ponte
Castañeda (1992b) derived earlier in Sections 2.4 and 2.5, and the lower bounds of
Talbot and Willis (1994, 1996) for nonlinear composites with statistically isotropic
particulate microstructures (with n ≥ 1).
Finally, let us point out that Gibiansky and Torquato (1998b) derived cross-
property bounds that link the effective conductivity of nonlinear disordered
materials to their effective elastic moduli. Such cross-property bounds were already
described in Section 7.9 of Volume I for linear materials, and will be presented in
Chapter 4 for nonlinear composites.

Summary
In this chapter, we described and discussed general procedures for estimating
the effective conductivity and dielectric constant of nonlinear materials. These
procedures, which represent the generalization of those described in Chapters 4 and
7 of Volume I for linear materials, provide bounds and estimates for the effective
conductivity and dielectric constant. One procedure leads to rigorous bounds and
estimates that are exact to first order in the phase property contrast, while the
second technique yields estimates that are exact to second order in the contrast.
One important difference between the results obtained by the two procedures
must be emphasized. By design, the results presented in Section 10.6 are exact
to second-order in the phase contrast, and thus are consistent with the asymptotic
results of Blumenfeld and Bergman (1991b), whereas the results presented in Sec-
tions 10.4 and 10.5 are nonlinear estimates that are exact only to first order in the
phase contrast. On the other hand, while the first-order results provide rigorous
bounds for the effective energy function of nonlinear materials (and hence their
 2.6. Second-Order Exact Results 61

generalized effective conductivity and dielectric constant), the second-order esti-

mates do not lead to any bound, either the lower or upper bound. Nevertheless,
comparison of these estimates with the numerical results and the known bounds
suggests that the second-order results provide accurate estimates for the effec-
tive conductivity and dielectric constant of general nonlinear materials, and in
particular, strongly nonlinear, power-law type composites.
3
Nonlinear Conductivity, Dielectric
Constant, and Optical Properties:
The Discrete Approach

3.0 Introduction
In this chapter we study nonlinear transport and optical properties of heteroge-
neous materials, representing their morphology by a discrete model. In particular,
we consider two-phase materials with percolation disorder which represents a
strong type of heterogeneity, although all the theoretical developments that are
described in this chapter (and throughout this book) are equally applicable to other
types of disorder. As we emphasized in Volume I, we believe that if a theory can
provide accurate predictions for transport and optical properties of materials with
percolation disorder, i.e., materials in which the contrast between the properties of
its two phases is strong, it should also be able to do so for almost any other type
of disorder.
There are many transport processes in which the current density is not related
to the applied field through a linear relation. Such nonlinearities, in the limit of
zero frequency, play an important role in many phenomena, including dielectric
breakdown, field dependence of hopping conductivity in heavily-doped semicon-
ductors, and many others. They are, at finite frequencies, the basis of nonlinear
optical phenomena in many disordered materials. By suitably tuning of the mate-
rial’s parameters, such as the volume fraction of the conducting material and its
nonlinear susceptibility, one can design a wide variety of composite materials with
specific properties that have important industrial applications. Chapter 2 described
the theoretical methods for estimating the effective conductivity and dielectric
constant of nonlinear disordered materials, based on the continuum models. In the
present chapter we consider several classes of nonlinear transport processes and
describe and discuss, based on the discrete models of heterogeneous materials, the
progress that has been made in understanding such phenomena.

3.1 Strongly Nonlinear Composites

In most of our discussions in this chapter we use a resistor network model for
describing transport in heterogeneous materials. Strongly nonlinear composites
are those in which the relation between the current i and the voltage v for any
 3.1. Strongly Nonlinear Composites 63

bond of the network is of power-law type and is given by

i = gv 1/n , (1)
where, as in Chapter 2, we interpret g as a generalized conductance of the bond.
Equation (1) defines a power-law resistor. If we replace i and v with q and P , the
flow rate and pressure drop in a tube or pore of a porous material, then Eq. (1) also
defines a power-law fluid, widely used for modeling flow of polymers (Bird et al.,
1987). Experimentally, Eq. (1) has been observed in certain classes of conductors,
such as ZnO ceramics. More generally, Eq. (1) describes the response of a material
when the magnitude of the applied field is very large, so that a linear relation
between i and v breaks down completely.
In theoretical analyses of a nonlinear materials, certain subtleties, that are not
encountered in linear systems, arise that must be addressed. For example, the nature
of boundary conditions that are imposed on the system is very important to the
solution of the transport problem. In our discussions in this chapter, we consider
only two-terminal networks, i.e., those into which one injects a constant current
at one node and extracts it at another node. Little is known, at least in the context
of the problems that we discuss here, about multiterminal networks (i.e., those
with more than one injection and one extraction node). It can be shown (see, for
example, Straley and Kenkel, 1984) that an equation similar to (1) is also valid for
two-terminal networks made of such nonlinear resistors. That is, if I , ge and V
are the macroscopic current, effective generalized conductivity and voltage drop
in the network, then I = ge V 1/n . To prove this, one proceeds as follows (Straley
and Kenkel, 1984). One defines a function
 v
Gkj (v) = ikj (v)dv (2)
0
for each bond kj and constructs a function


F = Gkj (vk − vj ). (3)
k,j

Because Gkj has a lower bound, so does the function F , and therefore it has
a minimum. The existence of this minimum is equivalent to the existence of a
solution to Kirchhoff’s equations for the resistor network. This can be easily shown
by calculating ∂F /∂vk and showing that it vanishes at node k, hence demonstrating
that the net current reaching node k is zero.
However, because this is a nonlinear system, the proof is complete only one
also proves that, in addition to existing, the solution to Kirchhoff’s equations is
also unique. This can also be proven (Straley and Kenkel, 1984) by assuming that
(1) (2)
the function F has two minima for the voltage distributions {vk } and {vk }. If
(s)
so, then F must also have a saddle point at {vk }, because along any path in the
(1) (2)
voltage space that connects the two distributions {vk } and {vk }, F must have a
maximum. If this saddle point exists, it must be a solution to Kirchhoff’s equations,
∂F /∂vk = 0. However, if the function i(v) [e.g., one that is defined by Eq. (1)]
is differentiable with a positive derivative, then it is not difficult to show that the
64 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

saddle point cannot exist, since F can be expanded in a series,

(s)
F = F (s) + ∂Gkj /∂(vk − vj )|s (vk − vj − vj
k,j
(s)

(s) (s)
(4)
+ vj ) + ∂ Gkj /∂(vk − vj )2 |s (vk − vj − vj + vj )2 + · · ·
2

k,j

In this equation, the linear term must vanish as ∂Gkj /∂vk = 0, and the quadratic
terms are all positive since we assumed that, di/dv > 0, and therefore the saddle
point does not exist, implying that the function F has a unique minimum, i.e., the
solution to Kirchhoff’s equations is unique. We note that Larson (1981) showed
that for slow flow of a power-law fluid in a porous medium with one injection
point and one producing point (which is the analogue of a two-terminal network)
in which flow in each pore is governed by Eq. (1), an equation similar to (1) is
also valid at the macroscopic scale, i.e., one has, at the macroscopic scale, Q =
GP 1/n , or, Q = G(P /L)1/n , where L is the length of the porous medium.
The reason that the general form of power-law (1) survives at the macroscopic
scale is that, such power-laws are self-similar and therefore they preserve their
identity under a microscopic-to-macroscopic transformation (that is, power laws
propagate self-similarly).
Calculating the voltage distribution in a nonlinear resistor network is a difficult
task, since the nonlinear Kirchhoff’s equations may have multiple solutions (all
but one of which would be unphysical), and thus one must be careful with the
numerical technique used in the simulation (see Uenoyama et al., 1992, for a
discussion of this point).

3.1.1 Exact Solution for Bethe Lattices

The simplest non-trivial discrete model of strongly nonlinear composites that can
be analyzed exactly is a Bethe lattice of coordination number Z, which is an
endlessly branching network without any closed loops, an example of which is
shown in Figure 3.1. We assume that each bond of the Bethe lattice is a power-law
conductor. If the lattice contains percolation-type disorder, then the solution of
the problem corresponds to the mean-field limit of percolation, i.e., the limit in
which the dimensionality of the system is d ≥ 6. To derive the solution we need
the appropriate rules for determining the equivalent conductance of power-law
resistors that are in series or parallel. It is not difficult to show that for N power-law
resistors in series or parallel, the equivalent conductivity gN is given by
⎧ N
⎨ i=1 gi , parallel,
gN =  −1/n (5)
⎩ N −n
, series.
i=1 gi

Suppose now that the bonds’ conductances are distributed according to a proba-
bility density distribution f (g). We derive an integral equation, from the solution
of which all the properties of interest can be computed (Sahimi, 1993a). Consider
 3.1. Strongly Nonlinear Composites 65

Figure 3.1. A Bethe lattice of coordina-

tion number Z = 3.

a branch of a Bethe lattice of coordination number Z which starts at the origin O.

The conductance of the branch can be computed by simply realizing that, it is the
equivalent conductance of one bond, say OA, of the branch that starts at O with
conductance gi in series with the branch that starts at A and has a conductance
Gi . Suppose now that the lattice is grounded at infinity and that a unit voltage has
been imposed at O. Then, the total conductance G of the network between O and
infinity is that of (Z − 1) branches that are in parallel. Therefore,
Z−1 $ %−1/n

1 1
G= + n . (6)
gin Gi
i=1

For an infinitely large Bethe lattice, G and Gi are statistically equivalent. Thus, if
H (G) represents the statistical distribution of G, we must have
⎧ Z−1 $ ⎫
  ⎨
1 %−1/n ⎬ Z−1
'
1
H (G) = · · · δ G − + n f (gi )H (Gi )dgi dGi .
⎩ gin Gi ⎭
i=1 i=1
(7)
If we now take the Laplace transform of both sides of Eq. (7), we obtain (Sahimi,
1993a)
 ∞
H̃ (s) = exp(−sG)H (G)dG
0
   $ %−1/n  Z−1
1 1
= exp −s + n f (g)H (G)dgdG . (8)
gn G

From the numerical solution of integral equation (8) we obtain all the properties of
interest. Note that, in the limit n = 1, Eq. (8) reduces to the corresponding integral
equation for Bethe lattices with linear resistors which was analyzed in Chapter
66 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

5 of Volume I. To our knowledge, no exact solution of Eqs. (8) and (9), for any
distribution f (g) and any value of n, has been derived.

3.1.1.1 Microscopic Versus Macroscopic Conductivity

In general, one may calculate two different effective conductivities for a Bethe lat-
tice. One is gm , the microscopic conductivity of the lattice, obtained by grounding
the lattice at infinity, imposing a unit voltage at site O of the lattice, and calcu-
lating gm as the current that flows out along one of the outgoing bonds connected
to O. It is not difficult to see that, aside from a constant factor, gm is the average
G of the distribution H (G), gm = ZG/(Z − 1). Using the properties of the
Laplace transform, one can then show (Sahimi, 1993a) that for a Bethe lattice of
coordination number Z,
  $ % Z−2
1 1 −1/n
gm = Z f (g)H (G) + n dg dG , (9)
gn G
which reduces to the equation given by Stinchcombe (1974) and Heinrichs and
Kumar (1975) for the n = 1 limit, derived in Chapter 5 of Volume I.
Equation (9) is valid for any f (g), the statistical distribution of the bond
conductances. Consider then percolation-type disorder, i.e., one for which
f (g) = (1 − p)δ(g) + ph(g), (10)
where p is the fraction of the conducting bonds with conductances that are selected
from h(g), which can be any normalized probability density function. It is then
not difficult to show that near the percolation threshold pc
 1/n
2c(Z − 1)2+1/n 1
gm = [n(J − n − 1)]1/n (p − pc )1+1/n .
1/n
hn (Z − 1) (Z − 1) n−1
(11)
In Eq. (11), c is a constant of order unity,  is the gamma function, J = 2 + [n],
where [n] denotes the integer part of n, pc = 1/(Z − 1) is the percolation threshold
of the Bethe lattice, and
 ∞
h(g)
hn = dg.
0 gn
The power law implied by Eq. (11) for the dependence on p of gm near pc was
first derived by Straley and Kenkel (1984), except that they did not provide the
exact form of the numerical factor given by Eq. (11). Equation (11) predicts that,
for the linear (n = 1) limit, one has
gm ∝ (p − pc )2 . (12)
On the other hand, the macroscopic or effective conductivity ge , which is what
one usually calculates for 2D or 3D networks, is the average current density per
unit applied field. The difference between the two cases is due to the geometry of
the Bethe lattice, which has a peculiar structure (lacking any closed loops while
keeping the length of the bonds constant), and the boundary conditions at infinity
 3.1. Strongly Nonlinear Composites 67

(Straley, 1977). To estimate ge one may proceed as follows (Straley and Kenkel,
1984). The average power P dissipated per unit volume is given by
$ %1+1/n
IV V
P= = ge , (13)
AL L
where A and L are, respectively, the surface area and linear size of the sample.
The voltage difference across a chain of the lattice is controlled by the geometrical
distance ξp between the ends of the chain, where ξp is the correlation length of
percolation. In general, ξp is less than L, the length of the chain, since the chain is
twisted. However, in a Bethe lattice, the chain performs a random walk in space,
implying that, ξp2 ∼ L, and therefore the current Ic that is carried by a chain is
given by
$ % $ %1/n
ξp V 1/n 1/2n V
Ic = = (p − pc ) . (14)
LL L
However, the chain will carry no current at all unless its ends are connected to the
sample-spanning percolation cluster. To calculate the probability of this connec-
tion, we note that the probability that a given site is connected by a particular bond
to the sample-spanning cluster is P (p), the percolation probability, and therefore,
near pc , the two ends of the chain are connected to the cluster with a probability
P 2 (p) ∼ (p − pc )2β . As β = 1 for the Bethe lattice, we find that the probabil-
ity that the chain is connected to the sample-spanning cluster is proportional to
(p − pc )2 . Therefore, the dissipated power is given by
P = P 2 (p)[(p − pc )1/2n (V /L)1/n ]1+n = (p − pc )(5+1/n) (V /L)1+1/n , (15)
which, when compared with Eq. (13), implies that, near pc ,
ge ∼ (p − pc )(5+1/n)/2 . (16)
Observe that the critical exponents that characterize the near threshold behavior
of both gm and ge depend on n. In particular, Eq. (16) indicates that if, in general,
near pc one has
ge ∼ (p − pc )µ(n) , (17)
where µ(n) is the analogue of the conductivity critical exponent µ for the linear
case; that is, for linear resistor networks near pc one has
ge ∼ (p − pc )µ , (18)
then in the mean-field approximation (the solution of which is obtained by solving
the problem on a Bethe lattice) µ(n) = µn is given by
1
µn = (5 + n−1 ), (19)
2
which implies that, in the linear (n = 1) limit, one has
ge ∼ (p − pc )3 , (20)
in agreement with the result derived in Chapter 5 of Volume I.
68 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

The effective conductivity of 3D linear resistor networks near pc follows the

power law (18) with µ  2.0. Therefore, Eq. (12) is similar to the power-law
behavior of the effective linear conductivity of 3D networks. Because by vary-
ing the coordination number Z of the Bethe lattice, its percolation threshold,
pc = 1/(Z − 1), can be adjusted to closely match that of a 3D network (for ex-
ample, the percolation threshold of a Bethe lattice with Z = 5 is pc = 1/4, which
is essentially the same as the bond percolation threshold of a simple-cubic lattice,
pc  0.249), it is clear that for linear transport (the limit n = 1) gm should pro-
vide an excellent approximation to the conductivity of 3D networks (Heiba et al.,
1982, 1992) and this has been shown to be indeed the case (Sahimi, 1993b). For
power-law transport considered here one may also use gm as an approximation
to the effective nonlinear conductivity of 3D networks. Figure 3.2 compares the
conductivity gm obtained from the numerical solution of Eqs. (8) and (9) with
Z = 5 with that of a simple-cubic network obtained by Monte Carlo calculations,
and it is clear that the agreement between the two is very good.

Figure 3.2. Comparison of the microscopic conductivity of a Bethe lattice of coordination

number Z = 5 with the effective conductivity of a simple-cubic network obtained by Monte
Carlo simulations (dashed curve). The bonds of the two lattices are power-law resistors with
a power-law exponent n = 0.4. The other two curves are, from top to bottom, the predictions
of Eqs. (32) and (31) (after Sahimi, 1993a).
 3.1. Strongly Nonlinear Composites 69

3.1.1.2 Effective-Medium Approximation for Bethe Lattices

Using Eq. (8), one can also construct an effective-medium approximation (EMA)
for power-law electrical transport in a Bethe lattice. As pointed out in Section 5.3.2
of Volume I, in the effective-medium approach, the probability distribution H (G)
is expected to achieve its maximum around a mean value G∗ , and thus we may
approximate H (G) by H (G)  δ(G − G∗ ), so that H̃ (s) = exp(−sG∗ ). Then,
Eq. (8) becomes
   $ %−1/n  Z−1
∗ 1 1
exp(−sG ) = exp −s + f (g)dg . (21)
gn (G∗ )n
To determine G∗ , we take the derivative of Eq. (21) with respect to s and evaluate
the result at s = 0; we find that
 ∞  −1/n 
1 1 G∗
+ − f (g)dg = 0. (22)
0 gn (G∗ )n Z−1
The effective conductivity ge of the network is obtained if we set in Eq. (22),
f (g) = δ(g − ge ) (because in the EMAapproach, all bonds of the network have the
same conductance ge ), in which case Eq. (22) yields, (G∗ )n = gen [(Z − 1)n − 1].
Substituting this result in Eq. (22) yields the desired EMA (Sahimi, 1993a):
 ∞ 
(Z − 1)g
− 1 f (g)dg = 0. (23)
0 {g n + [(Z − 1)n − 1]gen }1/n
Typical of all the EMAs, and similar to the EMAs derived in Volume I for the
effective linear properties, Eq. (23) provides accurate estimates of ge if the disorder
is not too strong, implying that the EMA cannot be very accurate near pc .

3.1.2 Effective-Medium Approximation for Three-Dimensional

Materials
Unlike the EMA for linear electrical transport which was derived and discussed in
Chapters 5 and 6 of Volume I, derivation of an EMAfor the nonlinear transport is not
unambiguous. In particular, several of such approximations have been proposed in
the past in order to estimate the effective conductivity of random resistor networks
with power-law conductances, all of which are purported to represent some sort
of an EMA. We should point out, however, that any reasonable EMA (and similar
approximations) should possess two important properties.
(1) It should reduce, in the limit n = 1, to the well-known EMA for linear random
resistor networks derived and analyzed in Chapters 5 and 6 of Volume I:
 ∞
g − ge
f (g)dg = 0. (24)
0 g + (Z/2 − 1)ge
(2) It should predict the same bond percolation threshold, pc = 2/Z, that the linear
EMA predicts, as the location of the percolation thresholds is independent of n.
70 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

One of the first EMAs for resistor networks with power-law conductances was
proposed by Sahimi (1993a), and is given by
 ∞ 
gZ/2
− 1 f (g)dg = 0, (25)
0 [g n + ((Z/2)n − 1)Gn ]1/n
which reduces to Eq. (24) in the limit n = 1, as it should. Another EMA was de-
rived by Tua and Bernasconi (1988) for a 2D isotropic continuum (with circular
inclusions), which was extended (Sahimi, 1993a) to networks of random conduc-
tances with coordination number Z. In this approach one first defines a tangent or
differential conductance σ by
di
σ = , (26)
dv
which, in the limit n = 1, yields the usual σ = g. Equation (26), when combined
with (1), yields
g
σ = v (1−n)/n . (27)
n
Consider now a two-phase material with its phase tangent conductances being σ1
and σ2 , both of which depend on the voltage v. Recall from Chapter 5 of Volume I
that in the EMA approach one inserts in the effective medium a bond with its true
conductance and determines the voltage fluctuations along this bond, i.e., the extra
voltage in the effective medium generated by the replacement of the conductance
of the bond in the effective medium by its true value. Carrying out this replacement
for component j (j = 1, 2) yields
σe (v1 , v2 )Z/2
vj = ve , (28)
σj (vj ) + σe (Z/2 − 1)
where ve is the voltage along the bond in the effective medium, and σe is the
effective value of σ . If we now apply the usual idea of an EMA, namely, that the
average of vj must be equal to ve (or that the average of the voltage fluctuations
must be zero), we obtain
 ∞
σj (vj ) − σe
f (σj )dσj = 0, (29)
0 σj + σe (Z/2 − 1)
which is the same as Eq. (24) except that the conductances σj and σe are functions
of the voltage. If the composite consists of two phases with (volume) fractions p
and (1 − p), then
pv1 + (1 − p)v2 = ve . (30)
The generalization of Eq. (30) to an N-component system is obvious. Equations
(29) and (30) are then used for determining σe . Having determined this quantity,
we calculate ge using Eq. (27).
To test the accuracy of these two approximations, let us consider a simple
case, namely, a resistor network with a percolation-type conductance distribu-
 3.1. Strongly Nonlinear Composites 71

tion, f (g) = (1 − p)δ(g) + pδ(g − 1). In this limit, Eq. (25) reduces to (Sahimi,
1993a)
 
(pZ/2)n − 1 1/n
ge = , (31)
(Z/2)n − 1
while Eqs. (29) and (30) predict that (Sahimi, 1993a)
$ %
p − 2/Z 1/n
ge = p(n −1)/n
2
. (32)
1 − 2/Z
Equations (31) and (32) do meet the two criteria that we set above, namely, that they
both reduce to the linear EMA for n = 1, and their predictions for the percolation
threshold are the same as in the case of linear transport: Both equations predict
that ge vanishes at p = pc = 2/Z, the same as that predicted by Eq. (24) for linear
transport. We can also compare the predictions of these EMAs with those for the
effective microscopic conductivity of the Bethe lattice. For example, for n = 1/2
Eq. (7) predicts that µn = 3, whereas the numerical estimate for 3D systems (see
below) for n = 1/2 is µn  2.35. However, unlike the two EMAs described above,
the region near pc in which the conductivity of a Bethe lattice is different from
that of a 3D network is so narrow that it can hardly be detected (see Figure 3.2).
Consider now the case in which the nonlinear composite material obeys a
current-field response of the following form
I = g|E|1/n E, (33)
which is a slight generalization of Eq. (1). Bergman (1989) and Lee and Yu (1995)
developed an EMA for computing the effective conductivity of this type of com-
posite materials. Bergman developed an EMA for any value of n, while Lee and Yu
considered only the n = 1/2 limit. In both cases a 2D continuum model (but with
percolation disorder) in which inclusions, consisting of long cylinders (or circles)
of nonlinear conductance gα (α = i, h), representing the inclusion and the host
matrix, were embedded in an effective medium with a nonlinear conductance ge .
As usual, one applies a uniform far field E0 , calculates the local field Eα , and in-
sists that Eα  = E0 . We supplement Eq. (33) by the usual electrostatic equations,
namely,
∇ · I = 0, ∇ × E = 0. (34)
Then, there exists a potential ϕ such that
E = −∇ϕ. (35)
If the potential ϕ is known, then, one can calculate Eα . Trial functions of the
following form,
ϕα (r, θ ) = −E0 (1 − bα )r cos θ, r < R, (36)

ϕe (r, θ ) = −E0 (r − bα R 2 /r) cos θ, r > R, (37)

72 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

are now selected, where bα is a variational parameter, and R is the radius of the
cylinder. With these choices, the energy functional of the composite is given by
 $ % 
1
Hα = ge + pα ge −1 + 4bα + 4bα2 + bα4 + pα ge (1 − bα )4 V04 , (38)
3
where pα is the volume fraction of material of type α. If we now define yα = gα /ge
and minimize the energy functional, we obtain
(1 + yα )bα3 − 9yα bα2 + 3(2 + 3yα )bα + 3(1 − yα ) = 0, (39)
which provides an equation for bα and ϕα , and hence Eα . If the system is such
that inclusions of nonlinear conductivity gi and volume fraction pi are randomly
distributed in a host of conductivity gh with volume fraction ph (pi + ph = 1.0),
then the EMA equation is simply given by
pi bi (yi ) + ph bh (yh ) = 0. (40)
Figure 3.3 compares the predictions of this EMA with the results of numerical
simulation, demonstrating the accuracy of the predictions.

0.0

-0.5

-1.0

[ ]
(n)
ge -1.5
log
(l )
ge

-2.0

-2.5

-3.0

0.0 0.2 0.4 0.6 0.8 1.0 1.2

F
(n)
Figure 3.3. Effective nonlinear conductivity ge , normalized by the effective conductivity
of the system in the linear regime, versus the fraction p of the good conducting bonds.
Solid curves are the predictions of the EMA, Eqs. (39) and (40), while symbols show the
results of numerical simulations. The results are, from top to bottom, for conductivity ratios
y = 0.5, 0.1, 0.01 and 0.001 (after Lee and Yu, 1995).
 3.1. Strongly Nonlinear Composites 73

3.1.3 The Decoupling Approximation

Equation (29) is quite general and can be used with a variety of composites. For
example, Wan et al. (1996) analyzed a general two-phase composite consisting
of materials a and b with volume fractions p and (1 − p), respectively, such that
the constitutive equation that related the current density I to the electric field E
was given by Eq. (33). The effective generalized conductivity ge is then defined
by the usual equation, I(x) = ge |E0 |1/n E0 , where · denotes an average over
the volume of the system. For each region i of the composite (i = a or b), the
I-E relation is approximated by, I(x) = gi |E(x)|1/n i E(x) ≡ σi E(x), where ·i
denotes an average over volume of region i. Similarly, for the composite as a
whole, one can define, I = ge |E(x)|1/n E(x) ≡ σe E(x). Therefore, similar to
our discussion presented above, the composite is treated as a linear material, but
with field-dependent conductivities σa and σb . It is not difficult to show that
1 ∂σe 2
E 2 i = E , (41)
pi ∂σi 0
where pi is the volume fraction of phase i [pi = p or (1 − p)]. One can also use
a decoupling approximation (Stroud and Wood, 1989) according to which,
1/2n
|E|1/n i  |E|2 i , (42)

so that the right-hand side of Eq. (41) is only a function of |E|2 i . Therefore,
Eq. (41), when written for both phases a and b, forms a set of coupled self-
consistent equations, the solution of which yields E 2 a /E02 and E 2 b /E02 . Given
these two quantities and Eq. (29), the effective generalized conductivity ge is then
estimated.
As an example, consider a 2D system. With f (σ ) = pδ(σ − σa ) + (1 − p)
δ(σ − σb ) and Z = 4, Eq. (29) yields
σe
ge = 1/n
E0
  1/2 
1 1/2n
= 1/n
(1 − 2p)(Xb − Xa ) + (1 − 2p)2 (Xb − Xa )2 + 4Xa Xb , Xi = gi |E|2 i ,
2E0
(43)
and from Eq. (41) one obtains, for example,
 
E02 2Xb − (1 − 2p)2 (Xb − Xa )
E a =
2
(2p − 1) + . (44)
2p [(1 − 2p)2 (Xb − Xa )2 + 4Xa Xb ]1/2
It can then be shown that Eq. (43) is identical with the Hashin–Shtrikman lower
bound for ge , derived by Ponte Castaneda et al. (1992) and described in Chap-
ter 2. Numerical simulations of the problem indicated close agreement with the
predictions of Eq. (43).
Two other methods that have been proposed for treating the problem of conduc-
tivity of a nonlinear material embedded in a matrix are the perturbation expansion
and the variational approach. Normally, these methods are described as part of the
74 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

continuum approach to these problems. However, since they were developed for
materials with percolation disorder, we describe them here, rather than in Chapter
2. What follows is a brief description of each method.

3.1.4 Perturbation Expansion

In this approach, which was developed by Gu and Yu (1992), Yu and Gu (1992),
and Yu et al. (1993), the expansion parameter is the nonlinear conductance gh of
the host or the matrix into which the inclusions are embedded. The electrostatic
potentials ϕ i and ϕ h for the inclusion and the host are expanded as
ϕ i = ϕ0i + gh ϕ1i + gh2 ϕ2i + · · · (45)

ϕ h = ϕ0h + gh ϕ1h + gh2 ϕ2h + · · · (46)

If ϒ = |E|1/n , one writes down an expansion for ϒ h = ϒ0h + gh ϒ1h + gh2 ϒ2h +
· · ·, with a similar expansion for ϒ i . For example, for n = 1/2 one obtains
ϒ h = (∇ϕ0h )2 + 2gh (∇ϕ0h ) · (∇ϕ1h ) + gh2 (∇ϕ1h )2 + · · · (47)
with a similar expression for ϒ i . The current density functions Ih and Ii can also
be expanded in powers of gh , Ih = I0h + gh I1h + gh2 I2h · · ·, with a similar expression
for Ii , since they can be expressed in terms of ϕjh and ϕji . When all of the expansions
are substituted into Eqs. (34), one obtains sets of simultaneous equations for the
functions ϕji and ϕjh for j = 1, 2, · · · Then, specifying the shape of the inclusion
and the boundary conditions allows one to solve for these functions, and thus
obtain the overall nonlinear effective conductivity of the material. However, such
perturbation expansions are not very accurate, particularly for percolation disorder,
unless many terms of the expansion are computed. In fact, they break down and
predict unphysical results if the nonlinearity is strong, e.g., if the applied field E0
is very large, since in this case the linear response vanishes identically in some
regions of the composite.

3.1.5 Variational Approach

Yu and Gu (1994,1995) considered a class of strongly nonlinear composites that
follow Eq. (33) with n = 1/2, where the nonlinear conductance g takes on different
values in the inclusions and in the host. Their approach is different from what we
described in Chapter 2, and is closer to what is of interest to us in the present chapter.
Yu and Gu considered the dilute limit in which a single cylindrical inclusion of
volume i is inserted in a host medium with a larger volume . With n = 1/2,
Eqs. (33)–(35) yield
∇ · [g(x)|∇ϕ(x)|2 ∇ϕ(x)] = 0. (48)
One now invokes the variational principle (see Chapter 2) to minimize the energy
functional,

H[ϕ] = I · E(x)d, (49)

 3.1. Strongly Nonlinear Composites 75

with respect to an arbitrary variation δϕ(x) away from the solution of Eq. (48),
provided that δϕ vanishes on the surface of the inclusions. When the minimum
condition is satisfied by a trial function ϕ̂, the effective nonlinear conductivity is
obtained from

ge E0  = Ĥ =
4
g(x)|Ê(x)|4 d, (50)


where Ê = ∇ ϕ̂. Thus, it remains to develop suitable trial potential functions ϕ̂.
The trial functions must be selected so as to satisfy the symmetry of the system
and the boundary conditions that are imposed on it. Thus, if the inclusions are
cylindrical, then, the trial functions, similar to Eqs. (36) and (37), are expansions
in cos mθ (with m = 1, 3, 5, · · ·), whereas for spherical inclusions one must use
Legendre functions. If the trial functions are selected to be Eqs. (36) and (37)
(which involve only the parameter bα ), then Eq. (39) is obtained again. Yu and Gu
(1995) improved the accuracy of the method by using higher-order terms in the
expansions. Hence, for a cylindrical inclusion of radius R, they used
ϕi (r, θ ) = (c11 r + c13 r 3 /R 2 + c15 r 5 /R 4 ) cos θ

+ (c31 r + c33 r 3 /R 2 + c35 r 5 /R 4 ) cos 3θ

+ (c51 r + c53 r 3 /R 2 + c55 r 5 /R 4 ) cos 5θ, r < R, (51)

for the inclusion phase, and
ϕh (r, θ ) = r cos θ + (b11 R 2 /r + b13 R 4 /r 3 + b15 R 6 /r 5 ) cos θ

+ (b31 R 2 /r + b33 R 4 /r 3 + b35 R 6 /r 5 ) cos 3θ

+ (b51 R 2 /r + b53 R 4 /r 3 + b55 R 6 /r 5 ) cos 5θ, r > R, (52)

where the external voltage has been set to be, E0 = 1. Thus, the problem involves
determining 18 variational parameters, the bi and ci . By using the boundary con-
dition for the potential ϕ on the surface of the cylinder (at r = R), three relations
between the 18 coefficients are found. Then, Eq. (49) is used to compute H, and the
result is then minimized with respect to the remaining 15 parameters. Compared
to the case in which Eqs. (36) and (37) are used, this procedure with 18 variational
parameters improves the accuracy of the predictions by about 10%.

3.1.6 Exact Duality Relations

In Chapters 4 and 5 of Volume I we described duality relations for the effective
conductivity of linear materials. We now consider the same relations for nonlinear
materials that are characterized by Eq. (33). Note that, in the notation of Eq. (33),
the limit n = ∞ corresponds to the linear conduction case [whereas Eq. (1) reduces
to the linear problem in the limit n = 1]. Recall that duality relations exist only
for 2D systems, and therefore only such materials (for example, thin films) are
considered here. We consider composites in which the nonlinear conductivity g
76 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

varies from phase to phase, but the exponent n is the same for all the components.
The duality relations that we describe here is due to Levy and Kohn (1998).
Consider a two-phase composite in which the local conductivities are defined
by
gj (|V|) = gj |V|1/n , j = 1, 2. (53)
The dual composite is another two-phase material with the same morphology, but
with phases that have the following local conductivity,
1 −n/(n+1) −1/(n+1)
gj (|I|) = = gj |I| , j = 1, 2. (54)
gj (|V|)
The effective conductivities of the two materials are expressed as

g ∗ [g1 (|V|), g2 (|V|); V0 ] = ge V0

1/n
, (55)
and
−1/(n+1)
gd∗ [g1 (|I|), g2 (|I|); I0 ] = ge(d) I0 , (56)
(d)
where ge is the effective conductivity of the dual composite, and
I0 = g ∗ [g1 (|V|), g2 (|V|); V0 ]V0 , (57)
is the magnitude of the current that flows through the primal composite, which
is also the magnitude of the volume-averaged electric field in the dual com-
1/n
posite. Since the effective conductivities of the dual materials satisfy, ge V0 =
(d) −1/(n+1)
1/[ge I0 ], we obtain an exact duality relation for heterogeneous (2D)
materials made of power-law conductors:
n/(n+1) 1
ge = (d)
. (58)
ge
We may consider the consequences of duality for percolation composites by
studying two limiting cases:
(1) A mixture of good conductors (nonlinear conductivity gM , exponent n) and
perfect insulators (nonlinear conductivity gI = 0). Then, an equation similar
to (17) must hold near the percolation threshold pc of the good conductor.
(2) A mixture of normal conductors (nonlinear conductivity gI , exponent n) and
superconductors (nonlinear conductivity gM = ∞). Then, similar to linear
resistor networks of conductors-superconductors for which one has, near pc ,
ge ∼ (pc − p)−s , we expect to have
ge ∼ (pc − p)−sn , (59)
where sn = s(n) is the analogue of the exponent s, defined above. Therefore, if
we take phase 2 to be a perfect insulator, then, the dual composite is a mixture
of normal conductors and superconductors. Using Eq. (59), we then find that
 3.1. Strongly Nonlinear Composites 77

(Straley and Kenkel, 1984; Levy and Kohn, 1998)

$ %
n 1
µ(n) = s − , (60)
n+1 n+1
which, in the limit n → ∞, reduces to the well-known relations, µ = s, for
2D linear percolation conductivity which was already mentioned in Chapters
2 and 5 of Volume I. Let us emphasize that the exponent µn = µ(n) used in
Eq. (60) is slightly different from that in Eq. (17).
When the ratio of the conductivities of the two components is finite, we expect,
similar to linear resistor networks studied in Chapters 2, 5, and 6, to have a scaling
representation of ge :
gI /gM
ge ∼ gM (p − pc )µn ± (z), z= , (61)
(p − pc )µn +sn
where the plus (minus) sign is for p > pc (p < pc ). Thus, returning to our two-
phase composite with conductivities g1 and g2 , we find that when g1  g2 > 0,
then, the primal composite has an effective nonlinear conductivity given by
 
g2 /g1
ge ∼ g1 (p − pc )µn  . (62)
(p − pc )µn +sn
−n/(n+1)
The dual of this composite has local nonlinear conductivities g1 
−n/(n+1)
g2 , and therefore
 
−n/(n+1) (g2 /g1 )n/(n+1)
ge ∼ g 2
(d)
(p − pc ) −n/(n+1)
µn
. (63)
(p − pc )µn +sn
Therefore, using the duality relation, Eq. (59), we find that the scaling functions
for the primal composite and its dual satisfy an exact relation:
 n/(n+1)
z
−1/(n+1) [zn/(n+1) ] = , (64)
1/n (z)
with the understanding that if the left-hand side of Eq. (64) uses −1/(n+1) with
a plus sign [see Eq. (61)], then, the right-hand side uses 1/n with a minus sign,
and vice versa.

3.1.7 Scaling Properties

The critical exponent µ, defined by Eq. (18), that characterizes the power-law
behavior of the effective linear conductivity ge of percolation composites near the
percolation threshold [see Eq. (2.74)], can be expressed as
µ = (d − 2)ν + ζ, (65)
where ζ is the linear resistance exponent (that is, the resistance R of a sample of
linear size L < ξp scale as, R ∼ Lζ /ν ), ν is the critical exponent of percolation
78 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

correlation length, ξp ∼ |p − pc |−ν , and d is the dimensionality of the composite,

with ν = 4/3 and 0.88 for d = 2 and 3, respectively. From our analysis of conduc-
tion in Bethe lattices with power-law conductors presented above and Eqs. (12),
(16) and (17), it should be clear to the reader that for any d-dimensional random
resistor network with power-law conductors, the exponent µn , defined by Eq. (17),
which is the analogue of µ, must depend on n. This is indeed the case. One can
rewrite Eq. (65) in a more general form (Kenkel and Straley, 1982)
1
µn = µ(n) = (d − 1)ν + [ζ (n) − ν], (66)
n
indicating explicitly that the n-dependence of µ must be through the resistivity
exponent ζ as ν is a purely topological property, independent of the transport
process. Numerical simulations and scaling analyses discussed below show that
this is indeed the case. In fact, extensive analysis of random resistor networks
with power-law conductors indicates that, for certain limits and values of n, the
exponent ζ̃ (n) = ζ (n)/ν is related to various topological properties of the network.
We now describe these relations which provide insight into the n-dependence of
ζ (n) and hence µn .
In general, as Eq. (66) indicates, µn is larger than µ, and therefore near pc the
conductivity curve for power-law transport is flatter than that of the linear transport.
Several exact relations between ζ (n) and the topological exponents of percolation
networks have been derived. We present the proof of one of these relations to give
the reader some idea about how they are derived. Blumenfeld and Aharony (1985)
proved that
ζ̃ (n = ∞) = Dr , (67)
where Dr = 1/ν is the fractal dimension of the red bonds in the sample-spanning
cluster, i.e., those that, if cut, split the cluster into two parts. If Mr is the number
of the red bonds, then for length scale L < ξp the fractal dimension Dr is defined
by, Mr ∼ LDr . To prove this relation, consider a two-terminal blob of bonds (a
subcluster of multiply-connected conducting bonds) near pc , and suppose that the
current through the blob is I , while the voltage drop between its two terminals is
V . Thus, the resistance R of the blob is given by, R = V 1/n /I . Now, ifwe select
any transport path between the two ends of the blob, we can write, V = j Rj ij n ,
where Rj is the resistance of bond j along the path, and ij is its current. Therefore,
⎡ ⎤1/n

$ ij %n
R=⎣ Rj ⎦ . (68)
I
j

However, ij < I , and therefore (ij /I )n should vanish as n → ∞, implying that the
blob resistance will be zero, and thus all of the resistance of the cluster (material) is
offered by the red bonds, hence proving Eq. (67). By similar arguments Blumenfeld
and Aharony (1985) also proved that
ζ̃ (n = 0+ ) = Dmin , (69)
 3.1. Strongly Nonlinear Composites 79

where Dmin is the fractal dimension of the minimum or chemical path between
two points of a percolation cluster, i.e., the shortest path between the two points.
Thus, for L < ξp , the minimum length Lmin scales with L as, Lmin ∼ LDmin , with
Dmin  1.13 and 1.34 in 2D and 3D, respectively. Moreover, Blumenfeld et al.
(1986) showed that
ζ̃ (n = 0− ) = Dmax . (70)
Here Dmax is the fractal dimension associated with the longest self-avoiding walk
(that is, a random walk in which the walker never visits any point more than once)
between the two terminals of the percolation network; if Lmax is the length of the
walk, then Lmax ∼ LDmax . Blumenfeld et al. (1986) also proved that
ζ̃ (n = −1) = Dbb , (71)
with Dbb being the fractal dimension of the backbone of percolation clusters. Note,
however, that it has not been possible to relate ζ (n = 1) to any of the topological
exponents. Blumenfeld et al. (1986) also proved that ζ (n) decreases monotonically
with n, and therefore dζ (n)/dn ≤ 0, with the equality holding at n = ∞. Using
values of the various exponents and fractal dimensions given in Table 2.3 of Volume
I, we see that in 2D, ζ (n = ∞) = 1, and ζ (n = −1)  2.18, whereas in 3D ζ (n =
∞) = 1, and ζ (n = −1)  1.6. Therefore, ζ (n) is a slowly-varying function of n.
In addition to direct numerical simulations, there are at least two other methods
for estimating µn and its dependence on n. These methods are generalizations of
those discussed in Chapter 5 for the linear conductivity, and in what follows we
describe them briefly.

3.1.7.1 Series Expansion Analysis

Meir et al. (1986) used a series expansion method to calculate ζ (n) for several
values of n. As discussed in Chapter 5 of Volume I for linear conduction, in this
method one defines a percolation susceptibility χp by
, -


χp = sij , (72)
j

where sij = 1 if the two sites i and j belong to the same percolation cluster and
sij = 0 otherwise, and the averaging is over all configurations of the occupied
sites (probability p) and unoccupied ones [probability (1 − p)]. We now define a
resistive susceptibility χR (n; C) for a cluster C of sites via



χR (n; C) = Rij (n), (73)
i∈C j ∈C

where Rij (n) is the nonlinear resistance between sites i and j . Then, the total
resistive susceptibility χR (n), defined by
, -


χR (n) = Rij (n) , (74)
j
80 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

is obtained by summing χR (n; C) over all cluster, weighting each cluster by its
probability of occurrence. This is usually done in terms of cumulants, whereby
one writes


χR (n) = N (C; d)pnb (C ) χRc (n; C). (75)
C
In this equation nb (C) is the number of bonds in the cluster, N (C; d) is the number
of ways per site a diagram, topologically equivalent to C, can be realized on a d-
dimensional simple-cubic lattice, and the sum is over all topologically inequivalent
diagrams C. Moreover, χRc (n; C) is the cumulant defined by


χRc (n; C) = χR (n; C) − χRc (n; γ ), (76)
γ ∈C

where the sum is over all subdiagrams γ of C. Then, the average resistance R(n)
is defined by
χR
R(n) = ∼ |p − pc |−ζ (n) . (77)
χp
Therefore, the procedure for series analysis of resistance of random resistor
networks with power-law conductors is as follows. For each cluster C, and fixed
values of n (the power-law exponent) and nb (the number of bonds in the cluster),
the resistance Rij (n) is computed (by solving the Kirchhoff’s equations). These
computations are carried out for all such clusters, from which χR (n; C) and hence
χR (n) are obtained. Writing



χR (n) = A(k, l)d l p k , (78)
k l
one obtains a power series in p for χR (n). Since, in practice, the number of possible
cluster configurations increases very rapidly with nb , the computed power series
cannot be very long. For example, Meir et al. (1986) calculated the first 11 terms
of the series. Another series is obtained for χp , the computation of which is very
simple since it involves only counting of the number of clusters’ configurations.
The resulting two power series are then analyzed by a Padé approximation method,
from which the average resistance R(n) and hence the resistivity exponent ζ (n)
are computed. Using the results of Meir et al. (1986) and Eq. (66), we present
in Figure 3.4 the variations of µn = µ(n) with n. This figure indicates that µn
decreases very rapidly with increasing.

3.1.7.2 Field-Theoretic Approach

Harris (1987) developed a field-theoretic approach to power-law transport, a
generalization of what we described in Chapter 5 of Volume I for the linear con-
duction problem, and derived an -expansion (where  = 6 − d, with d being the
dimensionality of the system) for ζ (n) which, to linear order in , is given by
 
 7(n − 1)
ζ (n) = 1 + 1− + O( 2 ). (79)
42 72
 3.1. Strongly Nonlinear Composites 81

Figure 3.4. Dependence of the power-law conductivity exponent µ(n) on the power-law
exponent n (after Sahimi, 1993a, plotted based on the results of Meir et al., 1986).

Since (Harris et al., 1975), ν = 1/2 + 5/84 + O( 2 ), we obtain, using Eq. (66),
  
5 17 1 1  7(n − 1)
µ(n) = − + − 3+ , (80)
2 84 n 2 84 36
which reduces, in the limit n = 1, to Eq. (5.233) of Volume I for linear conductivity.
Such -expansions, while predicting the correct general trends in the n- and d-
dependence of the exponent µ(n), are not very accurate for the practical cases of
d = 2 and 3.

3.1.8 Resistance Noise, Moments of Current Distribution, and

Scaling Properties
As discussed in Section 5.16 of Volume I for linear conduction, in a conducting
composite resistance noise manifests itself as voltage fluctuations, when the sample
is subjected to constant current bias, or as current fluctuations in content voltage
bias. The low-frequency power spectrum of the resistance fluctuations often varies
as 1/f , where f is the frequency. This is the so-called flicker or 1/f noise. [In
the literature on this subject, frequency is often denoted by f , instead of ω, so
that the resistance noise is often referred to as 1/f noise. Therefore, we depart in
this section from our standard notation in this book, and use f , instead of ω, to
denote the frequency so as not to confuse the reader.] The magnitude of resistance
82 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

noise depends on the morphology of the conducting sample. Resistance noise was
studied in Chapter 5 of Volume I for the case of linear composites. In this section,
we consider the same problem for power-law conductors described by Eq. (1).
Consider a sample composite in which each of the conducting elements of the
nonlinear resistors has the same average value, but is fluctuating independently
with a correlation δra δrb  = ρ 2 , where ra and rb are two resistances. Then, similar
to what was discussed in Chapter 5 of Volume I, the relative noise SR is calculated
from
 2(n+1)
δRδR ρ2 b ib
SR = =  , (81)
R2 r 2 ( b ibn+1 )2
where R is the resistance of the sample, ib is the current in the bonds, and
the sums are over all the current-carrying bonds. Note that the voltage noise
SV = δV δV /V 2 itself is given by, SV ∼ I 2n , and that for a homogeneous,
d-dimensional lattice of linear size L, SR = (ρ 2 /r 2 )/Ld .
For the sample-spanning percolation cluster at pc (or, equivalently, at length
scales L < ξp above pc ) the resistance noise scales with L as
SR ∼ L−bn , (82)
where bn = b(n) is the analogue of the exponent b for linear conduction,
Eq. (5.250) of Volume I. Rammal and Tremblay (1987) showed that
ζ̃n ≤ bn ≤ Dbb , bn ≤ 2ζ̃n − Dr , (83)
where, as before, ζ̃n = ζ̃ (n) = ζ (n)/ν, and Dbb and Dr are the fractal dimensions
of the backbone and the red bonds, respectively. As discussed in Chapter 5 of
Volume I, these bounds are also satisfied in the linear conduction case. Near the
percolation threshold pc ,
SR ∼ (p − pc )−κn , (84)
where, similar to the case of linear conduction, κn = κ(n) is a completely new
exponent independent of all the percolation exponents. Of course, κn and bn are re-
lated, κn = ν(d − bn ), and therefore the above bounds for b(n) can be immediately
converted to bounds for κn .
While SR is related to the 4th moment of the current distribution, one can, similar
to linear conduction discussed in Chapter 5 of Volume I, construct the general
moments Mq (x, x ) of the current distribution between two points x and x ;

(n+1)q
Mq (x, x ) = ib , (85)
b

where, as before, the sum is over all the current-carrying bonds of the network.
Then, for self-similar morphologies, such as the sample-spanning percolation clus-
ter at pc (or at length scales L < ξp above pc ), one can define an infinite hierarchy
of exponents τq (n) for |x − x | ∼ L:
Mq (x, x ) ∼ L−τ̃q (n) , (86)
 3.2. Nonlinear Transport Caused by a Large External Field 83

Similar to the case of linear conduction, the exponents τq (n) are independent of
each other. Moreover, Rammal and Tremblay (1987) proved that τ0 − τq (n) is a
decreasing convex function of q that satisfies the following inequalities,
q 1
τq−1 (n) ≤ τq (n) ≤ τq−1 (n) − τ0 , (87)
q −1 q −1
where the last of the two inequalities is valid only for q ≥ 1. For the sample-
spanning percolation cluster at length scales L < ξp , one has, τ̃q (n) = τq (n)/ν.
Rammal and Tremblay (1987) obtained approximate (but not particularly accurate)
estimates of these exponents.

3.2 Nonlinear Transport Caused by a Large External

Field
Another type of nonlinear transport process arises as a result of applying a large
external potential gradient or driving force to a disordered material. Examples are
abundant and include flux lines in superconductors (see, for example, Larkin and
Ovchinnikov, 1979; Brass et al., 1989; Feigel’man and Vinokur, 1990; Fisher et
al., 1991), various fluid flow phenomena in porous materials (for reviews see,
for example, Sahimi, 1993b,1995b), and sliding charge-density waves (see, for
example, Fisher, 1985; Gorkov and Grüner, 1989). Dielectric breakdown, to be
studied in Chapters 5 and 6, also belongs to this class of phenomena.
In general, one must distinguish between two different types of systems in which
transport is driven by a large external field. In one type the disorder is weak, and
thus the interactions between the transport carriers produce an elastic structure
that will be distorted but will not break. Charge-density waves, and invasion of a
porous material by a wetting front belong to this class of systems. In the second
type, disorder is strong and the elastic medium can break up, giving rise to transport
processes that are plastic or fluid-like. An important example is strongly-pinned
vortex lines in the mixed state of superconducting films. This type of systems,
unlike the first type, has not received the attention that it deserves, despite its
practical importance, and is the subject of this section.
When a large potential gradient or driving force is imposed on a material, it
induces bias in it in the sense that, in a d-dimensional system there will be an
“easy” or longitudinal direction which is the direction of the external potential
gradient, and along which transport takes place “easier” than the remaining (d − 1)
transverse directions. This bias also induces anisotropy in the material so that one
must introduce two correlation lengths, instead of one as in isotropic systems,
which are the longitudinal correlation length ξL and the transverse correlation
length ξT (see Figure 3.5). It is not unreasonable to assume that there is a critical
value of the external potential or force Fc such that for F ≥ Fc macroscopic
transport occurs. Suppose now that an external driving force F > Fc is imposed
on the system. The dimensionless potential, χ = (F − Fc )/Fc , plays the same
role as (p − pc ) in percolation. Because Fc represents a type of critical point or
84 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

Figure 3.5. A strong external potential induces dynamic anisotropy in a material, giving
rise to two correlation lengths ξL and ξT . Circle denotes the point at which the potential is
applied to the system (after Sahimi, 1993a).

threshold, it is not unreasonable to assume that near F = Fc one must have

ξL ∼ |F − Fc |−νL , (88)

ξT ∼ |F − Fc |−νT . (89)
The problem studied here has certain similarities with directed percolation (Kinzel,
1983; Duarte, 1986,1990,1992; Duarte et al., 1992). In directed percolation, the
bonds of a network are directed and diode-like. Transport along such bonds is
allowed only in one direction. If the direction of the external potential is reversed,
then there can be no macroscopic transport in the new direction. Similar to the
present problem, in directed percolation one also needs two correlation lengths to
characterize the shape of the percolation clusters. However, there is an important
difference between what we study here and directed percolation: The anisotropy
in our system is dynamically induced, whereas the bias and anisotropy in directed
percolation are static and fixed.
An example of such nonlinear systems is the model proposed by Narayan and
Fisher (1994) (see also the somewhat related model proposed by Herrmann and
Sahimi, 1993, and Herrmann et al., 1993). They considered a randomly-rough
surface onto which a fluid or a charge carrier is poured into isolated “lakes,” such
that initially a sample-spanning cluster of connected lakes does not exist. The
surface is then slowly tilted at an angle θ , such that the fluid spills out of the filled
lakes and feeds unfilled lakes further downhill. For θ < θc , where θc is the critical
value of the tilt angle, the filled lakes cluster together. The characteristic size of
such clusters increases as θ does, and diverges at θ = θc . Above θc the system
becomes depinned, so that the fluid or the charge carrier can flow from the top
to the bottom of the system. Near and above θc the transport process is highly
inhomogeneous and confined to narrow and well-separated channels, somewhat
 3.2. Nonlinear Transport Caused by a Large External Field 85

similar to Figure 3.5. Note that, under the influence of gravity, a force builds up
at the terminus of a cluster, rather than being uniform everywhere in it. Therefore,
when θ increases, clusters grow from their terminus sites, with a higher probability
of growing if they are already large. This implies that, the dominating flow paths
cannot be determined by a local analysis that searches for weak links in the system.
Rather, one must consider the entire system, i.e., the phenomenon is non-local.
The above description is a continuum one, but has a well-defined lattice coun-
terpart. In the lattice model, the sites represent the lakes, while the bonds are the
transport paths that connect the lakes. A force F is imposed on the lattice, and it
suffices for each site i to have outlets connecting it only to its d nearest neighbors
iα in the next plane downhill, where d is the dimensionality of the system. It is
assumed that the current flowing in a path depends only on the depth above the
lip of the lake it emerges from. Thus, a barrier biα is assigned to each outlet α
emerging from a site i which controls the current flowing through the outlet. The
barriers are selected randonmly and independently from a distribution. At each
site i of the lattice there is a depth of fluid hi . The current Iiα flowing through an
outlet α from a site i is zero if hi < biα − F , and
Iiα = (hi − biα + F ) if hi > biα − F. (90)
The exponent  characterizes the transport over the barrier lip. Narayan and Fisher
(1994) presented arguments that indicate that  = 3 + d/2 for a d-dimensional
system. Note that an increase in F is equivalent to uniformly lowering all the
barriers biα . √
Narayan and Fisher (1994) argued that ξT ∼ ξL . That is, we can imagine that
the consecutive events in which the bonds are filled with the flowing current are
in fact consecutive steps of a random walk in the (d − 1) transverse directions.
If so, the longitudinal direction acts as the time axis, and therefore the distance
that the random walker travels in the transverse direction should increase with √ the
square root of time (the usual law of random walks), implying that ξT ∼ ξL ,
and thus νT = νL /2. The random-walk argument can also be used to estimate
the upper critical dimension du of the system at and above which the mean-field
theory is exact. The clusters perform random walks in the (d − 1)-dimensional
transverse space, with the longitudinal direction acting as the time coordinate.
From the theory of random walks (Hughes, 1995) we know that if d − 1 > 2, then
two walks that start out close to each other have a finite probability of not crossing
each other, whereas for d − 1 < 2 they are certain to cross. Therefore du − 1 = 2
and hence du = 3. This immediately implies another significant difference between
this model and directed percolation for which du = 5 (Obukhov, 1980), and also
with isotropic percolation for which du = 6.
Narayan and Fisher (1994) studied various topological and transport properties
of this model. One surprising aspect of this phenomenon is that the critical expo-
nents that characterize the power-law behavior of the properties of interest above
and below, but near, the threshold Fc are not equal. Consider first the system be-
low the threshold. We write ξL ∼ |F − Fc |−νL , where superscript b signifies the
b

fact that the critical exponent is associated with the regime below the threshold. The
86 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

fraction of the sites Pb (F ) which are in clusters of length ∼ ξL scales as

−β̃nn
Pb (F ) ∼ ξL ∼ |F − Fc |βnn , (91)
where β̃nn = βnn /νLb . The mean distance p  travelled by a charge carrier from
its initial position at F = 0 (also called the polarization density) scales as
p  ∼ |F − Fc |1−γnn . (92)
The clusters of the sites (lakes) are fractal objects at length scales L  ξL with
a fractal dimension Df . The two exponents βnn and γnn are related through the
following scaling law,
γnn = νL (1 − β̃nn ), (93)
both above and below the threshold Fc . One can show that, in the mean-field
approximation, i.e., at d = 3, one has
3 4 2 1
νLb = , Df = , β̃nn = , γnn = . (94)
2 3 3 2
Consider next the regime above the threshold. An important property is the
fraction Pa (F ) of sites that feed charge carriers into the transport paths, i.e., the
analogue of Pb (F ) above the threshold. Near Fc ,
Pa (F ) ∼ (F − Fc ) , (95)
and it is clear that,  = β̃nn νLa ,
where superscript a indicates that the critical
exponent is associated with the regime above the threshold. In general, one has
the following scaling laws (Narayan and Fisher, 1994)
1+
νLa = , (96)
d −1
1 βnn
Df = (d + 1) − a , (97)
2 νL
Near Fc the mean current density I  flowing through the system obeys the
following power law
I  ∼ (F − Fc )µnn . (98)
The transport exponent µnn is then given by
1
µnn = (1 +  )(1 + ). (99)
2
Scaling law (99) is an interesting feature of this model for two reasons. First, it
implies that, unlike percolation, in this model the transport exponent is related to
the topological exponent . Secondly, it indicates a sort of non-universality, since
 is a local or microscopic quantity. In the mean-field approximation
3 1 3
νLa = ,  = , µnn = (1 +  ). (100)
4 2 4
Note that νLb
= νLa . In 1D the problem can be solved exactly and one obtains,
β̃nn = 0, νLb = 2, and γnn = 2 (note that in 1D only the regime below the threshold
 3.3. Weakly Nonlinear Composites 87

is physically meaningful). Since the upper critical dimension is du = 3, d = 2 is

the only physical dimension for which exact results are not known. Numerical
simulations of Narayan and Fisher (1994) yielded the following estimates
νLb  1.76, νLa  1.41, β̃nn  0.29,   0.41, Df  1.21. (101)
Note the significant difference between νLb and νLa .
Note also that, similar to con-
ventional percolation, all the exponents can be estimated from any two exponents,
e.g., νLa (or νLb below the threshold) and . The low value of Df implies that,
a large external field and the associated dynamical bias and anisotropy give rise
to transporting paths that are essentially restricted to a narrow cone (see Figure
3.5). Moreover, the fractal dimensions Df is considerably smaller than that of
2D percolation clusters, Df = 91/48  1.896. This can be understood if we con-
sider the problem on the Bethe lattice, i.e., the mean-field limit. In this lattice any
large external potential makes the network completely directed, since there are no
closed loops in the lattice. As a result, the backbone is made of directed branches
that have a quasi-1D structure, and thus the fractal dimension of the backbone is,
Dbb = 1 (for percolation Dbb = 2), implying that only a small subset of all the
bonds participate in the transport process.

3.3 Weakly Nonlinear Composites

We now consider a more general composite in which a material with nonlinear
I − V characteristics is embedded randomly in a host with either linear or nonlinear
I − V response. To our knowledge, the suggestion for theoretical consideration
of such composites was first made by Fleming and Grimes (1979) and Mantese
et al. (1981) (see also Yagil et al., 1994, for an interesting experimental study of
this problem). A concrete step toward this goal was taken by Gefen et al. (1986)
who proposed and studied the following problem. Consider a random resistor
network near the percolation threshold pc , which is driven by an external current
I . If I is sufficiently weak, then the response of the system is linear, and its linear
conductivity g () follows a power law similar to Eq. (18). If the external current I is
gradually increased, then for some critical current Ic the conductivity of the system
()
deviates significantly from its linear value ge . Gefen et al. (1986) suggested that
if L, the linear size of the sample, is greater than ξp , the percolation correlation
length, then
 x
Ic ∼ ge() , (102)
and that x = 3/2 in 2D. To confirm this prediction, they measured the electrical
conductivity of thin gold films near pc and found that x  1.47, in good agreement
with their prediction. If, however, L  ξp , then Ic would depend on L and Gefen
et al. (1986) proposed that
 −y
Ic (L) ∼ ge() (L) . (103)
Both x and y are supposed to be universal.
88 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

To explain their theoretical predictions and experimental measurements, Gefen

et al. (1986) considered a percolation resistor network in which each conducting
bond satisfied the following relation between the current i flowing through it and
the voltage v:
v = r i − rn i n , (104)
where r and rn are, respectively, the linear and nonlinear resistances, and n > 1.
Note that, in materials with inversion symmetry, the lowest value of n is 3. For small
enough i, the second term of the right-hand side of Eq. (104) is much smaller than
the first term, and therefore the resistor behaves linearly. For sufficiently large
i the second term becomes important, and the resistor is nonlinear. The critical
current ic at which the crossover occurs is found by equating the two terms of the
right-hand side of Eq. (104), resulting in
$ %1/(n−1)
r
ic = . (105)
rn
Composites that are described by Eq. (104), or by similar equations (see below),
are what we refer to as weakly nonlinear materials, since the leading order term
is still linear.
Let us now discuss important properties of nonlinear composites modeled as
a system of nonlinear elements with an I − V characteristic that is described by
Eq. (104) or by a similar equation. We do not discuss numerical simulations of such
phenomena which, although somewhat difficult, is conceptually straightforward
and requires no particular explanation

3.3.1 Effective-Medium Approximation

As the first problem in this class of composites, we describe the development of an
effective-medium approximation (EMA) for predicting the macroscopic behavior
of the composite. As usual, we use the terminology of a resistor network, although
all the discussions presented here are also applicable to continuum models (with
spherical inclusions). Consider a resistor network in which a fraction 1 − p of
the bonds are linear conductors with an I − V characteristic given by, i = gA v,
where gA is the conductance. The rest of the bonds, with a fraction p, are weakly
nonlinear conductors with a current-voltage characteristic given by
i = gB v + g (n) v 3 , (106)
which is another version of Eq. (104), written explicitly for the current i (rather than
the voltage v). We assume that g (n) v 2 /gB  1. To derive an EMA for this problem
(Stroud and Hui, 1988; Zeng et al., 1988; Zeng, Hui, Bergman and Stroud, 1989;
Hui, 1990a; Yang and Hui, 1991) we replace the resistor network by a uniform ef-
fective network of identical conductors with a current-voltage characteristic given
by
I = ge() v + ge(n) v 3 , (107)
 3.3. Weakly Nonlinear Composites 89

() (n)
where ge and ge are the effective linear and nonlinear response of the network,
respectively. In general, as our discussion throughout this book should have made
()
it clear, the effective linear conductivity ge in a binary random network with
components gA and gB can always be written as
ge() = F (gA , gB , p), (108)
where F is a function which, in general, depends on the geometry of the system.
(n)
Then the effective nonlinear response ge of the system is given by
 2
g (n) ∂F
ge =
(n)
()
. (109)
p ∂ge
That is, the effective nonlinear response is estimated based on an estimate of the
effective conductivity of the same material but in the linear regime. Recall from
Chapter 2 that the same sort of idea was developed by Ponte Castañeda (1992b) in
the context of the continuum models. The derivation of Eq. (109) will be discussed
in detail in Section 3.4, where we describe the derivation of a similar equation for
the dielectric constant of the same type of composites. Therefore, if the function
(n)
F can somehow be calculated, ge will also be determined from Eq. (109). Since
F is an estimate of the effective conductivity of a linear binary composite, we
may use the EMA, Eq. (24) (or, for example, the Maxwell–Garnett or any other
approximation), for linear resistor networks which for our binary network with
f (g) = pδ(g − gB ) + (1 − p)δ(g − gA ) is given by
() ()
gA − g e g B − ge
(1 − p) ()
+p ()
= 0. (110)
gA + ge (Z/2 − 1) gB + ge (Z/2 − 1)
(n)
Thus, the procedure for calculating ge by an EMA is as follows. One first solves
() ()
Eq. (110) for ge . This equation, which is quadratic in ge , defines the function F .
() (n)
Having determined ge , one utilizes Eq. (109) to calculate ge . Figures 3.6 and 3.7
compare the results of computer simulations in the square network in two limiting
cases with the EMA predictions. The numerical results in Figure 3.6, which are
for gA = 10, gB = 20, and g (n) = 0.1, are in excellent agreement with the EMA
predictions. The reason for the agreement is that the difference gB − gA is not large
and thus, as discussed in Chapter 5 of Volume I, the function F (i.e., the EMA
()
estimate) provides accurate predictions for ge . On the other hand, the numerical
results shown in Figure 3.7, which are for gA = 5000, gB = 10, and g (n) = 0.1,
agree only qualitatively with the EMA predictions because, as discussed in Chapter
5 of Volume I, in this case, due to the large difference between gA and gB , F (i.e.,
()
the EMA estimate) cannot provide accurate predictions for ge , which is consistent
with the general properties of the EMA.
It is clear that the development of an EMA for this class of composites involves
two stages. More generally, one may consider composites with more complex
I − V characteristics and develop a similar, but multistage, procedure for an EMA-
based computation of their effective transport properties. For example, one may
90 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

1.0

0.8

0.6

(n)
ge
(l )
g e 0.4

0.2

0.0
0.0 0.2 0.4 0.6 0.8 1.0
F
(n)
Figure 3.6. The effective nonlinear conductivity ge , normalized by the effective conduc-
tivity of the system in the linear regime, versus the fraction p of nonlinear conductors in
the square network. Solid curve shows the EMA predictions, while the symbols show the
results of numerical simulations for gB = 2gA = 20 and g (n) = 0.1 (after Yang and Hui,
1991).
250

200

150
(n)
ge
(l )
ge
100

50

0
0.0 0.2 0.4 0.6 0.8 1.0
F
Figure 3.7. Same as in Figure 3.6, but for gA = 5000, gB = 10, and g (n) = 0.1 (after Yang
and Hui, 1991).
 3.3. Weakly Nonlinear Composites 91

consider a composite material (Yu and Gu, 1993) in which a fraction p of the
system has an I − V characteristic given by, i = gB v + gn1 v 3 + gn2 v 5 , while the
rest of the composite, with fraction 1 − p, is made of linear conductors, i = gA v.
One may compute the effective linear and nonlinear response of such a composite
() (n1) (n2)
defined by, I = ge v + ge v 3 + ge v 5 , by first solving the EMA equation for
()
the effective linear conductivity of the composite ge . Then, an equation similar
(n1)
to (108) is used for computing the first nonlinear conductivity ge . The two
() (n1)
conductivities ge and ge so obtained are then used in a higher-order equation
(n2)
in order to compute ge .

3.3.2 Resistance Noise, Moments of Current Distribution, and

Scaling Properties
To explain the experimental data of Gefen et al. (1986) (see above) for their
weakly nonlinear conducting materials, Aharony (1987) established a relation be-
tween Gefen et al.’s problem and the distribution of currents in a linear random
resistor network. Consider first the regime L  ξp , which is equivalent to p = pc .
The total dissipated power P in the network, the bonds of which have an I − V
characteristic given by Eq. (104), is
1
1

P= r |ib |2 − rn |ib |n+1 , (111)
2 n+1
b b

where ib is the current in bond b, which depends implicitly on n, and the sums
are over all the conducting bonds of the network. Blumenfeld et al. (1986) had
already proved that
.
∂P .. 1
0 n+1
. = |ib | , (112)
∂n rn =0 n+1
b

where ib0 = ib (rn = 0). Therefore, to linear order in rn , we can replace ib by ib0
and write
1 rn
P = r M1 I 2 − M(n+1)/2 I n+1 , (113)
2 n+1
where I is the total current in the network, and
 2q

i0
Mq = b
, (114)
I
b

is the 2qth moment of the current distribution in the linear random resistor net-
work. As already discussed in Section 3.1.8 [see Eq. (86)] for the case of strongly
nonlinear composites, and in Section 5.16 of Volume I for linear systems, for
L  ξp the moments of the current distribution scale with L as

Mq ∼ L−τ̃q , (115)
92 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

where all the τ̃q s are distinct. This means that the current distribution in a linear
random resistor network is multifractal, i.e., each of its moments scales with L
with a distinct exponent, which is similar to the moments of the force distribution
in elastic and superelastic percolation networks described in Chapter 8 of Volume
I (see Stanley and Meakin, 1988, for a review of general properties multifractal
()
systems and distributions). Therefore, the effective linear resistance Re of the
()
network, which is obtained via Re = ∂ 2 P/∂I 2 , shows deviations from a constant
value for n > 1 and
 1/(n−1)
M1
I > Ic (L) ∼ ic ∼ ic Ly τ̃1 ∼ [ge() (L)]−y , (116)
M(n+1)/2
and therefore (Aharony, 1987)
τ̃(n+1)/2
1−
τ̃1
y(n) = . (117)
n−1
Since τ̃q is a monotonic and convex function (see, for example, Blumenfeld et
al., 1986), so also is y(n). For example, for d = 2 and 3 one has y(3)  0.08 and
0.06, and y(0)  0.18 and 0.1, respectively. This means that 0 < y(n) < y(1), and
therefore the linear regime I < Ic (L) extends to larger currents for larger linear
sizes L, implying that even a narrow nonlinear regime will be enhanced (see also
below) in a percolation network. A similar analysis for L  ξp yields (Aharony,
1987)
d − 1 − y(n)τ̃1
x(n) = , (118)
d − 2 + τ̃1
and therefore for d = 2 one finds that x(n) = 1.03 − y(n). Since y(n) > 0,
Eq. (118) does not agree with the experimental result of Gefen et al. (1986) for
any n, and therefore a simple percolation network in which each conducting bond
follows Eq. (104) cannot explain Gefen et al.’s data.
To study scaling properties of weakly nonlinear composites near the percolation
threshold pc , we must consider resistance and conductance fluctuations in linear
resistors networks. Recall from Section 5.16 of Volume I that, for a percolation net-
()
work near pc , the relative linear resistance noise, SR = δRδR/[Re ]2 , follows
the following power law [see also Eq. (84) for strongly nonlinear composites]
SR ∼ (p − pc )−κ , (119)
which defines the critical exponent κ. One can, in a similar fashion, consider con-
ductance fluctuations SG of a linear superconducting percolation network below
pc . In this case

SG ∼ (pc − p)−κ . (120)
It can be shown (Wright et al., 1986) that in 2D, κ = κ  . Given Eqs. (119) and (120),
we can discuss some of the scaling properties of weakly nonlinear composites near
pc .
 3.3. Weakly Nonlinear Composites 93

Stroud and Hui (1988) considered a composite with the following characteristic,
I(x) = g () (x)E(x) + g (n) (x)|E(x)|n E(x), (121)
where n ≥ 1, and g () and g (n) are the linear and nonlinear conductivities of the
medium, respectively, which depend, in general, on the spatial position x, and
the applied electric field (or voltage) E. Equation (121) is just another version
of (104), written explicitly for the current. As mentioned earlier, if one assumes
that all the components in the disordered composite have inversion symmetry, then
n = 2, which was the case studied by Stroud and Hui (1988). The volume-averaged
current I is defined by
I = ge() E0 + ge(n) |E0 |2 E0 , (122)
with E = E0 . Consider now the dissipated power for this composite which, in a
continuum formulation, is given (for n = 2) by
  
P = I · E d =  ge() |E0 |2 + ge(n) |E0 |4 . (123)

This equation, in which  is the volume of the composite, is the continuum analog
of Eq. (111). Using Eq. (121), we rewrite Eq. (123) as
  
P= g () (x)E · E + g (n) (x)|E|4 d = P2 + P4 . (124)

Then, to first order in g (n) (x), the second term of Eq. (124) is rewritten as,
P4 = g (n) (x)|E|4  = P4  , (125)
where the subscript  indicates that the electric field must be calculated from
the solution of the linear problem, i.e., in the limit, g (n) (x) = 0. In reality, the
difference E − E is of first order in g (n) , and therefore will contribute to P4 only
a second-order term. By a similar argument, one can show that
P2 = P2  . (126)
() (n)
Therefore, to first order in g (n) (x), the effective conductivities ge and ge are
given by (Stroud and Hui, 1988)

1 g () |E |2 
ge =
()
g ()
(x)|E  | 2
d = , (127)
|E0 |2 |E0 |2

1 g (n) |E |4 
ge(n) = g (n)
(x)|E  | 4
d = . (128)
|E0 |4 |E0 |4
Observe that Eq. (128) is the same as (50) for strongly nonlinear composites. Equa-
tions (127) and (128) are manifestations of an important result: The effective linear
and nonlinear conductivities of a weakly nonlinear composite can be calculated
from the behavior of the electric field in the linear problem.
Utilizing a similar line of analysis, Stroud and Hui (1988) proved another im-
portant property of weakly nonlinear composites, namely, that to first order in
94 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

(n)
g (n) (x), ge is essentially given by the mean square conductivity fluctuations in a
linear composite,
[g () ]2
ge(n) = , (129)
c
where g () is the root mean square conductivity fluctuations in the linear compos-
ite, and c is a constant with dimensions of energy. Note that, since the conductivity
fluctuations cause corresponding fluctuations in the current, which in turn are
related to the 4th moment of the current distribution (see above), Eq. (129) is
consistent with, but much more general than, Aharony’s result, Eqs. (113)–(118),
discussed above.
Using Eq. (129), one can now deduce the power-law behavior of the nonlinear
(n) (n)
conductivity ge near the percolation threshold pc . According to Eq. (129), ge is
given by conductivity, or resistivity, fluctuations of the linear conductivity problem.
Therefore (Stroud and Hui, 1988), using Eq. (119), we can write
(n)
ge
()
∼ (p − pc )−κ , (130)
[ge ]2
which, when combined with the power-law behavior of the effective linear
()
conductivity ge near pc , Eq. (18), yields
ge(n) ∼ (p − pc )2µ−κ , (131)
where µ is the critical exponent of the effective linear conductivity near pc . Note
that in a composite in which a fraction p of the material is superconducting and
the rest is made of weakly nonlinear conducting material, one has

ge(n) ∼ (pc − p)−2s−κ . (132)
With the help of Eqs. (131) and (132), one can construct a general scaling repre-
sentation for the effective conductivity of a composite, a fraction pM of which is a
() (n)
good weakly nonlinear conductor characterized by, I = gM V + gM V 3 , while the
rest of the composite, with a fraction (1 − pM ), is a poor weakly nonlinear conduc-
() (n) () () (n) (n)
tor which follows, I = gI V + gI V 3 , with gM  gI and gM  gI . Then,
() (n)
with z = [gI /gM ]/(p − pc )µ+s , p = |p − pc |, and considering Eqs. (131)
and (132), one can write (Levy and Bergman, 1994b)

ge(n)  gI p −2s−κ I (z) + gM p 2µ−κ M (z).
(n) (n)
(133)
The properties of the two scaling functions I and M vary in three distinct
regimes.
(1) In regime I, which is for pM > pc and |z|  1, the scaling function M must
be constant in order for one to be able to obtain Eqs. (130) and (131). It is then
straightforward to see that I must also be a constant.
(2) In regime II, which is for pM < pc and |z|  1, the scaling function I
must be constant so that one can recover Eq. (132). It is not difficult to see
 3.3. Weakly Nonlinear Composites 95

that in this case, M ∼ z4 . The morphology of the composite consists of a

nearly insulating matrix (dominated by the I phase) that contains conducting
inclusions (made of the M phase).
(3) In regime III, which is for pM  pc and |z|  1, the scaling functions I and
(n)
M must be such that the dependence of ge on p is cancelled.
(n)
In regime I, the contribution of the good conductor to ge decreases as pM →
pc+ (since 2µ − κ > 0), whereas the poor conductor’s contribution increases.
Therefore, if the contribution of the poor conductor happens to be dominant, we
(n)
will have a non-monotonic dependence of ge upon pM , with a maximum very
close to pc , in regime III, and a minimum somewhere above it, in regime I. On the
other hand, in regime II (pM < pc ), the contributions from both components in-
crease as pM → pc− . Therefore, one cannot in general determine which component
(n) () () (n) (n)
makes the dominant contributions to ge without specifying gI /gM , gI /gM
and p.
Based on such considerations, then, one can write
⎧ 
⎪
⎪ gM p 2µ−κ + gI p −2s−κ ,
(n) (n)
regime I,
⎪
⎨

ge(n)  gM [gI /gM ]4 p −2µ−4s−κ + gI p −2s−κ ,
(n) () () (n)
regime II,
⎪
⎪
⎪
⎩ (n) () () (2µ−κ)/(µ+s) 
+ gI [gI /gM ]−(2s+κ )/(µ+s) , regime III.
(n) () ()
gM [gI /gM ]
(134)
These scaling function representations are very similar, in their general form,
to those for low-field Hall conductivity described in Section 5.17 of Volume I.
Numerical simulations of Levy and Bergman (1994b) confirmed the validity of
these scaling laws.

3.3.3 Crossover from Linear to Weakly Nonlinear Conductivity

Equations (122), (130) and (131) enable us to derive the critical current for the
crossover from linear to weakly nonlinear regime. As discussed above, where we
derived Eq. (105), the critical voltage Vc or electric field Ec is obtained by equating
the two terms of the right hand side of Eq. (122). This yields
 
() 1/2
ge
Vc ∼ (n)
, (135)
ge
from which the critical current Ic is obtained (Blumenfeld and Bergman, 1991a):
 (1+κ/µ)/2
Ic ∼ ge() . (136)

We may interpret Eq. (136) as meaning that, the exponent x defined by Eq. (102),
is given by, x = 12 (1 + κ/µ). In 2D, where κ  1.12, Eq. (136) predicts that,
x  0.93, which still does not agree with Gefen et al.’s measurement, x  1.47,
but is closer to it than the prediction of Eq. (118).
96 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

More generally, let us consider a weakly nonlinear composite with percolation

disorder. Specifically, we consider two limiting cases.
(1) A composite in which a (volume) fraction p of the material is made of weakly
nonlinear conductors that follow Eq. (121), while the rest of the composite,
with a fraction (1 − p), is insulating. Then, near the percolation threshold pc ,
the critical current Ic and voltage Vc (or, equivalently, the critical electric field
Ec ) follow the following power laws,
Ic ∼ (p − pc )w , (137)
Vc ∼ (p − pc ) .v
(138)
(2) We also consider a composite a fraction p of which is made of superconducting
materials, while the rest of the system, with a fraction (1 − p), is made of a
weakly nonlinear conducting material with an I − V (or I − E) characteristic
that is given by Eq. (121). Then, we define the critical exponents w and v  by

Ic ∼ (pc − p)w , (139)
v
Vc ∼ (pc − p) . (140)
For the first limiting case, we use Eqs. (132) and (136) and the appropriate
() (n)
scaling laws for ge and ge to obtain
1 1
v= (κ − µ), w = (κ + µ). (141)
2 2
Since κ + µ > 0 while κ − µ < 0, Eq. (141) implies that, as pc is approached, the
nonlinear effect is enhanced, so that very close to pc , even a very small Ic would
be enough for a crossover from linear to weakly nonlinear conductivity behavior.
For the second limiting case (Yu and Hui; 1994; see also Hui, 1990b, 1994) one
has
1 1
v  = (κ  + s), w  = (κ  − s), (142)
2 2
where s is the critical exponent that characterizes the power-law behavior of the
effective linear conductivity of conductor-superconductor percolation composites
near pc , ge ∼ (pc − p)−s , and κ  is defined by Eq. (120). Using the numerical
estimates of the exponent s  1.3 and 0.73, and κ   1 and 0.4 for d = 2 and 3,
respectively, we find again that Ic vanishes as p → pc− , so that the nonlinear effect
is enhanced.
More generally, if one replaces the insulating material with a linear material
with conductivity g0 (the first limiting case described above), and let h = g0 /g () ,
then one has a general scaling equation for Ic (Yu and Hui, 1994):
Ic = (p − pc )(κ+µ)/2 I [h(p − pc )−(s+µ) ], (143)
which is completely similar to Eqs. (61) and (62). The universal scaling function
I (z) has the properties that, I (z) → constant as z → 0, while it behaves for
 3.3. Weakly Nonlinear Composites 97

large z as a power law in z. For length scales L  ξp (which is equivalent to

p = pc ), where ξp is the percolation correlation length, one can write

Ic = h(κ+µ)/2(s+µ) I [hL(s+µ)/ν ], (144)

where I (z)is another universal scaling function such that I
→ constant as
z → ∞, while I has a power-law dependence on z for z → 0. A similar scaling
function representation can also be derived for Vc . Hence

Vc = (pc − p)(κ +s)/2 V [h(pc − p)−(µ+s) ]. (145)
We note here that one may use the EMA to not only obtain estimates of the
exponents v, w, v  and w , but also explicit expressions for the scaling functions
I , I , and V . All one must do is using Eq. (24) to estimate ge and Eq. (109)
()

to compute ge . Then, it can easily be shown that, w = v  = 1/2.

(n)

The foregoing scaling laws are valid when one has cubic nonlinearity, i.e., when
n = 2 in Eq. (121). Zhang (1996a) and Gao et al. (1999) generalized these results
to any n. For the first limiting case, i.e., a composite of insulating and weakly
nonlinear conducting materials near pc , Gao et al. (1999) obtained the following
estimates,
νd − ζ − µ 1 + (νDbb − 1)−n/2 (ζ − 1)n/2+1 − ζ
v= + , (146)
2 n
νd − ζ + µ 1 + (νDbb − 1)−n/2 (ζ − 1)n/2+1 − ζ
w= + , (147)
2 n
where Dbb is the fractal dimension of the backbone of the percolation cluster,
and ζ is the resistivity exponent defined by Eq. (65). For the case of a composite
of superconducting and weakly nonlinear conducting materials, Zhang (1996a)
obtained the following estimate,
s 12−n κ  [(n + 2)/2]
v = + νd + , (148)
2 2 n n
where κ  [(n + 2)/2] is the exponent associated with the conductance fluctuations
below the percolation threshold defined above. In the limit n = 2 Eq. (148) reduces
to (142). Numerical simulations for testing the validity of these predictions were
reported by Levy and Bergman (1993, 1994b) and Zhang (1996b).

3.3.4 Exact Duality Relations

Similar to linear and strongly nonlinear conducting composites, weakly nonlinear
heterogeneous materials also satisfy some exact duality relations in 2D which we
now describe. These relations were derived by Levy and Kohn (1998), and parallel
those already described for strongly nonlinear composites in Section 3.1.6.
Consider a two-phase weakly nonlinear composite material for which the I − V
characteristic is given by Eq. (121) (written in terms of the voltage V rather than
98 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

the electric field E). The local conductivities of the two-phase material are given
by
() (n)
gj (V) = gj + gj |V|n , j = 1, 2. (149)
The dual composite is another two-phase material with the same microgeometry,
but with its phases having the following local conductivity,
(n)
1 1 gj
gj (|I|) = = () − () |I|n . (150)
gj (|V|) gj [gj ]n+2
The effective conductivities of the two components can be expressed as

g ∗ [g1 (|V|), g2 (|V|); V0 ] = ge() + ge(n) V0n ,

() ()
(151)
for the primal composite, and

gd∗ [g1 (|I|), g2 (|I|); I0 ] = ge(,d) + ge(n,d) I0n ,

() ()
(152)
for the dual composite, with
 
I0 = g ∗ g1 (|V|), g2 (|V|); V0 V0 ,
() ()
(153)

being the magnitude of the current that flows through the primal composite, which
is also the magnitude of volume-averaged electric field in the dual composite.
All the notations have the same meaning as for the strongly nonlinear composites
discussed earlier. To first order in the local nonlinear conductivity g (n) , the effective
conductivities satisfy
(n,d)
1 ge
ge() + ge(n) V0n = (,d)
− I n.
(,d) 2 0
(154)
ge [ge ]
This relation leads us to
1
ge() = (,d)
, (155)
ge
which is the same as the well-known duality relation for linear composites, and
(n,d)
ge
ge(n) V0n = − I n.
(,d) 2 0
(156)
[ge ]
Equation (156) implies immediately that for cubic nonlinearity (n = 2),
(n) (n,d)
ge ge
()
=− (,d) 2
. (157)
[ge ]2 [ge ]
Similar to the case of strongly nonlinear composites described in Section 3.1.6,
we can extend this analysis to weakly nonlinear materials near the percolation
threshold and investigate its consequences. As discussed in Section 3.3.2, if we
 3.3. Weakly Nonlinear Composites 99

() (n)
have a mixture of good conductors [conductances gM and gM ] and perfect
insulators, then near pc we expect to have [see Eq. (130)]
(n) (n)
ge gM
()
∼ ()
(p − pc )−κ , (158)
[ge ]2 [gM ]2
where κ is the exponent for the resistance noise introduced and described above.
() (n)
Similarly, for a mixture of normal conductors [conductances gI and gI ] and
superconductors near pc , one must have [see Eq. (132)]
(n) (n)
ge gI 
()
∼ ()
(pc − p)−κ , (159)
[ge ]2 [gI ]2
where the exponent κ  was also defined above. Using the duality relations described
above, one can then show that κ = κ  which, as discussed above and in Chapter 5
of Volume I, also holds for linearly conducting composites.
The foregoing discussions can be extended to the case in which the ratio gI /gM ,
for both the linear and nonlinear conductivities, is finite. In this case Eq. (133)
should hold for the primal composite and its dual, both above and below the
percolation threshold pc . Then, using the above duality relations, one can show
that the scaling functions ±,I and ±,M [where the plus (minus) sign is for
p > pc (p < pc )] and their dual counterparts satisfy the following relations
(d) (d)
M = I , I = M , (160)
with the understanding that if the left-hand side of Eqs. (160) uses the scaling
function with the plus sign, then, the right-hand side uses the function with the
minus sign, and vice versa.

3.3.5 Comparison with the Experimental Data

The relevance of the above models of weakly nonlinear composites and their
properties to modeling real materials was established by experimental studies of
Lin (1992), who measured I − V characteristics of PrBa2 Cu3 O7−δ , a compound
thought for a long time to be superconducting, although it now appears that it
is a normal conductor, even at very low temperatures. Figure 3.8 presents the
results for four different experiments with the same compound, indicating highly
nonlinear behavior beyond a current of about Ic  0.02 A. If we assume that
Eq. (121) describes the I − V behavior of the material, then one may estimate the
exponent n by fitting the data to this equation. Lin found that n = 1 and 2 both
represent the data relatively well. When the critical current Ic was plotted versus the
()
linear conductivity ge , the data shown in Figure 3.9 were obtained. The straight
line passing through the data has a slope x  0.6. On the other hand, Eq. (136)
predicts that, x = (1 + κ/µ)/2, which implies that x  0.93 in both 2D and 3D,
if we use µ  1.3 and 2.0, and κ  1.12 and 1.60 in 2D and 3D, respectively.
This estimate of x does not agree with Lin’s measurements. However, if we use
100 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

Figure 3.8. A typical nonlinear I − V curve for a PrBa2 Cu3 O7−δ compound at 300 K.
Symbols show the data for four different samples (after Lin, 1992).

Figure 3.9. Logarithmic plot of the critical current Ic versus the effective linear conductivity
(l)
ge for four samples of PrBa2 Cu3 O7−δ . Solid circles show the data for an Ag-] added
(l)
sample. The straight line represents Ic ∼ [ge ]0.6 (after Lin, 1992).
 3.4. Dielectric Constant of Weakly Nonlinear Composites 101

µ  2.5 and κ  5.14 for the 3D Swiss-cheese model, i.e., the model in which
spherical inclusions are distributed randomly in a uniform matrix, then Eq. (133)
predicts that x  0.74, which is only about 20% larger than Lin’s measurements
which, given the scatter in the data shown in Figure 3.9, is quite acceptable.

3.4 Dielectric Constant of Weakly Nonlinear

Composites
Most of our analysis of the effective conductivity of nonlinear composites is equally
applicable to the problem of computing the effective dielectric function of the same
materials, with the effective conductivities replaced by the effective dielectric con-
stant e . Thus, in this section we summarize the most important results and discuss
their ramifications for the static case. Frequency-dependent dielectric constant will
be described in the next section.
Consider a two-component composite material in which each component is
described by a weakly cubic nonlinear relation between the electric displacement
field D and the electric field E given by
() (n)
Di = i Ei + i |Ei |2 Ei , i = 1, 2. (161)
(n) ()
In the analysis that follows we assume that, i |E|2  i . We wish to compute
(n)
the effective nonlinear dielectric function e defined by
D = e() E + e(n) |E|2 E, (162)
()
where e is the effective linear dielectric function of the composite when the
electric field is small enough, and · denotes an average over the volume of the
composite.
A general approximate scheme for this problem was proposed by Zeng et al.
(1988) which we now summarize and discuss. As in the case of the effective
conductivity, the linear effective dielectric function can always be written as
 
() ()
e() = F 1 , 2 , p1 , (163)

which is the analogue of Eq. (108). Here p1 is the volume fraction of the 1
component, and F is an estimate of the effective dielectric constant which, in
general, depends on the morphology of the composite. We initially assume that
()
only component 1 is nonlinear, so that 2 = 2 , and therefore we can invoke an
approximate nonlinear form of Eq. (163):
e = F (1 , 2 , p1 ), (164)
() (n)
where, i = i + i |Ei |2 , and |Ei |2  is the mean square of the electric field
in the ith component in the linear limit. We must keep in mind that Eq. (164) is
valid only if 1 and 2 are constant in their respective component, implying that E
is uniform in the nonlinear component.
102 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

()
The function F is now expanded in a Taylor series around e :
   
e  F 1 , 2 , p1 + F  1 , 2 , p1 1 |E1 |2 ,
() () () (n)
(165)

where F  = ∂F /∂1 . However, one can express F  exactly in terms of the average
squared electric field in component 1 in the linear limit:
 
|E1 |2 
()
∂e  
 () ()
p1 = ≡ F 1 , 2 , p1 , (166)
E02 ∂1

where E0 is the external field. Therefore,
(n)
1
e = e() + F  |F  |2 E02 , (167)
p1
which means that, by the definition of the effective nonlinear dielectric function
(n)
e , we obtain
(n) $ % . .
1 ∂e .. ∂e ..
e =
(n)
. (168)
p1 ∂1  . ∂1 .
Equation (168) is the analogue of Eq. (109) for the nonlinear conductivity.
We can generalize this result to composites in which both components are weakly
nonlinear. Hence, we write
(n) (n)
1 
e = e() + F  |F  |E 2 + 2 F2 |F2 |E02 , (169)
p1 1 1 0 p2
where Fi = ∂e /∂i (i = 1, 2). Therefore,
(n) (n)
 
e(n) = 1 F1 |F1 | + 2 F2 |F2 |. (170)
p1 p2
Equation (170) also suggests an analogous generalization for the effective nonlin-
(n)
ear conductivity ge , which would then represent a generalization of Eq. (109).
One can now use this general method of approximation and study its properties in
certain limits.

3.4.1 Exact Results

() (n)
There are a few simple morphologies for which e , and hence e , can be com-
puted exactly. In one such morphology the two components are arranged in the
form of cylinders that are parallel to the external field. The cylinders do not have
to have circular cross sections. For this model,
() ()
e() = p1 1 + p2 2 . (171)
Then, it is not difficult to see that,
(n) (n)
e(n) = p1 1 + p2 2 . (172)
 3.4. Dielectric Constant of Weakly Nonlinear Composites 103

The second morphology for which the effective nonlinear dielectric function can
be exactly computed is one in which the components are arranged in the form of
flat slabs perpendicular to the external field. For this case,
1
e() = () ()
, (173)
p1 /1 + p2 /2
from which one obtains, using Eq. (170),
(n) (n)
1 2
e(n) = p1 () ()
+ p2 () ()
. (174)
[p1 + 1 p2 /2 ]4 [p2 + 2 p1 /1 ]4
A perturbation expansion, similar to what we described in Section 3.1.4 for the
effective conductivity of strongly nonlinear composites, was also developed by Yu
et al. (1993).

3.4.2 Effective-Medium Approximation

As the reader probably knows by now, according to the EMA, the effective
dielectric constant is one of the solutions of the following quadratic equation,
() () () ()
1 − e 2 − e
p1 () ()
+ p2 () ()
= 0. (175)
1 + (Z/2 − 1)e 2 + (Z/2 − 1)e
() ()
If both 1 and 2 are real and positive, then the physically relevant solution
()
of the EMA is also the positive one. Equation (175) is now solved for e , from
() ()
the solution of which the functions, Fi = ∂e /∂i , are computed which, when
substituted in Eq. (170), yield the EMA prediction for the effective nonlinear
(n)
dielectric constant e .

3.4.3 The Maxwell–Garnett Approximation

In Chapter 2, as well as Section 4.9.4 of Volume I, we described the Maxwell–
Garnett (MG) approximation for the effective linear conductivity and dielectric
constant of composite materials based on the continuum models. As discussed
there, the MG approximation is most appropriate for a heterogeneous solid in
which one of the components plays the role of a matrix, while the other acts
as an inclusion. Therefore, assuming that component 2 is the matrix, the MG
approximation takes the following form:
() ()
1 (2p1 + 1) + 22 (1 − p1 ) ()
e() = () ()
2 (176)
1 (1 − p1 ) + 2 (2 + p1 )

Using Eq. (176), the functions Fi = ∂e /∂i are computed which, when sub-
()

stituted in Eq. (170), yield the MG estimate for the nonlinear dielectric function
104 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

(n)
e . Hui (1990a) extended the MG approximation to a more general composite
for which, D =  () E +  (n) |E|n E
We should emphasize again that in the type of nonlinear problems that we are
discussing here, the geometry of the system and the boundary conditions are very
important and have a profound influence on the overall behavior of the system.
As a matter of fact, every result described so far is valid only for two-terminal
systems, and essentially nothing is known for multi-terminal ones.

3.5 Electromagnetic Field Fluctuations and Optical

Nonlinearities
In this section we continue the discussion that we began in Chapter 4 of Volume
I and describe and discuss advances in understanding optical properties of disor-
dered materials, and the effect that constitutive nonlinearities may have on such
properties. The main conceptual framework for our discussions are the discrete
models, in the form of disordered resistor networks. Hence, we are particularly
interested in the optical properties of composite materials with percolation-type
disorder. In general, as our discussions in Section 3.3 made it clear, disordered
solid materials with percolation-type disorder are very sensitive to the magnitude
of the external electric field because, (1) their macroscopic transport and optical
properties are controlled by their backbone, i.e., the current-carrying part of the
network, and (2) because of the sparse morphology of the backbone, and in partic-
ular its low fractal structure at length scale L  ξp (Dbb  1.675 and 1.8 in 2D
and 3D, respectively), the effect of the external field accumulates around its weak
points, i.e., its red bonds which are those that, if cut, would split the backbone
into two pieces. Therefore, such materials should have, and indeed do have, much
larger nonlinear macroscopic response than those of their constitutes.
Even when there is no apparent constitutive nonlinearities in the conduction
properties of the phases of a disordered material, percolation disorder may lead to
nonlinear macroscopic response. An interesting manifestation of this phenomenon
was provided by theAC and DC conductivities of a percolation composite of carbon
particles embedded in a wax matrix (Bardhan, 1997). In this composite, neither the
carbon particles nor the wax matrix exhibits any nonlinearity in their conduction
properties; nevertheless, the macroscopic conductivity of the composite increases
significantly when the applied voltage increases by only a few volts. Such a strong
nonlinear response can be attributed to quantum tunneling between the conducting
carbon particles, a distinct feature of electrical transport in disordered solids near
the percolation threshold pc .
Likewise, local fluctuations in the electromagnetic field and the resulting
enhancement of nonlinear optical properties in disordered solids, such as metal-
dielectric composites with percolation disorder, especially near pc , constitute an
important set of phenomena, since such composites have high potential for various
applications. Nonlinear effects manifest themselves in two distinct ways:
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 105

(1) If the applied electric field or current exceeds a critical threshold, then, at
zero frequency, strong nonlinearity results in the breakdown of the conducting
elements of a composite. The critical field decreases to zero as the volume
fraction of the conducting component approaches pc , hence indicating that
such composites become progressively more responsive to the external field
as pc is approached. This phenomenon is what we have referred to as threshold
nonlinearity; it will be studied in Chapter 5.
(2) Alternatively, although increasing the external voltage or current may not result
in electric or dielectric breakdown of a composite, it can lead to very large
enhancements of the nonlinearities as the volume fraction of the conducting
component approaches pc . We already described this phenomenon in Sections
3.1 and 3.3 in terms of the crossover from a linearly conductive material to a
weakly nonlinear one, and our goal in this section is to do the same for optical
properties of the same type of composite solids.
Following our discussions in Section 3.3, we consider in this section weak non-
linearities so that the field-dependent conductivity g(E) can be written as a power
series in the applied electric field E, with the leading term, i.e., the linear con-
ductivity g () , being much larger than the higher-order terms, a situation which is
typical of various nonlinearities in the optical and infrared spectral ranges of in-
terest to us. As discussed in Section 3.3, despite this weakness, such nonlinearities
lead to qualitatively new phenomena, such as enhancement of higher harmonics in
percolation composites, and the occurrence of bistable behavior of the composite
(Bergman et al., 1994; Levy et al., 1995) in which the conductivity switches be-
tween two stable values. In such disordered materials, especially those that contain
metal particles that are characterized by a dielectric constant with negative real
and small imaginary parts, the fluctuations in the local field are strongly enhanced
in the optical and infrared spectral ranges, leading to enhancement of various
nonlinear properties. If the disorder in the morphology of such solid materials is
of percolation-type, then they are potentially of great practical importance (see,
for example, Flytzanis, 1992) as composites with intensity-dependent dielectric
functions and, in particular, as nonlinear filters and optical bistable elements. The
optical response of such nonlinear composites can be easily tuned by, for example,
controlling the volume fraction and morphology of their constitutes.
More generally, optical properties of fractal aggregates of metal particles have
been studied. These studies indicate that a fractal morphology results in very large
enhancement of various nonlinear responses of the aggregates within the spectral
range of their plasmon resonances. The typical size, a ∼ 10 nm, of the metal
particles in such fractal aggregates is much smaller than the wavelength λ > 300
nm in the optical and infrared spectral ranges. Since the average density of particles
in fractal aggregates is much smaller than in non-fractal materials, and approaches
zero with increasing size of the aggregates, it is possible to consider each particle in
the aggregate as an elementary dipole and introduce the corresponding interaction
operator. If this is done, then, solving the problem of the optical response of
metal fractal aggregates reduces to diagonalizing the interaction operator for the
106 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

light-induced dipoles. If the size of the fractal aggregate is not very large, the
diagonalization can be done numerically (and efficiently) and thus the local electric
field can be calculated (see, for example, Stockman et al., 1995, 1996; Stockman,
1997; Shalaev et al., 1993; Markel et al., 1999). Computations of this type indicate
that large field fluctuations are localized in some small parts of the fractal aggregate
and change with the wavelength. These predictions and numerical computations
of large enhancements of optical nonlinearities in metal fractals have also been
verified experimentally for degenerate four-wave mixing and nonlinear refraction
and absorption. In these experiments, aggregation of silver particles (which were
initially isolated) into fractal clusters led to six orders of magnitude enhancement
of the efficiency of the nonlinear four-wave process and about three orders of
magnitude enhancement in the nonlinear refraction and absorption. The localized
and strongly fluctuating local fields in these fractal aggregates were imaged by
means of the near-field scanning optical microscopy (Shalaev et al., 1993; Markel
et al., 1999). A similar pattern was obtained for the field distribution in self-affine
thin films (Shalaev et al., 1996a,b; Safonov et al., 1998). As discussed in Chapter
1, such self-affine films possess a fractal surface with different scaling properties
in the plane of the film and normal to it.
Despite such progress, the distribution of the local field and the corresponding
nonlinearities were, until recently, poorly understood for metal-dielectric com-
posites with percolation-type disorder, especially in the most interesting spectral
range where the plasmon resonances occur in the metal grains. As shown in Sec-
tion 3.3, if a small volume fraction p  1 of a nonlinear material is embedded in
a linear host, the effective nonlinear response of the composite can be calculated
explicitly. As one may expect, the nonlinearities are enhanced at the frequency
ωr corresponding to the plasmon resonance of a single metal grain. Numerical
calculations (Stroud and Zhang, 1994; Zhang and Stroud, 1994) for a finite p also
indicate considerable enhancement in the narrow frequency range around ωr and,
moreover, the system sizes that can currently be used in the computations are not
large enough for drawing quantitative conclusions about the nonlinear properties
for frequencies ω  ωr . However, we should recognize that a small system size
L may act as an artificial damping factor that cuts off all the fluctuations in the
local field when the spatial separation is larger than L, hence resulting in a cor-
responding decrease of the nonlinearities which may otherwise not be seen in a
large enough sample.
An alternative method to numerical simulations is the effective-medium approx-
imation (EMA) that has the virtue of mathematical and conceptual simplicity. We
already described in Sections 3.1 and 3.3 such EMAs for nonlinear composites
near pc . As discussed there, for the static case the predictions of the nonlinear EMA
(Wan et al., 1996; Hui et al., 1997) are in good agreement with numerical simu-
lations for 2D percolation composite. However, despite this success, application
of any type of nonlinear EMA is suspect for the frequency range corresponding
to the plasmon resonances in metal grains. This is due to the fact that both com-
puter simulations and experimental data for the field distribution in percolation
composites indicate that the distribution contains sharp peaks that are separated
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 107

by distances that are much larger than the metal grain size. Thus, the local electric
field cannot be assumed to be the same in all the metal grains of the composite,
implying that the main pillar of the EMA, i.e., the assumption of a uniform field,
fails for the frequency range corresponding to the plasmon resonance in the films.
To address this problem, a new theory of the distribution of the electromagnetic
field and nonlinear optical processes in metal-dielectric composites was devel-
oped (Sarychev and Shalaev, 1999; Sarychev et al., 1999). The theory is based
on the concepts of percolation processes, and takes advantage of the fact that the
problem of optical excitations in percolation composites can be mapped onto the
Anderson localization problem. It predicts localization of surface plasmons (SP)
in composites with percolation disorder, and describes in detail the localization
pattern. It also indicates that the SP eigenstates are localized on length scales that
are much smaller than the wavelength of an incident light. The eigenstates with
eigenvalues that are close to zero (resonant modes) are excited most efficiently
by the external field. Since the eigenstates are localized and only a small portion
of them is excited by the incident beam, overlapping of the eigenstates can typi-
cally be neglected, a fact that significantly simplifies the theoretical analysis and
allows one to derive relatively simple expressions for enhancement of linear and
nonlinear optical responses.
The purpose of this section is to describe and summarize this progress. An excel-
lent comprehensive review of this subject was presented by Sarychev and Shalaev
(2000). This section is patterned closely after their review and represents a sum-
mary of their discussions. Since the languages of nonlinear currents/conductivities
and nonlinear polarizations/susceptibilities, or dielectric constants, are completely
equivalent, they will be used interchangeably in this section.

3.5.1 Scaling Properties of Moments of the Electric Field

As already demonstrated in Chapters 2, 5, and 6 of Volume I and earlier in the
present chapter, in metal-dielectric percolation composites the effective static (DC
or zero frequency) conductivity ge decreases with decreasing volume fraction p of
the metal component, and vanishes at p = pc . Since for p < pc the effective DC
conductivity ge = 0, the material is dielectric-like. Therefore, a metal-insulator
transition takes place at the percolation threshold pc . However, although the tran-
sition at pc is second-order, the pattern of the fluctuations in percolation composites
appears to be quite different from that for a second-order phase transition, the fluc-
tuations of which are usually characterized by long-range correlations, with their
relative magnitudes being of the order of unity. In contrast, for (DC) percolation
conductivity, the local electric fields are concentrated on the edges of large metal
clusters, so that the field maxima (large fluctuations or peaks) are separated by
distances that are of the order of the percolation correlation length ξp . Since ξp
diverges at pc (recall that near pc , ξp ∼ |p − pc |−ν ), the implication is that the
distance between the field maxima or peaks also increases as pc is approached.
To obtain insight into the high-frequency properties of metals, consider first a
simple model—the Drude model (already utilized in Chapters 4 and 6 of Volume
108 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

I)—that reproduces semi-quantitatively the basic optical properties of a metal.

According to this model, the dielectric constant m of metal grains is given by
(ωp /ω)2
m (ω) = b − , (177)
1 + iωτ /ω
where b is the contribution to m due to the inter-band transitions, ωp is the
plasma frequency, and ωτ = 1/τ  ωp is the relaxation rate (in Chapters 4 and 6
we took b = 1). In the high-frequency range considered here, losses in the metal
 + i  , then
grains are relatively small, ωτ  ω. Therefore, if we write, m = m m
  
|m |/m  ω/ωτ  1. Moreover, one has, m < 0 for the frequencies ω < ω̃p ,
where ω̃p is the renormalized plasma frequency which is given by
ωp
ω̃p = √ . (178)
b
Therefore, the metal conductivity, gm = −iωm /4π  (b ω̃p2 /4π ω)[i(1 − ω2/ω̃p2 )
+ ωτ /ω], is characterized by the dominant imaginary part for ω̃p > ω  ωτ , i.e.,
it is of inductive character. In this sense, the metal grains can be thought of as in-
ductances L, while the dielectric gaps between the metal grains can be represented
by capacitances C. Then, the percolation composite represents a set of randomly
distributed L and C elements. The collective surface plasmons, excited by the
external field, can be thought of as resonances in different L − C circuits, and the
excited surface plasmon eigenstates represent giant fluctuations of the local field.

3.5.1.1 Distribution of Electric Fields in Strongly Disordered Composites

Before embarking on discussing the properties of the distribution of local electric
field in a composite, let us recall from Chapters 5 and 6 of Volume I how the dielec-
tric constant of a disordered material is computed via a discrete, percolation-type
model. Suppose that a percolation composite is illuminated by light and consider
the local optical field distributions in the material. A typical metal grain size a in
the composite is much smaller than λ, the wavelength of the light in the visible and
infrared spectral ranges. If so, then one can introduce a potential φ(r) for the local
electric field and write the local current density I as, I(r) = g(r)[−∇φ(r) + E0 ],
where E0 is the external field, and g(r) is the local conductivity at r. In the quasi-
static limit, computation of the field distribution reduces to finding the solution of
the Poisson’s equation since, due to current conservation, ∇ · I = 0, one has
∇ · {g(r)[−∇φ(r) + E0 ]} = 0, (179)
where the local conductivity g(r) = gm or gd for the metal and dielectric compo-
nents, respectively. We rewrite Eq. (179) in terms of the local dielectric constant,
(r) = 4π ig(r)/ω, so that
∇ · [(r)∇φ(r)] = E, (180)
where E = ∇ · [(r)E0 ]. The external field E0 can be real, while φ(r) is, in general,
a complex function since m is complex in the optical and infrared spectral ranges.
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 109

Since Eqs. (179) and (180) are difficult to solve analytically, one discretizes them in
order to solve them by numerical simulations. If, for example, a standard 5-point (in
2D) or 7-point (in 3D) finite-difference discretization is used, then, a discrete model
on a simple-cubic lattice is obtained in which the metal and dielectric particles
are represented by metal and dielectric bonds of the lattice. Thus, Eq. (180), in
discretized form, takes on the form of Kirchhoff’s equations defined on a lattice.
Assuming that the external electric field E0 is directed along the z-axis, one obtains



ij (φj − φi ) = ij Eij (181)
j j

where φi is the electric potential at site i of the lattice, and the sum is over the nearest
neighbors j of the site i. For the bonds ij in the ±z-direction, the electromotive
force Eij is given by, Eij = ±E0 a0 (where a0 is the spatial period of the lattice),
while Eij = 0 for the other bonds that are connected to site i. Thus, the composite
material is modeled by a resistor-capacitor-inductor network in which the bond
permittivities ij are statistically independent and a0 is equal to the metal grain size,
a0 = a. In the case of a two-component metal-dielectric random composite, the
permittivities ij take values m and d with probabilities p and 1 − p, respectively.
To make further progress, we use a simple-cubic lattice which has a very large but
finite number of sites N and rewrite Eq. (181) in a matrix form:

Hφ = E, (182)

where φ = {φ1 , φ2 , . . . , φN }, and the elements of the vector E are, Ei = j ij Eij .
Here H is a N × N matrix such that for i
= j , Hij = −ij = d > 0 and m =
(−1 + iκ)|m  | with probabilities p and 1 − p, respectively, and H =
ii j ij ,
where . . j refers to nearest neighbors of site i, and κ is the usual loss factor, κ=
 / .  .  1. The diagonal elements H are distributed between 2d and 2d ,
m m ii m d
where d is the dimensionality of the space.
Similar to the dielectric constant, we write H = H + iκH , where iκH rep-
resents losses in the system. The Hamiltonian H formally coincides with the
Hamiltonian of the problem of metal-insulator transition (Anderson transition) in
quantum systems, i.e., it maps the quantum-mechanical Hamiltonian for the Ander-
son transition problem with both on- and off-diagonal correlated disorder onto the
present problem. Hereafter, we refer to H as the Kirchhoff’s Hamiltonian (KH).
Thus, the problem of determining the solution of Kirchhoff’s equation, Eq. (181)
or (182), is equivalent to the eigenfunction problem for the KH, H n = n n ,
whereas the losses can be treated as perturbations.
Since m  < 0, and the permittivity  of the dielectric matrix is positive, the set
d
of the KH eigenvalues n contains eigenvalues with real parts that are equal (or
close) to zero. Then, eigenstates n that correspond to eigenvalues n such that,
|n |  |m |, and |d |, are strongly excited by the external field and are seen as
giant field fluctuations, representing the resonant surface plasmon modes. If one
assumes that the eigenstates excited by the external field are localized, then they
should look like the peaks of the local field with the average distance between
110 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

them being about a(N/n)1/d , where n is the number of the KH eigenstates excited
by the external field.
Consider now the special case when m  = − , which corresponds to the plas-
d
mon resonance of individual particles in a 2D system. Since a solution to Eq. (181)
does not change if m and d are multiplied by the same factor, we normalize the
system and set d = −m = 1. We also suppose, for simplicity, that the metal con-
centration is p = 0.5. In this case, the eigenstates n are all localized. On the other
hand, computer simulations (Müller et al., 1997) showed that there is a transition
from chaotic (Berry, 1977) to localized eigenstates for the 2D Anderson problem,
with a crossover region between the two. Consider first the case when the metal
volume fraction p = pc = 1/2 for the 2D bond percolation problem. Then, the
diagonal disorder in the KH is characterized by, Hii   = 0, and H2  = 4, which
ii
correspond to the chaos-localization transition (Müller et al., 1997). Moreover, H
also possesses strong off-diagonal disorder, Hij   = 0, which favors localization

(see, for example, Verges, 1998). There is therefore strong evidence that the eigen-
states n are localized for all n in the 2D system, although one cannot rule out
the possibility of inhomogeneous localization, similar to that obtained for frac-
tal clusters (see, for example, Stockman et al., 1994), or power-law localization
(Kaveh and Mott, 1981; Kramer and MacKinnon, 1993).
In the case of d = −m = 1 and p = 1/2, all parameters in H are of the
order of unity, and therefore its properties do not change under the transforma-
tion d ⇐⇒ m . Therefore, the real eigenvalues n are distributed symmetrically
around zero in an interval of the order of one. The eigenstates with eigenvalues
n  0 are effectively excited by the external field and represent the giant local
field fluctuations. When p decreases (increases), the eigenstates with eigenvalues
n  0 are shifted from the center of the distribution toward its lower (upper)
edge, which typically favors localization. Because of this, one may assume that in
2D the eigenstates, or at least those with eigenvalues n  0, are localized for all
metal volume fractions p.
The situation in 3D is much more complex. Despite the great effort and the
progress that has been made, the Anderson transition in 3D is not yet fully
understood. Computer simulations (Kawarabayashi et al., 1998) of Anderson lo-
calization in 3D [with d = −m = 1, p = 1/2, the diagonal matrix elements wii
distributed uniformly around 0, −w0 /2 ≤ wii ≤ w0 /2, and the off-diagonal ele-
ments wij = exp(iφij ), with phases φij also distributed uniformly in 0 ≤ φij ≤
2π ] show that in the center of the band the states are localized for the disorder
w0 > wc = 18.8. In the 3D H Hamiltonian discussed here, the diagonal elements
are distributed as −6 ≤ Hii ≤ 6, and therefore the diagonal disorder is smaller than
the critical disorder wc , but the off-diagonal disorder is stronger than in the calcu-
lations of Kawarabayashi et al. (1998). It has been shown (Verges, 1998; Elimes
et al., 1998) that even small off-diagonal disorder strongly enforces localization,
and thus one may conjecture that, in the 3D case, the eigenstates corresponding to
the eigenvalues n  0 are also localized for all p.
If we express the potential φ in Eq. (182) in terms of the eigenfunctions n
of H as, φ = n An n , and substitute it in Eq. (182), we obtain the following
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 111

equation for coefficients An :

. . 
(iκbn + n )An + iκ n .H . m Am = En , (183)
m
=n
. . 
where bn = n .H . n , and En = (n |E) is a projection of the external field
onto the eigenstate n . Since all the parameters in H are of the order of unity, the
bn are also of the order of unity and can be approximated by some constant b  1.
Sarychev and Shalaev (2000) suggested that the eigenstates n are localized within
spatial domains ξA (), where ξA () is the Anderson localization length. Then,
the sum in Eq. (183) is convergent and can be treated as a small perturbation. The
first two coefficients in the approximation are then given by
En
n =
A(0) , (184)
n + iκb
whereas

. . 
n = −iκ
A(1) n .H . m A(0)
m . (185)
m
=n

In Eq. (185), the most important eigenstates in the sum, in the limit κ → 0, are
those with eigenvalues |m | ≤ bκ. Since the eigenvalues n are distributed in
an interval of the order of unity, the spatial density of the eigenmodes with
|m | ≤ bκ vanishes as a −d κ → 0 as κ → 0, implying that A(1) n is exponen-
  0 1
tially small, |An | ∼ | m
=n n |H |m Em /bm | ∝ exp −[a/ξA (0)]κ −1/d ,
(1)

and can be neglected when κ  [a/ξA (0)]d . Then, the local potential φ is given
 (0) 
by, φ(r) = n An n = n En n (r)/(n + iκb), and the fluctuating part of
the local field Ef = −∇φ(r) is given by


Ef (r) = − En [∇n (r)/(n + iκb)] , (186)
n

where ∇ is understood as a lattice operator. The average field intensity is then

given by
, -
. .2

En E ∗ [∇n (r) · ∇ ∗ (r)]
|E| = .Ef + E0 . = |E0 | +
2 2 m m
, (187)
m n
(n + iκb)(m − iκb)

where we used the fact that Ef  = E∗f  = 0.

Consider now the eigenstates n with eigenvalues n within a small interval
|n − | ≤   κ centered at , which we denote them by n (r). Recall
that the eigenstates are assumed to be localized so that eigenfunctions n (r) are
well-separated in space, with the average distance l between them being, l() ∼
a[N ()]−1/d , where
ad

N () = δ( − n ), (188)
 n
112 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

is the dimensionless density of states for the Kirchhoff Hamiltonian (KH) H , and 
is the system’s volume. We assume that the metal volume fraction p  1/2, so that
all quantities in the KH H are about unity, and therefore the density of states N ()
is also about unity at  = 0. Hence, the distance l() can be arbitrary large as
 → 0, while it is still much smaller than the system size. It is further assumed
that the total number of eigenstates n (r) is large. When l()  ξA (), the
localized eigenfunctions n (r) are characterized by spatial positions of their
centers rn , so that n (r) = (r − rn ) and Eq. (187) becomes
 


m
∗ ∗
n En Em [∇(1 , r − rn ) · ∇ (2 , r − rm )]
|E|2 = |E0 |2 + ,
(1 + iκb)(2 − iκb)
1 2
(189)
where the first sums are over positions of the intervals |n − 1 | and |m − 2 |
in the  space, whereas the sums in the numerator are over spatial positions rn and
rm of the eigenfunctions. For each realization of a macroscopically-homogeneous
disordered film, the positions rn of the eigenfunctions (r − rn ) take on new
values that do not correlate with . Therefore, we can independently carry out the
averaging in the numerator in the second term of Eq. (189) over positions rm and
rn of eigenstates m and n . Since, ∇n (r) = 0, we obtain
∗
2 3
En Em ∇ (1 , r − rn ) · ∇ ∗ (2 , r − rm )
(190)
 |En |2 |∇ (1 , r − rn )|2 δ1 2 δnm ,

which, when substituted in Eq. (187), results in



n |En |2 |∇n (, r)|2
|E|2 = |E0 |2 + . (191)

2 + (bκ)2

The localized eigenstates are not in general degenerate, so that the eigenfunctions
n can be selected to be real, i.e., n = n∗ (where ∗ denotes the complex con-
. .2 4
jugate). Then, |En |2 = |(n |E)|2 = . N .
i=1 n,i Ei ∼ a
−2d |  Edr|2 , which,
n
after using (180) and (181), yields
. .2 . .2
4−2d .
. . . .
|En | ∼ a . 4−2d . .
. n (E0 · ∇)dr. = a . (E0 · ∇n )dr. .
2
(192)

Since the local dielectric constants || are of the order of unity, one can write,
∇n ∼ n /ξA (), and therefore,
. .2 .N .2
|E0 |2 a 4 .. . .
.
. ∼ |E0 | a
2 4
. .
|En | ∼
2
.  n (r) dr . . n,i . . (193)
a 2d ξA2 () ξA2 () . .
i=1

Using the fact that, n |n  = N
i=1 |n,i | = 1, and that n are localized within
2

ξA (), one obtains n,i ∼ [ξA ()/a] −d/2 in the localization domain which, when
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 113

substituted in Eq. (193), yields

|En |2 ∼ |E0 |2 a 2 [ξA )/a]d−2 . (194)
One can estimate in a similar way,


N
. .
|∇n (, r)|2 ∼ ξA−2 () |n (, r)|2 ∼ ξA−2 ()N −1 .n,i .2 ∼ ξ −2 (),
A
i=1
(195)
where N = /a d is the total number of sites. Using these estimates and taking
into account the fact that the total number of the eigenstates within interval  is
equal to N N (), one finally obtains

N ()[a/ξA ()]4−d
|E|2 ∼ |E0 |2 + |E0 |2 d. (196)
2 + (bκ)2
Since all matrix elements in the Hamiltonian H are of the order of unity (in fact, the
off-diagonal elements are ±1), the density of states N () and localization length
ξA () vary significantly within an interval of the order of one, while the denom-
inator in Eq. (191) has an essential singularity at  = ±ibκ. Then, the second
moment of the local electric field, M2 ≡ M2,0 = |E|2 / |E0 |2 , is estimated as

1
M2 ∼ 1 + N (a/ξA )4−d d ∼ N (a/ξA )4−d κ −1  1, (197)
2 + (bκ)2
provided that κ  N (a/ξA )4−d [we set ξA ( = 0) ≡ ξA , N ( = 0) ≡ N and
b  1]. Thus, in this case, the field distribution is described as a set of the KH
eigenstates localized within ξA , with its peaks having the amplitudes

Em ∼ E0 κ −1 (a/ξA )2 , (198)
which are separated by the field correlation length ξe ,
ξe ∼ a(N κb)−1/d ∼ a(N κ)−1/d . (199)
All the assumptions that led us to Eqs. (197)–(199) hold when  ξA , which is
ξe
fulfilled in the limit κ → 0.
Hereafter by superscript  we mark the fields, while the spatial scales are given
for the special case −m  =  = 1 considered here (note that  should not be
d
confused with the complex conjugation denoted by ∗), while for ξA and N we
omit the  sign in order to avoid complex notations; it is implied that their values
are always taken at −m  =  = 1, even if the case of | / |  1 is considered.
d m d
The assumption that the localization length ξA is proportional to the eigenstate
size might not, in general, be true for the Anderson systems, although it has been
confirmed well by numerical calculations for 2D percolation composites. It was
also assumed that the metal volume fraction p  1/2, which corresponds to the
2D percolation threshold, and that the density of states N () is finite and about
unity for  = 0. The latter assumption is, however, violated for small values of p
when the distribution of the eigenvalues shifts to the positive side of , so that the
eigenstates with eigenvalues   0 are shifted to the lower edge of the distribution,
114 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

and the density of states N in Eq. (197) becomes a function of p. In the limit p → 0,
the number of states effectively excited by the external field is proportional to the
number of metal particles, and hence N (p) ∼ p. The same consideration holds
in the opposite limit, p → 1, and therefore N (p) ∼ 1 − p. When N decreases,
localization becomes stronger and one can write, ξA ( = 0, p → 0) ∼ ξA ( =
0, p → 1) ∼ a. When p → 0 or p → 1, the number of the field maxima decreases
while the peaks become progressively sharper. Equation (197) also indicates that
strong field fluctuations (M2 > 1) exist in a metal-dielectric composite with d =
−m  in a wide range of concentrations,

κ < p < 1 − κ, κ  1. (200)

Although the above local fields were estimated for the special case of d = −m  ,

all the above results, which are based on the assumption that the eigenstates of
the Kirchhoff Hamiltonian are localized, hold in a more general case, when the
 of the metal dielectric constant is negative and its absolute value is
real part m
of the order of d . The important case of |m |  d will be considered in the next
subsection.

3.5.1.2 Moments of the Electric Field

Consider now the moments of the local electric field of arbitrary order, defined as

1
Mn,m = |E (r)|n E m (r) dr, (201)
ωE0m |E0 |n
where, as above, E0 ≡ E (0) is the amplitude of the external field, and E(r)
is the local field at r. We denote, for simplicity, Mn,0 = Mn , and assume that
a volume-averaged quantity is equivalent to its ensemble-averaged value, i.e.,
Mn,m = |E|n E m /E0m |E0 |n .
The high-order moment M2k,m ∝ E k+m E ∗k  represents a nonlinear optical
process in which in one elementary act k + m photons are added and k photons
are annihilated (see, for example, Boyd, 1992). This is because the complex-
conjugated field in the general expression for the nonlinear polarization implies
photon annihilation, so that the corresponding frequency enters the nonlinear sus-
ceptibility with a minus sign. Enhancement of the Kerr optical nonlinearity GK is
proportional to M2,2 , the enhancement of the third-harmonic generation is given
by |M0,3 |2 , and surface-enhanced Raman scattering is represented by M4,0 (see
below).
An important case is when Mn,m  1, i.e., when the fluctuating part of the local
electric field Ef is much larger than E0 . Suppose, for simplicity, that E0 is real
and that |E0 | = 1. We can write, for the moment M2p,2q (p and q are integers),
the following equation
,   

En1 En2 ∇n1 · ∇n∗2 · · · En2p−1 En2p ∇n2p−1 · ∇n∗2p
M2p,2q =    
n1 + ibk n2 − ibk · · · n2p−1 + ibk n2p − ibk
n1 ,n2 ,···n2p ;m1 ,m2 ,···m2q
 -
Em1 Em2 ∇m1 · ∇m2 · · · Em2q−1 Em2q ∇m2q−1 · ∇m2q
×     ,
m1 + ibk m2 + ibk · · · m2q−1 + ibk m2q + ibk
(202)
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 115

where · denotes an ensemble average (which, as discussed above, is equivalent

to the volume-average), and the sums are over all eigenstates of the KH H . We
now average Eq. (202) over spatial positions of eigenstates n (r) ≡  (r − rn )
to obtain
 

| −|≤ |En |2p En2q ∇n · ∇n∗ p (∇n · ∇n )q
M2p,2q ∼ n
2 3p ,

2 + (bk)2 ( + ibk)2q
(203)
where the summation in the numerator is over eigenfunctions n =  (, r − rn )
with eigenvalues within the interval |n − | ≤   κ, while the external sum
is over positions  of the intervals that cover the entire range of eigenvalues n .
Following the same line of arguments that was used for deriving Eq. (197), one
can show that (Sarychev and Shalaev, 2000)

N () [a/ξA ()]4(p+q)−d
M2p,2q ∼ 2 3p d. (204)
2 + (bκ)2 ( + ibκ)2q

Assuming that the density of states N () and the localization length ξA () are
both smooth functions of  in the vicinity of zero, and taking into account the fact
that all parameters of the Hamiltonian H for the case d = −m  = 1 are of the

order of one, the following estimate for the moments of the local field is obtained

Mn,m ∼ N (p)[a/ξA (p)]2(n+m)−d κ −n−m+1 , (205)

for n + m > 1 and m > 0 (for simplicity we set b = 1). We remind the reader once
again that N (p) and ξA (p) should be understood as N (p) = N (p,  = 0) and
ξA (p) = ξA (p,  = 0), i.e., they are given at the eigenvalue  = 0.
The maximum of the Anderson localization length ξA () is typically at the
center of the distribution of the eigenvalues  (Kawarabayashi et al., 1998). When
p
= 1/2,  = 0 moves from the center of the -distribution toward its tails where
the localization is typically stronger (i.e., ξA is smaller). Therefore, it is plausible
that ξA (p) reaches its maximum at p = 1/2 and decreases toward p = 0 and
p = 1, so that the absolute values of the moments of the local field may have a
minimum at p = 1/2. In 2D composites the percolation threshold pc is typically
close to 0.5. Therefore, in such composites the moments Mn,m do have a local
minimum at pc as a function of the metal volume fraction p, and the amplitudes of
various nonlinear processes, while much enhanced, have a characteristic minimum
at pc . It is important to note that the magnitude of the moments in Eq. (205) do not
depend on the number of annihilated photons in one elementary act of the nonlinear
scattering. However, when all photons are added (i.e., when all frequencies enter
the nonlinear susceptibility with the plus sign) and n = 0, one cannot estimate the
moments M0,m by Eq. (205), since the integral in Eq. (204) is no longer determined
by the poles at  = ±ibκ. However, all the functions of the integrand are about
unity and M0,m ∼ O(1) for m > 1. The moment M0,m is an important quantity
since it yields the enhancement GnH G of the nth order harmonic generation through
the relation, GnH G = |M0,m |2 (see below).
116 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

3.5.1.3 Field Fluctuations at Frequencies Below the Resonance

So far we have assumed that |m |/d  1, which corresponds to the plasmon
resonance in the metal grains. To estimate the fluctuations in the local field in per-
colation composites for |m |/d  1, the renormalization approach developed by
Shalaev and Sarychev (1998), Sarychev and Shalaev (1999) and Sarychev et al.
(1999) can be utilized. Let us briefly recall the main concepts of the renormalization
method (see also Chapter 5 of Volume I). Consider a percolation composite with
the metal volume fraction p = pc . The system is divided into cubic cells of size b,
each of which is considered as a new renormalized element. The cells are classi-
fied into two types: Those that contain a continuous path of metallic particles are
considered as conducting, while those without such a sample-spanning cluster are
considered as non-conducting, or dielectric. The effective dielectric constant m (b)
of a conducting cell decreases with increasing its size b as, m (b)  (b/a)−µ/ν m ,
whereas the effective dielectric constant d (b) of a dielectric cell increases with b
as d (b)  (b/a)s/ν d , where µ, s and ν are the usual percolation critical expo-
nents for the conductivity, dielectric constant, and percolation correlation length,
respectively (see above and Chapters 2, 5 and 6 of Volume I). The cube size b is
now taken to be
b = br = a(|m |/d )ν/(µ+s) . (206)
Let us recall that the exponent s also characterizes the power law behavior of
the effective conductivity of a conductor-superconductor composite near the per-
colation threshold. Then, in the renormalized system the dielectric constant of
µ/(µ+s)
the new elements either takes a value, m (br ) = d |m |s/(µ+s) (m / |m |),
µ/(µ+s)
for the renormalized conducting cell, or d (br ) = d |m |s/(µ+s) , for the
renormalized dielectric cell. The ratio of the dielectric constants of these new
elements is then, m (br )/d (br ) = m /|m |  −1 + iκ, where the loss-factor
κ = m  /| |  1 is the same as in the original system. As discussed in Chap-
m
ter 5 of Volume I, at p = pc , the volume fraction of conducting and dielectric
elements does not change under a renormalization transformation. Since the field
distribution in a two-component system depends on the ratio of the dielectric
permittivities of the components, after the renormalization the problem becomes
equivalent to what was discussed above for the case d = −m  = 1. Taking into

account the fact that the electric field renormalizes as E0 = E0 (br /a), one obtains


from Eq. (198) the following expression for the field’s peaks in the renormalized
system:
$ % $ %
|m | ν/(µ+s) |m |
Em  E0 (a/ξA )2 (br /a)κ −1  E0 (a/ξA )2 
, (207)
d m
where ξA = ξA (pc ) is the localization length in the renormalized system. Each
maximum of the field in the renormalized system is in a dielectric gap in a dielectric
cube of linear size br or in between two conducting cells of the size br that are
not necessarily connected to each other. There is not a characteristic length in the
original system which is smaller than br , except the grain size a. Therefore, it is
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 117

plausible that the width of a peak of the local field in the original system is about
a. Then, values of the field maxima Em do not change when returning from the
renormalized system to the original one. Hence, Eq. (207) yields values of the field
maxima in the original system.
Equation (207) provides the estimate for the local field extrema when the real part
 of the dielectric constant is negative. For metals  increases in absolute value
m m
with the wavelength, when the frequency ω < ω̃p . Therefore, the field maxima
Em (ω) increase strongly with the wavelength. For a Drude metal the steep growth
of the peaks Em (ω) occurs for the frequencies ω < ω̃p , when the dielectric constant
m can be approximated as
b  b ωτ
m (ω < ω̃p )  2(ω − ω̃p ) +i , (208)
ω̃p ω̃p
which, when substituted in Eq. (207), yields
 . . (ν+µ+s)/(µ+s)
. .
2 2b ω − ω̃p ω̃p
Em (ω < ω̃p )  E0 (a/ξA ) ν/(µ+s)
.
ω̃p ωτ b d
(209)
Since in a typical metal, ωτ  ω̃p , the amplitudes of the field’s peak first increase
steeply and then saturate (see below) at Em  E0 (a/ξA )2 (b /d )ν/(µ+s) (ω̃p /ωτ ) ∼
E0 ω̃p /ωτ , when ω  0.5ω̃p . Therefore, the intensity maxima Im exceed the
2
intensity of the incident wave I0 by a factor Im /I0 ∼ ω̃p /ωτ  1.
Consider now the case ω  ωp , when for a Drude metal
$ %2 
ω ωτ 
m (ω  ωp )  − 1−i , ω  ωτ (210)
ωp ω
which, when substituted in Eq. (207), yields
$ %2 $ % $ %
a ωp 2ν/(µ+s) ω
Em (ω  ωp )  E0 √ . (211)
ξA d ω ωτ
For 2D percolation, the critical exponents are, µ = s  ν = 4/3, √and thus
√
Eq. (211) yields, Em ∼ E0 (a/ξA )2 ωp /( d ωτ ) = E0 (a/ξA )2 (ω̃p /ωτ ) b /d ∼
E0 (ω̃p /ω), which coincides with the estimate obtained from Eq. (209) for ω =
0.5ω̃p , implying that the local field’s peaks increase steeply when m  , the real

part of m , is negative and then remains essentially constant in the wide frequency
range, ω̃p < ω < ωτ .
For 3D percolation composites, we roughly have, ν  (µ + s)/3, and thus
2/3
Eq. (211) yields, Em ∼ E0 (b /d )1/3 ω̃p ω1/3 /ωτ , implying that the local field
peaks increase up to Em /E0 ∼ ω̃p /ωτ when m  < 0, and then decrease as

Em /E0 ∼ (ω̃p /ωτ )(ω/ω̃p ) , if the frequency decreases further.

1/3

To obtain Mn,m we consider first the spatial distribution of the maxima of the
field for |m |  d . The average distance between the maxima in the renormalized
system is ξe , given by Eq. (199). Then, the average distance ξe between the maxima
118 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

in the original system (provided that N ∼ 1) is

$ % $ %
|m | ν/(µ+s) |m | 1/d
ξe  (br /a)ξe ∼ a 
, (212)
d m
which in 2D (with µ = s  ν = 4/3) reduces to a simple form,
|m |
ξe ∼ a ! . (213)
d m

In the renormalized system a typical ”area” of a peak of the field corresponds to

ξAd , implying that in the original system each maximum is stretched over (ξA /a)d
clusters of the size br . In each of these clusters the field maximum splits into n(br )
peaks of amplitude Em , distributed along a dielectric gap in the dielectric square
of size br . Since the gap area scales as the capacitance of the dielectric square, one
has
n(br ) ∝ (br /a)d−2+s/ν , (214)
and therefore
$ %n+m
Em n(br )
Mn,m ∼ (ξA /a) d
E0 (ξe /a)d
$ %[(n+m−2)ν+s]/(µ+s) $ %n+m−1
|m | |m |
∼ N (ξA /a) d−2(n+m)

, (215)
d m
for n + m > 1 and n > 0. Since |m |  d and |m | /m   1, M
n,m  1 in the
visible and infrared spectral ranges. We emphasize that the localization length ξA
in Eq. (215) corresponds to the renormalized system with d = −m  = 1. The

localization length in the original system, i.e., a typical size of the eigenfunction,
is about (br /a)ξA  a, i.e., the eigenstates become macroscopically large when
|m | /d  1, and consist of sharp peaks separated in space by distances much
larger than a.
It is then not difficult to show, using Eq. (215), that
$  % $ %
m n(br ) m |m | (m−2+s/ν)ν/(µ+s)
M0,m ∼ M0,m (br /a)

∼ , (216)
(ξe /a)d |m | d
which holds when M0,m > 1. In 2D, if we use µ = s  ν = 4/3, Eqs. (215) and
(216) are simplified to
 n+m−1
|m |3/2
Mn,m ∼ N ! , (217)

(ξA /a)2 d m
for n + m > 1 and n > 0, and
 | |(m−3)/2
m m
M0,m ∼ (m−1)/2
, (218)
d
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 119

 . The moments M
for m > 1, n = 0 and (m |/d )(m−1)/2 > |m |/m n,m (n
= 0) are
strongly enhanced in 2D Drude metal-dielectric composites since they reach the
maximum value
 n+m−1
ωp
Mn,m ∼ N √ , (219)
ωτ d (ξA /a)2
when ω  ωp . Thus, in a 2D percolation composite the moments Mn,m are
independent of frequency if ω  ωp . For metals this typically takes place in
the red and infrared spectral ranges. For example, for a semi-continuous sil-
ver film on a glass substrate, the moments Mn,m can be estimated as, Mn,m ∼
[3 × 102 (a/ξA )2 ]n+m−1 , for ω  ωp .
It follows from Eq. (215) that for 3D metal-dielectric percolation composites,
for which the dielectric constant of the metal component can be estimated by the
Drude formula, the moments Mn,m (n
= 0) achieve their maximum at frequency
ωmax  0.5ω̃p . Since, as mentioned above, for 3D percolation, ν/(µ + s)  1/3,
the maximum value of Mn,m is roughly given by
 n+m−1
Mn,m (ω = ωmax ) ∼ N (ξA /a) (a/ξA )2 (b /d )1/3 ω̃p /ωτ , (220)

whereas for ω  ωp ,
 2/3
n+m−1
 (a/ξA )2 ωp ω1/3
Mn,m ω  ωp ∼ N (ξA /a) 1/3
. (221)
 d ωτ
Figure 3.10 compares the results of numerical and theoretical calculations for

Figure 3.10. Moments Mn,m of the electric field in semicontinuous silver films versus the
wavelength λ at the percolation threshold. On the left are the moments Mn = Mn,0 , from
the bottom to the top, for n = 2, 3, 4, 5 and 6. The solid curves are the predictions of the
scaling theory, Eq. (215), while the symbols are the numerical simulation data. Shown on
the right are the moments M4,0 (upper solid curve predicted by the scaling theory, versus ∗,
the numerical data), M0,4 (upper dashed curve), M2,0 (lower solid curve predicted by the
scaling theory, versus +, the numerical data), and M0,2 (lower dashed curve predicted by
the scaling theory, versus circles, the numerical data) (after Sarychev and Shalaev, 2000).
120 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

Mn,m in a 2D semi-continuous silver film on glass, indicating excellent agree-

ment between the scaling theory and numerical simulations, where ξA  2a
was used. The small value of ξA indicates that, at least in 2D, there is strong
localization of surface plasmons in percolation composites. Note that, as dis-
cussed above, nonlinear optical processes are, in general, phase dependent with
their phase dependence being through the term E m and their enhancement being
Mn,m = |E/E0 |n (E/E0 )m .According to the above analysis, Mn,m ∼ Mn+m,0 ≡
Mn+m , for n ≥ 1. Thus, for example, enhancement of the Kerr-type nonlinearity,
IK = M2,2 , is proportional to the enhancement of the Raman scattering, IRS  M4 .
So far, it has been assumed that, when analyzing the case of m   0, the metal

volume fraction p equals pc . We now consider the range p = p − pc , where the

above estimates for Mn,m are valid. First, note that the above expressions for the
local field and the average moments Mn,m of the field hold for almost all values of p
given by Eq. (200) when m  −d . The metal volume fraction range p shrinks,
however, where the local electric field is strongly enhanced and m   0. The above

analysis was based on the finite-size scaling analysis (see Chapter 2 of Volume I for
description of the finite-size scaling method), which holds provided that lr < ξp ,
where ξp is the percolation correlation length. Since at pc the correlation length
ξp diverges, these estimates are valid in the wide frequency range ωτ < ω < ω̃p ,
which includes the visible, infrared, and far-infrared spectral ranges for typical
metals. For any particular frequency from this interval, one can estimate the range
p, where the giant field fluctuations occur, by requiring that, lr = ξp , which
results in, |p| ≤ (d / |m |)1/(µ+s) . Therefore, the local electric field fluctuates
strongly for such volume fractions and its moments Mn,m are much enhanced.

3.5.1.4 Computer Simulation

Anumber of EMAs, as well as position-space renormalization group (PSRG) meth-
ods, of the type described in Section 5.11 of Volume I, have been proposed for
calculation of optical properties of semi-continuous disordered films. However,
none of these methods allows one to calculate the field fluctuations and the effects
resulting from them. Because semi-continuous metal films are of great theoretical
and practical interest, it is important to study statistical properties of the electro-
magnetic fields in their near zone. To simplify the theoretical considerations, one
may assume that the electric field is homogeneous in the direction perpendicu-
lar to the √
film plane, implying that the skin depth δ for the metal grains is large,
δ  c/(ω |m |)  a, where a is the grain size, so that the quasi-static approxima-
tion holds (see also Chapter 4 of Volume I). Note that the role of the skin effect can
be very important, resulting, in many cases, in strong alterations of the electromag-
netic response found in the quasi-static approximation (see, for example, Sarychev
et al., 1995; Levy-Nathansohn and Bergman, 1997). At the same time, the quasi-
static approximation simplifies significantly theoretical considerations of the field
fluctuations and describes well the optical properties of semi-continuous films, pro-
viding qualitative, and in some cases, quantitative, agreement with experimental
data.
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 121

In the discussion that follows, the skin effect is neglected so that a semi-
continuous film can be considered as a 2D material. In the optical frequency range,
where the frequency ω is much larger than the relaxation rate τ −1 of the metallic
component, a semi-continuous metal film can be thought of as a 2D L − R − C
lattice (see, for example, Brouers et al., 1993). As before, the capacitance C rep-
resents the gaps between metal grains that are filled by the dielectric material
(substrate) with a dielectric constant d . The inductive elements L − R represent
the metallic grains that for the Drude metal have the dielectric function m (ω)
given by Eq. (177). In the high-frequency range considered here, the losses in the
metal grains are small, ω  ωτ . Therefore, m     (in modulus) and   < 0
√ m m
for frequencies ω < ω̃p = ωp / b . Thus, the metal conductivity is almost purely
imaginary and the metal grains can be modeled as L-R elements, with the active
component being much smaller than the reactive one. If the skin effect cannot be
neglected, i.e., if the skin depth δ < a, the simple quasi-static presentation of a
semi-continuous film as a 2D array of the L − R and C elements is not valid. One
can still use the L − R − C model in the other limiting case, when the skin effect
is very strong (δ  a). In this case, the losses in the metal grains are small, regard-
less of value of ω/ωτ , whereas the effective inductance for a metal grain depends
on the grain size and shape rather than on the material constants for the metal.
It is instructive to consider first the properties of the film at p = pc , where
the duality relation (see above and also Chapters 4 and 5 of Volume I) predicts
that, the effective dielectric constant e in the quasi-static case is given exactly by,
√
e = d m . If we neglect the metal losses and set ωτ = 0, the metal dielectric
constant m < 0 for ω < ω̃p . We also neglect possible small losses in a dielectric
substrate, assuming that d is real and positive, in which case e is purely imaginary
for ω < ω̃p . Therefore, a film consisting of loss-free metal and dielectric grains
is absorptive for ω < ω̃p . The effective absorption in a loss-free film means that
the electromagnetic energy is stored in the system and thus the local fields could
increase without limit. In reality, due to losses the local fields in a metal film are,
of course, finite. However, if the losses are small, one may expect very strong
fluctuations in the field. To calculate Rayleigh and Raman scattering, and various
nonlinear effects in a semi-continuous metal film, one must know the field and
current distributions in the film.
Although, as discussed in Chapters 4 and 5 of Volume I, there are several very
efficient numerical methods for calculating the effective conductivity of composite
materials, they typically do not allow calculations of the field distributions. Brouers
et al. (1997) developed a PSRG method, a generalization of what was described
in Chapter 5 of Volume I, using a square lattice of the L − R (metal) and C
(dielectric) bonds. A fraction p of the bonds were metallic (L − R bonds) and had
a conductivity gm = −im ω/4π, while the dielectric (C) bonds, with a fraction
1 − p, had a conductivity gd = −id ω/4π. The applied field E0 was E0 = 1,
whereas the local fields inside the system were of course complex quantities. In
this method, after each RG transformation, an external field E0 is applied to the
system and the Kirchhoff’s equations are solved in order to determine the fields
and the currents in all the bonds of the transformed lattice. The self-dual PSRG cell
122 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

F F

F F
F
F F

F F F

F

F F

Figure 3.11. 2 × 2 renormalization group cells in two and three dimensions.

of Figure 3.11 was used which, because of its hierarchical structure, allows these
equations to be solved exactly. Then, the one-to-one correspondence between the
elementary bonds of the transformed lattice and the bonds of the initial square
lattice was used for determining the field distributions, as well as the effective
conductivity, of the initial lattice. The number of operations for obtaining the full
distributions of the local fields is proportional to b2 [to be compared with O(b7 )
operations needed in the transform-matrix method and O(b3 ) operations needed
in the Lobb-Frank algorithm that was described in Sections 5.14.2 and 5.14.3 of
Volume I]. The Drude formula for metal dielectric functions was used, and thin
films of silver (for which b = 5, the plasma frequency ωp = 9.1 eV, and the
relaxation frequency ωτ = 0.021 eV) and gold (for which b = 6.5, ωp = 9.3 eV,
and ωτ = 0.03 eV), deposited on a glass substrate with the dielectric constant
d = 2.2, were modeled.
All the numerical results obtained with this method were in agreement with the
predictions of the scaling theory discussed above, as well as with experimental
data, described below.

3.5.1.5 Comparison with the Experimental Data

Optical properties of metal-insulator thin films have been intensively studied, both
experimentally and by computer simulations. Semi-continuous thin metal films are
usually produced by thermal evaporation or sputtering of metals onto an insulating
substrate. At first, small metallic grains are formed on the substrate. As the film
grows, the metal volume fraction increases and irregularly-shaped clusters are
formed on the substrate, resulting in 2D fractal morphologies. The size of these
structures diverges at pc where a percolating cluster of metal is formed, and a
continuous conducting path appears between two opposite ends of the sample.
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 123

The metal-insulator transition is very close to this point, even in the presence of
quantum tunneling. At higher surface coverage, the film is mostly metallic, with
voids of irregular shapes. As coverage increases further, the film becomes uniform.
Optical properties of such metal-dielectric films exhibit anomalous phenomena that
are absent for bulk metal and dielectric components. For example, the anomalous
absorption in the near-infrared spectral range leads to unusual behavior of the
transmittance and reflectance in that, the transmittance is much higher than that of
continuous metal films, whereas the reflectance is much lower.
The predictions of the PSRG computations have been compared with the experi-
mental data for gold-on-glass films at various wavelengths (Sarychev and Shalaev,
2000). There is good qualitative agreement between the two. The data for such dis-
ordered metal-dielectric films near pc suggest localization of optical excitations
in small nm-scale hot spots. The hot spots of a percolation film represent very
large local fields (fluctuations); spatial positions of the spots strongly depend on
the light frequency. Near-field spectra observed and calculated at various points of
the surface consist of several spectral resonances, the spectral locations of which
depend on the probed site of the sample. These features are observable only in
the near zone. In the far zone, one observes images and spectra in which the hot
spots and the spectral resonances are averaged out. The local field enhancement is
large, which is especially important for nonlinear processes of the nth order, and
are proportional to the enhanced local fields to the nth power. This opens up a fas-
cinating possibility for nonlinear near-field spectroscopy of single nano-particles
and molecules.

3.5.2 Anomalous Light Scattering from Semicontinuous Metal

Films
A quantitative analysis of the spatial distribution of the local field fluctuations, and
light scattering induced by such fluctuations, are now carried out. The resonance
frequency ωr , corresponding to the condition m  (ω ) = − is considered first
r d
which, for a Drude metal, is fulfilled at the frequency
# $ %2
1 ωτ ωp
ωr = ωp − √ , (222)
b +  d ωp b +  d
where it has been assumed that ωτ = 1/τ  ωp , which is the case for a typical
metal. Then, the metal dielectric function is, m (ωr ) = d (−1 + iκ), where the loss
factor κ is given by, κ  (1 + b /d )ωτ /ωr  1. In modeling the distribution of
the local field fluctuations, we take advantage of the fact that, since this distribution
does not change when bond conductances are multiplied by the same factor, it is
convenient to consider a lattice in which a bond conductance is gm = −1 + iκ
with probability p (the L bonds) and gd = 1 with probability 1 − p (the C bonds).
Since the absolute values of gd and gm are very close, the standard method based
on the percolation theory and scaling analysis cannot be used for estimating the
spatial distribution of the field. One may, however, use the PSRG method described
124 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

above to carry out the analysis, which yields interesting results. For example, using
a system of size b = 1024, p = pc = 0.5, and ω = ωr , Sarychev and Shalaev
(2000) calculated the electric field in all the bonds for 10−4 ≤ κ ≤ 10−1 with the
external field being E0 = 1. The distribution of the field intensity, I (r) = |E(r)|2 ,
was found to be close to the well-known log-normal distribution, with its values
spread over many orders of magnitude, even for κ = 10−1 . For κ = 10−4 , I (r)
was distributed essentially uniformly in (0, 104 ). The average intensity, I  =
|E0 |2 M2 , increased as, I  ∝ κ −1 , in agreement with Eq. (205). Thus, the field
fluctuations lead to enhanced light scattering from the film.
It should be pointed out that the fluctuations considered here, and the corre-
sponding light scattering, do not arise because of the fractal morphology of the
metal clusters, but are due to the distribution of local resonances in a disordered
metal-dielectric film, which is homogeneous on a macroscopic scale. The local
intensity of the electric field is strongly correlated in space, and the distribution is
dominated by the field correlation length ξe introduced by Eqs. (199) and (212),
and defined as the length scale over which the field fluctuations are small. As the
L− (metallic) component becomes loss-free (κ → 0), ξe diverges according to

ξe ∼ κ −νe , (223)

where νe is a new critical exponent which has been estimated by several numerical
methods. For example, in 2D the PSRG method described above yields νe 
0.45 ± 0.05, while the scaling theory, Eq. (208), predicts that νe = 1/d, where d
is the space dimension, a result that was also conjectured by Hesselbo (1994). For
small loses at resonance, the correlation length ξe is the only relevant length scale
of the system at pc since |m |/d  1.

3.5.2.1 Rayleigh Scattering

We consider now Rayleigh scattering induced by the giant field fluctuations
(Brouers et al., 1998) discussed above. Suppose that a semi-continuous film is
illuminated by a wave normal to the film plane. The space between the metal
grains is filled by a dielectric material. Therefore, the film can be considered as a
2D array of metal and dielectric grains that are distributed over the film’s plane.
The incident electromagnetic wave excites the surface current I in the film. Con-
sider the electromagnetic field induced by these currents at some distant point R.
The origin of the coordinates is fixed at some point in the film. Then, the vector
potential A(R) of the scattered field defined by, H(R) = ∇ × A(R) [where H(R)
is the magnetic field], arising from the surface current I(r), is such that if
I(r) exp(ik |R − r|)
A(R, r) dr = dr, (224)
c |R − r|
4
where k = ω/c is a wavevector, then A(R) = A(R, r) dr, where the integration
is over the entire film. In experiments, the dimensions of the film are small enough
that r  R, and therefore, ik |R − r|  ikR − ik(n · r), where n is the unit vector
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 125

in the direction of R. Thus,


ik
H(R)  exp(ikR) [n × I(r)] exp[−ik(n · r)] dr, (225)
cR
and the electric field is given by

i −ik
E(R) = [∇ × H(R)]  exp(ikR) {n × [n × I(r)]} exp[−ik(n · r)] dr.
k cR
(226)
It follows from Eqs. (225) and (226) that H(R) is perpendicular to E(R), and that
|E(R)| = |H(R)|, implying that the scattered field can be considered locally as a
plane wave when the distance from the film is large. The total intensity It of the
light scattered in the direction n = R/|R| is given by
 c 1 c 2
It (n) = R2 Re{[E(R) × H∗ (R)]} = R E(R) · E∗ (R)
4π 2 8π

c k2
= [n × I(r1 )] · [n × I∗ (r2 )] exp[ikn · (r1 − r2 )] dr1 dr2 , (227)
8π c2
where the angular brackets indicate an ensemble averaging. The semi-continuous
metal films that are considered here are much larger than any characteristic intrinsic
spatial scale, such as the field correlation length ξe , and therefore the ensemble
average can be included in the integrations over the film area in Eq. (227) without
changing the result. It is assumed, for simplicity, that the incident light is natural
(unpolarized), and that its direction is perpendicular to the film plane. Then, the
averaging [n × j(r1 )] · [n × j∗ (r2 )] should be carried out over the polarizations
of the incident wave, yielding, I(r1 ) · I∗ (r2 )[1 − sin2 (θ/2)], where θ is the angle
between n and the normal to the film plane.
If we replace in Eq. (227) the local currents I(r) by their average values I(r),
we obtain the specular scattering Is . The scattering I (θ) = It − Is in all other
directions is then obtained as
$ %  
c k2 1 2
I (θ ) = 1− sin θ I(r1 ) · I∗ (r2 ) − |I|2 exp[ikn · (r1 − r2 )] dr1 dr2 .
8π c2 2
(228)
The natural correlation length for the local field fluctuations, and therefore for the
current-current correlations, is ξe . If ξe  λ, where λ = 2π/k is the wavelength
of the incident light, Eq. (228) is simplified by replacing the exponential by unity,
hence yielding
$ %   
c k2 1 2 I(r1 ) · I∗ (r2 )
I (θ) = 1 − sin θ |I| 2
− 1 dr1 dr2 . (229)
8π c2 2 |I|2
Note that for macroscopically-homogeneous and isotropic films the current-current
correlations I(r1 ) · I∗ (r2 ) depend only on r = |r2 − r1 |. We now introduce the
correlation function
I(r1 ) · I∗ (r2 ) ReI(0) · I∗ (r)
C(r) = − 1 = − 1. (230)
|I|2 |I|2
126 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

in terms of which the intensity of the scattered light is given by

$ %  ∞
c k2 1 2
I (θ ) = A 1 − sin θ |I| 2π 2
C(r)r dr, (231)
8π c2 2 0
where A is the film area. I (θ ) should be compared with the integral intensity
(power) of the incident light, I0 = A(c/8π ) |E0 |2 , where |E0 | is the amplitude
of the incident wave. For the normal incident light, E = T E0 , where T is the
transmittance of the film. For semi-continuous metallic films at p = pc one has
|T |2  0.25 in a wide spectral range from the visible to the far infrared spec-
tral range (Yagil et al., 1992). One also has, I = age E = age T E0 , where
ge = −ie ω/(4π ) is the effective conductivity, and thickness of the film has been
approximated by the size a of a metal grain.
Substituting I = ge aT E0 in Eq. (231), the ratio, I˜(θ ) = I (θ )/I0 is obtained,
$ %  ∞
˜ (ka)4 1 2 2 1
I (θ ) = 1 − sin θ |T e | 2 C(r)r dr, (232)
8π 2 a 0
which is independent of the film’s geometry. It follows from Eq. (232) that the
portion of the incident light that is not reflected, transmitted or adsorbed, but is
scattered from the film is given by
  ∞
(ka)4 1
Itot = 2π I˜(θ ) sin θ dθ = |T e |2 2 C(r)r dr. (233)
3 a 0
The behavior of C(r) depends on the frequency, and also on the behavior of |T e |2 ,
which achieves large values, |T e |2  1, in the infrared spectral range.
We can compare Eq. (233) with the scattering for the case when the metal grains
interact with the electromagnetic field independently. The cross section σR of
Rayleigh scattering from a single metal grain is estimated as, σR = (8π/3)(ka)4 a 2
for |m |  1. The portion of the light which would be scattered if the grains were
R  p(8/3)(ka)4 . Assuming p = 1/2, the following
independent is given by Stot
estimate is obtained for the enhancement Ig = Itot /Stot R of the scattering due to

the field fluctuations,


|T e |2 ∞
Ig ∼ C(r)r dr. (234)
4a 2 0
If the integral in (234) is determined by the largest distances where field correlations
are most important, i.e., where, r ∼ ξe , the scattering can even diverge if losses
vanish and ξe → ∞. This is certainly the case for 2D metal-dielectric films. The
above formalism holds if Itot  1. Otherwise, it is necessary to take into account
the feedback effects, i.e., the interaction of the scattered light with the film.

3.5.2.2 Scaling Properties of the Correlation Function

Using the PSRG approach described above, Brouers et al. (1998) calculated
the correlation function C(r) for a 1024 × 1024 L − C system for gold semi-
continuous metal films at p = pc = 1/2 and for the resonance frequency ωr , so
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 127

 (ω ) = − , and for several values of κ =   /|  |. For each κ, the results

that m r d m m
were averaged over 100 different realizations of the system. Their calculations
indicated that for a < r < ξe , the correlation function decays as (the distance r is
measured in units of the metal grain size a)
C  (r) ∼ M2 (r/a)−(1+η) ∼ κ −1 (r/a)−(1+η) , (235)
where M2 [see Eq. (197)] is the second moment of the local field in the system
 =
with m −d , and η = 0.8 ± 0.1 is a new critical exponent that determines the
spatial correlation of the local electric field. If we substitute Eq. (235) in (232)
and (233), we find that the integrals diverge at the upper limit, implying that
the scattering is determined by values of the correlation function C(r) at large
distances, i.e., at r ∼ ξe . This means that the field fluctuations with spatial distances
of the field correlation length ξe  1 are responsible for the anomalous scattering
from semi-continuous films.
We now consider the dependence of scattering on the frequency of an incident
electromagnetic wave. We first consider frequencies just below ω̃p where the
metal dielectric function can be estimated for a Drude model as, m = m  + i  
m
 
2b (ω − ω̃p )/ω̃p + ib ωτ /ω̃p , i.e., m < 0. For such frequencies, |m |/d ≤ 1,
while the loss factor κ  ωτ /2(ω̃p − ω) decreases rapidly with frequency ω, and in
particular decreases below the renormalized plasma frequency ω̃p . For |m  |/ 
d
1, the correlation function C(r) is estimated by Eq. (235) which, when substituted
in Eq. (233) and integrated up to ξe ∼ aκ −1/d , yields
$ %1+(1−η)/d $ %
(ka)4 ω̃p ω 2+(1−η)/d
Itot ∼ |T |2 d b 1− , ω < ω̃p ,
3 ωτ ω̃p
(236)
√
where the exact result, e (p = pc ) = d m , which is a result of the duality rela-
tion for 2D percolation systems (see above and also Chapters 4 and 5 of Volume
I), was used.
Consider now the limit ω  ω̃p , assuming again that ω  ωτ , for which the
dielectric constant for a Drude metal is approximated as, m  (ωp /ω)2 (−1 +
iωτ /ω), yielding |m  |/  (ω /ω)2 /  1 and κ =   /|  |  ω/ω  1. To
d p d m m τ
estimate the correlation function C(r), the system is divided into squares of size b
and the procedure described above is followed by taking b = br , where br is given
by Eq. (206). Then, the correlation function C  in the renormalized system has the
same form as Eq. (235), while in the original system, C(r)  (br /a)1+η C  (r) for
r  br , and C(r) ∝ r −µ/ν for r  br . By matching these asymptotic expressions
at r = br , the following ansatz emerges,

M2 (br /r)µ/ν ∼ κ −1 (br /r)µ/ν , a  r < br ,
C(r) ∼ (237)
M2 (br /r)1+η ∼ κ −1 (br /r)1+η , br < r < ξe ,
where ξe is given by Eq. (212). Equation (237) allows one to estimate the sec-
ond moment of the local electric current, Mj ≡ |I(r)|2 , at pc . From Eq. (230),
one can write, Mj = |E0 |2 |ge |2 C(0) = (ω/4π )2 |E0 |2 |e |2 C(0). At pc one has
e ∼ d (m /d )s/(s+µ) (see Chapter 5 of Volume I). The correlation function C(r)
128 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

for r ∼ a is given by, C(0) ∼ C(a) ∼ M2 (lr /a)µ/ν ∼ M2 (|m |/d )µ/(s+µ) , and
hence
$ %
|m | (2s+µ)/(s+µ)
Mj ∼ (ω/4π )2 |E0 |2 d2 M2 ∼ (ω/4π )2 |E0 |2 d |m | M2 ,
d
(238)
where M2 ≡ M2,0  and M ≡ M , as defined earlier. Equation (238) holds for
2 2,0
arbitrary spatial dimension.
We now consider light scattering from semi-continuous metal films for ω 
ωp , where the metal dielectric constant for a Drude metal is approximated as,
m  −(ωp /ω)2 (1 − iωτ /ω). By substituting Eq. (237) into (233) and taking into
account the fact that at pc , |e |2  d |m |  d (ωp /ω)2 (using µ = s  ν = 4/3),
the following result is obtained
(ka)4
Itot ∼ |T |2 |e |2 κ −1 br 1+η ξe 1−η
3
(ka)4  ω a 4 $ ω %1+(1−η)/2
2 −1−(1−η)/2 p
∼ |T | κ |m | ∼ 0.1
2
, (239)
3 c ωτ
where the experimental result, |T |2  0.25, which holds for p = pc and ωτ 
ω  ωp , was used. Thus, the scattering first increases as ω1+(1−η)/2 with increas-
ing ω according to Eq. (239) and then vanishes as (ω̃p − ω)2+(1−η)/2 as ω → ω̃p
[see Eq. (236)].
The enhancement of the scattering due to the field fluctuations can be estimated
from Eqs. (234) and (237) as, Ig ∼ |T |2 d |m |br2 κ −1−(1−η)/2 /4, which yields, for
a Drude metal and ω  ωp , the following equation
$ % $ %
|T |2 ω̃p 4 ω 1+(1−η)/2
Ig ∼ . (240)
4 ω ωτ
Using typical values, |T |2 = 1/4 and d = 2.2, the enhancement Ig can become as
large as 5 × 104 at wavelength λ = 1.5 µm and continues to increase towards the
far infrared spectral range. Note that Rayleigh scattering decreases as ω4 with de-
creasing frequency, whereas the anomalous scattering varies as, I ∼ ω1+(1−η)/2 
ω1.1 , and therefore the enhancement increases as Ig ∼ ω−2.9 ∼ λ2.9 in the infrared
part of the spectrum.

3.5.3 Surface-Enhanced Raman Scattering

We now consider surface-enhanced Raman scattering (SERS), one of the most in-
triguing optical effects discovered over the past 20 years (Moskovitz, 1985; Markel
et al., 1996; Kneipp et al., 1997; Nie and Emory, 1997), and describe a theory of
Raman scattering (see also Chapter 6 of Volume I) enhanced by strong fluctua-
tions of the local fields (Brouers et al., 1997). In rough thin films this phenomenon
is commonly associated with excitation of surface plasmon oscillations which
are typically considered in two limiting cases: (1) Oscillations in non-interacting
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 129

roughness features of various shapes, and (2) surface plasmon waves (polaritons)
that laterally propagate along the metal surface. In practice, there are strong light-
induced interactions between different features of a rough surface, and therefore
plasmon oscillations should be treated as collective surface excitations (localized
surface plasmons) that depend strongly on the surface morphology.

3.5.3.1 General Formulation

The formulation of the problem and the solution that are discussed here are due to
Brouers et al. (1997), as described by Sarychev and Shalaev (2000). We consider
optical properties of a semi-continuous metal film consisting of metal grains, ran-
domly distributed on a dielectric substrate. The space between the metal grains
are usually filled by dielectric material of the substrate. As before, the local con-
ductivity g(r) of the film takes on either the metallic value, g(r) = gm , in the
metal grains, or the dielectric value, g(r) = −iωd /4π, outside the metal grains,
where ω is the frequency of the external field. We assume that the wavelength λ is
much larger than the grain size a, the linear size of the space between the grains,
percolation correlation length ξp , and the local field correlation length. Hence, the
local field E(r) is given by Eq. (179).
It is instructive to assume first that the external field E0 (r) is step-like, E0 (r) =
E1 δ(r − r1 ), where δ(r) is the Dirac delta-function. The current density at an
arbitrary point r2 is then given by
I(r1 , r2 ) = (r2 , r1 )E1 , (241)
where (r2 , r1 ) is the non-local conductivity matrix representing the system’s
response at point r2 to a source at the point r1 , such that if an external field E0 (r)
is applied to the system, the local current at the point r2 will be given by

I(r2 ) = (r2 , r1 )E0 (r1 ) dr1 , (242)

where the integration is over the total area of the system.

In view of our discussion in Chapters 5 and 6 of Volume I, it should be clear
that  can be expressed in terms of the Green function G of Eq. (179):
∇ · {g(r2 )[∇G(r2 , r1 )]} = δ(r2 − r1 ), (243)
where a differentiation with respect to the coordinate r2 is assumed. Comparing
Eqs. (179) and (243), the following equation for the element of  is obtained
∂ 2 G(r2 , r1 )
αβ (r2 , r1 ) = g(r2 )g(r1 ) , (244)
∂r2,α ∂r1,β
where the Greek indices denote x and y. It is clear that, because of the symmetry
of the Green function,
αβ (r1 , r2 ) = βα (r2 , r1 ). (245)
Since we assumed that the wavelength of the incident electromagnetic wave is
much larger than all spatial scales in a semi-continuous metal film, the external
130 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

field E0 is constant in the film plane. The local field E(r2 ), induced by the external
field E0 , is obtained by using Eq. (144) for the non-local conductivity ,

1
E(r2 ) = (r2 , r1 )E0 dr1 , (246)
g(r2 )
and excite Raman-active molecules that are (assumed to be) uniformly distributed
in the composite. Such molecules, in turn, generate the Stokes fields, Es (r2 ) =
αs (r2 )E(r2 ), oscillating at the shifted frequency ωs , where αs (r2 ) is the ratio of
the Raman and linear polarizabilities of the Raman-active molecules at r2 . The
Stokes fields Es (r2 ) induce in the composite the currents Is (r3 ) that are given by

Is (r3 ) = (r3 , r2 )Es (r2 ) dr2 . (247)

Since the frequency ωs is typically close to the external field’s frequency, i.e.,
|ω − ωs |/ω  1, the non-local conductivities  appearing in Eqs. (246) and (247)
are essentially the same.
The intensity I of the electromagnetic wave scattered from any inhomogeneous
material is proportional to the current fluctuations inside the system:
,. .2 -
. .
I ∝ .. [I(r) − I] dr.. , (248)

where the integration is over the entire system, and · denotes an ensemble average.
For Raman scattering, · also includes averaging over the fluctuating phases of
the incoherent Stokes fields generated by Raman-active molecules. Therefore, the
average current densities oscillating at ωs is zero, Is  = 0, and hence the intensity
IR of Raman scattering from a semi-continuous metal film is given by
, . .2 -
. .
IR ∝ . I(r) dr..
.


∗ ∗ ∗
= αβ (r3 , r2 )αs (r2 )Eβ (r2 ) αγ (r5 , r4 )αs (r4 )Eγ (r4 ) dr2 dr3 dr4 dr5
(249)
where a summation over repeating Greek indices is implied, and the integration
is over the entire film plane. Equation (249) is now averaged over the fluctuating
phases of the Raman polarizabilities αs . Because the Raman field sources are
incoherent, we have αs (r2 )αs∗ (r4 ) = |αs |2 δ(r2 − r4 ), and therefore

∗ ∗
IR ∝ αβ (r3 , r2 ) µγ (r5 , r2 )δαµ |αs | Eβ (r2 )Eγ (r2 ) dr2 dr3 dr5 . (250)
2

If we now take advantage of the facts that, (1) a semi-continuous film is macroscop-
ically homogeneous, and thus its Raman scattering is independent of the orientation
of the external field E0 ; (2) due to (1), Eq. (250) can be averaged over the orien-
tations of E0 without changing the result, and (3) the non-local conductivity is
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 131

independent of the field orientations and is symmetric, we obtain


|αs |2
IR  ∝ |g(r2 )|2 |E(r2 )|2 0 |E(r2 )|2 0 dr2 , (251)
|E0 |2
where ·0 denotes the orientation averaging. It is not difficult to show that, for
macroscopically-isotropic materials, Eq. (251) can be rewritten as

|αs |2
IR  ∝ |g(r2 )|2 |E(r2 )|4 dr2 . (252)
|E0 |2
In the absence of any metal grains on the film, the local fields would not fluctuate
and one would obtain

IR0 ∝ |gd |2 |αs |2 |E0 |2 dr2 . (253)

Therefore, the enhancement IRS = IR /IR0 of Raman scattering due to presence of

metal grains on a dielectric substrate is given by
|g(r)|2 |E(r)|4  |(r)|2 |E(r)|4 
IRS = = . (254)
|gd |2 |E0 |4 d2 |E0 |4
Note that the derivation of Eq. (254) is essentially independent of the dimension-
ality and morphology of the material. Therefore, the enhancement IRS should hold
for any heterogeneous material, provided that the field fluctuations take place in-
side of it. In particular, Eq. (254) yields the enhancement for Raman scattering
from a rough metallic surface, provided that the wavelength is much larger than
the roughness spatial scales. It can also be used for calculating the enhancements
in a 3D percolation composite. The present theory indicates also that the main
source for the Raman scattering is the currents excited by Raman molecules in
metal grains, hence explaining why a large IRS is obtained even for relatively flat
metal surfaces (Moskovitz, 1985).

3.5.3.2 Raman and Hyper-Raman Scattering in Metal–Dielectric Composites

Since, as discussed above, the local electric field in materials with percolation
disorder is distributed mainly in the dielectric space between the metal clusters,
the SERS enhancement IRS may be estimated as, IRS ∼ M4,0 = |E(r)/E0 |4 .
Hence, in view of Eq. (215), we obtain
$ %(2ν+s)/(µ+s) $ %
d−8 |m | |m | 3
IRS ∼ N (p)[ξA (p)/a] 
, (255)
d m
indicating that, when the states are delocalized, ξA → ∞, IRS vanishes very
rapidly. Equation (255) can now be used for investigating the frequency and
volume fraction dependence of Raman scattering. For 2D metal-dielectric com-
posites with the critical exponents, µ = s  ν = 4/3, the Drude metal dielectric
function can be used for frequencies ω  ωp , and therefore Eq. (255) predicts
132 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

3/2
that, IRS ∼ N (p)[a/ξA (p)]6 (ωp /ωτ )3 /d , independent of the frequency. For
example, for silver-on-glass percolation films at pc , the Anderson localization
length ξA is about ξA  2a, the density of state, N (pc )  1, and therefore, IRS ∼
106 . For 3D composites at ω  ωp , IRS decreases with decreasing ω as IRS ∼
N (p)(ξA /a)−5 ωp2 ω/ωτ3 ∼ 106 ω/ωp , where the 3D critical exponents have been
approximated as, ν  s  (µ + s)/3, and the data, ωp = 9.1 eV and ωτ = 0.021
eV, for silver dielectric constant have been utilized.
Consider now hyper-Raman scattering when n photons of frequency ω are con-
verted to one hyper-Stokes photon of the frequency ωhRS = nω − ωsf , where
ωsf is the Stokes frequency shift corresponding to the frequency of molecule os-
cillations (electronic or vibrational). Thus, following the same line of reasoning
outlined above, the surface enhancement of hyper-Raman scattering (SEHRS)
IhRS is given by
|ghRS (r)|2 |EhRS (r)|2 |E(r)|2n  |hRS (r)|2 |EhRS (r)|2 |E(r)|2n 
IhRS = . .2 = . .2 ,
|gd |2 .E0,hRS . |E0 |2n |d |2 .E0,hRS . |E0 |2n
(256)
where EhRS (r) is the local field excited in the system by the uniform probe
field E0,hRS , oscillating with ωhRS , and ghRS (r) and hRS (r) are the local con-
ductivity and dielectric constant at frequency ωhRS . For n = 1 Eq. (256) describes
the conventional SERS. To estimate IhRS , we must keep in mind that the spatial
scales br for the field maxima at the fundamental frequency ω and the hyper-Stokes
frequency ωhRS are significantly different. Therefore, the average in Eq. (256) can
be decoupled and approximated as, |hRS (r)|2 |EhRS (r)|2 |E(r)|2n  ∼ |hRS (r)
EhRS (r)|2 |E(r)|2n = |hRS (r)EhRS (r)|2 M2n |E0 |2n , where M2n (ω) is the
2nth moment. It follows from Eq. (238) that, |hRS (r)EhRS (r)|2 ∼ d |m (ωhRS )
|M2 |E0,hRS |2 , where M2 (ωhRS ) is the second moment of the field EhRS (r). Using
the expressions for M2 and M2n given above, and taking into account the fact
that for p  pc the density of states N is about unity, one obtains the following
equation for enhancement of hyper-Raman scattering,
$ % $ %
|m (ωhRS )| (µ+2s)/(µ+s) |m (ωhRS )|
IhRS ∼ (ξA /a)2d−4(1+n)  (ω
d m hRS )
$ %[2ν(n−1)+s]/(µ+s) $ %2n−1
|m (ω) | |m (ω)|
×  (ω)
, (257)
d m
with n ≥ 2. For a Drude metal and frequencies ω  ω̃p , ωhRS  ω̃p the metal
dielectric constant can be approximated as, |m (ωhRS )| ∼ |m (ω)| ∼ (ωp /ω)2 and
 (ω)/| (ω)| ∼ ω /ω, and therefore Eq. (257) becomes
m m τ
 2[2ν(n−1)+3s+µ]/(µ+s) $ ω %2n
2d−4(1+n) ωp
IhRS ∼ (ξA /a) , (258)
ω ωτ
which, in 2D (using µ = s  ν = 4/3), simplifies to
 2(n+1) $ ω %2n  2 $ ω %2n
4n ωp 4n ωp p
IhRS ∼ (a/ξA ) ∼ (a/ξA ) . (259)
ω ωτ ω ωτ
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 133

Figure 3.12. Comparison of experimental data (points with error bars) for normalized
SERS, I¯ = IRS (p)/IRS (p = pc ), for a semicontinuous silver film, versus the theoretical
computations (curve) (after Sarychev and Shalaev, 2000).

3.5.3.3 Comparison with the Experimental Data

As discussed earlier, the localization radius ξA of the eigenstates n with eigenval-
ues   0 decreases when one shifts from p = pc toward p = 0 or p = 1, because
the eigenvalue  = 0 shifts from the center of the -distribution to its tails, where
localization of the eigenstates is stronger. Therefore, according to Eq. (255), Ra-
man scattering must have a minimum at pc , as a result of which IRS (p) must have
two maxima, with one maximum below pc and a second one above pc . Figure 3.12
presents (Gadenne et al., 1998) experimental data for the dependence of SERS on
the metal volume fraction p, and compares them with the theoretical predictions.
It is clear that there is good qualitative agreement between the predictions and the
data. In particular, in agreement with the theory, there is a minimum near pc .

3.5.4 Enhancement of Optical Nonlinearities in Metal–Dielectric

Composites
The next subject we consider is enhancement in heterogeneous materials with
percolation-type disorder of various nonlinear optical processes, such as the Kerr
optical effect and generation of high harmonics.

3.5.4.1 Kerr Optical Nonlinearities

These are third-order optical nonlinearities that result in an additional term in the
electric displacement D given by
Di (ω) = ij kl (−ω, ω, ω, −ω)Ej Ek El∗ ,
(3) (3)
(260)
(3)
where ij kl (−ω, ω, ω, −ω) is the third-order nonlinear dielectric constant (see,
134 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

for example, Boyd, 1992), E is an electric field at frequency ω, and summation

over repeated indices is implied. The Kerr optical nonlinearity results in nonlinear
corrections, which are proportional to the light intensity, for the refractive index
and the absorption coefficient.
We consider disordered materials that are macroscopically homogeneous
and isotropic. For such materials, the third-order term in the average electric
displacement is given by

D(3) (r) = α|E0 |2 E0 + β|E0 |2 E∗0 , (261)

where |E0 | is the amplitude of the external electric field at frequency ω, and α and
β are some constants. Note that, for an isotropic film, the second term in Eq. (261)
results in change of the polarization of the incident light. Moreover, for the case
of linear and circular polarization of the incident light, Eq. (261) can be simplified
since for linear polarization the complex vector E0 reduces to a real vector. Then,
|E0 |2 E0 = E02 E0 , and Eq. (261) becomes

D(3) (r) = e(3) |E0 |2 E0 , (262)

(3)
where the effective nonlinear dielectric constant e is now a scalar quantity. Let
us consider, for the sake of simplicity, the linearly polarized incident wave. We
write Eq. (262) in terms of the nonlinear average current I(3) (r) and the effective
(3) (3)
Kerr conductivity ge = −iωe /4π:

I(3) (r) = ge(3) |E0 |2 E0 . (263)

(3)
We consider first the limit in which the nonlinearities in metal grains gm and
(3) (3) (3)
dielectric gd are approximately equal, gm  gd , which can be caused by, for
example, molecules that are uniformly covering a semi-continuous film. Then
. .2
I(r) = g () (r)E (r) + g (3) .E (r). E (r), (264)

where E (r) is the local fluctuating field. Then, current conservation law takes the
following form
  
g (3)  ..  ..2
∇ · g (r) −∇φ(r) + E0 + () E (r) E (r)
()
= 0, (265)
g (r)

where −∇φ(r) + E0 = E (r) is the local field. The second and third terms of
Eq. (265) can be thought of as a renormalized external field

g (3)  ..  ..2
Ee (r) = E0 + Ef (r) = E0 + E (r) E (r) , (266)
g () (r)
where the field Ef (r) may change over the film but its average, Ef (r), is collinear
to E0 , in which case the average current density I(r) is also collinear to E0 and
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 135

can be written as

E0 E0 1
I = (E0 · I) = 2 E0 · I(r) dr, (267)
E02 E0 A
where A is the total area of the film, the integration is over the film area, and
E02 ≡ (E0 · E0 ). Expressing I(r) in terms of the non-local conductivity matrix
defined by Eq. (241) yields

E0 1
I = 2 [E0 (r, r1 )Ee (r1 )] dr dr1 . (268)
E0 A
If we integrate Eq. (268) over the coordinates r and use the symmetry of the
non-local conductivity matrix , we obtain

E0 1
I = 2 [I0 r · Ee (r)] dr, (269)
E0 A
where I0 (r) is the current induced at r by the constant external field E0 . Using
Eq. (266) and carrying out the integration, Eq. (269) becomes
⎡ 2 3. .2 ⎤
g (3) E(r) · E (r) .E (r).
I = E0 ⎣ge() + ⎦, (270)
E02
()
where ge and E(r) are the effective conductivity and local fluctuating field in
the linear approximation (i.e., for g (3) ≡ 0). Comparison of Eqs. (270) and (263)
yields an expression for the effective Kerr conductivity:
2 3. .2
g (3) E(r) · E (r) .E (r).
ge(3) = . (271)
E02 |E0 |2
Equation (271) is general and applicable to weak as well as strong nonlinearities.
In the former case, E (r)  E(r), and Eq. (271) becomes
g (3) E 2 (r)|E(r)|2
ge(3) = , (272)
E02 |E0 |2
(3)
yielding ge in terms of the linear local field. Note that Eq. (272) is the analogue
(3)
of (128). In the absence of metal grains, ge = g (3) . Therefore, the enhancement
IK of the Kerr nonlinearity is given by
E 2 (r) |E(r)|2
IK = = M2,2 , (273)
E02 |E0 |2
where M2,2 is the fourth moment of the local field.
(3) (3) (3)
Equations (272) and (273) were derived assuming that gm  gd . If gm
=
(3)
gd , the above analysis can be repeated in order to derive the following equation,
E 2 (r) |E(r)|2 (3) E 2 (r)|E(r)|2
ge(3) = pgm
(3) m
+ (1 − p)gd d
, (274)
E02 |E0 |2 E02 |E0 |2
136 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

where ·m and ·d represent averaging over the metal and dielectric grains, re-
spectively. Note that, for the case of cubic nonlinearity in the conductivity of
materials with percolation disorder, Eq. (274) was already derived and discussed
in Section 3.2 [see Eq. (128)]. For the case of Kerr conductivity, Eq. (274) was first
derived by Shalaev et al. (1998). According to Eq. (273), the Kerr enhancement
IK is proportional to the fourth power of the local field, averaged over the sample,
which is similar to the case of SERS with the enhancement factor IRS given by
Eq. (254). Note, however, that while IK is complex, IRS is a real and positive
quantity.
The enhancement of the Kerr nonlinearity can be estimated analytically using
the methods described above. Consider first the case when g (3) (r) in the dielectric
component is of the same order of magnitude or larger than in the metal component.
Then,
. . . .
. . . .
IK ∼ .ge(3) / g (3) (r) . = .e(3) /  (3) (r) . ∼ |M2,2 |
$ % $ %
|m | (2ν+s)/(µ+s) |m | 3
∼ N (ξA /a)d−8 
, (275)
d m
where Eq. (215) has been used for the moment M2,2 . For ω  ωp , the Kerr en-
hancement for 2D composites is estimated as, IK ∼ N (ξA /a)−6 (ωp /ωτ )3 , if the
Drude formula is used for the metal dielectric constant m . For example, as dis-
cussed above, for silver-on-glass semi-continuous films, Anderson localization
length ξA  2a and density of states, N  1, and therefore, IK ∼ 105 − 106 . As
discussed by Sarychev and Shalaev (2000), for d = 2 a plot of IK versus the metal
volume fraction p has a two-peak structure, which is similar to the case of Raman
scattering shown in Figure 3.12. However, in contrast to IRS , the dip at p = pc
is much more pronouced and is proportional to the loss factor κ, implying that at
p = pc the enhancement is actually given by, IK ∼ κM2,2 . This result is presum-
ably a consequence of the special symmetry of a 2D self-dual system at p = pc .
If one .moves. slightly away from p = pc , the enhancement IK increases such that,
IK ∼ .M2,2 . ∼ IRS ∼ M4,0 . The fact that the minimum at p = pc is much smaller
for SERS than for the Kerr process is presumably related to the latter being a phase
sensitive effect. Moreover, as already discussed above, the local field maxima are
concentrated in the dielectric gaps where |m |  d . Therefore, Eq. (275) is valid
when the Kerr nonlinearity is located mainly in such gaps.
Consider now the case when the Kerr nonlinearity is due to metal grains (see,
for example, Ma et al., 1998; Liao et al., 1998). Provided that m   − , the
d
local electric fields are equally distributed in the metal and dielectric components,
implying that the Kerr enhancement is still given by Eq. (274) with |m |/d = 1.
However, if m |  d , the local field will be concentrated in the dielectric space
between the conducting clusters with a value Em given by Eq. (207). The total
current Is of the electric displacement flowing in the dielectric space between
two resonate metal clusters of size br is given by, Is = aEm e brd−2 . Because of
the current continuity, the same current should flow in the adjacent metal clusters
where it is concentrated in a percolating channel. The electric field Emc in the metal
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 137

channel, which spans over the cluster, is given by, Emc ∼ Is /(m a d−1 ), where
a d−1 represents the cross-section of the channel. Then the nth moment of the local
electric field in a metal cluster of size br is, Emc
n  = E n La d−1 /bd , where L =
mc r
−d+2
a(m /e )br is the effective length of the conducting channel. Keeping in mind
that only a fraction κ = m  /| |  1 of the metal clusters of size b are excited
m r
by the external electric field, we obtain, Mnmet = |E|n met /E0n = κEmc n /E n , for
0
the moments of the electric field in the metal component,
$ % $ %
|m | n−1 |m | [(d−1)(n−2)ν−µ(n−1)]/(µ+s)
Mn ∼
met

, (276)
m d
where Eq. (206) was used for the size br of the resonant clusters. Then,
enhancement IKmet of the Kerr nonlinearity is given by
$ % $ %
|m | 3 |m | [2(d−1)ν−3µ]/(µ+s)
IKmet ∼ M4met ∼ 
. (277)
m d
 )3
In 2D (for which µ = s  ν = 4/3) Eq. (277) yields, IKmet ∼ M4met ∼ (|m | /m
(d / |m |) . As expected, IK  IK , and in fact for 2D systems near pc ,
1/2 met

$ %
IK |m | 2
∼ . (278)
IKmet d
Since in optic and infrared spectral ranges, |m |  d , the enhancement due to
the Kerr nonlinearity is much larger than when the initial nonlinearity is located
in the dielectric gaps where the local fields are much larger than in the metal. It
follows from Eq. (278) that the Kerr enhancement IKmet may become less than one,
implying that, on average, the local electric field in the metal component can be
smaller than the external field. For example, for semi-continuous silver films on a
glass substrate, IKmet < 1 for wavelengths λ > 10µm.

3.5.4.2 Enhancement of Nonlinear Scattering from Strongly Disordered Films

The next subject we consider is percolation-enhanced nonlinear scattering (PENS)
from a random metal-dielectric film at the metal volume fraction p near pc . Specif-
ically, we consider the enhanced nonlinear scattering which is due to local field
oscillation at frequency nω, while a percolation metal-dielectric film is exposed to
an electromagnetic wave of frequency ω. Since at pc a self-similar fractal metal
cluster forms and the metal-dielectric transition occurs in a semi-continuous metal
film, optical excitations of the self-similar cluster result in giant, scale-invariant,
field fluctuations. As before, we assume that a semi-continuous film is exposed to
the light that propagates normal to the film, with the wavelength λ larger than any
intrinsic length scale in the film. The space between the metal grains are filled by
the dielectric substrate so that a semi-continuous metal film can be thought of as a
2D array of metal and dielectric grains that are randomly distributed over a plane
for which we consider nth-order harmonic generation (nHG) for an incident wave
of frequency ω. The nHG is generated by the semi-continuous metal film that is
138 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

covered by a layer possessing a nonlinear conductivity g (n) . The layer can be made
of nonlinear organic molecules, semi-conductor quantum dots, or a quantum well
on top of a percolation film. The local electric field Eω (r), induced in the film by
the external field E0 , generates in the layer the nω current g (n) Eω Eωn−1 . Note that,
strictly speaking, this expression is valid only for the scalar nonlinear conductiv-
ity and odd n. However, for obtaining order-of-magnitude estimates, we can use
this formula for arbitrary n. The external field, oscillating at frequency ω, is still
denoted as E0 , though the frequency is indicated explicitly for other fields. The
nonlinear current g (n) Eω Eωn−1 , in turn, interacts with the film and generates the
initial nω electric field with an amplitude E(n) = g (n) Eωn−1 Eω /g () , where g () is
the linear conductivity of the nonlinear layer at frequency nω. The electric field
E(n) can be thought of as an inhomogeneous external field exciting the film at
frequency nω.
The nHG current I(n) induced in the film by the initial field E(n) can be de-
termined in terms of the non-local conductivity matrix (r, r ) introduced by
Eq. (241):

Iβ (r) =  βα (r, r )Eα(n) (r ) dr ,
(n) (n)
(279)
(n)
where  βα is the conductivity matrix at frequency nω, the integration is over the
entire film area, the Greek indices represent {x, y}, and summation over repeated
indices is implied. It is I(n) that eventually generates the nonlinear scattered field at
frequency nω. By using the standard approach of the scattering theory adopted to
semi-continuous metal films (Brouers et al., 1998), and assuming that the incident
light is unpolarized, the integral scattering in all directions but the specular one is
given by
 $ . . %
4k 2 . (n) .2
I= Iα (r1 )Iα (r2 ) − . I .
(n) (n)∗
dr1 dr2 , (280)
3c
where the integrations is over the entire area A of the film, k = ω/c, and ·
indicates an ensemble average. As in the case of Rayleigh scattering, we have
assumed that the integrand vanishes for r  λ, where r = r2 − r1 [therefore, the
term exp(ik · r) was omitted]. Using Eq. (279), we can write

Iα(n) (r1 ) Iα(n)∗ (r2 ) dr1 dr2

 '
4
Eβ (r3 )Eα∗(n) (r4 )
(n) (n)∗ (n)
= γβ (r1 , r3 ) δα (r2 , r4 )δγ δ dri , (281)
0
i=1
where ·0 denotes an average over the light polarization. We now introduce the
(0)
spatially uniform probe field Enω which oscillates at frequency nω and is assumed
(0) (0)∗ (0)
to be unpolarized. For the unpolarized light, δγ δ = 2 Enω,gamma Enω,δ /|Enω |2 ,
0
which, when substituted in Eq. (281), the integration is carried out over the coor-
(0)
dinates r1 and r2 , and the averaging over independent polarizations of fields Enω
and E0 are performed, the following equation for the current-current correlation
 3.5. Electromagnetic Field Fluctuations and Optical Nonlinearities 139

function is obtained,

Iα(n) (r1 ) Iα(n)∗ (r2 ) dr1 dr2 =

1 ∗
2 3 
(0)
()
gnω (r3 )gnω (r4 ) Enω (r3 ) · E∗nω (r4 ) E(n) (r3 ) · E(n)∗ (r4 ) dr3 dr4 ,
|Enω |2
(282)
(0)
where Enω (r) is the local nω field excited in the film by the probe field Enω , and
()
gnω (r) is the film linear conductivity at frequency nω.
In macroscopically-homogeneous and isotropic films considered here, the inte-
(0)
gral in Eq. (282) does not depend on direction of the probe field Enω . Therefore,
(0)
Enω can be selected to be collinear with the external field E0 . Moreover, I(n) is
(0)
parallel to the external field E0 . If the probe field Enω is aligned with E0 , we have,
.
. (n) .2 . (0) (n) .2 .
. I . = . Enω · I . /|E(0) nω | . Then, using Eq. (279), we can write
2

. . . .2
. (n) .2 1 . .
.I . = . E (0) (n)
(r , r )E (n)
(r ) dr dr .
2. . (283)
(0) 2 . nω,β βα 1 2 α 2 1
A|Enω |
If the integration over coordinate r1 is carried out, one obtains
.  ..2
. .2 . ()
. gnω Enω · E(n) .
. (n) .
.I . = (0)
, (284)
|Enω |2
and therefore
. .2 5 6 ∞
8π k 2 .. g (n) .. .
. () .2
.
I= . () . A .gEnω . |Eω | |Eω | 2 2(n−1)
C (n) (r)r dr, (285)
3c|Enω |2 . g .
(0)
0

where C (n) (r) is the nonlinear correlation function defined as

C (n) (r) =
.  ..2
() ∗ (r )[E (r ) · E∗ (r )][E(n) (r ) · E(n)∗ (r )] − . g () E(n) · E
gnω (r1 )gnω 2 nω 1 nω 2 1 2 . nω nω .
5. .2 . 6 ,
. () . . (n) ..2
.gnω Enω . E
(286)
which, for macroscopically-homogeneous and isotropic films, depends only on the
distance r = |r1 − r2 |.
Equation (285) should be compared with the nω signal Inω from the
nonlinear layer on a dielectric film with no metal grains on it, Snω =
. .2 .. (0) ..2 .. (0) ..2(n−1)
(c 2 /2π )A .g (n) /g () . .Eω . .Eω .
d . Therefore, the enhancement factor for
PENS, IP EN S = I /Inω , is given by
 ∞
(ka)4 |nω Enω |2 |Eω |2 |Eω |2(n−1) n2
IP EN S = . . C (n) (r)r dr. (287)
3 . (0) .2 a2
d2 .Enω . |E0 |2 |E0 |2(n−1) 0
140 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

Note that for a homogeneous (p = 0 or p = 1) surface, C (n) (r) = 0, and therefore,

IP ENS = 0, so that the scattering occurs only in the reflected direction. According
to Eq. (287), the enhancement IP EN S is proportional to |E|2(n+1)  which, for
highly fluctuating local fields, is very large. Since a metal-dielectric transition at
pc is similar to a second-order phase transition, one may anticipate that local field
fluctuations are rather large and have long-range correlations near pc . However,
what is surprising is that the field fluctuations in the optical spectral range discussed
here are quite different from those for a second-order phase transition. The reason
may be the following. The fluctuations in the local electric field that result in
PENS are of the resonant character and their variations can be over several orders
of magnitude. Therefore, the field correlation function C (3) (r) decreases very
rapidly for r > a, and has a negative minimum, regardless of the magnitude of
the local field correlation length ξe ; this anticorrelation occurs because the field
maxima have different signs. Moreover, the power-low decrease of C (3) (r), which
is typical for critical phenomena, occurs only in the tail and deviates from it for
r > ξe . The magnitude of ξe can be estimated from Eq. (213) as, ξe (λ)  5, 20 and
30 (in units of a, the grain size) for λ = 0.34, 0.53 and 0.9 µm, respectively. For a
typical size of a metal grain in a semi-continuous film, a  2 − 20 nm, the intrinsic
spatial scale of the local field inhomogeneity ξe  λ, as assumed
4∞ in advance. Based
on such considerations, the dimensionless integral a −2 0 C (n) (r)r dr should be
of the order of one for all n. Thus, one may anticipate that, in contrast to harmonic
generation from conventional metal surfaces, PENS is characterized by a broad-
angle distribution, with the total (in all directions) scattering being much larger
than the coherent scattering in the reflected direction.
To obtain√ a more accurate estimate of PENS, we note that the typical size
br (ω) ∼ a |m (ω)| of the local field maxima increases with decreasing ω, and
thus for a Drude metal, br (ω) ∝ ω−1 if ω  ωp . Since the spatial scales for Enω
and Eω are different, the average [|Enω |2 |Eω |2 |Eω |2(n−1) ]2  in Eq. (287) 2 can
be decoupled and approximated roughly as, |nω Enω |2 |Eω |2 |Eω |2(n−1) ∼
(0)
|nω Enω |2 |Eω |2n ∼ |nω d |M2,nω M2n |Enω |2 |E0 |2n , where M2,nω and M2n are
the spatial moments of the local fields Enω and Eω , respectively. Using this de-
coupling in Eq. (287) and taking into account the fact that, as discussed above, the
integral there is of the order of unity, Eq. (287) simplifies to
. .
IP EN S . m (nω) .
B. . . M2,nω M2n , (288)
(ka)4 d .
where B is an adjustable pre-factor. Finally, using Eq. (217) for the moments M2,nω
and M2n , and assuming that the localization length ξA ∼ a, and that the density of
states N ∼ 1, one obtains

IP EN S |m (nω)|5/2 |m (ω)|3(n−1)/2

 B , (289)
(ka)4  (nω)   (ω)2n−1
dn+1 m m

where it was assumed that the generated frequency nω is less than ωp , so that
 3.6. Electromagnetic Properties of Solid Composites 141

 (nω) < 0; otherwise, I (2n)

P EN S  B(ka) Mω , since the local nω fields are
m 4
 (nω) > 0. For the Drude metal and nω  ω , Eq. (289) is
not enhanced for m p
simplified to
$ %2n  
4 1 ωp ωp 2
IP EN S ∼ B (ka) n+1 . (290)
d ωτ ω
Equation (290) states that PENS increases with increasing the order of a nonlinear
process and decreases toward the infrared part of the spectrum as IP EN S ∝ λ−2 ,
in contrast to the well-known λ−4 law for Rayleigh scattering. Moreover, it is
interesting to note that, for high-harmonic scattering, PENS is proportional λ−2 ,
independently of the order n of optical nonlinearity.

3.5.4.3 Comparison with the Experimental Data

The diffusive scattering of the second harmonic from metal-dielectric films has
been observed in experiments with C60 -coated semi-continuous silver films (Akt-
sipetrov et al., 1993) and from thin but continuous silver films (Kuang and Simon,
1995). One may argue that the diffusive scattering of 2ω field is due to the anoma-
lous fluctuations of local electric fields on the rough features of the surface with
the spatial scale a being much smaller than wavelength λ of the incident light. If
so, then the scattering data reported by Kuang and Simon (1995) are similar to
PENS from percolation films.
To summarize, large field fluctuations in random metal-dielectric composites
near pc result in a new physical phenomenon: Percolation-enhanced nonlin-
ear scattering which is characterized by giant enhancement and a broad-angle
distribution.

3.6 Electromagnetic Properties of Solid Composites

In the preceding discussions, the skin effects in the metal grains was neglected. We
now consider electromagnetic properties of metal-dielectric materials, character-
ized by percolation disorder and irradiated by a high-frequency electromagnetic
field under the conditions that the skin effect in the metal grains is strong. The goal
of this section is to show that electromagnetic properties of random composites
can be understood in terms of the effective dielectric constant and magnetic per-
meability, provided that the wavelength of an incident wave is much larger than the
intrinsic spatial scale of the system. The wavelength inside a metal component can
be very small. The most interesting effects are expected in the limit of the strong
skin effect. Thus, one must go beyond the quasi-static approximation employed in
the analyses presented above.
Propagation of electromagnetic waves in percolation composites with wave-
length λ < ξp , where ξp is the correlation length of percolation, may be
accompanied by strong scattering. On the other hand, wave propagation for λ  ξp
142 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

can be described by Maxwell’s equations with effective dielectric constant e and

effective magnetic permeability Ke . In order to calculate these effective param-
eters, the approach suggested by Panina et al. (1990), as developed by Sarychev
and Shalaev (2000), is described.
√ We restrict our attention to the optically-thin
systems of size L  λ/ |e Ke |, which are still macroscopically homogeneous,
so that L  ξp . We already described in Section 4.13 of Volume I the theoretical
treatment of this problem for linear materials, and what follows is the extension
of that discussion to nonlinear composites.

3.6.1 Effective-Medium Approximation

Suppose that a percolation composite is placed inside a resonator, where electro-
magnetic standing waves are excited. The change in the field when a composite
is placed inside of the resonator is determined by superposition of the fields scat-
tered from individual metal and dielectric particles that have dielectric constants
m and d , respectively. The interaction between the particles can be taken into
account by an effective-medium approximation (EMA). As discussed in the pre-
vious sections, and in Volume I, in this method, the interaction of a given metal or
dielectric particle with the rest of the system is determined by replacing the latter
by a homogeneous medium with the effective parameters e and Ke . Assuming
that the composite grains are spherical, the electric fields Ein,m and Eout,m , excited
by the external electric field E0 , are calculated inside and outside of a metal grain
of size a, yielding the following equations (see also Chapter 4 of Volume I) for the
electric field inside the metal grain:
Ein,m (r) = Ein,m0 + 4πL(r), (291)
where
3e
Ein,m0 = E0 , (292)
2e + ˜m
and ˜m is the renormalized dielectric constant of the metal defined as
2F (ym a) 1 cot(x)
˜m = m , F (x) = 2 − , (293)
1 − F (ym a) x x
√
with k = ω/c, ym = k m Km , a being the radius of a metal grain. The skin
(penetration) depth δ is given by, δ = 1/Im(ym ). When the metal conductivity gm
is a real√quantity (i.e., in the microwave and radio frequency range), the skin depth,
δ = c/ 2π Km gm ω. In Cartesian coordinate system with the z-axis directed along
the field E0 , the local electric field L in Eq. (291) is determined by
1 ik
∇ × L(r) = ∇ × Ein,m (r) = BE , (294)
4π 4π
where
akm e sin(ym r)F (ym r) 7y x 8
BE = −3iE0 ,− ,0 (295)
(2e + ˜m ) sin(ym a) [F (ym a) − 1] r r
is a rotational magnetic induction generated in a metal particle by the electric
 3.6. Electromagnetic Properties of Solid Composites 143

current. Therefore, the inside electric field consists of uniform curl-free part Ein,m0
(i.e., ∇ × Ein,m0 ) and the rotational part L(r) that depends on the coordinate. The
field outside the metal particle is given by
$ %
e − ˜m E0 · r
Eout,m = E0 + a 3 ∇ . (296)
2e + ˜m r3
√
The local wavelength inside a dielectric grain, λd = λ/ d , is assumed to be much
larger than the grain size a. Then, the electric fields inside and outside a dielectric
particle are given by the following well-known equations, already familiar from
Chapters 4 and 5 of Volume I:
3e
Ein,d = E0 , (297)
2e + d
$ %
3 e −  d E0 · r
Eout,d = E0 + a ∇ . (298)
2e + d r3
Similar equations can be obtained for the magnetic field excited by a uniform
magnetic field H0 inside and outside a metal (dielectric) particle:
Hin,m = Hin,m0 + 4πM, (299)
where
3Ke
Hin,m0 = H0 , (300)
2Ke + K̃m
and the renormalized metal magnetic permeability K̃m is given by
2F (ym a)
K̃m = Km , (301)
1 − F (ym a)
where the function F is defined by Eq. (293). Note that the renormalized metal
magnetic permeability K̃m is not equal to one, even if the metal is non-magnetic and
the seed magnetic permeability Km = 1. The local magnetic field M in Eq. (299)
is the solution of
1 ik
∇×M = ∇ × Hin,m = − DH , (302)
4π 4π
with
akKm Ke sin(ym r)F (km r) 7y x 8
DH = 3iH0 ,− ,0 , (303)
(2Ke + K̃m ) sin(ym a)[F (ym a) − 1] r r
where DH is the electric displacement induced in the metal particle by high-
frequency magnetic field H0 . The displacement DH can be written as DH =
i(4π/ω)I, where the eddy electric current I is called the Foucault current. The
field Hin,m0 is the potential (curl-free) part, while M is the rotational part of the
local magnetic field. The magnetic field outside the metal particle is irrotational
and equals to
$ %
3 Ke − K̃m H0 · r
Hout,m = H0 + a ∇ . (304)
2Ke + K̃m r3
144 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

We assume, for simplicity, that the dielectric component of the composite is non-
magnetic, i.e., the dielectric magnetic permeability Kd = 1. Then,
3Ke
Hin,d = H0 , (305)
2Ke + 1
and
$ %
Ke − 1 H0 · r
Hout,d = H0 + a 3 ∇ . (306)
2Ke + 1 r3
As in all the EMAs described so far in this book, the effective parameters e and
Ke are determined by the self-consistent condition that the fluctuations in the fields
should vanish when averaged over all the (spherical) inclusions, i.e., Eout  =
pEout,m + (1 − p)Eout,d = E0 , and Hout  = pHout,m + (1 − p)Hout,d = H0 ,
where · indicates a volume averaging. Therefore, when these averagings are
carried out, they results in the following equations
e − ˜m  e − d
p + (1 − p) = 0, (307)
2e + ˜m 2e + d
Ke − K̃m Ke − 1
p + (1 − p) = 0. (308)
2Ke + K̃m 2Ke + 1
These equations are completely similar to the traditional EMAs discussed in the
previous sections and Volume I. It can be seen that the skin effect results in renor-
malization of the dielectric constant and magnetic permeability of the conducting
component. Specifically, the metal dielectric constant m and magnetic permeabil-
ity Km are replaced by ˜m and K̃m given by Eqs. (293) and (301), respectively.
This fact has an important effect on the frequency dependence of the effective
parameters. For example, it is commonly accepted that the effective conductiv-
ity ge = −iωe /(4π ) of a composite is dispersion-free, when the conductivity of
metal component gm is independent of frequency and gm  ω (which is typical
for the microwave and far-infrared ranges). Thus, as shown in Chapter 5 of Volume
I [see Eq. (5.62) there], the traditional EMA predicts that, ge = gm (3p − 1)/2 for
p > pc . Equation (307) yields the same result for the effective conductivity ge ,
but with the metal conductivity being renormalized according to Eq. (293), which
results in, ge = gm F (ym a)(3p − 1)/[1 − F (ym a)]. Thus, the effective conduc-
tivity has a dispersive behavior, provided that the skin effect in metal grains is
important. In the limit of very strong skin effect, δ  a,√the effective conductivity
decreases with the frequency as, ge ∼ gm (δ/a) ∼ gm / ω.
Another interesting prediction is that percolation composites exhibit magnetic
properties, even if such properties are absent in each component, i.e., even if
Km = Kd = 1. In this case, the real part Ke of the effective magnetic permeability
Ke is less than one and decreases with frequency, while its imaginary part Ke has
its maximum at frequencies such that, δ ∼ a.
One can now show that the effective parameters e and Ke determine propaga-
tion of an electromagnetic wave in the metal-dielectric composites. The average
 3.6. Electromagnetic Properties of Solid Composites 145

electric field is equal to

E = pEin,m + (1 − p)Ein,d = pEin,m0 + 4π L + (1 − p)Ein,d . (309)
When Eqs. (291) and (297) are substituted in Eq. (309) and Eq. (307) is taken into
account, Eq. (309) simplifies to
E = E0 + 4π L , (310)
where · indicates an average over the volume of the system. Therefore, the
irrotational part of the local field, being averaged over the volume, gives the field
E0 , while the second term of Eq. (310) results from the skin effect in metal grains.
In a similar fashion, we obtain
H = pHin,m + (1 − p)Hin,d = H0 + 4π M , (311)
where the rotational field M in the metal grains is given by Eq. (302), and M = 0
in the dielectric grains.
Consider now the average electric displacement D induced in the system by
the electric field E0 , which can be written as
D = m pEin,m0 + 4π m L + (1 − p)d Ein,d . (312)
It follows from Eq. (292) for Ein,m0 and Eq. (294) for L that the sum, m pEin,m0 +
4π m L, in Eq. (312) can be written as
$  %
3e 4π
m pEin,m0 + 4π m L = m p E0 + L dr
2e + ˜m 
  
3e k 3e ˜m
= m p E0 + i (r × BE ) dr = p E0 , (313)
2e + ˜m 2 2e + ˜m
where the integration is over the volume  = 4π a 3 /3 of a metal particle, and BE
is given by Eq. (295). Substitution of Eqs. (313) and (297) into (312) yields
D = e E0 . (314)
Therefore, the average electric displacement is proportional to the irrotational part
of the local field, and the proportionality coefficient is exactly equal to the effective
dielectric constant. In a similar way, we obtain
B = Ke H0 . (315)
Equations (314) and (315) can be considered as definitions of the fields E0 and
H0 . Indeed, if the local fields were known in the composite, the fields E0 and
H0 could be determined from these equations. Then, Eqs. (314) and (315) can be
used to determine the effective dielectric constant e and the effective magnetic
permeability Ke of a composite. These equations replace the usual constitutive
equations, D = e E and B = Ke H, which hold only in the quasi-static
case.
We now derive the governing equations for the macroscopic electromagnetism in
metal-dielectric composites. Equation (314) provides the average electric displace-
ment excited by the electric field E0 , but the local magnetic field also excites the
146 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

Foucault currents. Adding the electric displacement DH given by Eq. (303) to the
average displacement given by Eq. (314) yields the complete electric displacement,
4π i
Df = e E0 + ∇ × M. (316)
k
Note that the second term of Eq. (316) disappears when the skin effect vanishes,
i.e., when |ym |a → 0. We are still assuming that the linear size of the sample is
much smaller than the wavelength λ. Similarly, the average magnetic induction
Bf is given by
4π i
Bf = Ke H0 − ∇ × L. (317)
k
At this point, the Maxwell’s equations are averaged over macroscopic volume
 ∼ L3 , centered at point r, such that ξp  L  λ, yielding
∇ × E = ikBf = ikKe H0 + 4π∇ × L, (318)
∇ × H = −ikDf = −ike E0 + 4π ∇ × M. (319)
The order of the curl operation and the volume averages in Eqs. (318) and (319) can
be interchanged, as is usually done for derivation of the macroscopic Maxwell’s
equations. For example, ∇ × E = ∇ × [E(r)], where (r) indicates that the dif-
ferentiation is over the position r of the volume . Then, the Maxwell’s equations,
Eqs. (318) and (319), become
∇ × E0 (r) = ikKe H0 (r), (320)
∇ × H0 (r) = −ike E0 (r), (321)
which have the typical forms for macroscopic electromagnetism, describing
propagation of electromagnetic waves in composite media.
It is important to recognize that all quantities in Eqs. (310), (313), (314), (320),
and (321) are well-defined and do not depend on the assumptions made in the
course of their derivation. Thus, for example, M in Eq. (311) can be determined
as a magnetic moment of the Foucault currents per unit volume, so that
 
ik 1
M = (r × DH ) dr = (r × jH ) dr, (322)
8π  2c
where the integration now is over the volume . This definition of M is in
agreement with Eq. (302), except that it is not required that the currents IH be the
same in all the metal particles. In a similar way, one may write

ik
L = (r × BE ) dr, (323)
8π 
where the integration is still over the volume , and BE = −(4π i/k)∇ × E, with
E being the local electric field. Note that L has no direct analogue in the classical
electrodynamics, since there is no such thing as loop magnetic currents in atoms
and molecules.
 3.7. Beyond the Quasi-static Approximation: Generalized Ohm’s Law 147

3.7 Beyond the Quasi-static Approximation:

Generalized Ohm’s Law
The analysis presented above cannot be used for describing the optical properties of
semi-continuous films in the important case in which skin effects in the metal grains
are strong. Sarychev et al. (1994,1995) attempted to extend the above theoretical
analysis beyond the quasi-static approximation, which is based on the full set of
Maxwell’s equations. In their approach the quasi-static approximation is not used
because the fields are not assumed to be curl-free inside the film. In this section
we summarize their theoretical analysis and discuss its implications. We restrict
ourselves to the case in which all the external fields are parallel to the plane of the
film. This means that an incident wave, as well as the reflected and transmitted
waves, are travelling in the direction perpendicular to the film plane. The analysis is
focused on the electric and magnetic fields at certain distances away from the film
and attempts to relate them to the currents inside the film. We assume that the film’s
heterogeneities are over length scales that are much smaller than the wavelength
λ, but not necessarily smaller than the skin depth δ, so that the fields away from
the film are curl-free and can be expressed as gradients of potential fields. The
electric and magnetic induction currents, averaged over the film thickness, obey
the usual 2D continuity equations. Therefore, equations such as, ∇ × E = 0, and
∇ · I = 0, are the same as in the quasi-static case. The only differences are that
the fields and the average currents are now related by new constitutive equations,
and that there are magnetic as well as electric currents.
In contrast to the traditional analyses, it is not assumed that the electric and mag-
netic fields inside a semi-continuous metal film are curl-free and z-independent,
where the z-coordinate is perpendicular to the film plane. Let us consider first a
homogeneous conducting film with a uniform conductivity gm and thickness d,
and assume constant electric field E1 and magnetic field H1 at some reference
plane z = −d/2 − l0 behind the film, as shown in Figure 3.13. Under these con-
ditions, the fields depend only on the z-coordinate, and Maxwell’s equations for a
monochromatic field can be written as
d iω
E(z) = − K(z)[n × H(z)], (324)
dz c
d 4π
H(z) = − g(z)[n × E(z)], (325)
dz c
with boundary conditions
E(z = −d/2 − l0 ) = E1 , H(z = −d/2 − l0 ) = H1 , (326)
where E1 and H1 are parallel to the film plane. Here, the conductivity g(z) is
equal to the metal conductivity gm inside the film (−d/2 < z < d/2) and to
gd = −iω/4π outside the film (z < −d/2 and z > d/2). Similarly, the magnetic
permeability K(z) = Km and 1 inside and outside the film, respectively; the unit
vector n = {0, 0, 1} is perpendicular to the film plane. When solving Eqs. (324)
and (325), we must take into account the fact that the electric and magnetic fields
are continuous at the film boundaries. Then, electric (IE ) and magnetic (IH ) cur-
148 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

Figure 3.13. Schematics of the model used in the computations. Electromagnetic wave
of wavelength λ is incident on a thin metal-insulator film of thickness d. The wave is
partially reflected and absorbed, and the remainder passes through the film (after Sarychev
and Shalaev, 2000).

rents flowing in between the two planes at z = −d/2 − l0 and z = d/2 + l0 are
calculated as
  d/2  d/2+l0 
iω −d/2
IE = − E(z) dz + m E(z) dz + E(z) dz ,
4π −d/2−l0 −d/2 d/2

(327)
   
iω −d/2 d/2 d/2+l0
IH = H(z) dz + Km H(z) dz + H(z) dz ,
4π −d/2−l0 −d/2 d/2

(328)
where m = 4iπgm /ω is the metal dielectric constant. We assume, for simplicity,
that the magnetic permeability Km = 1. Since the Maxwell’s equations are linear,
the local fields E(z) and H(z) are linear functions of the boundary values E1 and
 3.7. Beyond the Quasi-static Approximation: Generalized Ohm’s Law 149

H1 defined at the plane z = −d/2 − l0 :

E(z) = a(z)E1 + c(z)(n × H1 ), (329)
H(z) = b(z)H1 + d(z)(n × E1 ). (330)
By substituting Eq. (329) for E(z) and (330) for H(z) in Eqs. (229) and (230), we
can express the currents IE and IH in terms of the boundary fields E1 and H1 :
IE = sE1 + g1 (n × H1 ), (331)
IH = mH1 + g2 (n × E1 ). (332)
Note that, Eq. (331) implies that, in contrast to the usual constitutive equations,
the current IE (which flows between the planes z = −d/2 − l0 and z = d/2 +
l0 ) depends not only on the external electric field E1 , but also on the external
magnetic field H1 , and similarly for the current IH . These equations are referred to
as the generalized Ohm’s law (GOL). The Ohmic parameters s, m, g1 and g2 have
the dimension of surface conductivity and depend on the frequency ω, the metal
dielectric constant m , the film thickness d, and the distance l0 between the film
and the reference plane z = −d/2 − l0 . We assume that the films are invariant
under reflection through the plane z = 0. In this case (Sarychev et al., 1995),
g1 = g2 = g. The Ohmic parameter g is then expressed in terms of parameters s
and m as
& 
c c 2
g=− + − ms. (333)
4π 4π
Thus, the GOL equations take the following forms
IE = sE1 + g(n × H1 ), (334)
IH = mH1 + g(n × E1 ). (335)
The Ohmic parameters s and m can be expressed in terms of the film refractive
√
index η = m and its thickness d:
c 7 8
s= exp(−idkη) [η cos(adk) − i sin(adk)]2 − exp(idkη) [η cos(adk) + i sin(adk)]2 ,
8nπ
(336)
c 7 8
m= exp(−idkη) [i cos(adk) + η sin(adk)]2 − exp(idkη) [−i cos(adk) + η sin(adk)]2 ,
8ηπ
(337)
where k = ω/c. We still assume, for simplicity, that  = 1 outside the film, and
introduce a dimensionless parameter a ≡ l0 /d. In these notations, the skin (pen-
etration) depth δ is, δ = 1/k[Im(n)]. In the microwave spectral range, the metal
conductivity
√ is real while the dielectric constant m is purely imaginary, so that,
δ = c/ 2πgm ω. On the other hand, the dielectric constant is negative for a typ-
ical metal
√ in the optical and infrared spectra ranges, and therefore, in this case,
δ  1/k |m |.
In the case of laterally heterogeneous films, the currents IE and IH , as well as
the fields E1 and H1 , are functions of the 2D vector r = {x, y}. It follows from
150 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

Maxwell’s equations that the fields and currents are connected by linear relations
given by
IE (r) = sE1 + g(n × H1 ), (338)

IH (r) = mH1 + g(n × E1 ), (339)

in which s, m and g represent integral operators. The metal islands in semi-
continuous films usually have an oblate shape, so that the grain diameter D is
much larger than the film thickness d. When the thickness d of a conducting grain
(or the skin depth δ) and the distance l0 are much smaller than the grain diameter D,
the relations between the fields E1 and H1 on one hand and the currents on the other
hand become completely local in Eqs. (338) and (339). The local Ohmic parameters
s = s(r), m = m(r), and g = g(r), given by√ Eqs. (333), (336) and (337), are deter-
mined by the local refraction index, η(r) = (r), where (r) is a local dielectric
constant. Equations (338) and (339) are the local GOL for semi-continuous films.
For binary metal-dielectric semi-continuous films the local dielectric constant is
equal to either m or d . The electric (IE ) and magnetic (IH ) currents given by
Eqs. (338) and (339) lie in between the planes z = −d/2 − l0 and z = d/2 + l0 ,
and satisfy the usual 2D continuity equation, ∇ · IE (r) = 0, and ∇ · IH (r) = 0,
which follow from the 3D continuity equations when the z-components of E1 and
H1 are neglected at the planes z = ±(d/2 + l0 ), made possible by the fact that
these components are small since the average fields E1  and H1  are parallel
to the film plane. Since we are considering semi-continuous films with a scale of
heterogeneities much smaller than the wavelength λ, the fields E1 (r) and H1 (r)
are still the gradients of potential fields when considered as functions of x and y
in the fixed reference plane z = −d/2 − l0 , i.e.,
E1 (r) = −∇ϕ1 (r), H1 (r) = −∇ψ1 (r). (340)
By substituting these expressions and Eqs. (338) and (339) in the continuity
equation, one obtains
∇ · [s∇ϕ1 + g(n × ∇ψ1 )] = 0, ∇ · [m∇ψ1 + g(n × ∇ϕ1 )] = 0. (341)
Equations (341) must be solved with the conditions that
∇ϕ1  = E1 , ∇ψ1  = H1 , (342)
where the constant fields E1  and H1  are external (given) fields, and · indicates
an average over coordinates x and y.
Summarizing, the basic idea behind the GOL is as follows. The properties of
a 3D heterogeneous layer, which are described by the complete set of Maxwell’s
equations, are reduced to a set of quasi-static equations in a 2D reference plane, with
the price being the introduction of coupled electric/magnetic fields and currents
and dependence on one adjustable parameter, namely, the distance l0 from the
reference plane. Comparison of numerical calculation and the GOL approximation
for metal films with periodic corrugation (Levy-Nathansohn and Bergman, 1997)
indicate that the GOL results are not sensitive to l0 . The original choice l0 = 0.25D
 3.7. Beyond the Quasi-static Approximation: Generalized Ohm’s Law 151

(Sarychev et al., 1995) [i.e., the parameter a = D/4d in Eqs. (336) and (337)]
allows one to reproduce most of the computer simulations’ results, except when a
surface polariton is excited in the corrugated film.
To simplify (341), the system of the basic equations, the electric and magnetic
fields on both sides of the film must be analyzed, namely, one must consider
these fields at a distance l0 behind the film, E1 (r) = E(r, −d/2 − l0 ), H1 (r =
H(r, −d/2 − l0 ), and at the same distance in front of the film, E2 (r) = E(r, d/2 +
l0 ), and H2 (r) = H(r, d/29 + l0 ). Then, Maxwell’s second equation, ∇ × H =
(4π/c)I, can be written as, H dl = (4π/c)(n1 · IE ), where n1 is perpendicular
to the integration contour, and the integration is over a rectangular contour which
has sides d + 2l0 and , such that the sides with length d + 2l0 are perpendicular
to the film and those with length  are in the planes z = ±(d/2 + l0 ). In the limit
 → 0 this equation takes the following form
4π 4π
H2 − H1 = − (n × IE ) = − [s (n × E1 ) − gH1 ] . (343)
c c
The same procedure, when applied to Maxwell’s first equation, ∇ × H = ikH,
yields
4π 4π
E 2 − E1 = − (n × IH ) = − [m (n × H1 ) − gE1 ] . (344)
c c
Now, the electric field E1 can be expressed, using Eq. (343), in terms of the
magnetic fields H1 and H2 , while the magnetic field H1 can be expressed, using
Eq. (344), in terms of the electric fields E1 and E2 . If we substitute the resulting
expressions in the GOL, Eqs. (338) and (339), and use Eq. (333), we obtain
IE = uE, IH = wH, (345)
where E = 12 (E1 + E2 ), H = 12 (H1 + H2 ), and parameters u and w are given by
c g c g
u=− , w=− , (346)
2π m 2π s
implying that the GOL is diagonalized by introducing new fields E and H, such that
Eqs. (345) have the same form as constitutive equations of macroscopic electro-
dynamics, but with the difference that the local conductivity has been replaced by
the parameter u and the magnetic permeability K has been replaced by −4iπ w/ω.
It is then straightforward to show that, the new Ohmic
√ parameters u and w can be
expressed in terms of the local refractive index η = (r) as
c tan(Dk/4) + η tan(dkη/2)
u = −i , (347)
2π 1 − η tan(Dk/4) tan(dkη/2)
c η tan(Dk/4) + tan(dkη/2)
w=i . (348)
2π η − tan(Dk/4) tan(dkη/2)
√ √
In these equations, the refractive index η takes on values ηm = m and ηd = d
for metal and dielectric regions of the film, respectively. In the quasi-static limit,
when the optical thickness of the metal grains is small, dk|ηm |  1, but the metal
152 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

dielectric constant is large in magnitude, |m |  1, the following estimates are

obtained
$ %
ωm ω 1
um  −i d, wm  i d+ D , (d/δ  1) (349)
4π 4π 2
for the metal grains. In the opposite limit, when the skin effect is strong, i.e., when
δ = 1/k[Im(ηm )]  d, and the electromagnetic field does not penetrate the metal
grains, we have
2c2 ωD
um = i , wm = i . (350)
π Dω 8π
In the dielectric region, when the film is thin enough that, dkηd  1 and d ∼ 1,
we obtain
$ %
ωd ω 1
ud = −i D, wd = i d+ D , (351)
8π 4π 2
where d = 1 + 2d d/D.
Potentials for the fields E2 (r) and H2 (r) may be introduced for the same reason
as for E1 (r) and H1 (r). Therefore, the fields E(r) and H(r) in Eqs. (345) can in turn
be represented as gradients of some potentials, E = −∇φ  , and H = −∇ψ  . By
substituting these expressions into Eqs. (345) and then in the continuity equation,
we obtain two equations for φ  (r) and ψ  (r) that can be solved independently
under the conditions that, ∇φ1  = E ≡ E0 , and ∇ψ1  = H1  ≡ H0 , where
the constant fields E0 and H0 are external (given) fields that are determined by
the incident wave. When E, H, IE , and IH are determined from the solution of
these equations, the local electric and magnetic fields in the plane z = −l0 − d/2
are given by, E1 = E + (2π/c)(n × IH ), and H1 = H + (2π/c)(n × IE ). Note
that the field E1 (r) is usually measured in a typical near field experiment. Then,
the effective parameters ue and we are defined in a way similar to Eqs. (345),
viz., IE  = ue E0 ≡ 12 ue (E1  + E2 ), which, when substituted in Eqs. (343)
and (344) (which are averaged over the {x, y} coordinates), yield
2π
[n × (H2  − H1 )] = ue (E1  + E2 ), (352)
c
2π
[n × (E2  − E1 )] = we (H1  + H2 ). (353)
c
Suppose now that the wave enters the film from the right-half space, such that its
amplitude is proportional to exp(−ikz). The incident wave is partially reflected and
partially transmitted through the film. The electric field amplitude in the right-half
space, away from the film, can be written as e[exp(−ikz) + r exp(ikz)], where r
is the reflection coefficient and e is the polarization vector. Then, the electric com-
ponent of the electromagnetic wave well behind the film will be e[t exp(−ikz)],
where t is the transmission coefficient. We assume for simplicity that the film has
no optical activity, which means that the wave polarization e remains the same
before and after the film. At the planes z = d/2 + l0 and z = −d/2 − l0 the av-
erage electric field is E2  and E1 , respectively. On the other hand, the wave
 3.7. Beyond the Quasi-static Approximation: Generalized Ohm’s Law 153

away from the film is matched with the average fields in the planes z = d/2 + l0
and z = −d/2 − l0 , i.e., E2  = e {exp[−ik(d/2 + l0 )] + r exp[ik(d/2 + l0 )]}
and E1  = e{t exp[ik(d/2 + l0 )]}. The same matching, but with the magnetic
fields, yields, H2  = (n × e) {− exp[−ik(d/2 + l0 )] + r exp[ik(d/2 + l0 )]} and
H1  = −(n × e)t exp[ik(d/2 + l0 )] in the planes z = d/2 + l0 and z = −d/2 −
l0 , respectively. Substitution of these expressions for E1 , E2 , H1 , and H2 
in Eqs. (352) and (353) yields two scalar, linear equations for reflection (r) and
transmission (t) coefficients, the solution of which yields the reflectance,
. .2
. 2π .
. (ue + we ) .
. .
R ≡ |r| = .
2 . $ c % $ % . , (354)
.
. 1 + 2π u 1−
2π
we ..
. e
c c
the transmittance
. $ %2 .2
. 2π .
. 1+ .
. ue we .
. c .
T ≡ |t| = . $
2 %$ %. , (355)
. 2π 2π .
. 1+ ue 1− we ..
. c c
and the absorbance
α =1−T −R (356)
of the film. Therefore, the effective Ohmic parameters ue and we determine com-
pletely the optical properties of heterogeneous films. This analysis indicates that,
the problem of the field distribution and optical properties of the metal-dielectric
films reduces to uncoupled quasi-static conductivity problems for which extensive
theoretical analyses have already been carried out. Numerous analytical as well as
numerical methods, developed for heterogeneous media with percolation disorder
(see Chapters 4–6 of Volume I), can be employed for determining the effective
parameters ue and we of the film.
We can now consider the case of strong skin effect in the metal grains and
study the evolution of the optical properties of a semi-continuous metal film,
as the volume fraction p of the metal is increasing. When p = 0, the film is
purely dielectric and ue = ud and we = wd , where ud and wd are the dielectric
Ohmic parameters given by Eqs. (351). If we substitute ue = ud and we = wd
in Eqs. (354)–(356) and assume that the dielectric film has no losses and is op-
tically thin (i.e., dkd  1), we obtain the reflectance R = d 2 (d − 1)2 k 2 /4, the
transmittance T = 1 − d 2 (d − 1)2 k 2 /4, and the absorbance α = 0, well-known
results for a thin dielectric film (see, for example, Jackson, 1998). The losses are
also absent in the limit of full coverage, i.e., when the metal volume fraction p = 1.
Indeed, substituting the Ohmic parameters ue = um and we = wm from Eqs. (350)
in Eqs. (354)–(356) yields, R = 1, T = 0, and α = 0. Note that in the limits p = 0
and p = 1, the optical properties of the film do not depend on the particle size D,
because properties of the dielectric and continuous metal films should not depend
on the shape of the metal grains.
154 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

Next, we consider the film at p = pc with pc = 1/2 for a self-dual system. A

semi-continuous metal film may be thought of as a mirror, which is broken into
small pieces with typical size D  λ. At pc , the exact equations (see Sections 3.1
√ √
and 3.3, and also Chapters 4 and 5 of Volume I), ue = ud um and we = wd wm ,
which result from the exact duality relation, hold. Thus,
" &
2π  2π Dk 2d
ue (pc ) = d , we (pc ) = i 1+ , (357)
c c 4 D
from which it follows that |we /ue | ∼ Dk  1, and hence, compared with ue , we
can be neglected. Under this condition, we obtain
d
R(pc ) =  " 2 , (358)
1 + d

1
T (pc ) =  " 2 , (359)
1 + d
"
2 d
α(pc ) =  " 2 , (360)
1 + d

where, d = 1 + 2d d/D, as before. When metal grains are oblate enough that
d d/D  1 and d → 1, one obtains the universal result
R = T = 1/4, α = 1/2, (361)
implying that there is effective absorption in semi-continuous metal films even for
the case when neither dielectric nor metal grains absorb light energy. The effective
absorption in a loss-free film means that the electromagnetic energy is stored in
the system, and that the amplitudes of the local electromagnetic field can diverge.
In practice, due to non-zero losses, the local field saturates in any semi-continuous
metal film.
To determine the optical properties of semi-continuous films for arbitrary metal
volume fraction p, the EMA can be used which then yields the following equations,
u2e − pue (um − ud ) − ud um = 0, (362)
we2 − pwe (wm − wd ) − wd wm = 0, (363)
where p = (p − pc )/pc (pc = 1/2). Equation (363) indicates that, when the
skin effect is strong and wm and wd are given by Eqs. (350) and (351), then
|we |  c for all metal volume fractions p, and therefore we can neglect we in
Eqs. (354) and (355). Moreover, compared with um , ud can also be neglected in the
second term of Eq. (362). Thus, introducing the dimensionless Ohmic parameter
ue = (2π/c)ue allows us to rewrite Eq. (362) as
λp 
ue − 2i u − d = 0.
2
(364)
πD e
 3.8. Piecewise Linear Transport Processes 155
"
At p = pc = 1/2 (i.e., where p = 0), Eq. (364) predicts that, ue (pc ) = d ,
which coincides with the exact result, Eq. (357), and those given by Eqs. (358)–
(360). For p
= pc , Eq. (364) predicts that
#
$ %
 λp  λp 2
ue = i + d − , (365)
πD πD
"
which becomes purely imaginary for |p| > πD d /λ. Then, α = 1 − |ue |2 /
|1 + ue |2 − 1/|1 + ue |2 = 0 "
(recall that we was neglected). In the vicinity of pc ,
namely, for |p| < (π D/λ) d , the effective Ohmic parameter ue has a non-
vanishing real part, and therefore
"
2 d − [λp/(π D)]2
α= " , (366)
1 + d + 2 d − [λp/(π D)]2
which has a well-defined maximum at pc , with the width of the maximum being
inversely proportional to the wavelength. These predictions were confirmed by
extensive numerical simulations. They are also in agreement with the experimen-
tal data (see Sarychev and Shalaev, 2000, for detailed discussions). Note that the
parameters ue and we can be determined experimentally by measuring the am-
plitudes and phases of the transmitted and reflected waves using, for example, a
waveguide technique (see, for example, Golosovsky et al., 1993 and references
therein), or by measuring the film reflectance as a function of the fields E1 and H1 .

3.8 Piecewise Linear Transport Processes

The last nonlinear transport process that we describe and analyze is not caused
by strong morphological disorder and its interplay with a transport process, rather
it has to do with the constitutive relation between the current and the potential
gradient, augmented by a threshold in the potential gradient. Such nonlinear trans-
port phenomenon are typically piecewise linear, or possibly nonlinear, and are
characterized by at least one threshold. Several possible I − V characteristics of
such materials are shown in Figure 3.14. Because of the threshold, of course, even
a piecewise linear transport is in fact a highly nonlinear process. In many cases,
the regime below the threshold is degenerate in the sense that, nothing interesting
happens if the driving potential applied to the material is below its threshold. The
applications of this type of nonlinear transport process are numerous. For example,
bipolar Zener diodes (which are commercially called varistors) switch from being
a non-conducting link to a conducting one at an onset voltage threshold vc . More
generally, a network of such diodes can become conducting only if the voltage
applied to it is larger than a critical value Vc . In brittle fracture, which will be
studied in Chapters 6–8, no microcrack nucleation and propagation take place un-
less the external stress or strain applied to a solid material exceeds a critical value
156 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

Figure 3.14. Twelve types of physically realiz-

able nonlinear I − V characteristics, seven of which
are also characterized by a threshold (after Sahimi,
1998).

which depends on the size of the sample. Bingham fluids are viscous and behave
like Newtonian fluids if the shear stress applied to them is larger than a critical
value, but do not flow if the stress is less than the threshold value. An example
of such fluids, already described in Section 9.3 of Volume I, is foam. In order to
mobilize foam and force it to flow, the applied pressure must exceed a critical
value; otherwise it will not flow.
Let us consider a 2D or 3D resistor network in which every bond is characterized
by the following current-voltage relation,

g(v − vc )n , v > vc ,
i= (367)
0 v ≤ vc ,

where vc is the critical voltage or threshold for the onset of transport. As in the
case of strong and weak nonlinearities, we take g to be a generalized bond con-
ductance which, in general, can vary from bond to bond. On the other hand, in any
physical situation involving a disordered material, one expects a distribution of
the thresholds vc , because due to a variety of factors, different parts of a material
may become conductive beyond different thresholds. Therefore, one may make
the simplification that, instead of assuming g to be a statistically-distributed vari-
able, vc is assumed to be randomly distributed which, for the sake of simplicity, is
assumed to be distributed uniformly in (0, 1). The conductivity g is then the same
for all bonds, and therefore its numerical value is irrelevant (we assume g = 1).
The questions that we ask are:
(1) What is the critical voltage Vc in order to have macroscopic transport in the
network, and
(2) how do the macroscopic current I and the effective conductivity ge of the
network vary with the applied voltage? The piecewise linear process that we
study here is reversible, i.e., if I is lowered the conducting bonds become
 3.8. Piecewise Linear Transport Processes 157

insulating again. This is an important assumption since, if we assume that the

process is irreversible, then converting even one insulating bond to a conduct-
ing one triggers an avalanche effect: The conversion of the first bond makes
consecutive conversions easier. Such irreversible and nonlinear models have
been used to model brittle fracture and electrical and dielectric breakdown of
disordered materials, which will be discussed in Chapters 5–8.
It is clear that for any applied voltage V less than a critical threshold Vc no
macroscopic current can flow. Therefore, it should also be clear that
 


Vc = min vci , (368)
i

where vci is the critical voltage of bond i, and the sum is taken over all paths
between the two terminals of the network. Equation (368) immediately necessitates
the concept of an optimal path between the two terminals of the network (see, for
example, Cieplak et al., 1994,1996; Porto et al., 1997). Obviously, if the applied
voltage is larger than a final or the last voltage threshold Vl , all bonds of the network
will be conducting, and one will have the usual linear transport in which the
current I is simply proportional to V . Therefore, one generally has three regimes
of interest:
(1) If V < Vc , then enough bonds have not become conducting to form a sample-
spanning cluster, and therefore no macroscopic transport takes place. Hence,
I = 0 and ge =0.
(2) If Vc < V < Vl , then enough bonds have become conducting that make macro-
scopic transport possible, while some of the bonds are still not conducting. We
expect I to depend nonlinearly on V − Vc , because this is precisely the regime
in which the effect of nonlinearity (random voltage thresholds) should manifest
itself. As we show below, this is indeed the case (note that in linear transport
above pc , I always varies linearly with V ).
(3) If V > Vl , then every bond of the network is conducting, the normalized
effective conductivity is ge = 1, and I depends linearly on V again.

3.8.1 Computer Simulation

Computer simulation of this problem, even for n = 1, is difficult, and thus deserves
to be discussed here. At the beginning of the simulations, one distributes the critical
thresholds vc of the bonds and applies a large enough external voltage to the
network, such that every bond becomes conducting (i.e., the voltage across it
exceeds its critical voltage). The external voltage is then decreased gradually, and
the nodal voltage distribution and hence the current distribution in the bonds are
computed. As a result of lowering the applied voltage, some of the conducting
bonds become insulating (since the voltage across them will be less than their
critical voltage). The new voltage and current distributions are calculated, the
newly-insulating bonds are identified, and so on.
158 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

3.8.2 Scaling Properties

Roux and Herrmann (1987) used accurate numerical simulations, and Gilabert et
al. (1987) utilized an analogue resistor network, to find that in 2D and near Vc ,
I ∼ (V − Vc )δ , (369)
with δ  2 ± 0.08. The power-law (369) is the only scaling property of piecewise
linear transport that has been studied so far.

3.8.3 Effective-Medium Approximation

We now describe the predictions of an EMA for piecewise linear transport and
compare its predictions with simulation results. We consider only the case n = 1
and present the final results; complete details are given by Sahimi (1993a). Sup-
pose that p is the fraction of the bonds that have become conducting. Then, in the
non-conducting regime, i.e., before a sample-spanning conducting path has formed
between two opposite faces of the network and p < pc = 2/Z (recall from Sec-
tions 3.1 and 3.2 that, since the problem is treated within an EMA, the percolation
threshold is pc = 2/Z), the applied voltage V varies with p according to
1
V = p − p2 , p < 2/Z. (370)
2
Equation (370) predicts how the applied voltage V varies with p before a sample-
spanning conducting path is formed. At p = pc = 2/Z the first sample-spanning
conducting path is formed and therefore
2 2
Vc = − 2, (371)
Z Z
which is obtained by substituting p = pc = 2/Z in Eq. (370). For p > 2/Z we
have a conducting system for which
Z−2 2 2
V = p + 2, p≥ . (372)
Z Z Z
At p = 1 all the bonds are conducting, so that the corresponding last voltage for
converting the last bond to a conducting bond is given by
Z−2 2
Vl = + 2. (373)
Z Z
The corresponding equations for the effective conductivity ge of the network
are as follows. Clearly, for V < Vc we must have ge = 0. For Vc ≤ V ≤ Vl we
have
Z2 2 2
ge = (V − 2 ) − , Vc ≤ V ≤ Vl . (374)
(Z − 2) 2 Z Z−2
Obviously, ge = 1 for V ≥ Vl .
 3.8. Piecewise Linear Transport Processes 159

We can also determine the macroscopic I − V characteristic of the material. For

V < Vc there is no macroscopic transport and therefore, I = 0. For Vc ≤ V ≤ Vl
we have
Z2 2 2 2 2
I= (V − 2 )2 − (V − 2 ) + 2 , Vc ≤ V ≤ Vl . (375)
2(Z − 2) 2 Z Z−2 Z Z
For V ≥ Vl , we have ge = 1, and therefore the current I is related to the applied
voltage through a simple equation
1
I =V − , (376)
2
independent of Z. Thus, the EMA predicts correctly the existence of the three
transport regimes discussed above and, in particular, it predicts that for Vc ≤ V ≤
Vl , I depends quadratically on V − Vc , where Vc = 2/Z − 2/Z 2 .
Figure 3.15 presents the dependence of ge on the applied voltage V in a square
network. All the qualitative features of the transport process are correctly predicted
by the EMA, except that the numerical simulations indicate smooth variations
of ge with V, whereas the EMA predicts a sharp, discontinuous, transition at
V = Vl . Figure 3.16 presents the variations of the macroscopic current I with
the applied voltage V in the same system and, unlike ge , both the numerical
calculations and the EMA predict no sharp transition at V = Vl . However, the
numerical value of the critical voltage Vc does not agree with the prediction of the

Figure 3.15. Effective conductivity of the square network with piecewise linear resistors
with a threshold, versus the applied voltage (after Sahimi, 1993a).
160 3. Nonlinear Conductivity, Dielectric Constant, and Optical Properties

Figure 3.16. The I − V characteristics of the square network of Figure 3.15 (after Sahimi,
1993a).

EMA. While computer simulations indicate that, Vc  0.29, the EMApredicts that,
Vc = 3/8 = 0.375. Roux et al. (1987) used a transfer-matrix method described
in Section 5.14.2 of Volume I and estimated that for a square network, tilted at
45◦ , one
√ has, Vc  0.23 [in general, for the square network, Vc (tilted)=Vc (non-
tilted)/ 2]. This difference can be explained by the fact that, because the resistor
network that Roux et al. (1987) used in their simulations was tilted, their network
is different from a non-tilted one, since the distribution of currents in the bonds
of their network is isotropic, whereas the same distribution is anisotropic in a
non-tilted network. The difference is due to the fact that the bonds of a non-
tilted network that are perpendicular to the direction of the applied voltage receive
much less current than those that are aligned with it. As a result, formation of a
sample-spanning conducting cluster is easier in a tilted network than in a non-
tilted one, implying that the critical voltage Vc for a tilted network should be
smaller than that of a non-tilted one. Thus, such local anisotropies, which usually
have no consequence for macroscopic properties of linear transport processes,
are important in a nonlinear system, such as what is described here. Moreover,
according to Eq. (375), in the nonlinear regime, the macroscopic current I varies
quadratically with V − Vc , which is in agreement with the simulations of Roux
and Herrmann (1987), Eq. (369).
 3.8. Piecewise Linear Transport Processes 161

Summary
Using the discrete models, we described and analyzed several types of nonlinear
transport and optical properties of disordered materials. As our analyses indicate,
the interplay of nonlinearity and the disordered morphology of a material gives
rise to a rich set of phenomena that are absent in linear transport processes in
the same material. In particular, strong heterogeneity, such as percolation-type
disorder, enhances the nonlinear response of a material, and shrinks the range of
the parameter space in which the material behaves linearly, and hence opens up
the possibility of developing composite materials with highly unusual and useful
properties.
4
Nonlinear Rigidity and Elastic Moduli:
The Continuum Approach

4.0 Introduction
In this chapter we consider nonlinear mechanical properties of heterogeneous
materials. This class of problems has many applications that will be described
throughout this chapter. However, to give the reader an interesting and somewhat
unusual application of this class of phenomena, we consider the following prob-
lem. It has been observed (Gordon, 1978) that extensible biological tissues, such
as skin, are difficult to tear, even though their specific work of fracture (see the
discussions in Chapters 6 and 7) is not large compared to those of materials that
tear easily. For example, the fracture toughness of animal membranes is around
1-10 kJm−2 , an order of magnitude smaller than aluminum foil which tears easily.
Gordon reasoned that this difference is due to the markedly different shape of the
stress-strain diagram of such materials. Figure 4.1 presents schematic stress-strain
curves for extensible biological tissues, rubber, and the standard Hookean solid for
which the diagram is a straight line. The small-strain portion of the J-shaped curve
of the biological material is indicative of lack of shear connection in the material,
i.e., absence of shear stiffness in what are anisotropic solids. This diagram provides
an explanation as to why such materials are difficult to tear, because it is difficult
to concentrate energy into the path of a putative crack. Note also the difference be-
tween the stress-strain diagrams for rubbers and the biological materials: For small
strains, the rubber’s curve is not J-shaped, which may also explain why we cannot
replace human body arteries or veins by rubber tubes. We also remind the reader
that when Nature does want fracture and tear to happen, as in, for example, amniotic
membranes and eggshells, the stress-strain diagrams are Hookean linear elastic!
Studies of heterogeneous materials with nonlinear constitutive behavior go back
to at least Taylor (1938) who studied the plasticity of polycrystals, and to the
subsequent work by Bishop and Hill (1951a,b) and Drucker (1959) who investi-
gated the behavior of ideally plastic polycrystals and composite materials. Over
the past decade or so, numerical simulations of nonlinear materials with periodic
microstructures have been carried out (see, for example, Christman et al., 1989;
Tvergaard, 1990; Bao et al., 1991), as well as materials with more general mi-
crostructures (see, for example, Brokenborough et al., 1991; Moulinec and Suquet,
1995). Such efforts will be briefly described in this chapter where we make com-
parison between the theoretical predictions and the numerical simulation results.
 4.1. Constitutive Relations and Potentials 163

Figure 4.1. Schematic representation of

different stress-strain relations.

5JHAII
Strain

The main advantage of such simulations is that they provide accurate description
of the system under study, and yield useful insight into their properties. Their main
disadvantage is that they require very intensive computations, especially when the
material’s microstructure is disordered.
In this chapter we describe and discuss recent advances in understanding
the effective mechanical properties of disordered materials with constitutive
nonlinearity. Although one may argue that numerical techniques, such as the
finite-element methods, represent some form of discrete approach to this class
of problems, to our knowledge very little work has been done using the discrete
network models of the type that we have so far described and discussed for esti-
mating various transport properties of disordered materials. Therefore, the main
focus of this chapter is on the theoretical developments based on nonlinear contin-
uum models of disordered materials. These theoretical approaches represent the
mechanical analogues of those described in Chapter 2 for estimating the effective
conductivity and dielectric constant of nonlinear materials. Thus, the methods that
we describe in this chapter are based on rigorous variational principles which, in
addition to possessing mathematical rigor, have the advantage of leading to bounds
and relatively accurate estimates for the mechanical properties. As described and
discussed in Chapter 2, such variational principles allow one to obtain estimates of
the effective energy densities of nonlinear materials in terms of the corresponding
information for linear composites with the same microstructure. A large portion
of our analyses and discussions in this chapter is based on an excellent review by
Ponte Castañeda and Suquet (1998).

4.1 Constitutive Relations and Potentials

Similar to Chapter 2, where we analyzed the effective nonlinear conductivity and
dielectric constant of disordered materials, we also assume in the present chapter
that the constitutive behavior of the individual phases of the material is governed
by a potential, or strain-energy function, w(), in such a way that the (infinitesimal)
164 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

strain  and stress σ fields are related by

∂w
σ = . (1)
∂
Although Eq. (1) is intended for nonlinear elastic behavior of materials in the limit
of small strains, by interpreting  and σ as the Eulerian strain rate and Cauchy
stress, it can also be used for modeling finite viscous deformations. Assuming
then that w is a convex function of , Eq. (1) is inverted with aid of the Legendre
transformation:
u(σ ) = sup{σ : τ − w()}. (2)

Equation (2) defines a convex stress-energy function u, such that
∂u
= . (3)
∂
The functions w and u are dual potentials and are related by the classical reciprocity
relations. As in Chapter 2, the notation u = w∗ is used to express the relation
between these two quantities.
For isotropic materials, general forms of w and u are given by
9 2
w() = K + ϕ(eq ), (4)
2 m
and
1 2
u(σ ) =σ + ψ(σeq ), (5)
2K m
where ϕ and ψ are dual convex potentials, σm and m are the hydrostatic stress
and strain given by
1 1
σm = tr(σ ) ; m = tr(), (6)
3 3
and σeq and eq are the Von Mises equivalent stress and strain,
$ %1/2 $ %1/2
3 2
σeq = σd : σd , eq = d : d , (7)
2 3
with σ d and  d being the stress and strain deviators (see also Chapter 7 of Volume
I) given by, σ d = σ − σm U and  d =  − m U. Thus, one can write
σm = 3Km , σ d = 2µ(eq ) d , (8)
with
1 σeq 1 ϕ  (eq ) 1 σeq
µ(eq ) = = = . (9)
3 eq 3 eq 3 ψ  (eq )
Therefore, each phase is assumed to be linear for purely hydrostatic loadings,
characterized by a constant bulk modulus K and nonlinear in shear, characterized
by a strain-dependent shear modulus µ.
 4.1. Constitutive Relations and Potentials 165

Consider, as an example, high temperature creep of metals, which is commonly

characterized in terms of a power-law constitutive relation. If we neglect elastic
effects and assume incompressibility, then, the dissipation ϕ and stress potential
ψ of the material are given by
σ 0  0  eq m+1 σ 0  0  σeq n+1
ϕ(eq ) = , ψ(σ eq ) = , (10)
m + 1 0 n + 1 σ0
where  0 and σ 0 denote a reference strain rate and stress, respectively, m and n
are two exponents such that m = 1/n, and m = 0. For example, for Newtonian
viscous materials, n = m = 1, where η = σ 0 /3 is the viscosity, while the Von
Mises rigid, ideally plastic materials correspond to the limit m → 0 (n → ∞),
where σ 0 now denotes the flow stress in tension. In the latter case, the stress
potential becomes unbounded for stresses that exceed σ 0 . It is then useful to
introduce the strength domain P , defined by the set
P = {σ : σeq ≤ σ 0 }. (11)
The creep of crystalline materials can also be described within this frame-
work. We consider a single crystal that undergoes creep on a set of M preferred
crystallographic slip systems, and is characterized by the second-order tensors
µi , i = 1, . . . , M, defined by
1
µi = (ni ⊗ mi + mi ⊗ ni ), (12)
2
where ni and mi are the unit vectors normal to the slip plane and along the slip
direction in the ith system, respectively, and ⊗ denotes the tensorial product of
two vectors. If a stress σ is applied to the crystal, then, the resulting shear stress
acting on the ith slip system is given by
τi = σ : µi , (13)
while the strain rate  in the crystal is the superposition of the strain rates on each
slip system,


M
= γi µi , (14)
i=1
where γi is the shear strain rate acting on the ith system, which is given by
∂ψi
γi = , (15)
∂τi
with the functions ψi being convex. An equation commonly used for ψi is
 ni +1
γ 0 τi0 |τ |
ψi (τ ) = , (16)
ni + 1 τi0

with ni ≥ 1 and τi0 being the creep exponent and reference stress of the ith slip
system, respectively, and γ 0 is a reference strain rate. The constitutive relations
166 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

(14) and (15) can then be expressed in terms of the convex potential for the crystal:

M
u(c) (σ ) = ψi (σ : µi ), (17)
i=1

such that
∂u(c)
= . (18)
∂σ
The limit ni → ∞ corresponds to a rigid, ideally plastic crystal, with a strength
domain given by
P = {σ , τi ≤ τi0 , i = 1, . . . , M}. (19)
We can assume, more generally, that the potential w can be expressed by
w() = F (E), (20)
where F is an appropriately-selected function, and E is a fourth-rank tensor which
is defined by
1
E=  ⊗ , (21)
2
and possesses the usual diagonal symmetry and positive-definitiveness property
of an elasticity tensor. The function F is then defined on the space of fourth-rank
tensors P that have diagonal symmetry, so that the constitutive relation (1) can be
written as
σ = Ls (E) : , (22)
with
∂F
Ls (E) = , (23)
∂P
being the secant modulus tensor of the material, which also has diagonal symmetry.
Given Eq. (20), the dual potential u can be expressed as
1
u(σ ) = G(S), S= σ ⊗ σ, (24)
2
where G is a function of fourth-rank tensors S. In terms of the secant compli-
ance tensor of the material, the constitutive relation (3) may be expressed in the
following form
∂G
 = Ms (S) : σ , Ms (S) = . (25)
∂S
As an example, consider crystalline materials. First, note that
τi2 = 2Mi :: S, Mi = µi ⊗ µi , (26)
 4.2. Formulation of the Problem 167

so that


M
√
u(σ ) = gi (2Mi :: S) = G(S), gi (x) = ψi ( x), (27)
i=1
and the compliance tensor is given by


M
Ms (S) = 2 αi Mi , αi = gi (τi2 ). (28)
i=1

4.2 Formulation of the Problem

We now consider a representative volume element  of a heterogeneous material,
such that the size of its heterogeneities is small compared to . The material
consists of N homogeneous phases i , i = 1, . . . , N, the distribution of which is
defined by indicator functions mi (x), which are 1 when x belongs to the phase i,
and zero otherwise. One can define two spatial averages, one over  and another
one over i , so that, for example

1
i = (x)dx, (29)
|i | i

N
1
 = (x)dx = φi i , (30)
|| 
i=1
where φi is the volume fraction of phase i. All the phases are assumed to be ho-
mogeneous with potentials wi and ui , and to be perfectly bonded at the interfaces.
The total potentials w and u are then given by


N

N
w(x, ) = mi (x)wi (), u(x, σ ) = mi (x)ui (σ ). (31)
i=1 i=1
As an example, consider a polycrystalline material, which we regard it as an
aggregate of a large number of identical single crystals with different orientations,
so that it can be treated as a composite, where phase i is defined as the region
occupied by all grains of a given orientation, relative to a reference crystal with
potential u(c) given by (17). If Qi denotes the rotation tensor that defines the
orientation of phase i, the corresponding potential ui is given by
 
M  
(k)
ui (σ ) = u(c) QTi · σ · Qi = ψ k τi , (32)
k=1
where
(k) (k) (k)
τi = σ : µi , µi = QTi · µk · Qi . (33)
The microscopic problem is one in which the local stress and strain fields
within  solve a local problem that consists of the constitutive relation (1), the com-
168 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

patibility conditions satisfied by , and the usual equilibrium equations satisfied

by σ :
∂w 1 
σ = , = ∇u + (∇u)T , ∇ · σ = 0, (34)
∂ 2
subject to one of the two classes of boundary conditions on ∂. One is in terms
of affine displacements,
u(x) = E · x, (35)
while the second one is in terms of uniform traction,
σ (x) · n(x) =  · n(x). (36)
Here E and  are the averages of the local strain and stress fields:


N
E =  = φi i , (37)
i=1

N
 = σ  = φi σ i , (38)
i=1

and satisfy (Hill, 1963)

 : E = σ : (u). (39)

4.3 The Classical Variational Principles

As discussed in Chapter 2, and also Chapters 4 and 7 of Volume I, the solutions u
and σ of the local problem can be given two equivalent variational representations.
One is in terms of the minimum potential energy which states that u is the solution
of the problem
inf w[(v)), (40)
V ∈S1 (E)

where
S1 (E) = {v = E · x on ∂}, (41)
while the second one is in terms of the minimum complementary energy, according
to which τ is the solution of the problem
inf u(σ ), (42)
τ ∈S2 ( )
with
S2 () = {τ , ∇ · τ = 0, in , τ  = }. (43)
 4.3. The Classical Variational Principles 169

Then, since the infimum problem in (40) defines the average strain energy in the
material, the effective strain-energy potential He is defined as
He (E) = inf w[(v)], (44)
v∈S1 (E)

so that
5 $ %6 5 $ %6
∂He ∂w ∂u ∂u
= [(u)] :  = σ : . (45)
∂E ∂ ∂E ∂E
However, since ∂u/∂E = U · x, where U is the identity tensor in the space of
fourth-rank tensors, it follows from Eq. (39) that
∂He
= σ  : U = , (46)
∂E
which defines the effective stress-strain relation for the material. Similarly, the
effective stress-energy potential He∗ is defined as
He∗ () = inf u(τ ), (47)
τ ∈S1 ( )
in terms of which,
∂He∗
E= . (48)
∂
Both He and He∗ are convex functions. Furthermore, it can be shown (Suquet,
1987; Willis, 1989a) that they are in fact the (Legendre) dual functions, such that
He (E) + He∗ () = w() + u(σ ) = σ :  =  : E, (49)
and that they correspond to the boundary conditions (35). Adopting the boundary
condition (36) would lead to different pairs of dual potentials. However, under
the assumption that the potentials w and u are strictly convex, the two types of
boundary conditions are equivalent for the representative volume element, and
are also equivalent to the periodic boundary conditions used in the theory of
homogenization (see, for example, Sanchez-Palencia, 1980).
As an example, consider the limiting case of rigid, ideally plastic materials for
which the potentials are convex, but not strictly. In this limit, which requires special
treatment (Bouchitte and Suquet, 1991), He is a positively-homogeneous function
of order one in E, usually referred to as the plastic dissipation function. It may
also be useful to introduce the effective strength domain of the material, defined
as (Suquet, 1983)
Pe = { such that there exists σ (x) with σ  =  and

∇ · σ (x) = 0, with σ (x) ∈ Pi , for x in phase i}. (50)

Note that
He (E) = sup { : E}
 ∈Pe
170 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

and that

0 if  ∈ Pe
He∗ () =
+∞ otherwise

4.3.1 One-Point Bounds

As already discussed in Chapter 2, the minimum energy principles can be utilized
for deriving rigorous bounds for the effective potentials He and He∗ for rigid, ideally
plastic polycrystalline materials (Bishop and Hill, 1951a,b) and for materials with
elastic, ideally plastic phases (Drucker, 1966). If we use uniform trial fields in
the variational principles, the following rigorous bounds of the Voigt (1889) and
Reuss (1929) type are obtained:


N
He (E) ≤ u(E) = φi wi (E), (51)
i=1

and


N
He∗ () ≤ u() = φi wi (), (52)
i=1

or, equivalently,
N ∗ N 



φi ui (E) ≤ He (E) ≤ φi wi (E), (53)
i=1 i=1

where superscript ∗ denotes the convex dual function. In the context of polycrys-
talline materials, the bounds (51) and (52) are commonly referred to as the Taylor
(1938) and Sachs (1928) bounds, respectively. For example, the Reuss and Voigt
bounds for incompressible, isotropic power-law phases are given by
$ % $ %
(σ 0 )−n −m  0 Eeq m+1 σ 0  0 Eeq m+1
≤ H e (E) ≤ . (54)
m+1 0 m+1 0
Since the Voigt and Reuss bounds incorporate only limited information on the
morphology of a material—the volume fractions of the phases—they are not very
useful, particularly when the contrast between the phases is large. In fact, they can
be shown to be exact only to first order in the contrast between the properties of
the phases.

4.3.2 Two-Point Bounds: The Talbot–Willis Method

We have already described and discussed in Chapter 2, as well as Chapters 4 and
7 of Volume I, the variational procedure of Hashin and Shtrikman (Hashin and
Shtrikman, 1962a,b, 1963). A generalization of the Hashin–Shtrikman variational
principles, suitable for nonlinear materials, was developed by Talbot and Willis
 4.3. The Classical Variational Principles 171

(1985), following the earlier work of Willis (1983), which we now describe and
discuss.
Let w 0 be the potential function of a linear, homogeneous reference material
with uniform modulus tensor L0 , such that
1
w0 () =  : L0 : , (55)
2
and assume that the difference potential (w − w0 ) is a concave function, so that
the concave polar of this difference is defined as (see Ponte Castañeda and Suquet,
1998)
7  8
(w − w0 )∗ (x, τ ) = inf τ :  − w(x, ) − w0 () .

The concavity of (w − w 0 ) results in
7 8
w(x, ) − w 0 () = inf τ :  − (w − w 0 )∗ (x, τ ) , (56)
τ
Substituting (56) for w in Eq. (47) and interchanging the order of the infima over
 and τ , one arrives at
 7 8
He (E) = inf inf w 0 [(v)] + τ : (v) − (w − w 0 )∗ (x, τ ) . (57)
τ v∈S1 (E)
It then follows that minimizing the displacement field u is equivalent to finding
the solution to the following boundary value problem:
 
∇ · L0 : (u) = −∇ · τ , u ∈ S1 (E). (58)

If one utilizes the Green function G0 associated with the system (58) in the domain
, one obtains the following expressions for the strain tensor,
 = E − 0 ∗ τ , (59)
where, as before, E is the average strain over , and

2 3
 ∗τ =
0
 0 (x, x ) : τ (x ) − τ  dx , (60)

with
 
∂ 2 G0ik
ij0 kl = .
∂xj ∂xl
(ij ),(kl)

Note that the concavity of (w − w0 )

is essential in attaining the equality in (57).
Typically, however, (w − w 0 ) is neither concave nor convex, as in the case of,
for example, a power-law material. In such a case, the equality in (57) must be
replaced by an inequality [either ≤ or ≥, depending on whether (w − w 0 ) grows
weaker-than-affine or stronger-than-affine at infinity, respectively].
However, as already pointed out in Chapter 2, as well as Chapters 4 and 7
of Volume I, it is very difficult, if not impossible, to determine the exact τ that
172 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

satisfies (57). Because of this difficulty, an approximation of the following form

(the so-called piecewise constant polarization approximation)


N
τ (x) = mi (x)τ i (61)
i=1

is usually used. Since φi = mi (x) denotes the volume fraction of the phase
i, and given the fact that the average of a tensor T over phase i is given by,
Ti = (mi /φi )T, it follows from (59) and (60) that

1

N
i = E − ij : τ j , (62)
φi
j =1

where
5 6
2 3
ij = mi (x) mj (x ) − φj  0 (x, x )dx , i, j = 1, · · · , N (63)


are tensors that depend only on the microstructure of the material and L0 , and ij
are symmetric (Kohn and Milton, 1986) in i and j and are not all independent,
since they satisfy the relations


N

N
ij = ij = 0.
i=1 j =1

After some algebra, one obtains

⎧
⎨

N
He (E) ≤ inf w 0 (E) + τ  : E − φi (wi − w 0 )∗ (τ i )
τ l , l=1,...,N ⎩
i=1
⎫
1


N N ⎬
− τ i : ij : τ j , (64)
2 ⎭
i=1 j =1

where τ  = N i=1 φi τ i . Then, optimizing over τ i (with i = 1, . . . , N), one
obtains the governing equations for the τ i :

1

N
∂
(wi − w )∗ (τ i ) +
0
ij : τ j = E, i = 1, . . . , N, (65)
∂τ i φi
j =1

so that, from Eqs. (46), (64), and (65) one finally obtains [by replacing the
inequality in (64) by an equality] an approximate stress-strain relation:
 =  0 : E + τ , (66)
where the τ i are obtained from Eqs. (65).
The upper bound (64) for He (E), which was first given by Ponte Castañeda
and Willis (1988), can be written in an alternative form (Willis, 1991) by noting,
 4.4. Variational Principles Based on a Linear Comparison Material 173

through the use of (62), that the optimality conditions (65) can be rewritten in the
form
∂  
i = wi − w 0 (τ i )
∂τ i ∗

which, when inverted, yield

∂  
τi = wi − w 0 (i ), (67)
∂i ∗

so that the (Legendre) dual variables i satisfy the conditions

1

N
∂
i + ij : (wj − w 0 )∗ () = E, (68)
φi ∂j
j =1

for i = 1, · · · , N. Then, the bound (64) may be rewritten as

N
He (E) ≥ w0 (E) + φi [2τ i : (E − i ) + (wi − w 0 )∗∗ (i )], (69)
i=1

where τ i are given in terms of i by Eq. (67).

An upper bound for He is obtained from (69) for any choice of w 0 ; the sharpest
bound is obtained by minimizing over L0 . The resulting bound is finite only if (w −
w 0 ) has weaker-than-affine growth at infinity, which would be the case for, for
example, power-law materials. The minimization with respect to L0 is complicated
by the fact that computation of (wi − w 0 )∗∗ is difficult. Ponte Castañeda and
Willis (1988) and Willis (1989a, b) obtained non-optimal bounds by utilizing
values of L0 for which (wi − w 0 )∗∗ = (wi − w 0 ). Willis (1991,1992) then showed
that improved bounds, agreeing with those of the variational procedure of Ponte
Castañeda (1991a), are obtained by eliminating this unnecessary restriction.
The estimate for He provided by Eqs. (64) and (65), or (68) and (69), after
optimizing over the choice of L0 , is explicit except for the microstructural param-
eters ij , which must be determined separately for each class of morphologies.
Explicit expressions for these parameters were derived by Willis (1977,1978) and
Ponte Castañeda and Willis (1995) for various classes of disordered morphologies
with prescribed two-point correlation functions for the distribution of the phases,
including particulate and granular materials (see below).

4.4 Variational Principles Based on a Linear

Comparison Material
Variational methods for deriving improved bounds and estimates for the effective
properties of nonlinear materials, utilizing the effective modulus tensor of suitably
selected linear-elastic comparison materials, were introduced by Ponte Castañeda
174 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

(1991a) for materials with isotropic phases and by Suquet (1993a) for compos-
ites with power-law phases. Moreover, a hybrid of the Talbot–Willis and Ponte
Castañeda procedures, using a linear thermoelastic comparison material, was pro-
posed by Talbot and Willis (1992). These procedures can in fact be shown to be
equivalent under appropriate hypotheses on the local potentials. An important ad-
vantage of the variational procedures that involve linear comparison materials is
that, they can not only produce the nonlinear Hashin–Shtrikman-type bounds of the
Talbot–Willis procedure directly from the corresponding linear Hashin–Shtrikman
bounds, but also yield higher-order nonlinear bounds, such as Beran-type bounds,
as well as other types of estimates. The application of this technique to deriving
bounds and estimates for the effective nonlinear conductivity and dielectric con-
stant of materials was described and discussed in Chapter 2. We now describe the
analogous results for the effective nonlinear mechanical properties of materials.

4.4.1 Materials with Isotropic Phases

The potential w of a nonlinear material with isotropic phases is written as
9
w(x, ) = 2
K(x)m + f (x, eq
2
),
2
where


N

N
K(x) = mi (x)Ki , 2
f (x, eq ) = 2
mi (x)fi (eq ), (70)
i=1 i=1

with the functions fi , characterizing the deviatoric behavior of the material (see
Chapter 7 of Volume I), being defined by the relations, fi (p) = ϕi (eq ) for p =
2 . The functions f are assumed to be concave functions of p, such that f (p) =
eq i i
−∞ for p < 0, fi (0) = 0, and fi → ∞ as p → ∞. By definition, the concave
dual function of fi is given by
fi∗ (q) = inf {pq − fi (p)} = inf {pq − fi (p)} .
p p>0

It then follows from the concavity hypothesis that

0 1 0 1
fi (p) = inf pq − fi∗ (q) = inf pq − fi∗ (q) . (71)
q q>0

Note that the above hypotheses on fi are consistent with weaker-than-quadratic

growth for wi at infinity, in agreement with the physical requirements for plasticity
and creep. For example, for power-law materials characterized by Eq. (10), ϕi ∼
1+m (0 ≤ m ≤ 1), so that f ∼ p (1+m)/2 is a concave function in the interval
eq i
[0, ∞], even if ϕi is itself convex.
We now introduce a linear comparison material with potential w 0 , such that
9 3
w0 (x, ) = 2
K(x)m + µ0 (x)eq
2
. (72)
2 2
 4.4. Variational Principles Based on a Linear Comparison Material 175

Then, using (71) with q = 3µ0 /2, one finds that the potential of the nonlinear
material w is given by the exact equation,
7 8
w(x, ) = inf w 0 (x, ) + v(x, µ0 ) , (73)
µ0 (x)>0

where


N
3
v(x, µ0 ) = mi (x)vi [µ0 (x)], with vi (µ0 ) = −fi∗ ( µ0 ), (74)
2
i=1

Note that (see also Chapter 2)

7 8
vi (µ0 ) = sup wi () − wi0 () , (75)

so that
7 8
v(x, µ0 ) = sup wi (x, ) − w 0 (x, ) . (76)

If one substitutes Eq. (73) into (44) for the effective potential He , one obtains,
 7 8
He (E) = inf inf w [x, (v)] − v(x, µ ) ,
0 0
v∈S1 (E) µ0 (x)

from which one obtains, by interchanging the order of the infima over  and µ0 ,
7 8
He (E) = inf He0 (E) + V (µ0 ) , (77)
µ0 (x)

where V (µ0 ) = v[x, µ0 (x)], and He0 is the effective potential of the linear
comparison material (Ponte Castañeda, 1992a):

He0 (E) = inf w 0 [x, (v)]. (78)

v∈S1 (E)

It must be emphasized that, under the concavity hypothesis on fi , the variational

representation (77) and the usual representation (44) are exactly equivalent.
One can also start from the complementary energy representation (47) for He∗
to derive a corresponding dual version of the variational representation (77). In
this case
7 8
He∗ () = sup (He∗ )0 () − V (µ0 ) , (79)
µ0 (x)

where

(He∗ )0 () = inf u0 (x, σ ) (80)

σ ∈S2 ( )
is the effective stress potential of the linear comparison material.
176 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

4.4.2 Strongly Nonlinear Materials

We now consider similar variational principles for strongly nonlinear materials
characterized by power-law constitutive equations (Suquet, 1993a). Such materi-
als consist of power-law phases with strain-energy functions (10) with the same
exponent n and the reference strain  0 , but with different flow stresses σ 0 . For
such materials, the variational representation of the effective strain potentials is
given by
 
1 1
He (E) = inf σ (x)eq [v(x)] .
0 m+1
v∈S1 (E) m + 1 ( 0 )m

One can then show that (Ponte Castañeda and Suquet, 1998)
He (E) =
⎧ ,$ -(1−m)/2 ⎫
⎨ %(m+1)/(m−1) ⎬
1 1 3
inf H 0
(E) (m+1)/2
µ 0
(σ 0 2/(1−m)
) ,
m + 1 ( 0 )m µ0 (x)>0 ⎩ e
2 ⎭

(81)
and that
⎧ , -(1−n)/2 ⎫
0 ⎨ (σ 0 )2n/(n−1) ⎬
He∗ () = sup (He∗ )0 ()(n+1)/2 .
n+1 µ0 (x)>0 ⎩ 6(µ0 )(n+1)/(n−1) ⎭

(82)

4.4.3 Materials with Anisotropic Phases

To derive analogous results for nonlinear composite materials with anisotropic
phases, we assume that the functions Fi , which define the strain potentials wi via
Eq. (20), are concave on the space of positive, symmetric fourth-rank tensors P,
i.e., they satisfy
Fi [tP1 + (1 − t)P2 ] ≥ tFi (P1 ) + (1 − t)Fi (P2 ), ∀ P1 and P2 , 0 ≤ t ≤ 1,
(83)
which implies weaker-than-quadratic growth for the potentials wi on the strain ,
when P is set equal to E, as defined by (21). The concave dual function of Fi is
defined by
Fi∗ (L) = inf {L :: P − Fi (P)} ,
P
based on which one defines F (x, P) as


N
F (x, P) = mi (x)Fi (P).
i=1
Because of definition of E by Eq. (21), one has
7 8
F {x, E[v(x)]} = inf L0 (x) :: E[v(x)] − F ∗ [x, L0 (x)] .
L0 (x)
 4.4. Variational Principles Based on a Linear Comparison Material 177

Therefore, from definition (44) one obtains

7 8
He (E) = inf F [E(v)] = inf inf L0 :: E(v) − F ∗ [x, L0 (x)] .
v∈S1 (E) v∈S1 (E) L0 (x)>0
(84)
Then, introducing a linear comparison material with a local potential,
1
w 0 [x, (v)] = L0 :: E(v) = (v) : L0 (x) : (v), (85)
2
and interchanging the infima in (84), one obtains the following exact variational
representation for the effective potential,
7 8
He (E) = inf He0 (E) + V (L0 ) , (86)
L0 >0

where He0 is the effective potential of the linear comparison material defined by
the local potential (85), and V (L0 ) = v[x, L0 (x)], given by

v[x, L0 (x)] = −F ∗ [x, L0 (x)] = sup[F (x, P) − L0 (x) :: P]. (87)

P

Equation (86) expresses the nonlinear effective properties of the material in terms
of two functions which are, (1) He0 , the elastic energy of a fictitious linear het-
erogeneous solid, called the linear comparison material, that consists of phases
with stiffness L0 (x), and (2) v(x, ·), the role of which is to measure the difference
between the non-quadratic potential w(x, ·) and the quadratic energy of the linear
comparison solid. The linear comparison solid is selected from amongst all the
possible comparison materials by solving the optimization problem (86).
Equation (86), which is exact, is strictly equivalent to the variational represen-
tation of He given by (44). However, determining the exact solution of (86) is not
possible. The difficulty lies in the precise determination of the energy He0 for a
linear comparison solid consisting of infinitely many different phases, about which
very little is known. For this reason, except for very simple microstructures, such
as laminates considered in Section 2.3, the optimal solution of (86) is not known,
and only sub-optimal solutions can be determined. We now consider application
of these principles to a few classes of materials.

4.4.3.1 Polycrystalline Materials

If the individual phases of a material are single crystals, the functions Gi [see
Eqs. (24) and (25)] are given by

M  
(k)
Gi (S) = g(k) 2Mi :: S ,
k=1

where, as before
(k) (k) (k)
Mi = µi ⊗ µi .
178 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

It can then be shown that

M 

  
(k) ∗
G∗i (Mi ) = inf
(k)
gi αi ,
(k)
αi >0 k=1

with
 N (k) (k) (k)
(k) 2 k=1 αi Mi , αi > 0,
Mi = (88)
+∞, otherwise.
Therefore, the corresponding local stress potential for the linear comparison
polycrystalline material is given by


M . . . .
1 (k) . (k) .2 . (k) .2
u0i (σ ) = σ : Mi : σ = αi .τi . , .τi . = 2Mi :: σ . (89)
2
k=1

Hence, for polycrystalline materials one obtains

7 8
He∗ () = sup (He∗ )0 () − V [αi ] , i = 1, · · · , N, k = 1, · · · , M (90)
(k)
(k)
αi >0

where (He∗ )0 is the effective potential of the linear comparison polycrystalline

material with grain potentials (89), and


N

M  
φi [gi ]∗ αi
(k) (k) (k)
V [αi ] = ,
i
i=1 k=1

which was first derived by deBotton and Ponte Castañeda (1995). The functions
(k)
αi (x), k = 1, . . . , M, are defined over the region in space that is occupied by
the crystals with fixed orientation i.

4.4.3.2 Strongly Nonlinear Materials

If the individual phases of a composite are power-law materials with the same
exponent m (with 0 ≤ m ≤ 1), the composite itself is also a power-law material.
That is, the local potentials and the effective macroscopic potential are given by
wi (λ) = λm+1 wi (), He (λE) = λm+1 He (E), ∀λ ≥ 0,
i.e., they are homogeneous function of order m + 1. The function F that defines the
strain potential w is itself a power-law function of degree 12 (m + 1), and its dual
is a power-law function of degree (m + 1)/(m − 1). If we let L0 (x) = t L̂0 (x) for
any t > 0, and note that He0 and V = −F ∗  > 0 are homogeneous functions of
orders 1 and (m + 1)/(m − 1) in L0 , respectively, it follows from the variational
statement (86) that
7 8
He (E) = inf inf t Ĥe0 (E) + t (m+1)/(m−1) V (L̂0 ) .
L̂0 >0 t>0
 4.4. Variational Principles Based on a Linear Comparison Material 179

where Ĥe0 is the same as He0 in (86), but with L0 replaced by L̂0 . Evaluating the
minimum over t yields an exact representation for He :
  (1−m)/2 
2 (m+1)/2 1 + m
He (E) = inf He (E)
0 0
V (L ) , (91)
m + 1 L̂0 >0 1−m

where the hat notation has been deleted for simplicity. The analogous representa-
tion for He∗ is given by
  (1−n)/2 
2 (n+1)/2 n + 1
He∗ () = ∗ 0
sup (He ) () 0
V (L ) . (92)
n + 1 L0 >0 n−1

4.4.3.3 Materials with Isotropic and Strongly Nonlinear Phases

In this case, it is sufficient to consider isotropic linear comparison materials. If
such materials are governed by Eq. (10), then, the functions f , f ∗ , g and g ∗ are
given by
   2/(1−m)
σ 00 |x| (m+1)/2 ∗ m−1 σ0
f (x) = , f (y) = |y|(m+1)/(m−1),
m + 1 ( 0 )2 m + 1 2( 0 )m
   2/(n−1)
σ 00 |x| (n+1)/2 ∗ n − 1 2(σ 0 )n
g(x) = , g (y) = |y|(n+1)/(n−1),
n + 1 (σ 0 )2 n+1 0
(93)

and
5 $ %6
3 0
V (L0 ) = − f ∗ µ
2
 2/(1−m) ,$ % -
1−m 1 3 0 (m+1)/(m−1) 0 2/(1−m)
= µ (σ ) .
1 + m 2( 0 )m 2

4.4.3.4 Strongly Nonlinear Polycrystalline Materials

The corresponding result for power-law polycrystalline materials with potentials
(16) is obtained directly from Eq. (92), with the result being

He∗ () =
⎧ N M , ⎫
⎨

  2n/(n−1) - (1−n)/2 ⎬
γ0 (k)
sup (H∗ )0 ()(n+1)/2
(k)
φi [αi ](n+1)/(n−1) τ 0 ,
n + 1 α (k) >0 ⎩ e i ⎭
i i=1 k=1 i
(94)
for i = 1, · · · , N and k = 1, · · · , M.
180 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

4.4.3.5 Ideally Plastic Materials

In the ideally plastic limit, m → 0, the variational representations (91) and (92)
reduce to
7 81/2
He (E) = 2 inf He0 (E)V (L0 ) , (95)
L0
and

0 if (He∗ )0 (E) ≤ V (L0 ) ∀L0 = (M0 )−1 > 0,
He∗ () = (96)
+∞ otherwise.

4.5 Bounds with Piecewise Constant Elastic Moduli

The exact computation of He requires the determination of the effective potential
of a linear material with infinitely many different phases, an extremely difficult
problem, which may be simplified by restricting the optimization over L(x) to the
set of piecewise constant moduli,


N
L0 (x) = mi (x)L0i , (97)
i=1

where the tensors L0i are assumed constant. In this manner an upper bound for He ,
given by
 


N
He (E) ≤ inf He0 (E) + φi vi (L0i ) , (98)
L0i >0, i=1,···,n i=1

is obtained in which He0

is the effective potential [see Eq. (78)] of a linear com-
posite with the same microstructure as the nonlinear material with the domains
i occupied by linear phases with stiffness L0i . The comparison material has an
effective stiffness L0e , such that
1
He0 (E) = E : L0e : E, (99)
2
and the functions vi are defined by
vi (L0i ) = −Fi∗ (L0i ). (100)
The bound (98) is a generalization for materials with anisotropic phases of a
corresponding bound for composites with isotropic phases, introduced by Ponte
Castañeda (1991a).
A bound equivalent to (98) can be derived by considering the stress poten-
tial He∗ and its variational representation. Thus, utilizing the piecewise constant
compliances M0i , one obtains
 

N  
∗ ∗ 0 0 −1
He () ≥ sup (He ) () − φi vi (Mi ) , (101)
M0i >0, i=1,···,N i=1
 4.5. Bounds with Piecewise Constant Elastic Moduli 181

with (He∗ )0 now being the effective stress potential associated with the same linear
comparison material as for He0 given above, i.e., one with the same microstructure
as the nonlinear material, but with the domains i occupied by linear phases with
compliances M0i . From Eq. (101) one obtains
∂He∗
E= () = M0e (M0i ) : . (102)
∂
It also follows from (100) that
N 


N
∗ ∗
He () ≥ inf φi Gi (Si ) = φi Gi (S)i , (103)
τ ∈S2 ( ) i=1 i=1

where Si∗ = Si = 12 σ ⊗ σ i is the second moment of the stress field in phase i
of the linear comparison material. The compliances M0i of the comparison mate-
rial are determined as the solution of the optimization problem (100), which can
alternatively be written in terms of the solution of the following nonlinear problem
for the variables Si∗ :
∂Gi ∗ 1 ∂M0e
M0i = (S ), Si∗ = : : . (104)
∂S i 2φi ∂M0i

4.5.1 Materials with Isotropic Phases

If the nonlinear phases are isotropic, then the constituent phases of the linear com-
parison material can also be selected to be isotropic. The effective bulk modulus
is then equal to the bulk modulus Ki of the nonlinear constituent phase i, and
therefore the only modulus that must be determined is the shear modulus µ0i of
each phase. Thus, the bound (98) reduces to
 
1

N
He () ≥ inf E : Le (µi ) : E +
0 0 0
φi vi (µi ) , (105)
µ0i >0, i=1,···,N 2 i=1

where the functions vi are defined by Eq. (75). The upper bound (105), as well as
the analogous lower bound,
 
1

N
∗
He () ≥ sup  : Me (µi ) :  −
0 0 0
φi vi (µi ) , (106)
µ0 >0, i=1,···,N 2
i i=1

were first derived by Ponte Castañeda (1991a). Based on the associated optimality
conditions (deBotton and Ponte Castañeda, 1992,1993), it can be shown that

9

N  
N
(m) 2 eq 2
He ≤ φ i K i i + φi ϕi i , (107)
2
i=1 i=1

φi  (m) 2

N N
1 eq 2
He∗ ≥ σ + φi ψi σi , (108)
2 Ki i
i=1 i=1
182 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

where
 1/2  1/2
(m) 1 ∂L0e (m) 1 ∂M0e
i = E: :E , σi = : : , (109)
9φi ∂Ki0 φi ∂(1/Ki0 )
and
 1/2  1/2
eq 1 ∂L0e eq 3 ∂M0e
i = E: :E , σi = : : . (110)
3φi ∂µ0i φi ∂(1/µ0i )
These simplified bounds were first given by Suquet (1995,1997).
For power-law materials, one obtains,
⎧ N (1−m)/2 ⎫
1 1 ⎨
$ 3 %(m+1)/(m−1) ⎬
He (E) ≤ inf H 0
(E)(m+1)/2
φ i µ0
(σ 0 2/(1−m)
) ,
m + 1 ( 0 )m µ0i >0 ⎩ e
2 i i ⎭
i=1
(111)
and
⎧ N (1−n)/2 ⎫
0 ⎨
⎬
He∗ () ≥ ∗ 0
sup (He ) ()(n+1)/2 0 (n+1)/(1−n) 0 2n/(n−1)
φi (6µi ) (σi ) ,
n + 1 µ0 >0 ⎩ ⎭
i i=1
(112)
which were also derived by Suquet (1993a).

4.5.2 Polycrystalline Materials

In this case, one restricts the optimization in (101) to compliance tensors that
yield finite values for the functions vi , which then leads to (deBotton and Ponte
Castañeda, 1995)
 

N
M    
∗ 1 (k) ∗ (k)
He () ≥ sup  : Me (α) :  −
0
φi g i αi , (113)
(k)
α >0
2
i i=1 k=1

where the suprema should be performed for i = 1, · · · , N and k = 1, · · · , M.

(k)
Here, α denotes the entire set of positive slip compliances αi , and Me is the
effective compliance tensor of the linear comparison polycrystalline material with
grain compliances Mi , as given by Eq. (88) in terms of the slip compliances α.
One may also approximate the effective stress-strain relation of the polycrystalline
material by the relation (102) of the linear comparison material, with the optimal
(k)
M0i replaced by the optimal αi . In that case, the nonlinear optimality relations
are given by

∂gi  
M (k)
(k) (k) (k) (k)
Mi = 2 αi Mi , αi = 2Mi :: σ̄i ,
∂τ
k=1
1 ∂Me
σ̄i = : : . (114)
2φi ∂Mi
 4.5. Bounds with Piecewise Constant Elastic Moduli 183

These nonlinear equations can be expressed more explicitly in terms of the slip
(k)
compliances αi and the corresponding second moment of the resolved shears,
 1/2
(k) (k)
τ̄i = 2Mi :: σ̄i .

with the result being

∂gi  (k) 2 1 ∂ψi  (k) ∗

(k) (k)
(k)
αi = τ̄i = (k) τ̄i ,
∂τ 2τ̄i ∂τ
 1/2
(k) 1 ∂Me
τ̄i = : (k)
(α) :  , (115)
2φi ∂α i

in terms of which the bound is rewritten as

N

M  
He∗ ()
(k) (k)
≥ φi ψ i τ̄i , (116)
i=1 k=1

For power-law polycrystalline materials, one obtains the following result,

⎧ N M  ⎫
γ 0 ⎨

 (k) (n+1)/(n−1)  2n/(n−1) (1−n)/2 ⎬
He∗ ≥ sup (H∗ )0 ()(n+1)/2
(k)
φi αi (τ 0 )i ,
n + 1 α (k) >0 ⎩ e ⎭
i i=1 k=1
(117)
where the suprema must be carried out over i = 1, · · · , N and k = 1, · · · , M.
It should be emphasized that any estimate for the effective modulus tensor of
a linear elastic material can be used to generate, by the variational procedures
described above, a corresponding estimate for a nonlinear material with the same
microstructure. This is in contrast to several other schemes which are closely
connected with specific types of estimates. For example, the Talbot–Willis method
described above provides only estimates of the Hashin–Shtrikman-type. Moreover,
similar to the case of the effective conductivity and dielectric constant of nonlinear
materials discussed in Chapter 2, if the estimate for the effective modulus tensor
of the linear elastic material is an upper bound to Le , then an upper bound is
obtained for He . If, on the other hand, the linear estimate is a lower bound, then,
the variational method cannot, in general, be used for deriving a lower bound
for the nonlinear material. However, if accurate estimates (but not necessarily
bounds) are available for a specific type of linear material, such as those provided
by the effective-medium approximation, then the above variational methods can
be utilized for generating the corresponding estimates for a nonlinear material with
the same microstructure. The resulting estimates for He would tend to err on the
high side, because of the nature of the approximations intrinsic to the variational
method. In addition, these variational methods can be used for deriving higher-
order (≥ 2) bounds, such as Beran-type bounds (Ponte Castañeda, 1992a), as well
as other types of estimates, such as the generalized self-consistent estimates of
Suquet (1993b). Let us also mention that Smyshlyaev and Fleck (1995; see also
184 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

Fleck and Hutchinson, 1997) proposed extensions of the variational methods in

the context of strain gradient plasticity.
We remind the reader that the above variational methods use the concavity
hypothesis on the function F associated with the local strain potential w. Except
for some pathological cases, this mild hypothesis is satisfied by the standard models
of plasticity and creep (Willis, 1992; Ponte Castañeda and Willis, 1993). When
this hypothesis is satisfied, Ponte Castañeda (1992c) showed, in the context of
materials with isotropic phases, that the Talbot–Willis variational method (see
Section 4.3) can be directly derived, via variational principles, from the Hashin–
Shtrikman variational principles for linear materials.An alternative, simpler way of
analyzing materials for which the concavity hypothesis is violated was proposed by
Ponte Castañeda (1996b, 1997); see also Kohn and Little (1997) and Bhattacharya
and Kohn (1997) in the context of polycrystalline materials.

4.6 Second-Order Exact Results

We now describe and discuss exact results for the effective mechanical properties of
weakly heterogeneous nonlinear materials, and also estimates for arbitrary contrast
of the phases. This analysis represents an extension of a similar theory for linear
elasticity, for which it is well-known that the effective moduli tensor of a weakly
heterogeneous material can be determined exactly to second order in the contrast
(see Chapter 7 of Volume I). The present theory also represents an extension of the
analogous theoretical developments for the effective nonlinear conductivity and
dielectric constant of heterogeneous materials that were described and discussed
in Section 2.6. The analysis that follows also establishes that the above variational
estimates are exact only to first order in the phase contrast, when estimates that are
exact to second order are used to evaluate the mechanical properties of the linear
comparison material.

4.6.1 Weak-Contrast Expansion

It is assumed that the contrast between the properties of the phases is small. To
incorporate this assumption into the analysis, the potential w is assumed to depend
on a small parameter t that characterizes the contrast between the properties of
the material and those of a homogeneous nonlinear reference material with energy
function w0 (), such that
w(x, , t) = w 0 () + tδw(x, ). (118)
The effective potential also depends on the parameter t:
He (E, t) = w[x, (ut ), t], (119)
where ut and (ut ) are the local displacement and the associated strain fields
induced by appropriate boundary conditions that generate an average strain E in
. Furthermore, it is assumed that He (., t) and ut are continuously differentiable
 4.6. Second-Order Exact Results 185

functions of t. Since t is small, one can write down a perturbation series expansion
of He about t = 0, given formally by
∂He 1 ∂ 2 He
He (E, t) = He (E, 0) + t (E, 0) + t 2 (E, 0) + O(t 3 ). (120)
∂t 2 ∂t 2
The problem to be solved for ut is given by
 
∂w
∇· [x, (ut ), t] = 0, ut ∈ S1 (E). (121)
∂
If we differentiate (121), we find that u̇t = ∂ut /∂t is the solution of the following
system of equations
∇ · [Lt : (u̇t )] + ∇ · τ t = 0, u̇t ∈ S1 (0). (122)
where
∂ 2w ∂ 2w ∂
Lt = [x, (ut ), t], τ t = [x, (ut ), t] = (δw)[x, (ut ), t].
∂∂ ∂t∂ ∂
Therefore,
5 6 5 6
∂He ∂w ∂w
(E, t) = [x, (ut ), t] : (u̇t ) + [x, (ut ), t] . (123)
∂t ∂ ∂t
The first term of (123) vanishes due to Eq. (39) (the so-called Hill’s lemma), and
therefore,
∂He
(E, t) = δw[x, (ut )]. (124)
∂t
Using Eq. (121), one obtains
5 6
∂ 2 He ∂
(E, t) = (δw)[x, (ut )] : (ut ) = − (u̇t ) : Lt : (u̇t ) . (125)
∂t 2 ∂
Because the material is homogeneous for t = 0, u0 = E · x, and therefore
He (E, 0) = w 0 (E),
∂He
(E, 0) = δw(E), (126)
∂t
∂ 2 He
(E, 0) = −(u̇0 ) : L0 : (u̇0 ),
∂t 2
where
∂ 2 w0
L0 = (E). (127)
∂∂
Here, u̇0 is the solution of the linear elasticity problem,
∇ · [L0 : (u̇0 )] + ∇ · τ = 0, u̇0 ∈ S1 (0), (128)
with
∂
τ (x) = (δw)(x, E).
∂
186 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

Since the modulus tensor L0 is constant, the problem posed by (128) is a linear
elasticity problem for a homogeneous material with a distribution of body forces
determined by τ .
If the material consists of N homogeneous phases, then τ is piecewise constant,
i.e., it is a constant in each phase, with


N
∂
τ (x) = mi (x) τ i , τi = (δwi )(E),
∂
i=1

in terms of which one has

∂ 2 He

N N

2
(E, 0) = − τ i : ij : τ j , (129)
∂t
i=1 j =1

where the microstructural tensors ij are defined by Eq. (63). Therefore (Suquet
and Ponte Castañeda, 1993),

1 2


N N
He (E, t) = w(E) − t τ i : ij : τ j + O(t 3 ). (130)
2
i=1 j =1

As an example, consider a material with N isotropic phases, with strain potentials

defined by Eq. (118) and
9 0 2
w0 () = K m + f 0 (eq
2
),
2
9
δwi () = δKi m 2
+ δfi (eq
2
).
2
Then, it is straightforward to show that (Ponte Castañeda and Suquet, 1995),


N  
9 9
He (E, t) = K 0 Em 2
+ f 0 (Eeq
2
)+t φi 2
δKi Em + δfi (Eeq
2
)
2 2
i=1
N 
N



1
− t2 9δKi δKj U : ij : UEm
2
+ 4δµi δµj Ed : ij : Ed + O(t 3 ).
2
i=1 j =1
(131)
where Ed = E − Em U is the average strain deviator.

4.6.2 Strong-Contrast Expansion

Another method for estimating the effective mechanical properties of nonlinear
materials was proposed by Ponte Castañeda (1996a). His method uses a linear
heterogeneous comparison material and the associated tangent modulus tensors
of the constituent phases. This choice of comparison material ensures that the
resulting nonlinear estimates are exact to second order in the contrast, and thus are
in agreement with the small-contrast asymptotic results of the last section.
 4.6. Second-Order Exact Results 187

Similar to the case of the effective conductivity and dielectric constant of nonlin-
ear heterogeneous materials that was discussed in Chapter 2, this method is based
on a Taylor expansion for the phase potentials wi . Thus, introducing reference
strains E(i) , the Taylor expansion for wi about E(i) is given by
1
wi () = wi [E(i) ] + ρ i : [ − E(i) ] + [ − E(i) ] : Li : [ − E(i) ], (132)
2
where ρ i and Li are, respectively, an internal stress and a tangent modulus tensor,
with components
∂wk (k) ∂ 2 wm
(ρk )ij = [E ], (Lm )ij kl = [Ẽ(m) ]. (133)
ij ∂ij ∂kl
Li depends on the strain Ẽ(i) = λ(i) E(i) + [1 − λ(i) ], where λ(i) depends on 
and is such that 0 < λ(i) < 1.
In terms of the average E and fluctuating   components of  = E +   , Eq. (132)
is rewritten as
1
wi (E +   ) = νi + τ i :   +   : Li :   , (134)
2
where
1
νi = wi [E(i) ] + ρ i : [E − E(i) ] + [E − E(i) ] : Li : [E − E(i) ], (135)
2
τ i = ρ i + Li : [E − E(i) ]. (136)
Making the approximation that the strains Ẽ(i)
are constant in each phase, the
effective potential He of the nonlinear material is then estimated as


N
He (E)  H̃e (E) = φ i νi + P , (137)
i=1
where
5 6
1
P = 
inf (v ) : L : (v ) + τ : (v ) . (138)
v ∈S1 (0) 2
τ and L(x) are defined by equations similar to (61). The advantage of approxi-
mation (137), relative to the exact result (44), is that it requires only the solution
of a linear problem for an N -phase thermoelastic material, as defined by the
Euler–Lagrange equations of the variational problem P in (138):
∇ · [L : (u )] = −∇ · τ , u ∈ S1 (0). (139)
Estimates for N-phase linear-thermoelastic materials can, in general, be obtained
by appropriate extension of the corresponding methods for N -phase linear-elastic
composites (see, for example, Willis, 1981). Similar, but not equivalent, repre-
sentations for the effective mechanical properties of nonlinear materials, which
also utilize heterogeneous thermoelastic reference materials, were proposed by
Molinari et al. (1987) and Talbot and Willis (1992).
188 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

Equation (137) provides an estimate for He for any choice of E(i) and Ẽ(i) , if
we supply it with an estimate for P . A plausible approximation for the E(i) is to set
them equal to the averages of the strain field over the phases i. However, because
the exact strain field is not known, the approximate field , as determined by (139),
is used, so that
E(i) = i , (140)
where, i = E +   i . Equation (140) is a reasonable choice because the strain
 in phase i is expected to fluctuate about its average in phase i in such a way that
large deviations would only be expected in regions of relatively small measure.
The following identity, obtained from Eq. (139),
1 ∂P
  i = , (141)
φi ∂τ i
which can be used to obtain i directly from P , via
1 ∂P
i = E + , (142)
φi ∂τ i
is also useful, since the reference strains E(i) may also be computed from P by
means of Eqs. (140) and (142). It can also be shown that Eq. (140) provides the
optimal choice for E(i) in the sense that, estimate (137) for He is stationary with
respect to the E(i) .
One important consequence of stationarity of Eq. (140) is that the overall stress-
strain relation (46) for the material may be approximated as
 

N
1 ∂ Ẽ(i)
= φi ρ i + ( − i ) : Ni : ( − i )i : , (143)
2 ∂E
i=1

where i are determined by Eq. (142), and

∂ 3 wm
(Nm )ij klpq = [Ẽ(m) ],
∂ij ∂kl ∂pq
which can be derived by taking the derivative of Eq. (137) with respect to E, with
E(i) held fixed (because of stationarity), and enforcing (140).
Equation (140) also allows simplification of estimate (137) for He . Note that
the Euler–Lagrange equations, Eqs. (139), of problem (138) for P imply that
τ : (u ) = −(u ) : L : (u ), (144)
which, together with Eq. (140) and the definition (136) of τ i , are used to rewrite
the estimate (137) in the following simpler form,

N  
1 ∂wi
He  H̃e (E) = φi wi i + (i ) : (E − i ) , (145)
2 ∂
i=1

with i being determined by Eqs. (142).

 4.6. Second-Order Exact Results 189

The choice of Ẽ(i) in definition (133) of Li is not as straightforward, and, in

particular, stationarity of H̃e with respect to Ẽ(i) cannot be implemented. For this
reason, Ponte Castañeda (1996a) proposed the following physically motivated
equation for Ẽ(i) :
Ẽ(i) = i = E(i) , (146)
an interesting consequence of which is that it implies that
∂ 2 H̃e
= 0. (147)
∂E(i) ∂E(i)
It is now not difficult to show that


N
1
H̃e (E) = φi wi (E) − t 2 (u̇0 ) : L0 : (u̇0 ) + O(t 3 ), (148)
2
i=1
in agreement with the small-contrast expansion (120) together with (126).
As an example, consider two-phase materials, for which a well-known result
due to Levin (1967) allows further simplification of the thermoelastic problem P ,
and hence of the corresponding estimate for He . The result for P , which depends
only on the effective modulus tensor Le of a two-phase, linear elastic material with
phase modulus tensors L1 and L2 , is given by
1
P =(τ ) : (L)−1 : (Le − L) : (L)−1 : (τ ), (149)
2
where L = L1 − L2 and τ = τ 1 − τ 2 . It then follows from Eqs. (142) and
(146) that
E(i) = Ẽ(i) = i = E + (Ai − U) : (L)−1 : (τ ), (150)
where Ai denote the strain-concentration tensors (Hill, 1965a) for the linear elastic
material problem, such that
φ1 A1 + φ2 A2 = U, Le = φ1 L1 : A1 + φ2 L2 : A2 , (151)
which can be solved for the tensors Ai in terms of the Li and Le .
It must be emphasized that any estimate of any type for Le can be used for
generating the corresponding estimates for He , that the second-order term in the
above expansion depends only on the two-point statistics of the material and com-
pletely specifies its effective properties (to second order in the contrast), and that
by comparison with this exact result, it becomes clear that the variational estimates
described above are exact only to first order in the contrast. In addition, the second-
order theory does produce estimates that are exact to second order in the contrast.
However, the approximations involved in the second-order theory are such that it
is not possible to control the sign of the error, so that the resulting estimates, unlike
the earlier variational estimates, cannot be guaranteed to be bounds to the effective
properties. Another important limitation of the second-order theory is the existence
of a duality gap, i.e., it can be shown that, H̃e∗
= (H̃e )∗ . As a practical matter, in
plasticity and creep, as in conductivity and dielectric constant (see Chapter 2),
190 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

the second-order estimates based on the estimate He are more accurate than the
analogous estimates for He∗ .

4.7 Applications of Second-Order Exact Results

We now consider some applications of the above theoretical results to modeling of
mechanical properties of porous materials, and composites with a superrigid phase,
as well as more general two-phase power-law and perfectly plastic composites. In
each case, we first describe the various bonds that can be obtained from the above
general formulation, and then discuss the application of the second-order theory.

4.7.1 Porous Materials

We consider porous materials with isotropic matrix phases, so that the strain and
stress potentials of the matrix phase are given by Eqs. (4) and (5). Designating the
matrix as phase 1, the bound (107) becomes
 
9 (m) 2 eq 2
He (E) ≤ φ1 K1 1 + φ1 ϕ1 1 (152)
2
where
  

eq
 1 ∂L0e 0
1 = E: (µ , K1 ) : E, (153)
3φ1 ∂µ01 1
#  0 
(m) 1 ∂Le 0
1 = E: (µ , K1 ) : E. (154)
9φ1 ∂K1 1
eq
Equation (153) must be solved for 1 with
1 ∂ϕ1 eq 
µ01 = eq eq 1 .
31 ∂1
If we utilize any upper bound on, or estimate for, the effective modulus tensor of a
linear porous material with an isotropic matrix, then, the bound (152) would lead
to a corresponding upper bound or estimate for the effective strain potential of the
nonlinear porous material. If the matrix phase is incompressible (K1 → ∞), so
that the effective modulus and compliance tensors of the linear comparison porous
material can be written as
 −1
L0e = µ01 L̂, M0e = µ01 M̂,

where L̂ and M̂ are two microstructural tensors that are independent of µ01 , then,
the estimate (152) for He and the corresponding estimate for He∗ reduce to
eq  eq 
He (E) ≤ φ1 ϕ1 1 , He∗ () ≥ φ1 ψ1 σ1 , (155)
 4.7. Applications of Second-Order Exact Results 191

where
# #
eq 1 eq 3
1 = E : L̂ : E, σ1 =  : M̂ : . (156)
3φ1 φ1

4.7.1.1 Two-Point Bounds

When the distribution of the pore phase is statistically isotropic, the linear Hashin–
Shtrikman bound (see Chapter 7 of Volume I) leads to a corresponding upper
(lower) bound for He (He∗ ) with
#
$ %−1
eq 4 2 2
1 = Em + 1 + φ 2 2 ,
Eeq (157)
φ2 3
# $ %
eq 1 9 2 2
σ1 = + 1 + φ 2
2
eq , (158)
φ2 4 m 3
which was first derived Ponte Castañeda (1991a) and Suquet (1992). It was also
derived as an ad hoc estimate (not a bound) by Qiu and Weng (1992) by estimating
the stress in the matrix from the energy in the porous material. If, on the other hand,
the voids’ shapes and distribution are cylindrical with circular cross section, one
obtains a Hashin–Shtrikman-type bound given by (152) with (Suquet, 1992)
#  
eq 1 3
σ1 = eq + φ2 ( 11 + 22 ) + 3φ2 ( 13 + 23 + 12 ) , (159)
2 2 2 2 2 2
1 − φ2 2
where the axis of symmetry has been taken to be aligned with the x3 direction.
Another important case is when one of the aspect ratios of the voids approaches
zero, leading to cracks, in which case, φ2 → 0. When the cracks are penny shaped,
aligned, and distributed isotropically, one obtains the Hashin–Shtrikman-type
bound (152) with
 $ 
 % $ %
eq
 3ρ 32 ρ −1 2 4 4 ρ −1 2
σ1 = eq +
2 1− 33 + 3 1 − 15 π ( 13 + 23 ) ,
2
π 15 π
(160)
where ρ = 43 π n2 a 3 is the crack density corresponding to n2 cracks of mean radius
a per unit volume. The corresponding results for flat distributions of cracks, i.e.,
when the crack interactions are weak, which are obtained by linearizing (with
respect to α2 ) Eq. (160), were first given by Suquet (1992) and Talbot and Willis
(1992). When the cracks are randomly oriented and distributed isotropically, the
following upper bound is obtained:
#
$ %$ %
eq 3π 2 12 ρ 8 ρ −1 2
1 = E + 1− 1+ Eeq , (161)
ρ m 25 π 25 π
which was derived by Ponte Castañeda and Willis (1995) in the linear context.
192 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

If we use the self-consistent or effective-medium approximation estimate of

Hill (1965b) and Budiansky (1965), described in Chater 7 of Volume I, it follows
(Ponte Castañeda, 1991a) that expression (152) provides an estimate for He with
# $ %$ %
eq 1 1 − 2φ2 4 2
1 = Em + Eeq2 . (162)
φ1 1 − φ2 /3 φ2

4.7.1.2 Three-Point Bounds

One can also obtain third-order, Beran-type bounds for this class of materials. If one
uses the Milton (1982) simplified form of the third-order bounds for linear elastic
materials (see also Chapter 7 of Volume I), it is straightforward to derive a third-
order upper bound for porous materials with statistically isotropic microstructures,
given by (152), with
#
$ %−1
eq 4ζ1 2 2
1 = E + η 1 η1 + φ2 2 ,
Eeq (163)
φ2 m 3
where η1 and ζ1 are two microstructural parameters defined and described in
Chapters 4 and 7 of Volume I. Bounds of this type for nonlinear materials were
first proposed by Ponte Castañeda (1992a, 1997). Note that when ζ1 = η1 = 1,
the bound (152), together with Eq. (163), reduce to the Hashin–Shtrikman upper
bound [together with Eqs. (157) and (158)], but the bound (152) is generally tighter
than the Hashin–Shtrikman bound for ζ1
= 1 and η1
= 1.
One may also utilize the second-order theory of Ponte Castañeda (1996a) in
order to derive certain results for porous materials with an incompressible matrix
and statistically isotropic microstructures (or isotropic distributions of spherical
pores), provided that the pores are also incompressible, so that the material as a
whole is incompressible (Em = 0). This would be the case if the pores are saturated
with an incompressible fluid. Assuming isotropy of the matrix, as characterized
by the function f1 in
9
wi () = 2
Ki m + fi (eq
2
), (164)
2
where fi characterizes the shear modulus of phase i, and letting K1 → ∞, the
second-order estimate (145) for such fluid-saturated porous materials is written as
   
H̃e (E) = φ1 f1 (1 + φ2 ω)2 Eeq
2
− φ1 φ2 ω (1 + φ2 ω) (f1 ) (1 + φ2 ω)2 Eeq
2 2
Eeq ,
(165)
where 1 = (1 + φ2 ω)E is obtained from Eq. (150) for the average strain in
the matrix phase. In a similar manner, a self-consistent or an effective-medium
approximation estimate can also be obtained (Ponte Castañeda and Suquet, 1998).
To see the application of these results, consider, as an example, the Hashin–
Shtrikman-type variational bounds [i.e., using Eqs. (157) and (158) in (155) and
(156)] and second-order estimates (165) for statistically-isotropic porous materials.
The behavior of the incompressible matrix is characterized by the power-law rela-
tion (10), so that for purely deviatoric loading conditions (Em = 0), the effective
 4.7. Applications of Second-Order Exact Results 193

potential He can be written in the form

$ %m+1
σ 0 (θ ) 0 Eeq
He (E) = e , (166)
m+1 0
where θ depends on the determinant of the strain, with θ = 0 corresponding to
axisymmetric deformation and θ = π/6 to simple shear. In general, one finds that
the second-order estimates lie below the variational bounds. Moreover, although
the variational bounds are independent of the type of loading, the corresponding
second-order estimates are different for such cases as uniaxial tension and simple
shear, with the shear results always lying below the tensile results. In addition,
the difference between the shear and tensile results becomes progressively larger,
as the level of nonlinearity increases, with the second-order estimates remaining
close to the variational estimates for tension, but predicting sharper drops in the
load-carrying capacity of the porous material in shear. As first pointed out by
Drucker (1959), the sharper drop for large values of n (tending to perfectly plastic
behavior) is possible under shear loading because of the availability of localized
deformation modes (i.e., slip bands) passing through the pores. There is also ex-
perimental evidence for this type of behavior (Spitzig et al., 1988). On the other
hand, for the axisymmetric deformation mode, the plastic deformation is diffused
through the matrix (Duva and Hutchinson, 1984), and the differences between
the variational and second-order estimates are relatively small (Ponte Castañeda,
1996a). It must, however, be emphasized that the second-order procedure can cap-
ture more accurately the anisotropy of the localized deformation fields by means
of the use of the anisotropic tangent modulus tensors (Ponte Castañeda, 1992a).

4.7.2 Rigidly Reinforced Materials

Let us now discuss composite materials with isotropic nonlinear matrix phases,
reinforced by a rigid phase. The phase strain and stress potentials are assumed to
be given by Eqs. (4) and (5). Designating the matrix as phase 1, Eq. (108) becomes
φ1  (m) 2
He∗ () ≥
eq
σ + φ1 ψ1 (σ1 ), (167)
2K1 1
where
  

eq
3 ∂M0e
σ1 = : (µ0 , K1 ) : , (168)
φ1 ∂(1/µ01 ) 1
#  
(m) 1 ∂M0e
σ1 = : (µ01 , K1 ) : . (169)
φ1 ∂(1/K1 )
eq
Equation (168) must be solved for σ1 with
1 3 ∂ψ1 eq
= eq (σ ).
µ01 σ1 ∂σ eq 1
194 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

Use of any lower bound or any estimate for the effective compliance tensor of a
rigidly-reinforced material with an isotropic matrix then leads to corresponding
lower bounds and estimates for the effective stress potential of the corresponding
nonlinear, rigidly-reinforced composites. When the matrix phase is also incom-
pressible (i.e., when K1 → ∞), the resulting material is also incompressible and
the corresponding estimates for He and He∗ can be written in a form similar to
(155) and (156), with Em = 0, in terms of appropriate microstructural tensors L̂
and M̂ = (L̂)−1 . For example, the Hashin–Shtrikman estimates can be interpreted
as appropriate variational estimates for particulate microstructures, and thus the
corresponding nonlinear results can be thought of as appropriate variational esti-
mates for particulate microstructures. Thus, (155) and (156), with the inequality
replaced by an approximate equality, yield estimates for He and/or He∗ . In particu-
lar, for spherical particles that are distributed with statistically-isotropic symmetry,
the following estimate should be used:
&
eq 1 3
1 = 1 + φ2 Eeq , (170)
φ1 2
This “lower estimate” was proposed by Ponte Castañeda (1991b, 1992a) for
isotropic, rigidly-reinforced composites and generalized by Talbot and Willis
(1992) for anisotropic materials. Talbot and Willis (1992) and Li et al. (1993)
also presented predictions for aligned spheroidal inclusions. Gărăjeu and Suquet
(1997) also discussed an application to rigidly-reinforced materials.

4.7.2.1 Two-Point Bounds

As discussed in Chapter 7 of Volume I, in the case of statistically-isotropic mor-
phologies, the Hashin–Shtrikman upper bounds for linear elastic materials with
arbitrary microstructures are unbounded, and therefore the corresponding upper
bounds for He are also unbounded. Physically, this is due to the fact that statistical
isotropy does not exclude the possibility of formation of a sample-spanning perco-
lation cluster of rigid materials. However, for particulate microstructures (which,
at least for small enough volume fractions of the inclusions, exclude the possibil-
ity of formation of rigid percolation clusters), one can obtain finite upper bounds
for the effective modulus tensor of rigidly-reinforced materials. Linear Hashin–
Shtrikman bounds of this type were derived by Hervé, Stolz, and Zaoui (HSZ)
(1991) for coated-spheres models, and for more general morphologies by Bornert
et al. (1996). In the coated-spheres model (see also Chapters 3, 4 and 7 of Volume
I) the material consists of composite spheres that are composed of a spherical core
of elastic stiffness tensor C2 and radius a, surrounded by a concentric shell of
elastic stiffness tensor C1 with an outer radius b > a. The ratio a/b is fixed, and
the volume fraction φ2 of inclusions in d dimensions is given by φ2 = (a/b)d . The
composite spheres fill the space, implying that there is a sphere size distribution
that extends to infinitesimally-small spheres.
Bornert (1996) pointed out that in fact the (lower) bounds for coated-spheres
model can be interpreted as rigorous bounds for materials with the larger class
 4.7. Applications of Second-Order Exact Results 195

of particulate microstructures considered by Ponte Castañeda and Willis (1995).

When both the shapes of the rigid inclusions and their distribution are spherical,
the upper bound can be explicitly computed from the corresponding linear bound
of Hashin (1962) and HSZ (which are identical in this case), and is given by
 ⎧

 ⎨   ⎫
2/3 2 −1 ⎬
eq  1 2 21φ 2 (1 − φ )
1 = 1 + φ2 (1 − φ2 ) − 2
E , (171)
φ1 ⎩ 5 10(1 − φ )
7/3 ⎭ eq
2

which was first derived by Suquet (1993a).

For fiber-reinforced materials with cylindrical inclusions that have circular cross
sections, the following result (Li et al., 1993) is obtained:
#  
eq 3 1
σ1 = 12 + 4 ( 11 − 22 ) + 13 + 23 ,
2 2 (172)
1 + φ2
where the axis of symmetry was assumed to be along the x3 direction. Such ma-
terials are inextensible along the fiber direction and can only support shear in the
transverse and longitudinal directions. Another important case is one in which one
of the aspect ratios of the rigid inclusions approaches zero, leading to disk-like
inclusions (in this limit, φ2 → 0). If the disks have circular cross sections, and are
aligned and distributed isotropically, the following result is obtained:
  $ %  2
eq 4ρ 1 4 ρ −1 1
σ1 = eq −
2
1+ 33 − ( 11 + 22 )
π 3 5π 2
$ %−1  2 1/2
24 ρ 1
+2 1 + 12 + ( 11 − 22 )
2 2
. (173)
15 π 2

where, as before, ρ = 43 π n2 a 3 is the disk density corresponding to n2 disks (per

unit volume) of mean radius a. The corresponding results for flat distributions
of disks (i.e., when the disk interactions are weak) are obtained by linearizing
Eq. (173), and were first given by Talbot and Willis (1992) and Li et al. (1993). If
the disks are randomly oriented and distributed isotropically, one obtains (Ponte
Castañeda and Willis, 1995)
#
$ %$ %
eq 12 ρ 8 ρ −1
1 = 1+ 1− Eeq . (174)
15 π 15 π

4.7.2.2 Three-Point Bounds and Estimates

Utilizing the third-order bounds for linear, elastic materials (see Chapter 7 of
Volume I), it is straightforward to derive the following third-order estimates for
rigidly-reinforced materials with statistically isotropic microstructures:
# $ %
eq 1 3 11ζ1 + 5η1
1 = 1+ φ2 Eeq , (175)
φ1 2 21η1 − 5ζ1
196 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

which, in the limit, ζ1 = η1 = 1, reduces to the Hashin–Shtrikman estimate. Be-

cause for these values of ζ1 and η1 , the Beran upper and lower bounds coincide,
Eq. (175) is a rigorous upper bound for nonlinear materials with ζ1 = η1 = 1.
One may also obtain the self-consistent or effective-medium approximation
estimates for statistically-isotropic microstructures by utilizing the estimates of
Hill (1965b) and Budiansky (1965) (see also Chapter 7 of Volume I). The result is
then given by (Ponte Castañeda, 1991)
#
$ %−1
eq 1 5
1 = 1 − φ2 Eeq . (176)
φ1 2
The results presented so far represent rigorous bounds for the effective mechan-
ical properties of rigidly-reinforced materials. The corresponding second-order
Hashin–Shtrikman estimate, Eq. (170), and the self-consistent estimate, Eq. (176),
for statistically-isotropic microstructures (or isotropic distributions of spheri-
cal voids) were derived by Ponte Castañeda (1996a) and Ponte Castañeda and
Nebozhyn (1997) for materials with an isotropic, incompressible matrix phase, as
characterized by the function f1 in Eq. (164).

4.7.3 Completely Plastic Materials

Another class of nonlinear composites for which explicit analytical results are
available consists of two-phase, rigid, perfectly plastic materials with isotropic
constituents. In certain limits of this class, the associated nonlinear equations for
the comparison moduli or reference strain in the phases can even be solved exactly.
For example, consider a two-phase material with isotropic, ductile phases governed
by the Von Mises criterion,
σ eq (x) ≤ σi0 , in phase i. (177)
Then, the variational representation (95) can be utilized for deriving explicit results
for some cases of practical interest, which are now briefly discussed.
If the material is isotropic, the dissipation potential He depends only on the
second and third invariants of the strain and, due to homogeneity, can be written as
He (E) = σe0 (θ )Eeq , (178)
where θ depends on the determinant of the normalized deviatoric strain. Use of a
piecewise constant shear modulus µ0 (x) in (95) then leads to the following upper
bound for σe0 :
⎧  ⎡  2 ⎤⎫
⎪
⎨  0 0 0 ⎪
⎬
0
σe  µe ⎣ σ1 µ2
⎦
≤ inf φ1 + φ 2 , (179)
µ01 /µ02 ≥0 ⎪ ⎪
0 0 0 0
σ2 ⎩ µ2 σ2 µ1 ⎭

which is independent of θ and therefore of the third invariant. Rigorous upper

bounds for the effective flow stress σe0 of isotropic, two-phase materials can then
be obtained by incorporating upper bounds for the effective shear modulus µe of
the linear comparison material in (179). For example, assuming that σ10 ≥ σ20 , the
 4.7. Applications of Second-Order Exact Results 197

Hashin–Shtrikman upper bound for a d-dimensional material is given by (Ponte

Castañeda and deBotton, 1992; Suquet, 1993a; Olson, 1994)
 ⎡  ⎤
 2
 2
σe0 (d + 2)φ2 dφ1  σ10 2 ⎣ σ10
= + + φ2 − 1⎦. (180)
σ20 d + 2φ2 d + 2φ2 σ20 d σ20

Similarly, the Hashin–Shtrikman estimates for spherical inclusions, distributed

with statistical isotropy, can also be derived. In this case, estimates for the effec-
tive flow stress of the material can be obtained by using the appropriate estimates
for µe for this class of microstructures. If the estimate for µe is accurate for arbitrary
contrast µ1 /µ2 , then, the resulting expression for σe0 is likely to be an upper bound
for the same class of microstructures. For example, the Hashin–Shtrikman lower
bound is appropriate for describing the effective shear modulus of dispersions of
spherical inclusions (phase 2) in a matrix (phase 1) at moderate volume fractions
of inclusions which, as mentioned earlier in this chapter (see also Chapter 7 of
Volume I), is a rigorous upper bound for materials with the microstructural param-
eters ζ1 = η1 = 1. When used in (179), the optimization procedure can be carried
out analytically. Assuming that σ20 ≥ σ10 , the estimate for the overall flow stress
resulting from this calculation is given by (Ponte Castañeda and deBotton, 1992)
 ⎡

  2 ⎤
σe0 (d + 2)φ2 σ2 0
dφ1  2φ2 ⎣ σ20
= + 1+ 1 − 1 − 0 ⎦, (181)
σ10 d + 2φ2 σ10 d + 2φ2 d σ1

with
&
σ10 2 1
≥ 1 + dφ2 ,
0
σ2 d +2 2
and
&
σe0 1
= 1 + dφ2 , (182)
σ10 2
where
&
σ10 2 1
≤ 1 + dφ2 .
0
σ2 d + 2 2
These results, which may be interpreted as approximate estimates for materials
with particulate microstructures, are upper bounds for composites with morpholo-
gies for which the Hashin–Shtrikman lower bound for µe is exact (for example,
sequentially-laminated composites; see Chapter 2). Note that the estimate (182)
predicts that the strengthening effect of the inclusions (when they are stronger than
the matrix) saturates after a certain finite increase in the strength of the inclusions.
This is a consequence of the non-hardening character of the matrix phase, which
would be expected to carry all the deformation, for sufficiently strong (but still
non-rigid) inclusions.
198 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

For unidirectional materials with transverse isotropy (or for fiber-reinforced

composites with circular fibers of phase 2 dispersed isotropically), the expression
for the effective yield function reduces to
⎧⎡ ⎫
⎪  2 ⎤−1 ⎪
⎨ σ10  2 ⎬
Pe () ≥ max ⎣φ2 + φ1 y ⎦  · M̂0e (y) ·  − σ20 , (183)
y≥0 ⎪
⎩ σ20 ⎪
⎭

where y = µ01 /µ02 , and the tensor M̂0e = µ01 M0e is the (normalized) effective com-
pliance of the fiber-reinforced linear comparison material with incompressible and
isotropic phases. In general, this result requires numerical computation, but for
transverse and longitudinal shear, the result simplifies to expressions similar in
form to (180)–(182) with d = 2. Similarly, for (axisymmetric) uniaxial tension,
one obtains
σe0 = φ1 σ10 + φ2 σ20 , (184)
in agreement with the Voigt estimate. These results are due to Ponte Castañeda
and deBotton (1992) and Moulinec and Suquet (1995); see also deBotton (1995).
In a similar way, one may obtain second-order estimates for the effective me-
chanical properties of this class of nonlinear materials. For example, for two-phase,
rigid, perfectly plastic materials with statistically-isotropic microstructures [or
with isotropic distributions of spherical inclusions (phase 2) in a matrix (phase
1)], the second-order estimates (145) for He can be simplified. The result, for
simple shear loading conditions, is given by
⎧  
⎪
⎪ 1 σ20 σ20
⎪
⎨ 1 − (1 + φ 2 ) 1 − , if < 1,
σe0 2 σ10 σ10
= (185)
σ10 ⎪
⎪ σ20
⎪
⎩ 1, if ≥ 1.
σ10
An identical result is obtained for fiber-reinforced microstructures with transverse
isotropy loaded in transverse shear. We should point out that the small-contrast
expansion described in Section 4.6.1 diverges for simple shear loading, whereas,
as indicated by Eq. (185), the corresponding second-order estimate does not.
Finite-element computations carried out by Suquet (1993a), for particle-
reinforced materials with inclusion volume fraction φ2 = 0.15, indicate that,
although the two types of nonlinear estimates obtained from the linear Hashin–
Shtrikman lower bound exhibit the same general trends, the second-order estimates
are in closer agreement with the numerical results. Moreover, the variational
estimates lie above the numerical results, consistent with the fact that the vari-
ational estimates are expected to overestimate the effective yield strength of the
composite at this value of φ2 . The nonlinear estimate obtained from the linear
Hashin–Shtrikman upper bound lies below the microstructure-independent Voigt
(one-point) upper bound (see Section 4.3.1), and is such that the second-order es-
timate lies below the variational estimate, which is known to be a rigorous bound
for all statistically-isotropic microstructures.
 4.7. Applications of Second-Order Exact Results 199

One may also compare the results of numerical simulations by Moulinec and
Suquet (1995) for the effective yield strength of fiber-reinforced materials with
the corresponding predictions (183) obtained from the variational method. These
authors considered cylindrical fibers (phase 2) with circular cross section and
aligned with the x3 axis, distributed randomly in a matrix (phase 1). The overall
stresses considered by these authors consisted of the superposition of uniaxial
tension and transverse shear,
= 11 (e1 ⊗ e 1 − e 2 ⊗ e2 ) + 33 e3 ⊗ e3 .
Various contrast ratios for the strengths of the two phases were investigated:
σ20 /σ10 = 0.5, 1.1, 2, 3, 5, and 10. For σ20 /σ10 = 2, 11 different realizations were
used, while for the other ratios, the computations were performed on a single
realization, representative of the average of the predictions over the entire set of
configurations for σ20 /σ10 = 2, a configuration that approaches transverse isotropy,
with its overall strain/stress response being close to the mean response of all the
realizations, both under multiaxial loading and uniaxial tension. The results are
shown in Figure 4.2. The agreement between the numerical simulation results and
the variational estimates (183) is good. In particular, the variational estimates (183)
capture rather well the flat sectors on the yield surfaces.
For the cases that involve sufficiently strong fibers, the shape of the observed ex-
tremal surfaces was found to be bimodal in character. Bimodal surfaces were used
by Hashin (1980), Dvorak and Bahei-El-Din (1987), and de Buhan and Taliercio
(1991) for describing the initial yield or the flow surface of unidirectional com-
posites. The numerical and variational results are consistent with these models and
with experimental observations (Dvorak et al., 1988). The numerical calculations

Figure 4.2. Effective yield strength 11 of composites with cylindrical fibers aligned in
the x3 -direction (perpendicular to the plane of this page) with volume φ2 . The curves are,
from left to right, for σ20 /σ10 = 0.5, 1.1, 2, 3, 5, and 10. Symbols represent the results of
numerical simulations for randomly isotropic configurations (averaged over 11 realizations),
while the curves show the predictions of the variational method in which the Hashin–
Shtrikman lower bound for the linear comparison material has been used (after Moulinec
and Suquet, 1995).
200 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

also suggest closed-form expression for the bimodal surface:

 $ %2 1/2  $ %2 1/2
11 11
| 33 | ≤ φ1 σ10 1 − + φ2 σ20 1 − , (186)
K1 Ke
√
where K1 = σ10 / 3 is the in-plane shear strength of phase 1, and Ke is the in-
plane shear strength of the composite, which can either be fitted to the numerical
simulations (Moulinec and Suquet, 1995), or be taken from the prediction of the
variational procedure used with the Hashin–Shtrikman
√ lower bound (Ponte Cas-
tañeda and deBotton, 1992): Ke = (1/ 3)σe0 , with σe0 being given by (181) and
(182) with d = 2. In the second case, the agreement with the predictions of the
variational procedure for the full yield surface was found to be quite good.

4.8 Other Theoretical Methods

In addition to what was discussed above, several other theoretical methods have
been proposed over the past 30 years for predicting the overall effective mechan-
ical properties of nonlinear materials. Two noteworthy of such methods are the
(classical) secant method developed by Chu and Hashin (1971), Berveiller and
Zaoui (1979), and Tandon and Weng (1988), and the incremental method orig-
inally proposed by Hill (1965a) in conjunction with the self-consistent or the
effective-medium approximation method. Briefly, the secant method consists of
writing down the constitutive relation in phase i with the secant tensor of phase i,
evaluated at the average strain i . In the incremental method, one writes down
(t) (t)
the constitutive law of phase i in the form σ̇ = Li (i ) : ˙ i , where Li is now
the tensor of instantaneous or tangent moduli of the phase, given by the second
derivative of the energy wi with respect to the strain.
Two-phase, incompressible, power-law materials with the same exponent pro-
vide an important test for comparing the different models. Particulate power-law
materials were considered by Ponte Castañeda and Willis (1988) in the context
of the Talbot–Willis procedure, by Ponte Castañeda (l99la) and Suquet (1993a) in
the context of the variational method with a linear comparison material, and by
Ponte Castañeda (1996a) in the context of the second-order procedure. Granular
microstructures were also considered by these groups, as well as by Gilormini
(1995), who compared the different methods using the self-consistent method for
estimating the effective properties of the linear comparison material. He pointed out
that the predictions of the incremental and classical secant method can violate the
rigorous variational upper bound for isotropic materials. Michel (1996) proposed
a nonlinear extension of the self-consistent method for power-law materials.
Consider, as an example, two-phase materials with particulate microstructure.
Both phases are characterized by Eq. (10) with the same exponent m but different
stresses σi0 . Suppose that the material consists of inclusions (phase 2) that are
distributed randomly in a softer matrix (phase 1). If the volume fraction φ2 of the
inclusions is not too large, the Hashin–Shtrikman lower bound provides accurate
 4.9. Critique of the Variational Procedure 201

estimates for the effective linear properties of the comparison material with the
same microstructure as that of the nonlinear material. The material itself is a
power-law composite with the same exponent as the individual phases, and is, in
addition, incompressible. Under the assumption of statistical isotropy, the effective
potential is a function of the second and third invariant of the average strain E and,
by homogeneity, is given by Eq. (166). The variational bounds, derived above for
power-law materials provide bounds for σe0 that are independent of the parameter θ
of Eq. (166), whereas the estimates provided by the second-order theory do depend
on this parameter. It can then be shown (Ponte Castañeda and Suquet, 1998) that
the incremental and secant procedures lead to the stiffest predictions, whereas
the variational and second-order methods provide more compliant predictions.
In particular, since, as already noted in Chapter 2 (see also Chapters 4 and 7
of Volume I), the linear Hashin–Shtrikman lower bound is attained by certain
particulate microstructures, the variational estimates are actually upper bounds for
the nonlinear composites with the same type of microstructure. Therefore, both
the corresponding secant and incremental estimates violate this bound, whereas
the second-order estimates do not. In fact, the incremental estimates violate even
the Hashin–Shtrikman upper bound for statistically-isotropic microstructures, at
sufficiently large values of the exponent n. This is somewhat unexpected, as this
type of bound is known to correspond to the opposite type of microstructure, with
the stronger material occupying the matrix phase.
A similar observation was made by Gilormini (1995) in the context of the self-
consistent estimate (instead of the Hashin–Shtrikman lower bound). These results
indicate that the tendency of the incremental model to approach the Voigt (one-
point) bound (see Section 4.3.1) when m → 0 is not due to the approximate nature
of the self-consistent method, but is because of the shortcomings of the incremental
method itself. Let us emphasize again that of the four nonlinear homogenization
procedures described above, only the second-order theory yields estimates that
are exact to second order in the contrast between the properties of the phases. The
other three (variational, secant, and incremental) provide estimates that are exact
only to first order in the contrast.
Finally, Gibiansky and Torquato (1998b) derived approximations for the ef-
fective energy of d-dimensional nonlinear, isotropic, elastic dispersions. These
approximations are similar to those described in Sections 2.2.2.1 and 2.2.2.2,
derived by Gibiansky and Torquato (1998a), for the effective conductivity of the
materials with the same morphology. In addition, Gibiansky and Torquato (1998b)
derived cross-property relations that link the effective energy of nonlinear materials
with their effective conductivity.

4.9 Critique of the Variational Procedure

A valid criticism of the variational procedures is that they rely, from the very be-
ginning, on the assumption that the mechanical behavior of the constituent phases
can be described by a potential, which is not the case for many nonlinear (usually
202 4. Nonlinear Rigidity and Elastic Moduli: The Continuum Approach

elasto-plastic) materials. A partial response to this criticism was provided by Ponte

Castañeda and Suquet (1998) who argued that, at least for certain loading condi-
tions of practical interest, it is possible to use a deformation theory of plasticity,
instead of a flow theory, to describe the mechanical properties of the constituent
materials. This substitution is rigorous only when the loading is radial and mono-
tonic at every point x in the volume element , but it may also be appropriate
for small deviations from proportionality (Budiansky, 1959). The assumption of
proportionality is rarely met and deviations from radial paths are likely to be the
rule. Nevertheless, numerical simulations of the transverse response of nonlinear
matrices, reinforced by aligned continuous fibers, suggest that, even though local
deviations from this assumption are actually observed and found to affect the local
stress and strain fields, they seem to have little influence on the overall stress-
strain response of the material under monotonic loading, implying that using a
deformation theory for the constituents can be a good approximation for materials
that are subjected to a monotonic radial loading, such as uniaxial tension. Strictly
speaking, although this model is not applicable to general loadings, its predictions
for those loadings to which it is applicable are much more accurate (Suquet, 1997)
than those of theories that allow for more general loadings, such as the incremental
method or the transformation field analysis (Dvorak, 1992).
However, use of a deformation theory for non-monotonic loadings is not ap-
propriate. Instead, one must use a flow theory for which a variational method
cannot be utilized. The variational method can still yield useful insight into how to
construct approximate effective constitutive relations, expressed in terms of two
thermodynamic potentials, the free energy for reversible effects and the dissipation
potential for irreversible phenomena (see, for example, Rice, 1970; Mandel, 1972;
Germain et al., 1983).

Summary
Several continuum approaches to estimating the effective nonlinear mechanical
properties of multiphase materials were described and discussed. One method, due
to Talbot and Willis, is based on a nonlinear extension of the Hashin–Shtrikman
variational principles, while the second method, developed by Ponte Castañeda
(for nonlinear isotropic materials) and Suquet (for power-law composites) uti-
lizes new variational principles that involve a linear comparison material with the
same microstructure as that of the nonlinear composite. These methods provide
at least one type of rigorous bounds (i.e., upper or lower bounds). The Talbot–
Willis procedure yields the bounds of the Hashin–Shtrikman type, while the Ponte
Castañeda–Suquet method provides bounds and estimates of any type, given the
corresponding bounds and estimates for the linear comparison material. In both
cases, the resulting bounds and estimates are exact to first order in the contrast
between the properties of the phases. A third method, also developed by Ponte
Castañeda, yields estimates that are exact to second order in the contrast. The
resulting estimates are not, however, bounds of any type.
 4.9. Critique of the Variational Procedure 203

Despite this considerable progress, much remains to be done, especially since

it appears that the constitutive laws that characterize the behavior of many ma-
terials are rather complex. In addition, true second-order bounds, i.e., those that
are exact to second order in the phase contrast, remain to be derived. When the
deformations are finite, a material may undergo microstructural evolution. An ex-
ample is deformations that are present in metal-forming processes. Little is known
about modeling and predicting the mechanical properties of such evolving mate-
rials. Finally, no discrete model of the types that have been described throughout
this book has been developed for studying the mechanical properties of nonlinear
heterogeneous materials. This research field is wide open.
 Part II

Fracture and Breakdown of

Heterogeneous Materials
5
Electrical and Dielectric Breakdown:
The Discrete Approach

5.0 Introduction
Beginning with this chapter, and in the next three, we study and analyze failure and
fracture of heterogeneous materials. In the present chapter, electrical and dielectric
breakdown of composite materials, which constitute a set of complex, nonlinear,
and non-local transport processes, are described. Their nonlinearity stems from
the existence of a threshold: Below and far from the threshold nothing particularly
complex happens. The laws of linear (or constitutively nonlinear) transport hold,
and the electrical properties of the materials are described by the models that were
described in the previous chapters and in Volume I. However, at the threshold,
the materials’ behavior and their transport properties abruptly change and become
very complex. Note that, unlike the percolation threshold, the threshold in electrical
or dielectric breakdown is not geometrical but dynamical although, as discussed
below, the interplay between the heterogeneities and the dynamical threshold gives
rise to a rich set of phenomena that are completely absent in the linear transport
regime in the same system.
Dielectric breakdown in gases, liquids, and solids is a complex problem and has
been studied for a long time. Many breakdown phenomena in gases are relatively
well-understood (see, for example, Meek and Craggs, 1978), while some, such as
atmospheric lightning, are more difficult to analyze, because the density, conduc-
tivity, and humidity of air are distributed inhomogeneously. Another well-known
example, in addition to lightning, is surface discharges, also known as Lichten-
berg figures. These phenomena are beyond the scope of our book and will not be
considered.
In dielectric breakdown in solids, the material is initially non-conducting when
an electric field is applied across the sample. If the field exceeds a certain threshold,
the material breaks down and becomes conducting. The microscopic mechanisms
of dielectric breakdown in solid materials are much more complex than those in
gases since, in addition to dielectric effects, mechanical and chemical effects can
also intervene and make the problem more difficult. From a practical view point,
dielectric breakdown is an important phenomenon, since it limits the application of
dielectrics as insulators. For this reason, dielectric breakdown in solids has received
much attention over the past several decades, and has been especially studied inten-
sively over the past decade.Awell-known example of such phenomena is formation
and growth of electrical trees (as in, for example, discharge treeing in polymers).
208 5. Electrical and Dielectric Breakdown: The Discrete Approach

Figure 5.1. Schematic representation of a tree growing between two electrodes on two
parallel planes (after Hill and Dissado, 1983a).

The trees themselves may not cause breakdown unless they grow so large that
they span the thickness of the material. A diagrammatic representation of this
phenomena is shown in Figure 5.1. We will come back to this phenomenon shortly.
Another important example is dielectric breakdown in metal-loaded dielectrics,
which are disordered materials consisting of a mixture of conducting and non-
conducting components. For example, solid-fuel rocket propellant is a mixture of
aluminum and perchlorate particles in a polymer binder (Kent and Rat, 1985). It has
been reported that the breakdown field of this material decreases significantly by
the presence of the aluminum particles, and is also a strong function of the volume
fraction of the constituent particles. Dielectric breakdown of such composite solids
is dominated by space charge effects due to the large electric fields near any sharp
metal tips occurring in the composite, and thus the composite is unusually sensitive
to breakdown. Recall that about a decade ago the solid fuel of a United States Air
Force rocket experienced dielectric breakdown, with the fuel becoming electrically
conductive, setting the rocket on fire.
Electrical breakdown occurs when the current through a conducting medium
causes an irreversible resistance change in the medium. In this phenomenon, the
material is initially conducting. The failure occurs when the current density flowing
in the material exceeds a threshold value at and beyond which the material becomes
insulating. Unlike dielectric breakdown, the mechanism of electrical failure is
well-understood; it is merely Joule effect which causes degradation of metallic
interconnects (or the metal lines) which, due to electromigration phenomenon,
lose their conducting properties. Note that in this phenomenon the material be-
haves precisely like a fuse, which is broken when the applied voltage exceeds a
certain limit. Electrical breakdown is a major obstacle to development of nano-
size devices. Experimental realizations of electrical and dielectric breakdown in
 5.1. Continuum Models of Dielectric Breakdown 209

metal-insulator films, with a view to explain them in terms of the statistical physics
of disordered media, were reported by Yagil et al. (1992, 1993) among others. Hill
and Dissado (1983b) analyzed the older experimental data. We will come back to
these experiments later in this chapter.
Another important phenomenon that belongs to this class of problems is elec-
tromigration failure in polycrystalline metal films (see, for example, Huntington,
1975; Ho and Kwok, 1989). If a high current density passes through a thin metal
film, collisions between the conduction electrons and the metal ions result in drift-
ing of the ions and their electromigration. If there is a divergence in the flux of
the ions at some points, voids nucleate, grow and overlap with each other until
conduction ceases and the film suffers electrical breakdown (see, for example,
Rodbell et al., 1987). This phenomenon is particularly important in integrated cir-
cuits, where the continuing miniaturization of the circuits exposes the conducting
thin metal films to increasingly large current densities. Under such conditions,
electromigration failure decreases the circuit lifetime which is unacceptable from
an economical view point.
Throughout this book, both in Volume I and in the present Volume, we have
grouped the models for any phenomenon of interest to us into two classes—
the continuum models and the discrete models. In this chapter, we deviate from
this general approach because the continuum models of electrical and dielectric
breakdown of heterogeneous solid materials are well-documented (see, for ex-
ample, Whitehead, 1951; O’Dwyer, 1973; see also Niklasson, 1989a; Dissado
and Fothergill, 1992; Ohring, 1998, for more recent references); hence, the best
we could do would be summarizing these works, an unwise action. In addition,
as will be discussed in this chapter, many phenomena associated with electrical
and dielectric breakdown have a vector analogue in brittle fracture of solids, for
which many continuum models have been developed that will be described and
discussed in detail in Chapter 7. Thus, we restrict our discussion of the continuum
models to a few recent efforts that utilized extensive numerical solution of the
discretized continuum equations in order to study the breakdown phenomena in
strongly-disordered solids. On the other hand, over the past several decades several
discrete models of breakdown of heterogeneous materials have been developed.
These models are either stochastic or completely deterministic. Their general fea-
tures for modeling both the electrical and dielectric breakdown are the same, and
in fact, with appropriate modifications, a model for one of the phenomena can be
used for studying the other one. In this chapter, we describe these models in detail,
discuss their predictions, and, whenever possible, compare the predictions to the
relevant experimental data.

5.1 Continuum Models of Dielectric Breakdown

Typical of continuum models of dielectric breakdown are those of Garboczi (1988),
who studied the problem analytically, and of Gyure and Beale (1989, 1992) who
carried out a numerical study of the problem. What follows is a brief description
of each model.
210 5. Electrical and Dielectric Breakdown: The Discrete Approach

5.1.1 Griffith-like Criterion and the Analogy with Brittle Fracture

Garboczi (1988) extended the analysis of Griffith (1920) for brittle fracture (see
Chapters 6 and 7) to dielectric breakdown, and derived the criterion for nucleation
and development of a single conducting “crack” in an isotropic dielectric (insu-
lating) material. The problem that one solves is one of an elliptical inclusion with
dielectric constant   placed in an isotropic linear dielectric material with dielec-
tric constant . A far-field electric field E0 is then applied to the material, and the
Laplace equation, ∇ 2 V = 0 is solved for the distribution of the voltage V in the
material, subject to the boundary conditions that far from the inclusion the electric
field E → E0 , and that the normal component of the displacement field D = E is
continuous at the inclusion boundary. In the limit   → ∞ and fixed , the latter
boundary condition becomes V = 0 at the boundary of the inclusion.
This problem is easily solved by using elliptical cylindrical coordinates (u, θ, z)
(see, for example, Jackson, 1998), where we assume that all the quantities in the
z-direction are uniform. Then, the transformation between the (x, y) and (u, θ )
coordinates is given by

x = c cosh u cos θ, y = c sinh u sin θ, (1)

valid for 0 ≤ u < ∞ and 0 ≤ θ ≤ 2π . The inclusion’s surface is defined by u = β,

where β is a constant. If β → 0, then the inclusion degenerates into a “crack” of
length 2c with its tip at x = ±c. The solution of the problem is given by

cE 0
V = −cE 0 cosh u cos θ + exp(β − u)(  − ) sinh(2β) cos θ, u > β,
2C
(2)
cE 0
V =− (cosh β + sinh β) cosh u cos θ, u < β, (3)
C
with

C =  cosh β +   sinh β.

From this solution, the components of the electric field, namely, Eu = −τ −1 ∂V/∂u,
and Eθ = −τ −1 ∂V /∂θ, are computed, where τ = c(sinh2 u + sin2 θ )1/2 . One
then defines a field multiplication factor, Eu (β, 0)/E 0 = Ex (β, 0)/E 0 , which is
given by
Ex (β, 0) a(  − )
= 1 + , (4)
E0 a + b 
where a and b are the semi-major and semi-minor axes of the elliptical inclusion,
respectively.
The critical question to be answered is: What is the difference H in the electro-
static energy between a material with and without the inclusion? If the sources of
the applied field are fixed, then for the elliptical inclusion embedded in an infinite
medium, one has, H = − 12 px E 0 , where px is the x-component of the dipole
 5.1. Continuum Models of Dielectric Breakdown 211

moment p of the inclusion. It is straightforward to show that

$ %
2π c2 E 0   1
px = ( − )(cosh β + sinh β) sinh β . (5)
C 2
Therefore, in the limit of a conducting (  → ∞ with  held fixed) crack (β → 0),
one obtains
1
H = − π c2 (E 0 )2 , (6)
2
which is negative, indicating that the presence of the conducting crack lowers the
energy of the system. Had we made the same computations but for a fixed potential
(the common situation in practice), we would have obtained the same H, but
with the opposite sign.
Now, suppose that Hb is the breakdown energy required to create a unit area of
conducting crack (per unit length in the z-direction). Then, the surface energy of
the crack is 4Hb c > 0. Hence, the total energy difference between a cracked and
uncracked material is given by
1
H = − π c2 (E 0 )2 + 4Hb c. (7)
2
The linear term of Eq. (7) will dominate if c is small, implying that it is energetically
unfavorable to have the conducting crack nucleate or propagate. The reverse is
true for large enough c. The equilibrium point is thus found from dH/dc = 0,
yielding
&
4Hb
Ec =
0
, (8)
π c
for the critical value of the applied far-field. Equation (8), which was first derived
by Horowitz (1927), is the analogue of the Griffith’s prediction for brittle fracture,
which will be described in detail in Chapter 7. It is easy to show that the point
represented by Ec0 is a point of unstable equilibrium, and therefore for any applied
field E0 > Ec0 dielectric breakdown will occur spontaneously.
Similar to brittle fracture of materials, of great interest is the region around the
tip of the conducting crack where the most intense electric fields are located, and
where the dielectric breakdown actually takes place. For simplicity, consider the
limits   → ∞ and β → 0, and consider the x = c crack tip. One can then use a
new coordinate system consisting of r, the distance from the crack tip, and , the
angle from the x-axis. Then, in the limit (u, θ ) → 0, we obtain
& $ %
c 1
Er = E 0
cos  ,
2r 2
& $ %
c 1
E = −E 0
sin  , (9)
2r 2
$ %
√ 1
V = −E 0 2cr cos  .
2
212 5. Electrical and Dielectric Breakdown: The Discrete Approach

In analogy with brittle fracture, which is associated with a quantity referred to as the
stress-intensity factor (see Chapters 6 and 7), we define an electric field-intensity
factor KI or, more simply, field-intensity factor,
√
KI = π c E 0 ,
in terms of which one has
$ %
KI 1
Er = √ cos  ,
2π r 2
$ %
KI 1
E = − √ sin  , (10)
2π r 2
& $ %
2r 1
V = −KI cos  .
π 2
Physically, KI is the amplitude of the r −1/2 electric field singularity at the tip of
the conducting crack. One may also define the electrostatic energy release rate
HR by
d[ 12 π c2 (E 0 )2 ]
HR = = π c(E 0 )2 , (11)
dc
where HR dc is the amount of electrostatic energy released when the crack extends
by dc, with its critical value being, HcR = 4Ec0 . Moreover,
KI c
Ec0 = √ , (12)
πc
where KI c represents the critical value of KI .
Finally, Rice (1968) developed a line integral, usually called the J -integral,
which is independent of the contour. This quantity was originally developed for
fracture of material, and its usefulness becomes evident when the contour encloses
the tip of the fracture. Thus, J yields HR , the elastic energy release rate. The
J -integral for the elasticity problem is defined by
:  $ % 
∂u
J = −(σ · n) · ds + He dy , (13)
∂x
where u is the displacement vector, σ is the stress tensor, n is the unit vector normal
to the contour, and He is the elastic energy density. Since the analogue of the stress
tensor is the displacement field D, then, the J -integral for the electrostatic problem
is given by
:
J = [−(D · n)Ex ds + Hdy], (14)

where H is the electrostatic energy. Garboczi (1988) showed that, similar to me-
chanical fracture, the J integral for the electrostatic energy is independent of the
path. Equation (14) was also suggested by Hoeing (1984).
 5.2. Discrete Models of Electrical Breakdown 213

The above discussions should make it clear that, many results that have been
derived for brittle fracture of materials, based on the continuum models and de-
scribed in Chapter 7, can be directly translated into analogous results for dielectric
breakdown of materials.

5.1.2 Computer Simulation

Gyure and Beale (1989,1992) developed two-dimensional (2D) and 3D models of
breakdown of metal-loaded dielectric materials. Their model consisted of a ran-
dom array of perfectly conducting cylinders (in 2D) or spheres (in 3D), embedded
in a uniform dielectric. The same type of boundary conditions that were used in
Garboczi’s work (described above), i.e., continuity of the normal component of
the displacement field at the inclusions’ boundaries, and the far-field condition,
E → E0 , were also utilized by Gyure and Beale. In their model, the numerical
solution of the Laplace equation was obtained by the boundary element method
(Kim and Karrila, 1991) described in Section 7.8.2 of Volume I. After determining
the solution of the Laplace equation—the voltage distribution in the composite
material—those regions of the system that are vulnerable to breakdown are iden-
tified by using the fact that the largest electric fields lie along lines joining the
centers of the (cylindrical or spherical) inclusions which are closely spaced, with
center-to-center lines that are nearly parallel to the applied field. It is then assumed
that local breakdown occurs only between the pair of inclusions that has the largest
electric field between them, and that, as a result of the breakdown, an electrical
connection between the two inclusions is established, so that the two conductors
attain the same electrical potential. This assumption is based on the experimental
observation that such local breakdowns occur by vaporization of a portion of the
metallic particles followed by resolidification as a single conductor. The voltage
distribution of the new (defected) system was then calculated, the next region to
suffer breakdown was identified, and so on. Various quantities of interest, such
as the breakdown field, the geometry of the breakdown path, and the dielectric
constant of the medium, as a function of the packing fraction were calculated by
Gyure and Beale (1989,1992). These properties are further discussed below, where
we describe the discrete models.

5.2 Discrete Models of Electrical Breakdown

We first describe and discuss discrete or lattice models of electrical breakdown of
materials with percolation-type disorder. As we have been emphasizing through-
out this book, the reason for considering percolation-type heterogeneities is that,
they represent strong disorder and therefore any theory that provides reasonable
predictions for a material with percolation disorder should be at least as accurate
for other less extreme types of disorder. We will, however, discuss the effect of
other types of disorder.
214 5. Electrical and Dielectric Breakdown: The Discrete Approach

Thus, the problem that we wish to study is the following. We are given a dis-
ordered material, represented by a lattice in which the conductance of every bond
is selected from a probability density function f (g). In this state, the material is
completely conducting (it contains no insulating region). We now select at random
a fraction 1 − p of the bonds and convert them to insulators; that is, the fraction
of the conducting bonds is p. So long as p  pc , where pc is the percolation
threshold of the lattice, the material will still be conducting, albeit with a smaller
effective conductivity than when p = 1. We now apply a voltage V across the
material. If V is small enough, then there would be no change in the conductivity
state of the material. We now increase V by an amount large enough that the first
microscopic failed region (or the first failed bond in the lattice model) appears in
the material. Then, the material may behave according to one of the two scenarios.
(1) As soon as the first failed region appears, the entire material may fail rapidly
by an avalanche of local failed regions, without any need for increasing the
applied voltage V.
(2) The state of the material may be such that the macroscopic failure of the ma-
terial is more gradual, as the disorder distributes the current in an “equitable”
way, rather than concentrating it in a few weak regions. In this case, after the
first failed region appears, nothing further happens, unless we increase the
applied voltage so that new failed regions can emerge.
Corresponding to any applied voltage, there exists a current that flows through
the material. Since in practice macroscopic failure of the material is what one is
interested in, we consider the behavior of the macroscopic current and its influence
on the material. If this current exceeds a threshold If , then, the material as whole is
converted to an insulator and fails. Two important questions that must be addressed
by any model are as follows.
(1) How does If depend on p?
(2) How does the breakdown process take place? In other words, how does the
first sample-spanning path of the failed regions (or bonds in the lattice models)
appear for the first time?
Let us analyze the problem in detail for two limiting cases, namely, the
dilute limit when p  1 (very few insulating regions), and the opposite limit,
p  pc (most of the sample being insulating).

5.2.1 The Dilute Limit

Consider first the dilute limit. In a completely conducting material (no insulating
regions), the current lines are more or less parallel to each other and perpendicular
to the electrode surface. Suppose now that there is only one insulating defect in
the material which, for simplicity, is assumed to be spherical (or circular in 2D). In
the lattice model, the corresponding defect consists of a few insulating bonds that
form a cell with a regular shape, placed at the lattice’s center. Then, the current
lines around the defect are “deformed,” leading to a current enhancement. If id
 5.2. Discrete Models of Electrical Breakdown 215

and iu are, respectively, the current densities around the defect and far from it in
the unperturbed state, then, one can write
id = iu (1 + E), (15)
where E is the enhancement factor, the magnitude of which depends on the ma-
terial’s morphology. For example, for an elliptical defect with major and minor
axes 2a and 2b, E = a/b. The total current flowing through the material is then,
I = Siu = Sid /(1 + E), where S is the surface area of the electrode. The first fail-
ure happens when id = iw , where iw is the threshold current density for the failure
of the sample without the defect. Therefore,
Siw
If = , (16)
1+E
implying that the current enhancement decreases the failure current If . Typically,
the current for the complete first failure is also the current for failure of the sample,
since as soon as the regions in the vicinity of the defect fail, the current density
around the new defect is further enhanced, leading to a rapid failure of the entire
material. Clearly, the most damaging defects are those that are perpendicular to
the current lines, and are in the form of long cylinders or rods. The probability of
developing a defect depends on its shape.
In the context of the lattice models, the simplest and smallest defect is one
insulating bond which is positioned parallel to the direction of the current lines
and is far from the lattice’s boundaries (see Li and Duxbury, 1987, for the effect
of the defects that are near the boundaries of the lattice). If no defect is present in
the lattice, then, If = Liw , where L is the linear size of the lattice. For a defect of
size one (i.e., one bond), it is not difficult to show that, E = π/4, and therefore in
this case,
π
If = Liw . (17)
4

5.2.2 The Effect of Sample Size

The most damaging defect consists of N neighboring insulating bonds that are in
the same plane which is perpendicular to the current lines. Thus, in 2D the most
damaging defect is a line of N of such insulting bonds, while in 3D it is a set of
such bonds with roughly the shape of a disk. Since in 3D the current that is diverted
by the N bonds√should be distributed over the perimeter of the defect, which is
proportional to N , one obtains

iw (1 + a2 N ), 2D,
id = √ (18)
iw (1 + a3 N ), 3D.
The next issue to be addressed is the relation between N and L, the linear size of
the lattice. Since the probability that a bond has failed is proportional to (1 − p),
then, PN , the probability that N bonds are insulating, is given by
PN ∼ (1 − p)N Ld , (19)
216 5. Electrical and Dielectric Breakdown: The Discrete Approach

where Ld represents the volume of the system. The most probable, most damag-
ing defect is formed when PN ∼ 1, and therefore the critical number Nc for the
formation of such a defect is given by
d
Nc ∼ − ln L. (20)
ln(1 − p)
Therefore, the corresponding current density id is given by
⎧  
⎪ −2 ln L
⎪
⎪ iw 1 + a 2 , 2D,
⎨ ln(1 − p)
id =     (21)
⎪
⎪ −3 ln L 1/2
⎪
⎩ wi 1 + a 3 , 3D.
ln(1 − p)

Because the total current in the system is iLd−1 , the failure current is obtained by
setting id = iw , resulting in
⎧ iw L
⎪
⎪ 2D,
⎨ 1 + 2a ln L/ ln(1 − p) ,
2
If = (22)
⎪
⎪ iw L2
⎩ , 3D.
1 + 3a3 [ln L/ ln(1 − p)]1/2
The most interesting aspect of Eq. (22) is its prediction for the size-dependence
of the failure current. According to this equation, the failure current per bond,
if = If /Ld−1 , decreases with the linear size of the sample in a complex way (in
practice, L is the ratio of the linear size of the actual sample and the typical size
of the insulating defects). If ln(1 − p) is not too large, then

(ln L)−1 , 2D,
If ∼ (23)
(ln L)−1/2 , 3D.

Thus, for a fixed size of the insulating defect, the larger the sample, the smaller
the failure current.

5.2.3 Electrical Failure in Strongly Disordered Materials

In the limit, p  pc , where the material is strongly heterogeneous, the distribution
of the current in the materials is controlled by the links or the red bonds of the
percolation lattice model (see Chapter 2 of Volume I) that connect two multiply-
connected conducting clusters. These are the bounds that, if cut, would break the
sample-spanning clusters into two parts. They break down and become insulating
by only a small current. Therefore, it is reasonable to assume that as p → pc , the
critical current If vanishes. The number of the links is proportional to ξpd−1 , where
ξp is the correlation length of percolation and d is the Euclidean dimensionality
of the system. Since near pc , ξp ∼ |p − pc |−ν , if  is the thickness of the links,
 5.2. Discrete Models of Electrical Breakdown 217

then we must have


If ∼ iw ∼ (p − pc )(d−1)ν . (24)
ξpd−1
On the other hand, If = ge Vf , where ge is the effective conductivity of the sample,
and Vf is the failure voltage. Since near pc one has, ge ∼ (p − pc )µ , where µ is
the critical exponent of the effective conductivity near pc , we obtain
Vf ∼ (p − pc )(d−1)ν−µ . (25)
Equation (25) can also be derived by the following more detailed analysis. If, for
length scales L  ξp , we cut one red bond, it splits the sample-spanning conduct-
ing cluster (and hence the backbone) of the material into two pieces, and therefore
the total critical current for breakdown is I ∼ O(1) (because all the current must
go through this red bond), and thus the failure (breakdown) current density (cur-
rent per length of the sample) is If = I /L ∼ 1/L. Therefore, the failure voltage
Vf is given by Vf ∼ If /Ge , where Ge is the effective conductance of the sam-
ple. As Ge ∼ Ld−2 L−µ/ν for a d-dimensional system (note that the factor Ld−2 is
included to convert the effective conductivity to the effective conductance), we ob-
tain Vf ∼ Lµ/ν−(d−1) . For L  ξp , we replace L by ξp ∼ (p − pc )−ν and obtain
Eq. (25). Equation (25) indicates that there is a qualitative difference between 2D
and 3D materials. In 2D where µ  1.3 and ν = 4/3, (d − 1)ν − µ > 0, and there-
fore Vf vanishes as pc is approached, in agreement with the results of computer
simulations (see below). On the other hand, in 3D where µ  2.0 and ν  0.88,
(d − 1)ν − µ < 0, and therefore Vf diverges as pc is approached. Therefore, a
thin (2D) conducting film (attached to a substrate) suffers electrical breakdown
quite differently than a bulk (3D) material.
If, instead of a lattice model, we utilize a continuum one, then, the exponent that
characterizes the power-law (25) will be different from its lattice counterpart (d −
1)ν − µ. For example, this exponent for the Swiss-cheese model in which spherical
or circular holes are distributed in an otherwise uniform conducting matrix, is
given by ν + d − 1 + δ, where δ = 1 and 3/2 for d = 2 and 3, respectively, so
that the voltage Vf for a continuum near its percolation threshold is smaller than
the corresponding value for a discrete system.
Consider now the effect of the sample size which, in the context of the lattice
model, leads us to the size of the most damaging defect which is an inclusion of
size ξp , in the direction parallel to the macroscopic voltage, and ⊥ perpendicular
to it. The total probability P of having a defect of size  is, P = P (L/ξp )d , where
P is the probability density of defect clusters of size . Percolation theory predicts
(Stauffer and Aharony, 1992; Sahimi, 1994a) that
$ %

P ∼ ξp−1 exp − , (26)
ξp
and therefore, P = exp(−/ξp )(Ld /ξpd+1 ). The linear size ⊥ is that value of 
for which P ∼ 1, which then yields
⊥ ∼ ξp ln L. (27)
218 5. Electrical and Dielectric Breakdown: The Discrete Approach

Since the current that flows through the side link of the defect is proportional to
(⊥ )d−1 I , one obtains
(p − pc )(d−1)ν
If ∼ , (28)
(ln L)d−1
implying that, finite size of the sample generates a (weak) correction to Eq. (24).
Thermal effects also modify Eq. (24) which will be described in Section 5.2.7.
Duxbury and Li (1990) proposed that one may combine the above results for
the dilute limit and the region near pc , Eqs. (22) and (24), into a single unified
equation, given by
$ %
p − pc φ
1 − pc
If = Iw   , (29)
ln(L/ξp ) ψ
1+c
ln(1 − p)
where c is a constant, and ψ is an exponent, the precise value of which is not
known, but can be bounded by
1
< ψ < 1. (30)
2(d − 1)
Thus, in general, there exist three regimes.
(1) For p = 1, one has If = Iw , as expected.
(2) For p  1, the numerator of Eq. (29) is essentially a constant of order unity,
and one recovers Eq. (22).
(3) For p  pc , the denominator of Eq. (29) is of the order of 1, and one recovers
Eq. (24) with φ = (d − 1)ν.

5.2.4 Computer Simulation

One of the first computer simulations of a discrete model of electrical breakdown
problem was carried out by de Arcangelis et al. (1985). In their model, a fraction
p of the bonds are conducting, while the rest, with fraction (1 − p), are insulating.
A voltage is then applied to the lattice. Once the current in one bond reaches the
failure value, the failed bond is removed (its conductance is set to zero), a voltage
is applied again, and the next bond to fail is looked for. This procedure is repeated
until the system fails macroscopically and its effective conductivity vanishes. A
slightly more general version of this model was studied by Duxbury et al. (1995) in
which each bond of a lattice, with probability p, is a conductor with conductance
g1 and failure current threshold i1 , while the rest of the bonds, with a fraction
(1 − p), have a conductance g2 and a threshold i2 . Söderberg (1987) and Stephens
and Sahimi (1987) suggested another model in which each bond burns out and
becomes insulating if the dissipated Joule heat in it exceeds a threshold value.
In general, as more bonds fail, the necessary applied voltage for failing a bond
decreases. de Arcangelis et al. (1985) determined two voltages: One, Vi , is the
 5.2. Discrete Models of Electrical Breakdown 219

Figure 5.2. Size dependence of the fail- #

ure current If in the square network. The
curves from top to bottom are, respec-
tively, for p = 0.6, 0.7, 0.8 and 0.9 (after 
Duxbury et al., 1987).

L / If
#


$  !  

voltage at which the first bond fails, and a second one, Vl , causes the last bond,
and hence the sample, to fail. The two voltages exhibit very different behaviors as
p, the fraction of the conducting bonds in the original lattice, was varied. Vi first
decreases up to p  0.7, and then increases again. On the other hand, Vl increases
monotonically with p until, in the vicinity of pc , it becomes roughly equal to Vi .
Duxbury et al. (1987) employed the same model and analyzed the dependence of
the failure current If on the sample size L. Figure 5.2 presents the results, where
L/If is plotted versus ln L. The linear dependence of L/If on ln L, for several
values of p, is in agreement with Eq. (29). In addition, when If was determined as
a function of p, it was found to follow Eq. (24) [or Eq. (29)], although when their
data are fitted to this equation, the exponent φ is about 1, rather than the theoretical
prediction (for d = 2), φ = ν = 4/3.
de Arcangelis and Herrmann (1989) studied a model of electrical breakdown in
which each conducting bond is characterized by a voltage threshold, such that if the
voltage along the bond exceeds the threshold, the bond breaks down and becomes
an insulator. This model can be thought of as the scalar analog of brittle fracture
of materials, in which a microscopic portion of a material behaves elastically
until the stress or the force that it suffers exceeds a threshold, in which case the
material breaks. The thresholds in the model of de Arcangelis and Herrmann (1989)
were distributed according to a probability distribution function. Interesting scaling
properties, in addition to what we described above, were discovered for the model.
For example, the total current I that passes through the network, as the conducting
bonds burn out, scales with the linear size L of the network as
I ∼ Lζ h(Nb /LDf ) (31)
where Nb is the number of burnt-out bonds, and h(x) is a universal scaling function.
Numerical simulations in 2D indicated that ζ  0.85 and Df  1.7. Note that
Df represents the fractal dimension of the set of all the burnt-out bonds. If one
considers only those burnt-out bonds that form a sample-spanning cluster, then
one finds that, Df  1.1, indicating that the cluster is almost like a straight line.
Moreover, de Arcangelis and Herrmann (1989) found that the distribution of the
local currents in the network just before it fails macroscopically is multifractal, so
that each of its moments is characterized by a distinct exponent (which is similar
220 5. Electrical and Dielectric Breakdown: The Discrete Approach

to the distribution of currents in random resistor networks studied in Chapter 3,

and also Chapter 5 of Volume I), whereas the same distribution obeys constant-
gap scaling (i.e., there is a constant difference between the exponents so that
from one exponents all other exponents are computed) before the catastrophic
failure sets in (i.e., the point beyond which the network burns out very quickly
and becomes insulating). Since, as pointed out above, many properties of such
models of electrical breakdown have analogues in the problem of brittle fracture,
we postpone a more detailed discussion of these properties to Chapter 8 where
we describe and discuss the discrete models of brittle fracture and other types of
mechanical breakdown. Simulation of large 3D models of this type was carried
out by Batrouni and Hansen (1998) who found that their results follow Eq. (31).

5.2.5 Distribution of the Failure Currents

Equations (22), (24), and (29) predict the value of the most probable failure current.
In practice, this quantity is not a self-averaged property. That is, nominally identical
samples have different failure currents. Therefore, there is a distribution of such
currents, which also depends on the linear size of the sample. Duxbury et al.
(1987) determined this distribution by computing PL (N ), the probability that in
a sample of linear size L, no defect of insulating configuration with a size larger
than N (bonds) is formed. In order to accomplish this, the lattice is divided into
smaller elementary cubes (or squares in 2D) of linear size Lc . Due to the statistical
independence of the elementary cubes, the probability that no defect of size larger
than N forms is [PL (N )]n , where n is the number of the elementary cubes of linear
size Lc . Since the distribution functions must have the same form on the lattice
and its elementary cubes or sublattices, one must have
[PL (N )]n = PL (an N + bn ), (32)
where an and bn are scaling functions that remain finite as n → ∞. Two general
solutions can now be derived.
(1) an = 0, in which case one has
PL (N ) = exp[−x1 exp(−x1 N )], (33)
where x1 > 0 and x2 > 0 are two parameters to be determined.
(2) bn = 0, in which case one obtains
PL (N ) = exp(−r/N m ), (34)
with r > 0 and m > 0. To determine the constants x1 and x2 , we note that the
probability that a defect of size N is formed is given by dPL /dN , and the
maximum of this probability is obtained when N = Nc , where Nc is given
by Eq. (20). Consequently, one finds that, x1 = cLd and x2 ∼ − ln(1 − p),
where c is a constant which depends on the dimensionality d of the system.
Thus, the numerical value of x1 is large, ensuring that PL (0)  0. Combining
these results with Eqs. (18) and (22), the cumulative probability of failure,
 5.2. Discrete Models of Electrical Breakdown 221

FL (If ) = 1 − PL (N ) is then given by

  $ %d−1 
Iw /I − 1
FL (If ) = 1 − exp −cL exp −dc
d
ln L . (35)
Iw /If − 1
Distribution (35), which was derived by Chakrabarti and Benguigui (1997),
is a double exponential distribution and is normally referred to as the Gumbel
distribution (Gumble, 1958). We remind the reader that Iw is the current for the
failure of the pure sample (without any insulating region, or the limit p = 1).
If L is large enough, then the current If that appears in Eq. (35) is indeed
the most probable failure current. Note also that FL (∞) → 1 only when L is
large enough. How large is large enough cannot be answered very precisely,
because the constant c depends on the dimensionality d. Duxbury et al. (1987)
derived the probability FL in terms of the failure voltage Vf . Their equation
is given by
 $ %
kLd−1
FL (Vf ) = 1 − exp −cLd exp − , (36)
Vf
where k is a constant.
On the other hand, Eq. (34) does not lead straightforwardly to a corresponding
cumulative probability of failure. However, it is often stated that, the cumulative
probability distribution FL that corresponds to Eq. (34) is the Weibull distribution,
given by
 $ %m 
I
FL (If ) = 1 − exp −rL d
. (37)
If
If the parameter m is large enough, then If that appears in Eq. (37) is indeed the
failure current. Two points are now worth mentioning.
(1) Distributions (35)–(37) are valid if the material is far from its percolation
threshold. It has been proposed that, near pc , the following cumulative failure
distribution should be valid,
⎧ ⎡ ⎤⎫
⎨ k (p − pc ) ⎬
 ν
F (If ) = 1 − exp −c Ld exp ⎣− 1/(d−1) ⎦ , (38)
⎩ I ⎭
f

c k
where and are two constants, the precise values of which are not known.
The distribution (38) is similar to (35) (in the sense of being double expo-
nential) although, unlike (35), it has never been checked against the results of
computer simulations or experimental data.
(2) It is difficult to test the validity of the Gumbel distribution against the Weibull
distribution by simply fitting the data to them. However, if one defines a
quantity A by
 
ln[1 − FL (Vf )]
A = − ln − , (39)
Ld
222 5. Electrical and Dielectric Breakdown: The Discrete Approach

then, the corresponding quantity, for example, for the distribution (37) (when
written for the failure voltage Vf ) is given by
$ %
1
AW = a1 ln + b1 , (40)
Vf
and thus a plot of AW versus ln(1/Vf ) must be linear. On the other hand, for
the Gumbel distribution, Eq. (35) or (36) [or (38)], one has
$ %
1
AG = a2 + b2 , (41)
Vf
which predicts linear variation of AG with 1/Vf . In this way, one can clearly
determine which cumulative distribution provides a better fit of the data.
Duxbury et al. (1987) found, using this method, that the Gumbel distribu-
tion provides a more accurate fit of their numerical data. In Chapter 8 we will
utilize this method in order to test the accuracy of analogous distributions for
the failure stress of brittle materials.

5.2.6 The Effect of Failure Thresholds

In practice, different parts of a material may exhibit different resistance to electrical
breakdown. Therefore, a more realistic model may be one in which one character-
izes the conducting bonds by a threshold in the voltage or current, beyond which
it breaks down and becomes insulating. The thresholds can be selected from a
probability density function, which then introduces into the model a heterogeneity
that is different from percolation disorder. Kahng et al. (1988) considered such
a model in which each bond is characterized by a failure voltage uniformly dis-
tributed over the range v− = 1 − 12 w to v+ = 1 + 12 w, where 0 < w ≤ 2. All the
bonds have the same resistance. The limit w = 0 represents a system without any
disorder, while the limit w = 2 corresponds to a uniform distribution in (0,2). A
voltage is applied to the system and is increased until the first bond fails. The con-
ductance of the failed bond is set to zero, the applied voltage is kept content, and
the voltage distribution in the network with its new configuration is recalculated.
If another bond fails, its conductance is set to zero, and the procedure is repeated.
If, at some stage, no more bond fails, the applied voltage is increased gradually
until the next bond fails. This procedure is repeated until the entire sample fails.
This model represents a slow breakdown process, since the characteristic time for
a “hot” bond to suffer breakdown and become an insulator is assumed to be much
larger than the time that it takes the system to relax and reach equilibrium.
Despite its apparent simplicity, the behavior of the system depends crucially
on the value of w and the linear size L of the network, and exhibits interesting
phenomena. Similar to the models with percolation disorder, the value of the
external voltage to cause the network to fail decreases as L increases, but at a
rate that depends on w. If w is small enough, then one of the first few bonds
that fails triggers a path of failed bonds that propagates across the system. This
is somewhat similar to brittle fracture of relatively homogeneous solids (in which
 5.2. Discrete Models of Electrical Breakdown 223

mechanical failure of the first few atomic bonds generates a path of broken bonds
that eventually spans the materials), and hence we refer to this case as the “brittle”
regime. In this case, the failure of the material is governed by the weakest (or at
least one of the weakest) bonds in the initial system. For larger w, the disorder is
stronger, and therefore the breakdown of the material is more gradual, as there is
a large range over which individual bonds’ failure is driven by an increases in the
applied voltage. This situation somewhat resembles ductile fracture, and therefore
we refer to it as such (without claiming that it actually represents the scalar analogue
of ductile fracture). In a d-dimensional network of volume Ld , ductility is expected
if the number of failed bonds exceeds Ld/2 . Then, the behavior of the breaking
voltage in the ductile regime parallels that of materials with percolation disorder.
However, the average breaking voltage cannot be less than v− , and therefore this
leads to an eventual crossover to the brittle regime as the linear size L of the
network increases, except when w = 2.
Whether the network behaves as in the brittle or ductile regime depends on w
(1)
and L. Kahng et al. (1988) showed that there exists a critical value wc of w such
that, regardless of L, the material always fails in the brittle regime. The failure of
(1)
the system in this case is trivial. For w > wc , the network’s failure is brittle for
large L and ductile for small L. The two regimes are separated by another critical
(2) (2)
value of w, wc (L), which is a function of L. For L → ∞, one has wc → 2,
and failure of the system is brittle.
More quantitatively (but approximately), we consider the sequence of the weak-
est bonds. The average failure voltage for the N th weakest bond to break can be
shown to be (Kahng et al., 1988)
wN
V1 = vweak (N ) = v− + , (42)
L2
which predicts a linear dependence of V1 on N , since the distribution of the thresh-
olds is uniform and must be equal to v− for N = 0 and to v+ for N = L2 . We
now suppose that N bonds have failed and formed 2N edge bonds, where there
is an increase of the current due to enhancement effect (see above). It can then be
shown that the average failure voltage for the 2N failed bonds is given by
w
V2 = vedge (N ) = v+ + . (43)
2N + 1
Observe that V2 is a decreasing function of N, because as N increases, the proba-
bility that a weak bond is included in the 2N edge bonds increases. An approximate
criterion for brittleness of the system is then given by
EV1 (N) > V2 (N ), (44)
where E is the enhancement factor described earlier. Then, two possible situations
may arise:
(1) If we plot EV1 (N ) and V2 (N ) versus N , the two curves do not cross each other.
In this case, the network becomes unstable (behaves as in the brittle regime)
after the first bond fails, regardless of the network size L. For this to happen,
224 5. Electrical and Dielectric Breakdown: The Discrete Approach

one must have Ev− > v− + w = 1 + 12 w, and therefore,

E −1
wc(1) = 2 . (45)
E +1
For example, as mentioned above, for the square network, E = 4/π , and there-
(1) (1)
fore wc  0.24. For w < wc the effect of the randomness is trivial, since the
minimum voltage to break the first bond is just v− = 1 − 12 w, which generates
a voltage Ev− at its edge.
(2) In the second case, the curves EV1 (N ) and V2 (N ) do cross each other. The
crossing point defines the critical value Nc of N for failure of the system.
During breakdown of the first Nc bonds, the system is stable and behaves in
the ductile regime, but it becomes unstable beyond Nc and fails. However, if
L is small enough, then the system may stay in the ductile regime.
The mean failure voltage (per bond) Vf was also determined by Kahng et al.
(1988). In the brittle regime, one has
αw
Vf = v− + 2 , (46)
L
where α is a constant. Since Vf can never be less than v− , Eq. (46) indicates clearly
that by increasing L the system will always eventually behave as brittle. For the
ductile regime, we have
Vf ∼ (ln L)−y , (47)
where y  0.8 for 2D systems. Equation (47) was confirmed by the numerical
simulations of Leath and Duxbury (1994).
Two points are worth mentioning here. One is that the qualitative features of
the above results hold for a large class of voltage thresholds (see, for example, de
Arcangelis and Herrmann, 1989). However, Stephens and Sahimi (1987) (see also
Chan et al., 1991) showed that, if the conductances of the bonds are distributed
according to a probability density function, and if this function is of power-law
type, then many of the above results do not hold, and the problem is more complex.
The second point is that these qualitative features are also observed in discrete
models of mechanical fracture, and in fact, prior to Kahng et al. (1988), had been
predicted by Sahimi and Goddard (1986), who were the first to propose a class of
discrete models for mechanical fracture.

5.2.7 Dynamical and Thermal Aspects of Electrical Breakdown

All the breakdown models discussed so far are quasi-static models, since they do
not have an explicit time scale built in them. However, time-dependent effects
in breakdown phenomena are very important. In particular, a highly important
characteristic of a conducting material is its failure time, i.e., the time that it takes
to suffer breakdown and become insulating. Similar to failure current and failure
voltage, failure time is also not a self-averaged property of a material, as nominally
identical samples exhibit completely different failure times. In practice, what is
 5.2. Discrete Models of Electrical Breakdown 225

usually done is to select a priori a distribution, such as log-normal or the Weibull

distribution, and fit the experimental data for the failure time in order to estimate
the distribution’s parameters (see, for example, Ohring, 1998). However, failure
time data measured in given test conditions are often sufficiently well fitted by
several distributions, with the drawback that different distributions may predict
widely different failure times when they are extrapolated to a specific application,
hence resulting in serious error. In addition, due to cost limitations, the number
of samples tested usually represents only a small fraction of the entire ensemble,
and therefore there may be significant uncertainties in the estimated values of the
distributions’ parameters. Therefore, a dynamic model that can provide accurate
predictions for failure times and other dynamical properties is of considerable
interest. Another important dynamical aspect of the problem that has not been
discussed so far is the behavior of the material in an AC field, whereas use of an
AC field in experiments is very common. The question is, how does a material
suffer electrical breakdown if one applies an AC voltage across it?
In addition, all the results presented so far have been derived based on purely
geometrical considerations, whereas thermal (Joule) effects are in fact the main
driving force for electrical breakdown of composite materials. The purpose of this
section is to address these issues.

5.2.7.1 Discrete Dynamical Models

A few dynamical models have already been developed. We describe and discuss
three of these models, one of which is deterministic, while the other two are
stochastic. The deterministic model is due to Sornette and Vanneste (1992) and
Vanneste and Sornette (1992) (see also Sornette and Vanneste, 1994), which is a
generalization of the fuse model of de Arcangelis et al. (1985b) described above.
In their model, the temperature T of each conducting bond at time t satisfies the
following equation
dT
Cp = Ri b − aT , (48)
dt
where Cp is the specific heat of the material at constant pressure, R is its resis-
tance, i is the current in the bond, and a and b are two constants. The Ri b term
accounts for a generalized Joule heating of the bond (b = 2 for real fuses), while
aT represents the heat lost to the substrate. To each conducting bond a critical
temperature Tc is assigned, such that the bond burns out and becomes an insulator
once its temperature exceeds Tc . A current I is injected into the system, and the
current distribution throughout the network is calculated. Each bond’s current is
then used in Eq. (48) to calculate the time evolution of its temperature. The first
bond burns out when its temperature reaches Tc . The current distribution in the new
network is calculated and the next bond is allowed to burn out. Thus, one essential
assumption of the model is that, the redistribution of the currents in the network is
either instantaneous or happens much faster than the temperature evolution of the
bonds. The limit b → ∞ corresponds to the fuse model described earlier, because
in this limit only the bond that carries the largest current is heated significantly
226 5. Electrical and Dielectric Breakdown: The Discrete Approach

and reaches its critical temperature faster than any other bond. The opposite limit,
b → 0, corresponds to a percolation model, because in this limit the heating rate
becomes independent of the current, and therefore the sequence in which the bonds
burn out is essentially random. Note that there are two characteristic time scales
in the system which are t1 = Tc /Ri b and t2 = 1/a. If Ic is the critical current for
the emergence of the first sample-spanning cluster of the burnt-out bonds, then,
three distinct regimes can be recognized.
(1) If the current I through the network is very close to Ic , then one has a number
of growing clusters of burnt-out bonds, all nucleating from the same center,
which is the first burnt-out bond in the network. The degree of branchiness of
the clusters depends on the quenched disorder of the network (for example,
the distribution of the resistances). The larger the disorder in the network, the
more branched the clusters are.
(2) If I  Ic , then there is only one relevant time scale, t1 , in the system. Initially,
the bonds burn out more or less at random, a process that is dominated by the
quenched disorder, and then at later times the growth of the burnt-out clusters
becomes correlated as they become connected.
(3) The third regime corresponds to a crossover between (1) and (2). In this case,
the behavior of the system is extremely sensitive to the applied voltage or
current. The model produces a hierarchy of evolving failure patterns at various
length scales, as the applied current I is varied. The breakdown patterns are
also fractal with a fractal dimension Df which is a strong function of the
parameter b. Experimental realization and confirmation of this model will be
described and discussed shortly.
A stochastic model that takes into account the Joule effect was developed by Pen-
netta et al. (2000), which was intended for electrical breakdown of thin conducting
films. An external current I , which is held constant, is injected into a 2D lattice.
Each bond of the network is a resistor with a resistance r(T ) = r0 [1 + α(T − T0 )],
where r0 is a constant resistance, T is the resistor’s present temperature, T0 is a con-
stant reference temperature, and α = (1/r)dr/dT is the temperature coefficient
of resistance. A bond breaks down and becomes an insulator with a probability pb
given by
$ %
H0
pb = exp − , (49)
kB T

where H0 is an activation energy characteristic, and kB is the Boltzmann’s constant.

The temperature in the j th resistor is updated according to the following equation
 
a2

N
Tj = T0 + a1 rj ij2 + (rk ik − rj ij ) ,
2 2
(50)
N
k=1

where ij is the current in, and N is the number of nearest neighbors of, the j th
resistor. The parameter a1 describes the heat coupling of each resistor with the
 5.2. Discrete Models of Electrical Breakdown 227

substrate to which the thin film is attached, and measures the importance of Joule
heating effects. a2 is a constant which was taken to be 3/4.
Hence, starting from a resistor network in which all the bonds are conducting,
the current and temperature distributions in the network are calculated. Conducting
bonds are then converted to insulating ones with a probability given by Eq. (49).
The current and temperature distributions are then recalculated, the next bonds to
fail are identified, and so on. The simulations stop when a sample-spanning cluster
of the failed bonds is formed. Computer simulations indicated that the effective
resistance Re (t) of the sample at time t follows the following power law,

Re (t) ∼ (t − tf )−µd , (51)

where µd  1/4. Note that the failure time tf can be estimated from two
measurements of Re (t) at two different times, namely,
ct1 − t2
tf = , (52)
c−1

where c = [Re (t1 )/Re (t2 )]1/µd represents the ratio of the two measured resistances
at two different times, raised to the power 1/µd . Therefore, once again, the concepts
of scaling and universality seem to be quite useful to modeling of an important
phenomenon, namely, electrical breakdown of thin solid films. Let us mention
that another deterministic model that takes into account the Joule effect, but uses
nonlinear, power-law, resistors (see Chapters 2 and 3) was developed by Martin
and Heaney (2000).
The second stochastic dynamical model that we describe was developed by
Hansen et al. (1990), and is a generalization of the dielectric breakdown model of
Niemeyer et al. (1984) which will be studied shortly, but also has some similarities
to the fuse model of de Arcangelis et al. (1985) described above. In their model, a
conducting bond breaks down and becomes an insulator with a probability pb ∼
η
iij , where η is a parameter of the model, and iij is the current in the bond ij .
Initially, all the bonds in the network are conducting. A macroscopic voltage drop
is applied to the network, and the current distribution in the bonds is computed.
The bond that breaks first is selected from among all the conducting bonds. The
current distribution in the network with its new configuration, including the failed
bond, is calculated, the next bond to be broken is selected, and so on.
This model provides some interesting predictions. Hansen et al. (1990) found
that there is a critical value ηc = 2 of η, such that the breakdown patterns are
qualitatively different for η < ηc and η > ηc . For η < ηc the breakdown pattern
resembles a percolation cluster, in the sense that a finite fraction of the conducting
bonds must breakdown before the system fails and becomes insulating. On the
other hand, for η > ηc the breakdown pattern is a fractal object with a fractal
dimension that depends on η. The vector analogue of Hansen et al.’s model, i.e.,
one in which the bonds represent elastic elements that break with some probability
(which might be applicable to mechanical fracture), was analyzed in detail by
Curtin and Scher (1991,1992).
228 5. Electrical and Dielectric Breakdown: The Discrete Approach

5.2.7.2 Breakdown in an AC Field: Thermal Effects

Suppose that the initial resistance of a sample material is R0 . If a current I is
injected into the material, its resistances will change by R = R0 αT , where α
is the temperature coefficient of the resistance, and T is the temperature rise in
the sample as a result of injecting the current into the material. Since T ∼ R0 I 2 ,
one obtains R ∼ (R0 I )2 . Corresponding to the current I there exists a voltage
V across the material which is given by
V = R0 I + cT02 I 3 , (53)
where c is a constant. Then, if I = I0 cos(ωt) = i0 cos(2πf t), the voltage V
becomes
V = R0 I0 cos(ωt) + V3f cos(3ωt), (54)
where V3f ∼ RI0 is the third harmonic voltage. The third harmonic coefficient
(THC) B is then defined by
V3f
B= . (55)
I03
As discussed by Dubson et al. (1989), the THC results from local Joule heating.
Therefore, in effect B measures the local temperature rise at the hot spots that are
developed as a result of Joule heating.
If the material is a two-phase composite a fraction p > pc of which is conducting
and the rest is insulating, then, as was pointed out by Yagil et al. (1992,1993), the
failure current If is related to the THC B. Yagil et al. (1992) suggested that
breakdown occurs when a hot spot in the material reaches the melting temperature
Tm of the metallic (conducting) grains, at which a weak link in the system breaks
down, an irreversible change occurs in the material, and its resistance is modified.
To derive the relation between If and B (Yagil et al., 1992), one notes that the
temperature rise due to a weak link with resistance r0 and current i is T = r0 i 2 R,
where R is the ratio of the temperature rise and the dissipated power at the hot
spot. The resulting change in the local resistance is δr = r0 αT , where α is the
temperature coefficient of resistance. If one applies an AC current, I = I0 cos(ωt),
to the material, it results in the generation of a third harmonic voltage component
V3f , given by
1
2
V3f = ij δrj , (56)
4I0
j

where the sum is over all the hot spots. If we assume that the resistance r and the
ratio R are the same for all the links (in the percolation material), we obtain
αr 2 R
4
B= ij (57)
4I04 j

which implies that B is related to the fourth moment of the current distribution in
the material, a subject that was discussed in Section 5.16 of Volume I. The current in
 5.2. Discrete Models of Electrical Breakdown 229

each resistor of an L × L resistor network is I /L (where L is measured in units of

the bonds). For a resistor network near the percolation threshold, the current in the
red bonds (i.e., those that, if cut, would break the backbone into two pieces) is much
larger than the rest of the bonds. Since near pc the resistance follows the power law
(p − pc )−µ , and the fourth moment of the current as (p − pc )−2κ (see Chapter 5
of Volume I), the third harmonic follows the power law (p − pc )−(2µ+κ) , where,
as discussed in Chapter 5 of Volume I, the exponent κ is independent of all the
percolation exponents.
As discussed in Chapter 5 of Volume I, one may obtain upper and lower bounds
for the exponent κ. Similar ideas can be used for deriving a bound for B (Yagil
et al., 1992). Consider, for example, deriving a lower bound to B for a thin (2D)
film. The bound is obtained by taking into account only the red bonds. For L  ξp ,
where ξp is the correlation length of percolation, the number Mr of the red bonds
follows the power law, Mr ∼ (p − pc )−1 , and the current through each of such
bounds is Ir = (ξp /L)I0 . Therefore,
1
B≥ αr02 RMr Ir4 . (58)
4I04
On the other hand, the average AC component of the temperature increase in each
of the red bonds is, Tr = 12 r0 RI02 , and thus
1
B≥ αr0 Tr (p − pc )−1 . (59)
2I02
Suppose now that Tm is the temperature rise that the material needs to reach
the melting temperature of its conducting portion. If one defines the failure or
breakdown current If as the current at which the melting temperature is reached,
then
&
1
If ≥ αr0 Tm (p − pc )−1/2 B −1/2 . (60)
2
The THC for the pure material (with no insulating region) is given by, B0 =
4 αr0 R/L , while its failure current is, If = L(2Tm /r0 R)
1 2 2 0 1/2 . Since, (p −

pc ) = (B/B0 )
1/2 1/2(2µ+κ) , one obtains the final result:
If0
If ≥ 1/2(2µ+κ)−1/2
B 1/2(2µ+κ)−1/2 . (61)
B0
If we substitute the 2D lower bound, κ = 2ν + 1 − 2µ (see Chapter 5 of Volume I),
we obtain If ∼ (p − pc )ν , in agreement with Eqs. (24) and (28). Thus, taking the
thermal effects into account, one obtains a refinement to Eqs. (24) and (28) which
were derived earlier based on geometrical considerations alone. Since,
 typically,
only a fraction of the red bonds contribute significantly to the sum i 4 , we expect
to have
1 1 1
If ∼ B −w , with − ≤w≤ . (62)
2 2(2µ + κ) 2
230 5. Electrical and Dielectric Breakdown: The Discrete Approach

We are now ready to compare the above theoretical predictions to the relevant
experimental data.

5.2.7.3 Comparison with the Experimental Data

An experimental realization of the dynamical model of Sornette and Vanneste
(1992) was provided by Lamaignere et al. (1996). In their experiment, insulating
epoxy resin was mixed with spherical carbon microbeads. The matrix was ob-
tained by heating the solution for 2 hours, yielding a conducting composite with
quenched disorder. The I − V characteristic of the composite is linear when the ap-
plied voltage is small, V < V1 , signifying the fact that the connectivity properties
of the composite are independent of the voltage V . For V1 < V < Vc , where Vc is
the critical threshold, the I − V curve bends over and the tangential conductivity
decreases, indicating a significant change in the connectivity of the beads which
is the result of local breakdown caused by Joule heating. If the volume fraction of
the beads is above the percolation threshold, and if the temperature of the system
is in the range 20 − 30◦ C above 120◦ C, an additional factor decreases the conduc-
tivity of the composite. This factor is due to the thermal expansion of the polymer
matrix that entails strain growth, leading to a redistribution of the stress field and
modification of the connectivity, and thus the conductivity. Beyond Vc and its cor-
responding current Ic the tangential conductivity vanishes, and I deceases as V
increases. For I ≥ Ic macroscopic breakdown occurs. These data are summarized
in Figure 5.3. As I ≥ Ic increases the resistance of the composite also increases,

Figure 5.3. I − V characteristics obtained under applied voltage (circles) or current

(squares). Dashed line indicates the linear I − V behavior, while the thick solid line indicates
the critical value of the current (after Lamaignere et al., 1996).
 5.2. Discrete Models of Electrical Breakdown 231

signaling the breakdown of more and more conducting fraction of the composite.
Suppose that at time tf the composite fails and becomes insulating. Lamaignere
et al. (1996) found that
tf ∼ I −2 , (63)
and that the effective conductivity ge of the composite at times close to tf follows
the power law,
ge ∼ (tf − t)µd , (64)
with µd , which is a sort of dynamical analogue of the percolation conductivity
exponent µ, being about 2/3 for their 2D material.
Yagil et al. (1992,1993) measured failure current If and the THC of thin, semi-
continuous Ag and Au percolating films. The films were evaporated in vacuum at
a rate of 0.1 nm/sec onto room temperature glass substrate. Several samples with
different surface coverage (i.e., different fraction of the conducting material) were
employed. The samples were then removed from the vacuum and measured at
room temperature. The measured I − V characteristic indicated Ohmic behavior
at low currents, and nonlinear behavior at high currents, due to Joule heating. The
failure current If was defined as the current at which the first irreversible change
in the resistance was measured. Figure 5.4 presents a sample of their results for
the failure current If versus the THC B, measured for the Ag samples.

Figure 5.4. Failure current If versus the third harmonic coefficient B, indicating the slope
w (If ∼ B −w ). The inset presents the data for the relation, B ∼ R 2+w with a slope of 3.2
(after Yagil et al., 1992).
232 5. Electrical and Dielectric Breakdown: The Discrete Approach

The 2D value of the exponent w defined by (62) is bounded in 0.36 ≤ w ≤ 0.5,

if we use µ  1.3 and κ  1.12. Furthermore, if Tm = 103 K, r0 = 1 , and α =
10−3 K−1 , one obtains, 12 αr0 Tm  0.5, and If = If0 (B/B0 )−w with If0 B0w 
0.7, which is in good agreement with the measured value for both the Ag samples,
w = 0.48 ± 0.05 and If0 B0w  0.4, and for the Au materials, w = 0.41 ± 0.01 and
If0 B0w  0.6.
The experiments of Yagil et al. (1992,1993) shed light on the mechanism of
electrical breakdown of composite materials. If the initial material has a low re-
sistance, the breakdown usually results in an insulating composite, implying that
all the links or red bonds that carry high currents burn out and become insulators.
Applying a high voltage, on the other hand, protected by a very low current limit,
causes the material to become reconnected again and produce a composite with a
high resistance and a very low failure current, indicating that only a few of the red
bonds were re-established. Thus, such a material is dominated by the red bonds.
On the other hand, according to Yagil et al.’s experiments, the breakdown of a
high resistance material may result in higher, lower, or infinite (insulating) resis-
tance, implying that a few red bonds are either burnt out, established (dielectric
breakdown), or improved (i.e., their width increases).

5.3 Electromigration Phenomena and the Minimum

Gap
A dynamical model of electromigration was proposed by Bradley and Wu (1993)
and Wu and Bradley (1994) which was intended for electromigration failure in
polycrystalline metal films. In their model each bond of a lattice is either a con-
ducting wire with probability 1 − p or an insulator with probability p. Suppose
that a certain mass mw leaves a wire before it fails. The mass flux jm in the wire
is given by
ρD ∗
jm = Z eE (65)
kB T
where ρ is the atomic density, D is the diffusivity, Z ∗ e is the effective charge, kB
is the Boltzmann’s constant, and E is the electric field. The total mass mw out of
the wire is proportional to the magnitude of the current I , mw = a(T )I , where
a(T ) is a temperature-dependent constant. Since Joule heat in the wire is rapidly
conducted away by the substrate, one can assume that the temperature of the wire
and that of the substrate are equal. The lifetime t of the wire is given by
 t
mw
|I (t)|dt = = q(T ). (66)
0 a(T )
Therefore, once a charge q(T ) has flowed through the wire, it fails irreversibly
and becomes an insulator. To a good degree of approximation, the charge q(T )
has an Arrhenius-type temperature dependence. Thus, the essentials of this model
are as follows. A macroscopic voltage is applied to the network and the current
 5.3. Electromigration Phenomena and the Minimum Gap 233

distribution in it is calculated. The wire that carries the most current fails first, after
which the current distribution is calculated again, the next wire to fail is identified,
and so on.
Several interesting results emerge from this model. For example, suppose that
at time t = 0 a “crack” (i.e., insulating material) of length 2c is inserted in the
metal film and its growth is monitored. Suppose also that v∞ (x, c) is the speed
of the crack tip when the crack’s length is 2x. If a constant external current flows
through the film, then for x  c (Wu and Bradley, 1994)
i0 2
v∞ (x, c)  x , (67)
c
where i0 is the current density far from the crack. Thus, as the crack grows, its
speed of propagation increases quadratically. The dependence
4 x on the time t of the
crack tip location for x  c is obviously found from t = c dz/v∞ (z, c). Near
pc the mean failure time tf obeys the following power law
tf (p) ∼ (pc − p)ν , (68)
where ν is the exponent of percolation correlation length; clearly, tf = 0 for
p ≥ pc .
Electromigration motivates the introduction of a new percolation quantity, which
is called the minimum gap. Consider a random resistor network in which a fraction
(1 − p) of the bonds are insulating. Suppose now that a random walker starts its
walk from one side of the lattice, and jumps from one cell to an adjacent cell by
crossing the bonds, regardless of whether these bonds are conducting or insulating.
We also assume that the walk is self-avoiding, i.e., the walker never visits a cell
more than once. After some steps, the walker finally arrives at the opposite face
of the network; its path consists of all the bonds that were visited. Suppose then
that the path consists of Nc conducting and Ni insulating bonds. The connection
between this concept and electromigration becomes clear if we assume that, any
bond that is crossed by the walker breaks down and becomes an insulator. Thus,
in a 2D system, for example, when the walker has crossed the sample, the system
breaks down and becomes an insulator. The shortest path is one that corresponds
to the smallest number of resistors that burn out during the walk.
We now introduce the concept of minimum gap gm which, in an insulating
material, is the minimum number of conducting bonds (per length of the system)
that must be added to the system (or to the trail of the random walk) in order for
the material to become conducting. Clearly, gm depends on p, the fraction of the
conducting bonds already in the material. Chayes et al. (1986) and Stinchcombe
et al. (1986) studied the properties of the minimum gap gm (p). Figure 5.5 present
the dependence of gm (p) on p in the square network. For p  1, the minimum
gap decreases from 1, with the slope dgm /dp  3 in the square network. Near pc ,
the minimum gap vanishes according to the power law,
gm ∼ (p − pc )ν . (69)
Thus, Eqs. (68) and (69) suggest that the failure time is proportional to the mini-
234 5. Electrical and Dielectric Breakdown: The Discrete Approach

Figure 5.5. Dependence of the minimum gap gm , normalized by the linear size of the
square lattice, on the fraction p of the conducting bonds (after Manna and Chakrabarti,
1987).

mum gap gm (p) of the network. On an intuitive ground, the relation between the
minimum gap and the time to failure in the electromigration problem is expected.
A problem related to electromigration phenomenon is one in which the line
width of the metallic interconnects is comparable to, or smaller than, the grain size
of the film. In this case, referred to as the bamboo regime, the grain boundaries
no longer provide connected diffusion paths along the conductor line. Instead,
electrical breakdown occurs due to intergranular voids which nucleate at the edges
of the line, migrate in the current direction, and finally collapse into a slit which
disconnects the conductor.
This problem was studied in detail by Schimschak and Krug (1998), and later
by Mahadevan et al. (1999), whose analysis we briefly describe. The shape of the
void changes due to the current I along its inner surface. Two factors contribute
to the current, the electromigration and capillary smoothing. Thus, one writes
 
∂Y(L)
I =γ σ + qE(L) , (70)
∂L
where γ and σ are, respectively, the atom mobility and the surface tension, L is
the arc length along the surface, Y is the surface curvature, q is the charge, and E
is the tangential local electric field. Because of conservation of the void area (in
 5.4. Dielectric Breakdown 235

2D), the inner surface must move with a normal velocity vn which is given by
∂I
vn + = 0. (71)
∂L
Due to the growth of the void, this is a moving boundary-value problem, the
numerical solution of which is typically difficult to obtain.
One must first determine the electric field E by solving the Laplace’s equation
in the domain outside the void, subject to the boundary conditions that the normal
electric field vanishes at the void surface, and a constant electric force E0 is applied
to the system far from the void. It is√not difficult to see that the only relevant
length scale in the problem is s = σ/(qE0 ), and therefore the natural time
scale is given by, ts = 4s /(σ γ ), with which the governing equations can be made
dimensionless. After determining the distribution of the electric field, Eq. (71) is
iterated. A breakup procedure is triggered if two points that belong to different
surface segments are closer than half the distance between neighboring points
along the surface. In a similar way, merging of two voids can be treated.
Numerical simulations of this model indicated that, typically, the void disin-
tegrates at long times by one of the two routes. If the void is initially elongated
along the current direction, then, a protrusion develops at the leading end of the
void, which subsequently forms a daughter void. Because the daughter void is
smaller than the initial void, it moves more rapidly ahead of the mother void. If,
on the other hand, the void is initially deformed perpendicular to the current, an
invagination develops which eventually splits the void horizontally.

5.4 Dielectric Breakdown

We consider a heterogeneous material, consisting mostly of an insulating (dielec-
tric) phase, in which a conducting material has been dispersed. The (volume)
fraction of the conducting phase is p < pc , so that, macroscopically, the mate-
rial is insulating. The electric field E and its corresponding displacement field
D = (r)E(r) satisfy the usual equations that we have used so far in this book:
∇ · D = 0, ∇ × E = 0, (72)
where, as usual, (r) is the dielectric constant of the insulating phase.

5.4.1 Exact Duality Relation

The duality relations described in Chapter 2, and also in Chapters 4 and 5 of Volume
I, can also be used here to relate the problem of dielectric breakdown in 2D to the
electrical breakdown in 2D (see, for example, Bowman and Stroud, 1989). With
r = (x, y), Eq. (72) implies that
   
∂ ∂φ ∂ ∂φ
(r) + (r) = 0, (73)
∂x ∂x ∂y ∂y
where the potential φ is defined such that, E = −∇φ.
236 5. Electrical and Dielectric Breakdown: The Discrete Approach

Consider now the dual of the 2D material which is obtained by replacing the
conducting phase by the insulating material and vice versa. We also assume that
the conductivity g of the formerly-insulating parts is given by, g = 1/. The dual
material is conducting since the original material was assumed to be insulating or
dielectric, and therefore the current I must satisfy the continuity equation, ∇ · I =
0, because of which one can write, I = ∇ × ψ, where the potential vector ψ is
selected such that only its z-component ψz (x, y)
= 0. As I = g(r)E, we must have
   
∂ 1 ∂ψz ∂ 1 ∂ψz
+ = 0, (74)
∂x g(r) ∂x ∂y g(r) ∂y
or,
   
∂ ∂ψz ∂ ∂ψz
(r) + (r) = 0. (75)
∂x ∂x ∂y ∂y
In view of Eq. (73), we see that the conductivity problem in the dual material is
identical with the dielectric problem in the original composite, if ∂ψz /∂x = ∂φ/∂x
and ∂ψz /∂y = ∂φ/∂y. If so, one has, Ix = ∂ψz /∂y = Ey and Iy = −∂ψz /∂x =
−Ex . Therefore, the magnitude of the current density I in the dual material is
equal to that of E in the original composite, but its direction is rotated by 90◦ from
the dielectric problem. Physically, while in the electrical breakdown problem the
current is zero inside an insulating inclusion, in the dielectric breakdown problem
the electric field is zero inside a conducting region. Moreover, the regions that
experience an enhancement of the current (in the electrical breakdown problem)
are perpendicular to those that feel the enhancement of the electric field (in the
dielectric breakdown problem). The conclusion is that, in 2D, most of the results
that were described above for the electrical breakdown problem can be immediately
translated to corresponding predictions for the dielectric breakdown problem. We
will discuss this important point shortly, but let us first describe discrete models
of dielectric breakdown.

5.4.2 Stochastic Models

The main stochastic model of dielectric breakdown was proposed by Niemeyer
et al. (1984). In their model, the central site of a square lattice was designated as
one of the electrodes, while the other electrode was placed on a circle at a large
distance from the center. The rules of the model were as follows.
(1) The electric potential distribution in the lattice is obtained by solving the
Laplace equation for V , ∇ 2 V = 0, with the boundary conditions that V =
V0 = 0 for all the sites that belong to the dielectric pattern, and V = V∞ = 1
outside the external circle.
(2) At each step one bond suffers dielectric breakdown and is added to the devel-
oping dielectric pattern. The failing bond is selected from amongst those that
are at the interface between the dielectric pattern and the rest of the system,
 5.4. Dielectric Breakdown 237

with a breakdown probability pb given by

η
pb ∼ Vij , (76)

where Vij = Vi − Vj is the potential or voltage difference between sites i and

j of the interface bond ij , with i being on the interface and j outside of, but
next to, the interface. Since Vi = 0, Vij is simply the potential Vj at j , and
is proportional to the current in the bond ij . In this model, η is an important
parameter, so much so that this model is popularly known as the η-model.
(3) After a bond suffers breakdown, the potential distribution in the system with
its new configuration is recalculated, a new bond is selected for breakdown,
and so on.
Niemeyer et al. (1984) showed that their model leads to fractal breakdown pat-
terns which, for η = 1, are similar to diffusion-limited aggregation (DLA) model
of Witten and Sander (1981) (for a review of aggregation models see Meakin,
1998), who had already pointed out the similarity between their model and the
breakdown patterns. To see the similarity between the two models, let us describe
briefly the DLA model.
In the DLAmodel one starts with an occupied site (the “seed”) of a lattice, located
either at the center of the lattice or on its edges. Random walkers are released, one
at a time, far from the seed particle and are allowed to move randomly on the
lattice. If they visit an empty site adjacent to an occupied one, the aggregate of
the occupied sites advances by one site and absorbs the last site visited by the
walker (in effect one bond is added to the aggregate). The walker is removed,
another one is released, and so on. After a large number of particles have joined
the aggregate, it takes on a disordered structure with many branches, very similar to
the dielectric pattern with η = 1. To see the analogy between the two models, note
that the original seed particle represents the point at which dielectric breakdown
starts. Since the particles perform their random walks on the empty sites, the
probability P (r) of finding them at a position r in this region satisfies the Laplace’s
equation, ∇ 2 P = 0, the same as the governing equation for the nodal potentials or
voltages in the dielectric breakdown model. Because the walkers never move into
the aggregate, the probability of finding them there is zero, P = 0, the same as the
boundary condition, V = V0 = 0 in the dielectric breakdown model. Finally, the
probability with which the aggregate grows is proportional to the flux of particles
between the empty region and the aggregate front, i.e., ∇P  Pi − Pj , the same
as Eq. (76) in the limit η = 1.
In Niemeyer et al.’s model, the fractal dimension of the dielectric pattern depends
on η. In 2D one has Df  2.0, 1.9, and 1.7 for η = 0., 0.5, and 1.0, respectively.
The resulting 2D pattern for η = 1.0 is very similar to a Lichtenberg figure. Earlier,
Sawada et al. (1982) had used a similar model, except that they had assigned a
priori a larger probability for the growth of the tips with respect to side branching.
This is, however, not realistic as the discharge pattern depends non-locally on the
potential distribution throughout the system, which in turn is controlled by the
distribution of the heterogeneities in the material.
238 5. Electrical and Dielectric Breakdown: The Discrete Approach

However, Niemeyer et al.’s model does not have an explicit rule for breakdown.
A bond with even a small probability pb can break down, which is not realistic.
Moreover, the physical reason for Eq. (76) is not clear. Pietronero and Wiesmann
(1988) did attempt to give a theoretical justification for Eq. (76) based on the
time required for the establishment of a filamentary projection of the discharge
as a sort of a “conducting fluid” in a given region of the local field. While their
argument may justify use of Eq. (76), in the limit η = 1, for dielectric patterns
in gases, its generality is not clear, and in addition, whereas the structure of the
simulated discharge patterns is highly sensitive to η (Barclay et al., 1990; Sánchez
et al., 1992), the physical origin or significance of η is not clear. Moreover, the
breakdown patterns in solid materials are propagating damage structures, not the
advancing front of an injected charge “fluid,” as in Niemeyer et al.’s model. As
such, their model is not, in general, suitable for dielectric breakdown in solids.
Wiesmann and Zeller (1986) (see also Noskov et al., 1995) modified the η-
model by incorporating two new features in it. One was that a critical field Vc
for the growth of the dielectric pattern was introduced, such that the breakdown
probability pb is non-zero if Vij ≥ Vc , and pb = 0 otherwise, an assumption that
makes the model somewhat similar to the deterministic models discussed in the
next section. The second feature was the introduction of an internal field Vs in
the structure, such that the potential in it is no longer V0 but V0 + sVs , where s
is the length of the path (measured as the number of sites that it contains) along
the structure which connects the point to the central electrode. The structure of
the resulting dielectric pattern now depends on Vc and Vs . Figure 5.6 shows two
of the fractal patterns generated by this model which are somewhat similar to
treeing in polymers. However, the accumulation of damage, which is known to be
required for electrical tree formation in AC fields, is not allowed in the Wiesmann–
Zeller model, and therefore their model is probably more appropriate for nano-

Figure 5.6. Dielectric trees with the ground plate and the needle voltage V = 0 and the top
plate at V = V0 . The threshold field for growth is zero for the left pattern, and about the
original field at the tip for the right pattern (after Wiesmann and Zaller, 1986).
 5.4. Dielectric Breakdown 239

second impulses. Even then the damage pattern situation is not fractal (Knaur and
Budenstein, 1980), whereas the Wiesmann–Zeller model predicts it to be fractal.
Dissado and Sweeney (1993) argued that fractal tree-like patterns should form only
when the fields at the growth tips can fluctuate around their values obtained from
the solution of the Laplace’s equation. They showed that if one treats the local-field
enhancement factor as a white noise generated by the breakdown mechanism itself,
the amount of branching in the dielectric pattern would depend only on the range
of the fluctuations allowed. Thus, the Wiesmann–Zeller model, though interesting,
is not also completely suitable for modeling dielectric breakdown in solids.

5.4.3 Deterministic Models

Several, very similar, discrete deterministic models of dielectric breakdown have
been proposed over the past decade. These models assume percolation-type dis-
order, and their essential features are as follows. Each bond of a lattice is either a
conductor with probability p or a capacitor (an insulator) with probability 1 − p.
Each capacitor can sustain a fixed voltage drop, say 1 volt, beyond which it breaks
down and becomes a conductor. A macroscopic voltage drop is then applied to
the lattice, and the voltage distribution throughout the lattice is computed. The
capacitor that sustains the largest voltage drop greater than its threshold fails first.
The voltage distribution is then recalculated, the next capacitor to fail is identified,
and so on. If at any stage the applied voltage drop is not large enough to cause
breakdown of any capacitor, it is increased gradually. The simulation stops when
a sample-spanning conducting cluster is formed. The breakdown or failure field
Eb is defined as the minimum external voltage required to cause formation of a
sample-spanning cluster of failed capacitors (conductors), divided by the length L
of the lattice. One important result of this model is that Eb → 0 as p → pc . This
is of course due to the tortuous nature of the percolation cluster near pc . Another
significant prediction of this model is that Eb is smaller for larger lattice, so that
very large samples break down easier than the smaller ones (see also below).
Various versions of this basic model (Beale and Duxbury, 1988) have been stud-
ied, the first of which was probably suggested by Takayasu (1985). In his model,
the resistance of the lattice bonds are distributed randomly. Each bond breaks
down if it suffers a voltage greater than a critical threshold voltage vc . If a bond
does break down, its resistance r is reduced to δr, where δ is a small number.
After a bond breaks down, it remains in that state forever. The breakdown pattern
was found to be fractal with a fractal dimension Df  1.6 in 2D. In the model of
Family et al. (1986), which is essentially a deterministic version of the Niemeyer
et al.’s, the bonds are insulating and carry a breakdown coefficient B which is
randomly distributed in [0,1]. The voltage distribution throughout the lattice is
then computed, with the boundary conditions that V = 0 on the conducting dis-
charge and V = 1 far from the interface between the conducting and insulating
parts. Two versions of the model were investigated. In one model, at each time
η
step an interface bond ij with the largest BVij breaks down, whereas in the sec-
η
ond model an interface bond breaks down with a probability BVij /pbm , where pbm
240 5. Electrical and Dielectric Breakdown: The Discrete Approach

0.20 Figure 5.7. Initial breakdown field Eb in the

square lattice versus the fraction p of the con-
ducting bonds. Squares and circles show the
0.15
results for 50 × 51 and 100 × 101 samples
(after Bowman and Stroud, 1989).
Eb

0.10

0.05

0.00
0.8 0.9 1.0
p/pc

η
is the largest value of BVij among all the interface bonds. The second model is
clearly very similar to the model of Niemeyer et al. (1984). Breakdown patterns
were found to be fractal again, with a fractal dimension that depended sensitively
on η. In the model of Manna and Chakrabarti (1987), each bond or site of the lattice
is either conducting with probability p or insulating (dielectric) with probability
1 − p. After determining the voltage distribution throughout the lattice, all the
insulating bonds or sites break down if the voltage that they suffer is larger than
a threshold voltage. Chakrabarti et al. (1987) and Barbosa and de Queiroz (1989)
studied this model with small-cell position-space renormalization group approach.
Bowman and Stroud (1989) studied the same model, except that in their work the
insulating bond with the largest voltage difference between its end sites breaks
down first. In a somewhat different model, Benguigui (1988) considered the case
in which after a bond breaks down it becomes a superconductor. This was achieved
by inserting light emitting diodes as the insulators in a host of conductors.
The most critical questions in dielectric breakdown phenomenon, that any
reasonable model should be able to address, are as follows.
(1) How does the initial breakdown field Eb (or the corresponding voltage Vi )
depend on the volume fraction p of the conducting material (bonds) in the
initial dielectric material? A typical example is shown in Figure 5.7.
(2) How does the final voltage Vf vary with p? For small p one expects the
final breakdown voltage Vb = Eb L to be different from the initial breakdown
voltage, but as p increases the difference between the two decreases until very
near pc where they are essentially identical. This has an important consequence
in that, when these two voltages are equal, the breakdown proceeds by an
avalanche (see the discussion above) in that, many bonds break down without
any need for further increase in the applied macroscopic voltage drop.
(3) How do the two voltages Vi and Vb depend on the linear size L of the sample?
To understand the importance of the sample size, recall that breakdown starts
 5.4. Dielectric Breakdown 241

Figure 5.8. Path length l(p) of

breakdown versus the fraction of con-
ducting sites in the square lattice.
Symbols are the same as in Figure 5.7
(after Bowman and Stroud, 1989).

near the critical defect of the system, which is (roughly speaking) the largest
pair of strongly interacting conducting clusters which are oriented parallel
to the macroscopic electric field. The breakdown field is of the order of the
inverse of the linear size of the defect, and since the largest defect in a large
system is larger than the largest defect in a small sample, the breakdown field
is smaller in the larger sample.
(4) How does the path length, i.e., the number of bonds in the breakdown path,
vary with p? An example is shown in Figure 5.8.
(5) Do power laws govern the important properties of dielectric breakdown (such
as the breakdown field Eb ) near pc , and if so, are such laws universal?
We now discuss the scaling laws that govern the dependence on p of various
properties of interest near the percolation threshold, and also on the sample size L.

5.4.3.1 Scaling Properties of Dielectric Breakdown

Before discussing scaling properties of dielectric breakdown, let us emphasize a
very important point. Unlike percolation and similar types of critical phenomena,
some of the scaling properties of electrical and dielectric breakdown phenomena
are valid over a wide range of the parameter space, and therefore are very useful
from a practical point of view. For example, as already described and discussed
for electrical breakdown, one can consider, for a fixed p, the scaling properties
of breakdown phenomena in terms of the linear dimension L of the system. Not
242 5. Electrical and Dielectric Breakdown: The Discrete Approach

only are such scaling properties important, but are in fact measured routinely in
practical situations, and therefore a scaling theory of breakdown phenomena in
terms of the sample size L is a very useful tool for interpreting the experimental
data.
The scaling properties of dielectric breakdown phenomena have been studied
extensively. Let us first recall that, as discussed in Chapter 6 of Volume I, the static
dielectric constant 0 follows the following power law (Efros and Shklovskii,
1976) near the percolation threshold pc ,

 ∼ (pc − p)−s , (77)

where s is the critical exponent that characterizes the effective conductivity of

conductor-superconductor percolation composites near pc , utilized extensively in
Chapters 5, 6 and 9 of Volume I. The root mean square Erms of the electric field
is given by Erms = |E|2 1/2 ∝  1/2 |E0 | ∼ (pc − p)−s/2 , where E0 is the applied
electric field on the external surface of the system. The maximum field Em in the
system is certainly larger than Erms . Suppose that Em ∼ (pc − p)−y . Because
Em > Erms , we must have y > s/2 (Bowman and Stroud, 1989), and

Eb ∼ (pc − p)y . (78)

To estimate y, Beale and Duxbury (1988) used an argument based on the idea of
the critical defect mentioned above. Suppose that the total length of the critical
defect (the conducting path), made up of a pair of the largest interacting clusters of
conducting material, separated by a small distance, is . The electric field between
these two clusters is enhanced by a factor of the order of  times the applied
macroscopic field. Far from pc the probability of finding a percolation cluster of
linear size  is given by Eq. (26). The largest cluster in a d-dimensional percolation
system of volume Ld is of the order of m ∼ ξp ln Ld [see Eq. (27)]. Since Eb ∼
1/m , we obtain (Beale and Duxbury, 1988)

(pc − p)ν
Eb ∼ , (79)
ln L
and therefore y = ν, which is certainly greater than s/2. Equation (79) can also
be derived based on the argument (Stinchcombe et al., 1986) that Eb should be
proportional to the minimum gap gm which is proportional to ξp−1 . The ln L term of
Eq. (79) can also be derived from the fact that (Li and Duxbury, 1987) the maximum
current Im in a percolation network of linear size L that leads to its failure is given
by Im ∼ (ln L)ψ , where ψ is the same exponent that appears in (29) and (30)
(for the problem of the largest currents in a random resistor network see also
Machta and Guyer, 1987). Numerical simulations (Manna and Chakrabarti, 1987;
Benguigui, 1988; Beale and Duxbury, 1988; Bowman and Stroud, 1989) seem to
confirm Eq. (79). Lobb et al. (1987) and Chakrabarti et al. (1988) extended this
analysis to the Swiss-cheese model of continuum percolation, in which spherical
or circular grains of dielectric are distributed randomly in a conducting matrix,
 5.4. Dielectric Breakdown 243

and showed that

1
y=ν+ (80)
2
in any dimension. One may also consider the inverted Swiss-Cheese model (see
Chapter 2 of Volume I) in which the metallic grains that can freely interpenetrate
are randomly distributed in a dielectric matrix. For this case Lobb et al. (1987)
showed that
y = ν + 1. (81)
Equations (80) and (81) are both different from y = ν for lattice models, Eq. (79),
and indicate that, as far as dielectric breakdown is concerned, a continuum is
weaker than a discrete system. This is understandable since in a lattice model
the conductivity of the bonds is independent of p, whereas the state (geometrical
configuration) of a continuum depends on p.
The next important issue is the size dependence of Eb . Beale and Duxbury
(1988) proposed that
1
Eb ∼ , (82)
A(p) + B(p) ln L
where A(p) and B(p) are simple functions. If we compare Eq. (82) to Eq. (79), we
infer that B(p) ∼ (pc − p)−ν , and numerical simulations of Beale and Duxbury
(1988) in 2D confirmed this expectation; see Figure 5.9.

5.4.3.2 Distribution of Breakdown Fields

Similar to electrical breakdown of solids, the breakdown field for dielectric
breakdown is not a self-averaged property, because different materials with differ-
ent types of heterogeneity, or even nominally identical materials, have different

Figure 5.9. Breakdown field Eb versus

the linear size L of the square lattice. The
results, from top to bottom, are for p =
0.4, 0.35, 0.25, and 0.1 (after Beale and
Duxbury, 1987).
244 5. Electrical and Dielectric Breakdown: The Discrete Approach

breakdown fields Eb . Therefore, there should be a distribution of such fields for

given values of p and L. In a series of papers, Duxbury and co-workers (Duxbury et
al., 1986, 1987; Duxbury and Leath, 1987; Beale and Duxbury, 1988) derived this
distribution for the dielectric (and electrical) breakdown. The resulting distribution
is very similar to what we derived for the electrical breakdown problem.Asummary
of their derivation is as follows. Suppose that PL (m ) is the probability that no
defect (conducting region) larger than size m exists in a d-dimensional cubic
lattice of volume Ld . We divide the cubic network into smaller cubes of linear
dimension Lc , and assume that the characteristic size of the largest defect is much
smaller than Lc . Then
d
PL (m ) ∼ [PLc (m )](L/Ls ) .
Solving this equation and using the fact that for p  pc and L  ξp the cluster
size distribution of percolation systems is an exponentially decaying function of
, we obtain
 
PL (m ) = exp −cLd exp(−km ) . (83)

Near pc , the cluster size distribution is of power-law type, in which case

PL (m ) = exp(−cLd −m
m ), (84)
which is of the same form as the classical Weibull distribution, and is appropriate
for length scales L  ξp , where m is a constant parameter. Since the breakdown
field is of the order 1/m , the distribution of the breakdown fields is given by
 $ %
k
FL (Eb ) = 1 − exp −cL exp −d
, (85)
Eb
a Gumble distribution which is appropriate for length scales L  ξp . In Eq. (85)
the parameter c depends only weakly on p, and k ∝ ξp−1 . If we now define E1/2
as that value of Eb for which half of the system fails, we obtain
 $ %
1 k
FL (E1/2 ) = = 1 − exp −cLd exp − , (86)
2 E1/2
which, when solved for E1/2 , yields an equation similar to (82) with A(p) =
[ln c − ln(ln 2)]/k ∼ ξp and B(p) = d/k = dξp . The equivalent Weibull forms
are
FL (Eb ) = 1 − exp(−cLd Ebm ), (87)

E1/2 ∼ L−d/m . (88)

To determine which one of the two distributions, Eq. (85) or (87), can fit the exper-
imental data more accurately, we proceed as in the case of electrical breakdown,
namely, we compute the quantities AW and AG , analogous to Eqs. (40) and (41),
and fit the data to them. We note that Sornette (1988) argued that in a continuum
 5.4. Dielectric Breakdown 245

system with percolation-type disorder, such as the Swiss-cheese model, Eq. (86)
is no longer valid. Instead, one has a simple exponential, Weibull-like distribution.

5.4.4 Comparison with the Experimental Data

The above theoretical results have been tested against (at least) two sets of experi-
mental data. Coppard et al. (1989) studied the dielectric breakdown of polyethylene
plaques that contained a fixed volume fraction of aluminum particles. Each plaque
was compression molded to a disc of thickness 0.7 mm with a depressed inner
region of diameter 54 mm. The particles had a well-defined range (53–75 µm),
and were distributed randomly within the polyethylene. The breakdown statistics
were collected by stressing the metal-loaded plaques under a uniform AC field and
ramping the field amplitude at a fixed rate until breakdown took place. Their data
confirmed the validity of Eq. (79), and indicated that Eq. (85) is at least as accurate
as Eq. (87).
Benguigui (1988) and Benguigui and Ron (1994) carried out experiments using
a square lattice of random resistors and light-emitting diodes. The diodes had a
very large resistance up to a voltage threshold Vb , but their resistance decreased
very significantly above Vb , converting them to conductors. The transition between
the two states was relatively sharp, but beyond Vb the voltage across the diodes
remained essentially constant (which is in contrast to the usual insulator-conductor
transition in which the voltage after the transition would be almost zero). The
advantages of using such diodes are that, (1) their breakdown is reversible, and
(2) the breakdown becomes visible as the diodes, after becoming conductors, emit
light.
Suppose that the lattice consists of resistors with fraction p < pc and the diodes
with fraction (1 − p). Figure 5.10 presents the dependence of the voltage Vb on
(pc − p). If we fit these data to Eq. (79), we obtain an exponent y  1.1 ± 0.05,
which reasonably close to the theoretical prediction y = ν = 4/3. The difference
is presumably due to the small size of the lattice (L = 20) used in the study.

Figure 5.10. Failure voltage Vb as a function

of the fraction of the resistors p in a system
with light-emitting diode (after Chakrabarti
and Benguigui, 1997).
246 5. Electrical and Dielectric Breakdown: The Discrete Approach

Summary
Discrete models of electrical and dielectric breakdown of composite solids have
provided very useful insights into the properties of these important phenomena, by
demonstrating the significant role that defects of heterogeneities play in them. In
particular, they have provided the important prediction that the statistics of these
breakdown phenomena depend critically on the volume fraction of the defects or
the broadness of the distribution of the heterogeneities. If the volume fraction of
the defects is low, then, the probability distribution of the failure fields (voltage or
current) is of Gumble type, rather than the classical Weibull distribution. Moreover,
the discrete models have enabled us to obtain important predictions for the effect
of sample size on breakdown properties of heterogeneous materials.
6
Fracture: Basic Concepts and
Experimental Techniques

6.0 Introduction
In Chapter 5 we studied electrical and dielectric breakdown of materials—
phenomena that are well-known examples of nonlinear scalar transport processes
with their nonlinearity manifested by the existence of a threshold in the linear
(or possibly nonlinear) constitutive law that describes the relation between the
flux and the potential gradient. Beginning with this chapter, we study a nonlinear
vector transport process which is of immense significance to materials, and leads
to their mechanical failure. This type of failure, which is a result of nucleation
and propagation of fractures in materials, varies anywhere from brittle fracture,
that represents a nonlinear vector transport process characterized by a threshold
in the otherwise linear elasticity equations that govern the elastic behavior of the
material, to ductile yielding and flow. Such failure phenomena are some of the
most complex sets of phenomena in science and technology. The range of natural
and industrial systems in which mechanical fracture occurs is very broad. Under
a large stress or strain, a crack opens up in soils which grows with time, lead-
ing to complex phenomena such as soil liquefaction and eventually earthquake.
Natural or man-made fractures in oil and geothermal reservoirs and aquifers are
crucial to the flow of oil, heat and vapor, or groundwater, especially in those reser-
voirs that have a very small porosity, such as many oil fields in the Middle East.
Other rock-like materials, such as concrete and asphaltenes, often develop large
fractures, causing considerable damage to highways and buildings. Propagation of
cracks in airplane wings and fuselages can cause an airliner to crash. An important,
and undesirable, property of many high-temperature superconducting materials is
their brittleness and mechanical instability. Polymers, glasses and ceramics often
develop microcracks under a large enough stress or strain which can lead to their
mechanical failure and eventual fragmentation. Composite materials can develop
cracks due to thermal mismatch between their various constituents. Pressurized
nuclear reactors can develop cracks in their structure which can create tremendous
safety problems. Thus, a comprehensive understanding of fracture nucleation and
propagation has tremendous practical implications.
248 6. Fracture: Basic Concepts and Experimental Techniques

6.1 Historical Background

Most of us have been familiar with the phenomenon of fracture of materials since
our childhood, since most of us broke something like a glass or a doll when we
were very young. Even if we did not break anything during our childhood, at least
some of us might have heard a song like the following in a nursery:
Humpty Dumpty sat on a wall
Humpty Dumpty had a great fall
All the king’s horses and all the king’s men
couldn’t put Humpty together again.1
Any child who heard this song in a nursery was in fact introduced to the phe-
nomenon of fracture, without, of course, knowing it. This simple song also points
out two important aspects of fracture phenomenon, namely,
(1) a material develops fracture in response to a driving force which, in the case
of Humpty, was the collapse of the church tower, and
(2) fracture is irreversible, since not even all the king’s horses and men could put
Humpty together after it had been broken into pieces!
The story about Humpty Dumpty also points out another important aspect of
fracture of materials, namely, that because of its huge practical significance, the
development of an understanding of how materials fracture and break has been
of great interest for many centuries, and goes back at least 500 years to Leonardo
da Vinci who studied fracture of iron wires and showed that a long wire breaks
more easily than a short one. That is, long wires are, on average, weaker than short
wires. Today, this behavior is known as the size effect and is a manifestation of the
fact that often fracture is initiated by rare flaws in a material. Since a larger piece
of a material is more likely to contain a rare defect, it is also more likely to break
under an applied force than a smaller piece of the same material.
Marder and Fineberg (1996) presented a delightful discussion of the historical
background of the development of solid mechanics that has led us to the present
continuum fracture mechanics. According to them, this development goes back
to at least Galileo Galilei who was almost 70 years old when he was working
on this subject. His life had been nearly ruined by a trial for heresy before the
Inquisition, when he retired in 1633 to his villa near Florence to construct the
Dialogues Concerning lluo New Sciences. His first science was the study of the
forces that hold objects together and the conditions that cause them to fall apart—
the dialogue taking place in a shipyard, triggered by observations of craftsmen
building the Venetian fleet. His second science concerned local motions—laws

1According to legends, Humpty Dumpty was a powerful cannon that was mounted on top of St.
Mary’s at the Wall Church in Colchester, defending the city against siege in the summer of 1648, during
the English Civil War (1642–1649). The church tower was hit by the enemy, with its top blown off,
hence sending Humpty to the ground.
 6.1. Historical Background 249

governing the movement of projectiles. As we now know, these two subjects have
fared differently over the centuries. The first subject, now known as the strength of
material, is an integral part of the basic education that most engineering students
receive, while the second one has become a core subject that physicists learn at
the beginning of their education. Although now, as in Galileo’s time, shipbuilders
need good answers to questions about the strength of materials, the subject has
never yielded easily to basic analysis. Galileo identified the main difficulty when
he wrote: One cannot reason from the small to the large, because many mechanical
devices succeed on a small scale that cannot exist in great size. Over 350 years
after Galileo wrote these lines science reached the atomic scale and began to
answer the questions that he had posed on the origins of strength and the relation
between large and small. These wise words of Galileo also pointed out an important
aspect of fracture of materials, namely, the fact that this is an inherently multiscale
phenomenon, ranging from atomic to macroscopic length scales. While the vast
majority of the theoretical and computer simulation studies of fracture have been
concerned with only one of these length scales, the past few years have witnessed
development of multiscale modeling approaches to fracture propagation in solid
materials. We will describe such approaches in Chapter 10.
However, huge accidents in the 1800s and the first half of the twentieth century,
that were caused by catastrophic fracture of materials, provided the motivation for
intensive study of fracture phenomena. For example, the boiler of the Soltana, a
steamboat that carried the Union soldiers during the American Civil War, exploded,
resulting in the death of over 1,000 soldiers. In 1919, a molasses tank 50 feet high
and 90 feet wide burst in Boston, killing 12 people and several horses. The court
auditor concluded that, the only rock to which he could safely cling was the obvious
fact that at least one-half of the scientists must be wrong.
One of the most important cases of material fracture in the twentieth century,
that helped to establish the significance of fracture mechanics, occurred during
World War II. Wartime demands for ocean freighters led to the production of the
Liberty ship, the first to have an all-welded hull. Of the nearly 4,700 ships of the
Liberty class launched during the war, over 200 suffered catastrophic failure, some
splitting in two while lying at anchor in port, and over 1,200 suffered some sort
of severe damage due to fractures. The discipline of fracture mechanics emerged
from these catastrophes. The all-welded ships were redesigned, eliminating, for
example, sharp corners on hatches, and systematic procedures were developed for
testing the fracture resistance of materials. In the early 1950s, failure by fracture
cursed the British airline industry’s efforts to establish passenger service using
jet aircraft. Ill-placed rivet holes destroyed two of Britain’s Comet aircraft, and
played an important role in transferring the center of gravity for building civilian
jet aircrafts from Britain to the United States. Aircrafts are now subjected to a
systematic program of inspection that acknowledges that every structure has flaws,
but that flaws greater than a certain size are intolerable. Testing procedures have
continued to evolve in response to accidents, most recently after an incident (in
the 1980s) in which part of the top of the fuselage of an Aloha airliner separated
during flight, killing two people.
250 6. Fracture: Basic Concepts and Experimental Techniques

6.2 Fracture of a Homogeneous Solid

The strength of a material is its ability to resist an applied load without breaking
or changing its shape. Therefore, let us first ask the seemingly simple question,
how does a perfect (defect-free) solid break? To answer this question, consider
a block of material of height h and cross-sectional area S, pulled by a force F .
The block separates into halves when its atoms are pulled beyond the breaking
point. To estimate the force Fc , or the corresponding stress σc , required to reach
the breaking point, we recall that the Young’s modulus Y relates the stress σ on a
material to its extension δh through the relation
F δh
σ =− = Y. (1)
S h
The ideal or cohesive strength of a perfect solid, i.e., the critical stress to reach the
breaking point, is typically
1
σc = Y. (2)
10
If the material is under shear, the same estimate of σc should be used, except
that the Young’s modulus Y should be replaced by the shear modulus µ. Except
for some rather exotic materials, such as micrometer-sized whiskers, however,
most solids have strengths in the range 10−2 Y to 10−4 Y . This lower strength is
caused by various defects, such as vacancies, interstitials, impurity atoms (point
defects), dislocations (line defects), grain boundaries, heterogeneous interfaces,
microcracks (planar defects), chemically-heterogeneous precipitants, twins, and
other strain-inducing phase transformations (volume defects). The defects promote
plasticity and premature fracture (see below). The mechanisms of crack nucleation
that are described below provide insight into the phenomena involved.
However, the lower strength of certain materials, such as silicate glasses, which
represent three-dimensional (3D) covalent networks, cannot be explained by the
above deformation processes, since their microstructure is homogeneous except
perhaps at very small length scales, of the order of 10 nm. In this case, the smooth
surface of the glass, when it comes into contact with another solid material,
produces sub-microscopic cracks, as point contacts generate very large local-
ized stresses that cannot be relieved by plastic (or viscoelastic) deformation. The
severity of the contact also determines the length and distribution of the cracks.
This example demonstrates the fact that, in order to find the best material to
build, for example, a house, it is not enough to simply pull out the Periodic Table
and find the element with the highest bonding strength and melting point, as this
“exercise” will point to diamond, too expensive a material to build a house with! If
one were to use, for example, vitreous mixture of silicon and oxygen, raw materials
that are abundant and safe and form strong bonds, the attempt will again be a failure
as soon as the material is hit with, say, a piece of stone. The failure of the Periodic
Table in telling us which material to use is due to the fact that the relation between
bonding energies and strength of materials is far from direct.
 6.3. Introduction of Heterogeneity 251

6.3 Introduction of Heterogeneity

In most engineering materials (as well as natural materials, such as rock) the pres-
ence of flaws or defects with various sizes, shapes and orientations makes fracture
a very complex phenomenon. In fact, disorder comes into play in many ways
during a fracture process. The effect of even small initial disorder can be enor-
mously amplified during fracture. This makes fracture a collective phenomenon in
which disorder plays a fundamental role. Due to disorder, brittle materials gener-
ally exhibit large statistical fluctuations in their fracture strengths, when nominally
identical samples are tested under identical loading, giving rise to a distribution of
fracture strengths (similar to distribution of the breakdown fields in the electrical
and dielectric breakdown phenomena described in Chapter 5). Because of these
statistical fluctuations, it is insufficient, and indeed inappropriate, to represent the
fracture behavior of a disordered material by only its average properties, an idea
which, as the previous chapters should have made clear, is usually used in mean-
field and effective-medium approximations: Fluctuations are important to fracture
nucleation and propagation and cannot be neglected
The traditional approaches to fracture mechanics (see, for example, Ewalds and
Wanhill, 1986; Freund, 1990; Lawn, 1993) have certainly provided the frame-
work for analyzing a wide variety of phenomena without considering the effect
of disorder. These approaches are based on continuum fracture mechanics, some
of the most important contributions of which will be summarized, described and
discussed in Chapter 7. The basis for most of these traditional approaches is the im-
portant criterion developed by Griffith (1920; see below and also Chapter 7). The
analogue of Griffith’s analysis for the dielectric breakdown problem was already
described in Section 5.1.1. He proposed that a single crack becomes unstable to
extension when the elastic energy released in the crack extension by a small length
dc becomes equal to the surface energy required to create a length dc of crack sur-
face. However, Griffith’s criterion was derived under quasi-static conditions and,
moreover, it is presumably valid for materials that are essentially homogeneous,
so that strong disorder plays no important role. Once the crack begins to move,
the prevailing dynamical conditions render this criterion useless. In addition, the
extension of this criterion to heterogeneous materials, as simple as polycrystalline
ceramics with various crystalline orientations and/or grain boundary energies, is
not obvious.
Accompanying the traditional phenomenological theories has been direct nu-
merical modeling using the finite-element method (FEM). With a combination
of computers and adroit mesh constructions, the stress field of a configuration of
grains, fibers or cracks may be calculated by the FEM. The mesh size of the FEM
must be smaller than the scale on which the stress field is expected to vary, which
is therefore much smaller than the relevant length scale of the disorder. Therefore,
only a small portion of a disordered material can be analyzed using the FEM,
and full calculations must be performed for each of the many local configurations
which are required to understand the statistical nature of the problem. To extend
such small-scale FEM studies to larger length scales is still a formidable, if not
252 6. Fracture: Basic Concepts and Experimental Techniques

impossible, computational problem. We will describe in Chapter 7 some typical

FE simulations of fracture propagation in solid materials.
Such difficulties have inspired further development of continuum mechanics
approach to fracture on one hand, and development of many discrete models of
fracture of materials on the other hand. The discrete models are typically based
on lattices of elastic elements, such as springs and beams. The advantage of such
models is that, at least over certain length scales, they allow disorder to be explicitly
included in the models. We will study such discrete models in Chapter 8. Another
type of discrete model of fracture and failure of materials is based on molecular
dynamics simulations that consider propagation of a fracture at the atomic scale.
We will describe and study this approach in Chapters 9 and 10. Both the lattice
models and the MD simulations have also necessitated use of large-scale computer
simulations.
In the present chapter we lay the foundations for our discussions of fracture
phenomena, and describe the basic concepts that will be employed heavily in the
subsequent chapters. We also describe and discuss the experimental techniques for
measuring the most important properties of interest in facture of materials, so that
when in the subsequent chapters we compare the theoretical predictions with the
relevant experimental data and mention the technique by which the data have been
collected, the reader will have a clear understanding of, and familiarity with, the
technique. Also described in this chapter are the basic features of several important
classes of materials, as they relate to their fracture properties.

6.4 Brittle Versus Ductile Materials

The most important qualitative fact in the mechanical properties of solid materials
is that some are brittle and shatter in response to an external force, while others
are ductile and merely deform in response to the blow. If we take a piece of a solid
material, make a saw cut in it, and pull it, then, if the material is brittle, the tip of
the saw cut sharpens spontaneously down to atomic dimensions and, similar to a
knife blade one atom wide, it slices its way forward. In a ductile material, on the
other hand, the tip of the saw cut blunts, broadens and flows, so that great effort is
required to make the cut progress. The question is, why? Posing the question in a
new guise, we ask, what makes a crack grow and propagate?
There is no completely satisfactory answer to the question of why some mate-
rials are brittle and others are ductile, as the giant stars of the Milky Way Galaxy,
the long-dead true manufacturers of atoms, forgot to specify this property when
writing down their technical specifications. The most well-developed investiga-
tion of this problem considers stationary, atomically sharp cracks in otherwise
perfect crystals, and asks what happens when slowly increasing stresses are im-
posed on them. Rice and Thomson (1974) were probably the first to show how
to estimate whether the crack will move forward in response to such a stress,
or whether, instead, a crystal dislocation (i.e., a line of defects) will pop out of
 6.5. Mechanisms of Fracture 253

the crack tip, causing the tip to become blunt. Brittleness and ductility are not,
in fact, inherent in the atoms that make up a solid. For most solid materials there
is a definite temperature at which they make a transition from brittle to ductile
behavior which for example, is about 500◦ C for silicon. In Chapter 7 we will
briefly describe theories that attempt to predict this transition temperature.

6.5 Mechanisms of Fracture

To understand how a fracture propagates in a solid material, it is essential to under-
stand how a fracture is nucleated. At the atomic level, a crack or fracture is the result
of breaking the interatomic bonds of a material. However, the answer to the all
important question, “when do the atomic bonds break,” is mostly material-specific,
and depends critically on the morphology of a material. Normally, fractures are
generated as a result of a stress or strain imposed on a material which causes its
deformation and breakage of its interatomic bonds. The stress or strain can be
applied externally, or can be generated internally by differential changes within
the material. The cracks in the latter case are usually referred to as the pre-existing
cracks. The differential changes can be caused by a temperature gradient, a trans-
port process such as diffusion, chemical changes and reactions, or by shrinkage.
One must also distinguish between the nucleation of a crack and its propagation.
In some cases, a crack propagates by growing alone, while in other cases the prop-
agation process is the result of coalescence of a multitude of smaller cracks. What
follows is a brief discussion of several mechanisms of deformation of a material
which leads to nucleation of cracks.

6.5.1 Elastic Incompatibility

If a solid material consists of rigid phases or grains, then cracks nucleate at the
interface between the grains (and also in the grains themselves). This is due to the
elastic incompatibility of the neighboring grains, caused by the differences in their
composition and orientations. These differences result in different elastic strains
in the grains, when a stress is applied to the material, leading to formation of local
high-stress areas in the material that can be relieved only by formation of a crack.

6.5.2 Plastic Deformation

First introduced as a mathematical concept in the 19th century, the idea of a dis-
location as a crystal defect was hypothesized simultaneously by Orowan (1934),
Polanyi (1934), and Taylor (1934), mainly to explain the less-than-ideal strength
of crystalline materials. Only much later, in the 1950s, was the existence of dislo-
cations experimentally confirmed (Hirsch et al., 1956). Currently, such ubiquitous
crystal defects are routinely observed by various means of electron microscopy.
254 6. Fracture: Basic Concepts and Experimental Techniques

Low-temperature shear deformation of crystalline materials (e.g., ceramics) oc-

curs by gliding of individual dislocations or the coordinated movement of arrays
of partial dislocations. The shear can be localized in a narrow band which, if it
meets some sort of a microstructural barrier (e.g., a grain boundary or a particle
from another phase of the material), leads to very high local stresses at the band’s
tip, resulting in the nucleation of a crack. The direction of the shear as well as
the location of the crack are both influenced very strongly by the crystal structure
and the strength of the interface between the shear band and the barrier. However,
instead of nucleating a crack, the high stresses can also be relieved by some sort of
generalized plastic deformation. Many materials are unable to relieve high stresses
caused by plastic deformation, and therefore form cracks.
Over the last seven decades, experimental and theoretical developments have
firmly established the principal role of dislocation mechanisms in defining material
strength. It is now universally accepted that the macroscopic plasticity properties
of crystalline materials are derivable, at least in principle, from the behavior of their
constituent defects. However, this fundamental understanding has not translated
into a quantitative theory of crystal plasticity based on dislocation mechanisms.
One difficulty is the multiplicity and complexity of the mechanisms of dislocation
motion and interactions, which leave little hope, if any, for a quantitative analytical
approach. The situation is further exacerbated by the need to trace the evolution
of a large number of interacting dislocations over long periods of time, which is
required for any calculation of plastic response in a representative volume element
of the material.

6.5.3 Coalescence of Plastic Cavities

An operating mechanism for crack nucleation, especially in ductile materials that
contain rigid inclusions, is the coalescence of cavities. When a stress is applied to
the material, the ductile matrix is deformed, with its mechanism of deformation
being either slip (as in crystalline materials) or shear deformation (as in amorphous
materials). The rigid inclusions do not deform, and therefore the interface between
them and the matrix separates, followed by development of plastic cavities around
the inclusions. Further deformation of the matrix forces the cavities to grow. Al-
ternatively, if the temperature of the system is high enough, the cavities grow by
a diffusion process. At some point the local cavities begin to interact with each
other, and eventually merge and form a crack.

6.5.4 Cracks Initiated by Thin Brittle Films

If a strong material is covered by a thin brittle film, fracture of the film can lead
to the fracture of the material itself in the bulk, even if the material is ductile. An
example is a nitride layer on steel. In this case, the deformation of the film causes
its fracture which then propagates at high speeds, penetrating the material itself.
Degradation of the surface of materials can lead to the same effect.
 6.6. Conventional Fracture Modes 255

6.5.5 Crazing
Crack nucleation by crazing occurs in amorphous polymeric materials. When such
materials are deformed by an applied stress, the polymeric chains rotate and, if the
strain is large enough, become aligned in the direction of the maximum extensional
strain. Crazing then involves formation of planar arrays of fine voids that are normal
to the tensile stress. The distance between the voids is filled by ligaments of aligned
polymer chains. If the deformation is strong enough, the ligaments eventually break
and help the voids to merge.

6.5.6 Boundary Sliding

If a material contains rigid blocks (as in polycrystalline materials), and if the
temperature of the system is high enough, then, it is deformed by sliding of the
rigid blocks. The sliding is stopped at the triple point grain corners, and cracks that
are wedge-shaped are formed. In addition, rigid particles can help nucleate plastic
cavities during sliding which then grow, coalesce and form cracks.

6.6 Conventional Fracture Modes

There are three symmetrical ways of loading a solid material with a crack. These
are known as modes, and are illustrated in Figure 6.1. A generic loading situa-
tion produced by some combination of forces without any particular symmetry is
usually referred to as mixed mode fracture. Although understanding mixed-mode
fracture is obviously of practical importance, our focus will primarily be upon the
physics of fracture propagation rather than upon engineering applications. There-
fore, we will restrict our attention to the cases in which the loading has a high
degree of symmetry, but will also briefly discuss the mixed mode case.
The fracture mode that we will mainly deal with in this book is Mode I (opening
mode), where the fracture faces, under tension, are displaced in a direction normal
to the fracture plane. In Mode II (sliding mode), the motion of the fracture faces is
that of shear along the fracture plane. Mode III (tearing mode) fracture corresponds
to an out of plane tearing motion where the direction of the stresses at the fracture
faces is normal to the plane of the sample. One experimental difficulty of Modes
II and III is that the fracture faces are not pulled away from one another, and
thus contact along the fracture faces still occurs. The resulting friction between
the fracture faces contributes to the forces acting on the crack, but its precise
measurement is difficult.

Figure 6.1. The three basic fracture modes.

256 6. Fracture: Basic Concepts and Experimental Techniques

For these reasons, Mode I corresponds most closely to the conditions used in
most experimental and theoretical work on brittle fracture of solids, since there
is always a tendency for a brittle crack to seek an orientation that minimizes the
shear loading. This is consistent with crack extension by progressive stretching
and rupture of cohesive bonds across the crack plane. In 2D isotropic materials,
Mode II fracture cannot easily be observed, because slowly propagating fractures
spontaneously orient themselves so as to make the Mode II component of the
loading vanish near the crack tip (Cotterell and Rice, 1980). Mode II fracture is,
however, observed in strongly anisotropic materials. For example, friction and
earthquakes along a pre-defined fault are examples of Mode II fracture where the
binding across the fracture interface is considerably weaker than the strength of
the bulk of the material. Pure Mode III fracture, although experimentally difficult
to achieve, is sometimes used as a model system for theoretical studies, since
in this case the equations of elasticity simplify considerably. Analytical solutions
obtained in this mode (some of which will be described in Chapter 7) have provided
considerable insight into the fracture process.

6.7 Stress Concentration and Griffith’s Criterion

Inglis (1913) analyzed the stress distribution in a uniformly-stressed plate con-
taining an elliptical cavity at its center. His work, which represents an important
precursor to that of Griffith (1920), showed that the stress around a sharp notch
or corner may be many times larger than the applied stress, hence providing the
important clue that even sub-microscopic voids or flaws can weaken a material.
Most importantly, his analysis established that the limiting case of an infinites-
imally narrow ellipse can be considered as representing a crack. We summarize
Inglis’ analysis here.
Consider a plate that contains an elliptical cavity of semi-axes c and b, which
are small compared to the dimensions of the plate. We apply a uniform tension
σ 0 along the y-axis. The system is shown in Figure 6.2. The cavity’s boundary is
stress-free, and Hooke’s law of linear elasticity holds everywhere in the plate. The
equation for the ellipse is given by
x2 y2
+ = 1, (3)
c2 b2
based on which it is easy to show that the radius of curvature of the ellipse’s
boundary given by, Y = b2 /c, achieves its maximum at point A shown in Figure
6.2. Point A is also where the stress is maximum and is given by
$ % $ & %
2c c
σm = σ 0 1 + = σ0 1 + 2 , (4)
b Y
which, in the limit b  c that the cavity represents a crack, reduces to
&
σm 2c c
= =2 . (5)
σ 0 b Y
 6.7. Stress Concentration and Griffith’s Criterion 257

Figure 6.2. The elliptical cavity in a plate, subjected to

a uniform applied stress. Point A represents the notch
tip.

&
σ?

$ 5JHAII

"

σOO
> σNN
ρ )

Figure 6.3. Stress concentration at the elliptical cavity for c = 3b.

The ratio σm /σ 0 is called the stress-concentration factor, which is the mechanical

analogue of the field-multiplication factor defined in Section 5.1.1 for the problem
of dielectric breakdown with an elliptical conductor. Since as b → 0 the radius of
curvature becomes very small, it is clear that σm can become much larger than the
applied stress σ 0 .
Of particular interest is the local stresses along the x-axis. This is shown in
Figure 6.3 for c/b = 3, where we present the stresses σxx and σyy . The stress σyy
decreases from its maximum value of 7σ 0 at point A to an asymptotic value of σ 0 ,
while σxx rises from a zero value at A, reaching its maximum value at a point very
near the boundary of the cavity, beyond which it approaches 0 at large distances.
Note that the value of the stress depends on the shape of the cavity rather than
its size. Therefore, although it appeared that Eq. (5) can be used for estimating
the stress-concentration factors of such systems as the surface notch, a nagging
258 6. Fracture: Basic Concepts and Experimental Techniques

Figure 6.4. Incremental extension of a fracture of

length c through dc, under the applied stress.

question hindered further progress in understanding of fracture mechanics at Inglis

time: If the analysis of Inglis is applicable to a crack system (predicting a size-
independent stress), then why in practice large cracks appear to grow and propagate
more easily than the small ones? In addition, since the result of Inglis was in terms
of the radius of the curvature, the natural question to ask was, what is the physical
significance of the radius of curvature at the tip of a real crack?
Inglis’ work was followed up by Griffith (1920) who was interested in the
strength of inorganic glasses. He showed that the low strength of these materials,
compared to the theoretical estimates described earlier, was due to the presence
of sub-microscopic cracks. To reach this conclusion, Griffith analyzed the system
shown in Figure 6.4 which shows an elastic body that contains a plane-crack
surface S of length c, subjected to loads applied at its outer boundary. Griffith’s
main idea was to analyze this problem as a reversible thermodynamic system,
seeking the configuration that minimizes the total free energy of the system. Under
this condition, the crack would be in a state of equilibrium, and thus on the verge
of propagation.
If the crack undergoes extension, the energy H of the system associated with
this motion is the sum of the mechanical and surface energies. The mechanical
energy HM is itself the sum of two terms, the strain potential energy stored in
the elastic material, and the potential energy of the outer applied loading system
(which, in magnitude, is equal to the work associated with the displacements of
the loading points). The surface contribution HS is the free energy expended in
generating the new crack surfaces. Thus,
H = HM + HS . (6)
Thermodynamic equilibrium is reached when the mechanical and surface ener-
gies for a virtual crack extension dc (see Figure 6.4) are balanced. However, the
mechanical energy favors the crack extension (i.e., dHM /dc < 0) while the sur-
face energy opposes it (dHS /dc > 0). Thus, the Griffith energy-balance concept
is expressed through the equilibrium requirement that
dH
= 0. (7)
dc
Therefore, a crack would extend or contract reversibly for small displacements
from the equilibrium length, according to whether dH/dc is negative or positive,
 6.8. The Stress Intensity Factor and Fracture Toughness 259

respectively. For over 80 years, Eq. (7) has remained a pillar of the classical
continuum theory of brittle fracture.
To develop his theory further, Griffith took advantage of the Inglis’ solution
for an elliptical cavity described above. It can be shown that for a system under
constant applied stress (during crack formation), HM = −HE , where HE is the
strain potential energy stored in the elastic material, mentioned above, and the
negative sign is due to the fact that crack formation reduces the mechanical energy.
Using the solution of Inglis, it is not difficult to compute the strain energy density,
from which one obtains (by integrating the energy density over dimensions that
are large compared with the length of the crack), HE = −HM = π c2 (σ 0 )2 /Y  ,
where Y  is equal to the Young’s modulus Y in plane stress (thin plates), and
Y  = Y /(1 − νp2 ) in plane strain (thick plates), with νp being the Poisson’s ratio.
Since, for a unit width of the crack front, one has HS = 4c, where  is the free
surface energy per unit area, one obtains
π c2 (σ 0 )2
H = 4c − . (8)
Y
If we now apply Griffith’s criterion, Eq. (7), and identify σ 0 = σc0 as the critical
stress, we obtain
&
2Y  
σc =
0
. (9)
πc
Equation (9) is the famous Griffith relation, and is the mechanical analogue of
Eq. (5.8), the critical value of the far-field electric field for dielectric break-
down. Griffith also succeeded in qualitative verification of Eq. (9) by carrying
out experiments on an inorganic glass.
Because d 2 H/dc2 < 0, the energy of the system at equilibrium is maximum,
and therefore its configuration is unstable. That is, for σ 0 < σc0 the crack remains
stationary at its initial size c, whereas for σ 0 > σc0 it propagates spontaneously
without limit. Note, however, that an unstable crack may ultimately be arrested at
some point, which is often the case with cracks around contacts and inclusions. In
this case, further increase in the applied loading may lead to a second, catastrophic
instability configuration.

6.8 The Stress Intensity Factor and Fracture Toughness

An alternative, but equivalent, approach to determining the critical stress σc0 was
developed by Irwin (1958). He was the first to note that the stress field at a point
(r, θ) near the fracture tip, measured in polar coordinates with the crack line corre-
sponding to θ = 0, can be determined analytically. This problem will be discussed
in detail in Chapter 7, but for now it suffices to record the solution for the stress
components for Mode I fracture:
&  $ % $ % $ %
c 1 3 1
σxx = σ 0
1 − sin θ sin θ cos θ , (10)
2r 2 2 2
260 6. Fracture: Basic Concepts and Experimental Techniques
& $ % $ % $ %
c 1 3 1
σyy = σ 0
1 + sin θ sin θ cos θ , (11)
2r 2 2 2
& $ % $ % $ %
c 1 1 3
σxy = σ 0
sin θ cos θ cos θ , (12)
2r 2 2 2
and σzz = ν  (σxx + σyy ), where ν  = 0 for plane stress and ν  = νp for plane
strain, with νp being the Poisson’s ratio. The other components of the stress tensor
are zero. These results can also be written in terms of σrr , σθ θ , σrθ , etc. Qualitatively
similar equations also hold for Mode II fracture. The results for Mode III fracture
are particularly simple, as only σxz and σyz are non-zero. Therefore, Eqs. (10)–(12)
can be written in a general form:
&
c
σij = σ 0 fij (θ ). (13)
2r
Irwin introduced the quantity K = σ 0 (π c)1/2 as the stress intensity factor. Since,
in general, the stress intensity factor and the function f depend on the fracture
mode (I, II, or III), and as f also depends on the instantaneous crack velocity v,
Eq. (13) is written in a very general form:
Kβ β
σij = √ fij (v, θ ), (14)
2π r
where β indicates the fracture modes, β =I, II, and III. For each of the three
β
symmetrical loading configurations, fij (v, θ ) in Eq. (14) is a known universal
function. The stress intensity factor Kβ contains all the detailed information about
sample loading and history, and is determined by the elastic fields that develop
throughout the material, but the stress that locally drives the fracture is one which
is present at its tip. The stress intensity factors are related to the flow of energy
into the crack tip. A fracture can be viewed as a sort of sink that dissipates built-
up energy in a material. Therefore, the amount of energy flowing into a fracture
tip influences its behavior. The theoretical aspects of this view will be discussed
in Chapter 7. Thus, Kβ determines entirely the behavior of a fracture, and much
of the study of fracture processes is focused on either calculating or measuring
this quantity. The universal form of the stress intensity factor allows a complete
description of the behavior of the tip of a fracture where one need only carry out
the analysis of a given problem within the universal elastic region (see Chapter 7).
What happens if the material contains complicating factors, such as hetero-
geneity and anisotropy? Such complications destroy the symmetry that exists in
homogeneous and isotropic materials. For example, for a material in which the
elastic properties on opposing sides of a plane-crack interface are asymmetric, the
crack tip fields will also be asymmetric. Therefore, for example, a crack inter-
face between two dissimilar materials, subjected to tensile loading, will exhibit
not only Mode I behavior, but some Mode II and Mode III as well. Despite such
complications, it is now generally accepted that the essential r −1/2 singularity that
Eqs. (10)–(14) exhibit is not changed by such complexities, and therefore the stress
intensity factors can still be superposed. Therefore, for arbitrary loading configu-
rations, the stress field around the crack tip is given by three stress intensity factors
 6.9. Classification of the Regions Around the Crack Tip 261

Kβ which lead to a stress field that is a linear combination of the pure modes:

3
Kβ β
σij = √ fij (v, θ ). (15)
β=1
2π r
The critical condition for crack propagation can now be expressed in terms of
the critical value Kc of the stress intensity factor, which is usually referred to as the
fracture toughness. Thus, in terms of the critical energy Hc , the fracture toughness
is given by
!
Kc = Hc Y  . (16)
We should emphasize, as already mentioned above, that the Griffith–Irwin pre-
diction for the critical stress σc0 (or the fracture toughness Kc ) is valid for the onset
of growth under static conditions, and for homogeneous materials. As soon as the
crack begins to grow, the stress field around it changes dynamically. In particular,
if the crack propagates at high speeds, the inertial effects substantially change the
stress field. The Griffith–Irwin approach has nothing to offer for these changes.
In other words, the Griffith–Irwin criterion can tell us when a brittle crack may
extend, but has nothing to say about how it will extend. In addition, the r −1/2 sin-
gularity at the tip of the crack cannot be reconciled with any real fracture process,
as there is no solid that can resist an infinite stress anywhere in its structure. The
root of this singularity is in the assumptions that the Hooke’s law (linear elasticity)
is operative everywhere in the material, and that a continuum approximation can
describe the state of the system. These assumptions break down for the region in
the vicinity of the crack tip, and necessitate a reclassification of the region around
the tip; this is discussed in the next section.

6.9 Classification of the Regions Around the Crack Tip

Many complex phenomena are active in the vicinity of a crack tip that vary, de-
pending on the material, from dislocation formation and emission in crystalline
materials to the complex unraveling and fracture of intertangled polymer strands in
amorphous polymers. Fracturing and the complex dissipative processes occurring
in the vicinity of the crack tip occur due to large values of the stress field as one
approaches the tip. As discussed above (and will also be considered in detail in
Chapter 7 where we describe formation of fracture nucleation and propagation by
continuum mechanics), if the material around the crack tip were to remain linearly
elastic until fracture, the stress field at the crack tip would be singular. Since a real
material cannot support such singular stresses, the assumption of linearly elastic
behavior in the vicinity of the tip must break down and material-dependent dis-
sipative processes must begin playing an important role. Thus, at first glance, a
universal description of fracture, in terms of the stress intensity factor and the func-
tion fij described above, may seem a hopeless task. However, a way for attacking
this problem was proposed by Orowan (1955) and Irwin (1956) who suggested
independently that the region around the crack tip should be divided into three
262 6. Fracture: Basic Concepts and Experimental Techniques

separate regions:
(1) The cohesive zone, which is the region immediately surrounding the crack tip,
in which all the nonlinear dissipative processes that allow a crack to move
(forward) are assumed to occur. In continuum fracture mechanics, detailed
description of this zone is avoided, and is simply characterized by the energy
, per unit area of crack extension, that it will consume. The size of the
cohesive zone is material-dependent, ranging from nanometers in glass to
microns in brittle polymers. Its typical size is the radius at which an assumed
linear elastic stress field surrounding the fracture tip would equal the yield
stress of the material.
(2) The universal elastic region, which is the region outside of the cohesive zone
for which the response of the material can be described by linear continuum
elasticity. Outside of the cohesive zone, but in the vicinity of the fracture tip,
the stress and strain fields take on universal singular forms which depend only
on the symmetry of the externally applied loads. In 2D the singular fields
surrounding the cohesive zone are completely described by the three stress
intensity factors which incorporate all the information regarding the loading
of the material. As discussed above, the stress intensity factors are related to the
energy flux into the cohesive zone. The larger the overall size of the material
containing the crack, the larger this region becomes. Roughly speaking, for
given values of √ the stress intensity factors, the size of the universal elastic
region scales as L, where L is the macroscopic length scale on which forces
are applied to the material. Thus, as L increases, the assumptions of continuum
fracture mechanics become progressively more accurate.
(3) Outer elastic region, which is the region far from the crack tip in which stresses
and strains are described by linear elasticity. Details of the solution of the
equations, describing fracture propagation, in this region depend only on the
locations and strengths of the loads, and the shape of the material. For some
special cases, analytical solutions are available, but in general one must resort
to numerical simulation. That deriving these solutions is possible is because, so
far as linear elasticity is concerned, viewed on macroscopic scales, the cohesive
zone shrinks to a point at the fracture tip, and the fracture itself becomes a
branch cut. Thus, replacing the complex domain in which linear elasticity holds
with an approximate one that needs no detailed knowledge of the cohesive zone
is another approximation that becomes increasingly accurate as the dimensions
of the sample, and hence the size of the universal elastic region, increase. The
assumption that the cohesive zone in a material is encompassed within the
universal elastic region is sometimes called the assumption of small-scale
yielding.
The dissipative processes within the cohesive zone determine the fracture energy
. If no dissipative processes other than the direct breaking of the atomic bonds
take place, then  will be a constant that depends on the bond energy. In general
though,  is a complex function of both the crack velocity and history, and differs
by orders of magnitude from the surface energy—the amount of energy required
 6.10. Dynamic Fracture 263

to sever a unit area of atomic bonds. No general first principles description of

the cohesive zone exists, although numerous models have been proposed (see, for
example, Lawn, 1993). We will come back to this important issue in Chapter 7.

6.10 Dynamic Fracture

Our discussion so far has been limited to static fractures. However, in practice
dynamical effects are important and must be considered. To understand how a
dynamic situation may come about, suppose that an unbalanced force acts on any
volume elements within a material that contains cracks. Then, that element will
be accelerated, thereby acquiring kinetic energy. The system will then be in a
dynamic state so that, as pointed out and emphasized above, the Griffith–Irwin
static equilibrium condition will no longer apply. Under certain conditions, the
growth of the crack may be slow (for example, when, compared to the mechanical
energy, the contribution of the kinetic energy is insignificant), in which case the
material may be considered as being in a quasi-steady-state condition.
There are two scenarios by which the state of a cracked material may become
dynamical. One is when a crack reaches an unstable state in its length: The material
receives kinetic energy contributed by the inertia of the material that surrounds
the rapidly-separating walls of the crack. One then has a running crack which is
characterized by a rapid acceleration toward a terminal velocity vc , and is governed
by the speed of elastic waves in the solid. As will be discussed in Chapter 7, the
prediction of linear continuum fracture mechanics for the value of the terminal
velocity vc did not agree with experimental observations, and because of this the
subject was controversial for a long time and was resolved only recently. In the
second scenario a dynamical state arises when the applied loading changes rapidly
with the time, as in, for example, impact loading.
Avery common dynamical effect is fatigue. It has been seen in many experiments
that, often a material that has resisted the same external load many times without
developing cracks, suddenly does so after the external load has been applied a
certain number of times. If the external load is applied periodically in time, then
the phenomenon is called cyclic fatigue, and the number of times that the external
load must be applied for the crack to develop is called failure life. It has been found
empirically that the number Nc of the cycles that the external load must be applied
scales with the amplitude A of the load as
Nc ∝ (A − Ae )−α , (17)
where Ae is called the endurance limit. Clearly, if the amplitude of the applied load
is less than Ae , then the material will not break at all. The value of the exponent
α has been found to be around 8 − 10. Equation (17) is usually called the Besquin
law.
Another important dynamical effect is stress corrosion cracking. Baker and
Preston (1946) first reported that the toughness of glass reduces considerably if it
is in a humid environment, since water penetrates the glass at the crack tip where
264 6. Fracture: Basic Concepts and Experimental Techniques

the crystalline structure of the glass is relatively open. Once inside the glass, water
forms a base with existing sodium ions which corrodes the region in the vicinity of
the crack tip, hence lowering its toughness and increasing the likelihood of brittle
fracture. Aluminum and titanium, two heavily-used metals in aircraft, suffer most
from stress corrosion cracking.
The mechanism that leads to stress corrosion cracking is either anodic or ca-
thodic. That is, the phenomenon can be suppressed in an electrolytic environment
by placing either the anode or the cathode on the material and the corroding agent
as electrolytic medium. For example, hydrogen embrittlement of metals is the
most common cathodic process. The anodic process also occurs in metals that are
coated by a layer of oxide to be protected from the environment. If the coating
is opened at the crack tip, the metal will be exposed to the anodic agent at the
tip. Under this condition, the velocity of fracture propagation would be controlled
by the rate of the chemical reactions. Since these reactions are typically slow, the
fracture propagates slowly, which is why, for example, it takes aircraft a long time
to develop stress corrosion cracks in their fuselage.

6.11 Experimental Methods in Dynamic Fracture

We now describe and discuss some of the main experimental methods that are used
in studies of dynamic fracture. These methods vary greatly and their use depends on
both the specific phenomenon that is under study and on the experimental resources
at hand. In a typical experiment stress is applied externally at the boundary of the
system and its response and the resulting behavior of the fracture are observed
and measured. During the time that the crack propagates one can measure its
position and velocity, the time-dependent stress field at the crack’s tip, the acoustic
emissions resulting from the crack motion, as well as the resulting fracture surface.
In what follows we describe the typical ways by which the various quantities
of interest are measured. Our discussion in this section follows closely that of
Fineberg and Marder (1999).

6.11.1 Application of External Stress

The externally-applied stress distribution determines the stress field in the close
vicinity of the crack tip or, equivalently, the stress intensity factor, and hence, is
the driving force for advancement of a fracture. Two basic types of loading are
typically used in fracture experiments, static and dynamic, and what follows is the
description of each type.

6.11.1.1 Static Stress

In such experiments either the boundary conditions or the applied stresses are con-
stant, thus imprinting an initial static stress distribution onto the sample material.
Depending on the applied loading and boundary conditions, the stress intensity
 6.11. Experimental Methods in Dynamic Fracture 265

Figure 6.5. Three typical experimental configurations.

factor (or stored energy density) along the prospective path of a crack can in-
crease, resulting in a continuously accelerating crack, or decrease, leading to a
decelerating and possibly arrested crack. A few examples of the common loading
conditions that are used are shown in Figure 6.5 where single-edge notched (SEN),
double-cantilever beam (DCB), and infinite strip (IS) loading conditions are pre-
sented. The SEN condition is sometimes used to approximate fracture propagation
in a semi-infinite system. When the external loading is a constant stress applied
at the vertical boundaries of the sample,√then for a large enough sample the stress
intensity factor KI is proportional to σ l, and therefore the energy release rate H
is given by, H ∝ σ 2 l (see Chapter 7 for additional theoretical details), where σ is
the applied stress and l is the length of the crack. This configuration is used, for
example, to study the behavior of an accelerating crack.
In the IS configuration, the sample is loaded by displacing its vertical boundaries
by a constant amount. Under this condition, the energy release rate is constant for
a crack that is sufficiently far from the horizontal boundaries of the sample, and
thus this loading configuration is amenable to the study of a crack moving in
steady-state.
In the DCB configuration, a constant separation of the crack faces is imposed
at l = 0, H ∝ l −4 is a decreasing function of l, and hence can be used to cause
crack arrest. How is the DCB configuration used to study dynamic fracture? An
initially imposed seed crack of length l = l0 would propagate as soon as H ex-
ceeds the limit imposed by the Griffith condition, i.e., when dH/dl = 0. Under
ideal conditions, the crack propagates for an infinitesimal distance and then stops,
because in DCB configuration H is a decreasing function of l. Although the Grif-
fith criterion assumes that the initial crack is as sharp as possible, what is prepared
in the laboratory by cutting rarely yields a tip that meets this condition. We may
view the initial seed crack as having a finite radius at its tip, thereby blunting the
stress singularity and allowing a substantially higher energy density to be imposed
in the system prior to fracture than what is allowed by a sharp crack. Thus, there
is excess elastic energy that drives the crack beyond the constraints imposed by
an initially sharp crack which, in the case of the DCB configuration with constant
266 6. Fracture: Basic Concepts and Experimental Techniques

separation imposed, can cause a crack to propagate well into the sample before
crack arrest occurs. Nonlinear material deformation around the tip, plastic flow
induced by the large stress build-up, and from crack-tip shielding that results from
the formation of either micro-cracks or small bridges across the crack faces in the
near vicinity of the tip, can also cause blunting of the singularity around a crack
tip. The DCB configuration can also generate an accelerating crack by imposing
a constant stress (instead of constant separation) at the crack faces. Under this
condition, the (quasi-static) energy release rate increases with the crack length l as

12σ 2 l 2
H= , (18)
Y w2 d 3
where σ is the stress applied at opposite points on the crack faces at the edge of the
sample, and w and d are, respectively, the thickness and half-width of the sample.

6.11.1.2 Initiation of Fractures

The stress singularity at the tip of a crack, as its radius of curvature approaches
zero (see above), implies that initiation of a fracture under static loading config-
urations is strongly dependent on the initial radius of the crack tip and hence on
the preparation of the initial crack. However, the stress build up that precedes
fracture initiation can be taken advantage of for loading a material with an initial
energy density before the onset of fracture. This is, however, extremely difficult
as experimental reproducibility of the stress at fracture initiation is non-trivial. In
some materials one can achieve a reproducible stress at fracture initiation by first
loading the system to the desired stress and then either waiting for some time for
the material to fracture as a result of noise-induced perturbations, or by sharpening
the initial crack, once the desired initial conditions have been reached. These tricks
do not, however, work very well in such brittle materials as ceramics.

6.11.1.3 Dynamic Stress

In some applications, such as the study of crack initiation before the material
surrounding the crack tip has had time to react to the applied stress, very high
loading rates are desirable. A common way to achieve this is by loading an initially
seeded sample by collision√ with a guided projectile. In this way loading rates as
high as K̇I ∼ 109 MPa m/s (Prakash and Clifton, 1992) have been achieved.
An alternative way for producing high loading rate is by sending a very large
current through a folded conducting strip, inserted between the two faces of an
initial crack, which induces magnetic repulsion between adjacent parts of the strip,
enabling direct loading of the crack faces (Ravi-Chandar and Knauss, 1982). A
high loading rate can also be produced by discharging a capacitor-inductor bank
through the strips. This technique has been utilized for producing a pressure pulse
with a step function
√ profile on the crack faces having loading rates of the order
of K̇I ∼ 105 MPa m/s in experiments designed to investigate the response of a
moving crack to rapidly changing stresses.
 6.11. Experimental Methods in Dynamic Fracture 267

6.11.2 Direct Measurement of the Stress Intensity Factor

The stress intensity factor can be directly measured by optical methods, which can
measure the energy release rate. Two common methods are the method of caustics,
and photoelasticity.

6.11.2.1 The Method of Caustics

The technique was originally proposed by Manogg (1966) with significant contri-
butions by Theocaris and Gdoutos (1972) and Kalthoff (1987) (who also provided
a review of this method) for transparent materials, and by Rosakis et al. (1984) for
opaque materials. The method, applicable to thin quasi-2D plates, uses the deflec-
tion of an incident collimated beam of light as it either passes through transparent
material, or is reflected by an opaque material, that surrounds the crack’s tip. Due
to so-called Poisson contraction generated by the high tensile stresses near the tip,
the initially flat faces of a plate will deform inwardly which creates a lensing effect
and diverts light away from the crack’s tip. The diverted rays form a 3D surface
in space in which no light propagates. When this light is imaged on a screen, a
shadow (hence the name shadow-spot that is sometimes used) is observed which
is bounded by a caustic surface or a region of high luminescence formed by the
locus of the diverted rays. From the shape of the caustic surface, which is recorded
by a high speed camera, the instantaneous value of the stress intensity factor is
estimated. This method works well with the caveat that estimating the stress inten-
sity factor is based on a certain assumption that, as discussed in Chapter 7 [see the
discussion before and after Eqs. (7.50)–(7.52)], must, in the immediate vicinity of
the crack tip, break down as the material’s yield stress is approached. Therefore,
care must be taken that the curve on the material that maps onto the caustic is well
away from the cohesive zone surrounding the crack tip; see, for example, Rosakis
and Freund (1981).

6.11.2.2 Photoelasticity
This method, coupled with high-speed photography, is also used for measuring the
stress distribution, and hence the stress intensity factor, induced by a moving crack
(Kobayashi, 1987). It is based on the birefringence induced in most materials un-
der an imposed stress, which causes the rotation of the plane of polarization light
moving through the material. The induced polarization depends on the properties
of the stress tensor which are rotationally invariant, and therefore can depend only
on the two principal stresses σ1 and σ2 . Moreover, there should be no rotation of
polarization when the material is stretched uniformly in all directions, in which
case the two principal stresses are equal, and therefore the angular rotation of the
plane of polarization must be of the form, c(σ1 − σ2 ), where c is a constant that is
determined experimentally. If stresses of a 2D problem are calculated analytically,
the results can be substituted into this expression and compared with experimental
fringe patterns obtained by viewing a reflected or transmitted beam of incident
polarized light through a polarizer. The observed intensity depends on the phase
268 6. Fracture: Basic Concepts and Experimental Techniques

difference picked up while traversing the material, and hence provides a quantita-
tive measure of the local value of the stress field. The application of this method to
transparent materials is straightforward. These methods have also been extended
to opaque materials by the use of birefringent coatings which, when sufficiently
thin, mirror the stress field at the surface of the underlying material. Dally (1987)
reviewed the applications of these methods. Similar to the method of caustics,
quantitative interpretation of these measurements is limited to the region outside
of the plastic zone.

6.11.3 Direct Measurement of Energy

Direct measurement of the energy release rate, as a function of the velocity of a
moving crack, can be obtained by constraining a crack to propagate along a long
and narrow strip; see Figure 6.5. The advantage of this method is that it relies
only on symmetry properties of the system, and hence does not require additional
assumptions regarding, for example, the size or properties of the cohesive zone.
A series of experiments, using a long strip geometry and varying the value of
δ (as shown in the Figure 6.5), results in a direct measure of H(v), where v
is the velocity of the crack. In the experiments of Sharon et al. (1995) using
polymethylmethacrylate (PMMA) (see Chapter 7), steady-state mean velocities
were attained when the crack length exceeded roughly half the strip height. Their
measurements of H(v) agreed well with results previously obtained with PMMA
by means of the methods of caustics (see above) reported by Pratt and Green
(1974).

6.11.4 Measurement of Fracture Velocity

Under dynamic conditions, the velocity of the tip of a crack generally accelerates
to values of the order of the sound speed in the material. Since the duration of
a typical experiment is of the order of 100 µs, one needs relatively high-speed
measurement techniques. Three common methods, based on either high-speed
photography, resistance measurements, or the interaction of a moving crack with
ultrasonic waves, have been used in the past which are now briefly described.

6.11.4.1 High-Speed Photography

This method is the most straightforward technique for measuring the velocity of
a moving crack. It can be used in conjunction with instantaneous measurements
of the stress intensity factor by means of the method of caustics or photoelasticity
discussed above. It also has some major shortcomings. For example, although the
frame rates of high speed cameras are typically between 200 kHz and 10 MHz, the
cameras are capable of photographing only a limited number, say 30, of frames.
Thus, this method can either provide measurements of the mean velocity (with
the average taken over the interval between the frames) at a few points, or can,
at the highest photographic rates, provide a detailed measurement of the crack
 6.11. Experimental Methods in Dynamic Fracture 269

velocity over a short, say about 3 µs, interval. Moreover, precision of this method
is obviously limited by the accuracy at which the location of the crack tip can be
determined from a photograph.
These problems can, to some extent, be overcome by using a streak camera
(Bergkvist, 1974). In this method, a film is pulled past the camera’s aperture at
high speed. The material is illuminated from behind so that, at a given instant, only
the light passing through the crack is photographed. Since one can force a crack to
propagate along an essentially straight line, the exposed film provides a continuous
record of its length as a function of time. The basic resolution of the measurements
depends on the film’s velocity and that of the high-speed film used, and on the
post-processing performed on the film in order to extract the velocity measurement
and the stability of the film’s travel velocity. The same type of experiments have
also been carried out (Döll, 1975) by high-speed measurements of the total beam
intensity that penetrates the material. If the crack does not change its shape, the
beam intensity depends linearly on the crack’s length.

6.11.4.2 Measurement of Resistivity

Another method of measuring the velocity of a rapidly moving crack is by adhering
a grid of thin, electrically-conductive strips to a sample material prior to fracture.
Crack propagation causes the crack faces, and therefore the conducting strips, to
separate. Therefore, if, for example, the strips are connected in parallel to a current
source, measurement of the grid’s electric resistance with time will provide a
jump at each instant that the crack tip traverses the end of a strip, yielding the
precise location of the crack tip at a number of discrete times. To ensure that the
crack tip is not significantly ahead of the fracture of the strip, the strip’s thickness
must be at least an order of magnitude less than the crack face separation. The
disadvantage of this method is that the discrete measurements can only provide
a measure of the mean velocity between the strips. By extending the method to a
continuous coating (instead of discrete strips), one has the advantage that the crack
tip’s location is obtained as quickly as the voltage drop across the coating can be
digitized. The precision of the measurements is limited only by the background
noise and the uniformity of the coating. It can be improved with an evaporated
coating which provides precise velocity data near the sample faces. Thinness of the
sample does not present a limitation and can, in fact, be taken advantage of if one
wishes to correlate the instantaneous velocity with localized features formed on the
fracture surface. This method has been used widely (see, for example, Brickstad
and Nilsson, 1980; Fineberg et al., 1991, 1992) with considerable success, using
a variety of materials.

6.11.4.3 Ultrasonic Measurements

In this method (Kerkhof, 1973), which has been used both with glass and brittle
polymers, a moving fracture is perturbed by an ultrasonic wave generated by a
sample boundary in a direction orthogonal to that of crack propagation. The inter-
action of the sound with the crack tip causes the sound to be deflected periodically
270 6. Fracture: Basic Concepts and Experimental Techniques

as it traverses the sample, the trace of which is imprinted onto the resulting fracture
surface. Since the temporal frequency of the modulation is that of the ultrasonic
driving, measuring the distance between neighboring surface modulations pro-
vides a nearly continuous data set for the instantaneous velocity of the crack tip.
The method’s precision is limited only by the ultrasonic frequency used, which
is typically in the MHz range, and also by the precision of the surface measure-
ment. The disadvantage of this method, relative to the other techniques, is that it
is a perturbative method, since the crack deflection is accomplished by altering
the stress field at its tip, and hence externally-induced oscillations can potentially
mask intrinsic, time-dependent effects.

6.11.5 Measurement of the Thermal Effects

A propagating fracture transforms the elastic energy stored in the material to either
kinetic energy, the energy needed for breaking the atomic bonds, or to dissipated
heat. The dissipated heat can be measured by two types of measurements. In one
method one places small temperature sensors at a given distance from the path
of a crack and measures the temperature rise in the material as a function of time
after fracture has occurred. Since the time scale of fracture is orders of magnitude
shorter than the typical times for thermal diffusion within the material, one can
approximate the problem by assuming that an instantaneous planar heat source
is created along the fracture plane, and that the radiative losses are negligible
over the period of measurements. Then, the measured time-dependence of the
temperature at a single point can be fitted to the solution of the heat conduction
equation. Measurements of this sort were carried out in PMMA (Döll, 1973),
in glass (Weichert and Schonert, 1974), and in steel (Zimmerman et al., 1984).
Moreover, it is possible to estimate the temperature rise in the vicinity of the
crack tip by use of IR detectors (Fuller et al., 1983; Zehnder and Rosakis, 1991;
Kallivayalil and Zahnder, 1994), assuming that the emission spectrum of a crack
corresponds to a black body spectrum, although this assumption may be suspect,
at least in the immediate vicinity of the tip.

6.11.6 Measurement of Acoustic Emissions of Fractures

Measurements of acoustic emissions have long been used (see, for example, Scott,
1991) as a means of detecting either the onset of, or the precursors to, fracture,
where the existence, the frequency of events and their locations can be measured.
Although these techniques, due to their relatively limited precision, have not been
used extensively in dynamic fracture experiments, they provide a sensitive method
for determining whether changes in the stress field are taking place during fracture,
because any rapid changes invariably release stress waves, and therefore can be
used for detection of fracture and its onset. Such methods utilize arrays of reso-
nant acoustic transducers since the advantage of their high sensitivity more than
offsets the loss of information about the signal’s spectral content. In fact, since
the spectral content of the acoustic signal broadcast by a moving crack carries
 6.12. Oscillatory Fracture Patterns 271

important information (see Chapter 7 for theoretical discussion of this point),

broadband transducers should be used together with relatively high amplification
to offset the transducers’ lack of sensitivity. The emissions are then correlated
(Gross et al., 1993; Boudet et al., 1995, 1996) with velocity and fracture surface
measurements. That this is a sensible method even when deflections of the 2D
sample normal to their surface that are measured are due to the fact that the probe
is sensitive to both longitudinal and shear waves due to mode conversion (Kolsky,
1953).

6.12 Oscillatory Fracture Patterns

One fundamental prediction of linear continuum fracture mechanics is that, as a
crack propagates, its speed should increase until it reaches its asymptotic value,
the Rayleigh sound speed cR —the speed of sound on a free surface. However, ex-
perimental observations of fracture propagation in many heterogeneous materials
indicate that, in the vast majority of cases, the ultimate velocity of a propagating
crack is not more than about 0.5cR (unless the material is strongly anisotropic; see
Chapter 7). For example, oscillatory fracture patterns that have been observed in
many materials strongly violate this fundamental prediction, and our goal in this
section is to briefly describe and discuss these patterns and how they have been
created in the laboratory.
These patterns were observed in the beautiful experiments of Yuse and Sano
(1993). They imposed a temperature gradient along a thin glass plate, from a hot
region to a cold one. A microcrack was introduced in the glass, and the glass was
pushed. As the plate started to move the crack jumped ahead of the thermal gradient
and stayed there. It was observed that if the plate moves slowly, the growing
crack remains straight and stable. However, increasing the velocity to a critical
value vc gives rise to a transition whereby the fracture path begins to oscillate
and an instability appears. At still higher velocities crack branching appears; see
Figure 6.6. Ronsin et al. (1995) also provided experimental data for brittle fracture
propagation in thin glass strips, using a thermally-induced stress field. In their
experiments the temperature field was controlled by the width w of the plate, and
induced thermal expansion in the sample. It was observed that for widths below a
critical value wc no fracture was formed. For wc < w < wo , where wo is a second
critical width for the onset of oscillatory cracks, straight fractures were formed and
propagated with a constant speed. For w > wo oscillatory fractures were generated
which became more irregular as w was increased beyond wo .
These predictions are in agreement with the results of several sets of spectacular
experiments by Fineberg et al. (1991,1992) and Gross et al. (1993). Many earlier
experiments had already reported several interesting features of dynamic crack
propagation in materials (see, for example, Mecholsky, 1985). For example, it
had been reported (see, for example, Döll, 1975; Kusy and Turner, 1977) that in
some brittle materials, such as PMMA, the fracture pattern exhibits characteristic
wavelength, that surface roughness increases with crack speed (see, for example,
272 6. Fracture: Basic Concepts and Experimental Techniques

Figure 6.6. Fracture pattern forma-

tion in the experiments of Yuse and
Sano (1993).

Langford et al., 1989, and references therein), and that periodic stress waves are
emitted from the tip of the rapidly moving cracks in a wide variety of materials
(see, for example, Rosakis and Zehnder, 1985; Dally et al., 1985, and references
therein). Fineberg et al. (1991,1992) carried out beautiful and precise experiments
to study fracture propagation in brittle plastic PMMA and showed, that there is
a critical velocity vc beyond which the velocity of crack tip begins to oscillate,
the dynamics of the crack changes abruptly, and a periodic fracture pattern is
formed. For v > vc the amplitude of the oscillations depends linearly on the mean
velocity of the propagating crack. Thus, the dynamics of cracks is governed by a
dynamical instability, and explains why the crack tip velocity does not attain the
limiting Rayleigh velocity predicted by the linear elastic theory. Although Yoffe
(1951) had already predicted the existence of a sort of dynamical instability in
fracture, showing that a fracture that moves along a straight line will branch off if
its speed becomes larger than a critical value, her predicted critical velocity was
too large, and therefore the type of instability that was considered by her could not
provide a complete explanation for Fineberg et al.’s experiments. The theoretical
studies of such fracture patterns will be discussed in Chapter 7.
In another set of beautiful experiments, Gross et al. (1993) used two materials,
the PMMAand soda-lime glass, to show that all features of dynamics of crack prop-
agation in the two materials, such as acoustic emission, crack velocity, and surface
structure, exhibit quantitative similarity with each other. Thus, there exists univer-
sal characteristics of fracture energy in most materials that are the result of energy
dissipation in a dynamical instability. Perhaps the most spectacular experiments
were carried out by Sharon et al. (1995) and Sharon and Fineberg (1996) using
the brittle plastic PMMA. They identified the origin of the dynamical instability
 6.13. Mirror, Mist, and Hackle Pattern on a Fracture Surface 273

during fracture propagation as being the nucleation and growth of the daughter
cracks which limit the speed of the propagating crack tip. The daughter fracture
carries away a fraction of the energy concentrated at the tip of the moving crack,
thus lowering the velocity of the tip. After some time, the daughter crack stops
growing, and thus the crack tip velocity increases, until a new daughter fracture
starts to grow, and so on. They also observed that the branching angle for a longer
daughter fracture was smaller than that of the shorter daughter fractures. Theoret-
ical modeling and computer simulations of dynamic fracture that can reproduce
these features will be described in detail in Chapters 7 and 8.

6.13 Mirror, Mist, and Hackle Pattern on a Fracture

Surface
Studies of fracture surfaces of amorphous brittle materials indicate that they have a
characteristic structure that is popularly referred to as mirror, mist, and hackle. This
pattern has provided an important tool for studying a number of important fracture
phenomena, and at the same time has raised a number of fundamental questions.
Figure 6.7 presents the original pattern reported by Johnson and Holloway (1966),
which is the fracture surface of an inorganic glass, soda-lime-silica glass rod with

1 mm

Figure 6.7. Light microscope photograph of mirror, mist, and hackle regions on fracture
surface of a 5 mm diameter soda-lime-silica glass rod, tested in uniaxial tension. The mirror
region is roughly circular, surrounded by the narrow band of mist that gradually develops
into the hackle (after Johnson and Holloway, 1966).
274 6. Fracture: Basic Concepts and Experimental Techniques

a diameter of 5 mm, tested in uniaxial tension. A crack nucleated at a small surface

flaw that was generated by contact during handling, and then propagated normal to
the tensile axis, i.e., under Mode I fracture. In the initial stages of the experiment,
the crack growth led to a very smooth fracture surface, which is called mirror. The
crossing of the rupture front with elastic waves can leave behind ripples in the
mirror zone which are called Wallner lines.
This region is surrounded by a slightly rougher and less reflective region, which
is referred to as the mist. It consists of fine striations that look like microscopic
blades that are oblique to the crack plane. This zone appears when the velocity of
the crack is about half of the velocity of transverse elastic waves. Finally, the mist
region merges into a very rough fracture surface with irregularly oriented facets,
which comprise the hackle region. The facets are separated by large steps that are
aligned parallel to the main direction of crack propagation. As Figure 6.7 indicates,
the transitions between the neighboring regions are not sharp; rather they represent
progressive changes in the surface roughness.
Since the transitions from mirror to mist to hackle regions are not sharp but
gradual and diffused, the answer to the question of where one region ends and
another one starts cannot be precise. Johnson and Holloway (1966), who analyzed
these regions for the first time, stated that, “The position assigned to the boundary
between mirror and mist zones depends upon illumination and the magnification
at which the fracture is examined, even within the range of the optical microscope.
With an electron microscope mist can readily be resolved in the region seen as mir-
ror under optical conditions.” However, a better way of distinguishing between
the three zones is by measuring the changes that occur in the surface roughness of
the fractured material. While the height of the roughness remains essentially con-
stant in the mirror region, it increases sharply and monotonically as the transition
to the mist zone is made. Measurement of roughness of fracture surfaces and its
significance will be discussed in the next section.
In the light of our discussions earlier in this chapter, it is not difficult to un-
derstand the development of the mirror, mist and hackle pattern. Suppose that the
length of the initial flaw is c. In uniaxial tension, the stress concentration is large at
the tip of the flaw. If the stress is large enough, the Griffith criterion is satisfied and
the fracture begins to grow. If the loading condition is held constant, the increase
in the fracture length implies fracture instability and the existence of excess energy
that drives the fracture. Thus, the crack accelerates very rapidly, with which the
rate of energy release also increases rapidly, resulting in higher stress intensities
at the tip. The large stress intensity and rate of energy release also imply a corre-
sponding increase in the micro-mechanical activity at the tip of the fracture, and
hence a corresponding increase in the roughness of the fracture surface. Note that,
depending on the test conditions, a fourth region of the fracture surface may also
develop. This region would be the result of having the main fracture bifurcate into
two or more branches. Normally, bifurcation occurs in high-stress failures.
The boundaries between the mirror, mist, and hackle regions are roughly circular,
implying that the crack accelerates outward in all directions with essentially the
same rate. Experiments have indicated that if R is the radius of a boundary between
 6.14. Roughness of Fracture Surfaces 275

two zones, then the fracture strength σf of the material, i.e., the stress at which
the crack starts to move (see also Chapters 7 and 8) is related to R through
√
σf R = a, R = Rmirror , Rmist , Rhackle , (19)
where a is a constant. Observe that Eq. (19) has √ the √same form as the Griffith
condition, Eq. (9) [if we rewrite Eq.√(9) as, σc c ∝ Y   ], and therefore the
constant a is related to the quantity Y   that appears in the Griffith condition.
Moreover, in view of Eq. (16), the constant a can also be related to the fracture
toughness Kc . Experiments have also indicated that if R0 is the radius of the
initial flaw at which the crack nucleates, then the radius Rmirror of the mirror
zone is related to R0 through, Rmirror /R0  10. Clearly, the circular boundaries
between the three zones will not develop if the crack cannot accelerate in all
directions with the same rate. The deviation from circularity depends partly on
the boundary conditions used in the test. For example, a material in a bending
experiment develops a stress distribution that is quite different from one that it
experiences in a uniaxial tension experiment. Moreover, the mechanism of crack
growth in amorphous materials is different from that of crystalline materials, so
that the shape of the boundaries between the mirror, mist and hackle zones also
depends on the material.
Before closing this section, let us point out that in the fracture literature one
often finds references to twist hackle and stress or velocity hackle. The former
refers to a rough surface that is generated by a Mode I/III fracture experiment,
whereas the latter is the result of a crack propagating at very high speeds or under
a large stress. The phrase mirror has also been used occasionally for describing the
initial stage of the development of a fracture surface, whereas careful examination
of the surface would reveal that it is too rough to be classified as the mirror zone.
To make the distinction between a mirror zone and a rougher region, one may
define the mirror region as the zone in which the average height of the roughness
is less that the wavelength of light.

6.14 Roughness of Fracture Surfaces

The development of mirror, mist and hackle zones makes it clear that, as a crack
propagates, the fracture surface develops roughness, the intensity of which in-
creases with the extent of the crack propagation, which in turn depends on the
loading condition, and the shape, morphology and composition of the material.
Therefore, measurement of the roughness of a fracture surface may provide in-
sight into dynamics of fracture propagation in a material. However, because of the
dearth of comprehensive experimental data, i.e., data sets that contain simultane-
ous measurements of the roughness, the (dynamic) stress intensity factor Kd and
the speed v of fracture propagation, the relation between the three quantities is not
clear at present, and is the subject of ongoing research by many groups around
the world. Arakawa and Takahashi (1991, where references to their earlier work
in the Japanese literature can also be found) carried out one such set of measure-
276 6. Fracture: Basic Concepts and Experimental Techniques

Figure 6.8. Dynamic stress intensity factor Kd , the fracture velocity v, and the roughness
of the fracture surface versus the fracture length for a brittle epoxy resin. KI c is the critical
value of the stress intensity factor KI . The data are from Arakawa and Takahashi (1991)
(after Hull, 1999).

ments which is summarized in Figure 6.8. In their experiments, they used 6 mm

thick plates of various transparent plastics, including a thermosetting epoxy and
a thermoplastic PMMA, and measured the velocity of the propagating crack, the
dynamic stress intensity factor Kd , and the roughness w of the surface. There seem
to be general correlations between the crack speed and the stress intensity factor
on one hand, and the roughness of the fracture surface on the other hand. At the
same time, another feature of this figure indicates that there may not be a unique
relation between the crack speed and the intensity factor, since the two quantities
have not reached their maximum at the same point, whereas the maxima of the
intensity factor and surface roughness seem to happen at the same crack length,
and therefore these two quantities are probably better correlated than Kd and v.
However, this is not a completely universal rule. Under certain circumstances,
the surface may become smoother as fracture propagation proceeds. An example
is provided by elastomers, where in some range of the crack speed their fracture
surface is rough at low speeds, while it is smooth and mirror-like at high speeds.
Thus, the increase or decrease in Kd and v is not directly linked to the roughness
of the surface. Moreover, it must be mentioned that many materials do not develop
mirror smooth surfaces at all. For example, if sharp pre-existing cracks are not
 6.14. Roughness of Fracture Surfaces 277

present on the surface, or are blunted by deformation, the mirror surface will not
develop. In addition, the presence of grain boundaries, multiple phases of the
material, and reinforcing particles force the crack paths into irregular shapes.
Systematic investigation of roughness of fracture surfaces and their scaling
properties were first undertaken by Mandelbrot et al. (1984), although Passoja
and Amborski (1978) and Chermant and Coster (1979) had already suggested
that fracture surface of metals may have fractal and scale-invariant properties. As
discussed in Chapter 1, if the width w of a rough surface follows the scaling law
(1.34), then the surface is a self-affine fractal with a fractal dimension Df which,
in d dimensions, is given by
Df = d − α, (20)
where α is the roughness exponent, which is usually the same as the Hurst ex-
ponent H introduced and discussed in Chapter 1, although, theoretically, the two
exponents can be different. Mandelbrot et al. (1984) studied fracture surface of
steel and concluded that the surface possessed fractal morphology. They estimated
the fractal dimension of the fracture surface of their material to be Df  1.28,
implying a roughness exponent α  0.72. If we assume that the roughness expo-
nent α is equivalent to the Hurst exponent H for the fractional Brownian motion
described in Section 1.4, a roughness exponent of 0.72 implies long-range positive
correlations on the fracture surface. Indeed, the profiles of such fracture surfaces
are very similar to fBm with a Hurst exponent H > 0.5 (see Figure 1.2). Since
the original work of Mandelbrot et al. (1984), many other measurements of frac-
tal and self-affine properties of fracture surface of a wide variety materials have
been reported. In particular, several experimental techniques have been used for
measuring and characterizing the roughness of fracture surfaces and estimating its
roughness exponent, which we now describe and discuss.

6.14.1 Measurement of Roughness of Fracture Surface

Underwood and Banerji (1986) measured fractal dimension of fracture surface of
AISI 4340 steels over the temperature range of 200 to 7000◦ C, and found that the
lowest value of Df is at 500◦ C, generally believed to correspond to temper brit-
tleness. Pande et al. (1987) disputed the accuracy of Mandelbrot et al.’s result, and
measured the apparent fractal dimension of fracture surfaces of titanium alloys.
Fractal dimensions of about 1.2 were obtained, implying a roughness exponent
α  0.8. This value is, however, in agreement with many other measurements on
a wide variety of materials discussed below, and with the Molecular Dynamics
simulation results described in Chapter 9, and thus it does not cast doubt on the
measurements of Mandelbrot et al. (1984). Wang et al. (1988) investigated the
relationship between the fractal dimension of a fracture surface and its fatigue
threshold using dual-phase steel, and found roughly a linear relation between the
two. Mu and Lung (1988) measured the fractal dimension Df of fracture surface
of 24SiMnCrNi2 Mo and 30CrMnSiNI2A steels under plane strain. A linear re-
lationship was found between the fractal dimension of fracture surface of these
278 6. Fracture: Basic Concepts and Experimental Techniques

metals and their fracture toughness, such that Df decreased smoothly as the frac-
ture toughness increased. These issues and the progress up to 1988 were reviewed
by Williford (1988).
Mecholsky et al. (1988,1989) and Passoja (1988) studied fracture surfaces of
many solid materials, including several different aluminum and five glass ceramics,
all of which had distinct microstructures. They found that as the toughness of the
materials increases, so does also the roughness of the fracture surface. The fractal
dimension Df was found to be in the range 1.15 − 1.30, with an average of about
1.22, implying an average roughness exponent α  0.78. They also investigated
the relation between fracture energy and the geometry of fracture surface in many
different brittle materials and proposed the following equation
1
= Y ξ(Df − 1), (21)
2
where  is the fracture energy, Y is an elastic modulus, and ξ is a characteristic
length scale of the material.
Dauskardt et al. (1990) undertook a systematic study of five samples of brittle
and ductile transgranular cleavage, intergranular fracture, microvoid coalescence,
quasi-cleavage, and intergranular microvoid coalescence in various steels. These
materials were fractured both at room temperature and also a very low tempera-
ture. They analyzed the measured length L of the surface versus the measuring step
1−D
length Ls which are related through, L ∼ Ls f . In many cases, a fractal dimen-
sion Df  1.2 was obtained, in agreement with the previous estimates discussed
earlier. However, in several other cases the relation between L and Ls was more
complex. Bouchaud et al. (1990) studied fracture of an aluminum alloy in 4 dif-
ferent heat treatment regimes. The fracture surface was elecro-coated with nickel,
then polished and digitized. The correlation function C(r), Eqs. (1.5)–(1.8), was
then constructed for the aluminum-nickel boundary for a large number of sam-
ples. Even though quite different mechanisms of fracture were dominant in these
materials, in all cases the roughness exponent was α  0.8.
Zhenyi et al. (1990) and Dickinson (1991) studied fracture surface of poly-
mers and ceramics, measuring both surface roughness and light emission signals.
Fractal dimensions of 1.2–1.3 were measured for the rough surfaces, resulting
in roughness exponents of about 0.7–0.8. The photon emission signals also had
fractal characteristics, and measurement of their fractal dimensions yielded values
between 1.24 to 1.42, implying roughness exponents in the range 0.6–0.75. Note
that, there appears to be a close relationship between the fractal dimensions of the
fracture surface and that of the emission signals. If the exact nature of this rela-
tionship can be identified, then photon emission signals may provide an accurate
probe of fracture surfaces and their morphology.
Fractures on carbon surfaces were analyzed by Miller and Reifenberger (1992),
who reported that α  0.75. Poon et al. (1992) studied fracture surface of natural
rock, such as sandstone, limestone, and carbonates. For each sample roughness
profiles of several thousand points were constructed, and for all cases studied
a roughness exponent of about 0.8 was obtained. Måløy et al. (1992) investi-
 6.14. Roughness of Fracture Surfaces 279

gated fracture surfaces of six different brittle materials, ranging from Al-Si alloy
AA4253 to porcelain. The materials were notched and then fractured at the tem-
perature at which nitrogen becomes liquid. Many profiles of the rough fracture
surfaces were then obtained and analyzed. Two methods of analysis, including
the power-spectrum method described in Section 1.4.1, were used. The rough-
ness exponent was estimated to be α  0.87 ± 0.07 for all the six samples. Baran
et al. (1992) analyzed fracture surface of several brittle materials, including glass
and dental porcelain, and reported large roughness exponents, ranging from 0.65 to
0.93. Poirier et al. (1992) studied deformation of regular packings of equal parallel
cylinders. The local stress-strain characteristics, at the contact between the cylin-
ders, exhibited a softening part which localized the deformation. The deformation
band was rough with a roughness exponent α  0.73 ± 0.07.
An interesting method for studying fracture surface was developed by Imre et
al. (1992) who determined the fractal dimension of the surface electrochemically
by measuring the diffusion current, also called Cottrell current, at a gold replica
of the fractured metal electrode. (It is interesting to find research groups that are
rich enough to afford gold in their investigations, while others starve for research
funds!) The replicas were prepared by pressing gold wafers into the fractured
steel surfaces in a hydraulic press at high pressure. The gold surfaces were then
cleaned, and the gold electrodes were immersed in an aqueous electrolyte with a
calomel reference electrode. The potential was switched from 0 V to 650 mV for
a short period of time, and then was switched back to 0 V. According to Nyikos
and Pajkossy (1985) the current I (t) should scale with the time t as
I (t) ∼ t (α−2)/2 , (22)
so that simple measurements of I (t) versus t should yield α (and hence Df ).
Roughness exponents of about 0.8 were measured by this method.
Another interesting method for measuring roughness properties of a fracture sur-
face was developed by Friel and Pande (1993). In their method pairs of electron
micrograph images of fracture surface of titanium 6211 at two different inclina-
tion angles (30◦ and 36◦ ) were constructed using a scanning electron microscope
(SEM). The surfaces were fractured under tension. The SEM images were obtained
under various magnifications, ranging from 50 to 10,000. The surface fractal di-
mension was then estimated by measuring the surface area as a function of the
length scale (or measurement resolution), and was found to be about 2.22, implying
a roughness exponent α = 3 − 2.22 = 0.78. Schmittbuhl et al. (1993) measured
roughness exponent of several granitic faults and found α  0.85, close to the val-
ues obtained by others for various materials. E. Bouchaud et al. (1993b) analyzed
the statistics of fracture surfaces of polycrystalline intermediate compound Ni3Al.
Such fracture surfaces also contain secondary branches, as opposed to most of
the fracture surfaces discussed above which had no side branches. Despite this,
E. Bouchaud et al. (1993) could define a roughness exponent for fracture sur-
face of these materials, and their measurements indicated that α  0.8. Lemaire
et al. (1993) put a viscoelastic paste made of sand and resin between two plates
which were driven away from each other at a given velocity until the paste broke.
280 6. Fracture: Basic Concepts and Experimental Techniques

Five different velocities were used, and after fracture the hardened paste was
sliced parallel to the tensile direction. The fractal dimension of the profiles was
then determined by two methods, the standard box-counting method, and by the
power-spectrum methods, both of which were described in Chapter 1. A roughness
exponent α  0.88 ± 0.05 was measured which was independent of the velocity.
Daguier et al. (1995) studied the morphology of fractures in two different metal-
lic alloys. The fractures had been stopped during their propagation by pinning
microstructural obstacles to the surface. One of the alloys was the 8090-Al-Li
which is very anisotropic, for which the roughness exponent was found to be
α  0.6 ± 0.04. The other alloy was Super α2 Ti3Al with a 3D fatigue fracture for
which α  0.54 ± 0.03. Daguier et al. (1996) used atomic force microscopy and
SEM methods to study fracture surface of Ti3Al-based alloys. They found that at
large length scales, and over several decades in length scales, the roughness expo-
nent was α  0.8, whereas at much shorter length scales the roughness exponent
was close to 0.5. Daguier et al. (1997) also studied fracture surface of a silicate
glass as a function of the fracture velocity. At large length scales the roughness
exponent was α  0.78, whereas at smaller length scales α  0.5. The crossover
length scale ξco that separated the two scaling regimes was shown to be propor-
tional to the inverse of the fracture velocity. If hmax is the difference between the
maximum and minimum heights h within a given window on the surface, then the
two scaling regimes could be combined into a single scaling law
hmax ∼ r 0.5 (r/ξco ), (23)
where  is a scaling function with the properties that (x) ∼ 1 as x → 0, and
(x) ∼ x 0.28 for x  1.
Thus, summarizing all the experimental data discussed so far, it appears that at
large enough length scales a roughness exponent α  0.8 represents a universal
value, regardless of the material or even the mechanism of fracture. The possi-
bility of universality of α was first pointed out by Bouchaud et al. (1990). We
should, however, point out that if a fracture surface is analyzed on relatively short
length scales, then the effective value of α may be smaller than 0.8. For example,
Mitchell and Bonnell (1990) analyzed fracture surface of fatigued polycrystalline
copper and reported that α  0.65, while for a single crystal silicon α  0.7 was
obtained. Metallic materials, the roughness exponents of which have been deter-
mined through scanning tunneling microscopy, usually operate in the nanometer
range and have α < 0.8. For example, Milman et al. (1993, 1994) reported a rough-
ness exponent of about 0.6 for fractured tungstene, and close to 0.5 for graphite.
Low cycle fatigue experiments on steel samples on micrometer scales yielded
a roughness exponent close to 0.6 (McAnulty et al., 1992). Low values of the
roughness exponents are interesting because they might be explained based on
models of minimum energy surfaces in disordered environments. Such concepts
were first discussed by Chudnovsky and Kunin (1987), Kardar (1990), Roux and
Francois (1991), and Ertas and Kardar (1992,1993,1994,1996). For example, Roux
and Francois (1991) argued that the path that is selected by a propagating fracture
should be such that the overall fracture energy is minimized. Their simulations un-
 6.14. Roughness of Fracture Surfaces 281

der such a condition led to a roughness exponent in the range 0.4–0.5. The apparent
length-scale dependence of the roughness exponent α may also be explained in
another way based on the velocity of fracture propagation, and whether one is in
the regime of quasi-static or rapid fracture (Bouchaud and Navéos, 1995). This
distinction, its theoretical treatment, and the corresponding roughness exponents
will be described in Chapter 7, where we will also discuss the implication of the
self-affine structure of fracture surface for crack propagation.
Now that there is little doubt that fracture surface of a wide variety of materials
is rough with well-defined characteristics, let us briefly describe how such sur-
faces are studied experimentally. This subject has been discussed in detail by Hull
(1999), and what follows is a summary of his discussion. Roughness is typically
characterized by measuring the height h of the roughness profile. A “primitive”
method would be based on using a raster scan of parallel traverses across the sur-
face using a stylus which traverses parallel to the x-axis—the axis that is parallel
to the mean position of the roughness profile—and measures the height. The stylus
is typically a fine, diamond-tipped needle which is in contact with the surface by a
small external load. The height of the needle is measured using a transducer. The
disadvantage of this method is that the stylus may damage the surface, and hence
create traces that do not belong to the original fracture surface.
Atomic force microscope can also be used which has a highly fine silicon nitride
stylus with a tip radius of about 20–30 nm. The probe is held at a fixed position
from the base of the rough surface, and the surface itself is moved parallel to this
base. The height of the probe is measured from the reflection of light from mirror
on the stylus beam. A powerful feature of this method is that it can determine
roughness parameters on specific sections of the roughness profile.
In a modern version of the stylus technique, the mechanical stylus is replaced
by a fine laser beam that is held at a constant distant from a references surface.
The size of the spot is typically 1 µm in diameter, and the rough surface traverses
under the beam light. The surface shape is then determined from the change in the
length of the light’s path that is reflected from the surface.

6.14.2 Mechanisms of Surface Roughness Generation

There are at least three main mechanisms that give rise to a rough fracture surface.
What follows is a brief description of each mechanism.

6.14.2.1 Growth of Microcracks

In thermoplastic polymers (as well as other materials) the high stresses around
the main crack cause micro-cracking in the material ahead of the main fracture.
These smaller cracks grow and eventually become connected to each other and
to the main crack. As the stress intensity increases, there are corresponding in-
creases in the size of the damage zone and the out-of-the plane crack nucleation.
The net result is a rough fracture surface. Natural materials, particularly rock, ex-
hibit intense micro-cracking and surface roughness (see Sahimi, 1993b, 1995b, for
282 6. Fracture: Basic Concepts and Experimental Techniques

detailed discussions), with the scale of their roughness being equal to at least the
scale of the microstructure. We will come back to this mechanism in Chapter 7,
where we discuss the relation between micro-cracking and dynamics of fracture
propagation.

6.14.2.2 Plastic Deformation

If plastic deformation occurs ahead of the tip of the main growing crack, crack
growth takes place in a zone of deformed material. If the deformation zone is not
homogeneous, the crack path is deflected out of the plane in which it is propagating,
leading to surface roughness. The interaction between deformation processes and
the growing crack depends on the dynamics of growth of the deformation zones
and cracks, which in turn depends on the stress field in the material, and the stress
level at which these phenomena are activated.

6.14.2.3 Macroscopic Branching and Bifurcation

Roughness of fracture surface in isotropic, homogeneous, amorphous and brittle
materials, such as inorganic glasses, might be the result of local changes in the path
of the growing crack. These changes are the result of local instabilities at the tip
of the growing crack. The nature of these instabilities will be discussed in detail in
Chapters 7 and 8. For now it suffices to say that micro-cracks are formed ahead of,
and interact with, the main crack, the nucleation of which can be explained based
on the Griffith criterion. Due to the high stresses that are distributed around the
main growing crack, the micro-cracks are deflected out of the plane of the main
crack by micro-branching or micro-bifurcation, hence giving rise to roughness in
the fracture surface.
However, the growth of micro-cracks ahead of the main crack in brittle glasses
has been disputed by some researchers, who argue that in such materials the stress
to activate very small flaws and grow them into micro-cracks approaches the
theoretical strength of the material, in which case only the main crack grows by
breaking the interatomic bonds. It has been suggested instead that local tilting of
the crack out of its main plane is the cause of micro-branching. These tilted cracks
grow a short distance, but their size increases with the dynamic stress intensity
factor Kd and the crack velocity v. When the dimensions of the tilted cracks
become comparable to the dimensions of the test sample, macroscopic bifurcation
takes place. Experimental evidence for this mechanism was reported by Johnson
and Holloway (1968) and Kulawansa et al. (1993).

6.15 Cleavage of Crystalline Materials

The discussions so far are mostly relevant to brittle fracture of amorphous materi-
als. Another important subject is cleavage of crystalline materials. Single crystals
are homogeneous, but they also contain a degree of anisotropy which assists their
cleavage. To understand this phenomenon, not only does one need information on
 6.15. Cleavage of Crystalline Materials 283

the effective properties of the material, such as their elastic moduli and fracture
toughness, but also an understanding of such micro-deformation processes as slip
that usually precedes and accompanies fracture in crystalline materials. The de-
gree of symmetry that the crystalline material exhibits also plays an important role,
because the strength of the anisotropy of micro-deformation processes depends on
such symmetries. The most important effect of anisotropy is that cleavage may
occur parallel to planes in a crystal that are not normal to the maximum tensile
stress. This is particularly true in crystalline materials that exhibit a low degree
of symmetry, such as mica in which cleavage is only in a single set of planes. In
addition, temperature and strain rates also play important roles by influencing the
mobility of dislocations.
The low surface energy of crystallographic planes, which in turn depends on
the strength of the interatomic bonds, is the main cause of cleavage in crystals. If
cleavage occurs along a single plane, it would produce a featureless surface. How-
ever, often one observes well-defined and crystallographically oriented features
on the fracture surface of a crystalline material. These features are usually caused
by the generation and presence of dislocations that interact with the propagating
fracture. In metals with body-centered cubic symmetry, such as chromium, tung-
sten, and iron, the main cleavage occurs on {001} planes, of which there are three,
(001), (010), and (100). If a cyrstal is tested in an arbitrary direction, the {001}
plane with the largest tensile stress normal to the plane is the most likely place
for cleavage. If a crystal is tested in tension parallel to [011], the (001) and (010)
planes have the same resolved normal tensile stress. In this orientation the stress
on the (011) plane is much greater than on the {100} planes. Thus, fracture may
occur either on an (011) plane, or along the two equally stressed {001} planes.
On the other hand, crystals with the zinc-blende structure, such as gallium ar-
senide, can be described as a cubic unit cell that consists of two interpenetrating
FCC lattices of the two elements (Ga and As). The center of one lattice is at the po-
sition (1/4,1/4,1/4) of the other. These materials are of great industrial importance
because of their use in producing semi-conductors. They cleave on {001} planes,
of which there are three equivalent pairs of orthogonal planes. Slip is restricted to
{111} planes. Such materials usually exhibit strong brittleness.
If polished (001) faces of GaAs crystals are coated with an epitaxial layer of
GaAs that contains a small amount of carbon, tensile stresses are generated in the
surface layers. These stresses then lead to the formation of very fine, atomically
sharp surface cracks (see, for example, Murray et al., 1996). The cracks form on
two orthogonal {011} planes that intersect the (001) surface at right angles, remain
sharp, and grow at very low stresses. The fracture surface is mirror smooth and flat.
However, if GaAs crystals are tested in complex loading conditions, the fracture
surface becomes very rough.
Layered materials usually have very strong bonding within the layers and weak
bonding between the layers. An example is muscovite mica that consists of an
ordered stack of double layers, about 2 nm thick, of strongly bonded planar arrays of
silica tetrahedra held together by Coulomb attraction caused by the potassium ions
between the layers. In such materials cleavage occurs between the weakly-bonded
284 6. Fracture: Basic Concepts and Experimental Techniques

layers, and may also occur through the center of the double layer. Deformation is
restricted to the sliding of layers over each other. The reader should consult Hull
(1999) for extensive discussions of other crystalline materials.
So far we have discussed the cleavage of single crystals. In practice, cleavage
of polycrystalline materials, such as ceramics and rock, is also very important.
Let us briefly discuss these phenomena. We assume that the bonding between the
crystals is very strong, and that the grain boundary interface does not experience
failure.
In polycrystalline materials, each grain is surrounded by many other grains of
different orientations. Therefore, such materials fracture by successive nucleation
and propagation of several cleavage cracks across the boundaries between neigh-
boring crystals. There is a change in the orientation at the grain boundary. If the
angle between the neighboring grains is small, the cleavage crack in a crystal can
propagate across the boundary between the neighboring crystals, in which case
the cleavage plane is tilted and twisted. However, if the orientations of the crystal
grains are very different, the propagation of cleavage from one crystal to another
depends on the relative orientation of the cleavage planes in the crystals.
Consider, for example, two adjacent grains with a common boundary between
them, and suppose that a crack in one of the grains reaches the boundary. Then, it
may stop there with no further crack propagation. Alternatively, the crack may stop
at the boundary, but the high stress at its tip may help nucleate another crack in the
adjacent grain with a different orientation. The two cracks have a common point at
the boundary. The third possibility is having a cleavage plane in the second grain
that is tilted relative to the cleavage plane in the first grain, in which case the crack
propagates continuously across the boundary. Therefore, fracture propagation in
polycrystalline materials depends critically on the distribution of their grains or
single crystals. Even if an array of grains is distributed randomly, the local direction
of crack propagation depends on the relative orientations of the grains at the crack
tip. On the scale of the single crystal size, the main crack path is not straight. It
is also possible that local regions of the crack “tunnel” ahead of the main crack
front because of the existence of a path of favorably oriented single crystals in
the region. If a polycrystalline material contains preferred orientations, then crack
growth in it is easier in some directions than others.

6.16 Fracture Properties of Materials

Let us now describe and discuss important fracture properties of several classes
of materials. In general, one may divide most materials into three distinct classes
which are polymeric materials, metals, and rock-like materials which include con-
crete, rock, glass, and ceramics. We already described fracture properties of glass
when we discussed the mirror, mist and hackle patterns. We do not consider con-
cretes here, and fracture properties of natural rock have been described in detail
elsewhere (Sahimi, 1993b, 1995b). What follows is a brief summary of the prop-
erties of the remaining important materials. Our discussion is not, and cannot be,
 6.16. Fracture Properties of Materials 285

exhaustive, as the mechanical properties of each of these materials are subjects of

separate books.

6.16.1 Polymeric Materials

We described in Chapter 9 of Volume I many important properties of polymeric
materials, and therefore the discussion in this section must be considered as com-
plementary to what was presented there. Since we already discussed the difference
between brittle fracture of amorphous materials and cleavage of crystalline ma-
terials, it is important to understand to what extent a polymeric material can be
crystalline. Although homopolymers are crystalline, due to the length of the chains
in their structure, polymeric materials do not usually have a completely crystalline
structure. Instead, they usually consist of a mixture of crystalline and amorphous
regions. On the other hand, many industrial polymers, such as PMMA, are com-
pletely amorphous, as already mentioned above. Moreover, generally speaking,
random copolymers and cross-linked polymers are also amorphous.
To discuss mechanical and fracture properties of polymers, we consider amor-
phous polymers below the glass transition temperature Tg and crystalline polymers
below the melting temperature Tm . Figure 6.9 shows typical stress-strain curves
(in tension) for polymeric materials. The top curve represents brittle behavior. The
tensile strain is typically about 1–5%. The middle curve exhibits a yield point
and represents ductile fracture. The lowest curve indicates that the yield point
is followed by a strain softening region in which the stress reaches a minimum,
beyond which one has stress hardening which then leads to brittle fracture. The
yield point σy defines the onset of irreversible plastic deformation, and is propor-
tional to the maximum of the true stress in a compression test. Its value depends, of
course, on the composition of the material and the stress configuration. It increases

Figure 6.9. A typical stress-strain diagram for polymers. The top curve corresponds to
brittle behavior, while the middle curve leads to ductile behavior. In the lowest curve, strain
hardening leads to brittle behavior.
286 6. Fracture: Basic Concepts and Experimental Techniques

logarithmically with the strain rate, and slowly decreases with increasing temper-
ature, eventually vanishing at Tg .
Between the elastic limit and the yield point, many polymers that are under ten-
sion exhibit a series of crazes that are normal to the tensile stress. Both amorphous
and crystalline polymers generate crazes with the same features. In particular, it
is easy to see crazes in amorphous polymers as they strongly scatter visible light.
The inside of a polymer craze is typically filled with polymer fibrils, as a result of
which the effective moduli of the material after crazing is only slightly smaller than
before, implying that the onset of crazing cannot be detected on the stress-strain
diagram.
Under tension or compression, polymeric materials can also develop shear
bands, i.e., zones of highly localized shear. The bands are diffuse at high tem-
peratures or low strain rates, but are localized at lower temperatures or higher
strain rates. If the diffuse bands are further deformed, it will lead to ductile frac-
ture, whereas deformation of localized shear bands leads to brittle fracture. If two
shear bands intersect, it usually leads to a craze. The stress at the craze tip can also
lead to the formation of shear bands.

6.16.2 Ceramics
The British Ceramic Society defines ceramic materials as, “All solid manufac-
tured materials or products that are chemically inorganic, except for metals and
their alloys, and which are usually rendered serviceable through high temperature
processing.” Ceramic materials include borides, carbides, halides, nitrides, oxides,
and cermets, which are ceramic metals. They usually have a crystalline structure,
but can also be found in amorphous form. The interatomic bonds in ceramics may
be ionic, covalent, metallic, and van der Waals. It is clear how the first two types of
bonds may form in ceramics. Metal transition carbides have bonds which have a
metallic characteristic in that, valence electrons are freely shared by all the atoms
in the structure.
Relative to metals, ceramics have large elastic moduli, ranging from 70 to 400
GPa. The moduli decrease very slowly with increasing temperature. They also
have a large cohesive strength which is due to the fact that their interatomic bonds
require high energies to be broken. However, as discussed earlier in this chapter,
the presence of defects, which results in stress concentration, reduces the actual
strength of these materials. In fact, the fracture strength σf of ceramics is very
sensitive to the presence of defects, the porosity, the shape and size of the grains, as
wells as the pore-crack combination. Most importantly, σf depends on the size of
the defects, for which there is a critical size that, at a given stress, leads to fracture.
For ceramics this size can be as large as a single crystal. The Weibull distribution
[see Eq. (5.37); see also Chapter 8] usually describes well the statistical distribu-
tion of the fracture strength of ceramics. Moreover, if the defects are uniformly
distributed in the material, the probability of having the critical condition in the
material for fracture is relatively large. Experiments have indicated that in many
ceramics, especially those that have a secondary phase, the crack velocity v is
related to the stress intensity factor KI by, v ∼ KIn , where n is a constant.
 6.16. Fracture Properties of Materials 287

Experiments by Buresch et al. (1983) and others have also shown that the fracture
strength of certain ceramics depends on the critical value σn of the notch fracture
stress, and also on the size of the cohesive zone (see above). The cohesive zone in
ceramics is somewhat similar to the plastic zone in metals in that, the microcracked
zone in the immediate vicinity of a crack tip causes the nonlinear behavior of
ceramics. In this zone, there is a constant stress σn for breaking either the grain
boundaries or the crystal themselves, which depends on the cohesive stress σc (see
above). If the average stress in the cohesive zone reaches σn , instability occurs in
the material and the main crack propagates.
The behavior of the fracture strength σf of ceramics with variations in the tem-
perature can be divided into two groups. In one group, σf decreases monotonically
with increasing temperature. Nitrides typically exhibit this behavior. In the second
group, the fracture strength either stays constant with increasing temperature, or
first experiences a small increase and then decreases. Ceramics that do not have a
secondary phase at their grain boundaries exhibit this behavior.

6.16.3 Metals
Most metals have simple crystalline structures in the form of BCC or FCC lattices
or a hexagonal close-packed (HCP). At the atomic scale the interatomic bonds
break either along crystallographic plane in Mode I fracture (i.e., in a direction
normal to the plane), or in Modes II and III fracture (i.e., in a direction parallel to
that plane), which is also the mechanism for cleavage fracture already described
above. Alternatively, metals fracture at high temperatures by coalescence of cavi-
ties. Single crystals of a HCP metal (for example, zinc) can slip on a single plane
until the two parts completely separate. Usually, however, multiple slip occurs
in single crystals which generates a neck in the material which is under tension.
These necks usually initiate at inclusions which do not deform in the same way
as the metallic matrix. In polycrystalline metals, necking occurs in a more diffuse
fashion, but can lead to the complete separation of the two halves of the material
when the neck’s cross section vanishes. This mechanism is, however, rare. In most
cases, the material breaks much sooner by developing a crack in the middle of
the neck which is perpendicular to the tensile axis, which at the end tilts to a 45◦
orientation. This crack is the result of coalescence of vacancies which grow due to
plastic deformation and elongate in the direction of the maximum principal strain.
This mechanism of fracture in metals is ductile because it involves large local
slip deformation. It also often corresponds to a large macroscopic plastic de-
formation. However, coalescence of the vacancies does not always need large
macroscopic deformation, such as when the volume fraction of the inclusions in a
metal-matrix composite is large. Hydrostatic pressure prevents the growth of the
cavities, whereas a tensile positive hydrostatic stress increases it, and thus reduces
greatly the fracture strain.
Another mechanism of fracture of metals is intergranular cracking, which hap-
pens when the grain boundaries are weaker than their interior. The weakness is
caused by impurities that have accumulated at the grain boundaries. In such a situ-
ation, the cracks preferentially follow the grain boundaries, leading to intergranular
288 6. Fracture: Basic Concepts and Experimental Techniques

fracture, sometimes referred to as dimple fracture. If the temperature of the system

is high enough, then the vacancies migrate by diffusion, and then coalesce to create
cavities and ultimately cracks. This mechanism is called creep cracking.
Cyclic straining of metals also results in fatigue fracture of metals which usually
starts on the surface, and is generated by irreversible localized shear deformations
(see above). The surface gradually develops roughness which, if strong enough,
develops into a crack which then penetrates into the material along the shear
direction.

6.16.4 Fiber-Reinforced Composites

These materials exhibit a wide variety of fracture modes, including rupture of in-
dividual fibers, interfacial debonding, matrix cracking and delamination. Various
experiments involving X-ray radiography and optical and scanning microscopy
indicate that if a unidirectional fiber-reinforced composite is loaded in the longi-
tudinal direction (parallel to the fibers), the fracture process consists of four main
stages.
(1) At less than 50% of the final load, individual fibers break at random.
(2) As the broken fibers accumulate, they join and form macroscopic cracks
throughout the material.
(3) Delamination begins parallel to the fibers, starting at the large cracks.
(4) Delamination propagates parallel to the direction of the fibers.
If the composite material is subjected to a static tensile load in the longitudinal
direction, the breaking of a fiber generates tensile stress concentration in the first
unbroken fiber, which may lead to their breaking. In addition, shear stress con-
centration is generated at the interface between the broken fiber and the matrix
which contributes to shear debonding along the fiber surface. Thus, breaking of
fibers induces two types of fracture modes that proceed simultaneously. The vol-
ume fraction of the fibers and their orientations control which of the two modes
is the dominating one. If the fibers are distributed closely, the fracture propagates
from fiber to fiber, whereas when they are relatively far apart, the failure process
proceeds along individual fiber surfaces in the shear fracture mode. Fiber mis-
alignment, or fiber waviness, also influences the tensile strength of the composite.
In fact, the broader the distribution of fiber misalignment, the smaller is the tensile
strength of the composite materials.

6.16.5 Metal-Matrix Composites

The fracture strength of metallic materials can be improved by inserting into them
short fibers or particles. A typical failure process in such materials involves,
(1) failure of the interface between the fibers and the matrix at the tip of the fiber;
(2) growth of a cavity within the matrix, beginning at the fiber tip;
 6.16. Fracture Properties of Materials 289

(3) coalescence of the cavities due to plastic deformation and formation of a crack,
and
(4) propagation of the crack.
Inside the metallic matrix the failure is ductile, but it appears brittle at the
macroscopic length scales.
An important factor is the aspect ratio of the fibers, i.e., the ratio of their lengths
and diameters. For example, fibers reinforce a material better than spherical parti-
cles. The larger the aspect ratio, the higher is the fracture strength of the composite.
However, if the fibers become too long, they will no longer influence the strength
of the material. The properties of the interface between the fibers and the matrix
also have a very strong effect on the strength of the composite. If the interface is
stronger, the composite material will have a lower ductility and a higher fracture
strength. During production of the composite, internal stresses may be produced
by mismatch between the thermal expansion coefficients of the matrix and the
fibers. Thus, when the temperature of the system is reduced, residual stresses are
produced in the composite which, however, disappear at high enough tempera-
tures. In addition, one may have chemical reaction at the interface. If, for example,
the fibers are oxidized, the fracture strength of the composite will reduce.

Summary
The aim of this chapter was to define the basic concepts of fracture mechanics, and
describe and discuss the basic phenomena that occur during fracture of materials.
These concepts will be utilized in the next few chapters where we describe and
discuss modeling of brittle fracture of heterogeneous materials and its transition
to ductile behavior. We also described the experimental techniques that are used
for measuring important characteristics of fracture of materials, such as the speed
of crack propagation, and measurement and analysis of roughness of a fracture
surface.
7
Brittle Fracture: The Continuum
Approach

7.0 Introduction
As discussed in the last chapter, fracture of brittle amorphous materials is a difficult
problem, because the way a large piece of a material breaks is closely related to
details of cohesion at microscopic length scales. For this reason alone, description
of brittle fracture of materials has been plagued by conceptual puzzles. What made
matters worse for a long time was the fact that many past experiments seemed to
contradict the most firmly-established theoretical results. However, considerable
progress has been made over the past decade, and one main aim of this chapter
is to demonstrate that the theory and experiments fit within a consistent picture.
This has become possible by the realization that dynamic instabilities of the tip of
a fracture play a critical role in determining the fracture behavior of amorphous
materials. To accomplish this goal, we follow our by-now-familiar path, namely,
we first describe and summarize the central results of continuum theories of linear
elastic dynamic fracture mechanics which provides an elegant and powerful de-
scription of fracture propagation. However, the continuum theory is unable to make
quantitative predictions without additional information that must be provided by
experiments, or be supplied by other types of theories. We already discussed in
the last chapter some of the most important experimental observations and data,
and the techniques that were used for obtaining them. These experiments teach us
that when the flux of energy to a fracture tip exceeds a critical value, the fracture
becomes unstable and hence propagates in an increasingly complex manner. As a
result, the moving crack cannot travel as quickly as the linear continuum theory
predicts or assumes, the fracture surface becomes rough and begins to branch out
and radiate sound, and the energy cost for the motion of the crack increases signif-
icantly. These observations are completely consistent with the continuum theory,
but cannot be described by it. Therefore, to complete the emerging theoretical
picture and the fundamental understanding of this phenomenon, we continue this
chapter with an account of theoretical and numerical work of the past decade or
so that attempts to explain the dynamic instabilities in fracture propagation. As
discussed in the last chapter, our current experimental understanding of instabili-
ties in fracture tip in brittle amorphous materials is fairly detailed. We also have
a rather detailed theoretical understanding of these instabilities in crystals which
reproduces many qualitative features of the experiments. Recent numerical work
 7.0. Introduction 291

is attempting to establish the missing connections between the experiments and

the theory.
Up until a decade or so ago, most engineers and materials scientists believed
that the development of continuum fracture mechanics is largely complete. Why?
Because this field is in fact one of the most heavily developed branches of engi-
neering science. We only need to consider how many books and review articles
have been written on this subject to appreciate this fact. The development of con-
tinuum fracture mechanics actually emerged from mathematical exercises in the
early part of the 20th century into a coherent collection of theoretical concepts and
experimental techniques that are now widely used to ensure the safety of critical
structures, ranging from aircraft to microelectronic devices. Despite considerable
progress, two important and puzzling features of the problem kept researchers
attracted to fracture of brittle materials. The first feature is that it is often stated
that propagating fractures do not reach the limiting velocity predicted by linear
continuum mechanics of fracture propagation, and that they have a seemingly
unexplained instability at a critical velocity of propagation which is between the
prediction of the linear theory and the experimental data. In fact, only about a
decade ago, Freund (1990) specifically mentioned in his book (pp. 37–38) in a
short list of phenomena (associated with dynamic fracture) entitled “not yet com-
pletely understood” the apparent terminal fracture speed well below the Rayleigh
wave speed in glass and some other very brittle materials. The Rayleigh wave
speed cR is the speed at which sound travels over a free surface. The root cause
of this apparent inconsistency is in the energy dissipation at the fracture tip and,
as we discuss in this chapter, recent work indicates that when energy flux into a
crack tip exceeds a certain critical value, efficient and steady motion of the tip
becomes unstable to the formation of microfractures that propagate away from the
main fracture. In fact, the tip undergoes a hierarchy of instabilities which increases
enormously its ability to absorb energy. The second feature is the need for under-
standing how materials break at the atomic length scale. To understand this aspect
of the problem one must resort to molecular dynamics (MD) simulations which
enable one to model generation of fracture and their motion one atomic bond at a
time. However, MD simulations require extensive and very time consuming com-
putations. To make the simulations efficient and cost effective, a sound strategy
is perhaps to study the existing analytical results so as to understand the quali-
tative effect of atomic discreteness on crack motion. Once this understanding is
acquired, many experimental results become understandable, the relation between
simulations and experiments becomes clearer, and therefore MD simulations will
be much more efficient. We will describe MD simulations of fracture propagation
in Chapters 9 and 10.
These puzzling, and theoretically challenging, features of dynamics of brittle
fracture of materials have motivated a considerable amount of work in this research
area, especially by physicists and their allied scientists. Their work has helped the
emergence of a much clearer picture of fracture dynamics which indicates that
the two puzzling features of fracture dynamics, at both atomic and macroscopic
length scales, are in fact manifestations of the same underlying phenomenon. One
292 7. Brittle Fracture: The Continuum Approach

goal of this chapter is to explain how these puzzles have arisen, and how to recast
them in new terms and explain them. We do not intend to provide a complete
review of fracture mechanics as it will require a book by itself. Instead, we focus
on brittle materials. Ductility and dynamic elasto-plastic fracture, which is a well-
developed field, have been described well by others (see, for example, Freund,
1990; Chan, 1997). Therefore, we will discuss only the brittle-to-ductile transition.
The emphasize in this chapter is first to describe and summarize the most important
predictions of the conventional continuum fracture mechanics, and then answer
some fundamental and interesting questions that this type of models do not ask or,
if they do, cannot answer. To write a significant portion of this chapter, we relied
heavily on the excellent review of this subject by Fineberg and Marder (1999).
Some of the developments that we discuss had been described in an earlier article
by the author (Sahimi, 1998), and thus have also been utilized in this chapter.

7.1 Scaling Analysis

Before embarking on a detailed analysis of fracture of materials, we carry out
some preliminary scaling analysis of this problem. Although our analysis is too
simple-minded, it does point to some fundamental properties of materials, and
does exhibit some basic problems that a detailed analysis of fracture propagation
must address. We consider both the static and dynamical cases.

7.1.1 Scaling Analysis of Materials Strength

Despite what most of us believe (and apparently feel), the world is farther from
equilibrium than we realize. To see this, consider a piece of rock of area S and
height h. Equilibrium principles teach us that the rock should not be able to sustain
its own weight under the force of gravity, if it becomes too tall. To estimate the
critical height, recall that the gravitational potential energy of the rock is 12 ρSh2 g
where ρ is the rock’s density. If we cut the rock into two equal blocks of height 12 h
and set them side by side, this energy is reduced to 14 ρSh2 g, resulting in an energy
gain of 14 ρSh2 g. The cost in energy of the cut is the same as the cost of creating
new rock surface, the characteristic value of which per unit area is, H = 1 J/m2 .
If we assume a typical value, ρ = 2000 3
√ kg/m , the critical height hc at which it
pays to divide the rock in two is, h = 4H/ρg ∼ 1.4 cm, so that every block of
rock more than about 2 cm tall is unstable under its own weight. Similar scaling
analyses are applicable to steel or concrete. Thus, although things fall apart when
they reach equilibrium, the time to reach this state is fortunately long.
Since the fact that most objects do not fall apart easily is an indication
that they are out of mechanical equilibrium, one must estimate the size of
the energy barriers that hold them in place. A rough estimate is obtained by
imagining what happens to the atoms of a solid material as one pulls it uni-
formly at two ends. Initially, the forces between the atoms increase, but they
eventually reach a maximum value, at which the material breaks into pieces.
 7.1. Scaling Analysis 293

As is well known, interatomic forces vary greatly between different elements and
molecules (see Chapter 9), but they typically attain their maximum value when
the distance between atoms increases by about 20% of its original value. The force
needed to stretch a solid material slightly is, F = Y Sδ/L, where Y is the Young’s
modulus, L is its initial length, and δ is the amount (in length) that the material
has been stretched. Therefore, the force per unit area needed to reach the breaking
point is about, σc = F /S = Y δ/L  Y /5, where we have used δ/L = 0.2. We
list in Table 7.1 values of Y for several materials, the theoretical strength σc , and
its comparison with the experimental data. As this table indicates, the theoretical
estimate of σc is in error by orders of magnitude. The scaling estimate of hc greatly
underestimates the practical resistance of solid materials to fracturing, while the
estimate of σc too large. What is the problem? The only way to discuss the correct
orders of magnitude is to account for the actual dynamical mode by which brittle
materials fail mechanically, which is by propagation of a fracture.
As described in Chapter 6, and shown later in this chapter, the presence of a
fracture in an otherwise perfect material results in a stress singularity at the frac-
ture tip. If the fracture tip is atomically sharp, a single fracture which is a few
microns long suffices for explaining the large discrepancies between the theoret-
ical and experimental material strengths that are shown in Table 7.1. The stress
singularity that develops at the tip of a fracture focuses the energy that is stored
in the surrounding material and uses it efficiently for breaking one atomic bond
after another. Thus, continuous fracture propagation provides an efficient way of
overcoming the energy barrier between two equilibrium states of the system that
have different amounts of mechanical energy. We now turn to a scaling analysis
of dynamic fracture (Fineberg and Marder, 1999).

7.1.2 Scaling Analysis of Dynamic Fracture

An analysis of rapid fracture was first carried out by Mott (1948) whose analysis
was slightly improved by Dulaney and Brace (1960). Mott’s work is a dimensional

Table 7.1. The experimental strength σe of a

number polycrystalline or amorphous materials,
and their comparison with the corresponding
theoretical strength σc . Y is the materials’
Young’s modulus (adopted from Fineberg and
Marder, 1999).
Material Y σc σe σe /σc
(GPa) (GPa) (GPa)
Iron 195-205 43-56 0.3 0.006
Copper 110-130 24-55 0.2 0.005
Titanium 110 31 0.3 0.009
Silicon 110-160 45 0.7 0.01
Glass 70 37 0.4 0.01
Plexiglas 3.6 3 0.05 0.01
294 7. Brittle Fracture: The Continuum Approach

Figure 7.1. Propagation of a fracture of length l at velocity

v in an infinite plate disturbs the material up to a distance
l (after Fineberg and Marder, 1999).

analysis which, despite being wrong in many of its details, clarifies the basic
physical processes. It consists of writing down an energy balance for the motion
of a fracture. Consider a fracture of length l(t) growing at time t at rate v(t) in a
very large plate to which a stress σ∞ is applied at its far boundaries; see Figure
7.1. As the fracture extends, its faces separate, causing the plate to relax within a
circular region centered at the middle of the crack with a diameter which is of the
order of l. The kinetic energy Hk involved in moving a piece of material of this
size is 12 mv 2 , where m is the total mass, and v is a characteristic velocity. Since
the mass of the moving material is proportional to l 2 , the kinetic energy should be
given by
Hk (l, v) = ck l 2 v 2 , (1)
where ck is a constant. The moving portion of the material is also where elastic
potential energy is being released as the crack propagates. This stress release results
in a gain in the potential energy which is given by
Hp (l) = −cp l 2 , (2)
where cp is another constant. Equations (1) and (2) are correct if the crack moves
slowly, but they fail even qualitatively if the fracture velocity approaches the speed
of sound, in which case Hk and Hp both diverge. Their divergence will be demon-
strated below, but let us assume for now that it is true. The final piece of the energy
balance is the contribution of creation of new fracture surfaces. This contribution
is l, where  is the fracture energy that, as described in Chapter 6 (see Section
6.7), accounts for the minimum energy needed to break the atomic bonds and any
other dissipative processes that the material may need in order for the fracture
to propagate, and is often orders of magnitude greater than the thermodynamic
surface energy. Therefore, the total energy H of the system containing a fracture
 7.1. Scaling Analysis 295

is given by
H(l, v) = ck l 2 v 2 + Hqs (l), (3)
where Hqs is the quasi-static part of the total energy given by
Hqs (l) = −cp l 2 + l. (4)
If a crack moves forward slowly, its kinetic energy will be negligible, and therefore
only Hqs will be important. For small fractures, l, the linear cost of fracture
energy, is always greater than the quadratic gain of the potential energy, Hp = cp l 2 .
In fact, such fractures would heal (move backward) if such irreversible processes
as oxidation of the crack surface did not prevent them from healing. The fact that
the fracture grows is due to additional irreversible processes, such as chemical
attack on the crack tip (see Chapter 6), or vibration and other irregular mechanical
stresses. Eventually, at a critical length lc , the energy gained by relieving elastic
stresses in the material exceeds the cost of creating new fracture surfaces, in which
case the crack is able to extend spontaneously. Clearly, at lc , the energy functional
Hqs (l) has a quadratic maximum. The Griffith criterion (Griffith, 1920; see Chapter
6 and also below) for the onset of fracture is that fracture occurs when the potential
energy released per unit crack extension equals the fracture energy . Thus, fracture
in this system occurs at a critical crack length lc such that, dHqs /dl = 0 at l = lc .
Using Eq. (4) we find that,

lc = , (5)
2cp
so that
Hqs (l) = Hqs (lc ) − cp (l − lc )2 . (6)
The most important issue in engineering fracture mechanics is calculating lc ,
given such information as the external stresses which, in the present case, is rep-
resented by the constant cp . Dynamic fracture begins in the next instant, and since
it is very rapid, the energy H of the system is conserved, remaining at Hqs (lc ).
Thus, from Eqs. (3) and (6) we obtain
$ % $ %
cp lc lc
v(t) = 1− = vm 1 − , (7)
ck l l
which predicts that fracture propagation will accelerate until it approaches the
maximum speed vm . Equation (7), and more generally the above scaling analysis,
cannot by themselves predict vm , but Stroh (1957) argued correctly that vm should
be the Rayleigh wave speed cR , although his suggestion was implicitly contained
in the earlier calculations of Yoffe (1951) (see below).
In this system, one needs only to know the length lc at which a fracture begins
to propagate in order to predict all the ensuing dynamics. As we discuss later
in this chapter, Eq. (7) is actually very close to anticipating the results of a far
more sophisticated analysis, which is surprising since the Eqs. (1), (3) and (4)
for the kinetic and potential energy are in fact incorrect because they actually
296 7. Brittle Fracture: The Continuum Approach

diverge as the speed of fracture propagation approaches the Rayleigh wave speed
cR . However, the success of Eq. (7) is due to the fact that it involves the ratio
Hp /Hk . Since the divergence of the kinetic and potential energy are according to
exactly the same forms, the errors involved in their estimation cancel each other
out. We now attempt to review and discuss the background, basic formalism and
underlying assumptions that form the basis for continuum fracture mechanics.

7.2 Continuum Formulation of Fracture Mechanics

The general strategy in continuum fracture mechanics is to solve for the dis-
placement fields in the material subject to both the boundary conditions and the
externally applied stresses. The elastic energy transmitted by the displacement
fields is then matched to the amount of energy dissipated throughout the material,
which results in an equation of motion. The only energy sink in a single moving
fracture is at the tip of the fracture itself. Thus, an equation of motion for a mov-
ing fracture is obtained if detailed knowledge of the dissiption mechanisms in the
vicinity of the fracture tip is available.

7.2.1 Dissipation and the Cohesive Zone

As discussed in Chapter 6, the processes that give rise to energy dissipation in
the vicinity of the crack tip are complex and, depending on the material, vary
from dislocation formation and emission in crystalline materials to the complex
unraveling and fracture of intertangled polymer strands in amorphous polymers.
Fracturing and the complex dissipative processes occurring in the vicinity of the
crack tip occur due to very large values of the stress field as one approaches the
tip. As discussed below, if the material around the crack tip were to remain linearly
elastic until fracture, the stress field at the crack tip would actually diverge. Since
a real material cannot support such singular stresses, the assumption of linearly
elastic behavior in the vicinity of the tip must break down and material-dependent
dissipative processes must begin playing an important role. Given the enormous
variety of materials, the emergence of material-dependent dissipative processes
might indicate that a universal description of fracture is impossible. However, as
described and discussed in Chapter 6, Orowan (1955) and Irwin (1956) developed
a way around this difficulty by suggesting independently that the region around
the fracture tip should be divided into three separate regions which, as described
in Section 6.9, are as follows.
(1) The cohesive zone (also called the process zone), which is the region im-
mediately surrounding the fracture tip in which all the nonlinear dissipative
processes that allow a crack to move (forward) are assumed to occur. In contin-
uum fracture mechanics detailed description of this zone is avoided. Instead,
this zone is simply characterized by the energy , per unit area of crack exten-
sion, that it consumes during fracture propagation. As discussed in Chapter 6,
 7.2. Continuum Formulation of Fracture Mechanics 297

the size of the cohesive zone depends on the material, ranging from nanometers
in glass to microns in brittle polymers.
(2) The universal elastic region, which is the region outside the cohesive zone
for which the response of the material can be described by linear continuum
mechanics. Outside the cohesive zone, but in the vicinity of the fracture tip,
the stress and strain fields take on universal singular forms which depend only
on the symmetry of the externally applied loads. In two dimensions (2D) the
singular fields surrounding the cohesive zone are completely described by
three constants which are the stress intensity factors introduced and discussed
in Section 6.8 (see also below). They incorporate all the information regarding
the loading of the material.
(3) The outer elastic region far from the crack tip in which stresses and strains are
described by linear elasticity. Details of the solution to the stress field in this
region of materials depend only on the locations and strengths of the loads,
and the shape of the material. For some special cases analytical solutions have
been derived. Deriving such solutions is made possible by the fact that, so far
as linear elasticity is concerned, viewed on macroscopic scales, the cohesive
zone can be represented by just a point at the fracture tip, while the fracture
itself is equivalent to a branch cut. In general, however, one must resort to
numerical simulations and solutions.
The dissipative processes within the cohesive zone determine the fracture energy
. If no dissipative processes other than the direct breaking of the atomic bonds
take place, then  is a constant which depends on the bond energy. In general
though,  is a complex function of both the fracture velocity and history, and
differs by orders of magnitude from the surface energy—the amount of energy
required to sever a unit area of atomic bonds. No general first principle description
of the cohesive zone exists, although numerous models have been proposed (see,
for example, Lawn, 1993).

7.2.2 Universal Singularities near the Fracture Tip

As one approaches the tip of a fracture in a linearly elastic material, the stress
field surrounding the tip develops a square root singularity (in the distance r).
As mentioned in Section 6.8, Irwin (1958) noted that the stress field at a point
(r, θ) near the fracture tip, measured in polar coordinates with the fracture line
corresponding to θ = 0, can be represented by
Kβ β
σij = √ fij (v, θ ), (8)
2π r
where v is the instantaneous crack velocity, and β is an index that represents
Modes I, II and III of fracture described in Section 6.6. For each of these three
β
symmetrical loading configurations, fij (v, θ ) is a known universal function. The
coefficients Kβ is the stress intensity factor, introduced in Chapter 6, that contains
all the detailed information about sample loading and history, and is determined by
298 7. Brittle Fracture: The Continuum Approach

the elastic fields throughout the material. However, the stress that locally drives the
fracture is one which is present at its tip. Thus, Kβ determines entirely the behavior
of a fracture, and much of the study of fracture processes is aimed toward either
calculating or measuring this quantity. The universal form of the stress intensity
factor allows a complete description of the behavior of the tip of a fracture where
one needs only carry out the analysis of a given problem within the universal elastic
region (see below). For arbitrary loading configurations, the stress field around the
fracture tip is given by three stress intensity factors Kβ which lead to a stress field
that is a linear combination of the pure Modes:

3
Kβ β
σij = √ fij (v, θ ). (9)
β=1
2π r

As mentioned above, the stress intensity factors are related to the flow of energy
into the fracture tip. Since a fracture may be viewed as a means of dissipating built-
up energy in a material, the amount of energy flowing into its tip must influence
its behavior. Irwin (1956) showed that the stress intensity factor is related to the
energy release rate H, defined as the amount of energy flowing into the crack tip
per unit fracture surface formed. The relation between the two quantities is given
by

3
1 − νp2
H= Aβ (v)Kβ2 , (10)
Y
β=1

where νp is the Poisson’s ratio of the material, and the three functions Aβ (v)
depend only on the fracture velocity v. Equation (10) is accurate when the stress
field near the tip of a fracture can be accurately described by Eq. (8), which is the
case as the dimensions of the sample increase.

7.3 Linear Continuum Theory of Elasticity

Since most of the theoretical work that we describe in this chapter is carried out
in 2D (or quasi-2D) systems, we follow the analysis presented by Fineberg and
Marder (1999) who performed a reduction of the full 3D elastic description of a
fracture to 2D in three important cases: For Mode III fracture, and Mode I fracture
in very thin and very thick plates. As noted in Section 6.6, Mode III fracture is
an important model system for which much analytical work has been carried out,
resulting in deeper gains in understanding qualitative features of fracture. The
second case, Mode I fracture of a thick plate, describes stress and strain conditions
of importance in describing the phenomenon in the immediate vicinity of the
fracture tip. The third case, Mode I fracture in thin plates, corresponds to much of
the experimental work that was described in Chapter 6, some of which will also
be considered in the present chapter.
 7.3. Linear Continuum Theory of Elasticity 299

As already described and discussed in detail in Chapter 7 of Volume I, the starting

point is the Navier equation of motion for an isotropic elastic material:
∂ 2u
ρ = (λ + µ)∇(∇ · u) + µ∇ 2 u, (11)
∂t 2
where u is the displacement field for each mass point relative to its original location
in an unstrained material, and ρ is the density. The constants µ and λ are the usual
Lamé constants (with dimensions of energy per volume and typical values of order
of 1010 erg/cm3 ). We also define the linear elastic strain tensor with components
$ %
1 ∂ui ∂uj
ij = + . (12)
2 ∂xj ∂xi
If a linear stress-strain relation exists in a homogeneous and isotropic material, the
components σij of the stress tensor are defined by


σij = λδij εkk + 2µij . (13)
k
The simplest analytical results are obtained for pure Mode III. The only non-
zero displacement is uz = uz (x, y) alone. Thus, the only non-vanishing stresses
are, σxz = µ∂uz /∂x, and, σyz = µ∂uz /∂y. The governing equation for uz is the
ordinary wave equation,
1 ∂ 2 uz
= ∇ 2 uz , (14)
c2 ∂t 2
√
where c = µ/ρ.
Consider now Mode I fracture in a sample material that is extremely thick along
the z-direction. All the applied forces are uniform in this direction. Because all the
derivatives with respect to z vanish, all the fields are functions of x and y alone, so
that one deals with a plane strain problem. The reduction of the problem to 2D is
simple, but this geometry is not convenient for experiments. A third case in which
the equations of elasticity reduce to 2D is the plane stress problem in which one
pulls on a thin plate in Mode I. If the length scale over which the stresses vary in x
and y is large compared with the thickness of the plate along the z-direction, then
we might expect the displacements in that direction to quickly reach equilibrium
with the local stresses. If the Poisson’s ratio is positive, then when the material
is stretched, the plate will contract in the z-direction, and if it is compressed, the
plate will thicken. (Counter-examples, when the material expands under stretching,
were described in Section 9.8 of Volume I.) Under this condition, ux and uy are
independent of z, and therefore it is reasonable to assume that,
uz = zf (ux , uy ). (15)
The function f can be found by realizing that the stress σzz must vanish on the
face of the plate, implying that at the surface of the plate we must have
$ %
∂ux ∂uy ∂uz
λ + + (λ + 2µ) = 0, (16)
∂x ∂y ∂z
300 7. Brittle Fracture: The Continuum Approach

which means that

$ %
∂uz λ ∂ux ∂uy
f (ux , uy ) = =− + , (17)
∂z λ + 2µ ∂x ∂y
so that
$ %
∂ux ∂uy ∂uz 2µ ∂ux ∂uy
+ + = + . (18)
∂x ∂y ∂z λ + 2µ ∂x ∂y
Therefore,
$ %
∂uγ ∂uα ∂uβ
σαβ = λ̃δαβ +µ + , (19)
∂xγ ∂xβ ∂xα
with
2µλ
λ̃ = , (20)
λ + 2µ
and α and β now run only over x and y. Therefore, a thin plate satisfies the equations
of 2D elasticity, with an effective constant λ̃, so long as uz is dependent upon ux
and uy according to Eqs. (15)–(17). In the following discussion, the tilde over λ is
dropped with the understanding that the relation to 3D materials properties is given
by Eq. (20). The equation of motion is still the Navier equation, but is restricted
to 2D.
Note that, as described in Chapter 7 of Volume I, materials are frequently de-
scribed by the Young’s modulus Y and the Poisson’s ratio νp , in terms of which
we have
Y νp Y Y
λ= , λ̃ = , µ= . (21)
(1 + νp )(1 − 2νp ) 2(1 − νp )2 2(1 + νp )
Moreover, note that


∇ · u = (λ + 2µ) σαα , (22)
α

and that from Eq. (11) one finds that

ρ ∂ 2 σαα
= ∇ 2 σαα . (23)
λ + 2µ ∂t 2
Therefore, ∇ · u satisfies the wave equation with the longitudinal wave speed
1
cl = (λ + 2µ), (24)
ρ
whereas, while ∇ × u also satisfies the wave equation, it does so with the shear
(compressional) wave speed
!
ct = µ/ρ. (25)
One must also consider the transition from 2D to 3D. Near the tip of a fracture in
a plate stresses become large enough that the approximations leading to 2D plane
 7.3. Linear Continuum Theory of Elasticity 301

stress elasticity fail (Nakamura and Parks, 1988). If the thickness of the plate along
the z-direction is denoted by d, then at distances from the fracture tip that are much
larger than d all fields are described by equations of plane stress. At distances from
the fracture tip that are much less than d, and away from the x − y surfaces of the
plate, the fields solve the equations of plane strain.

7.3.1 Static Fractures in Mode III

If one inserts an elliptical crack in a plate and pulls it, then, as discussed in Section
6.7, Inglis (1913) was the first to derive the expression for the stresses at the crack’s
narrow ends, and found that they are much larger than those exerted off at infinity.
Therefore, a crack acts as an amplifier of the stresses and causes the elastic energy
to be preferentially focused into its tip, implying that the existence of a crack leads
to a large decrease in the effective strength of a material. The ratio of the maximum
to the applied stress is
Maximum stress l
=2 , (26)
Applied stress Y
where l is the crack’s length and Y the radius of curvature at its tip. Thus, if one
assumes that typical solids have fracture tips of size 1 Å and length of 104 Å,
then one can account for the discrepancies shown in Table 7.1. To derive Eq. (26)
we assume that the stresses applied to the plate coincides with the conditions of
anti-plane shear stress, so that the only non-zero displacement is uz . From Eq. (14)
one sees that the static equation of linear elasticity is now simply the Laplace’s
equation, ∇ 2 uz = 0. For our boundary value problem conformal mapping is the
appropriate technique. Since uz is a solution of the Laplace’s equation, it can be
represented by
1
uz = [φ(ζ ) + φ(ζ )], (27)
2
where φ is analytic, ζ = x + iy, and φ̄ is the complex conjugate of φ.
Far from the crack, the displacement uz (x, y) increases linearly with y, and
therefore we must have the asymptotic property that
φ = −icζ. (28)
Although the constant c of Eq. (28) is dimensionless, in essence it measures the
stress in units of the Lamé constant µ. Because the crack’s edges are free, the stress
normal to the edge must vanish. It can then be shown that
φ(ζ ) = φ(ζ ), (29)
when ζ lies on the boundary. To illustrate the use of Eq. (29), let us define ω such
that
l
ζ = . (30)
2(ω + m/ω)
When ω lies on the unit circle (i.e., ω = eiθ , with θ real), ζ traces out an elliptical
302 7. Brittle Fracture: The Continuum Approach

boundary. When m = 0, the boundary is a circle of radius l/2, whereas when m =

1, the boundary is a cut, i.e., a straight fracture along the real axis extending from
−l to +l. The function φ(ω) has the properties that, φ(ω) = φ(ω) = φ̄(1/ω). The
last property is due to the fact that, on the unit circles, ω̄ = 1/ω. These properties
can be analytically continued outside the unit circle, where φ must be completely
regular except that, for large ζ , it should diverge as −icζ . From Eq. (30) we see
that for large ζ we must have ω  ζ and that φ ∼ −icω as ω → −∞, implying
(using the above properties of φ) that, as ω → 0, we must have φ(ω) ∼ −ic/ω,
and therefore φ(ω) = −icω + ic/ω. It is then straightforward to show (Fineberg
and Marder, 1999) that the displacement uz is finite as one approaches the fracture
tip, but the stress
∂uz
σyz = µ ∼ (z − 1)−1/2 , z → 1, (31)
∂y
diverges as one approaches the crack tip.
Although Eq. (31) was derived for a particular case, its main feature, namely, the
existence of a square root stress singularity at the fracture tip, is of general validity
and confirms Eq. (8), a feature that was already mentioned in Chapter 6. Thus, if a
fracture is given a finite radius of curvature, the singularity is effectively removed.
An amazing, and counterintuitive, application of this idea, that was pointed out by
Fineberg and Marder (1999), is to arresting the advance of a fracture in a damaged
material by drilling a small hole at the fracture tip, since the hole increases the
tip’s radius of curvature and hence blunts the singularity in the stress field. As a
result, the strength of the material increases sharply!
The conformal mapping method outlined above for Mode III cracks was ex-
tended to Mode I by Muskhelishvili (1953). The problem in this case is more
complex as one must solve the biharmonic equation rather than the Laplace’s
equation, and solve for two complex functions not one. Since Muskhelishvili’s
work hundreds of papers have been devoted to solutions of fracture problems
using these methods, a review of which will occupy a book by itself.

7.3.2 Dynamic Fractures in Mode I

According to Eq. (31), in an elastic material to which a uniform stress is applied at
its boundaries, the stress field at the tip of a static fracture is singular. Let us now
consider the case of a propagating fracture and examine the structure of the stress
field at its tip in Mode I. The dynamical equation for the displacement field u of
a steady state in a frame moving with a constant velocity v in the x-direction is
given by
∂ 2u
(λ + µ)∇(∇ · u) + µ∇ 2 u = ρv 2 . (32)
∂x 2
If we decompose u into longitudinal and transverse parts, u = ul + ut , with
$ %
∂vt ∂vt
ul = ∇vl , ut = − , (33)
∂y ∂x
 7.3. Linear Continuum Theory of Elasticity 303

it follows immediately that ul satisfies the following equation

 2  $ 2 %
2 ∂ 2 ∂
(λ + 2µ)∇ − ρv
2
ul = − µ∇ − ρv2
ut = f (x, y). (34)
∂x 2 ∂x 2
It can be shown that, f = 0. If
ρv 2 v2
α2 = 1 − = 1− 2, (35)
λ + 2µ cl
ρv 2 v 2
β2 = 1 − = 1− 2, (36)
µ ct
then, the general forms of vl and vt are (Fineberg and Marder, 1999)

vl = vl0 (z) + vl0 (z) + vl1 (x + iαy) + vl1 (x + iαy), (37)

vt = vt0 (z) + vt0 (z) + vt1 (x + iβy) + vt1 (x + iβy). (38)

However, it can be shown that, =
vl0 = 0. Therefore, if we define φ(z) =
vt0
∂vl1 /∂z and ψ(z) = ∂vt1 /∂z, the components of u = (ux , uy ) are given by,
ux = φ(zα ) + φ(zα ) + iβ[ψ(zβ ) − ψ(zβ )], (39)

uy = iα[φ(zα ) − φ(zα )] − [ψ(zβ ) + ψ(zβ )], (40)

where, zα = x + iαy, and zβ = x + iβy.
Equations (37) and (38) provide general expressions for steady-state elastic
problems in which a fracture propagates with a velocity v. If we define  =
∂φ(z)/∂z and  = ∂ψ(z)/∂z, then the stresses are given by
σxx = µ(1 + 2α 2 − β 2 )[(zα ) + (zα )] + 2iβµ[(zβ ) − (zβ )], (41)

σyy = −µ(1 + β 2 )[(zα ) + (zα )] − 2iβµ[(zβ ) − (zβ )], (42)

7 8
2σxy = 2µ 2iα[(zα ) − (zα )] − (β 2 + 1)[(zβ ) + (zβ )] . (43)

Equations (41)–(43) represent the general solutions in which the functions φ and
ψ must match the boundary conditions that are specified. Since one wishes to find
the potentials from given stresses at the boundaries,  must diverge as 1/v, and
the right-hand sides of Eqs. (41)–(43) turn into the derivative of  with respect
to α, implying that the static theory has a different structure than the dynamical
theory which is in fact more straightforward.
Let us now derive, as an application of Eqs. (37)–(43), the expressions for the
stresses around the tip of a fracture moving under Mode I loading. We assume that
the fracture lies along the negative x-axis (terminating at x = 0) and propagates
forward. The only assumption is that the problem is symmetric under reflection
about the x-axis. As discussed above (and also in Chapter 6), in the static case, the
stress fields have a square root singularity at the crack tip. We assume the same
to be true in the dynamic case (which can be verified in all cases for which the
304 7. Brittle Fracture: The Continuum Approach

expressions have been derived). Therefore, we assume that near the fracture tip
(Fineberg and Marder, 1999)
φ(z) ∼ (br + ibi )z−1/2 , (44)
−1/2
ψ(z) ∼ (dr + idi )z . (45)
Since we are considering Mode I fracture, then by symmetry the displacements
satisfy
ux (−y) = ux (y), uy (−y) = −u(y). (46)
If we substitute Eqs. (44) and (45) into (39) and (40) and use Eq. (46), we find that
bi = dr = 0, and therefore
(z) ∼ br z−1/2 , (47)
−1/2
(z) ∼ idi z . (48)
Observe that the square roots in Eqs. (44) and (45) must be interpreted as having
their cuts along the negative x-axis, where the fracture is located. Since on the
crack surface the stresses are relaxed, σxy and σyy vanish there. If we substitute
Eqs. (47) and (48) into Eqs. (41)–(43), we find that the condition for σyy is satisfied
identically for x < 0, y = 0, and that at y = 0
  √ √
σxy = iµ 2αbr − (β 2 + 1)di (1/ x − 1/ x), (49)
and therefore
di 2α
= 2 , (50)
br β +1
which, when used in Eqs. (41)–(43), (47) and (48), yields
 $ %  
KI 1 1 1 1
σxx = √ (β + 1)(1 + 2α − β ) √ + √
2 2 2
− 4αβ √ + ! ,
2π D zα z̄α zβ z̄β
(51)
   $ %
KI 1 1 1 1
σyy = √ 4αβ √ +! − (1 + β 2 )2
√ +√ , (52)
2 2π D zβ z̄β zα z̄α
 
2iαKI 1 1 1 1
σxy = √ (β + 1) √ − √ − √ + !
2
, (53)
2 2π D zα z̄α zβ z̄β
with
D = 4αβ − (1 + β 2 )2 . (54)
Note that the Rayleigh wave speed is in fact the root of D = 0, when Eqs. (35)
and (36) are used in (54). The most important physical feature of Eqs. (51)–(53)
is the overall scale of the stress singularity, which is characterized by the Mode I
stress intensity factor which, at y = 0, is given by
√
KI = lim 2π xσyy , (55)
x→0+
 7.3. Linear Continuum Theory of Elasticity 305

σ?
σ OO

σ?

Figure 7.2. Behind its tip, a fracture is pulled apart by two stresses (after Fineberg and
Marder, 1999).

which, as will be shown below, is directly related to energy flux into a fracture
tip. Moreover, Eqs. (51)–(53) contain information about the angular structure of
the stress fields which can be used in both theoretical and experimental analyses.
Theoretically, one can use these equations for predicting the direction of fracture
motion, and the conditions under which a fracture branches out. Experimentally,
one can utilize these equations for assessing the accuracy of the predictions of
continuum fracture mechanics, and for obtaining measurements of the stress fields
surrounding rapidly-propagating fractures; we will discuss these matters later in
this chapter. It is important to recognize, as pointed out by Freund (1990), that
although Eqs. (51)–(53) were derived for fractures moving at a constant speed, the
same equations are also true for those that, during propagation, accelerate and/or
decelerate, so long as the derivative dv/dt is small during the time needed for
sound to travel across the region of the universal elastic singularity.
We now suppose that a fracture is loaded by two stresses, located a distance l0
behind its tip, moving with it in steady state at velocity v, and of strength −σc (see
Figure 7.2) such that
lim σyy (x, y) = −σc δ(x + l0 ), x < 0. (56)
y→0+

If the fracture tip is at the origin, the stress and displacement fields are continuous
and differentiable everywhere, except along a branch cut starting at the origin and
running backwards along the negative x-axis. If we define ± (x) and ± (x) by
± (x) ≡ lim (x ± iy), ± (x) = ± (x ± iy), (57)
y→0+

then because of the branch cut, for x < 0, + (x) = −− (x). As shown above,
for Mode I loading, σxy = 0 for y → 0+ and ∀x. Therefore, from Eq. (43) we
obtain
¯ − ) = (β 2 + 1)(+ + 
2iα(+ −  ¯ − ), (58)
306 7. Brittle Fracture: The Continuum Approach

¯ − iε). The function

using the fact that (x + iε) = (x
f+ (x) = 2iα+ (x) − (1 + β 2 )+ (x) (59)
is defined for all x, and can be analytically continued above the x-axis, where
it is related to stresses and must be free of singularities, whereas f− , defined
in a manner similar to f+ , must contain no singularities below the real axis. If
two complex functions are equal, one without singularities for x > 0 and the
other without singularities for x < 0, the two functions must individually equal a
constant which, in fact, is zero since all the stresses are zero far from the fracture.
Therefore, f+ = f− = 0, and one has
¯ − (x) = (1 + β 2 )
2i+ (x) = (1 + β 2 )+ (x), 2i  ¯ − (x). (60)
The boundary condition for σyy for x < 0 is [see Eq. (42)]
σyy = −µ(1 + β 2 )(+ + ¯− ) − 2iβµ(+ − ¯− ) = −σc δ(x + l0 ). (61)
Using Eqs. (51)–(54), Eq. (61) becomes
σyy = −σc δ(x + l0 ) = µD(+ − ¯− )/(2iα). (62)
Since the delta function can be represented as
i 1
δ(x + l0 ) = (63)
π x + l0 + iε
one can argue that the only complex function that decays properly at infinity, has
a singularity no worse than a square root at the origin, and satisfies Eq. (61), is
&
iα σc l0
+ (x) = . (64)
π µD x + l0 + iε x
The function (z) can now be obtained by analytical continuation of + (x). In
particular, for x > 0 the stress σyy is easily found from Eq. (62) to be
&
1 l0 σ c
σyy = , (65)
π x x + l0
which means that the stress intensity factor associated with σyy is given by
!
KI = σc 2/π l0 . (66)

7.4 The Onset of Fracture Propagation: Griffith’s

Criterion
What are the conditions under which a fracture propagates? Calculations such as
those outlined above yield the value of the stress fields at the tip of a propagating
fracture, but have nothing to say about the conditions under which a fracture
actually propagates. As already discussed in Chapter 6, Griffith (1920) proposed
that fracture occurs when the energy per unit area released by a small extension of a
 7.4. The Onset of Fracture Propagation: Griffith’s Criterion 307

crack is equal to , the energy required for creating new fracture surface. Griffith’s
idea, which is the final assumption of continuum fracture mechanics, states that
the dynamics of a fracture tip depends only on the total energy flux H per unit area
into the cohesive zone, and that all the details about the spatial structure of the
stress fields are irrelevant. The energy H creates new fracture surfaces, and is also
dissipated near the fracture tip. In general, the fracture velocity v is a function of
H. It is common to use (v) for representing the energy consumed by a fracture
in the cohesive zone, in which case the equation of motion for a fracture is
H = (v). (67)
The central question of interest to continuum fracture mechanics is the conditions
under which a static fracture begins to move. For this to happen, a critical fracture
energy Hc , the minimum energy per unit area needed for a fracture to propagate
forward, irrespective of its velocity, is needed. The standard assumption is that the
velocity consuming the minimum energy is very small, although this assumption is
not necessarily correct. Equivalently [see Eq. (10) and Chapter 6], one may define
a critical stress intensity factor KI c at which the fracture first begins to propagate.
We now derive this equivalence, following Fineberg and Marder (1999).
In what follows, we adopt the summation convention for repeated indices.
Energy flux is found from the time derivative of the total energy:
  $ %
d d 1 1 ∂uα
(Hk + Hp ) = ρ u̇α u̇α + σαβ dxdy, (68)
dt dt 2 2 ∂xβ
where Hk and Hp are, respectively, the total kinetic and potential energies within
the entire system, and u̇α = duα /dt. Since the spatial integral in Eq. (68) is taken
over a region which is static in the laboratory frame (i.e., dx/dt = dy/dt = 0),
we have
  $ %
d ∂ u̇α
(Hk + Hp ) = ρ üα u̇α + σαβ dxdy, (69)
dt ∂xβ
where the symmetry of the stress tensor under interchange of indices has been used
for the last term. Use of the equation of motion, ρ üα = ∂σαβ /∂xβ , in Eq. (69),
yields,
  $ %  
∂ ∂ u˙α ∂ 
σαβ u̇α + σαβ dxdy = σαβ u̇α dxdy
∂xβ ∂xβ ∂xβ

= u˙α σαβ nβ dS, (70)
∂S
where ∂S is the surface boundary of the system, and n is an outward unit nor-
mal with components nβ . Equation (70) is a statement of the fact that energy is
transported by a flux vector j with components that are given by
jα = σαβ u̇β . (71)
As mentioned in Chapter 5 [see Section 5.1.1 and Eq. (5.13)], the total energy
flux J per unit time into the fracture tip is called the J -integral (see Cotterell
308 7. Brittle Fracture: The Continuum Approach

Figure 7.3. Dotted lines show the most convenient contour for integrating the energy flux
and calculating the energy that flows to a fracture tip. The contour runs below the fracture,
closes at infinity, and comes back just above the contour (after Fineberg and Marder, 1999).

and Atkins, 1996, for a discussion of the use of the J -integral to ductile fracture).
A convenient contour for the integration is shown in Figure 7.3. If, for a crack
loaded in pure Mode I, we use the asymptotic forms, Eq. (52) for σyy and the
corresponding expression for uy , we find that J is given by
α v(1 − β 2 )
J = K 2, (72)
2µ 4αβ − (1 + β 2 )2 I
where KI is the stress intensity factor defined by Eq. (55), with the subscript I
emphasizing that the result is specific to Mode I fracture. Thus, the energy release
rate H in the case of pure Mode I is
J α 1 − β2 1 − νp2
H= = K 2
≡ AI (v)KI2 . (73)
v 2µ 4αβ − (1 + β 2 )2 I Y
The corresponding result for pure Mode II fracture is
β 1 − β2 1 − νp2
H= K 2
≡ AI I (v)KI2I , (74)
2µ 4αβ − (1 + β 2 )2 I I Y
while for Mode III fracture one has
v
H= K2 . (75)
2αµ I I I
In the limit v → 0, each of the functions Aα (v) → 1 (α =I, II and III) and, for
example, Eq. (73) simplifies to
1 − νp2
H= KI2 . (76)
Y
In the general case of mixed mode fracture, Eq. (10) should be used.
The functions Aα (v) are universal in the sense that they are independent of
most details of the material’s loading or geometric configuration. Assuming that
there is no energy sink in the system other than the one at the tip of the fracture,
 7.5. The Equation of Motion for a Fracture in an Infinite Plate 309

Eqs. (73)–(75) relate the total flux of energy from the entire elastic material to the tip
which, when it is set to equal to the energy dissipated in the cohesive zone, yields
an equation of motion for the fracture. Note that, in order to derive Eqs. (73)–
(75), we have tacitly assumed that, given near-field descriptions of stress and
displacement fields [Eqs. (50)–(53)], Eqs. (37) and (38) are valid. If, for example,
the cohesive zone is of the order of 1 mm in a piece of a solid with dimensions
that are a few centimeters, the value of the stress field on the contour ∂S used in
Eq. (70) will not be approximated well by the asymptotic forms of the stress and
displacement fields, invalidating Eqs. (73)–(75). Since an energy balance provides
no information about a fracture’s path, we have assumed that the fracture travels
along a straight line (see below). Although the rules for determining paths of slowly
propagating fractures are known, they are not known for rapidly moving fractures.

7.5 The Equation of Motion for a Fracture in an Infinite

Plate
As discussed above, one can derive an equation of motion for a fracture by calcu-
lating either the energy release rate H or, equivalently, the dynamic stress intensity
factor K which depends on the fracture’s loading history, and its length and veloc-
ity. In what follows we derive an exact expression for K for a straight semi-infinite
fracture in an infinite plate with loads applied to the fracture’s faces. The derivation
follows closely those given by Willis (1990) and Fineberg and Marder (1999). The
calculation is, in the context of linear elasticity—a boundary-value problem—and
in the most general case is applicable to a system in which,
(1) the fracture is a semi-infinite straight-line branch cut in an infinite isotropic
2D elastic plate.
(2) The velocity 4v(t) of the is not in general constant, with the position of its tip
being l(t) = 0 v(t  )dt  , which, in the context of a boundary-value problem,
t

is assumed to be known. However, v(t) must be less than the relevant sound
speeds at all times.
(3) The external stresses σe are permitted only along the fracture, but are allowed
arbitrary time and space dependence. This can be realized by placing wedges
between the faces of the cracks in order to load them.
We derive the corresponding expressions for fracture Modes I, II and III. In the
calculations that follow u, σ and c denote the displacement, stress, and a sound
speed in each case, as listed in Table 7.2. By symmetry, u(x, t) = 0 for all x > l(t).
Due to the one-to-one relation between K and the energy flux H, we compute the
latter as a function of l(t) and v(t), and as a functional of the external load σe (x, t).
We look for a Green function G operating on the displacement field u defined by
the following convolution integral,
 
G∗u≡ G(x − x  , t − t  )u(x  , t  ) dxdt. (77)
310 7. Brittle Fracture: The Continuum Approach

Table 7.2. Notation convention for the solution of equation of

motion for a fracture in an infinite plate. The Rayleigh wave speed
cR is the roots of D = 0 [see Eq. (54)], and is typically about 90%
of the transverse wave speed ct .
Mode I Mode II Mode III
u denotes uy (x, y = 0+ , t) ux (x, y = 0+, t) uz (x, y = 0+ , t)
σ denotes σyy (x, y = 0+ , t) σyx (x, y = 0+ , t) σyz (x, y = 0+ , t)
c denotes cR cR ct

If
 
 
G(k, ω) = eikx −iωt G(x  , t  ) dxdt, (78)

denotes the Fourier–Laplace transform of G, we require that

G− (k, ω)
G(k, ω) ≡ , (79)
G+ (k, ω)
with the properties that G+ and G− vanish for x < cR t and x > −cR t, respec-
tively, where cR is the Rayleigh wave speed. Physically, this implies that G+ is
non-zero only for x large enough that a pulse beginning at the origin at t = 0 could
never reach it in the forward direction (with a similar condition for G− ). In fact,
for the cases to be discussed below, we have
G+ ∝ δ(x − cR t), G− ∝ δ(x + cR t). (80)
While it is not yet clear that G can be decomposed according to Eq. (79), or that
it even exists, for the moment we simply assume these to be true.
We decompose σ into two functions, σ = σ + + σ,− and define u = u,− where
σ vanishes for x < l(t), σ − vanishes for x > l(t), while u− does so for x > l(t).
+

Therefore, σ − describes the stresses along the fracture faces, while σ + is an, as
yet unknown, function. u,− on the other hand, is an unknown function along
the fracture faces and vanishes ahead of its tip. Using Eqs. (77)–(79), we write,
G ∗ u = σ , which after Laplace–Fourier transforming yields,
G(k, ω) u(k, ω) = σ (k, ω). (81)
Using Eq. (79) we obtain
G− (k, ω) u(k, ω) = G+ (k, ω) σ (k, ω) (82)
which, after inverting back to real space, yields
G+ ∗ σ = G− ∗ u. (83)
One can show that for x < l(t), G+ ∗ σ + = 0. Suppose that x > l(t). Since σ +
is zero behind the fracture, the integral
 
G+ ∗ u + = G+ (x − x  , t − t  )σ + (x  , t  ) dxdt (84)

is zero for x  < l(t  ). The only case for the integrand to be non-zero is for x  > l(t  ),
 7.5. The Equation of Motion for a Fracture in an Infinite Plate 311

˙ ∗ )(t  − t), where t < t ∗ < t  . However,

in which case, x  − x > l(t  ) − l(t) = l(t
this means, by the mean-value theorem, that
x − x  < cR (t − t  ), (85)
since cR is the largest value that v(t) can take on. On the other hand, (85) is
precisely the condition under which G+ (x − x  , t − t  ) vanishes. Therefore,
 
G+ (x − x  , t − t  )σ + (x  , t  ) dxdt = 0, x < l(t), (86)

and similarly
 
G− (x − x  , t − t  )u− (x  , t  ) dxdt = 0, x > l(t). (87)

From Eq. (83) one can show that

G+ ∗ σ + = −(G+ ∗ σ − ) H (x, t), (88)
where H (x, t) = #[x − l(t)], and # is the Heaviside step function. Equation (88),
which has now been shown to be true both for x > l(t) and x < l(t), yields, after
inverting it back to real space,
σ + = −(G+ )−1 ∗ [(G+ ∗ σ − )H ]. (89)
Since σ− is the (known) stress exerted at the back of the fracture tip, Eq. (89)
provides a formal solution to the problem. The stress intensity factor is given by
√
K = lim 2π ε σ (ε + l, t), (90)
ε→0+
√
which requires identifying the terms that lead to a divergence of the form 1/ ε
as x = l(t) + ε approaches l(t) from above. √
We now show that (G+ )−1 has a singularity for ε → 0, behaving as 1/ ε 3 ,
while G+ ∗ σ + is finite. To find the singularity of Eq. (89), G+ ∗ σ − is evaluated
at x = l(t) and pulled outside the convolution as a multiplicative factor. The stress
intensity factor can therefore be written as
K = K̃[l(t), σ ] · K(v), (91)
with
√
K̃(l, σ ) ≡ −( 2G+ ∗ σ − )(l,t) , (92)
and
√
K(v) ≡ lim [ π ε(G+ )−1 ∗ H ](l+ε,t) (93)
ε→0−

Physically, K̃(l, σ ), which is independent of the fracture velocity, is the stress

intensity factor that would emerge at the tip of a static fracture sitting at all times
at l [the tip is exposed to the load σ − (t)]. On the other hand, although K depends
on the instantaneous velocity v(t) of the fracture, it is independent of the crack’s
history, i.e., how it arrived at a particular position at time t.
312 7. Brittle Fracture: The Continuum Approach

7.5.1 Mode III

We now apply the general results, Eqs. (91)–(93), to the particular case of anti-plane
shear, which will also allow us to verify the general structure of the Green function
G used so far. By calculating the stress intensity factor using Eqs. (73)–(75), and
hence the energy release rate H, we derive the equation of motion for a Mode III
fracture by equating the energy release rate to the fracture energy. The starting
point is the wave equation for uz , Eq. (14), which after Fourier transforming in
both space and time yields
∂ 2 uz
= (k 2 − ω2 /c2 − 2ibω)uz , (94)
∂y 2
where a small damping b has been added to help us overcome some convergence
problems that will arise later. In an infinite plane, the only allowed solution is one
that decays as a function of y, and therefore Eq. (94) is solved by
 " 
uz (k, y, ω) = exp −y k 2 − ω2 /c2 − 2ibω u(k, ω). (95)

By taking u = uz (y = 0) and σ = σyz (y = 0), one has

"
σ
G(k, ω) = = −µ k 2 − ω2 /c2 − 2ibω. (96)
u
Using Eq. (79) we can write
!
G− = −µ ik − iω/c + b, (97)
and
!
G+ = 1/ −ik − iω/c + b. (98)
The decomposition, G = G− /G+ , satisfies the conditions of the preceding section
if we write
 
+ 1 e−ikx−iωt
G (x, t) = √ dk dω
(2π )2 −ik − iω/c + b
  −ipx−iω(t−x/c) (99)
1 e
= √ dp dω,
(2π )2 −ip + b
with p = k + ω/c, and therefore

δ(t − x/c) e−ipx
G+ (x, t) = √ dp. (100)
2π −ip + b
When x < 0, one must close the contour in the upper half plane, and since the
branch cut is in the lower half plane, the integral vanishes. When x > 0, we deform
the contour to surround the branch cut to obtain
 ∞
1 2e−px 1
√ dp = √ . (101)
2π 0 p+b πx
 7.5. The Equation of Motion for a Fracture in an Infinite Plate 313

Therefore
1
G+ (x, t) = √ δ(t − x/c)#(x). (102)
πx
By a largely similar analysis we find that
 
d #(x)
(G+ )−1 (x, t) = δ(t − x/c) √ . (103)
dx πx
Having calculated (G+ )−1 (x, t), we can now find the stress intensity factor
KI I I (l, t). From Eq. (93) we find that
  
√ d #(x1 )
K(v) = π ε δ(t1 − x/c) √ #[l(t) + ε − x1 − l(t − t1 )] dx1 dt1
dx1 π x1
  
√ d #(x1 )
= ε √ #[ε/(1 − v/c) − x1 ] dx1 . (104)
dx1 π x1
Since only very small x1 are important, we find that
 !
√ #(x1 )
K(v) = ε √ δ[ε/(1 − v/c) − x1 ] dx1 = 1 − v/c. (105)
π x1
Similarly,
√   #(x1 )
K̃(l, t) = − 2 δ(t1 − x1 /c) √ σ [−l(t) − x1 , t − t1 ] dx1 dt1
π x1
√  #(x1 )
=− 2 √ σ [l(t) − x1 , t − x1 /c]dx1 . (106)
π x1
In particular, when σ − does not depend on the time and, σ (x) = σ0 #(x), one
obtains
√ √
K̃ = −(4/ 2π )σ0 l. (107)
The minus sign arises because the stresses ahead of the fracture tip always act
against those applied on the fracture faces. Note that Eq. (104) reduces to unity
when v → 0, implying that in the case of time-independent loading, K̃ is indeed
the stress intensity factor one would have had if the fracture had been static at l
for all times. For the propagating fracture, we obtain
!
KI I I = 1 − v/c K̃[l(t), σ0 ]. (108)
We now compute the stress singularity that would have developed had we had
a static fracture of length l(t) at time t, and multiply the result by a function of the
instantaneous velocity. We should emphasize that all details of the history of the
crack motion are irrelevant; only the velocity and loading configuration are needed
for determining the stress fields sufficiently close to the tip. As a consequence, one
can use Eq. (75) to determine the energy flow to the tip of the crack:
K̃ 2
H = v(1 − v/c) . (109)
2αµ
314 7. Brittle Fracture: The Continuum Approach

The rate at which energy enters the tip of the fracture must be equal to v(v). There
is nothing to prevent the fracture energy  from being a function of the velocity, but
the notion of local equilibrium, which has prevailed until now, strongly suggests
that  should be a function of v alone. Therefore
K̃ 2 (l)
(v) = (1 − v/c) , (110)
2αµ
which, after rearranging and using Eq. (106), yields
π µ !
= (1 − v/c)(1 + v/c). (111)
4lσ02
If we define
π µ
l0 = , (112)
4σ02
Eq. (111) is rewritten as
l0 !
= (1 − v/c)(1 + v/c). (113)
l

7.5.2 Mode I
The preceding analysis can also be carried out for thin plates under tension. Al-
though all steps of the analysis proceed as before, it is not possible to derive
simple analytical expressions. This case has been discussed in detail by Freund
(1990) who finds that the energy flux per unit length extension of the fracture, to
an accurate approximation, is given by
(1 − v/cR )K̃ 2 (l)
H(v) = (v) = , (114)
2λ̃
where λ̃ is a Lamé constant defined by Eq. (20). Rearranging Eq. (114) yields
Y (v) v
=1− , (115)
K̃ 2 (l)(1 − vνp2 ) cR

where K̃ is still given by Eq. (106), using σyy on the x-axis for σ . In the case of
time-independent loading described by σ (x) = σ0 #(x), one obtains
l0 v
=1− , (116)
l cR
with
π  λ̃
l0 = . (117)
4σ02
Equation (116) is now written in the following form
v = cR (1 − l0 / l), (118)
 7.5. The Equation of Motion for a Fracture in an Infinite Plate 315

which is nothing but Eq. (7) obtained by the scaling analysis, with the difference
that  and hence l0 can depend strongly upon the crack velocity v. Hence, seem-
ingly large differences between the predictions of the theory and experimental data
are due to nothing more than assuming that l0 is a constant!
What are the practical implications of Eq. (10) for the design of experiments? As
discussed by Fineberg and Marder (1999), one may consider three experimental
situations. (1) One for which the assumptions of the theory hold well. (2) A second
experiment in which the theoretical assumptions are satisfied in an approximate
way, while (3) in the third experiment the assumptions clearly fail. The three cases
are as follows.
(1) A thin plate has a fracture running half-way through, and driven by wedging
action in the middle. For times less than that needed for sound to travel from
the point of loading to the material’s boundaries and back to the tip of the
fracture, all the assumptions of the theory are satisfied.
(2) A thin plate has a long fracture as before, but with uniform static stresses σ∞
applied at the outer boundaries, and the faces of the fracture being stress-free.
This problem is equivalent to one in which the upper and lower outer bound-
aries are stress-free, but uniform stresses −σ∞ are applied along the fracture
faces. The equivalence is due to the fact that an uncracked plate under uni-
form tension σ∞ is a solution of the equations of elasticity, so this trivial static
solution can be subtracted from the first problem to obtain the second equiv-
alent one. However, in the new problem, stresses are applied to the fracture
faces all the way back to the left-hand boundary of the material. Therefore, the
problem must be mapped onto one in which stresses are applied to the faces
of a semi-infinite fracture in an infinite plate, but the correspondence is only
approximate.
(3) Consider now a semi-infinite fracture in an infinitely long strip, shown in Fig-
ure 7.4, which is loaded by displacing each of its boundaries at y = ±w/2 by
a constant amount δ. Far behind the tip (as x → −∞), the fracture relieves
all the stresses within the strip. Far ahead of the tip (as x → +∞), the mate-
rial is unaffected by the fracture with the stress field being linear in y. Thus,
the energy per unit extension far ahead of the fracture has a constant value,
2Y δ 2 /[w(1 − νp2 )], where Y and νp are, respectively, the usual Young’s mod-
ulus and Poisson’s ratio of the material. The translational invariance of the
system along the x-direction implies that, for a given δ, the fracture should

Figure 7.4. A fracture in a semi-infinite strip.

316 7. Brittle Fracture: The Continuum Approach

eventually propagate at a constant velocity v. Writing an energy balance yields

2yδ 2
H== . (119)
w(1 − νp2 )
If we now assume that we can still use Eq. (114), then, since the stress intensity
factor K̃ of a static fracture in a strip loaded with constant displacements δ
cannot depend upon where the fracture is located, K̃ must be a constant and
Eq. (114) would predict that
(1 − νp2 )K̃ 2
H== (1 − v/cR ). (120)
Y
However, the velocity term of Eq. (120) contradicts Eq. (119), implying that
Eq. (114) has failed. The reason for this failure lies in the assumption that
the fracture tip does not feel the presence of the system’s boundaries, which
is clearly invalid. In fact, the translational invariance of the system depends
crucially on the presence of its vertical boundaries. Energy flows continuously
into the material as the amount of kinetic energy reaches a steady state. In
contrast, the kinetic energy within a system of infinite extent increases without
bounds as ever farther reaches of the material feel propagation of the fracture,
since elastic waves that carry this information propagate outward.

7.6 The Path of a Fracture

We now discuss briefly the path travelled by a propagating fracture. As discussed
above, energy balance provides an equation of motion for the tip of a fracture only
when its path or direction of its propagation is assumed. Although criteria for the
path of a slowly propagating fracture have been established, no such criteria have
been proven to exist for a rapidly moving fracture. We will discuss this issue later
in this chapter.

7.6.1 Planar Quasi-static Fractures: Principle of Local Symmetry

A fracture is considered to propagate slowly if the velocity v of its tip is much less
than the Rayleigh wave speed cR . Goldstein and Salganik (1974) proposed that
the path taken by slow cracks satisfies the principle of local symmetry, according
to which a crack propagates so as to set the component of Mode II loading to
zero. An immediate consequence of this proposal is that if a stationary fracture is
loaded in such a way as to experience Mode II loading, it forms, upon extension,
a sharp kink and moves at a new angle. This rule means that the fracture moves
perpendicular to the direction in which tensile stresses are maximum. Cotterell and
Rice (1980) showed that a fracture satisfying the principle of local symmetry also
chooses a direction so as to maximize the rate of energy release. The distance over
which a fracture must move so as to set KI I to zero is of the order of the size
of the cohesive zone (Hodgdon and Sethna, 1993). Cotterell and Rice (1980) also
 7.6. The Path of a Fracture 317

showed that the condition KI I = 0 has the following consequences for fracture
propagation. Consider an initially straight fracture, propagating along the x-axis.
The components σxx and σyy of the stress field are given by
KI
σxx = + σ + O(r 1/2 ), (121)
2π r 1/2
KI
σyy = + O(r 1/2 ). (122)
2π r 1/2
The constant stress σ is parallel to the fracture at its tip. If σ > 0, any small
deviations from straightness cause the fracture to diverge from the x-direction,
whereas if σ < 0 the fracture is stable and continues to propagate along the x-axis.
Yuse and Sano (1993) and Ronsin et al. (1995) conducted experiments described in
Section 6.12 by slowly pulling a glass plate from a hot region to a cold one across
a constant thermal gradient. The velocity of the fracture, driven by the stresses
induced by the non-uniform thermal expansion of the material, follows that of
the glass plate. At a critical pulling velocity, the fracture’s path deviated from
straight-line propagation and developed transverse oscillations. This instability is
completely consistent with the principle of local symmetry: The crack deviates
from a straight path if the stress σ in Eq. (121) is positive. The wavelength of
the ensuing oscillations has also been computed numerically (Adda-Bedia and
Pomeau, 1995).

7.6.2 Three-Dimensional Quasi-static Fractures

Hodgdon and Sethna (1993) generalized the principle of local symmetry to 3D
and showed that an equation of motion for a crack line involves, in principle, nine
different constants, although we are not aware of any experimental determination
of these constants. Larralde and Ball (1995) and Ball and Larralde (1995) carried
out stability analysis of cracks which are almost planar and have a tip which is
almost a straight line, so that the differences from a planar and straight edged crack
could be considered as perturbation parameters. They showed that, in agreement
with the proposal of Goldstein and Salganik (1974) (i.e., the principle of local
symmetry), at least under quasi-static conditions the mechanism underlying the
stability of planar fractures propagating under Mode I loading is the appearance
of a Mode II loading in the vicinity of the fracture edge associated with each out-
of-the plane perturbation mode. This Mode II loading, which tends to suppress the
perturbation in the quasi-static propagation of the fracture, is a consequence of the
global structure of the crack edge, and has nothing to do with the prior history of
the fracture, nor with the local geometry around any point of the edge. As such,
this stabilizing mechanism is an intrinsically three-dimensional effect.
Thus, the principle of local symmetry is consistent with all experimental tests
that have been performed so far on slowly propagating fractures. Nevertheless, it
is not based on a rigorous theoretical foundation, since there is no basic principle
that predicts that a fracture must extend perpendicular to the maximum tensile
stress, or that it must maximize energy release.
318 7. Brittle Fracture: The Continuum Approach

7.6.3 Dynamic Fractures: Yoffe’s Criterion

In the case of rapid fractures, there is no rigorous basis for deciding the direction in
which a fracture may propagate. A variety of criteria for path selection have been
proposed in the literature which can be divided into two types: Those proposing
that a crack propagates in the direction of a maximal stress, and those that are
based on a maximum dissipation of energy. In contrast to quasi-static fractures,
however, these criteria are not equivalent and, more importantly, none of them is
strongly supported by experiment.
An important work is that of Yoffe (1951), already mentioned in Chapter 6, who
proposed that one should check the stability of a rapidly propagating fracture by
examining the dynamic stress fields given by Eqs. (51)–(53), approaching the tip
of the fracture along a line at an angle θ relative to the x-axis, and computing
the stress perpendicular to that line. If we choose, zα = r cos θ + irα sin θ and
zβ = r cos θ + irβ sin β, and evaluate the stress in the polar coordinates, σθ θ =
σxx sin2 θ + σyy cos2 θ − σxy sin 2θ, we find that below a velocity vc  0.61cR
(which depends on the Poisson’s ratio), the maximum tensile stress occurs for
θ = 0. Above vc , the tensile stress σθ θ develops a maximum in a direction θ > 0,
and the angle of maximum tensile stress increases smoothly until it finally develops
a maximum at about ±60◦ relative to the x-axis, implying that above the critical
velocity a fracture might propagate off-axis.As pointed out by Fineberg and Marder
(1999), this spontaneous breaking of the axial symmetry of the phenomenon is due
to purely kinetic effects. Recall that in an elastic medium information is propagated
at the speed of sound, and that the stress field at the tip of a rapidly moving fracture
is analogous to the electric field surrounding a point charge moving at relativistic
velocities. The stress field then experiences a Lorentz contraction in the direction
of propagation as the fracture’s velocity approaches the speed of sound, resulting
in the formation of symmetric lobes around the x-axis of maximal tensile stress
(above the critical velocity).
Fracture branching stemming from the approach of the velocity of a crack tip to
Yoffe’s critical velocity was first thought to provide a rigorous criterion for crack
instability. However, many experiments have shown that large-scale branching oc-
curs in a variety of materials at velocities much less than 0.61cR , and that branching
angles of about 10◦ − 15◦ (see Section 7.8.10 below), instead of Yoffe’s predicted
value of 60◦ , are generally observed. To overcome the failure of the Yoffe’s cri-
terion for fracture branching, a number of other criteria have been proposed (see,
for example, Ramulu and Kobayashi, 1985, 1986) in which the form of the stress
field at the boundary of the cohesive zone near the tip is used for deriving criteria
for fracture branching and its angle (see, for example, Theocaris and Georgiadis,
1985; Ramulu and Kobayashi, 1985, 1986). To obtain the angle of branching one
determines the direction in which the local energy density, evaluated at the edge
of the cohesive zone, is maximum. The theoretical justification for this criterion
was originally suggested by Sig (1973) who proposed that fracture propagation
occurs in the radial direction along which the local energy density possesses a
stationary value. Experimentally-measured crack branching angles are consistent
 7.7. Comparison with the Experimental Data 319

with those predicted by variants of this criterion, although the same criteria predict
critical velocities for crack branching that are nearly identical to Yoffe’s prediction,
namely, 0.61cR . Adda-Bedia and Ben Amar (1996), for example, proposed that
one should draw contours of constant principal stress and search for points where
these contours are perpendicular to lines drawn from the crack tip, along which
the fracture travels. This criterion predicts the existence of two critical speeds.
The first is the velocity at which the fracture must choose between three possible
directions, whereas the second critical velocity is one at which the fracture must
choose between five possible directions. Although this criterion is plausible, there
is no experimental evidence indicating that this is in fact the preferred criterion.
The branching angles that are predicted by such criteria are not significantly dif-
ferent from those determined by the following condition, which is a type of static
condition. Consider the stress field formed ahead of a single propagating fracture,
from which one can compute the trajectories that satisfy the quasi-static condi-
tion, KI I = 0 (Kalthoff, 1972; Parleton, 1979). The angle that is determined by
this trajectory at a distance rc from the fracture tip, where rc is the typical size
of the cohesive zone, is in good quantitative agreement with the experimental
observations.

7.7 Comparison with the Experimental Data

It is instructive at this point to compare the predictions of linear continuum frac-
ture mechanics with the experimental data. A close inspection indicates that, as
long as the basic assumptions of linear fracture mechanics hold, the theory is quite
successful in predicting both fracture propagation and the behavior of the stress
field throughout the material. However, if one or more of these assumptions break
down, the linear theory loses its predictive power. For example, continuum frac-
ture mechanics has been successful in predicting the value of the stress intensity
factor at the tip of both static and dynamic fractures for both static and dynamically-
applied loads. Kim (1985) measured transient behavior of the stress intensity factor
and made a quantitative comparison with the predictions of Eqs. (91)–(93). In his
experiment, a step function loading was applied to the crack faces in a sheet of
Homalite-100 that was large enough to be approximated as an infinite system.
Homalite-100 is a thermoset polyester resin that, at room temperature, can be ac-
curately represented as a linearly elastic material with brittle fracture behavior. Of
particular importance is its property of birefringence that permits the use of optical
techniques, such as photoelasticity described in Section 6.11.2.2, for mapping the
stress field. Due to such desirable properties, Homalite-100 has been used in many
studies of dynamic fracture. The stress intensity factor was measured optically (see
Chapter 6), using a method developed by Kim himself in which the relation of the
transmitted light through the fracture tip with the stress intensity factor was used.
Kim’s data agreed well with the calculated time dependence of the stress inten-
sity factor. Similar agreement between the theory and experiments for PMMA was
reported by Vu and Kinra (1981) who measured the transient relaxation of the stress
320 7. Brittle Fracture: The Continuum Approach

Figure 7.5. Comparison of the predictions of linear continuum fracture mechanics (the
curve) with the experimental data (symbols) of Kobayashi et al. (1974).

field within the material. In their experiments strain gauges, with a temporal reso-
lution of about 1 µs, were placed throughout the sample to measure the temporal
behavior of the stress field surrounding a fracture at times immediately following
its arrest. Their data agreed with a prediction of Freund (1990), that the stress field
at a point directly ahead (behind) the fracture should reach its equilibrium value
(to within a few percent) as soon as the shear (Rayleigh) wave front passes.
However, the same type of favorable comparison between the theory and ex-
periment does not exist at high fracture velocities, and in fact experiments often
seem to disagree with Eq. (14). As an example we show in Figure 7.5 the data
of Kobayashi et al. (1974) with PMMA and compare them with the theoretical
predictions. Although the theory predicts that if the fracture energy is not a strong
function of the velocity, the fracture would smoothly accelerate from rest to the
Rayleigh wave speed cR , Kobayashi et al.’s data do not confirm this prediction:
After the fracture initially accelerates rapidly, it becomes increasingly sluggish and
eventually reaches a final velocity well below cR . However, if we suppose that
the fracture energy  is a function of the velocity, and specify in Eq. (118) [using
Eqs. (114)–(116)] that l0 is defined in terms of the minimum energy (0) at which
fracture propagation first happens, one obtains instead of Eq. (118):
 
(v) l0
v = cR 1− . (123)
(0) l

Therefore, if the fracture energy (v) increases rapidly with the velocity v, one can
obtain practically any functional dependence of the velocity on the fracture length.
One can also interpret Eq. (123) as a way of extracting the velocity dependence of
fracture energy from measurements of v. However, validation of the theory cannot
be accomplished without an independent measurement of the fracture energy ,
although even such validation would provide no fundamental explanation of the
origin of any measured velocity dependence of the fracture energy.
 7.7. Comparison with the Experimental Data 321

Figure 7.6. Experimental data (triangles) of Sharon and Fineberg (1999) and their
comparison with the theoretical predictions (rectangles).

Bergkvist’s (1974) beautiful experiments on crack arrest in PMMA provided

the first comparison of the theory and experiment where the velocity dependence
of fracture energy was explicitly taken into account. His experiments allowed
direct comparison of the calculated energy release rates with experimental data
for fracture velocities below 0.2cR (which are less than 200 m/s). He obtained a
continuous distribution of the fracture tip locations with a temporal resolution of
about 1 µs, and used independent measurements of the fracture energy of PMMAas
a function of the fracture’s velocity. Values of the fracture velocity were computed
by equating the measured value of the fracture energy to the calculated value of
the energy release rate. The predicted and measured velocities were in agreement
to within 10%.
A similar comparison between the theoretical predictions and experimental data
for PMMA was reported by Sharon and Fineberg (1999). They first carried out an
independent measurement of the fracture energy of a crack by the use of a strip
geometry. An additional series of experiments, which was carried out in 40 × 40
cm samples, yielded the velocity values which were then inserted into Eq. (123)
to yield values of (v) which were then compared to the direct measurements.
The results are shown in Figure 7.6. Their data agree with Eq. (123) for velocities
less than about 400 m/s  0.4cR . However, above 0.4cR there is a large difference
between the data and the predicted values of (v), which is due to the growth of
the cohesive zone around the crack tip to a length scale where the assumptions of
linear continuum fracture mechanics are no longer valid (see also below).

7.7.1 The Limiting Velocity of a Fracture

As the derivation of Eq. (118) indicated, an important prediction of linear contin-
uum fracture mechanics is that, disallowing divergent behavior of (v), a fracture
322 7. Brittle Fracture: The Continuum Approach

should accelerate until it arrives asymptotically at the Rayleigh wave speed cR .

However, in amorphous materials, such as PMMA and glass, the maximum mea-
sured velocity of a fracture hardly exceeds a value of about 0.5cR , whereas in
strongly anisotropic materials, such as LiF (Gilman et al., 1958), tungsten (Hull
and Beardmore, 1966; Field, 1971), and MgO (Field, 1971), a propagating fracture
attains a speed of up to 0.9cR , as cleavage through a weak plane takes place, hinting
that strong anisotropy in materials may be necessary for the fracture to attain the
limiting velocity cR . An interesting experiment by Washabaugh and Knauss (1994)
indicated that this may indeed be the case. In their experiment, plates of PMMA
were first fractured and then rehealed to form a preferred plane in the material that
was substantially weaker than the material on either side of it. Although the inter-
face did weaken the PMMA, the rehealed material still had between 40% and 70%
of the strength of the original material. Using an interferometer together with a
high-speed rotating mirror camera, interferograms of the fracture tip were recorded
at equal time intervals. In this way a fracture velocity of up to 0.9cR was measured.
Washabaugh and Knauss (1994) also noted that none of the fractures propagating
along the weakened interfaces produced branches beyond the point of fracture ini-
tiation. The same type of behavior takes place in strongly anisotropic crystalline
materials. Field (1971) noted that in experiments on MgO and rolled tungsten
(rolling in the preparation of tungsten induces a preferred orientation in the mate-
rial, hence making it anisotropic) branching of a fracture is suppressed until very
high velocities. Thus, in strongly anisotropic materials, where microscopic crack
branching is inhibited, fractures approach the predicted limiting velocity of cR .
Let us mention here that there have been some continuum models of dynamic
fracture that predict that a fracture tip may propagate with a speed even larger
than cR . For example, Langer (1992) investigated three 1D and 2D unsteady-state
models of fracture propagation. His 1D models had the following general form
∂ 2u ∂ 2u
= − αf2 (u − δ) − Fc (u) − F, (124)
∂t 2 ∂x 2
where u(x, t) is the displacement of the material at time t and position x along
the face of the fracture, and αf2 is some sort of a force constant representing a
linear elastic coupling between the4 ∞fracturing material and a fixed substrate. Here,
Fc (u) is the cohesive force, i.e., 0 Fc (u)du =  is the fracture energy, and F is
a function that depends on u̇ = ∂u/∂t. A fully relaxed configuration of the system
corresponds to u = δ. In model 1, F was a friction force given by F = c1 + 2c2 ut ,
where c1 and c2 are constant. For this model Langer (1992) [see also Langer
and Tang (1991)] found that, if c2 < 0, then slipping friction causes the fracture
to accelerate to the limiting wave velocity independent of loading strength. The
second model was a 2D version of model 1 with the same qualitative behavior. In
model 3, F = −η∂ 2 ut /∂x 2 , where η is a viscosity-like coefficient. It was found
that the steady-state solutions at large applied stresses exhibit oscillating fracture-
opening displacements which propagate at speeds that are comparable to, or higher
than, the nominal wave speed cR , i.e., the fracture propagates supersonically. We
will come back to this interesting prediction in Section 7.8.15.
 7.8. Beyond Linear Continuum Fracture Mechanics 323

7.8 Beyond Linear Continuum Fracture Mechanics

Let us now discuss some of the phenomena in rapid fracture that are outside the
realm of linear continuum fracture mechanics.

7.8.1 The Dissipated Heat

We already described in Sections 6.11.4 and 6.11.6 the technique for measuring
the heat generated by a propagating crack in PMMA (Döll, 1973; Zimmerman et
al. 1984), in glass (Weichert and Schonert, 1974), and in steel (Zimmerman et al.,
1984). These experiments indicate that heating accounts for most of the elastic
energy that drives the fracture. In particular, for fracture velocities ranging from
0.1cR to 0.6cR , the measured heat flux accounts for 50-60% of the energy release,
whereas for fractures velocities in the range 0.1cR − 0.3cR the measured heat flux
accounts for virtually the entire energy release.
Although these experiments tell us that almost all of the elastic energy is con-
verted into dissipated heat, a central question is where this dissipation takes place
within the material. Is it in, for example, the cohesive zone, or does it occur as
elastic waves that propagate away from the fracture are attenuated within the ma-
terial? Fuller et al. (1983) provided the answer to this central question by real-time
infrared visualization of the fracture tip during its propagation. Their experiments
were carried out on PMMA and polystyrene, and indicated that, in both materials,
temperatures at the tip were approximately constant, as a function of the frac-
ture’s velocity, with a temperature rise of the order of 500 K (see also Zehnder
and Rosakis, 1991; Kallivayalil and Zehnder, 1994, for similar data for AISI 4340
carbon steel and β-C titanium). These experiments also established that, in addi-
tion to the large temperature rise (in PMMA and polystyrene the temperatures at
the crack tip were well above the equilibrium melting temperature), the source of
the heating was within a few µm of the crack’s path, or well within the cohesive
zone, as defined by the material’s yield stress, implying that nearly all of the heat
dissipation in the material takes place in the vicinity of a fracture. The heat release
appears to be caused by the extreme plastic deformation induced by the fracture
process in the vicinity of the tip. This result is also supported by the experiments
of Kusy and Turner (1975) who investigated the fracture energy of PMMA. They
found that the fracture energy of high (> 105 ) molecular weight PMMA can be
over two orders of magnitude larger than the surface energy (i.e., the energy needed
to break a unit area of atomic bonds), which they explained it in terms of plastic
deformation of the polymer chains, whereas below a molecular weight of about
105 , no significant plastic deformation occurred in the fracture, and its energy was
comparable with the surface energy.
The dissipated heat and the associated temperature rise in the material can be cal-
culated. For example, Langer (1993) and Langer and Nakanishi (1993) considered
a 2D dynamic fracture model defined by
∂ 2u
= c∇ 2 u − m2 (u − δ), (125)
∂t 2
324 7. Brittle Fracture: The Continuum Approach

where c is a wave speed, m is the mass, and m2 δ is the applied force. The fracture
is essentially moving along the center line of a strip of finite width. The traction
applied to the fracture surface was assumed to be given by
. .
∂u .. ∂ 2 u̇ ..
µ . = σc (u) − η 2 . , (126)
∂y y=0 ∂x .
x=0
where µ is an elastic modulus (for example, the shear modulus), and σc is the
cohesive stress acting between the open fracture faces. The cohesive stress was
taken to be σc = σy for 0 ≤ u(x, 0, t) ≤ δc , and σc = 0, otherwise, where σy is the
yield stress, and δc is obviously the range of the cohesive force; note that u(x, 0, t)
is just the fracture-opening displacement. The second term on the right-hand side
of (126) is a viscous damping stress which acts on the fracture surface. The two
spatial derivatives preserve reflection and translational symmetries, and the time
derivative in (126) breaks time-reversal symmetry in order to produce energy
dissipation. The most interesting prediction of the model was a relation between
the velocity of fracture propagation and the externally applied stress, given by
$ %12
v/c K
 , (127)
[1 − (v/c)2 ]3/2 Ke
which is valid for 1 !K/KG  (w/δc )1/6 (σy /µ)1/6 , where w is the width of
the system, and KG = 2σy δc /µ. Here K is the stress-intensity (more precisely,
the strain-intensity) factor associated with the applied force, and
$ %1/12 $ %2/3
1/3 ηcR σy
Ke = (6δc ) . (128)
µ µ
The surprising aspect of these predictions is the unusual exponent 1/12. If Ke >
KG , then the fracture velocity v jumps from very small values to values near cR
as K passes through Ke , and therefore Ke plays the role of an effective Griffith
threshold at which the fracture makes a sharp transition from slow motion to rapid
propagation. Whether such predictions can be observed in an actual experiment
remains an open question. The dissipated energy is η(∂ u̇/∂x)2 , and assuming that
this energy is converted to heat, then the corresponding temperature rise T will
be
#
K 2 µ3 v
T  , (129)
3Cp η
where Cp is the specific heat of the material.
How do the thermal effects in the cohesive zone influence our basic understand-
ing of the fracture process? Since the fracture energy is an input into the theory of
linear continuum fracture mechanics, neither the large temperature rise observed
within the cohesive zone nor its cause(s) have any effect on the predictions of the
theory. This is true so long as the heat dissipation is localized within the cohe-
sive zone and does not spread out throughout the material. Otherwise, the entire
rationale behind Eqs. (91)–(93) would be invalid. Thus, the total fracture energy
 7.8. Beyond Linear Continuum Fracture Mechanics 325

is related to the amount of microscopic surface actually generated by the fracture

process, which in turn is related to instabilities that occur to a single fracture as a
function of the energy that it dissipates.

7.8.2 The Structure of Fracture Surface

Studies of fracture surfaces of amorphous brittle materials indicates that the sur-
face (which is generated by dynamic fracture) has a characteristic structure called
mirror, mist, and hackle. These characteristics, which were already described in
Section 6.13, have been reported to occur in materials as diverse as glasses and
ceramics, non-crosslinked glassy polymers such as PMMA, and crosslinked glassy
polymers, such as Homalite-100, polystyrene and epoxies. To summarize our de-
scription of the structure of a fracture surface given in Chapter 6, near the location
of the fracture onset, the fracture surface appears smooth and shiny, and is thereby
called the mirror region. As a crack propagates further, the fracture surface be-
comes cloudy in appearance, and is referred to as mist. When the fracture surface
becomes extremely rough, it is said to be in the hackle region.

7.8.3 Topography of Fracture Surface

It is often useful to make quantitative measurements of the topography of a fracture
surface, for which there are several techniques, each of which is appropriate for
a specific length scale. For length scales ranging from 1 to 100 µm, commercial
contact-type scanning profilometers is used for measuring such properties as the
root mean-square of roughness of a surface. However, the tip size of the contact
probe limits its resolution in resolving surface features that are under 10 µm, in
which case optical profilometers have been used (Boudet et al., 1995). Fracture
surfaces at submicron length scales have recently been studied (Milman, 1994;
Daguier et al., 1997), using both scanning tunneling and atomic force microscopy.

7.8.4 Properties of Fracture Surface

Analysis of fracture surfaces, usually called fractography (Hull, 1999), is con-
cerned with the determination of the location of the onset of fracture of a given
material together with the probable cause for its failure. Although every material
has its own fracture surface which is different from that of any other material, the
proven usefulness of analysis of fracture surface in the determination of different
fracture processes stems from the fact that, a close empirical relation exists be-
tween the deterministic dynamics of a fracture and the surface that it creates. The
mechanisms that give rise to characteristic surface features are, in many cases, not
known, but the fact that these features are at all general is strong evidence that they
are generated by a deterministic process, independent of details of the loading or
the initial conditions of the material under study.
326 7. Brittle Fracture: The Continuum Approach

Figure 7.7. Typical parabolic markings

formed on the fracture surface of PMMA
(after Ravi-Chandar and Yang, 1997;
courtesy of Professor K. Ravi-Chandar).

7.8.5 Conic Markings on Fracture Surface

Fracture surfaces of amorphous materials also contain small conic (or parabolic
on a surface) markings in the mist region; see Figure 7.7. They appear in all
three fracture regimes, namely, mirror, mist, and hackle, and are the result of mi-
croscopic defects opening up ahead of the main fracture front. To see the origin
of these markings, suppose that a microscopic void is placed directly ahead of
a fracture. The large stress field, generated at the fracture’s tip, causes the void
to propagate some distance before the main fracture catches up with it. Smekal
(1952) postulated that in the large stress field of the main crack, heterogeneities
trigger the initiation of a secondary fracture ahead of the primary crack. The sec-
ondary fracture may not be in the same plane as the primary front. When these
two fronts intersect in space and time, the ligament separating the two fractures
breaks up, leaving a conic marking on the fracture surface. Therefore, the conic
marking indicates a level difference boundary, marking the common space time
interaction of the two fracture fronts, with the focus of the conic identifying the
origin of the secondary fracture front. The existence of the conic markings indi-
cates an increase in the number of voids activated into growing along the fracture
path, and an increase in the nucleation distance at which the secondary microcracks
 7.8. Beyond Linear Continuum Fracture Mechanics 327

begin to grow (see Sections 7.8.9–7.8.11 for a discussion of fracture branching).

Carlsson et al. (1972) observed that the number of the markings increases with
the fracture velocity, which is consistent with the fact that an increasing number
of voids is nucleated ahead of the fracture tip as the stress at the tip increases. We
should, however, mention that Ravi-Chandar and Yang (1997) reported that there
is no one-to-one correlation between the number or density of the markings and
the mean velocity of fracture. This is similar to lack of a one-to-one correspon-
dence between the stress intensity factor and the fracture velocity, which will be
discussed below (see Section 7.8.11). Shioya and Ishida (1991) found the depth of
the conic markings in PMMA to be approximately 1 µm. Ravi-Chandar and Yang
(1997) carried out a comprehensive study of the development of the conic mark-
ings as a function of the velocity of a fracture for four polymeric materials which
were PMMA, Homalite-100, Solithane-113 and polycarbonate. Solithane-113 is
a polyurethane elastomer which exhibits brittle fracture behavior. Polycarbonate
is a non-crosslinked thermoplastic polymer which is capable of inelastic defor-
mation, since the mobility of the carbonate segments of its structure is relatively
high. However, at large rates of loading, it does exhibit brittle dynamic fracture.
Ravi-Chandar and Yang (1997) found that the markings in all of these materials
increase in density with increasing values of the stress intensity factor.

7.8.6 Riblike Patterns on Fracture Surface

In the mist and hackle regions of many brittle polymers, such as polystyrene (Hull,
1970), PMMA (Fineberg et al., 1992), and Solithane-113 and polycarbonate (Ravi-
Chandar and Yang, 1997), rib-like patterns on the fracture surface are commonly
observed. In such materials, the typical distance between the markings is of the
order of 1 mm, so that they can easily be seen by naked eye. In PMMA, for example,
on which extensive work has been carried out for characterizing such patterns, the
rib-like patterns have been found to initiate within the mist regime. The initial width
of these patterns is usually much less than the sample’s thickness, but it increases
with the fracture velocity and eventually, within the hackle zone, extends across
the entire thickness of the sample (Sharon and Fineberg, 1996). These patterns,
rather than being smooth undulations along the fracture surface, are discrete bands
of jagged cliff-like structures. Their height increases with the fracture velocity, and
they exist up to the point where a fracture undergoes macroscopic branching. The
spacing between the ribs is also strongly related to the molecular weight of the
monomers used to form PMMA (Kusy and Turner, 1975), with the typical spacing
increasing by over two orders of magnitude as the molecular weight was varied
between 104 and 106 . Moreover, the fracture energy was found to be a strongly
increasing function of the rib spacing.

7.8.7 Roughness of Fracture Surface

We already described in Chapter 6 the roughness of fracture surface of materials,
and how the associated roughness exponent is measured. In effect, the fracture
328 7. Brittle Fracture: The Continuum Approach

surface is a self-affine fractal (see Section 1.3), and studies of aluminum alloys,
steel, ceramics and concrete indicated (Bouchaut et al., 1990, 1991; Måløy et al.,
1992) that the local width w of the fracture surface scales as
w ∼ α , (130)
where  is the scale of observation within the fracture plane, and α (which is usually
the same as the Hurst exponent H defined in Chapter 1) is the roughness exponent.
Characterization of rough surfaces and measurement of the associated roughness
exponent α were discussed in Chapters 1 and 6. As discussed in Section 6.14.1,
it appears that for both quasi-static and dynamic fracture a universal roughness
exponent, α  0.8, is obtained for  > ξc , where ξc is a material-dependent length
scale (Daguier et al., 1996, 1997). For  < ξc a different roughness exponent,
α  0.5, has been measured (Milman, 1994). Narayan and Fisher (1992) inter-
preted α  0.5 as being the result of a crack front pinned by microscopic material
inhomogeneities in very slow fracture.
As already explained in Chapter 6, the apparent length-scale dependence of the
roughness exponent α may also be explained in another way based on the velocity
of fracture propagation.According to Bouchaud and Navéos (1995) (and somewhat
similar to the argument of Narayan and Fisher, 1992), one must distinguish between
quasi-static (slow) and rapid fracture. In the former case, corresponding to small
length scales, one may obtain a roughness exponent close to 0.5, whereas rapid
fracture, which corresponds to large length scales, leads to α  0.8. Bouchaud
and Navéos (1995) thus argued for the existence of a length scale ξqs , such that
for  < ξqs one is in the quasi-static fracture regime and thus a low roughness
exponent, while at length scales   ξqs rapid fracture is dominant and therefore
one should obtain α  0.8. As shown by Daguier et al. (1997), ξqs depends on the
velocity of fracture propagation, and thus should decrease as the velocity increases.
If this picture of fracture is correct, then models that are based on minimum energy
surfaces are in the quasi-static class. Bouchaud and Navéos (1995) also showed
that the data for both cases can be expressed by the following equation
hmax
= A1 + A2 r α−αms , (131)
r αqs
where hmax is the same as before, αqs is the roughness exponent corresponding
to the quasi-static limit, α is the universal roughness exponent corresponding to
rapid fracture, αms is the roughness exponent of minimum energy surfaces, and
A1 and A2 are two constants.
We should point out that, despite the considerable effort that had gone into
understanding the properties of self-affine fracture surfaces, up until recently, there
was little discussion of the fact that, in many of the experiments in which a non-
trivial roughness exponent had been measured for a fracture surface, the typical
length scales where the scaling behavior had been observed were several orders of
magnitude smaller that the typical sample size. For example, the largest length scale
observed in measurements performed on soda-lime glass (Daguier et al., 1997) was
of the order of 0.1 µm, which is well within the mirror regime. Thus, in the context
 7.8. Beyond Linear Continuum Fracture Mechanics 329

of continuum models of dynamic fracture, the roughness at such length scales

does not constitute a departure from straightline fracture propagation, although
it is conceivable that the observed scaling structure may affect the value of the
fracture energy. Although it is known that the root mean-square surface roughness
increases with the velocity of a crack within the mist and hackle regions in PMMA
(Fineberg et al., 1991; Boudet et al., 1995), Homalite-100 (Ravi-Chandar and
Knauss, 1984a), and crystals that are cleaved at high velocities (Field, 1971; Reidle
et al., 1994), we are not aware of any systematic measurements of the dependence
of the roughness on the velocity of a crack, at velocities that are of interest to
dynamic fracture.
Thus, as pointed out by Fineberg and Marder (1999), the length scales at which
the fracture surfaces have been found to be self-affine are, in general, well within
the cohesive zone. As a crack accelerates, however, the surface structure within
the mist and hackle regimes may, depending on the overall sample size, become
larger than the length scales at which the singular contribution to the stress field
in the medium is dominant. At this point the structure within the fracture surface
may no longer be swallowed up within the cohesive zone, and the description of
the dynamics of a crack will be beyond the realm of linear continuum fracture
mechanics.
More recent work by López and Schmittbuhl (1998) and Morel et al. (1998)
has addressed the scale dependence of the roughness of fracture surfaces, and
the associated roughness exponent α. It has been suggested that the apparently-
universal roughness exponent α  0.8 represents a local exponent (even though it
supposedly corresponds to rapid fracture at larger length scales). Moreover, even
if the local roughness exponent, which we now denote it by αloc , is universal, i.e.,
independent of the material, the range of length scales within which the scaling of
the width of the rough surface is observed depends strongly on the material mor-
phology. It has been shown that the scaling laws that govern the crack development
in the longitudinal and transverse directions are different and material dependent.
Consider, for example, the development of a fracture surface from a flat notch of
length L with no roughness. The mean plane of the fracture surface is marked by
the coordinates (x, y) where the x-axis is perpendicular to the direction of crack
propagation, while the y-axis is parallel to the crack propagation direction. It has
been found that the height fluctuations h of the fracture surfaces of two heteroge-
neous brittle materials—granite (López and Schmittbuhl, 1998) and wood (Morel
et al., 1998)—estimated over a window of size  along the x-axis and at a distance
y from the initial position exhibits scaling properties that are much more complex
than what is predicted by Eq. (130) for the transverse direction, and are described
by the following anomalous scaling properties,

αloc ξ(y)α−αloc , if   ξ(y),
h(, y)  A (132)
ξ(y)αloc , if   ξ(y),

where ξ(y) = By 1/z depends on the distance to the initial notch y and corresponds
to the crossover length along the x-axis below which the fracture surface is self-
330 7. Brittle Fracture: The Continuum Approach

affine with a local roughness exponent αloc . The quantity z is the dynamic exponent
for rough surfaces that was already introduced in Section 1.5.
The scaling laws (132) indicate that along the y-axis the roughness develops ac-
cording to two different regimes: For large length scales [  ξ(y)], the roughness
grows as h ∼ y α/z , where α is called the global roughness exponent, whereas
for small length scales [  ξ(y)] the roughness growth is characterized by the
exponent (α − αloc )/z. Unlike the local roughness exponent, the global exponent
α, as well as the dynamic exponent z and the prefactors A and B are material
dependent, and hence non-universal. Thus, despite exhibiting universality in the
transverse direction, roughening in the longitudinal direction is material dependent.
An important consequence of scaling laws (132) is that, when the global sat-
uration occurs, i.e., far from the notch for y  ysat [where ysat = (L/B)z ], the
magnitude of the roughness is not only a function of the window size  but also
of the system size L, since in this case, h(, y  ysat )  Aαloc Lα−αloc . It is
for this reason that scaling laws (132) are viewed as anomalous because in the
conventional scaling of rough surfaces that were described in Section 1.5 one has

αloc , if   ξ(y),
h(, y)  A (133)
ξ(y) , if   ξ(y).
α loc

If fact, the scaling laws (132) and (133) become equivalent only if we take the
global roughness exponent α to be equal to the local exponent αloc .
These anisotropic scaling laws have important implications for the Griffith
criterion which will be described shortly.

7.8.8 Modeling Rough Fracture Surfaces

Although, in addition to their experimental realization, self-affine fracture surfaces
have been clearly produced in molecular dynamics simulations of dynamic frac-
ture (see Chapter 9), an important unsolved problem, which is outside the realm
of linear continuum fracture mechanics, is a proper model that can generate self-
affine fracture surfaces with the roughness exponents that have been measured in
many experiments. As usual, this problem has been attacked by many, employing
many different ideas. For example, J.P. Bouchaud et al. (1993) proposed a model
based on directed percolation. In directed percolation (see, for example, Kinzel,
1983; Duarte, 1986, 1990, 1992; Duarte et al., 1992), the bonds of a lattice are
directed and diode-like. Transport along such bonds is allowed only in one direc-
tion. If the direction of the external potential is reversed, then, there may be no
macroscopic transport in the new direction of the external potential. Unlike the
regular percolation, there are two correlation lengths in directed percolation that
characterize the shape of the percolation clusters. One is the longitudinal corre-
lation length (in the direction of the external potential), while the second one is
the transverse correlation length, in the direction perpendicular to the direction of
the external potential. As a result, one must also have two critical exponents that
characterize the scaling of the correlation lengths near the percolation threshold.
One is νL which is associated with the longitudinal correlation length, while the
 7.8. Beyond Linear Continuum Fracture Mechanics 331

second exponent is νT , associated with the transverse correlation length. How-

ever, although the directed percolation model does provide a prediction for the
roughness exponent, namely, α = νT /νL , its numerical value in 2D, α  0.63,
or in 3D, α  0.57, is not in good agreement with the data discussed above. J.P.
Bouchaud et al. (1993) also proposed a set of coupled equations which do have
some of the required symmetries and properties appropriate to this phenomenon.
Their equations are given by
$ % $ %
∂x ∂ 2x λxx ∂x 2 λxz ∂z 2
=v+ + + + Nx (y, t), (134)
∂t ∂y 2 2 ∂y 2 ∂y
∂z ∂ 2z ∂x ∂z
= 2
+ λz + Nz (y, t). (135)
∂t ∂y ∂y ∂y
Here, x is the direction of fracture propagation, y is along the fracture front, z is the
tensile axis, v is the nominal fracture velocity, is the line tension, N represents
noise or disorder in the material, and the λs are coupling constants. The nonlinear
terms signify the fact that the local velocity of the fracture depends on its local
direction. They are designed to satisfy the required symmetries, namely, y → −y
and z → −z. The same type of equations were discussed by Ertas and Kardar
(1992, 1993, 1994, 1996) in the context of driven vortex lines in superconductors,
and the morphology of polymers in shear flows. In their model the flux lines are
pulled away by a constant force. Their equations are nonlinear, with the nonlin-
earity accounting for the variations of the local propagation speed with the local
orientation of the front. Depending on the values of the parameters, many distinct
scaling regimes are predicted by these equations. In particular, in a certain limit and
for a finite velocity, Ertas and Kardar found that α  0.75 at large length scales,
and α  0.5 at short length scales, quite close to the experimental values of α dis-
cussed above and in Chapter 6. However, the exact correspondence between the
problem discussed by Ertas and Kardar and self-affine fracture surfaces is not clear.
J.P. Bouchaud et al. (1993) and E. Bouchaud et al. (1993a) also suggested that a
fracture surface may be modeled as the trace that is left by the fracture front prop-
agating in a medium with randomly-distributed obstacles. The model proposed by
Hansen et al. (1991), based on an analogy with directed polymers in random me-
dia first proposed by Kardar et al. (1986)—the KPZ equation described in Section
1.6—also does not produce the experimentally-measured value of the roughness
exponent, since it predicts that α = 2/3. Schmittbuhl et al. (1995) proposed a
perturbative approach to describe the evolution of a fracture between two elastic
solids, in which the driving force was the stress intensity factor along the fracture
front. The resulting fracture surface was rough and self-affine, but the roughness
exponent was only α  0.35, which does not agree with any of the experimental
data described above. We will come back to this issue later in this chapter and also
in Chapter 8.
A completely different approach was suggested by Räisänen et al. (1998). They
suggested an analogy between quasi-static fracture surfaces and minimal energy
surfaces. Although both types of surfaces are rough, it may seem surprising that
332 7. Brittle Fracture: The Continuum Approach

the two can be related, since the minimal energy surfaces, such as those obtained in
the random-bond Ising model, seem to have little, if anything, to do with fracture
of a material. Nevertheless, Hansen et al. (1991) suggested, and Räisänen et al.
(1998) confirmed by extensive numerical simulations, that the roughness exponent
of the two types of surfaces in 2D are the same. In particular, Räisänen et al. (1998)
used a scalar approximation to model fracture of a brittle material—the random
fuse model described in Section 5.2—to provide strong numerical evidence for
this equality. However, in 3D the scalar quasi-static fracture model was found to
be rougher than the minimal energy surfaces.

7.8.9 Fracture Branching at Microscopic Scales

As described in Chapter 6, in an early study of fracture of glass rods, Johnson and
Holloway (1968) demonstrated, by progressive etching of the fracture surface in
the mist region, the existence of microscopic cracks that branch away from the
main fracture. Similar microscopic branched cracks were later observed by Hull
(1970) in polystyrene, by Ravi-Chandar and Knauss (1984b) in Homalite-100, and
by Anthony et al. (1970) during rapid fracture of tool steel. In fact, as we discuss
later in this chapter, formation and evolution of micro-branches strongly influence
the dynamics of a fracture.

7.8.10 Multiple Fractures Due to Formation and Coalescence of

Microscopic Voids
Experiments carried out on Homalite-100 by Ravi-Chandar and Knauss (1984a)
suggest that, one should not view dynamic fracture as the propagation of a single
fracture, but as the coalescence of microscopic voids that are formed ahead of a
fracture front. In their experiments fracture was generated via the electromagnetic
loading method described in Chapter 6 in which a trapezoidal pressure profile
with a 25 µs rise time and 150 µs duration was applied to the faces of a seed
microcrack. The sample material was large enough that the first reflected waves
from its boundaries would not interact with the fracture throughout the experi-
ment. It was observed that within the mist and hackle regions, a front of multiple
microscopic parallel cracks, instead of a single fracture, was formed. The cracks
in the mirror region tended to propagate within a single plane, whereas in the mist
region caustics due to the formation of multiple fractures tips (which were seen in
high speed photographs) were observed, the intensity of which increased within
the hackle regime as the secondary fractures increased in size. Ravi-Chandar and
Knauss (1984a) proposed that formation of the multiple micro-cracks was due to
the nucleation of microscopic material flaws or voids, the traces of which were in-
dicated by the conic markings left on the fracture surface. Earlier, Broberg (1979)
had in fact proposed that, these voids are nucleated by the large stresses ahead of
the fracture front, so that the dynamics of fracture propagation is dictated by the
interactions between these growing flaws and the fracture front. We have already
discussed this phenomenon, and therefore do not elaborate further.
 7.8. Beyond Linear Continuum Fracture Mechanics 333

7.8.11 Microscopic Versus Macroscopic Fracture Branching

If, relative to the size of a sample material, the crack branches remain small,
then they can be considered as part of the cohesive zone. In materials such as
Homalite-100, above a certain energy flux, fractures are made of many microscopic
cracks propagating in unison. Such microscopic multiple fractures are observed
in a variety of materials within the mist and hackle zones. However, in sample
materials of any given size, an increase in size of microbranches with the energy
release rate H will eventually make the size of the cohesive zone large enough
that the assumptions of continuum fracture mechanics break down. As soon as
a crack begins branching, single fracture models are, of course, no longer valid.
Therefore, theories that are based on formation of a single fracture can, at best,
provide a criterion for when fracture branching may begin. We already discussed a
few of such branching criteria, such as that of Yoffe (1951) and those that are based
on extremal energy density. However, as discussed above, the same criteria also
predict fracture velocities for the onset of branching that are much too large. Other
criteria, such as those that postulate a critical value of the stress intensity factor,
are not consistent with experiments (Arakawa and Takahashi, 1991; Adda-Bedia
and Ben Amar, 1996) since they indicate that there is considerable variation of the
stress intensity factor KI at the point of branching. Another criterion was suggested
by Eshelby (1971) according to which a fracture branches when the energy  that
creates a single propagating fracture is large enough to support two single cracks.
However, this criterion suffers from the fact that if  were not a strongly increasing
function of v, then once branching began, one should observe a large decrease in
the velocities of the branches relative to that of the single fracture that preceded
the branching event. In glass, however, the post branching velocities either do
not decrease at all (Schardin, 1959), or decrease at most by about 10% (Kerkhof,
1973). It should, however, be clear that the Eshelby criterion is a necessary, but
not sufficient, condition for fracture branching.
Yoffe’s proposal that there exists a universal critical velocity for macroscopic
fracture branching is not supported by experimental observations. For example,
branching velocities in glass are between 0.18cR and 0.35cR (Schardin, 1959),
in PMMA are consistently about 0.78cR (Cotterell, 1965), and in Homalite are
between 0.34cR and 0.53cR (Arakawa and Takahashi, 1991). In any experiment
on fracture branching, one must ensure that branching occurs at locations that are
far from the lateral boundaries so that the system can be considered as effectively
infinitely large. Otherwise, experiments have indicated (Ravi-Chandar and Knauss,
1984c) that branching can be created by the arrival of waves generated at the
onset of fracture and reflected at the lateral boundaries of the system back into
the fracture tip. Despite such difficulties, the consistent values of the measured
branching angles in many different materials indicate that there may be a degree
of universality in the macroscopic branching process. The branching angles have
been typically determined by measurement of the tangent of a branched fracture at
distances of the order of a fraction of a millimeter from the fracture tip. They range
from 10◦ in PMMA (Cotterell, 1965) and glass (Johnson and Holloway, 1968) to
334 7. Brittle Fracture: The Continuum Approach

14◦ in Homalite, 15◦ in polycarbonate (Ramulu and Kobayashi, 1985) and about
18◦ in steel (Anthony et al., 1970), all of which were measured for materials that
were under pure uniaxial tension.

7.8.12 Nonuniqueness of the Stress Intensity Factor

Another discrepancy between the theory and experiment was discovered by Ravi-
Chandar and Knauss (1984a) in experiments on Homalite-100. They took high
speed photographs of the caustic formed at the tip of a fracture initiated by electro-
magnetic loading at high loading rates. The velocity of the fracture was estimated
from the position of its tip in the photographs, and was compared with the instan-
taneous value of the stress intensity factor, which had been estimated from the
size of the caustic. In agreement with the theory, at low velocities (below about
300 m/s = 0.3cR ) a change in the value of the stress intensity factor resulted in
an instantaneous change in the fracture’s velocity. However, at higher velocities
significant changes in the stress intensity factor produced no measurable change
in the fracture’s velocity, indicating that the stress intensity factor is not a unique
function of fracture velocity.

7.8.13 Dependence of the Fracture Energy on Crack Velocity

Due to its fundamental importance, the fracture energy —the energy needed
for generating a unit fracture surface—and its dependence on the crack velocity
have been measured for many different materials, for which the most common
technique is the method of caustics described in Section 6.11.2.1. Measured values
of the fracture energy  in single crystals, which are necessary for initiating crystal
cleavage, agree well with the theoretical predictions (see, for example, Lawn, 1993,
for a review). In amorphous or polycrystalline materials, however, experiments
indicate that (v) is a strongly increasing function of a fracture’s velocity, the
form of which is known only empirically. Most of the fracture energy is dissipated
as heat within the cohesive zone, or is radiated from the crack as acoustic energy,
or is lost as the emission of photons from excited molecules along the fracture
surface—the so-called fracto-emission (Dickinson, 1991).
Figure 7.8 presents some typical measurements of fracture energy  versus
fracture velocity v for PMMA, Homalite-100 and AISI 4340 √ steel. Also shown are
the dimensionless velocity v/cR versus  = KI /KI c = H/Hc , a dimensionless
measure of loading which is the ratio of the stress intensity factor KI and the
critical value KI c of KI at which fracture first begins. Although these materials
are quite different, a common feature among them is the steep rise in  as the
fracture velocity v increases. For steels, the increase in  is due to the fact that the
cohesive zone acts as a plastically deforming region (Freund, 1990). However, in
the case of PMMA and Homalite-100, which are brittle amorphous materials, there
is no reason to expect the classical theory of plasticity to describe deformations
near the fracture tip.
 7.8. Beyond Linear Continuum Fracture Mechanics 335

Figure 7.8. The dependence of the fracture energy  on the fracture velocity v, for (top
row, left to right) AISI 4340 steel (Rosakis et al., 1984), and PMMA (Sharon et al., 1996).
The bottom row shows the rescaled data, where  = KI /KI c , and cR is the Rayleigh wave
speed.

Figure 7.8 does in fact reflect the view of Dally (1979) who studied extensively
dynamic fracture in amorphous polymers, and in steels. According to him,
(1) the proper way to characterize a dynamic fracture experiment is through pre-
senting the data by two dimensionless numbers which are v/c, the ratio of the
fracture velocity and a wave speed, and  = KI /KI c , the ratio of the dynamic
stress intensity factor and its critical value at the fracture onset. The relation
v/c = f () contains most of the information about the dynamics of fracture.
(2) The energy needed for fracture of brittle amorphous materials increases steeply
past a critical velocity, where the straight-line fracture becomes unstable to
frustrated branching events.
We will come back to these points later in this chapter.

7.8.14 Generalized Griffith Criterion for Fractures with

Self-Affine Surfaces
If fracture surfaces are self-affine fractals, then one must think about modifying
the Griffith criterion in order to accommodate this fact. Such a generalization was
first suggested by Mosolov (1993). Bouchaud and Bouchaud (1994), considered
the case in which no distinction was made between the growth of the fracture
surface in the longitudinal and transverse directions, and the local and global
336 7. Brittle Fracture: The Continuum Approach

roughness exponents were assumed to be the same. This case, as described in

Section 7.8.7, corresponds to an isotropic fracture surface at small length scales,
which we consider first. Thus, consider the case of non-fractal fracture surfaces and
derivation of, for example, Eq. (66). We assume quite generally that KI ∼ r −ζ ,
where KI is the stress intensity factor [ζ = 1/2 yields Eq. (66)]. If the fracture
path is smooth, then the surface energy is simply
H = 2 w, (136)
where is the surface tension, w is the width, and  is the fracture length increment.
The released elastic energy  is estimated by noting that, since the stress field is
relaxed on length scales r <  and unperturbed on larger scales, then

K 2  −2ζ wKI2
 I r wrdr  2−2ζ , (137)
2Y rc 4Y (1 − α)
where rc is a microscopic cutoff length scale below which the stress saturates, and
Y is the Young’s modulus. According to the Griffith’s criterion, at the onset of
fracture one must have H = , which results in ζ = 1/2, as expected.
We now suppose that the fracture surface is self-affine at length scales ξ and is
represented by a height profile h(r) given by
$ %α
r
h(r) = (r)hmax , rξ (138)
ξ
where (r) is a random variables of order 1. For r  ξ we must have h(r) =
(r)hmax . Following Griffith’s method, one must calculate the surface energy
corresponding to opening of the fracture along a distance   ξ , which is given
by
#
  $ %2
dh
H2 w 1+ dr. (139)
0 dr
Equation (139) indicates that there is a new length scale ξ ∗ at r = ξ ∗ such that one
has dh/dr  1; for r  ξ ∗ one has dh/dr  1. Bouchaud and Bouchaud (1994)
argued that
$ %
ξ∗ hmax 1/(1−α)
 (140)
ξ ξ
One must distinguish between two distinct cases:
(1) If hmax  ξ or ξ ∗  ξ , which is the regime in which the surface is a self-
affine fractal but shallow, i.e., it has a mean local angle of the crack profile
smaller than 45◦ , and there is no sample size effect. Then
$ % $ %α
hmax R
H  2 wξ ,  < ξ ∗. (141)
ξ ξ
However, as soon as  > ξ ∗ one has H  2 w, even if R < ξ , so that the
surface energy is similar to that needed to create flat surfaces, even though
 7.8. Beyond Linear Continuum Fracture Mechanics 337

the surface is rough, and thus the stress-field singularity is the usual Griffith’s
singularity, −1/2 . Equating (137) and (141) leads to α = 2 − 2ζ (yielding
the Griffith’s result, ζ = 1/2, when α = 1, i.e., when the fracture surface is
smooth). Thus, rougher fractures, i.e., those with smaller α, lead to a more
singular stress field.
(2) In the second regime, hmax  ξ or ξ  ξ ∗ . In this case the slope of the surface
over the entire fractal domain is larger than one, resulting in a spiky regime,
and hmax /ξ is a measure of this spikiness. Near the tip of the fracture (r < ξ )
the stress field is characterized by the exponent ζ = 12 (2 − α).
However, the above considerations are valid when the anisotropy in the growth
of rough fracture surfaces is not taken into account. As described in Section 7.8.7,
the height fluctuations in the longitudinal and transverse directions exhibit distinct
scaling properties that are characterized by Eqs. (130) and (132). In particular,
one has an anomalous, size-dependent scaling in the saturation regime, which
must be taken into account if one is to generalize the Griffith criterion for the
onset of fracture. Based on these scaling laws, Morel et al. (2000) proposed a
modified form of the Griffith criterion. To understand their proposal, consider a
semi-infinite linear elastic material of thickness L that contains an initial crack
at position a and in Mode I (i.e., under a uniaxial stable and low tension). In
the zone where the roughness of the fracture surface grows, i.e., for a  ysat
[where ysat = (L/B)z defined in Section 7.8.7], the critical energy release rate Hc
during fracture propagation (which, in Griffith’s approach, is set to be equal to the
energy required for generating the corresponding free surfaces at the microscale;
see above and Section 6.7) is given by

  2
 AB α−αloc
c = 2s 1 + 1−α
a 2(α−αloc )/z , a  ysat , (142)
0 loc
where 0 is the lower cut-off for the length scale over which the fracture surface is
a self-affine fractal, i.e., 0 is the characteristic size of the smaller microstructural
element which is relevant for the fracture process, and s is the specific surface
energy that characterizes the resistance of the material to fracturing. The quantities
A and B and the exponents α and αloc were already introduced and discussed in
Section 7.8.7.
On the other hand, when the crack increment is large (i.e., a  ysat ), which
corresponds to the saturation state of the roughness, one has

  2
 A
c = 2s 1 + 1−α
L2(α−αloc ) , a  ysat , (143)
0 loc
implying that the energy c is independent of a, but depends on the linear size
L of the sample, an important characteristic of brittle fracture of heterogeneous
materials. Equation (143) indicates that the size effect gives rise to two asymptotic
behaviors which are, c ∼ 2s and c ∼ Lα−αloc . The crossover between the two
338 7. Brittle Fracture: The Continuum Approach

1−α
occurs at a length Lco = (0 loc /A)1/(α−αloc ) . Hence, for L  Lco the fracture
surface is shallow, and there is no size effect, c  2s . In this case, the classical
results of linear continuum fracture mechanics are applicable to fracturing of the
material. However, for L  Lco one has a power law
c ∼ Lα−αloc > 2s . (144)
Equation (144) was found to agree with the experimental data for wood (Morel et
al., 1998). Note that, if the anomalous scaling is neglected, and the fracture surface
is described by scaling laws (133), then

  2
 A
c (a)  2s 1 + 1−α
, (145)
0 loc
that is, there is no dependence on the size of the material, which is the case for
purely elastic brittle materials.

7.8.15 Crack Propagation Faster Than the Rayleigh Wave Speed

Our discussions so far should have made it clear that linear continuum fracture
mechanics predicts that a crack cannot propagate with a speed larger than the
Rayleigh wave speed cR . Briefly, continuum mechanics predicts that for Mode I
tensile loading there is a forbidden velocity zone (FVZ) for fracture propagation
which is a zone in which the speed of the propagation cannot be larger than cR .
For Mode II shear loading, the FVZ exists only for speeds between cR and shear
wave speed ct . Therefore, in Mode I a crack’s limiting speed is also cR because
its FVZ between cR and ct acts as an impenetrable barrier for the shear cracks to
go beyond cR .
However, several experiments have been reported in which the cracks propa-
gated with a speed larger than cR . Winkler et al. (1970) reported supersonic crack
propagation along weak crystallographic planes in anisotropic single crystals of
potassium chloride, where the fracture tip was loaded by laser-induced expanding
plasma. Supersonic crack tip speeds are those that are larger than the dilatational
wave speed cl which itself is larger than ct . At much larger length scales, indirect
observations of intersonic (i.e., one with a speed v between ct and cl ) shear ruptures
have been reported for shallow crustal earthquakes (Archuleta, 1982; Olsen et al.,
1997). In this case, the fault motion is primarily shear dominated, and the material
is not strictly monolithic because preferred weak rupture propagation paths exist
in the form of fault lines.
Rosakis et al. (1999) carried out interesting laboratory experiments to determine
whether in-plane shear intersonic crack growth can be obtained in materials that
are under remote shear loading conditions. They utilized two identical plates of
Homalite-100 polymer, and introduced a weak plane ahead of the notch tip (used
for initiating crack propagation) in the form of a bond between the two identical
samples of the materials. The bonding process was done carefully so that the con-
stitutive properties of the bond were close to those of the bulk material. In this way,
fracture toughness along the line was lower. Dynamic photoelasticity described
 7.8. Beyond Linear Continuum Fracture Mechanics 339

2500
Cl

2000
√2Ct
Crack Speed (m/s)

1500
Ct

CR
1000

500
Notch Tip

0
0 20 40 60 80 100 120
Crack Length (mm)
Figure 7.9. Supersonic crack propagation velocity in Homalite-100 (after Rosakis et al.,
1999).

in Section 6.11.2.3 was used for recording the stress field near the propagating
fracture. The sample was subjected to asymmetric impact loading with a projectile
at 25 m/s, and sequences of isochromatic fringe patterns were recorded around a
shear fracture as it propagated along the interface between the two Homalite halves.
Crack tip speeds were measured independently from crack length history. Figure
7.9 shows the speed of the propagating crack versus the crack length. Initially,
the crack tip speed is close to the shear wave speed of Homalite-100, beyond
which it accelerates and becomes intersonic. Thereafter, it continues to accelerate
up to the plane stress dilatational wave speed
√ of the material, then decelerates and
approaches a steady-state value of about 2ct . As mentioned above, the speeds
between cR and ct are in the FVZ.
Observations of fast shear rupture during earthquakes have also provided the
impetus for a considerable amount of theoretical work. We already mentioned in
Section 7.7.1.1 the theoretical work of Langer (1992) which predicted the possi-
bility of supersonic fracture propagation. Theoretical analysis of Andrews (1976)
had already shown that a shear
√ fracture can have a terminal velocity either less than
cR or slightly greater than 2ct , depending on the cohesive strength of the fault
plane ahead of the fracture. Burridge et al. (1979)√ concluded from their theoreti-
cal analysis that the crack speed regime ct < v < 2ct is inherently unstable for
dynamic shear crack growth. Broberg (1989) showed that the crack speed regime
cR < v < ct is forbidden for both opening and shear mode cracks, a result that
was mentioned above. He also showed that the regime ct < v < c√ l is forbidden for
opening mode cracks only. Finally, Freund (1979) showed that 2ct is the only
speed permissible for a stable intersonic shear crack.
340 7. Brittle Fracture: The Continuum Approach

The existence of crack growth with a speed larger than cR has also been con-
firmed by large-scale molecular dynamics simulations of dynamic fracture. These
simulations will be discussed in Chapter 9.

7.9 Shortcomings of Linear Continuum Fracture

Mechanics
Our discussion so far has been an attempt for providing an overview of linear
continuum fracture mechanics. As discussed above, the general principle is that
by balancing the energy flowing into the vicinity of a fracture’s tip with what is
required for creating new fracture surface one can predict the motion of a straight,
smooth fracture. In addition, continuum fracture mechanics can predict both the
strength and functional form of the near-field stresses, and its predictions agree
well with the experimental data (see below). However, as Fineberg and Marder
(1999) pointed out, there still remain several issues that linear continuum fracture
mechanics cannot resolve:
(1) How does the fracture energy in brittle material vary with its velocity?
(2) What are the main processes happening in the cohesive zone?
(3) What controls a non-straight path of a rapidly propagating fracture?
(4) What controls branching of a crack into two macroscopic fractures?
(5) As discussed above (see also Chapter 6), fractures can develop rough, self-
affine surfaces. What is the controlling factor in the transition from a smooth
fracture surface to a rough one?
Many of these questions have been answered by the beautiful experimental and
theoretical work of the past decade by a few research groups, most notably by
Fineberg, Marder, and co-workers, published in a series of papers in the 1990s.
Therefore, we first discuss in the next section the essence of these experimental re-
sults and the definitive conclusions that one may draw from them. We then describe
in the next chapter the recent theoretical and computational work, the predictions
of which have turned out to be in excellent agreement with the experimental ob-
servations. These developments have helped the emergence of a coherent picture
of dynamic fracture in which instabilities caused by fracture branching play a key
role. Our discussion of the experimental results follows closely that presented in
the review by Fineberg and Marder (1999), while the discussion of the theoretical
and computational approaches is patterned after Sahimi (1998) and Fineberg and
Marder (1999).

7.10 Instability in Dynamic Fracture of Isotropic

Amorphous Materials
The experiments of Fineberg, Marder, and co-workers (Fineberg et al., 1991, 1992,
1997; Gross et al., 1993; Sharon et al., 1995; Marder and Gross, 1995; Sharon
and Fineberg, 1996, 1998, 1999; Hauch and Marder, 1999) used the conductive
 7.10. Instability in Dynamic Fracture of Isotropic Amorphous Materials 341

strip method described in Section 6.11.5.2 with high resolutions—up to ±5m/s

for the velocities and 0.2 mm for the spatial resolution. They used PMMA and
Homalite-100 and were able to measure the fracture’s velocity at 1/20 µs intervals
for about 104 points throughout the duration of an experiment, which allowed
them to follow the long-time dynamics of a fracture in considerable detail. What
follows is a discussion of their results as well as those of others. These experiments
have helped us understand and resolve a few of the outstanding issues in dynamic
fracture.

7.10.1 The Onset of Velocity Oscillations

Typical data for fracture propagation in PMMA are shown in Figure 7.10. The
fracture was initially at rest. Its tip had ample time to become slightly blunted,
hence making it difficult for the fracture to begin propagating. Note that the crack
first accelerates abruptly, over a very short a time (< 1µs), to a velocity of the or-
der of vc = 100 − 200 m/s, beyond which the dynamics of the fracture is no longer

Figure 7.10. Typical measurements of velocity (in m/s) of a fracture tip as a function of
its length in PMMA. The fracture velocity initially jumps to 150 m/s, and then accelerates
smoothly to the critical velocity vc (dotted line), beyond which strong oscillations set in.
The times are in µsec (after Fineberg and Marder, 1999).
342 7. Brittle Fracture: The Continuum Approach

smooth. Instead, one has rapid oscillations in the fracture’s velocity which increase
in amplitude as v does. On the other hand, Hauch and Marder (1999) carried out
experiments in which the energy available per unit length decreased slowly through
the length of the sample. In both PMMA and Homalite-100, fractures decelerated
gradually to zero velocity, supporting strongly the notion that initial trapping, rather
than any intrinsic dynamical effect, is responsible for the velocity jumps, such as
those in Figure 7.10, which are always seen when fractures begin to propagate.
Indeed, in the case of glass, it is possible to prepare very sharp initial cracks so
that their propagation can begin gradually and then continue steadily at velocities
that are only a small fraction of the Rayleigh wave speed cR .
The next question is whether the velocity oscillations are random fluctuations
or are periodic in time. A careful examination of the oscillations indicate that,
although they are not completely periodic, a well-defined time scale does exist with
a value that, in the case of PMMA, is typically between 2 and 3µs. Moreover, in
experiments in which the fracture accelerates continuously, the location of the peak
of the power spectrum of the data in the frequency domain is constant, although
the velocity varies by as much as 60% of its mean value. As Figure 7.10 also
indicates, there is a critical velocity vc beyond which the fracture velocity begins
to oscillate. Many experimental observations indicate that vc is independent of the
sample geometry and thickness, and the applied stress. The value of the critical
velocity for PMMA is about 0.36cR which, when surpassed, results in oscillations
in the fracture velocity and an increase in the fracture surface area.

7.10.2 Relation Between Surface Structure and Dynamical

Instability
We already described in Sections 7.8.2–7.8.7 the various features that appear in
the structure of a fracture surface. How are these features related to the dynamics
of fracture propagation? Experiments by Fineberg et al. (1997) indicate that the
surface structure appears in the close vicinity of vc . The initial surface structure is
apparent on only a relatively small amount of the fracture surface. To characterize
the amplitude of this structure obtained for PMMA, Fineberg et al. (1992) [see also
Boudet et al. (1995, 1996)] plotted the average height of the points not found in the
mirror-like regions within the fracture surface as a function of the mean velocity
of the fracture. The results are shown in Figure 7.11. This figure indicates that
a well-defined transition occurs where surface structure is created. This happens
when the fracture velocity has reached v = vc = 0.36cR . Moreover, the surface
structure is a well-defined and monotonically increasing function of the mean ve-
locity of the fracture. Finally, both the transition point and functional form of the
graph are independent of such details as the initial and boundary conditions utilized
in the experiment. They are, therefore, intrinsic to the fracture process. Thus, the
existence of a well-defined critical velocity vc for the onset of oscillatory behav-
ior of the fracture and the monotonic dependence of the surface structure created by
 7.10. Instability in Dynamic Fracture of Isotropic Amorphous Materials 343

Figure 7.11. The root mean square values of the surface heights (in µm) as a function of
the mean fracture velocity (in m/s) in PMMA. Different symbols are for various stresses
and sample geometries (after Fineberg et al., 1992).

the fracture for v > vc demonstrate the existence of a dynamical instability in

propagation of fracture beyond vc . The dynamical instability is not influenced by
either the boundary or initial conditions, and is only a function of the mean velocity
of the fracture or, equivalently, the energy release rate, and thus is intrinsic to the
system. Moreover, this dynamical instability is a general feature of brittle fracture.

7.10.3 Mechanism of the Dynamical Instability

Although there is little, if any, doubt about the existence of an intrinsic dynam-
ical instability during fracture propagation in brittle amorphous materials, the
mechanism that gives rise to this instability must be identified, a task that was
accomplished by Sharon et al. (1995). As already discussed above, experiments
indicate that microscopic branches appear within the mist region in a variety of
brittle materials, ranging from PMMA to hardened steels. The morphology of these
branches was analyzed by Sharon et al. as a function of fracture velocity. Their
analysis indicated that below the critical velocity vc no microbranches appear.
They begin to emerge at vc , and as the mean velocity of the fracture increases,
they become both longer and more numerous. Figure 7.12 presents the mean length
of a microbranch as a function of the mean velocity of the fracture, indicating that
this quantity is a smooth and well-defined function of the mean velocity. Moreover,
344 7. Brittle Fracture: The Continuum Approach

Figure 7.12. Mean branch length (in µm) as a a function of the mean fracture velocity
in PMMA. The critical velocity is about 340 m/s (the data are from Sharon and Fineberg,
1996, and Sharon et al., 1996).

similar to Figure 7.11, at v = vc there is a sharp transition from a state which

has no branches to one in which both the main fracture and its daughter cracks
are observed. This feature is independent of the initial state of the material. At
the same time, a single value of vc describes both the transition to formation
of microbranches and the emergence of the surface structure. Indeed, the surface
structure is a result of the crack branching process, and in fact the structure observed
on the fracture surface is, essentially, the initial stage of a microbranch which
subsequently continues in the material in a direction transverse to the fracture
plane.
The microbranching instability is also responsible for the increase in the size
of the velocity fluctuations. As a fracture accelerates, the energy released from
the potential energy stored in the surrounding material is utilized for generating
new fracture surface (i.e., the two new faces created by the fracture). At vc , the
energy flowing into the fracture tip is divided between the main fracture and its
daughters, resulting in less energy for each crack and a decrease in velocity of the
crack ensemble. However, the daughter cracks cannot win their competition with
the main fracture, and thus have a finite lifetime. This is presumably because the
daughter cracks are screened by the main fracture which, due to its straight-line
propagation, outruns them. Thus, after some time the growth of the daughter cracks
stops and the energy that was being diverted from the main fracture now returns
to it, causing it to accelerate again until the scenario repeats itself.
 7.10. Instability in Dynamic Fracture of Isotropic Amorphous Materials 345

7.10.4 Universality of Microbranch Profiles

For a given mean velocity both the lengths and distances between consecutive
microbranches are broadly distributed. Sharon and Fineberg (1996) showed that
in PMMA log-normal distributions characterize these quantities with a mean and
standard deviation that increase linearly with increasing mean fracture velocity.
However, although a given branch may select its length from a broad distribution,
such as a log-normal distribution, all microbranches propagate along a highly
well-defined trajectory. Indeed, Sharon and Fineberg (1996, 1998) found that these
trajectories in both PMMA and glass, when considered at the same mean velocity,
follow a power law of the form
y = 0.2x 0.7 , (146)
where x and y are, respectively, the directions parallel and perpendicular to the
direction of propagation of the main crack, with the origin being the point at which
the microbranch begins. Much earlier, Hull (1970) had obtained the same result for
fracture of polystyrene. These identical trajectories in highly different materials
suggest that the microbranch profiles in brittle materials are universal, caused by
the universal behavior of the stress field surrounding the fracture. Hull had also
proposed that the branch profiles follow the trajectory of maximum tangential
stress of the singular field created at the tip of the main fracture (see also the
numerical calculations of the stress field of a single static fracture by Parleton,
1979). Moreover, recall (see Section 7.8.11) that value of the branching angle
for macroscopic branching in various materials ranges from 11◦ to 15◦ , which
suggest that a smooth transition between microscopic and macroscopic fracture
branches takes place in brittle materials, and that the characteristic features of
fracture branches exhibit a high degree of universality. If this is true, then the
criterion for the formation of macroscopic fracture branches is identical with the
onset of the microbranching instability.

7.10.5 Crossover from Three-Dimensional to Two-Dimensional

Behavior
The next question to be taken up is the following. What are the circumstances
under which a fracture branch survives and continues to propagate away from
the main crack? Sharon and Fineberg (1996) proposed that a necessary condition
for a microbranch to develop into a full-fledged fracture is the coherence of the
microbranch over the entire thickness of the sample material. They showed in
their experiments on PMMA that, near the onset of the instability, the width of
a microbranch is quite small, but as the fracture velocity surpasses the critical
velocity vc , both the branch width and length increase. Sharon and Fineberg (1996)
used two methods to quantify the increase in the coherence width of the branches.
One method was based on a study of the velocity-dependence of the width of the
fracture patterns formed by the branches along the fracture surface, which indicated
that, beginning with fracture velocities that are close to vc , the width of the pattern
346 7. Brittle Fracture: The Continuum Approach

increases sharply with the mean velocity of the crack. When the velocity reaches a
value of about 1.7vc , the pattern of the growing branches becomes coherent across
the entire thickness of the sample. At still higher velocities, macroscopic branching
occurs. This phenomenon represents a crossover from a 3D behavior to a 2D one.
The second method for quantifying the coherence of the microbranches, which
also helps to further quantify the crossover between the 3D and 2D behavior, is
based on measuring the ratio of the total amount of fracture surface produced by the
crack and its branches located at the sample faces, and that produced at the center
of the sample. Sharon and Fineberg (1996) found that the difference in surface
production between the outer and center planes decreases continuously until the
fracture velocity is about 1.65vc , at which the ratio approaches 1, indicating that
microbranch production across the sample is homogeneous. These results are also
supported by the experiments of Boudet et al. (1996) on PMMA that indicated
that both the sound emissions and surface roughness diverge as the mean fracture
velocity approached 1.7vc , hence suggesting that a second transition may occur at
v ∼ 1.7vc . As the divergence of surface roughness is an indication of macroscopic
branching, the crossover from 3D to 2D may be considered to be a sufficient
condition for macroscopic branching to occur.

7.10.6 Energy Dissipation

As we discussed above, the fracture energy  increases sharply with the fracture
velocity. In PMMA, for example (see Figure 7.8), the energy release rate increases
by nearly an order of magnitude as the mean fracture velocity exceeds vc . Since for
v > vc the microbranching instability occurs, the total amount of fracture surface
created by the fracture front must also increase, thereby leading to an increase in
. Sharon and Fineberg (1996) and Sharon et al. (1996) measured, for PMMA
and as a function of the mean crack velocity, the relative surface area, defined as
the ratio of the total area per unit crack width created by both the main fracture
and microbranches, and that which would be formed by a single crack. They also
measured the energy release rate. Their data indicate that the amount of surface area
formed is a linear function of the energy release rate, implying that, both before
and after the onset of the instability, the fracture energy is nearly constant. Thus,
the fracture energy ”increase” shown in Figure 7.11 is entirely a direct result of the
microbranching instability. The rise in the fracture energy is due to the formation
of more surface by the microbranches. The cost of creating a unit fracture surface
remains, however, constant with a value which is close to the fracture energy
immediately preceding the onset of the instability.
In light of these results, the long-standing question of why in isotropic materials
a propagating fracture never seems to approach the Rayleigh wave speed cR can be
answered. A propagating fracture does not have to dissipate increasing amounts of
energy by accelerating, thereby increasing the amount of kinetic energy. Beyond
the critical velocity vc a fracture has the option of dissipating energy by generating
an increased amount of fracture surface at the expense of a reduction in the total
kinetic energy. As the amount of energy to its tip increases, a fracture forms a
 7.11. Models of the Cohesive Zone 347

corresponding amount of surface via microscopic branching, the mean length of

which also increases with increasing the energy flux to the tip. If this energy
increases further, a second generation of microbranches may also form (Sharon
and Fineberg, 1996) which are the daughters of the daughter cracks. The process
of formation of the second, third, · · ·, generation of the microbranches may very
well be the mechanism for the generation of a fractal structure.

7.10.7 Universality of the Dynamical Instability

An important question is whether the microbranching instability is a universal fea-
ture of dynamic fracture, or is limited to certain types of brittle materials. Much
of the experimental data (and also the theoretical work to be discussed later)
indicate that the instability is indeed a general feature of brittle fracture. We al-
ready mentioned that patterns on the fracture surface have been observed within
the mist region in a variety of brittle polymers. In addition to PMMA that was
used by Sharon and Fineberg, microscopic branches have also been observed in
polycarbonate, polystyrene, hardened steels, glass, as well as in brittle polymers.
Additional evidence for this universality is supplied by the fact that, as discussed
in Section 7.10.4, microbranches in glass and PMMA develop nearly identical
trajectories. Moreover, Irwin et al. (1979), Ravi-Chandar and Knauss (1984a,b,c)
and Hauch and Marder (1999) reported that microbranches are initiated in Homa-
lite beyond vc = 0.37cR , which is within 2% of the critical velocity observed in
PMMA, although the critical velocity for glass seems to be slightly higher (Gross
et al., 1993), vc  0.42cR , which is still within 20% of the critical velocity for
PMMA and Homalite. These results all point to the universal nature of the dynam-
ical microbranching instability in a wide variety of materials, which also makes
it possible to describe dynamic fracture of many heterogeneous materials by a
unified theory.

7.11 Models of the Cohesive Zone

Having described the experimental facts that have helped us understand the nature
and characteristics of the microbranching instability in dynamic fracture of amor-
phous materials, we are now in a position to discuss the theoretical developments,
the predictions of many of which agree with the experimental data. An important
task is development of a reasonable model of the cohesive zone. Although there has
been considerable work devoted to modeling of metals’ cohesive zone, our focus
in this chapter is on brittle materials. We describe in this chapter the progress that
has been made based on the continuum models. Chapter 8 will discuss the lattice
models and the insights that they have provided.
As discussed earlier in this chapter, linear continuum fracture mechanics predicts
that, as one approaches the tip of a fracture, the stress field diverges as r −1/2 .
However, a divergent stress field is not tenable in a real material. This has motivated
the development of many models, both simple and complex, of the cohesive zone
348 7. Brittle Fracture: The Continuum Approach

in order to explain how the apparent stress singularity actually joins smoothly a
region around the fracture tip where all the fields are finite.

7.11.1 The Barenblatt–Dugdale Model

One of the simplest models of the cohesive zone was proposed by Barenblatt
(1959a,b) and apparently independently by Dugdale (1960) (see also Langer,
1992). In their model, one assumes that, up to a certain distance L from the tip of
the fracture—the length of the cohesive zone—the faces of the fracture are pulled
together by a uniform stress σc , which then drops abruptly to zero when the sepa-
ration between the surfaces reaches a critical separation of lc , as shown in Figure
7.13. The energy absorbed by the cohesive zone can be determined easily, if the
fracture propagates in a steady state so that the cohesive zone and all the elastic
fields translate in the x-direction without changing their form, since in this case
translating the fracture by a distance x increases the length of the material by
x that has passed through the cohesive zone. The energy cost H for bringing
a length x of the material through the cohesive zone, per unit length along z,
4l
is given by H = x 00 σc dy = xl0 σc . If all the energy that flows into the
fracture tip is dissipated by the cohesive forces, then the energy release rate H
equals l0 σc . The main idea of this model of the cohesive zone is to select l0 and σc
in such a way that the singularities from the linear elastic problem are removed.
Therefore, the condition
H = l 0 σc , (147)
must coincide exactly with the condition for eliminating the stress singularities.
With the aid of Eq. (66), one can then determine the length L of the cohesive

Figure 7.13. Schematics of the cohesive zone model of Barenblatt and Dugdale. The faces
of the fracture are pulled apart by a cohesive stress σc until the faces are separated by a
critical distance lc . The fracture moves from left to right.
 7.11. Models of the Cohesive Zone 349

zone, since this zone can be viewed as a superposition of delta-function stresses

of the type considered in Section 7.3.2, but with tensile stresses σc , rather than the
compressive stresses used there. From Eq. (66), the stress intensity factor is given
by
 0 ! !
KI = − σc 2/π l0 dl0 = −σc 8L/π , (148)
−L

where the negative sign is due to the fact that the cohesive zone is pulling the
fracture faces together and cancelling out the positive stress intensity factor which
is being generated by other forces outside the fracture. Substituting Eq. (148) into
(73) and using Eq. (147) yield
 $ %
1 − νp2 2 8L
H= AI (v) σc = lc σc , (149)
Y π

from which L is determined. Since as discussed earlier, AI (v) diverges as the

fracture velocity v approaches the Rayleigh wave speed cR , the length L of the
cohesive zone must vanish, because the fracture opens more and more steeply as
v increases, and therefore it reaches the critical separation lc sooner and sooner.
This type of cohesive zone is frequently observed in fracture of polymers, since
behind the fracture tip of such materials, there still are polymers that are arrayed
in the craze zone (see Chapter 6) which stretch between the two fracture faces and
pull them together. The cohesive zone in metals, on the other hand, is viewed as a
simple representation of plastic flow around the fracture tip.
Although this model is simple and has had some success in explaining some
aspects of the physics of the fracture energy, as far as explaining the dynamical
microbranching instability that we discussed in the last section is concerned, it is
not useful at all because, in essence, it replaces one phenomenological parameter,
L, by the two phenomenological parameters, σc and l0 , and hence provides no new
information or even a clear sense of how the dissipated energy varies with fracture
velocity.
More realistic models of the cohesive zone have played an important role in
providing a better understanding of the dynamical instabilities in brittle fracture
of materials. Some of these models are first formulated based on the continuum
mechanics, but are then discretized using the finite-element method and simulation,
or utilize analytical or semi-analytical analysis. In what follows we discuss these
models and the implications of their predictions.

7.11.2 Two-Field Continuum Models

An important advance in the continuum formulation of dynamic fracture has been
the development of the so-called two-field models that couple the equation for
elastic deformation of materials to one for the order-parameter of the system. The
concept of an order parameter is borrowed from theories of thermodynamic phase
transitions in which this parameter represents, for example, the difference between
350 7. Brittle Fracture: The Continuum Approach

the densities in the gas and liquid phases that are in equilibrium with each other. This
concept is also well-defined for geometrical models, such as the percolation model
for which the order parameter represents the fraction of the uncut bonds or sites,
where the cut bonds represent the “defects.” Hence, at the percolation threshold pc ,
where the geometrical connectivity of the system is lost due to the presence of too
many defects, the order parameter is zero, slightly above pc is very small, while
far from pc the order parameter is nearly unity, since in this region the defects
are too few. In a similar spirit, the order parameter for dynamic fracture should be
related to the concentration of point defects in the material, hence characterizing
local order. In this formulation, the order parameter is (similar to the percolation
model) unity outside of the propagating fracture, but zero inside the crack where
all the atomic bonds have been broken. On the crack surface, the order parameter
varies continuously between 0 and 1, on length scales that are much larger than the
interatomic distances. This would then justify use of a continuum formulation of
dynamic fracture propagation, in which case one would need an equation for the
order parameter that couples it to the equation for the elastic deformation, hence
the name two-field models. The advantage of formulating the problem in terms of
an order parameter and coupling it to the displacement field is that, by allowing
the order parameter to vary in the cohesive zone, the stress singularity at the tip of
the fracture is avoided, hence removing one main deficiency of continuum fracture
mechanics.
One such two-field model was developed by Aranson et al. (2000). They focused
on 2D materials in Mode I fracture, and represented the elastic deformation of an
amorphous material by the usual wave equation, coupled to a term that represents
viscous damping:
$ %
∂ 2u ∂u
ρ0 2 = η∇ 2 + ∇ · σ, (150)
∂t ∂t
where the first term on the right-hand side accounts for viscous damping with η
being the viscosity, and ρ0 is the material’s density which is taken to be unity. The
stress tensor σ is related as usual to the strain tensor , except that their relation
now contains a term involving the order parameter P. This relation, in component
form, is given by
$ %
Y νp ∂P
σij = ij + I δij + a1 δij , (151)
1 + νp 1 − νp ∂t
where a1 is a constant, I is the trace of the strain tensor, and the rest of the notations
are as before. One must take into account the effect of the material’s weakening by
fracture which reduces the Young’s modulus Y . Therefore, Aranson et al. (2000)
assumed that, Y ∼ Y0 P, where Y0 is the initial Young’s modulus. In Eq. (151) the
term that couples the stress and strain tensors to the order parameter accounts for
the hydrostatic pressure that one must apply to the material in order to generate
new defects. Although one might be tempted to interpret this term as being due
to the material’s thermal expansion, such identification would be erroneous since,
as discussed in Chapter 6 and earlier in this chapter, during fracture propagation
 7.11. Models of the Cohesive Zone 351

the temperature at the tip of the crack will be high and therefore it is unlikely that
the tip will be in thermal equilibrium. Note that for P = 1—the area outside the
fracture surface—one has the usual equations of elasticity, while for P = 0 the
dynamics is trivial as nothing is happening inside the fracture.
The next step is to develop an expression for the order parameter. Aranson et
al. (2000) assumed that P is governed by purely dissipative dynamics. As such,
the order parameter may be derived from an free-energy functional H,
∂P δH
=− , (152)
∂t δP
which is the standard practice in thermodynamics. In the theory of phase transitions
one has (see, for example, Landau and Lifshitz, 1980)
  
H= a2 |∇P|2 + Hp (P) dxdy, (153)

where Hp is a local potential energy that has minima at P = 0 and 1. If we choose

Hp to be a polynomial in P, we arrive at (Aranson et al., 2000)
∂P ∂P ∂ul
= a2 ∇ 2 P − a3 P(1 − P)F (P, I ) + f (P) . (154)
∂t ∂xl ∂t
Therefore, the order parameter is coupled to the displacement field through
Eqs. (151) and (154) and the function F (P, I ), which is subjected to the con-
straint that it must have one zero in the interval 0 < P < 1, so that F (Pc , I ) = 0
for 0 < Pc < 1, and ∂F (Pc , I )/∂P = 0. The simplest functional form for F that
satisfies these constraints is given by F = 1 − (a4 − a5 I )P, where a4 and a5 are
material-dependent constants that can be set to 1 by rescaling of the time, t → a3 t
and the spatial coordinates xi → a1 xi with a12 = a2 /a3 .
The last term on the right-hand side of Eq. (154) couples the order parameter to
the speed du/dt, and represents the localized shrinkage of the fracture caused by
the motion of the material. Aranson et al. stated that the precise form of the function
f (P) is immaterial, and therefore they used a simple form, f = a6 P(1 − P),
where a6 is a dimensionless material constant (taken to be 1). This completes the
formulation of the problem.
However, we must point out that although these functional forms for F and
f facilitate the solution of the problem, they also lead to certain anomalies. For
example, the model predicts that the crack opening depends logarithmically on
the sample size, as opposed to the correct linear dependence. The root of this
anomalous dependence is in the fact that in this model the strain in the bulk of
the material is not fully relieved after passage of the fracture, and hence a more
sophisticated formulation of these functions is necessary. Despite this deficiency,
the model does predict dynamical instability of the type described above which
we now describe.
This model predicts crack branching, with the size of the branches being depen-
dent on the parameters of the materials. The angle of the branches with the main
propagating crack is around 30◦ , but increases with the crack speed. Figure 7.14
352 7. Brittle Fracture: The Continuum Approach

Figure 7.14. Fracture velocity v, normalized by the Rayleigh wave speed cR = 926 m/s in
PMMA, versus dimensionless energy H/Hc . Open circles correspond to stable propagation,
crosses to unstable propagation, while diamonds are experimental data of Sharon et al.
(1996). The inset shows the curvature for unstable propagation at ρ = 0.5, with the arrows
indicating the progression of time (after Aranson et al., 2000).

compares the predictions of the model for the crack velocity v in PMMA with
the experimental data of Sharon et al. (1995, 1996). The crack velocity has been
normalized by the Rayleigh speed cR , and is plotted versus the fracture energy H
normalized by its value at v = 0.2c√R . The parameters used in the simulations were
Y0 = 10, νp = 0.36, and η = 13/ Y0 . For PMMA, the Rayleigh wave speed is
cR = 926 m/sec. The model predicts that, depending on the material’s parameter, a
crack instability develops when its speed varies anywhere from 0.32cR to 0.55cR ,
with the instability manifesting itself as pronounced velocity oscillations, sound
emission from the crack tip and, of course, crack branching, as mentioned above.
The agreement between the predictions and the experimental data shown in Figure
7.14 is quite good, indicating the correctness of the model in having most of the
essential features of dynamic fracture. For a somewhat related model see Karma
et al. (2001).

7.11.3 Finite-Element Simulation

Johnson (1992,1993) and Xu and Needleman (1994) carried out extensive numer-
ical simulations of dynamic fracture in model isotropic elastic materials. Their
simulations, which were based on discretization of the governing equations with
 7.11. Models of the Cohesive Zone 353

the finite-element (FE) method, have the closest correspondence with experiments
in brittle amorphous materials. In particular, similar to the experiments discussed
above, these FE simulations produced frustrated crack branching, oscillations in
fracture velocities, and limiting crack velocities below the Rayleigh wave speed
cR . Let us describe and discuss these successful efforts.
The basis of Johnson’s work was the physical fact that the size of the cohesive
zone is not predetermined, but is adaptive and changes in accordance with the frac-
ture’s behavior. Since the main purpose of the work was to investigate material
weakening and the role of the cohesive zone, accurate modeling of the contin-
uum region outside the zone was not essential, and therefore Johnson assumed
the continuum to be linearly elastic. In addition, the material modeled was highly
idealized in the sense that, no viscoplastic flow or other rate-dependent properties
were incorporated in the cohesive zone. A planar stress model was used, and an
initial crack of length a0 = 0.6h was inserted in the system, where h is the length
of the plane. The material in the vicinity of the crack tip was assumed to have a
large number of sites where nucleation of defects, all being of the same type and
having the same size, occurs. The fractures were driven by loading their faces with
a number of different loads. Depending on the applied load, the FE simulations
produced maximum fracture velocities of 0.29cR , 0.44cR and 0.55cR . Moreover,
the simulations predicted that, at the lowest velocities, a fracture would accelerate
smoothly. As the external loading was increased, multiple attempts at microbranch-
ing were observed and, similar to the experiments discussed above, the length of
the attempted branches increased with the loading. Moreover, the experimental
observations of Ravi-Chandar and Knauss (1984a) (see Section 7.8.12) that the
stress intensity factor is not a unique function of the fracture velocity v, once v
exceeds a certain limit, were also reproduced by these FE simulations. None of
these results was dependent on the various parameters of the simulations.
More extensive FE simulations were carried out by Xu and Needleman (1994),
although their model of the cohesive zone was different from Johnson’s, and was
also much more elaborate. The continuum was characterized by two constitutive
relations; (1) a volumetric constitutive law that related stress and strain, and (2)
a cohesive surface constitutive relation between the tractions and displacement
jumps across a specified set of cohesive surfaces that were interspersed throughout
the continuum. The first constitutive law was that for an isotropic hyperelastic solid:
∂Hs
σ PK = , (155)
∂
where Hs is the strain energy density which is given by
1
Hs =  : C : , (156)
2
where C is the tensor of the elastic moduli. Here,  is the Lagrangian strain, and
σ PK is the so-called second Piola–Kirchhoff stress, given by

σ PK = σ · (F−1 )T , (157)
354 7. Brittle Fracture: The Continuum Approach

where σ is the non-symmetric nominal stress tensor, F is the deformation gradient,

and T denotes the transpose operation. If, relative to a fixed Cartesian coordinate
system, a material point was initially at x0 and in the current position is at x, then,
F = ∂x/∂x0 . In addition,
1 T
= (F · F − U), (158)
2
where U is the identity tensor.
The constitutive law for the cohesive surface was taken to be a phenomenological
mechanical relation between the traction T and displacement jump across the
surface. This constitutive law must be such that, as the cohesive surface separates,
the magnitude of T first increases, reaching a maximum, and then approaches zero
with increasing separation. Xu and Needleman (1994) assumed the constitutive
relation for each cohesive surface to be elastic, so that any dissipation associated
with the separation is neglected, in which case one has
∂φ
T= , (159)
∂
where φ is a potential which in 2D is given by
$ %  $ %  
n 1−q r − q n
φ( ) = φn + φn exp(−n /δn ) 1−r + − q+ exp(−2t /δt2 ) .
δn r −1 r − 1 δn
(160)
Here, n and t are unit vectors that are normal and tangent, respectively, to the
surface at a given point in the reference configuration, n = n · , t = t · ,
q = φt /φn , and r = ∗n /δn , where φn and φt are the work of normal and tangential
separation, respectively, ∗n is the value of n after complete shear separation
with Tn = 0, and δt and δn are characteristic
√ lengths. The two separation works
are given by, φn =eσn δn , and φt = e/2 σt δt , where σn and σt are the cohesive
surface normal strength and tangential strength, respectively, and e= exp(1). All
the physical parameters used in the simulations were made to correspond to an
isotropic elastic material with the properties of PMMA.
To model the fracture tip, Xu and Needleman (1994) used a model of the cohesive
zone similar to what we described in Section 7.11.1 that takes into account both
tensile and shear stresses, and also allows for the creation of new fracture surface
with no additional dissipation added to the system. In order to allow fractures to
branch off the main fracture line, an underlying grid of lines was used on which
material separation was allowed if a critical condition was reached. Therefore,
this type of simulation combines features of FE models with lattice models (see
Chapter 8), but is in some respect more realistic than the lattice models. The
computations were carried out for a center-fractured rectangular block, and plane
strain conditions were assumed to prevail. Since both the volumetric and surface
constitutive relations are elastic, no dissipation mechanism was incorporated into
the model. As a result, the work done by the imposed loading was partitioned into
kinetic energy, strain energy stored in the material volume, and elastic energy stored
 7.11. Models of the Cohesive Zone 355

in the cohesive surfaces. The FE discretization was based on linear displacement

triangular elements that were arranged in a cross-triangle quadrilateral pattern.
The results of these simulations were very much similar to the experiments in
PMMA. Beyond a critical velocity of 0.45cR , fracture velocity oscillations together
with attempted fracture branching were produced. The branching angle was 29◦ ,
which is close to the maximum branching angle of 32◦ that has been obtained in the
experiments. Moreover, when the fracture was constrained to move along a straight
line, it accelerated to velocities close to cR , in agreement with the experiments of
Washabaugh and Knauss (1994); see Section 7.7.1. Hence, these FE computations
produce results that can describe many, but not all, of the instabilities in the fracture
of PMMA observed in experiments and described above, and in this regard are
more successful than most approaches to dynamic fracture.

7.11.4 Fracture Propagation in Three Dimensions

Several investigations have explored the possibility that the instability of fracture
tip arises naturally from a wiggly fracture front that propagates through a het-
erogeneous material. Notable among these investigations are those of Rice and
co-workers (Rice et al., 1994; Perrin and Rice, 1994; Morrissey and Rice, 1998),
Willis and Movchan (1995, 1997), Movchan and Willis (1995) and Ramanathan
and Fisher (1997,1998), which we now discuss.
Rice et al. (1994) studied the stability of a straight-line, half-plane fracture
front propagating dynamically through an unbounded heterogeneous solid. We
provide here some details of their method for studying this problem, as a good
example of the type of effort that such problems require. They considered the
scalar approximation,
∂ 2u
= c2 ∇ 2 u, (161)
∂t 2
where u is a displacement field representing tensile opening or shear slippage, and
c2 = Em /ρ, with Em and ρ being an elastic modulus and density, respectively.
Equation (161) is easily derived by assuming that the material occupies a volume
 with an external surface S on which a load q is applied. If we then form the
Lagrangian L, i.e., the difference between the kinetic and potential energies of the
system,
 $ % $  %
1 ∂u 2 1
L= d − Em |∇u|2 d − qu dS , (162)
 2 ∂t  2 S
and use variational principles, Eq. (161) is obtained. Suppose now that x = (t)
is the growth history of the fracture in the 2D version of the problem in which
a straight line front propagates in the x-direction, that the loadings are such that
the static solution of the problem has a stress intensity factor K0 for any position
of the fracture front, and that, compared to length scales of interest, all loadings
are applied far from the fracture tip. Then, the 2D version of the model equations
become identical to those that govern anti-plane strain in actual elastodynamics.
356 7. Brittle Fracture: The Continuum Approach

Eshelby (1969) derived the following equation for anti-plane solution for arbitrary
fracture propagation:
& ! 
2 K0
u(x, y, t) = Im x − (tr ) + iy , (163)
π Em
where tr = tr (x, y, t) is a retarded time at which a signal arriving at position (x, y)
at time t was launched at the fracture tip, and satisfies the equation, c2 (t − tr )2 =
[x − (tr )]2 + y 2 . The actual analysis is for a finite body prior to the arrival back
at the fracture tip of waves that are reflected from boundaries or from another
fracture tip, in which case the 2D solution very near the tip is given by
& ! 
2 K
u(x, y, t) = Im x − (t) + iαy + higher order terms, (164)
π αEm
! !
where α = 1 − v 2 (t)/c2 = 1 − (d/dt)2 /c2 , and K is the instantaneous stress
intensity factor given by,
!
K = K0 1 − v(t)/c. (165)
The corresponding energy release rate E is then given by
!
H = H0 [1 − v(t)/c]/[1 + v(t)/c], (166)
where H0 = K02 /(2M).
Rice et al. (1994) derived the 3D solution as a linearized perturbation about the
2D solutions for a fracture propagating at a steady speed v0 [hence, (t) = v0 t].
Thus, if we use polar coordinates such that, r exp(iθ ) = x − v0 t + iα0 y, the 2D
solution becomes
& !  &2 K √ $ %
2 K0 0 1
u0 (x, u, t) = Im x − v0 t + iα0 y = r sin θ ,
π α0 E m π α 0 Em 2
(167)
which is consistent with that of actual elastodynamics for anti-plane strain, if we
identify Em and c with the shear modulus and shear wave speed. To develop the 3D
solution, one sets x = (z, t) = v0 t + f (z, t), a first-order expansion in  about
the 2D results corresponding to a straight fracture ( = 0) propagating along the
x axis with a constant velocity v0 . Thus, the shape of the fracture front can deviate
from being straight. The 3D solution is then of the form,
u(x, y, z, t; ) = u0 (x, y, t) + φ(x, y, z, t) + O( 2 ), (168)
where φ(z, y, z, t) = (∂u/∂)=0 . The singular part of the 3D solution must be of
the 2D character, but now relative to the local direction of fracture propagation,
so that for any  we must have
& ! 
2 K(z, t; )
u(x, y, z, t; ) = Im x − (v0 t + f ) cos γ + iα(z, t; )y + · · ·
π Em α(z, t; )
(169)
 7.11. Models of the Cohesive Zone 357
!
where α(z, t; ) = 1 − v 2 (z, t; )/c2 , v(z, t; ) = (v0 + ∂f/∂t) cos γ (z, t; ),
and cos γ (z, t; ) = [1 + (∂f/∂z)2 ]−1/2 , with γ being the angle between the local
normal to the fracture front and the x- axis. It is then easy to show that as r → 0
[i.e., as x → (z, t) and y → 0] one has
& $ %
√ 1 K0 1
lim [φ(x, y, z, t) r] = f (z, t) sin θ , (170)
r→0 2π α0 Em 2

so that φ(x, y, z, t) satisfies the same equation as (161), subject to the stress-
free boundary condition, ∂φ/∂y = 0 at y = 0 if x < v0 t. We also have, by
symmetry, φ = 0 at y = 0 when x > v0 t. In the harmonic case, f (z, t) =
F (k, ω) exp(−ikz + iωt), the solution for φ is written as

φ(x, y, z, t; k, ω) =
&
1 K0
F (k, ω) exp[i(ωt − kz)] exp[−iωv0 (x − v0 t)/α02 c2 ]ψ(x − v0 t, y; k, ω).
2π α0 Em
(171)
Since φ must satisfy Eq. (161), we find that ψ must satisfy the following equation
  $ 2 %
∂2 1 ∂2 ∂ 1 ∂ 1 ∂2
+ 2 2 ψ= + + 2 2 ψ = Q2 ψ, (172)
∂x 2 α0 ∂y ∂r 2 r ∂r r ∂θ

where
 1/2
|k| ω2
Q(k, ω) = 1− 2 , ω2 < α02 k 2 c2 , (173)
α0 α0 k 2 c 2

 1/2
iω α02 k 2 c2
Q(k, ω) = 2 1 − , ω2 > α02 k 2 c2 . (174)
α0 c ω2

Equation (173) corresponds to letting k approach the positive real axis through
Im(k) > 0 and the negative real axis through Im(k) < 0; these approaches are then
taken to be branch-cut portions of the Re(k)-axis where |k| > |ω|/(α0 c). Equation
(174) holds for any direction of approach. Note that the combination α0 c, which
often appears in solutions of fracture propagation problems, has a clear physical
interpretation: It is the speed at which information is transmitted transversely along
the propagating fracture front. That is, two points of the fracture front a distance
z apart do not influence each other before the time delay z/(α0 c).
The solution ψ must satisfy the asymptotic requirement (170) as r → 0.
Any more general fracture perturbation f (z, t) can be represented as a Fourier
superposition, so that
 +∞  +∞
F (k, ω) = f (z, t) exp[−i(ωt − kz)] dzdt. (175)
−∞ −∞
358 7. Brittle Fracture: The Continuum Approach

The general solution for φ(x, y, z, t) for any f (z, t) is then given by
&  +∞  +∞
r K0 sin( 12 θ ) ∂f (z , t  ) c(t − t  ) − v0 (x − v0 t)/α02 c
φ(x, y, z, t) =
2π 3 α0 Em −∞ −∞ ∂t  (x − v0 t)2 /(α02 ) + y 2 + (z − z )2
!
#[c(t − t ) −
(x − v0 t  )2 + y 2 + (z − z )2 ]  
× ! dt dz ,
c (t − t ) − (z − z )2 − y 2 − (x − v0 t  )2
2  2
(176)
where #[ ] is the Heaviside unit-step function. Once ψ is obtained, φ, and hence
the displacement field u, are also obtained.
One can now derive an expression for the stress intensity factor (and hence the
energy release rate). To do this, it is convenient to replace f (z, t) by (z, t) − v0 t
and ∂f (z, t)/∂t by v(z, t) − v0 in all the expressions. To obtain the first order
perturbation to the stress intensity factor at some location ζ along the z-axis, the
crack front (z, t) is written as
(z, t) = v0 t + [(ζ, t) − v0 t] + {(z, t) − (ζ, t)}, (177)
where the [ ] term describes a 2D perturbation, which is solvable exactly to all
orders by Eqs. (163), (165) and (166), while the { } term corresponds to a 3D
perturbation that vanishes at z = ζ for all t. The stress intensity factor at z = ζ ,
due to small deviations from straightness in other fractures,
√ is determined by
applying to solution (176) for φ the operator limr→0 Em 2π r∂/∂y. The result is
given by
!  ! !  7 ! 8
K(z, t) = K0 1 − v0 /c + K0 1 − v(z, t)/c − K0 1 − v0 /c + K0 1 − v0 /c I (z, t) ,

(178)
with
 +∞  t−|z−z |/(α0 c)
1 c(t − t  )[v(z , t  ) − v(z, t  )]
I (z, t) = PV ! dt  dz ,
2π −∞ −∞ (z − z )2 [α0 c(t − t  )]2 − (z − z )2
(179)
with PV denoting the principal value integral, and v(z, t) = ∂(z, t)/∂t being the
local velocity of the propagating fracture. Therefore, the dependence of the stress
intensity factor on the shape of the fracture front and its deviations from being
straight are expressed in terms of I (z, t). The [ ] term of Eq. (178) is actually
exact for arbitrarily large perturbations of v(z, t), but the { } term is exact only to
first order in the deviation v(z, t) − v0 . The choice of v0 is arbitrary so long as it
is in the range of “first-order difference” from v(z, t).
If we examine the expression for I (z, t), we see that when a segment of the frac-
ture front suddenly slows down relative to neighboring locations along the front, a
reduction in K radiates outward from that segment at speed α0 c. Similarly, when
a segment speeds up, an increase of K is radiated. Such elementary slow-down
and speed-up are due to the encounters of the fracture front with regions of higher
or lower resistance to fracture. Rice et al. also (1994) found that when a straight
fracture front approaches a slightly heterogeneous strip which lies parallel to the
fracture tip along an otherwise homogeneous fracture plane, it may be pinned by
asperities after some advancement into the heterogeneous region, if it is propagat-
 7.11. Models of the Cohesive Zone 359

ing with a relatively small velocity. If, however, the velocity is relatively high, the
asperities give way, the fracture front becomes curvy and propagates further into
the bordering homogeneous region, where it recovers a straight-line configuration
through slowly-damped space-time oscillations which, if they are in response to
spatially-periodic heterogeneities, decay as t 1/2 with time. Such a slow decay sug-
gests that the configuration of a straight fracture front may be sensitive to even
small but sustained heterogeneity in the fracture resistance (i.e., in the material).
Using the results of Rice et al. (1994), Perrin and Rice (1994) showed that a frac-
ture propagating through a heterogeneous material, in which the heterogeneities
are represented as randomly-distributed asperities with which the fracture front
interacts continually, will never reach a statistically steady state. Instead, het-
erogeneities in the fracture energy lead to a logarithmic divergence of the root
mean-squares deviations of an initially straight fracture front. In particular, the
variance V of the deviation of propagation velocity from the mean, was found to be
V ∼ log(2α0 v0 t). (180)
More interestingly, if the material is uniform over the remaining part of the fracture
plane, after the encounter with the heterogeneous portion of the material, the prop-
agating fracture becomes asymptotically (i.e., in the limit t → ∞) straight again.
These predictions suggest that perhaps the roughness of a fracture surface may be
the direct result of a continuous roughening of the surface that is driven by small
heterogeneities within the material. More recently, Willis and Movchan (1995)
and Movchan and Willis (1995) computed the coupling of the energy release rate
to random perturbations to the fracture front in the case of planar perturbations
to the crack in Mode I fracture, and in shear loading. Willis and Movchan (1997)
extended the analysis to the perturbations to the stress intensity factors induced by
a small 3D dynamic perturbation of a propagating, nominally planar, fracture.
Ramanathan and Fisher (1997, 1998) calculated the dynamics of planar pertur-
bations to a tensile crack front and found that, in contrast to the case of the scalar
model for which Perrin and Rice (1994) had obtained logarithmic instability of
the crack front, in Mode I fracture weak heterogeneity of the material can lead to
a non-decaying unstable mode that propagates along the fracture front. They pre-
dicted that this propagating mode occurs in materials having ∂/∂v ≤ 0, where a
constant value of  is a marginal case. For ∂/∂v > 0, the propagating mode was
predicted to decay, with the propagation velocity of the new mode being between
0.94cR and cR . These predictions are supported by the numerical simulations of
Mode I fracture in a 3D material with a constant , carried out by Morrissey and
Rice (1998), indicating that the propagating mode is highly localized in space, and
indeed propagates at the predicted velocities.
Ramanathan and Fisher (1997, 1998) and Morrissey and Rice (1998) both
showed that these localized modes lead to linear growth of the root mean-square
deviations of an initially straight fracture with its distance of propagation. They
suggested that this may provide a new mechanism for the roughness produced
by a propagating fracture in materials in which the fracture energy does not in-
crease rapidly with the velocity of a crack. Both the calculations and simulations
360 7. Brittle Fracture: The Continuum Approach

were performed for in-plane disturbances to a fracture front. Disturbances of this

type cannot, of course, generate the out-of-plane roughness typically seen along a
fracture surface.

7.11.5 Failure of Dynamic Models of Cohesive Zone

Langer and collaborators (Barber et al., 1989; Langer, 1992, 1993; Ching, 1994;
Ching et al., 1996a,b,c; Langer and Lobkovsky, 1998) carried out extensive the-
oretical studies of dynamic models of cohesive zone. They defined the cohesive
zone in a manner similar to what was described in Section 7.11, but did not assume
that cracks propagate at a constant rate, or always in a straight line, and therefore
the cohesive zone becomes a dynamical entity which interacts with the fracture in
a complex fashion. Their goal was to understand whether fracture tip instabilities
can be predicted by such models.
In a first set of calculations, Barber et al. (1989), Langer (1992, 1993) and Ching
(1994) studied the dynamics of cracks confined to straight lines, and found that
such cracks always propagate in a stable fashion, which is consistent also with the
predictions of Marder (1991), although there were also tantalizing hints of insta-
bilities. Therefore, Ching, Langer and Nakanishi (1996a,b,c) studied dynamics of
fractures that are allowed to follow curvy, out of plane, paths. In its most elaborate
version, their model allows the fracture to pursue an oscillating path, and the cohe-
sive zone to contain both tensile and shear components. In most, although not all of
these models, fracture propagation is violently unstable to very short-length oscil-
lations of the tip. Their general conclusion is that these cohesive-zone models are
inherently unsatisfactory for use in dynamical studies. They are extremely difficult
mathematically and they seem to be highly sensitive to details that, from a physical
view point, ought to be unimportant. Pathological short-wavelength instabilities
of fractures also emerge from their analysis which have a simple underlying ex-
planation, which is as follows. The logic of the principle of local symmetry (see
Section 7.6.1) states that atomic bonds under the greatest tension must break first,
and therefore cracks loaded in Mode I propagate straight ahead, at least until a
velocity, identified by Yoffe (see Section 7.6.3), is reached when a fracture is pre-
dicted to spontaneously break the symmetry inherent in straight-line propagation.
This logic has been called into question by a very simple calculation, first described
by Rice (1968).
To see this, let us look at the ratio σxx /σyy right on the fracture line. Using
Eqs. (51) and (52), we find that
σxx (β 2 + 1)[1 + 2(α 2 − β 2 ) − 4αβ] 2(β 2 + 1)(α 2 − β 2 )
= = − 1, (181)
σyy 4αβ − (1 + β 2 )2 4αβ − (1 + β 2 )2
which after a Taylor expansion for low velocities v becomes
σxx v 2 (ct4 + cl4 )
=1+ 2 2 + ···, (182)
σyy 2cl ct (cl − ct )(cl + ct )
indicating that σxx /σyy is greater than unity for all v [cl and ct are defined by
Eqs. (24) and (25)]. This result is surprising because it states that, in fact, as soon
 7.12. Brittle-to-Ductile Transition 361

as the fracture begins to propagate, the greatest tensile forces are perpendicular to
its tip and not parallel to it. Therefore, it is difficult to imagine how a fracture can
ever propagate in a straight line.
That Langer and collaborators found their dynamic models of the cohesive zone
to be unsatisfactory may imply that, such models must be replaced by those in
which plastic yielding is distributed across an area, and not restricted to a line.
The two-field continuum models of the type described in Section 7.11.2 represent
progress in this right direction. Another possibility is that calculations of Langer
and co-workers indicate a fundamental failure of the continuum formulation of
the type that they employ, and that the resolution must be sought either at the
atomic or molecular scale (see Chapters 9 and 10), or one should resort to two-
field continuum models that take into account the variations of the order parameter
in the fracture zone.

7.12 Brittle-to-Ductile Transition

The last topic that we would like to briefly discuss is the brittle-to-ductile (BTD)
transition that occurs in materials as the temperature is lowered and the strain rate
is increased. Kelly et al. (1967) and Rice and Thomson (1974) were probably the
first to offer a fundamental perspective on the class of materials that are capable
of this fracture transition. In particular, Rice and Thomson developed a theoretical
criterion for establishing the intrinsic brittle behavior and distinguishing it from
intrinsic ductility. According to their criterion, an atomically sharp fracture gov-
erns the behavior of a material in the absence of any other form of plastic response
in the background, by either (1) nucleating dislocations from its tip, or (2) by prop-
agating in a cleavage mode due to the presence of an energy barrier to the emission
of such dislocations. In the first class are intrinsically-ductile materials which can-
not undergo a fracture transition, whereas the materials in the second group are
usually considered as intrinsically brittle that are capable of making a transition
to ductility. The BTD transition takes place at a characteristic temperature TBTD ,
and one main goal of research in this area has been developing a theory for quan-
titative prediction of this transition temperature. A variety of factors affect TBTD ,
with chief among them being the rate of loading the material. Many experimental
studies (see, for example, Burns and Webb, 1970) indicate that mere nucleation of
some dislocations from the tip of a fracture may not ensure ductile behavior. De-
spite this evidence, the Rice–Thomson mechanism resembles a threshold process,
somewhat similar to the threshold nonlinearities that we have been considering
in this book, that triggers ductility in a class of intrinsically-brittle materials in
which the mobility of the dislocation is relatively high. Examples of such materi-
als include BCC transition metals and most alkali halides. However, completely
satisfactory confirmation of the Rice–Thomson criterion is rare.
Most models that are based on the Rice–Thomson criterion have been developed
based on the assumption that, while background plastic relaxation serves to lower
TBTD , the most important controlling factor of the transition temperature is the
ability of the fracture tip to emit dislocations that can shield the entire fracture front
362 7. Brittle Fracture: The Continuum Approach

and hence trigger extensive plastic deformation before the fracture can propagate
by cleavage. However, Argon (1987) showed that the Rice–Thomson-type models,
in which the activation configuration consists of a fully-developed dislocation line,
greatly over-estimate the energy barriers to nucleation of dislocations. This remains
true even if one considers a modified Rice–Thomson-type model developed by
Cheung et al. (1991) in which fracture tip nonlinearity and tension softening were
incorporated.
On the other hand, consider the response of silicon, and many other similar co-
valent compounds and materials, that have very sluggish dislocation mobility, and
hence are in contrast with high-mobility hypothesis and the nucleation-controlled
response of some materials. In such materials, the transition from brittleness to
toughness is governed by the mobility of groups of dislocations that are away from
the tip of the fracture (see, for example, St. John, 1975; Hirsch et al., 1989; George
and Michot, 1993). It is now well-established for both classes of materials that, the
emission of the dislocations from the tip of a fracture occurs preferentially from
specific sites on the tip, and that, in order to guarantee ductile behavior, the entire
fracture front must be shielded from local break-out of the cleavage fracture from
unprotected parts of the fracture front. Thus, it is now widely believed that the
fundamental BTD transition is governed by the behavior of a cleavage crack.
In addition to the experimental studies mentioned above, theoretical analyses
of fracture behavior of Si, carried out by Rice and Beltz (1994) and Xu et al.
(1995), indicate that the activation configuration of dislocation embryo is a double
kink of dislocation core matter. Thus, one may identify two distinct types of BTD
transitions:
(1) In the BCC transition metals, where barrier to kink mobility along the dislo-
cation are low, the BTD transition is governed by the formation of dislocation
embryos at the fracture tip, which then results in a nucleation-controlled
transition.
(2) By contrast, experimental work (see, for example, Yonenaga and Sumino,
1989) and theoretical modeling (Bulatov et al., 1995) suggest that, in semi-
conductors and compounds the kink mobility is hindered by substantial energy
barrier, hence rendering the BTD transition controlled by dislocation mobility
away from the tip of the fracture.
A complete understanding of the BTD transition can be obtained based on atom-
istic modeling of the formation and outward propagation of the dislocation embryo
at the tip of the fracture. Such atomistic modelings are based on ‘molecular dy-
namics simulation that will be described in Chapters 9 and 10. However, atomistic
models provide quantitative predictions for this phenomenon only if accurate po-
tentials for describing the interatomic interactions are available. Several promising
interatomic potentials have been developed over the past decade or so that will
be described in Chapter 9. Alternatively, one may utilize a multiscale modeling
approach—one that combines continuum modeling for the region away from the
fracture tip with atomistic simulations in the tip region—in order to study this phe-
nomenon. This represents a realistic and powerful approach that is rapidly gaining
 7.12. Brittle-to-Ductile Transition 363

popularity; Chapter 10 will describe this method. So far as the BTD transition is
concerned, Xu et al. (1995) have already developed a multiscale model for study-
ing this phenomenon. They showed that the energetics of the dislocation embryo
formation on inclined slip planes that contain the fracture tip, when compared with
an additional surface production resistance, is quite unfavorable and cannot ex-
plain the known BTD transition temperatures. Xu et al. conjectured that nucleation
may be more favorable on oblique slip planes, or may occur heterogeneously at
the edges of the fracture front. However, we must realize that, although disloca-
tion nucleation on oblique planes has often been suggested as a likely scenario,
approximate analyses that were based on the Rice–Thomson criterion have led to
estimates of TBTD that are several orders of magnitude larger than the experimental
values.
We note that, although experiments have established the ability of disloca-
tion nucleation at the fracture tip for accounting for the exceedingly sharp BTD
transitions in Si and similar materials, Khanta et al. (1994) questioned this well-
understood fact, and instead advocated an approach based on an analogy with
thermal phase transitions. Specifically, they considered, unlike the more tradi-
tional methods described above, the thermally-induced instability of many small
loops in the presence of an applied stress, and proposed that the creation of many
atomic-size loops by thermal activation induces a temperature-dependent coop-
erative screening effect that enhances the subsequent growth of the loops. This
cooperative effect is completely different from the dislocation shielding of frac-
ture tip stress described above. To develop their theory, they extended the concept
of dislocation screening, originally developed by Kosterlitz and Thouless (1973)
in an entirely different context, namely, 2D phase transitions. In the Kosterlitz–
Thouless (KT) theory, the generation of dislocations (which is an unstable process)
is driven by only thermal fluctuations, without the aid of an applied stress. The
KT transition occurs at a temperature close to the melting temperature, which then
gives rise to a dislocation-mediated melting transition (Nelson and Halperin, 1979;
Young, 1979). In the model developed by Khanta et al. (1994), both the external
stress and thermal fluctuations assist the growth of dislocation loops. The model
then predicts the existence of a KT-type instability, but not a phase transition in
the thermodynamic sense, at a temperature well below the melting temperature,
at a stress level that corresponds to the Griffith threshold that is needed for brittle
fracture propagation. This temperature is then identified with TBTD . If the transi-
tion temperature is zero and the applied load is equal to the Griffith threshold, the
model reduces to the Rice–Thomson model described above. Thus, one advantage
of this theory is that it is applicable to systems that are at a finite temperature, in
contrast with the Rice–Thomson model that is strictly valid for zero temperature.
Despite this success, there is not yet convincing evidence for the role of thermal
fluctuations advocated by Khanta et al. (1994). Indeed, the meticulous experi-
ments of George and Michot (1993), who used X-ray direct imaging of the stages
of evolution of the fracture-tip plastic response, starting from nucleation of crack
tip heterogeneities and followed by very rapid spread and multiplication of dislo-
cation length from such sources, demonstrate clearly the vast numbers of degrees
364 7. Brittle Fracture: The Continuum Approach

of freedom available to dislocation for populating the highly-stressed fracture tip,

but do not indicate any significant role for thermal fluctuations.
Finally, we note that there are many morphological aspects of a BTD transition
in polycrystalline materials in which microcracks, nucleation and crack arrest
at grain boundaries become very important, and modulate the actual TBTD . Our
understanding of such processes is still not complete, and therefore this is an active
research area (see, for example, Falk and Langer, 1998; Falk, 1999).

Summary
As stated at the beginning of this chapter, it was believed for a long time that there is
a conceptual problem with the continuum mechanical formulation of brittle fracture
of amorphous materials, as its prediction for the terminal velocity of propagating
fractures, i.e., the Rayleigh wave speed cR , had seemed to be experimentally
unattainable (apart from highly anisotropic materials). However, the discussions
of this chapter should have made it clear that the problem persisted not because of
a fault in the continuum mechanics, but because it had not been properly posed.
The correct question should have been about the nature of energy dissipation near
the fracture tip. However, such a problem was not studied for several decades,
because it had seemed natural to assume that, in a sufficiently brittle material,
energy will be consumed mainly for breaking the atomic bonds and generating
new fracture surface, a process that should depend only weakly on the fracture
velocity. However, by loading fractures in differing fashions, greatly-fluctuating
quantities of energy can be forced into the fracture tip. The tip must then find
some mechanism for dealing with the energy not needed to break a minimum set
of atomic bonds. A small fraction of the remaining energy is consumed by such
minor events as phonon emission, after which the tip begins consuming energy by
a sequence of dynamical instabilities, giving rise to ramified networks of fractures
(or broken atomic bonds) on small length scales.
Thus, there is actually no discrepancy between the conventional continuum
fracture mechanics and the experimental observations and data. In a large enough
amorphous material, the fracture-tip instabilities occur within the cohesive zone
where linear continuum fracture mechanics is not even an appropriate theoretical
framework for analyzing the instabilities, let alone predicting them. The finite-
element simulations, models of fracture propagation in 3D, the two-field continuum
models, the lattice models that will be described in Chapter 8, and many precise and
beautiful experiments carried out over the past decade, have now provided us with
a much better understanding of the structure and dynamics of energy dissipation in
the vicinity of the tip of a propagating fracture in a brittle material. It is now clear
that fracture in brittle materials is governed by a dynamic instability that gives
rise to repeated attempts for branching off of the main propagating fracture, hence
preventing the terminal fracture velocity from reaching the Rayleigh wave speed.
8
Brittle Fracture: The Discrete Approach

8.0 Introduction
As discussed in Chapters 6 and 7, theoretical and computer simulation studies of
fracture of materials are usually based on one of the following three approaches.
(1) The first approach formulates the problem using linear continuum fracture
mechanics. This approach, which was described in detail in Chapter 7, allows
one, in many cases, to derive the analytical solution of the problem of frac-
ture propagation in a given material, subject to certain initial and boundary
conditions. If, however, such analytical solutions cannot be derived, then the
governing equations must be discretized by, for example, a finite-difference
or finite-element method and solved by numerical simulations, in which case
the model reduces to a type of discrete or lattice model.
(2) The second approach is based on molecular dynamics (MD) simulation of
fracture propagation which studies the phenomenon at atomic length scales.
Molecular dynamics is a discrete approach in that, the system under study is
represented by a discrete set of atoms connected to one another by atomic
bonds. This approach will be described in Chapter 9.
(3) The third approach is based on lattice models which can be used for both quasi-
static and dynamic fracture phenomena. However, we must point out that there
is a major difference between lattice models of fracture that we describe and
discuss in this chapter and the MD approach to fracture. The difference is due
to the fact that, in MD simulation of fracture breaking of an atomic bond is a
natural outcome of the simulations, whereas in the lattice models described in
this chapter, how or when a bond breaks is an input of the models that must
be specified at the outset. There are, in general, two types of lattice models.

(i) One class of such models is intended for quasi-static fracture. Such mod-
els consist of a lattice of springs or beams, together with a criterion for
nucleation of local microcracks. In these models, each node of the lattice
is connected to only a finite number of other sites (which are usually the
nearest-neighbor sites), and a force balance is written down for each node,
resulting in a set of simultaneous equations that govern the nodal displace-
ments. Unlike the MD method, the nodes of the lattice do not represent the
material’s atoms, nor do the bonds represent the atomic bonds. Instead,
366 8. Brittle Fracture: The Discrete Approach

the lattice models represent a material at length scales much larger than
the distance between two neighboring atoms in the material, and therefore
one does not have to be concerned about developing accurate interatomic
potentials between the atoms, a subject that will be discussed in detail in
Chapter 9.
(ii) The second class of such models are intended for dynamic fracture. This
class of models is itself divided into two subclasses. (a) In one group are
models that represent generalization of the lattice models of quasi-static
fracture. The nodes of the lattice do not represent atoms. Some of such
models contain quenched (fixed in space) disorder, while others have been
developed for fracture of materials with annealed disorder (i.e., one that
may change with the time). (b) The lattice sites in the second group do
represent atoms. However, instead of assuming interatomic potentials be-
tween the atoms, as in MD simulations, one adopts, in a manner similar
to lattice models of quasi-static fracture, a simple force law between the
atoms, one in which the forces rise linearly up to a critical separation be-
tween the atoms, beyond which they abruptly vanish. If the lattice contains
no disorder, then exact calculations can be carried out (see below).

In essence, most of these models represent generalizations of the lattice models

for linear transport properties of heterogeneous materials (described in detail in
Volume I), and also those for the phenomena of electrical and dielectric breakdown
described in Chapter 5. Aside from the fact that for certain materials, such as fibrous
composites, lattice models are natural, the motivation for developing such models
of brittle fracture is twofold.
(1) In most materials, either manufactured (such as composite solids) or natural
(such as rock), the presence of heterogeneities in the form of either a distri-
bution of microscopic elastic constants, or in terms of flaws or defects with
various sizes, shapes and orientations, makes fracture a very complex phe-
nomenon. Thus, as already pointed out in Chapter 6, the effect of even small
initial disorder can be enormously amplified during fracture, with the result
being the fact that fracture is a collective phenomenon which is controlled by
the disorder. In fact, due to disorder, especially when it is strong, brittle mate-
rials generally exhibit large statistical fluctuations in their fracture strengths,
when nominally identical samples are tested under identical loading. Thus, as
is now well-understood, due to the fluctuations, it is inappropriate to analyze
the phenomena of fracture of a disordered material by a mean-field theory
or an effective-medium approximations. Incorporating the effect of disorder
in a continuum model of dynamic or even quasi-static fracture is, however,
a daunting task, especially when the heterogeneities are broadly distributed.
In addition, such lattice models allow one to investigate, in a convenient and
meaningful manner, various properties of the morphology of the networks of
microcracks that are formed, e.g., those that are formed in rock and rock-like
materials, such as concrete.
 8.0. Introduction 367

(2) Over the past fifteen years there has been considerable theoretical progress
towards understanding the dynamics of elastic manifolds moving through dis-
ordered media, such as charge density waves (see, for example, Narayan and
Fisher, 1992), fluid-surface contact lines (see, for example, Ertas and Kardar,
1992), and interfaces between two phases, such as those that are encoun-
tered in multiphase flow in a disordered porous medium (see, for example,
Sahimi, 1993b, 1995b), all of which exhibit a sort of non-equilibrium critical
phenomenon close to the onset of motion. Fracture of materials does have
similarities with these phenomena (although it has important differences too)
which have provided the impetus for developing some of the models that were
described in Chapter 7, and those that will be described in the present chapter.
In particular, one is interested to understand the extent of the similarities be-
tween these seemingly different phenomena, so that the possibility of a unified
approach to most, if not all, of them can be explored. Moreover, if such sim-
ilarities do exist, then the knowledge that already exists about some of such
phenomena can be immediately “transferred” into new insight about fracture
phenomena.
To make this point clearer, let us go back to Chapter 7 and recall the essentials
of brittle fracture phenomenon. Suppose that there exists a crack front in a material
and that an external load σ is applied to it. If σ is small, there is no steady-state
motion and the crack front is pinned by the heterogeneities of the material in one
of the many locally-stable configurations. As the external load increases, there are
a series of local instabilities that become larger as σ increases further. At a critical
load (stress) σc the crack front depins and begins to move. In a large enough
system, the transition from the stationary to the moving state exhibits features
of a non-equilibrium dynamic critical phenomenon which, to some extent, are
similar to those of second-order phase transitions, such as the percolation transition
emphasized in this book. For example, the mean velocity v of the moving fracture
just above σc obeys the following power law (Ramanathan and Fisher, 1997):
v ∼ (σ − σc )ζ , (1)
where ζ is a critical exponent which is, hopefully, independent of many micro-
scopic properties of the material. Moreover, in the quasi-static case, as σ increases,
segments of the crack front overcome the local toughness caused by the hetero-
geneities and move forward, causing other segments to jump, thereby triggering
an avalanche which will eventually be stopped by tougher regions. It has been
found that, up to a characteristic length ξ − , the avalanches exhibit a power-law
size distribution, where by size we mean roughly the extent l along the crack front
of an avalanche. This size distribution is given by
P (size > l) ∼ l −κ f (l/ξ − ), (2)
where κ is a characteristic critical exponent. The cutoff length scale ξ − itself obeys
the following power law near σc :
−
ξ − ∼ (σc − σ )−ν , (3)
368 8. Brittle Fracture: The Discrete Approach

where ν − is the critical exponent associated with ξ − . Note that the cutoff length
scale ξ − plays a role similar to ξp , the correlation length of percolation which, as
has been emphasized throughout this book, plays a fundamental role in determin-
ing the length scale over which materials with percolation heterogeneity can be
considered as homogeneous. Moreover, we expect that
 σc
l −κ f (l/ξ − ) dσc ∼ l −1 . (4)
0

Just above σc , the fluctuations in the crack velocity are correlated up to a length
scale ξ + which follows another power law given by
+
ξ + ∼ (σ − σc )−ν . (5)
In general, we expect ν − = ν + = ν (see Chapter 3 for examples for which this is
not true). As discussed in Chapters 6 and 7, at the threshold σc the fracture surface
has a self-affine structure with a roughness exponent α, so that the correlation
function C(r) scales as,
C(r) ∼ r 2α . (6)
Finally, the time scale tl that an avalanche of size l lasts is characterized by a
dynamic exponent z, similar to what was defined in Chapter 2:
tl ∼ l z . (7)
Not only are these exponents well-defined, but also satisfy certain scaling relations.
In fact, Ramanathan and Fisher (1997) showed that
ζ = (z − α)ν, ν = (1 − α)−1 , (8)
so that, similar to percolation and other second-order phase transitions, there are
only two independent exponents that characterize this transition. Two-dimensional
(2D) numerical simulations of Ramanathan and Fisher (1997) yielded, z  0.74,
α  0.34, ν  1.52, and ζ  0.34. The estimated α is smaller than the typical
value of the roughness exponent, α  0.8, that, as discussed in Section 7.8.7,
has been reported for several classes of materials. However, MD simulations of
fracture by Nakano et al. (1995), to be described in Chapter 9, indicate that, in
agreement with our discussion in Chapter 7, there may be two regimes of fracture
propagation, characterized by different roughness exponents. Nakano et al. found
that at the initial stages of fracture propagation, when the crack tip moves slowly,
α  0.44, which is reasonably close to the estimate of Ramanathan and Fisher
(1997), while at latter stages when fracture propagation proceeds at relatively high
speeds, α  0.8.
In addition, the lattice models that are described in this chapter have enabled
us to resolve the conflicts between the predictions of linear continuum fracture
mechanics and the experimental observations. In particular, the phenomena of
fracture instabilities, microbranching, and the inability of a propagating fracture
for reaching the Rayleigh wave speed cR (the experimental aspects of which were
 8.1. Quasi-static Fracture of Fibrous Materials 369

described in detail in Chapter 7) have been explained in a satisfactory manner by

such lattice models of dynamic fracture.
We begin this chapter by discussing important aspects of models of fibrous
materials and the predictions that they have provided. We then describe in detail
lattice models of quasi-static brittle fracture, and the considerable insight that they
have provided into the fracture of heterogeneous materials, after which lattice
models of dynamic fracture are described and discussed. As usual and whenever
possible, we compare the predictions of the models with the relevant experimental
observations and data.

8.1 Quasi-static Fracture of Fibrous Materials

As our discussions in Chapter 7 indicated, despite decades of effort, there are very
few exact results for fracture dynamics of disordered materials. Exact analytical
analysis of fracture of any type of material, regardless of whether a discrete model
is used or linear or nonlinear continuum mechanics is employed, is a complex
task. Moreover, quasi-static fracture processes are sensitive to the sample size, but
the approach to their asymptotic (large sample size) behavior is slow. At the same
time, numerical simulation of quasi-static fracture (of the type that is discussed
in this chapter) in very large systems is currently very difficult, if not impossible.
Thus, an exact solution of the fracture problem in any physically viable system
would be very valuable, as it would shed light on a very complex process.
Some of the early work on fracture phenomena concentrated on tensile failure
of continuous-fiber composites using relatively simple models (see, for example,
Daniels, 1945; Coleman, 1958). The reason for this was twofold. One was the wide
applications that such materials have, ranging from paper to glass-fiber mats. In
addition, many composite materials of industrial importance are reinforced by rigid
fibers. The second reason for these early studies was that some of the relatively
simple models developed for such materials, which could provide insight into their
fracture process, are amenable to analytical analysis. Hence, study of fracture of
such materials has remained an active research field (see, for example, Harlow
and Phoenix, 1978, 1991; Smith et al., 1983; Phoenix and Smith, 1983; Curtin,
1991; Phoenix and Raj, 1992; Åström et al., 1994, 2000; Kellomäki et al., 1996;
Räisänen et al., 1997). Some of these studies involved analytical computations
of mechanical and fracture properties of fibrous materials, while others, which
also used more realistic models of such materials, utilized large-scale computer
simulations. We aim to describe the important results that have emerged from such
studies, starting with the analytical results.
One of the rare models for which an exact analysis can be carried out is the
fiber-bundle model, the simplest example of which is shown in Figure 8.1. The
tensile stress is applied vertically. Suppose that p and q = 1 − p are the fractions
of the bonds that are present (unbroken) and absent (broken or failed) in the bun-
dle, respectively, and that each bond is characterized by a failure stress σf . One
can construct a 2D model of such fibers by putting together L of such bundles,
370 8. Brittle Fracture: The Discrete Approach

Figure 8.1. Fiber bundle (top) and chain-of-bundles

model (bottom).

which is also shown in Figure 8.1. The survival probability ps (the probability
that the bundle does not fail macroscopically) is then, (survival probability of a
1D bundle)L .
The model is physically viable only if the applied stress or strain is shared by
the bonds in a meaningful manner, and thus the issue of load sharing is critical. As
discussed by Duxbury and Leath (1994a), there are two classes of such load-sharing
models which we now describe and analyze.

8.1.1 Equal-Load-Sharing (Democratic) Models

In this class of models, also called the democratic models, the load carried previ-
ously by a failed bond is shared equally by all the remaining bonds in the system
(Daniels, 1945; Harlow and Phoenix, 1978). As simple as it may seem, this model
might be applicable to a variety of materials, such as cables or ropes made of
numerous fibers, and even geological faults that are locked by asperity barriers
sharing the total stress. The democratic model of failure of the material is a type
of an effective-medium or a mean-field approximation, and has been used in a
variety of situation, such as modeling of ceramic-matrix continuous-composites.
Because of their mean-field nature, such models can often be solved exactly. Here,
we briefly describe the solution for such models which is due to Sornette (1989).
Consider n independent vertical fibers with identical spring constant κ −1 but
random failure threshold Xi , i = 1, 2, · · · , n. Suppose that the total stress exerted
on this system is σ , and that the strengths X1 , X2 , · · · of the individual links are
independent and randomly distributed variables with the cumulative distribution
P (Xj < x) = F (x). Under a total load σ , a fraction F (σ/n) of the threads will
be submitted to more than their rated strength, and therefore will fail (break)
immediately, after which the total load will be redistributed by the transfer of
stress from the broken links to the unbroken ones, which will then induce secondary
failures, and so on. Thus, one has a cascade of induced failure which we would like
to describe. An important question to be answered is: Does the cascade stop at some
point or propagate until the entire system fails? The answer does, of course, depend
on the way the total stress is redistributed each time a link or bond fails. Although
 8.1. Quasi-static Fracture of Fibrous Materials 371

the democratic model may appear to be difficult but, as pointed out by Sornette
(1989), it can in fact be solved by using the theory of extreme order statistics which
was also used in our discussion of models of electrical and dielectric breakdown
of materials in Chapter 5. The key idea is that, the bundle will not break under an
external load σ if there are k links in it, each of which can withstand a load σ/k. In
other words, if X1;n ≤ X2;n ≤ · · · ≤ Xn;n is the way in which the strengths of the
individual links are ordered, then, if the first k − 1 weakest links fail, the bundle
will resist macroscopic failure under a stress σn ≤ (n − k + 1)Xk;n , because of the
remaining (n − k + 1) links of breaking strength ≥ Xk;n . Therefore, the strength
σn of the bundle is given by

σn = max{(n − k + 1)Xk;n ; 1 ≤ k ≤ n}. (9)

We now search for the strongest subgroup of the bonds. The variables Xk;n are
strongly dependent since they are correlated. However, regardless of the specific
form of F (x), there is a very general result for σn due to Galambos (1978) which
is as follows.

Theorem: Suppose that F (x) is an absolutely continuous function with finite

second moment, and that x[1 − F (x)] has a unique maximum at x = x0 > 0
such that y0 = x0 [1 − F (x0 )]. If F (x) has a positive second derivative in the
neighborhood of x0 , then as n → +∞, one has
 x $ %
√ 1
lim P (σn < ny + x n) = (2π )−1/2 exp − z2 dz, (10)
n→∞ −∞ 2
which is essentially a central-limit theorem. Equation (10) implies that

P (σn = σ ) ∼ (2π nx0 )−1/2 exp[−(σ − ny)2 /2nx02 ]. (11)

Equation (11) states that the density distribution of the global failure threshold
is Gaussian around the maximum σ = ny with a variance that scales as n, hence
implying that the typical strength of the system increases as σn ∼ n, if n is large.
Although by a naive argument one may predict that σn = nx, where x is the
mean one-link threshold, Eq. (11) shows that σn = ny, with y being in fact signif-
icantly smaller than x, and therefore the naive argument greatly overestimates
the global failure threshold.

The mechanical characteristics of the system under a given applied stress σ < σn
depend upon the history of the system, i.e., on the number and the way the links
have failed as the stress was increased from zero to σ . With each value of σ < σn
we associate an integer m(σ ) with 1 ≤ m(σ ) ≤ n such that

[n − m(σ ) + 2]Xm−1;n ≤ σ ≤ [n − m(σ ) + 1]Xm;n , (12)

which can be rearranged to

{1 − [m(σ ) − 2]/n}Xm−1;n ≤ σ/n ≤ {1 − [m(σ ) − 1]/n}Xm;n . (13)

372 8. Brittle Fracture: The Discrete Approach

Note that m(σ ) − 2 is the number of links which have failed under a stress ≤
σ/[n − m(σ ) + 2]. Moreover, by definition of F (x),
(m − 2)/n ≤ F [σ/(n − m + 2)] ≤ (m − 1)/n, (14)
which follows from the fact that, for large n, counting the number of links with fail-
ure threshold less than σ/[n − m(σ ) + 2] amounts to computing the cumulative
failure distribution F (x) at x = σ/[n − m(σ ) + 2]. Relations (13) and (14) indi-
cate, roughly speaking, that, as n → ∞, σ/n is increasingly better approximated
by x[1 − F (x)] with
σ
= x(σ ){1 − F [x(σ )]}. (15)
n
Note that Eq. (15), in the limit n → ∞, is a continuous function. It is then not
difficult to show that, for large n, the number of links which have failed under σ
is given by
k(σ ) = nF [x(σ )]. (16)
For large but finite n, σ (x) or x(σ ) is a staircase with plateaux of width decreasing
to zero as n → ∞. The width of each plateau, for a given σ , can be obtained
from (13), since the interval in σ is such that (13) holds with the same integer
m(σ ) = m.
Just before complete failure of the bundle, the total number of failed links is
given by
kn = k(σn ) = nF (x0 ), (17)
implying that a finite fraction of the links fail before global rupture occurs. If
we consider, for example, the (cumulative) Weibull distribution (WD) (see also
Chapter 5),
F (x) = 1 − exp[−(x/λ)m ], (18)
where λ and m are the parameters of the distribution, then
kn
= 1 − exp(−1/m), (19)
n
which for m = 2 yields kn /n = 0.393. For σ ≤ σn , x(σ ) is in neighborhood of x0
and may be expressed as
x(σ ) = x0 − A(y − σ/n)1/2 , (20)
where A is a constant with a value that depends on the shape of F (x). For example,
for the WD, A = [x0 exp(1/m)/m]1/2 . Then, the number of links that have failed
under the stress σ is given by
k(σ )
= F (x0 ) − B(y − σ/n)1/2 , (21)
n
where B is another constant. For example, for the WD, B is given by B =
(mx0 )−1/2 exp(−1/2m). Equation (21) indicates that k increases rapidly as σ →
σn , approaching nF (x0 ) with a square-root singularity.
 8.1. Quasi-static Fracture of Fibrous Materials 373

We can thus predict the strain-stress characteristics of the bundle of the fibers.
Suppose that each individual link is made of a brittle material, so that its strain-
stress relation is given by, l = κσl up to its failure point, where l is the strain.
Then,
(1) for σ ≤ σ1 , where σ1 is the strength of the weakest link (the first to fail), all
links are intact and the system has a linear stress-strain characteristic with
slope κ −1 . Note that for the WD, σ1 ∼ λn−1/m .
(2) For σ1 ≤ σ ≤ σn , some of the links have failed, and the system is elastic but
nonlinear, which can be established by the following argument. We see from
Eq. (16) that n{1 − F [x(σ )]} links support the total external stress σ , which
means that the stress per remaining link is given by
σ
σr = = x(σ ). (22)
n{1 − F [x(σ )]}
Thus, for every σr there is a corresponding strain per link r , which is equal
to the strain of the entire bundle of links associated in parallel, and is given by

r = κx(σ ), (23)

and therefore we have a strain-stress characteristic which becomes flat with

zero slope as the global failure threshold is approached, σ → σn . Hence, the
effective elastic modulus of the system decreases as σ increases. This nonlin-
earity is due to the fact that as σ → σn , more and more links fail and therefore
the total external stress is transferred to fewer and fewer links. The stress trans-
fer is of course a nonlinear process. The nonlinear behavior of the system is
characteristic of an irreversible process, with the irreversibility in the present
problem being the deterioration of the bundle as σ → σn .
Note that the failure transition in the democratic model is abrupt and hence it
represents a first-order phase transition. There is a rapid increase in the number
of the failed links as the global failure point is approached. If we assume F (x) to
be a WD with m = 2, then the value of F at the failure threshold is F = 0.168,
implying that, before the global failure threshold, few precursory failures have
taken place. Thus, in a sense, the system fails without any “warning.”

8.1.2 Local-Load-Sharing Models

In this class of models the stress carried previously by a failed bond is shared
locally by the remaining bonds in its vicinity, which is of course what happens in
most real materials. Suppose that the total number of bonds in a bundle, the sum
of the intact and failed ones, is L. Since a defect or vacant cluster grows as bonds
at its ends fail, catastrophic failure occurs as soon as a bond fails. Therefore, all
one must do is finding the bond that suffers the largest stress enhancement, and
adjusting the external stress until this bond fails. The adjusted stress is then the
fracture stress σf of the bundle as a whole. In practice, this is easier said than
374 8. Brittle Fracture: The Discrete Approach

done, because failure depends on the largest vacant cluster the statistics of which
are difficult to analyze.
An elegant solution of this problem was developed by Duxbury and Leath
(1994a) (for the solution of the problem in which the stress carried previously
by a failed fiber is shared by its nearest and next-nearest neighbors, see Phoenix
and Beyerlein, 2000). We present a brief description of their solution. With the
cluster-end-load-sharing rule, the bond which suffers the largest stress enhance-
ment is one at the end of the largest cluster of the absent bonds. Under this scenario
then, the survival probability is related to the probability PL (n) that there is no
cluster of vacant bonds of size greater than some prescribed value n. An important
load sharing rule is that, σt = σ (1 + 12 n), where σt is the stress at the tip of the
failed bond. Duxbury and Leath (1994a) calculated PL (n) following a method
proposed by Harlow (1991) in which one identifies the possible endings of a fiber
bundle of length L + 1, and the way by which these endings may be generated from
a bundle of length L. In essence, this method is similar to the transfer-matrix tech-
nique described in Section 5.14.2 of Volume I. Suppose that {1} stands for a present
(unbroken) bond and {0} for an absent (failed) one. If the size of the vacant sites
is restricted to be n, then the bundle endings that are allowed are (1), (10), (100),
(1000· · ·), where the number of zeros in the last probability is n. One now constructs
a transition probability matrix for going from each of these possible configurations
at the end of a bundle of length L to the same endings in a bundle of length L + 1,
by considering the probability of their occurrence. For example, the probability of
going from ending (1) to ending (10) is q, since the probability that the next bond
added is vacant is just q. We define PTL = [p(1) , p(10) , p(100) , · · · , p(100···0) ] as the
probability vector of having the set of possible endings on a fiber bundle of length
L. Then PL+1 is obtained from MPL = PL+1 = ML P1 , where
⎡ ⎤
p p p ··· p
⎢ q 0 0 ··· 0 ⎥
⎢ ⎥
M=⎢ ⎢ 0 q 0 ··· 0 ⎥
⎥ (24)
⎣ · · · 0 0 ⎦
0 0 ··· q 0
is called the transition matrix. Then, the probability PL (n) that there are no vacant
clusters of size larger than n is found from


PL (n) = (pl )L . (25)
l

One may use a variety of boundary conditions, the simplest of which is perhaps
the periodic conditions which require that the first and the last site of the bundle
to be equivalent, in which case

PL (n) = tr(ML ), (26)

where tr denotes the trace of the matrix. Thus, all one must do is studying the
eigenvalues of M. Let a1 = p/λ and a2 = q/λ, where λ is the eigenvalue of M.
 8.1. Quasi-static Fracture of Fibrous Materials 375

If we define a determinant Dn by
. .
. a1 − 1 a1 ··· a1 a1 .
. .
. a2 −1 0 ··· 0 .
. .
Dn = .. 0 ··· .
. (27)
. ··· ··· a2 −1 0 ..
.
. 0 ··· 0 a2 −1 .
then
Dn = −Dn−1 + (−1)n a1 a2n , (28)
with D0 = a1 − 1. The solution to the recursion relation (28) is
(−1)n Dn = a1 − 1 + a1 a2 + a1 a22 + · · · + a1 a2n = 0. (29)
It is then easy to see that
λn+2 − λn+1 + pq n+1 = 0. (30)
Because M is non-negative, then according to the Perron–Frobenious theorem (see,
for example, Noble and Daniel, 1977) its largest eigenvalue λ is real and unique.
Moreover, it is not difficult to see that λ → 1 as n becomes large. Therefore,
setting λ = 1 − δ, Eq. (30) yields
λ  1 − pq n+1 + O(q 2n ). (31)
and hence for periodic boundary conditions
PL (n) = tr(ML ) = λL
1 + λ2 + · · · + λn  λ + O(|λs | ),
L L L L
(32)
where λs is the second largest eigenvalue of M. We thus obtain
PL (n) = [1 − pq n+1 + O(q 2n )]L + O(|λs |L ). (33)
This result agrees with what Duxbury et al. (1986) derived for the electrical break-
down problem discussed in Section 5.2.5. We can now find the failure probability
pf when a stress σ is applied to the bundle by noting that, since failure of the bond
that carries the largest stress causes catastrophic failure, we must have
σf 1
pf = = 1 + n, (34)
σ 2
where σf is the failure stress. Therefore, the probability ps that the fiber bundle
will survive is
 L
ps (σ ) = 1 − pq 2σf /σ −1 . (35)

If L and n are large, then Eq. (35) is essentially equivalent to a double exponential
form, also called a Gumbel distribution, a result that was also obtained for electrical
and dielectric breakdown phenomena described in Section 5.2.5.
A more complex situation arises when an intact bond is between two clusters
of vacant bonds, in which case the bond suffers a large stress enhancement. Thus,
for a more complete analysis one must also consider this situation. The same
376 8. Brittle Fracture: The Discrete Approach

technique that was described above can be used to analyze this case, except that
some modifications must be made. For example, the distinct endings that must
be considered are (11), (110), (1100),· · ·,(110· · ·0); (101), (1010), (10100),· · ·,
(10100· · ·0); (1001), (10010), · · ·, and (10· · ·0), each of which occurs with a certain
probability analogous to p(10) , p(100) , and so on. Duxbury and Leath (1994a) then
showed that these more complex configurations do not change the essence of their
analysis described above. After some algebra one obtains
7 8L
PL (n)  1 − [(n + 1)p 2 − pq]q n+1 + O(q 3n/2 ) , (36)

and the probability of survival is given by

 $ % L
2σf 2 2σf /σ −1
ps (σ ) = 1 − p −p q . (37)
σ
Observe that, compared to (33) and (35), only some prefactors are different in (37).
The average strength of the fiber bundle can then be calculated as

σ 
2[PL (n) − PL (n − 1)]

L−1
=
σf n+2
n=0
(38)
2PL (n) L2 pq L+1
2PL (n − 1)
L−1
= − + ,
L+1 (L + 1)(L + 2) (n + 1)(n + 2)
n=1

where the second term on the right side of the second equation represents a cor-
rection term for preventing (38) from having unphysical behavior as L becomes
large.
In two other papers, Duxbury and Leath (1994b) and Leath and Duxbury (1994)
developed interesting recursion relations for calculating the failure probability and
average strength of the fiber-bundle model, so that one can numerically study the
behavior of the model [for a different approach, based on calculating the Green
functions, see Zhou and Curtin (1995); for a Green function analysis of fracture
in more general systems see also Zhou et al. (1993)]. As usual, suppose that
{1} denotes an intact (unbroken) bond and {0} a failed one. Then for L = 2 the
surviving configurations are {11, 10, 01}, while for arbitrary L there are 2L − 1
surviving configurations and one failure configuration {0 · · · 00}. The probability
4 (1+n/2)
psn that a bond with n failed neighbors survives is psn = 1 − 0 q(x)dx,
where q(x) is the differential failure probability of a bond. Duxbury and Leath
(1994b) separated the full set of 2L − 1 survival configurations into judiciously
selected subsets. Suppose that a lone surviving fiber is surrounded by failed fibers,
and let {A} be the set of all survival configurations which contain only failed
fibers, and lone fibers, and which are bracketed at both ends by lone fibers. Some
of such configurations are {101, 1001, 10001, 1010, · · ·}. From {A} construct {B},
the set of the configurations one specified end of which must be failed. The failed
configuration at the end can be on the left or the right end, but no distinction is
made between them. A third set {C} is also constructed out of {A} in which both
 8.1. Quasi-static Fracture of Fibrous Materials 377

ends of a configuration have failed, e.g., {010, 0100, · · ·}. Finally, suppose that
{P } is the set of configurations with no failed bond, e.g., {1, 11, 111, · · ·}. One
then defines generating functions
∞

∞

∞


A(z) = AL z L , B(z) = BL z L , C(z) = CL z L , (39)
L=3 L=2 L=3
where AL , BL , and CL are the sums, respectively, of the survival probabilities of
the sets {A}, {B}, and {C} for a fixed L. Likewise, a generating function for {P }
is also defined

∞
1
P (z) = (ps0 )L zL = , (40)
1 − ps0 z
L=0
where ps0 is the probability that a bond with no failed neighbors survives. Leath
and Duxbury (1994)
 showed that the generating function for the survival configura-
tions, S(z) = L psL zL , is given by (psL is the survival probability for a fixed L)
P (z)[1 + B(z)]2
S(z) = C(z) + . (41)
1 − P (z)A(z)
Since pf L = 1 − psL , where pf L is the failure probability for a fixed L, then
1
f (z) = − S(z) (42)
1−z

where f (z) = L pf L zL , with pf 0 = 0 and psL = 1. We thus obtain
(1 − z)[1 + B(z)]2 − [1 − ps0 z − A(z)]{1 − (1 − z)[f (z) + C(z)]} = 0.
(43)
Expanding identity (43) in powers of zL and setting the coefficient of the zL term
to zero, one finds the following recurrence relation
XL = XL−1 + ps0 DL−1 X − 2DL B − AL + pf 1 AL−1 − B2 BL−2


L−4 (44)
+ (Ai+2 DL−i−2 X − Bi+1 DL−i−1 B),
i=1
in which XL = pf L + CL , and DL Y = YL − YL−1 . Thus one needs AL , BL , and
CL to use recursion relation (44). These are found by defining new subsets {aL,l },
{bL,l }, and {cL,l }, where, e.g., {cL,l } is the set of survival configurations of length
L which end with exactly l failed bonds. Recursion relations are also found for
these new quantities. For example, aL,l = bL−1,l psl , and


L−l−2
bL,l = pf l psl δL−l−1 + bL−l−1,i ps,L+i pf l . (45)
i=1
These recursion relations can then be used efficiently for calculating various quan-
tities of interest. Because of their efficiency, the behavior of the system for large
L, of the order of several thousands, can be studied.
378 8. Brittle Fracture: The Discrete Approach

Figure 8.2. Dependence of the failure probability of the chain-of-bundles model on the
linear size L of the system (after Duxbury and Leath, 1994b).

An interesting and unexpected result of these calculations is that, the failure

probability possesses a deep minimum with respect to L. Figure 8.2 presents a
sample of the results (Duxbury and Leath, 1994b). For a large applied stress, the
failure probability increases monotonically with L. However, if the applied stress
is small, then the failure probability possesses a deep minimum at an optimal size
Lo , hence pointing to the intriguing possibility of designing fibrous materials that
operate near their minimum failure probability.
A similar, but simpler, exact recursive method was developed by Wu and Leath
(1999). They considered a bundle of parallel fibers in which the local fiber strengths
were distributed according to a statistical distribution f (σ ). Periodic boundary
conditions were imposed on the system. Their analysis indicated that there is a
critical size nc (measured in units of the number of fibers) at which there is a
transition from a tough material to a brittle-like one. More specifically, one has
one of the three following scenarios.
(1) If the size n of the system is less than nc , then the material is in the tough region
which is characterized by very small stresses and small system sizes. The
probability of failure of the material is a superposition of a very large number of
local distributions f (σ ). Since the failure of the material is path dependent, the
number of such local distributions can be as large as 2n−1 (n!). In this case, if the
statistical distribution of the local strengths is given by a Weibull distribution,
Eq. (18), then the cumulative failure probability Fn (σ ) = 1 − Pn (σ ) is given
by
 
Fn (σ ) = 1 − exp −(n!)γ (m)m σ mn , (46)
 8.1. Quasi-static Fracture of Fibrous Materials 379

where 0 < γ (m) < 1 is a parameter that depends on m. Equation (46) has the
general form of a Weibull distribution. Thus, the optimal sample size nmin that
corresponds to the minimum failure probability is obtained from Fn−1 = Fn ,
yielding
nmin ∼ σ −1/γ . (47)
(2) If n  nc  1, then the material is in the brittle region, where it is macro-
scopically brittle but microscopically tough. Roughly speaking, the failure of
the material depends on whether the size of the weakest region exceeds nc .
Since, as discussed in Chapter 5, the probability of finding a weak region of
size larger than nc decays exponentially (because in this case the statistics
of the weak or failed regions is described well by percolation statistics), the
cumulative failure probability is of the Gumbel type:
2 3
Fn (σ ) = 1 − exp −an exp(b ln σ/σ m ) , (48)
where a, b and m are fitting parameters. The size dependence of the mean
failure stress σf  can then be obtained by neglecting the slow-varying factor
ln σ and taking the median as the average, which then yield
σf  ∼ (ln n)−1/m . (49)
(3) For nc ∼ O(1) the material is in the super-brittle regime. This situation arises
when the applied stress is so large that the critical nuclei exist almost ev-
erywhere, and thus almost all the fibers fail simultaneously. The cumulative
failure probability is then simply
Fn (σ ) = 1 − [1 − f (σ )]n , (50)
where f (σ ) is the local strength distribution.
For related work on this problem see Wu and Leath (2000) and Kun et al. (2000).

8.1.3 Computer Simulation

Simulation of more realistic models of fiber networks (with interconnected fibers)
have also been undertaken by, for example, Åström et al. (1994, 2000) who used
a realistic model in which the fibers were linearly elastic beams, described in
Section 8.13 of Volume I, up to a threshold to be defined below, so that the fibrous
material can be considered as being brittle. Consider, as an example, a 2D system
of such fibers, each of which has a length lf . The network is constructed within a
rectangular surface of size Lx × Ly . The (x, y) coordinates of the fibers’ centers
are selected from uniform distributions in the intervals [−lf , Lx + lf ] and [0, Ly ],
respectively, while their orientations are chosen from a uniform distribution in the
interval [−π/2, π/2]. The cross sections of the beams are assumed to be squares
of width w with w  lf . The beams can be stretched and bent and are made of
a material with a Young’s modulus Yf . Two crossing fibers are rigidly bounded
together at their intersection, meaning that all the elastic energies are stored in the
380 8. Brittle Fracture: The Discrete Approach

Figure 8.3. A typical realization of a 2D model of a fibrous material with randomly

distributed and intersecting fibers.

beams and not at the intersections, and that when the network is deformed, the
angle between the crossing fibers will remain constant. Each fiber-fiber bond has
three degrees of freedom: Horizontal and vertical displacements, and rotations.
An example of a typical realization of such a model is shown in Figure 8.3. Two
distinct cases can be considered. (1) The beams are embedded by a background
material with specific elastic properties, as in, for example, a sheet of paper. (2)
Alternatively, the system consists of a network of the beams alone, as in, for
example, a polymer network.
The elastic properties of the model depend on the aspect ratio w/ lf , as well as
the density p of the fibers, defined as the average total length of fibers in an area
of lf2 . The percolation threshold, or the critical density of the fibers, is given by
pc  5.71lf . (51)
Each fiber contains a segment of length ls which is that part of the fiber that is
between the two intersections that the fiber has with two other fibers. Clearly, the
length of the segments is a random variable, as the fibers are distributed randomly
in the system. The average segment length is given by
ls  π pc
=  . (52)
lf 11.42(p/pc ) 3.6p
The elastic interaction between two connected bonds is characterized by a stiff-
ness matrix C. If the moment of inertia of the cross section is M = w4 /12, then
 8.1. Quasi-static Fracture of Fibrous Materials 381

the stiffness matrix for w  ls is given by

⎛ ⎞
Yf w2 / ls 0 0 −Yf w2 / ls 0 0
⎜ ⎟
⎜ 0 12Yf M/ ls3 6Yf M/ ls2 0 −12Yf M/ ls3 6Yf M/ ls2 ⎟
⎜ ⎟
⎜ ⎟
⎜ 0 6Yf M/ ls2 4Yf M/ ls 0 −6Yf M/ ls2 2Yf M/ ls ⎟
C=⎜
⎜
⎟
⎟
⎜−Yf w 2 / ls 0 0 Yf w 2 / ls 0 0 ⎟
⎜ ⎟
⎜ 0 −12Yf M/ ls3 −6Yf M/ ls2 0 12Yf M/ ls3 −6Yf M/ ls ⎟
2
⎝ ⎠
0 6Yf M/ ls2 2Yf M/ ls 0 −6Yf M/ ls2 4Yf M/ ls

(53)
The forces acting on the bonds at the segment ends are obtained by multiplying C
by the vector (ux1 , uy1 , ϕ1 , ux2 , uy2 , ϕ2 ), where u = (ux , uy ) is the displacement
vector and ϕ = (ϕ1 , ϕ2 ) is the rotation vector. If ls is short, the bending stiffness
12Yf M/ ls3 = Yf w 4 / ls3 should, as a first approximation, be replaced by the shear
modulus Yf w 2 /[2(1 + νp )ls ], where νp is the Poisson’s ratio of the material.
The fiber network is deformed by, for example, stretching it uniformly in the
x-direction, which means, for example, fixing the edge at x = 0 and pulling in the
positive x-direction the edge which is initially at x = Lx , with the fibers cross-
ing these edges rigidly tied to them. Periodic boundary condition is used in the
y-direction. Computations of the system deformation, when there is a background
matrix, is not straightforward. Typically, a finite-element method, of the type de-
scribed in Section 7.11.3, is used. Several commercial computer programs that
are capable of performing such computations are available. If the system consists
only of the fiber network (with no background material), then the computations
proceed in the same manner that was described for elastic percolation networks
(see Chapter 8 of Volume I). If the fiber density is too low, the system is not rigid
and the elastic stiffness is zero.
To study brittle fracture of the material, a failure criterion must be defined.
Although such criteria will be described in the next section where we discuss more
general discrete models of brittle fracture, we mention a few of them here. One
can, for example, consider a fiber as broken or failed if the axial tension or bending
of its corresponding beam exceeds a pre-set threshold. Alternatively, failure of a
fiber can be defined based on the shear-lag strain, defined as the magnitude of
the jump in the axial strain on a fiber across a bond. A combination of all such
criteria can also be considered, and in fact Åström et al. (1994) studied the case
in which fracture occurred by segment breaking due to axial tension and failure
at a critical value of shear-lag strain. Once the failure criterion is set, the fracture
simulations begin. Each time a fiber fails, the stress and strain distributions in the
network must be recomputed, as the network’s configuration changes dynamically.
As such, the computations are very intensive. Computer simulations indicated that,
the failure of the system at the initial stages occurs more or less randomly, and
thus the fracture process is similar to percolation. Figure 8.4 shows the stress-strain
diagram of the system when relatively few fibers have failed. As expected, up to a
certain strain, the stress-strain relation is perfectly linear which is what is expected
382 8. Brittle Fracture: The Discrete Approach

Figure 8.4. Stress-strain curve of the

fiber network shown in Figure 8.3.
The numbers refer to the iterations
(after Åström et al., 1994).

of brittle materials. Beyond the critical strain, however, the stress shows a generally
downward trend with increasing strains, accompanied by fluctuations that are the
result of having fibers failing at essentially random locations in the system.
However, as the number of the microcracks increases, the facture zone becomes
quasi-1D, populated mainly by such microcracks with no dominating fracture that
can propagate. The absence of a dominating fracture is presumably because of
the random orientations of the fibers that help distribute the applied stress in the
network more evenly than in a regular network where a dominating fracture usually
forms (see below). This behavior is also different from what is usually observed
during fracture of composite, but non-fibrous, materials described in Chapters 6
and 7.

8.1.4 Mean-Field and Effective-Medium Approximations

One may develop a mean-field or an effective-medium approximation for estimat-
ing the elastic and fracture properties of such fiber networks. The oldest of such
approximations for fiber networks appears to have been developed by Cox (1952).
The development of this type of approximation parallels those previously discussed
for linear conductivity and elastic moduli of materials described in Chapters 4–8
of Volume I, which we now describe.
Consider first the simplest possible approximation, which we refer to as the
EMA1. Suppose that a fiber is attached to an effective medium—a uniform sheet—
under tensile strain and is stretched via a number of links; see Figure 8.5. If the
fiber is uniformly stretched along with the sheet, the stress σf along it would be
constant. However, this is not possible as σf must vanish at the fiber’s ends. If the
strain is small, we can write down the following equation for σf :
dσf u f − us
ls  =c , (54)
dx ls 
where uf and us are the local displacements of the fiber and the sheet, respectively,
and c is a parameter that depends on w/ lf and p/pc . Since σ = Yf , where  is
 8.1. Quasi-static Fracture of Fibrous Materials 383

Figure 8.5. Schematics of EMA computation of stress along a fiber (after Åström et al.,
1994).

the strain, it can easily be verified that the following equation due to Cox (1952),
 
cosh[k( 12 − x/ lf )]
σf (x, k) = Yf x 1 − , (55)
cosh( 12 k)
which he derived by a mean-field approximation, satisfies Eq. (54) and √ the bound-
ary conditions that the stress vanishes at x = 0 and x = lf ; here, k = clf /ls .
The strain x that appears in Eq. (55) is in fact the strain f in the fiber, but use of
x indicates that the strain lies in the x-direction. Note that the average shear-lag
stress is simply ls dσf /dx ∼ kls / lf .
So far, we have assumed that the single fiber embedded in the effective medium is
aligned with the direction of the external strain along which the sheet is stretched
(see Figure 8.5). In general, however, the fibers are distributed randomly, and
therefore one must obtain the orientation dependence of σf . This is, however,
straightforward since, in the absence of transverse Poisson contraction, a rotation
by an infinitesimal field σf yields
σf (x, k, θ) = σf (x, k) cos2 θ, (56)
with θ being the angle of the fiber with respect to the direction of the external strain.
Figure 8.6 compares the predictions of Eqs. (55) and (56) with the simulation results
(Åström et al., 1994) in which k has been treated as an adjustable parameter. It is
clear that the predictions agree well with the simulation results. These simulations
also indicate that k = p(1 + aw/ lf )/pc , where a is a constant.
However, the foregoing treatment is not without problems, especially if it is
further developed in order to predict the elastic stiffness of the network, because
it actually makes the segment stresses correlated along the fibers with reduced
stress close to the fiber ends. A refined treatment of the problem, which we refer
to it as the EMA2, can be developed (Åström et al., 2000) if one combines the
384 8. Brittle Fracture: The Discrete Approach

Figure 8.6. Average distribution of axial stress along fiber for p/pc = 4 and lf /w = 18.8
(after Åström et al., 1994).

probability distribution for the segments’ length with the assumption that the fiber
segments deform only in the energetically most-favorable mode, with the modes
being bending, stretching, and shearing. Since the center of the fibers are distributed
at random in the simulation cell, the probability distribution for their segment
length is known, and is given by
$ %
2p 2p
P (ls ) = exp − ls , (57)
π lf π lf
and therefore the average segment length is, ls  = π lf /(2p).
If the deformation of the fiber network is quasi-static, then the fiber segments will
be deformed such that there is force equilibrium at all fiber-fiber bonds, which also
define the global minimum of the total elastic energy of the system. This implies that
the fiber segments will, in general, be deformed in a way that offers the least elastic
resistance. We may define the segments either by bending/shearing or by stretching.
According to the stiffness matrix (53), the bending stiffness modulus is Yf w 4 / ls3 ,
the shear stiffness modulus is Yf w 2 /[2(1 + νp )ls ], while the elongation stiffness
modulus is Yf w 2 / ls . Åström et al. (2000) assumed that a segment deforms only
by bending if the bending modulus is smaller than both the shear ! and elongation
modulus, i.e., if the segment length is such that, ls > lc ≡ w 2(1 + νp ). On the
other hand, if ls < lc , then the segments are assumed to deform by shearing and
stretching. The final ingredient of the model is the assumption that elongation of
a segment is proportional to cos2 θ [similar to Eq. (56)], while bending and shear
are proportional to sin 2θ. We note that the strain field in the effective-medium
treatment does not include any rotation.
 8.1. Quasi-static Fracture of Fibrous Materials 385

We can now compute the total elastic energy H of the system which is given by
$ %$ % π/2  $ %
Lx Ly 1 cos4 θ lc 2p Lx Ly
H = px2 Yf w2 dθ exp[−2pl/(πlf )]dl + px2
lf 2 −π/2 π 0 π lf lf

$ %  $ %$ %
1 π/2 sin2 (2θ ) lc 2p Lx Ly 1
× Gw2 dθ exp[−2pl/(π lf )]dl + px2 Yf w4
2 −π/2 4π 0 π lf lf 2

  ∞
π/2 sin2 (2θ ) 2p
× dθ exp[−2pl/(πlf )]dl, (58)
−π/2 4π lc π lf l 2

where G = Yf /[2(1 + νp )]. On the other hand, the elastic energy is related to the
effective stiffness Ce of the network by, H = (1/2)Ce x2 Lx Ly , which means that
the expression for Ce is given by
$ %     
pw 2 Yf 2pw 2 e−z 1 −z
Ce = − E1 (z) + 3 + (1 − e ) ,
8lf π lf z 2(1 + νp )
(59)
where z ≡ 2plc /(π lf ), and
 ∞ e−zx
En (z) = dx. (60)
1 xn

The first test of Eq. (59) is its ability for reproducing the known results in certain
limits. Hence, consider first the limit w → 0. If we rescale the network stiffness,
Ce → Ce /w 2 when w → 0, the fiber network becomes a central-force network,
i.e., a network of simple Hookean springs. Equation (59) then yields Ce ∝ w → 0,
which is expected since the average coordination number of the network is less than
4, and therefore, as explained in Section 8.7.3 of Volume I, the network cannot be
rigid. On the other hand, in the limit p → ∞, which is equivalent to w/ls  → ∞,
Eq. (59) predicts that Ce ∝ Yf w 2 p/ lf , implying that in the limit of high p the
network stiffness is simply proportional to Yf multiplied by the density of the
fibers in the network, i.e., the network behaves as an elastic continuum, which is
the expected behavior.
However, there remains one problem to be addressed. In writing down the ex-
pression for the total elastic energy H, Eq. (58), it was assumed that all segments are
deformed. However, below the percolation threshold of the network, Ce = 0, and
no segment is deformed. At, and just above, pc , there are also many segments that
carry no load, while for p  pc such segments appear only at the end of the fibers
with a density that can be shown to be about 0.55pc , independent of p (Åström
et al., 1994). Thus, for Eq. (59) to reproduce the correct percolation behavior, one
must make a transformation from p to the density pl of the loaded segments, and
Åström et al. (2000) suggested that p/pc = pl /pc + 0.55 + 0.45/(pl /pc + 1),
which is simply a crossover from p = pc when pl = 0 to pl → p − 0.55pc in
the limits p → ∞ and pl → ∞. Therefore, one should replace the first p on the
386 8. Brittle Fracture: The Discrete Approach

Figure 8.7. Comparison of the predictions of Eq. (62) (curve) with the results of numerical
simulations for w = 0.05 (+), w = 0.06 (×), w = 0.07 (∗), and w = 0.08 (2) (after Åström
et al., 2000).

right-hand side of Eq. (58) by pl given by

⎧ $ ⎫
⎨ %2 $ %1/2 ⎬
pc p p p
pl = − 1.55 + 1.55 − −4 1− . (61)
2 ⎩ pc pc pc ⎭
!
Finally, if we define zl = 2pl lc /(π lf ), and a reduced stiffness Cr = 16 2(1 + νp )
Ce /(πwYf ), we obtain
  −z    
z2 e 1 −z
Cr = zl − E1 (z) + 3 + (1 − e ) . (62)
2(1 + νp ) z 2(1 + νp )
Figure 8.7 compares the predictions of Eq. (62) with the simulation results for
various values of w, and it is clear that the agreement between the two sets is quite
good.
The shape of the stress-strain diagram for the fractured fiber network, shown
in Figure 8.4, can also be understood by appealing to the EMA. Here, we discuss
how this is accomplished by using the EMA1. The shape of a stress-strain diagram
of a fracturing material depends critically on the failure criterion. Suppose, for
example, that breaking occurs by axial tension. Then, Eq. (56) predicts that the
critical angle θf for failure is given by
$& %
f
θf = arccos , (63)
x
where f is the axial strain for failure. Equation (63) is obtained by writing
σf (x, k, θf ) = Yf f and σf (x, k) = Yf x and solving the resulting equation for
 8.1. Quasi-static Fracture of Fibrous Materials 387

θf . It can then be shown, using the EMA1 treatment described above, that one
obtains the following expression for the stress σ as a function of the strain x
(Åström et al., 1994):
$ %  π/2
2
σ = x w 2 p Yf [cos4 θ + (G/Yf ) sin2 (2θ )]dθ, (64)
π θf (x )

which is obtained from the total elastic energy of the system which, within the
EMA1, is given by
$ %  π/2
1 2 2 2
H = x w Yf p Lx Ly [cos4 θ + (G/Yf ) sin2 (2θ )]dθ. (65)
2 π 0
Equation (65) is of course a simplified version of Eq. (58).
As discussed above, as more fibers fail, the fracture zone becomes a narrow,
quasi-1D zone. Thus, in order to create such a zone, one assigns an infinitesimally
lower failure threshold to a band across the network, and then applies Eq. (64)
in this fracture zone which is given a unit width. No fiber fails in the rest of the
network, i.e., Eq. (64) is applied with θf = 0. The result is shown in Figure 8.8 in
which the dashed curve is the equilibrium curve. These predictions are in qualita-
tive agreement with the simulation results shown in Figure 8.4, except that there is
a discontinuity in the predicted stress-strain diagram after the elastic regime (the
regime of a linear relation between σ and x ) ends, whereas the simulations do
not indicate such a sharp and discontinuous change. The disagreement between
the simulation results and the EMA1 predictions becomes progressively stronger
as more fibers fail. The same qualitative trends would have been obtained, had
we used the EMA2 to derive the stress-strain diagram for the fracturing fiber
network. Therefore, although the EMA provides some qualitative insight into the

Figure 8.8. The stress-strain diagram as predicted by the effective-medium approximation

(solid curve).
388 8. Brittle Fracture: The Discrete Approach

early stages of a fracture process, it cannot be expected to be accurate as failure of

the fibers progresses.

8.2 Quasi-static Fracture of Heterogeneous Materials

Lattice models can represent the behavior of fracture of materials if the phenomeno-
logical coefficients and properties of the materials, such as their elastic constants
and failure threshold (see below), are properly defined and set. As discussed above
and also in Chapter 7, use of a finite-element (FE) method for discretizing the
continuum equations and studying fracture has been popular among engineers.
The discretized equations, and the associated mesh that one obtains in such ap-
proaches, resemble a lattice model. However, only very weak spatial disorder can
be incorporated into such a model, since strong disorder necessitates use of a very
fine structured FE mesh which makes the computations prohibitive. An alterna-
tive approach to the FE method is based on identifying the key microstructural
features associated with the disorder and relevant to the failure process. One then
subsumes all of the details of the mechanical behavior of that material region,
including the failure of a region of the material by the nucleation of a stable crack
of the same size, into a local constitutive law. Disorder is included by allowing the
phenomenological coefficients of the constitutive law to vary, from bond to bond,
according to some probability distribution. A network of such bonds is then used to
numerically calculate local stresses on, and interactions between, the bonds (and
sites) under the application of a macroscopic boundary condition. By allowing
for failure of such bonds under their local stress or strain (or a combination of
both), cracks are formed which may interact with each other, generate new cracks
via load transfer, and propagate to macroscopic sizes, leading to material failure.
Thus, one is able to account for the nucleation of cracks on the key length scales
and also the effect of disorder on such phenomena.
This approach was first used, in a rather primitive form, over 30 years ago.
Early efforts for developing discrete models of fracture of materials (Mikitishin
et al., 1969; Dobrodumov and El’yashevich, 1973) used lattices in which the
bonds were linear springs that could only be stretched (no bending or rotation
was allowed). However, because of the computational limitations of their times,
and the over-simplified nature of the models, they did not attract wide attention.
To our knowledge, modern lattice models of quasi-static mechanical fracture of
heterogeneous materials, of the type that are described in this chapter, were first
proposed by Sahimi and Goddard (1986).
Generally speaking, three variations of such lattice models have been developed
for studying mechanical breakdown in disordered materials. In the first approach,
which is completely deterministic, one uses a heterogeneous lattice each bond of
which describes the system on a certain length scale, with failure characteristics
described by a few key parameters. One then deforms the lattice gradually by
applying a boundary condition to the system that resembles what is used in an
experiment on fracture of a material, as a result of which the individual bonds
 8.2. Quasi-static Fracture of Heterogeneous Materials 389

break irreversibly in a certain manner. These models are either quasi-static so

that the process time enters the computations only as the number of Monte Carlo
steps (or as the number of bonds that are broken), or explicit time-dependence of
the fracture process is somehow built into them. This class of models is usually
appropriate for materials in which the disorder is quenched (fixed in time).
The second and third approaches are probabilistic. One of them (Louis and
Guinea, 1987; Hinrichsen et al., 1989; Meakin et al., 1989) draws on an analogy
between mechanical breakdown and the dielectric breakdown model of Niemeyer
et al. (1984) described in Section 5.4.1. As in Niemeyer et al.’s model, these mod-
els give rise to complex fractal crack patterns, and may be appropriate for systems
in which disorder is annealed; comparison between the predictions of such models
and fracture of materials with annealed disorder confirms this (see below). The
second class of probabilistic models was intended mainly for fracture of polymeric
materials. In these models, an elastic element breaks with a temperature-dependent
probability, hence taking into account the effect of the activation and elastic en-
ergies stored in the element. As we will see later in this chapter, many of the
probabilistic models have, in some sense, some type of dynamics built into them.

8.2.1 Lattice Models

Consider a 2D network, such as a L × L triangular or square lattice, or a 3D
network such as a L × L × L simple-cubic or BCC lattice. Every bond of the
lattice represents a Hookean spring or beam. In the former case, every site i of
the lattice is characterized by a displacement vector ui , while in the latter case,
in addition to its displacement ui , site i is also characterized by a rotation vector.
Hence, the nearest-neighbor sites are connected by springs or beams. The initial
(equilibrium) length of all the springs or beams is the same and, unlike the FE
method in which the mesh is made finer where the stress is larger, the initial
topology of the network is the same everywhere. The exception to this rule is
when one uses the lattice models for studying mechanical and fracture properties
of fibrous materials. Such models were already described above and also in Chapter
8 of Volume I, and therefore are not discussed any further.
We consider here the case of a brittle material for which a linear relation between
the stress and strain in the spring or beam is valid up to a threshold (defined
below). A force law of this type is not, of course, completely realistic, but has
long been thought of as a sensible approximation for brittle ceramics (see, for
example, the discussion by Lawn, 1993). The displacements ui (and the rotations)
are computed by minimizing the total elastic energy H of the system, the exact
form of which depends on the type of model that one wishes to study, and the
degree of microscopic detail that one incorporates into the model. For example,
lattices in which only the central- or stretching (Hookean) forces are operative
(Sahimi and Goddard, 1986; Beale and Srolovitz, 1988; Fernandez et al., 1988;
Srolovitz and Beale, 1988; Hansen et al., 1989; Arbabi and Sahimi, 1990b; Sahimi
and Arbabi, 1993), those in which the bond-bending or angle-changing forces, in
addition to central forces (see below) also act on the bonds of the lattice (Sahimi
390 8. Brittle Fracture: The Discrete Approach

and Goddard, 1986; Arbabi and Sahimi, 1990b; Sahimi and Arbabi, 1992, 1993,
1996; Sahimi et al., 1993), as well as the Born model described in Chapter 8 of
Volume I (see also below) in which the elastic energy of the system consists of the
contributions by the central forces and a scalar-like term (Hassold and Srolovitz,
1989; Yan et al., 1989; Caldarelli et al., 1994) have all been utilized.
Let us now describe these lattice models. In general, the elastic energy of the
bond-bending (BB) model is given by (Kantor and Webman, 1984)
1
1

H= α [(ui − uj ) · Rij ]2 eij + γ (δθ j ik )2 eij eik , (66)
2 2
ij  j ik

where α and γ are the central and BB force constants, respectively, j ik indicates
that the sum is over all triplets in which the bonds j -i and i-k form an angle with
its vertex at i, and eij = 1 if i and j are connected, and eij = 0 otherwise. The first
term on the right side of Eq. (66) represents the contribution of the stretching forces,
while the second term is due to BB forces. The precise form of δθ j ik depends on
the microscopic details of the model. In the most general form, if bending of all
pairs of bonds that have one site in common, including the collinear bonds, is
allowed, then (Arbabi and Sahimi, 1990a)

(uij × Rij − uik × Rik ) · (Rij × Rik )/|Rij × Rik |, Rij not parallel to Rik ,
δθ j ik =
|(uij + uik ) × Rij |, Rij parallel to Rik ,
(67)
where, uij = ui − uj . For all 2D systems, Eq. (67) is simplified to

δθ j ik = (ui − uj ) × Rij − (ui − uk ) × Rik .

The BB model has a well-defined continuum counterpart. For most materials to

which the BB model is applicable, one has γ /α ≤ 0.3 (Martins and Zunger, 1984).
In the Born model the associated elastic energy is given by
1
1

H= α1 µ[(ui − uj ) · Rij ]2 eij + α2 (ui − uj )2 eij , (68)
2 2
ij ij

where Rij is the unit vector along the line from i to j , and α1 and α2 repre-
sent, more or less, two adjustable parameters. The first term of Eq. (68) is the
energy of a network of central-force springs, i.e., Hookean springs that trans-
mit force only in the Rij direction, but do not transmit shear forces, whereas
the second term is a contribution analogous to scalar transport (for example,
the power dissipated in conduction), since (ui − uj )2 represents the magnitude
of the displacement difference ui − uj . The Born model can be derived from
linear continuum mechanics by discretizing the linear equation that governs the
elastic equilibrium of a solid, i.e., ∇ · σ = 0 (where σ is the stress tensor), and
using the usual relation, σ = λ(∇ · u)U + µ[∇u + (∇u)T ], where λ and µ are
the usual Lamé constants, and U is the identity tensor (see Section 8.4 of Vol-
ume I for details). If this is done, then one obtains, α1 = 2(1 − νp )/(1 + νp ), and
 8.2. Quasi-static Fracture of Heterogeneous Materials 391

α2 = 2(1 − 3νp )/[4(1 − νp )], where νp is the Poisson’s ratio. However, in this
form, the elastic energy given by Eq. (68) will not be rotationally invariant, thus
violating a fundamental physical requirement for an elastic energy representation
of a solid material. Therefore, Eq. (68), in which α1 and α2 are treated as adjustable
parameters, is a semi-empirical representation of materials.
The Born model may be considered as an analogue of a 3D solid in plane-stress
with holes normal to the x-y plane, or as a 2D solid with the Poisson’s ratio defined
as the negative of ratio of the strain in the y-direction to that in the x-direction,
when a stress is applied in the x-direction but none is applied in the y-direction.
Results for a 3D solid in plane-strain can be generated from those of this model
using the transformation νp = νp /(1 + νp ), where νp is the Poisson’s ratio for the
plain strain.
Let us mention another interesting way of generating a BB model. In their studies
of brittle fracture, Chung et al. (2001) generated a spring network by molecular
dynamics simulation, starting with a random distribution of spheres that interact
with each other through certain potentials. The system would then be allowed to
reach equilibrium, after which the centers of the spheres that were not separated by
a distance larger than a certain limit were connected by springs. Both the central
and BB forces were included in the network so obtained.
The spring lattices are suitable models for simulating a fracture process in mate-
rials that are under shear or tension. However, one should use the beam model (see
Chapter 8 of Volume I for more details) (Herrmann et al., 1989a; de Arcangelis et
al., 1989; Tzschichholz, 1992,1995; Tzschichholz et al., 1994; Tzschichholz and
Herrmann, 1996) when external compressional forces are imposed on the system,
since a spring cannot break under compression. In the beam model, in addition to
the central and BB or angle-changing forces, torsional forces also contribute to the
elastic energy H of the lattice. We believe, however, that, except when external
compressional forces are imposed on the system, the BB model is a completely
realistic representation of the elastic energy of disordered materials. Recall that, as
discussed in Chapters 8 and 9 of Volume I, the BB model is capable of describing
the elastic properties of polymers, glasses, ceramics and powders, and hence use
of more complex models for the elastic energy of the material is not necessary.
In addition to the above models, a model based on discretization of the following
equation (sometimes referred to as the Lamé equation)

(λ + µ)∇(∇u) + µ∇ 2 u = 0, (69)

where λ and µ are the usual Lamé constants, has also been used (Herrmann et al.,
1989b).
Sahimi and Goddard (1986) suggested that three general classes of disorder may
be incorporated into such model, which are as follows.
(1) Deletion or suppression of a fraction of the bonds either at random or in a
prescribed fashion, so that the material’s heterogeneity is of percolation-type.
The suppressed or deleted bonds may, for example, represent the microporosity
or some type of defect in the system before the fracture process began.
392 8. Brittle Fracture: The Discrete Approach

(2) Random or correlated distribution of the elastic constants eij of the bonds.
The idea is that in real heterogeneous materials the shapes and sizes of the
elastic zones through which stress transport takes place may be statistically
distributed, resulting in a different eij for each zone, or bond in the lattice
model, that follows some type of a statistical distribution. Such a model may
be appropriate for a composite material that, for example, consists of several
constituents, each of which has its own elastic properties.
(3) Random or correlated distribution of the critical thresholds at which the lin-
ear constitutive relation that describes the stress-strain relation in the beam
or spring breaks down. For example, in shear or tension each bond may be
characterized by a critical length lc , such that if it is stretched beyond lc , it
breaks irreversibly. Such a threshold can be estimated experimentally by eval-
uating macro tensile strength of the material. Alternatively, each bond can be
characterized by a critical force (stress) Fc (σc ), such that if it suffers a force
(stress) larger than Fc (σc ), it breaks irreversibly. Under compression, a beam
breaks if it is bent too much. The idea for using this type of disorder is that
a solid material made up intrinsically of the same material (the same elastic
constant eij everywhere) may contain regions having different resistances to
breakage under an imposed external stress or potential due to, for example, the
presence of defects during its manufacturing or formation process. Depending
on the intended application, we may use any combination of the three types
of disorder. For example, one may model the disordered material with fractal
lattices with bonds that have statistically-distributed properties (such as their
elastic constant). Because of their fractality, such models have low connectiv-
ities and large porosities, and may be relevant to transgranular stress corrosion
cracking of ductile metal alloys, such as stainless steel and brass (Sieradzki
and Newman, 1985). They may also be relevant to stress and crack propaga-
tion in weakly-connected granular media, such as sedimentary rocks. We do
not, however, consider them here as they have not been studied extensively.
Another important source of disorder in stressed materials is the so-called resid-
ual stress variations, which are caused by, among other things, thermal expansion
mismatch. The appropriate elastic lattice models with bond mismatch were de-
scribed in Section 9.7 of Volume I. We will not discuss the effect of this type of
disorder on fracture, although they can be analyzed by modification of the models
that are described here (see, for example, Curtin and Scher, 1990a,b; Sridhar et
al., 1994).
After selecting the lattice and the form of the elastic energy of the system (i.e.,
the types of forces that are operative in the lattice), we specify the type of the
heterogeneity that the material contains. If the disorder is of percolation-type (type-
1 heterogeneity described above), then its inclusion in the lattice is straightforward
and needs no discussion. For types-2 and 3 heterogeneities described above, their
statistical distribution must be specified. A statistical distribution that has been
used widely is
f (x) = (1 − ζ )x −ζ , (70)
 8.2. Quasi-static Fracture of Heterogeneous Materials 393

where x is any property of the lattice that is statistically distributed and repre-
sents its heterogeneity, and 0 ≤ ζ < 1. The advantage of the distribution (70) is
that, varying ζ allows one to generate distributions that are very narrow (ζ → 0)
or very broad (ζ → 1), and therefore one can study the extent to which such ex-
treme distributions affect failure phenomena. Note that ζ = 0 represents a uniform
distribution, while Roux et al. (1988) showed that, in the limit ζ → 1, fracture be-
comes equivalent to a type of percolation. A great advantage of the lattice models is
that, any type of statistical distribution f (x) of the heterogeneities can be used. For
example, de Arcangelis et al. (1989) used, in addition to (67), a Weibull distribution
2 3
f (x) = mλ−m x m−1 exp −(x/λ)m , (71)
where 2 ≤ m ≤ 10 supposedly describes many real materials.
After specifying the lattice type, the form of the elastic energy H, and the type
of disorder, the boundary conditions must be specified. One can, for example,
use shear, uniaxial tension or compression, uniform dilation (i.e., pulling a lattice
equally in all directions), or surface cracking which is used for simulating fracture
of a thin film of a material attached to a substrate (for example, thin polymeric
coatings, or paints, or even mud). In this case, each site of the lattice is connected
by a spring to the substrate which has a lattice constant larger than the original
lattice. In this way all the bonds are equally overstretched without having applied
any force on a boundary of the lattice, implying that no external boundary is in
fact needed, and one can use periodic boundary conditions in all directions.
The simulations can now begin. One must compute the distribution of the nodal
displacements (and rotations, if such motions are allowed), from which the forces
(and stresses) exerted on all the bonds are computed. The procedure for doing so
consists of minimizing the total elastic energy of the system with respect to the
displacements of the internal nodes of the lattice (and their rotations, if such motion
is allowed). Because of the assumption of brittleness, these equations are linear
and therefore, subject to the boundary conditions imposed on the system, can be
solved by one of several methods that are available for solving such equations. If
very high precision is needed, then the conjugate-gradient method (see Chapter 9
for a description of this method) is the best technique to use.
After computing the initial distribution of the stresses (and strains) in the lattice,
a criterion for nucleation of the microcracks must be specified. The criterion,
however, depends on the type of material that is being studied. For example, if
each elastic bond is a rubber band, then it will tear apart when stretched beyond
a certain limit. Thus, for example, we assign a threshold lc for the length of the
bonds, which is selected from the probability density functions described above.
Then, in terms of lc , the breaking criterion is that a bond breaks if its length in
the deformed lattice exceeds its lc . Alternatively, among all the bonds that have
exceeded their lc , the one with the largest deviation from its lc breaks first. The
idea is that in a deformed material, the weakest point of the system fails first.
However, if the elastic bond represents, for example, a glass rod, then it will
break if it is bent too much. One must of course use a lattice of beams for modeling
such a material. Therefore, a good strategy would be devising a breaking criterion
394 8. Brittle Fracture: The Discrete Approach

that is a combination of both stretching and bending. For example, in the beam
model one can use the criterion (de Arcangelis et al., 1989) that the beam with
largest value of the following quantity
$ %2  
F |Mi | |Mj |
pb = + max , , (72)
Fc Mc Mc
breaks, where F is the longitudinal force acting along the beam and Fc its critical
value, and Mi and Mj are the moments applied on the two adjacent sites i and
j of the beam, with Mc being the critical threshold of the moment. Both Fc and
Mc are distributed according to some probability density functions and can, in
fact, represent the heterogeneity of the material (type-3 disorder discussed above).
The first term of (72) describes breaking due to stretching, while the second term
is representative of breaking due to bending. One can distribute, for example, Mc
over a broader range than Fc (and vice versa) as a measure of susceptibility of the
material to breaking by bending as opposed to stretching.
Other failure criteria of this type can be, and have been, used. For example,
a combination of Fc and lc can also be used for setting up a breaking criterion
(Arbabi and Sahimi, 1990b): one breaks that bond for which the ratio R = lm l/ lc
is maximum, where l is the current length of the bond in the deformed lattice and
lm is the maximum length that any bond in the lattice has, or break the bond for
which U = F lc /Fm is minimum, where F is the total force that the bond suffers,
and Fm is the maximum force on any bond in the lattice. Sridhar et al. (1994) used
the criterion that a bond breaks if its strain energy exceeds a critical value. The
flexibility that the lattice models afford one in using almost any type of failure
criterion is one important advantage of such models.
Another advantage of such lattice models is that, depending on the intended
application, any failure criterion can be used without imposing any undue difficulty
on the computations. For example, often portions of a material are damaged but
do not break during the fracture process. The damaged area can be modeled by
lowering the breaking threshold of the bonds that are in the vicinity of the growing
crack. It is clear that the bonds that are damaged in this way are more likely to
break in the next step of the simulations than the undamaged, unbroken bonds.
One can also model short-lived damage by considering at step n of the simulations
the quantity
pb = pb (n) + β0 pb (n − 1), (73)
where pb (n) is the usual quantity that one uses for breaking criterion at step n
without considering the damage [for example, Eq. (72)], and 0 < β0 < 1. Then,
the breaking criterion must be based on pb , and pb (n − 1) indicates the effect of
“memory” of the damage that occurred during the previous step of the simulations.
In some fracture processes, such as stress corrosion, the damage accumulates
during the entire process, and therefore the breaking criterion must somehow
reflect this. A simple algorithm for taking this effect into account is as follows
(Herrmann et al., 1989b). One assigns a counter c(n) to each bond of the lattice
that is susceptible to damage, e.g., those in the “vicinity” of the growing crack,
 8.2. Quasi-static Fracture of Heterogeneous Materials 395

where the boundaries of this region must also be specified. For example, one can
consider all the unbroken bonds that are up to a certain distance from the cracked
area. At the beginning of the simulations, all the counters are set to zero, c(0) = 0,
for all the bonds that may be damaged. Then, at each step n of the simulations, one
calculates pb (n), the quantity based on which the breaking criterion is applied,
and computes
1 − c(n − 1)
α(n) = . (74)
pb (n)
The bond with the smallest value of α, namely αm , is then broken. Then, all the
counters are updated by
c(n) = αm pb (n) + βc(n − 1), (75)
so that c(n) “accumulates” the damages from all the previous steps of the process.
Clearly, if β = 1, then the damage is irreversible. The limit β → 0 corresponds to
criterion (73) with β0 = 1.
The failure process is then initiated by selecting the bond (or bonds) that must
break. Two different “dynamics” of fracture propagation can be studied. In model
1 only one bond is broken at each stage of the simulation, which is equivalent
to assuming that the rate at which the elastic forces relax throughout the network
is much faster than the breaking of one bond. In model 2 (and depending on the
failure criterion), all bonds that meet the failure criterion are broken. Most of the
studies so far have used model 1, and the properties of model 2 have not been
studied extensively. Breaking a bond is equivalent to removing it from the lattice
(by, for example, setting its elastic constant to zero), after which one recalculates
the stress and strain distributions for the new configuration of the lattice, select the
next bond(s) to break, and so on. If the external stress or strain imposed on the lattice
is not large enough to break any bond, it is gradually increased. The simulation
continues until a sample-spanning crack is formed. As mentioned above, instead of
removing a failed bond from the lattice, one may reduce its elastic constant. In this
case, one observes the interesting phenomenon of crack arrest (Li and Duxbury,
1988).

8.2.1.1 Shape of the Macroscopic Fracture

The number of the cracks, their size distribution, and the shape of the macroscopic
fracture, if one is formed, all depend on a variety of factors, including the broadness
of the distribution of the material’s heterogeneity, the dimensionality of the system,
the boundary conditions applied to the system, and the interplay between the
quenched disorder, which tries to delocalize the propagating microfractures, and
stress enhancement at the tip of the microfractures which attempts to localize the
fracture. For example, if an external strain is applied slowly, a great number of
microcracks form before a macroscopic fracture (network) is formed. The number
of the nucleated cracks depends on the broadness of the heterogeneity distribution
in the system. If this distribution is very broad, then the system is a mixture of
396 8. Brittle Fracture: The Discrete Approach

very strong and very weak regions. For the growing crack to take advantage of the
weak regions, it must find its path which could be quite tortuous, so much so that it
may result in fragmentation of the material (Åström and Timonen, 1997a). We will
discuss the phenomenon of fragmentation later in this chapter. On the other hand,
when weak disorder is present in a solid, a catastrophic crack is formed quickly
that spans the system, and therefore the mechanical failure of the system is very
fast.
The number and shape of the cracks also depend on the dimensionality of the
system. In 3D stress enhancement at the tip of the growing cracks is much weaker
than in 2D, so that even mild disorder can give rise to very complex fracture
pattern in 3D. Moreover, the type of the boundary conditions applied to the system
is also important. If a stress (instead of a strain) is applied to the material, then
the system would fail very quickly soon after the first microcrack is formed, even
if the stress is applied slowly to the system. Therefore, similar to real fracture
tests, there are significant differences between a stress-controlled and a strain-
controlled fracture test in such models. Figure 8.9 shows three different stages of
fracture of a 2D triangular lattice. The bonds of the lattice suffer only stretching
(central) forces, disorder was generated in the lattice by removing 10% of the
bonds at random before deformation of the lattice began, and the deformation was
caused by applying a strain to one face of the lattice, while the opposite face was
fixed. The top panel shows the initial configuration of the lattice before it was
deformed. The middle panel represents the system when it loses its rigidity. Recall
from Chapter 8 of Volume I that if only central forces act on the bonds, the lattice
loses its rigidity at a fraction of uncut springs which is significantly larger than the
connectivity threshold. For example, the connectivity threshold of the triangular
lattice is at pc = 0.347, while a triangular lattice of Hookean springs loses its
rigidity at p  0.641. However, the rotational freedom of the bonds (which do
not contribute to the total elastic energy of the system) is not lost yet. At this
point, the lattice contains a strip of bonds with zero shear modulus, which can be
thought of as a type of shear band. Note that the shear band can move at most
one bond before the shear modulus is re-established. Final failure of the lattice,
which occurs due to formation of a macroscopic crack that splits the system into
two pieces, occurs very close to the shear band. In contrast with Figure 8.9, Figure
8.10 presents an example of the microcracks that are produced in the same lattice
in which both the central and bond-bending (angle-changing) forces contribute
to the elastic energy of the system, and the distribution of the heterogeneities is
broad, following Eq. (70).
Other interesting crack shapes can also be produced by such models. For exam-
ple, Herrmann et al. (1989b) considered a square lattice of beams with periodic
conditions in the horizontal direction, while the top and bottom of the lattice was
sheared. They defined a probability pb slightly more general than Eq. (69), given by
pb = [F 2 + q · max(|Mi |, |Mj |)]η , (76)
where F is the traction (and/or compression) force applied on a beam, and q and
η are two free parameters, with q being a measure of the affinity of a beam to
 8.2. Quasi-static Fracture of Heterogeneous Materials 397

Figure 8.9. At the top is the ini-

tial configuration of the system in
which 10% of the bonds have been
removed at random. In the mid-
dle is the system after the rigidity
failure (when the fraction of bro-
ken bonds is less than the rigidity
threshold of the system), while
the bottom system is at the com-
plete fracture point (after Beale
and Srolovitz, 1988).

Figure 8.10. Fracture pattern in a triangu-

lar network with stretching and bond-bending
forces and a broad distribution of the critical
lengths lc .

breaking by bending, and η having no apparent physical meaning. A beam was re-
moved at the center of the lattice in order for microcracking to be initiated. Figure
8.11 presents three fracture patterns. The one at the top was produced by using a
breaking criterion based on Eqs. (74) and (75), where pb was assumed to be given
398 8. Brittle Fracture: The Discrete Approach

Figure 8.11. The top and the middle fig-

ures show the fracture patterns (rotated by
45◦ ) in a 60 × 60 network. The pattern at
the bottom is the experimental fracture pat-
tern (see the text) (after Herrmann et al.,
1989b).

by Eq. (76), and q = 0, β = 1 and η = 1, while the middle one was generated
based on Eq. (73) with β0 = 1 and η = 0.2. Evidently, decreasing η increases the
tendency of the growing fracture to branch out. Also shown at the bottom of Figure
8.11 is the fracture pattern in an alloy, Ti11.5 Mo6 Zr4.5 Sn, aged 100 hours at 750 K
and cracked under increasing stress intensity. There are certain similarity between
the experimental pattern and what is shown in the middle of the figure.
Let us point out that, under certain conditions, the set of the broken bonds (the
microcracks) can form a fractal network. The fractality of this set, and that of the
sample-spanning fracture, will be discussed later in this chapter.

8.2.1.2 Dependence of the Elastic Moduli on the Extent of Cracking

One of the most interesting bits of information, which is also experimentally ac-
cessible, is the behavior of the elastic moduli of the system as the breaking process
proceeds. Shown in Figure 8.12 are the Young’s moduli Y of three distinct models
undergoing fracture (Arbabi and Sahimi, 1990b) and their dependence on the frac-
tion p of the unbroken bonds. Three distinct sets of results are shown. One is for a
BCC lattice with central-force springs (recall from Chapter 8 of Volume I that in 3D
the rigidity percolation thresholds of the simple-cubic lattice of Hookean springs
with no BB forces is 1, and therefore it cannot be used in such simulations), while
the second set is for a simple-cubic lattice in which both the central and bond-
bending forces contribute to its deformation. The third set is the Young’s modulus
of a percolating simple-cubic network with central and bond-bending forces in
which a fraction 1 − p of the bonds has been removed at random. Clearly, a bond
breaking process in which the bonds are broken according to the force that they
suffer, or according to the deviation of their length from a threshold value, weakens
the system much faster than a random (percolation) bond-breaking process and,
as a result, the system fails much sooner than a percolation system. Also shown
in this figure are some experimental data on the Young’s modulus of ceramic and
glasses during their microcracking. The data were in terms of the microporosity
of the materials, which was taken to be proportional to (1 − p). As can be seen,
the predictions of the fracture model with the central and bond-bending forces are
 8.2. Quasi-static Fracture of Heterogeneous Materials 399

Figure 8.12. Young’s modulus of the central-force (triangles) and the bond-bending (2)
models (with γ /α = 0.04) versus the fraction p of unbroken bonds. Also shown are the
modulus of a simple-cubic lattice in bond percolation (circles), as well as the experimental
data (dashed curve) (after Arbabi and Sahimi, 1990b).

within the range of the experimental data. Moreover, except for p  1, the re-
sults produced by the central-force model do not agree with the experimental data
because such systems fail at high values of p. For p ≥ 0.5, the results predicted
by the random percolation model do not agree with the data as well as the lattice
model of fracture, presumably because the percolation threshold of the lattice is
low (pcB  0.25) and, as a result, the predicted modulus is somewhat large.
An interesting study by Curtin (1997) demonstrated the effect of disorder on
quasi-static fracture. He investigated the fracture toughness of heterogeneous ma-
terials using a simple-cubic lattice of springs with distributed toughness. He found
that the overall toughness of the lattice or, equivalently, the stress σf to initiate the
first microcrack, is a random variable that depends on the width of the toughness
distribution of the individual springs. This by itself is not surprising (see below).
What was interesting was the finding that, for narrow distributions, the toughness
was found to be controlled by the nucleation of the kinks at the weakest springs,
whereas for broad distributions the toughness was controlled by the highly rigid
regions of the system that pin the growth of the fracture front. However, the dif-
ference between the toughnesses of materials with narrow and broad distributions
400 8. Brittle Fracture: The Discrete Approach

was found to be small, hence suggesting that simple disorder alone (such as a
narrow and uniform distribution of the thresholds) cannot solely be responsible
for the variety of fracture behavior seen in experiments, and more complex factors
must play an important role. We will come back to this issue later in this chapter.

8.2.1.3 Fracture Strength of Materials with Strong Disorder

The failure strength σf of a solid is usually determined in a tensile test. There are
several definitions of the failure stress which depend on the nature of the tensile
test. In a stress-controlled test the sample fails at the highest stress in the stress-
strain diagram. In the context of network models considered here, this usually
occurs at the point where the first bond breaks. In a strain-controlled test, on
the other hand, the strain is incremented and stress is the dependent variable. As
the stress is finite for all strains, the failure stress in this case corresponds to the
point where the stress first drops to zero. We define stress, or fracture, strength
σf of a system as the lowest externally applied stress at which the system breaks
down. One can hypothesize that the eventual failure of the system is governed by
the most critical flaw in the system, i.e., the weakest part of the system. Hence,
calculation of the full distribution function of fracture strength σf reduces to the
computation of the distribution function of the most critical flaw in the system. It
can be shown that this is an excellent approximation for the failure stress of the
system in a stress-controlled tensile test. However, the fracture strength is not a
self-averaging property. That is, the distribution of the fracture strength of even
large samples of nominally the same material is not a delta function, implying that
the most probable fracture strength and its average are not equal. Therefore, the
full distribution must be computed or measured.
We first consider fracture strength of a strongly disordered material. The mate-
rial’s heterogeneity is generated by a percolation algorithm, i.e., before deformation
and cracking of the material begins, only a fraction p of the system is solid mate-
rial; the rest of it is pre-fracture vacancy or voids generated by some type of defect.
This means, in the context of the lattice models discussed in this chapter, that a
fraction p of the bonds (springs or beams) are present and the rest are cut before
deformation of the lattice begins.
Monte Carlo simulations of quasi-static fracture of triangular and simple-cubic
lattices of springs with central and bond-bending forces were carried out by Sahimi
and Arbabi (1993) to check whether the fracture strength of materials with strong
(percolation-type) disorder exhibits universality near the percolation threshold.
We assume that
σf ∼ (p − pc )Tf , (77)
where Tf is a new critical exponent that is not equal to f , the critical exponent
that characterizes the elastic moduli of linear elastic materials near the percolation
−T̂
threshold pc . Equation (77) is now rewritten as, σf ∼ ξp f , where ξp is the cor-
relation length of percolation and T̂f = Tf /ν, with ν being the critical exponent
of ξp . Thus, one may use the standard finite-size scaling method to estimate T̂f .
 8.2. Quasi-static Fracture of Heterogeneous Materials 401

Briefly, according to the finite-size scaling theory, for length scales L  ξp , we

must replace the correlation length ξp by L, the linear size of the network, and
hence in this regime σf ∼ L−T̂f . Thus, one carries out a series of simulations with
lattices of various sizes, at the percolation threshold pc of the infinite lattice, and
estimates σf for each L. Since only small and moderate L can be used, the results
are fitted to σf = L−T̂f [a0 + a1 h1 (L) + a2 h2 (L)], where a0 , a1 and a2 are three
constants, and h1 (L) and h2 (L) are correction-to-scaling functions that take into
account the effect of finite sizes of the network. The most accurate estimate of T̂f
is obtained with (Sahimi and Arbabi, 1991), h( L) = (ln L)−1 and h2 (L) = L−1 .
Monte Carlo simulations in 2D (Sahimi and Arbabi, 1993) yielded, Tf  2.42 ±
0.14, in good agreement with the measurement of Benguigui et al. (1987), Tf 
2.5 ± 0.4, described in Section 8.2.2, while in 3D (Sahimi and Arbabi, 1993),
Tf  2.64 ± 0.30. In addition to the Monte Carlo results of Sahimi and Arbabi
(1993), less extensive molecular dynamics computations of Ray and Chakrabarti
(1985b), and lattice simulations of Beale and Srolovitz (1988) also seemed to
support the validity of Eq. (77).
If, instead of a lattice, one considers a continuum in which the distribution of
the local transport properties may have certain singular properties, then one has
a new universality class for the transport exponents. The same is true about the
exponent Tf . For example, if one punches holes in the material at random, then
the material, before its deformation begins, resembles the Swiss-cheese model,
i.e., one in which spherical inclusions (or circular inclusions in 2D) are randomly
distributed in a uniform matrix. The elasticity exponent f of the Swiss-cheese
model, that characterizes the power-law behavior of its elastic moduli near the
percolation threshold pc , is larger than that of the lattice model. Therefore, we may
also expect higher values of Tf in order to describe the power-law dependence
of the fracture strength of the Swiss-cheese model near the percolation threshold,
and this is indeed true. In fact, using a scaling analysis, Chakrabarti et al. (1988)
proposed that for the Swiss-cheese model,

1 1
Tf  (f + dν + x) = (fsc + dν), (78)
2 2
where fsc = f + x is the elasticity exponent of the Swiss-cheese model, and
x = 3/2 and 5/2 for d = 2 and 3, respectively. Thus, for a 2D material, one obtains
Tf  4.06, in good agreement with the experimental measurement of Benguigui
et al. (1987) who also measured the fracture strength of a 2D Swiss-cheese model
(by punching holes into their material) and reported that, Tf  4.0 ± 0.1.
Rigorous upper and lower bounds for Tf have also been derived (Ray and
Chakrabarti, 1985a,1988). In a lattice near the percolation threshold, representing
a composite material with strong disorder, the nearest-neighbor uncut bonds or sites
form a type of “super-lattice”—a very large cluster made of tortuous links (long
strands made of many bonds)—that cross each other at nodes that are separated
by an average distance ξp , the correlation length of percolation. Therefore, in a
d-dimensional system, the external stress σ is shared by ξpd−1 number of parallel
402 8. Brittle Fracture: The Discrete Approach

links, implying that the stress per link is σl  σ ξpd−1 . Then, the total strain is,
f/ν
 = σ/Y  σ ξp , where we have used the power-law dependence of the Young’s
modulus on p near pc . The total strain is shared by ξp−1 number of links, which
1+f/ν
means that the strain per link is given by, l  σ ξp . Therefore, the elastic
energy per link is given by
d+f/ν
Hl  σl l  σ 2 ξp . (79)
We now recall that only the bonds in the backbone of the lattice, i.e., the multiply-
connected bonds of the sample-spanning cluster, contribute to stress and/or strain
transfer in the system, the total number M of which is, M ∝ ξpDbb , where Dbb is the
fractal dimension of the backbone. If Hl is shared equally by all such bonds, then
d−D +f/ν
the energy per bond Hl /M is of the order of σ 2 ξp bb . However, because
of the assumption of equal sharing by all the M bonds of the links, this value of
Hl /M underestimates the strain energy per bond, since many of the bonds either
do not support any stress at all or support very small stresses, because the backbone
of a percolation lattice is multiply connected. Thus, if we assume a fixed energy
threshold for breaking each bond, lattice fracture will occur for σf = σ for which
d−D +f/ν
Hl /M ∼ σ 2 ξp bb exceeds the threshold. Therefore, if σf follows Eq. (77),
then, because we underestimate Hl /M, we must have
1
Tf ≥ [f + (d − Dbb )ν]. (80)
2
On the other hand, recall that the number of singly-connected (red) bonds, i.e.,
1/ν
those that, if cut, would split the backbone into two pieces, is Mr ∼ ξp . If all the
strain energy is supported by such bonds, then one obtains an overestimate of the
d+f/ν−1/ν
elastic energy per bond, Hl /Mr ∼ σ 2 ξp . Using the same argument that
we utilized for the lower bound, we obtain
1 1
[f + (d − Dbb )ν] ≤ Tf ≤ (f + dν − 1). (81)
2 2
Using the numerical values of the various exponents, f  3.96 and 3.75, and
Dbb  1.675 and 1.87, for 2D and 3D systems, respectively, we obtain, 2.22 ≤
Tf ≤ 2.81 for d = 2, and 2.37 ≤ Tf ≤ 2.7 in d = 3. These bounds are perfectly
consistent with the experimental and numerical estimates of Tf given above.
Somewhat sharper bounds were proposed by Bergman (1986) who suggested
that
f − νDmin ≤ Tf ≤ f − 1, (82)
where Dmin is the fractal dimension of the shortest paths, or the chemical paths, on
the backbone of percolation clusters. These bounds, together with the estimates,
Dmin  1.13 and 1.34 for 2D and 3D systems, respectively, imply that
2.45 ≤ Tf ≤ 2.96, d = 2, (83)
2.58 ≤ Tf ≤ 2.76, d = 3, (84)
 8.2. Quasi-static Fracture of Heterogeneous Materials 403

which agree nicely with the simulation results. Moreover, in both 2D and 3D the
estimated Tf is close to the lower bound (82), and therefore the relation Tf =
f − νDmin cannot be ruled out.

8.2.1.4 Distribution of Fracture Strength

Traditionally, the Weibull distribution (WD) has been used in fitting the fracture
strength data for various materials. For a sample of a material of linear size L, the
cumulative WD is given by
FL (σf ) = 1 − exp(−cLd σfm ), (85)
where c and m are constant, and d is the dimensionality of the system [see Eq. (18)
which is the same as (85) with cLd = λ−m ]. However, as discussed in Sections
5.25 and 5.4.2.1, Duxbury and Leath (1987) formulated a new Gumble distribution
(GD) which is given by
  
−k
FL (σf ) = 1 − exp −cL expd
, (86)
σfδ
where k and δ are also constant. Although Eq. (86) is supposed to be valid for
materials that are far from their percolation threshold, i.e., those that have few
pre-fracture vacancies or voids, as our discussion below indicates, it is not clear
that this is actually the case. Both Eqs. (85) and (86) can be derived by appealing to
percolation statistics, or any other theory that can provide the statistics of clusters
of vacancies or voids in a material. Similar ideas were utilized to derive the failure
distribution for electrical and dielectric breakdown of heterogeneous materials (see
Chapter 5). We now describe the derivation of these two distributions for fracture
strength of materials.
Suppose that a solid material of linear size L contains n cracks, each with
failure probability qi (σ ), i = 1, 2 · · · , n, under an applied stress σ . To make the
analysis manageable, we further assume that the stress-released regions of each of
the cracks are separate and do not overlap, and that F (σ ) is the cumulative failure
probability of the entire sample under stress σ . Then (Ray and Chakrabarti, 1985a)
 
'n

1 − FL (σ ) = [1 − qi (σ )]  exp − qi (σ ) = exp[−Ld ρ(σ )], (87)
i=1 i

where ρ(σ ) is the density of cracks weaker than the stress σ or, equivalently, the
density of the cracks that will start propagating at and above the stress σ . Equation
(87) is due to the fact that the sample material survives if each of the cracks within
it survives. We now show that the WD and GD correspond to two limiting cases
of Eq. (87).
Consider first the derivation of the WD. This distribution arises if ρ(σ ) is given
by a power law in σ , which will be the case if the density ρ(l) of the linear cracks
is of power-law type. This type of distribution does not arise in materials with a
random distribution of vacancies or voids, unless the material is precisely at its
404 8. Brittle Fracture: The Discrete Approach

percolation threshold pc , or is in a state above pc but is viewed at length scales

L  ξp , where ξp is the correlation length of percolation. Moreover, such power-
law distributions do arise in real materials in which there are significant long-range
correlations, without necessarily being near pc . Suppose then that g(l) ∼ l −τ ,
where τ is a critical exponent that characterizes the power-law distribution of the
clusters of vacancies or voids (in the language of percolation theory, these are
clusters of broken or unoccupied bonds). We further assume that the breaking
stress (fracture strength) σf , beyond which a crack of length l is created, is related
to l by
1
σf ∼ , (88)
l 1/δ
where the value of δ is determined by the nature of the surface of the crack in
general, and its surface roughness in particular. Note that Eq. (88) represents a
generalization of the Griffith law (see Chapters 6 and 7) which predicts a 1/ l 1/2
singularity for σf (i.e., δ = 2). This generalization is justified on the ground that the
Griffith law was derived under the assumption that the cracks are smooth. However,
as discussed in Sections 7.8.7 and 7.8.14, when the cracks are not smooth and
have a rough and self-affine surface (which experiments show that this is often the
case), then one must use the generalized Griffith law in which case the parameter
δ can be treated as an adjustable parameter. If so, given Eq. (88), we can write
ρ(l) ∼ σfτ δ which, when substituted in Eq. (84), yields the Weibull distribution
with the Weibull exponent m = τ δ. If, for example, we define σq as the stress at
which a certain percentage q of the system fails (i.e., set F = q), we obtain
1
σq ∼ , (89)
Ld/m
so that the stress depends strongly on the system’s size.
Consider now the opposite limit in which the material is far from its percolation
threshold. Then, we know from percolation theory that ρ(l), i.e., the probability
of having a cluster of vacancies (cracks) of linear size l is given by
ρ(l) ∼ exp(−l/ξp ) ∼ exp(−1/ξp σfδ ), (90)
where the generalized Griffith law, Eq. (88), was again invoked. Equation (90),
when substituted in Eq. (87), yields an equation for FL (σf ) which is equivalent
to Eq. (86), which predicts that σq , the stress beyond which a certain fraction q of
the system fails, is given by
1
σq ∼ , (91)
(ln L)1/δ
which indicates a very weak logarithmic size dependence.
One can use the lattice models of fracture to directly test the accuracy of Eqs. (85)
and (86) and the conditions under which they may provide accurate fits of exper-
imental data. However, given that the results of such simulations may not be
accurate enough to distinguish between the two distributions, each of which con-
tains two adjustable parameters, a more sensitive test of the validity of these two
 8.2. Quasi-static Fracture of Heterogeneous Materials 405

distributions can be made if we rewrite Eqs. (85) and (86) in alternative forms. If
we define a quantity A by
 
ln[1 − FL (σf )]
A = − ln − , (92)
Ld
then the WD can be rewritten as
$ %
1
Aw = a1 ln + b1 , (93)
σf
while the GD is rearranged as
 
1
AG = a2 + b2 . (94)
σfδ
These two equations predict linear variations of Aw with ln(1/σf ) and of AG with
1/σfδ . The exact value of δ has not been determined, but in general 1 ≤ δ ≤ 2.
The conditions under which Eq. (85) or (86) may provide accurate representation
of fracture strength data for a given material are not completely clear yet. It appears
(Sahimi and Arbabi, 1993b) that in highly heterogeneous solids neither equation
is very accurate, although the WD appears to perform better. On the other hand,
in weakly-disordered materials, or those far from the percolation threshold (i.e., a
material with few vacancies or voids), the GD may be a better representation of
the distribution of fracture strengths. For example, Figure 8.13 presents the fit of
the fracture simulation results to a GD with δ = 1 for a triangular network with the
central and bond-bending forces, in which before deformation and fracture of the
lattice started, 10% of the bonds had been removed at random. Simulations with
the central force (Beale and Srolovitz, 1988) and the Born models (Hassold and

Figure 8.13. Fit of the simulation results for fracture of a triangular network with stretching
and bond-bending forces (with γ /α = 0.1) to the Gumbel distribution. Before fracturing,
10% of the bonds were removed at random (after Sahimi and Arbabi, 1993).
406 8. Brittle Fracture: The Discrete Approach

Srolovitz, 1989) also seem to indicate that the GD is accurate if the system is far
from pc , although Hassold and Srolovitz (1989) reported equally accurate fit of the
data with the WD, and Hansen et al. (1989) also reported some deviations from the
GD in their central-force model. Curtin and Scher (1992) discussed the conditions
under which a WD may be appropriate for representing the distribution of fracture
strength. We note, however, that as the percolation threshold is approached, neither
distribution seems to be very accurate.
Fracture behavior of a material at its percolation threshold pc , or equivalently
above its pc but at length scales L  ξp , depends on the broadness of the distri-
bution of its heterogeneities, and deserves to be discussed (Sornette, 1988). In a
percolation system far from pc , there are many multiply-connected paths, called
macro-links, which support stress transport. In such a system, the distribution of
fracture strength is the result of one or both of the following factors:
(1) Fluctuations of the individual characteristics of the microscopic regions (for
example, their breaking threshold lc and/or their elastic constants) of the
material.
(2) Fluctuations of the macro-link sizes L around the percolation correlation
length ξp . If the characteristics of the microscopic regions are all the same
(i.e., if the material is not made of different constituents with different proper-
ties), the first factor cannot contribute to the distribution of fracture strength.
As pc is approached, two changes take place: First, one has fewer macro-links
and, secondly, compared with those of the long macro-links, the contribu-
tions of the shorter macro-links to stress transport become negligible. Thus,
macro-link to macro-link fluctuations also decrease. At pc , there is only one
huge macro-link, and therefore all the fluctuations disappear completely and
the distribution of fracture strength must be a delta function. However, if,
for example, the elastic constants of the microscopic regions of the material
are statistically-distributed quantities (which is often the case in real hetero-
geneous solids), then region-to-region fluctuations exist and the distribution
of fracture strength is a meaningful quantity to define, calculate, or measure.
This is particularly important for disordered materials modeled by continuum
percolation which usually possess a broad distribution of the elastic constants
of the channels through which stress transport takes place. The distribution of
fracture strength in such materials is a Weibull-like distribution, rather than the
GD. This is supported by the lattice simulations of Sahimi and Arbabi (1993).

8.2.1.5 Size-Dependence of Fracture Properties

An important problem in fracture mechanics of materials is the dependence of the
various properties on the sample size. Moreover, equations such as (77), (88) and
(89) already state that certain features of fracture may have scale-invariant proper-
ties. Let us now summarize the equations that express the sample size dependence
of certain fracture properties, and discuss whether they are universal. We describe
here such properties for the discrete and deterministic models of fracture. Later in
this chapter, we discuss the same properties for other models of fracture.
 8.2. Quasi-static Fracture of Heterogeneous Materials 407

Similar to electrical and dielectric breakdown problems discussed in Chapter 5,

an important property is the average failure stress as a function of the sample size
L at a fixed p, where (1 − p) is the fraction of the vacancies or voids in the material
(or p is the fraction of the intact bonds in the lattice models) before fracture has
begun. To obtain this quantity, we first note that, in Eq. (86), the constants c and
k must depend on p. Thus, setting FL (σf ) to any constant value between 0 and 1
and solving for σf , we obtain from Eq. (83) the following expression

c
σfδ = , (95)
A(p) + B(p) ln L

where A(p) and B(p) are simple functions of p, and c is a constant. In particular,
B(p) is small for p close to unity, and B(p) → ∞ as p → pc . Simulations with
the central-force model (Beale and Srolovitz, 1988), the Born model (Hassold
and Srolovitz, 1989), and the central-force and bond-bending model (Sahimi and
Arbabi, 1993) all support the accuracy of Eq. (95).
In general, one is also interested in the scaling behavior of the external stress σ
or force F for breaking the lattice and its dependence on its linear size, since in
practice not only can this quantity be measured easily, but it also provides us with
important insight into a material’s structure (that is, whether the material is strong
and difficult to break versus being weak and easily breakable, both of which depend
on its morphology and size). Since this force is, for example, proportional to Y ,
where  is a displacement or strain, a plot of the stress σ , or the force F , versus
 can be compared with the traditional stress-strain diagrams that are measured
for composite solids. However, changing the parameters of the lattice model will
result in a wide variety of such diagrams. Therefore, instead of presenting the
results for each model and lattice size separately (which would be impossible), we
may try to collapse the data for all values of L, the linear size of the lattice, onto a
single curve. If the data collapse is possible, then such diagrams possess universal
properties which could be exploited for practical purposes.
Figure 8.14 presents (Sahimi and Arbabi, 1993) the results for the triangular
lattice with sizes L = 50 and 70, in which both the central and bond-bending
forces act on the springs. The thresholds were distributed according to Eq. (70) with
ζ = 0 (a uniform distribution). However, as can be seen, the data collapse is not
complete and three distinct regimes can be discerned. The first regime represents
the initial stages of crack growth and is far from the maximum of the curve. In
this regime microcracking propagates at a relatively slow rate and is more or
less similar to a percolation process, as the springs break essentially at random. As
microcracking proceeds, one eventually arrives in the second regime in the vicinity
of the maximum in which microcracking is intense and the lattice is relatively close
to its macroscopic failure point. Beyond the maximum, the system is in the so-
called post-failure regime, and is highly sensitive (unstable) to small variations
in the applied stress or strain. The general features and shapes of these curves
are in good qualitative agreement with direct experimental measurements and
observations for brittle fracture in various types of disordered solids.
408 8. Brittle Fracture: The Discrete Approach

Figure 8.14. Collapse of stress-strain data for fracture of a triangular network with
stretching and bond-bending forces (after Sahimi and Arbabi, 1993).

To obtain quantitative information on the scaling of F with L, we may assume

a scaling form. de Arcangelis et al. (1989) assumed that
F ∼ L1 φ(/L2 ), (96)
where 1 and 2 are two presumably universal exponents, and φ(x) is the scaling
function. An equation similar to (96) was also used for data collapse in the electrical
breakdown problem [see Eq. (5.31)]. Based on their simulations of the beam and
the central-force models, and using Eqs. (70) and (71) as the distribution of the
thresholds or heterogeneities, de Arcangelis et al. (1989) claimed that 1 = 2 .
Moreover, in 2D they found that, 1 = 2  0.75. On the other hand, Arbabi and
Sahimi (1990b) and Sahimi and Arbabi (1993) utilized the following equation
L1
F ∼ φ(/L1 ), (97)
(ln L)ψ
and argued that the logarithmic corrections, although seemingly weak, are neces-
sary because their existence is predicted by approximate analytical theories [see
Eq. (86) with ψ = 1/δ]. If both 1 and ψ are varied in order to obtain the most
complete data collapse, one obtains, in 2D, 1  1 ± 0.1 and ψ  0.1. The value
of ψ is small and may thus be subject to relatively large uncertainties. Since sim-
ulation of very large networks is currently not feasible, the most accurate way of
deciding whether Eq. (96) or (97) provides a more accurate fit of the data is by
fitting the data to both equations and calculating the squared residual errors (i.e.,
the difference between the data and their predictions by the fitted equation) that
each fit produces. Values of 1 and ψ are insensitive to the parameter ζ of the
distribution of the critical threshold, Eq. (70), unless ζ → 1. The results of de
Arcangelis et al. (1989) were also insensitive to the parameters of the distributions
that they used. The estimated values of the exponents for 3D systems are (Arbabi
 8.2. Quasi-static Fracture of Heterogeneous Materials 409

and Sahimi, 1990b; Sahimi and Arbabi, 1993), 1  2 ± 0.1, and ψ  0.2, which,
together with the results for 2D lattices, suggest that for a d-dimensional system
Ld−1
F ∼ φ(/Ld−1 ), (98)
(ln L)ψ
where 0 ≤ ψ ≤ 0.2. Our unpublished simulation results for quasi-static fracture
of a BCC lattice with central forces and lattice sizes of up to L = 32 indicated
that the results do follow Eq. (98). Note that Eq. (98) has a simple interpretation:
Ld−1 is the surface area on which the external force is applied, and (ln L)ψ is the
manifestation of the sample-size effect on the fracture process.
One can also study the variations of F with Nb , the number of bonds that break
during fracture. We assume that
F ∼ L3 φ(Nb /LDf ). (99)
An equation similar to (99) was used in the electrical breakdown problem; see
Chapter 5. Equation (99) implies that if the force F is plotted versus Nb /LDf , then
the results for various lattice sizes and parameters of the model should collapse onto
each other. Figure 8.15 presents such a data collapse (Sahimi and Arbabi, 1993). de
Arcangelis et al. (1989) found for their 2D models that, 3  0.75 and Df  1.7,
consistent with their results obtained with Eq. (96). On the other hand, Arbabi
and Sahimi (1990b) and Sahimi and Arbabi (1993) found that 3  1 ± 0.05, and
Df  1.7 ± 0.1 in 2D, and 3  2 ± 0.1, and Df  2.3 ± 0.2 in 3D, consistent
with their results and Eqs. (97) and (98). Note that Df represents the fractal
dimension of the set of all the broken bonds.
As discussed by de Arcangelis et al. (1989), there are other interesting scaling
features of these models. For example, one can study how the stress at the maximum
of the diagram of the type shown in Figure 8.14 scales with the linear size L of

Figure 8.15. Collapse of the data according to Eq. (99) (after Sahimi and Arbabi, 1993).
410 8. Brittle Fracture: The Discrete Approach

the sample. Similarly, one can look at the scaling of the number of the broken
bonds Nb that corresponds to this stage of the fracture at the maximum. Finally,
the number of broken bonds at the end of the fracture process also scales with L
as LDf , with Df  1.7 in 2D.

8.2.2 Comparison with the Experimental Data

Experimental measurements of the fracture strength of composite materials with
percolation disorder have been carried out. Benguigui et al. (1987) measured the
strain and stress of a perforated metal (aluminum or copper) sheets and of a 2D
diluted elastic network near pc . The samples were prepared by two different meth-
ods. In one, random holes, with a diameter slightly larger than the length of a bond
of a square lattice drawn on the material, were punched in the material. The ma-
terial so prepared was then used in a strain-controlled experiment in which it was
elongated by applying a strain to the material and increasing it monotonically and
continuously until it failed macroscopically. The stress was measured during the
entire experiment. In the second experiment the material was prepared by cutting
the interhole bonds at random such that the hole size was smaller than the lattice
unit cell. A stress was then applied to the material and increased monotonically,
during which the elongation was measured.
Benguigui et al. (1987) found that in strain-controlled experiments the stress
necessary to break the first bond was always larger than those for breaking the
subsequent bonds, and thus fracture occurred by a cascade effect. As discussed
earlier, in strain-controlled experiments, the fracture strength is defined as the stress
necessary to break the first bond, whereas in stress-controlled experiments it is the
stress by which the sample breaks. Figure 8.16 presents their experimental data.
Moreover, Benguigui et al. (1987) found that, as pc was approached, the fracture
stress σf vanishes according to power law (77) with Tf  2.5 ± 0.4, and therefore,
as expected, Tf is not identical with the elasticity exponent f of elastic percolation
networks f (d = 2)  3.96. This estimate of Tf is completely consistent with the
Monte Carlo estimate of Sahimi and Arbabi (1993), Tf  2.42 ± 0.14, mentioned
above, and also with the theoretical bounds (81) and (82).
Benguigui et al. (1987) also found that the stress for elongation of the sample
appeared to diverge at pc according to a power law with an exponent 1.4 ± 0.2.
Similar, but less precise, experiments were carried out by Sieradzki and Li (1986)
who measured the fracture stress of a system composed of a 2mm-thick plate of
aluminum with holes punched at positions corresponding to a triangular network
of 21 rows and 20 columns. The fracture stress was determined by obtaining the
full load-displacement curve for the sample to failure.
The validity of Eqs. (85) and (86) for representing the distribution of fracture
strength of composite materials was also tested against experimental data. van
den Born et al. (1991), who measured the mechanical strengths of highly porous
ceramics, reported that the size dependence of their data was well described by
both distributions, but for the failure pressure dependence, the GD with δ = 1 was
 8.2. Quasi-static Fracture of Heterogeneous Materials 411

Figure 8.16. Logarithmic plot of the experimental data for fracture load (open circles) and
yield load (filled circles) versus p − pc (after Benguigui et al., 1987).

found to be more accurate. Figure 8.17 presents their data and their fit to the GD.
Evidently, the pre-fracture porosity of the material used in these experiments was
very low, so that the porous ceramic was far above its percolation threshold.
In another set of interesting experiments, Li and Sieradzki (1992) studied me-
chanical breakdown of random porous Au, a new material specifically designed for
their experiments. They used digital image analysis to characterize the microstruc-
tures of their samples which varied by more than two orders of magnitude in length
scale. The porous Au underwent a microstructurally controlled ductile-brittle tran-
sition. Such a transition had already been predicted by Sahimi and Goddard (1986),
based on the broadness of the heterogeneity distribution in their lattice model of
fracture, and in the numerical simulations of Kahng et al. (1988) using the scalar
(fuse) model fracture processes (see Chapter 5). These results provide support for
usefulness of the lattice models for describing quasi-static fracture processes in
composite solids. Other relevant experimental data are discussed below.

8.2.3 Percolation Versus Quasi-static Brittle Fracture

In practice, in addition to the macroscopic properties of a fractured material, one
is also interested in the distribution of its local properties, such as the distribution
of the forces that are exerted on various parts of the material. This is an important
distribution as it identifies the regions of the material that may break, which can then
412 8. Brittle Fracture: The Discrete Approach

Figure 8.17. Fit of experimental data for fracture strength of porous silicon extrudates to
the Gumbel distribution (after van den Born et al., 1991).

be used in the design of tougher or better composites. Constructing this distribution

can also help one to understand the similarities between fractured and percolation
systems. Percolation is usually a static process in which failure of a bond has
nothing to do with the stress or strain field in the lattice. On the other hand,
the growth of cracks in a disordered solid is in general a dynamic and nonlinear
phenomenon which does not occur at random, but depends upon the stress or strain
field in the solid. However, under certain experimental conditions the accumulation
of damage and the growth of cracks in a solid occur essentially at random as
in, for example, a solid material that is under rapid thermal cycling, or a solid
in which the heterogeneities are broadly distributed (Sahimi et al., 1993), e.g.,
natural rock, in which case a percolation process may be able to describe fracture.
Moreover, in most cases, percolation phenomena typically represent second-order
(continuous) phase transitions, whereas many fracture processes that take place
in nature resemble first-order phase transitions, although the precise nature of the
geometrical phase transition that occurs during fracture—first order versus second
order—remains controversial (see, for example, Zapperi et al., 1997; Andersen
et al., 1997; Moreno et al., 2000). Therefore, it is important to understand the
extent of the similarities between fracture and percolation processes, because if
there are similarities between the two, then percolation phenomena, which are well-
understood and also much easier to study, may help us gain a deeper understanding
of fracture of disordered solids.
 8.2. Quasi-static Fracture of Heterogeneous Materials 413

There are at least two ways of comparing a fractured lattice with a percolating
one. The first method is based on comparing the force distributions (FDs), i.e., the
distributions of the forces that are exerted on the bonds of the lattice, and their
moments in the two systems. We already described in Chapter 8 of Volume I (see
Sections 8.6.3 and 8.11.4) the FD of elastic percolation networks, first computed by
Sahimi and Arbabi (1989). As discussed by Sahimi and Arbabi (1993), the initial
stages of brittle fracture and percolation processes in a lattice are more or less
similar. That is, during the initial stages of fracture growth, the bonds that break
are distributed essentially at random in the lattice, unless the lattice is uniform, or
its disorder is very weak. In these initial stages, the stress enhancement at the tip
of a given microcrack is not strong enough to ensure that the next bond that breaks
would be at the tip of the present microcrack. However, as more microcracks
nucleate the effect of stress enhancement becomes stronger, and deviations from
random percolation increase. Beyond a certain point in the growth of the cracks,
there will be no similarity between the two processes. Hence, one is naturally
interested to locate the point at which a fracture process starts to deviate from
a percolation phenomenon. The key clue is already provided in the stress-strain
diagrams discussed above. Equations (96)–(98) are manifestation of finite-size
scaling which represent the fracture data for various lattice sizes up to the maximum
of the stress, beyond which it breaks down. This type of finite-size scaling is also
valid for percolation networks for any p (the fraction of the intact bonds) in the
interval pc ≤ p ≤ 1 (so long as the linear size L of the network, L < ξp , where ξp
is the correlation length of percolation), albeit with different exponents and scaling
functions. Therefore, in the type of disordered lattices that we are considering here,
fracture and percolation are more or less similar up to the maximum in the stress-
strain diagram of the fractured system, i.e., in the regime in which finite-size scaling
is applicable to the fracture data, but they are not similar beyond this point.
Similar to elastic percolation networks analyzed in Chapter 8 of Volume I, one
can also calculate the moments M(q) of the FD in a fractured lattice and study
their scaling with the lattice’s linear size. Thus, one writes

q
M(q) = nFi Fi , (100)
i

where nFi is the number of bonds that suffer a force with a magnitude Fi . Similar
to elastic percolation networks, each moment of the FD scales with the linear size
L of the fractured lattice as

M(q) ∼ L−τ̃ (q) . (101)

Herrmann et al. (1989a) and de Arcangelis et al. (1989) calculated the correspond-
ing exponents τ̃ (q) for 2D fractured lattices at two different points in the system.
One was just before the lattices failed and a sample-spanning fracture was formed.
Figure 8.18 presents their results, indicating that each moment of the FD of frac-
tured lattices scales with the linear size L of the system with a different exponent.
The second point was at the maximum of the stress, where the system is entering
414 8. Brittle Fracture: The Discrete Approach

Figure 8.18. Rescaled moments m(q) = [M(q)/M(0)]1/2 versus the linear size L of the
lattice. The left figure gives the moments at the maximum current (where there is constant
gap scaling), whereas the right figure gives the moments right before the network fails (after
de Arcangelis et al., 1989).

the post-failure regime. In this case, the exponents that characterize the moments
of the FD near pc can be obtained from one of the exponents.
To understand better the difference between percolation and fractured lattices,
one important point to remember is that, while elastic properties of percolation
networks are controlled by the low moments of their FD (for example, the elastic
moduli are proportional to the second moment of the FD), fracture properties of
the same systems are controlled by the high moments of the FD. This is due to
the fact that fracture and breakdown occur where the largest loads (for example,
stress) are concentrated in the system, and the effect of such regions is manifested
only by the high moments of the FD.

8.2.4 Universal Fixed Points in Quasi-static Brittle Fracture

We now discuss another universal, and potentially very useful, aspect of quasi-
static fracture of disordered materials. In a series of simulations, Sahimi and Arbabi
(1992) computed three properties of a lattice as it underwent deformation and
brittle fracture. First, the thresholds lc were distributed according to Eq. (70) and
the elastic modulus C11 of the lattice during fracture was measured. Next, the same
fully-connected lattice (i.e., with the same values of lc ) was used to measure the
shear modulus µ of the lattice during fracture, caused by shearing the system. This
is equivalent to fracturing two identical samples and measuring their C11 and µ
 8.2. Quasi-static Fracture of Heterogeneous Materials 415

Figure 8.19. Dependence of the ratio r = C11 /µ on the fraction of unbroken bonds during
fracture of a triangular lattice with stretching and bond-bending forces. The results, from
top to bottom, are for γ /α = 0.0, 0.01, 0.3 and 1 (after Sahimi and Arbabi 1992).

independently. In Figure 8.19 we present typical results for the ratio r = C11 /µ
as a function of the fraction of unbroken springs, for various values of γ /α, the
ratio of the elastic constants of the bond-bending model [see Eqs. (66) and (67)].
The last points of these curves represent C11 /µ right before the system failed
macroscopically. We refer to this as the incipient fracture point (IFP). As can be
seen, even though the initial states of the systems (i.e., their initial values of r
with no spring broken) are different, they all approach the same value of r as
the IFP is approached. Note also that, initially r remains essentially constant,
implying that the initial value of r is not sensitive to the presence of a few cracks
or even a collection of localized cracks. However, as damage accumulates and
the cracks grow, a turning point (TP) appears and r changes drastically. Because
γ /α = 0 corresponds to a lattice in which only central forces are present, Figure
8.19 indicates that this behavior is independent of the microscopic force law of the
system. The behavior of the system for γ /α = 1 is particularly interesting. Initially,
r remains essentially constant. However, as damage accumulates a TP appears
beyond which r decreases and reaches a minimum. But near the IFP, r rises again
416 8. Brittle Fracture: The Discrete Approach

and approaches its value at the IFP which appears to be universal. Simulations
(Sahimi and Arbabi, 1992) indicated that the value of r at IFP is universal and
independent of γ /α and ζ , the parameter of the threshold distribution, Eq. (70)
(unless, of course, ζ → 1). For 2D isotropic lattices (for example, a triangular
lattice) simulations indicated that
C11 5
 . (102)
µ 4
The emergence of a universal fixed point at the IFP may mean that in a disordered
solid that undergoes quasi-static brittle fracture, the approach of r to its universal
value at the IFP may be interpreted as the “signature” of the material’s approach
to its global failure point. Although Figure 8.18 indicates that for certain values of
γ /α one may have a non-monotonic variation of r with the accumulated damage
(which, from an experimental view, implies that the closeness of r to its universal
value cannot be used for detecting the approach of the system to its global failure
point), for most real materials one has γ /α ≤ 0.3, and for such values of γ /α the
approach of r to its value at the IFP is always monotonic.
What is the theoretical explanation for the apparent universality of r? It is not
difficult to show that the dependence of C11 and µ on the fraction of unbroken
springs, as the IFP is approached, is similar to each other. As such, r represents an
amplitude ratio, and it is known (Aharony, 1980) from statistical mechanics that
certain amplitude ratios are universal. The apparent universality of r may mean
that, much like renormalization group theory of critical phenomena, universal
fixed points may be used for classifying various fracture processes. To do this,
we recall from Section 8.11.8 of Volume I that, it has been suggested (Bergman
and Kantor, 1984; Schwartz et al., 1985; Arbabi and Sahimi, 1988) that, in elastic
percolation networks C11 /µ takes on a universal value as pc is approached. For 2D
isotropic percolation networks near pc one has, C11 /µ  3, which is different from
the corresponding value for brittle fracture considered here. Sahimi and Goddard
(1986) suggested that brittle fracture of very heterogeneous networks is more or
less similar to a percolation process, and Roux et al. (1988) presented evidence
that in the limit of infinite disorder [i.e., the limit ζ → 1; see Eq. (70)] the quasi-
static lattice models of brittle fracture represent a type of percolation process.
For example, fracture in natural rock, a highly heterogeneous solid with scale-
dependent properties, may be a realization of this (Sahimi et al., 1993). On the
other hand, in most solids disorder is finite, and simulations indicated that even
for ζ = 0.8 the value of r at IFP is very different from that of elastic percolation
networks at pc , indicating that the limit ζ = 1 may be a type of singular point,
so that even for ζ = 1 − δ(δ  1), one should still obtain the value of r at the
IFP described here and not that of percolation networks at pc . Note that for 2D
isotropic systems, the Poisson’s ratio is given by νp = 1 − 2/r, implying that for
isotropic materials at the IFP, νp takes on a universal value.
Therefore, it has been proposed (Sahimi and Arbabi, 1992) that, value of the
Poisson’s point at the IFP may be used to classify various universality classes
 8.2. Quasi-static Fracture of Heterogeneous Materials 417

of quasi-static brittle fracture processes in disordered solids. Specifically, it has

been proposed that there are two distinct universality classes. One is for weakly-
disordered materials that are under a uniform external load (stress or strain). In
such solids the growth of a crack at a point depends on the environment around
that point, and therefore the damage accumulation is not at random. The univer-
sality class of such solids is described by the fixed point described by Eq. (102).
Examples include most engineering solid materials that are typically not too het-
erogeneous, with a few defects, or small (laboratory-scale) pieces of rock that
are microscopically disordered but macroscopically homogeneous with no large
scale variations in their elastic properties. The second fixed point is for systems
in which damage accumulates essentially at random. Such solids, which include
highly heterogeneous media, belong to the universality class of the fixed point of
elastic percolation networks at the pc . Examples may include natural rock at large
length scales (of the order of a few hundred meters or more) with spatially-varying
properties, and solids that undergo rapid thermal cycling.
An important question is whether the universality of the Poisson’s ratio at the
IFP can be tested experimentally. One system in which these ideas can be tested
is natural rock. It has been suggested (Sahimi and Arbabi, 1992,1996; Sahimi et
al., 1993; Robertson et al., 1995) that the quasi-static models of brittle fracture
that we have described so far may also describe fracture of natural rock, since
rock fracture is an extremely slow process. If so, then the fracture pattern and the
universal fixed point predicted by the model should be observable in rock. This
is in fact the case. Natural rock contains large fractures, in the form of a complex
and interconnected network. Despite their obvious significance, characterization
of fractured rock, and how the fractures are formed and become connected is not
as well-developed as that of unfractured porous media. However, this is changing
very fast now and such ideas as scaling, fractals, and percolation concepts are
beginning to find their proper place in the field of characterization of fractured
rock (for reviews see, for example, Sahimi, 1993b, 1994a, 1995b).
One of the first systematic studies of fractured rock was carried out by Barton
and co-workers (Barton and Larsen, 1985, Barton et al., 1987; Barton and Hsieh,
1989; Barton, 1992) as part of the effort by the United States Geological Survey to
characterize the geologic and hydrologic framework at Yucca Mountain, Nevada,
which is being considered by the United States Department of Energy as a poten-
tial underground repository for high-level radioactive wastes. Barton and Larsen
(1985) developed the pavement method of clearing a subplanar surface and map-
ping the fracture surface in order to measure connectivity, trace length, density and
fractal scaling of the fractures, in addition to their orientation, surface roughness
and aperture. Each of these parameters is important in predicting the hydraulic
characteristics of the network and in working out the history of its development
in relation to the regional tectonics. The most significant observation of the Yucca
Mountain study was that the fractured pavements have a fractal geometry, i.e., the
fracture pattern is scale-invariant. Thus, it was possible to represent the distribution
of the fractures ranging from 20 cm to 20 m by the fractal dimension Df defined
418 8. Brittle Fracture: The Discrete Approach

as
log N
Df = (103)
log(1/)
where N is the number of fractures of length . Using Eq. (103) (the standard
box-counting method described in Chapter 1), fractal dimensions of the fractured
surfaces at Yucca Mountain were found to be in the range 1.5–1.7.
Another study was undertaken for the Geysers geothermal field in northeast
California (Sahimi et al., 1993). This field, from which heat and vapor are extracted
for use in power plants generating electrical power, covers an area of more than
35,000 acres and is one of the most significant geothermal fields in the world. Using
the box-counting method, Sahimi et al. (1993) determined the fractal dimension of
the fracture surfaces of the Geysers field and found, as did Barton and co-workers,
that at small length scales the fracture pattern is fractal with Df  1.5 − 1.7,
whereas at much larger length scales, Df  1.9. These results were interpreted
with the help of the quasi-static models of fracture described above (Sahimi et al.,
1993). In particular, note that this range of fractal dimension Df is essentially the
same as what one finds with the quasi-static lattice models of fracture.
There is also convincing experimental evidence indicating that in fractured rock
the Poisson’s ratio νp may take on a universal value at the IFP, in agreement with
the prediction of the quasi-static lattice models of brittle fracture described above.
The existence of a universal fixed point in fractured rock can be directly tested by
experimental measurements, since for 3D systems
$ %
Vp C11 1 1/2
= + , (104)
Vs µ 3
where Vp and Vs are the velocities of the shear and compressional waves in the
medium, respectively, which can be measured by established experimental proce-
dures (Brace and Orange, 1968). Sammonds et al. (1989) fractured four sandstone
samples at four different confining pressures, and measured Vp and Vs . Differ-
ent confining pressures result in different fracture patterns since they control the
closure of pre-existing cracks and nucleation and growth of new microcracks.
At the three lowest confining pressures the corresponding fracture patterns were
found to be brittle-like, and from their results one finds (Sahimi and Arbabi, 1992)
that Vp /Vs  1.14 ± 0.04 at the IFP for all the three fractured sandstones, imply-
ing a universal value for C11 /µ. At the highest confining pressure fracture was
ductile-like, and although, as expected, the stress-strain diagrams of the sample
was not similar to that of brittle fracture, their results indicated that even for this
case Vp /Vs  1.1, beyond the point at which stress became independent of strain
(which is typical of stress-strain diagrams of ductile fracture), consistent again
with the value for brittle fracture at the lower confining pressures. These data pro-
vide strong experimental support for the existence of universal fixed points at the
IFP. Note that since for 3D isotropic systems, we have
3(C11 /µ − 1)
νp = , (105)
2 + 6(C11 /µ − 1)
 8.3. Dynamic Brittle Fracture 419

the experimental data of Sammonds et al. (1989) imply that for their samples, νp 
0.1 at the IFP. On the other hand, if we use C11 /µ  4/3, the corresponding value
for 3D isotropic elastic percolation networks at pc , we find νp  1/4. Thus, over
the length scale used in these experiments, the sandstones examined by Sammonds
et al. (1989) must have been relatively homogeneous, since their Poisson’s ratio at
the IFP (νp  0.1) is different from that of percolation networks at pc (νp  1/4),
which corresponds to highly heterogeneous systems. These results also support the
validity of classifying fracture processes according to the value of the Poisson’s
ratio at the IFP.

8.3 Dynamic Brittle Fracture

As discussed in Chapter 7, until a few years ago, there was a classical unsolved
problem in dynamic fracture: While classical analysis based on the linear contin-
uum mechanics would predict that brittle fracture in materials should speed up until
it reaches the Rayleigh wave speed cR , experiments indicated that the velocity of
fracture propagation never reaches this limit. It typically reaches about 40% of the
limit, and almost never more than 60% of it. The tip of the crack also heats up by
hundreds of degrees (see, for example, Green and Pratt, 1974; Fuller et al., 1975)
and, moreover, it emits high-frequency waves (Gross et al., 1993). Despite many
attempts and many proposed mechanisms for these experimental observations, the
problem remained unsolved for many years. Lattice models of quasi-static fracture
described and discussed earlier in this chapter cannot obviously shed any light on
the resolution of these issues. However, a few lattice models of dynamic frac-
ture, that have been developed over the past few years to address this problem,
have provided definitive answer to this classical problem. These models are either
probabilistic, in the sense that a bond of the lattice is broken with a certain prob-
ability, or are fully deterministic and therefore, in this sense, are similar to what
was described earlier in this chapter.
Why can such lattice models of dynamic fracture be useful? Aside from our
experience with lattice models of quasi-steady fracture which provided significant
insight into such phenomena, and in contrast to continuum models of cohesive
zone, where the correct starting equations are not yet known with certainty, and
instabilities that are in qualitative agreement with experiments are difficult to pre-
dict, calculations in a lattice or a crystal provide a framework within which the
starting point is unambiguous, and instabilities resembling those seen in experi-
ment arise naturally. Here, we describe and discuss some of the most important
theoretical results relating to the instability predicted by such models. If the lat-
tice contains no disorder, then it is possible to study the motion of a crack in a
macroscopic sample, but not describing the motion of every atom in detail. In
this way, questions about the behavior of cohesive zone and the precise nature of
crack motion can be resolved without any additional assumptions. A surprising
fact is that it is possible to obtain a large variety of analytical results for fracture
in arbitrarily large systems. Furthermore, the qualitative lessons following from
these calculations also seem to be quite general.
420 8. Brittle Fracture: The Discrete Approach

One may wonder about the appropriateness of such models for the dynamic
behavior of a crack in a crystal. The critical question, which could have also been
asked about the lattice models of quasi-static fracture is, is the lattice essential,
or can one make the underlying lattice go away by taking a continuum limit? To
our knowledge, all attempts for describing the cohesive zone of brittle materials
in a continuum framework have run into severe difficulties (see, for example,
the discussion by Langer and Lobkovsky, 1998). These problems do not arise in
lattice models described here, nor are they encountered in the atomic-scale MD
simulations that will be described in Chapter 9.
The simplicity of the ideal brittle crystal is somewhat misleading in a number
of respects. Therefore, before introducing the models, a few natural questions
regarding the generality of their predictions are posed and commented on.
(1) Does the simple force law employed between the nodes of the lattice (the
atoms in the crystal) neglect any essential aspect of the dynamics? Experience
with the MD simulations of dynamic fracture, to be described in Chapter 9,
indicates that the same qualitative results are also observed in a brittle ideal
crystal (that is, one without disorder). Moreover, we already know that, the
same type of force law, when utilized in the quasi-static case, produced very
insightful results.
(2) The most interesting calculations that have been carried out so far (see be-
low) are for a strip geometry. Is this geometry too restrictive? The general
formulation of fracture mechanics provides an answer to this question, since
it tells us that as long as the conditions of small-scale yielding are satisfied,
the behavior of a crack is entirely governed by the structure of the stress fields
in the near vicinity of the crack tip. These fields are solely determined by the
flux of energy to the crack tip. A given energy flux can be provided by an
infinite number of different loading configurations, but the resulting dynamics
of the crack will all be the same. As a result, the specific geometry used to
load the system is irrelevant and no generality is lost by the use of a specific
loading configuration. This fact is born out not only by the experimental work
described in Chapters 6 and 7, but also by the lattice models of quasi-static
fracture described earlier in this chapter.
(3) Are the predictions for a perfect (without disorder) crystal or lattice relevant
to amorphous materials? This is still an open question. However, the results of
the lattice calculations appear to be remarkably robust. Adding weak quenched
disorder to the crystal has little qualitative effect (recall that the same conclu-
sion was reached with the lattice models of quasi-static fracture). The effects
of topological disorder are not known. However, it is well-known (see, for
example, Holland and Marder, 1998) that when the temperature of a brittle
crystal increases above zero, the velocity gap (see below) becomes narrower,
and its behavior is reminiscent of what is seen in experiments performed on
amorphous materials, which were described in Chapters 6 and 7.
One of the earliest exact calculation with lattice models of dynamic fracture
was carried out by Kulakhmetova et al. (1984) who showed that it is possible to
 8.3. Dynamic Brittle Fracture 421

find exact analytical expressions for Mode I fractures moving in a square lattice.
They derived exact relationships between the energy flux to a crack tip and its
velocity, observed that phonons must be emitted by moving cracks, and calcu-
lated their frequencies and amplitudes. Later calculations by Marder and Gross
(1995) extended these results to other lattice geometries, allowed for a general
Poisson’s ratio, showed that there is a minimum allowed crack velocity (which
was found when steady fracture motion was linearly stable), computed the point
at which steady motion becomes unstable to a branching instability, and estimated
the spacing between branches. These calculations are extremely elaborate, with
the analytical expressions being so lengthy that they do not even fit on printed
pages (see Gross, 1995)! This is particularly true about the calculations for Mode
I equations. For this reason, we summarize some of the results for Mode I, and
then proceed to describe in detail how the calculations are carried out in the case
of anti-plane shear, Mode III, where the algebra is much less demanding, but most
of the ideas are the same. But, let us first describe a relatively simple lattice model
of dynamic fracture which could predict some aspects of dynamic fracture.
One of the earliest numerical simulation of dynamic fracture, based on a lattice
model, was carried out by Mori et al. (1991). A triangular lattice was used in their
work, each bond of which was assumed to be a Hooke’s spring, if the threshold
for its breaking was not exceeded. They assumed that a spring breaks irreversibly
if it is stretched beyond a given threshold. Each spring was also characterized by
a spring constant α. The nodes of the lattice were occupied by particles of mass
m. Suppose that e is the elongation vector of the springs that are connected to a
particle at position Ri . The equation of motion for the particle at time t is
d 2 Ri dRi
m = −D − ke, (106)
dt 2 dt
where the first term on the right-hand side is a damping term with D being the
damping constant; this term essentially represents some type of friction. D is not
a parameter that can be measured easily in any experiment, and thus should be
treated as a free parameter of the model. Setting dRi /dt = vi (where vi is the
velocity of node i), we obtain two equations that govern Ri and vi which, when
written in a finite-difference form, are given by
$ %
D
vi (t + t) = 1 − t vi (t) − t ke, (107)
m
t
Ri (t + t) = Ri (t) + vi (t), (108)
m
where t is the time increment.
To begin the simulations, an initial microcrack is inserted in the middle of the
lattice. As the initial condition, the lattice is stretched by an amount . Equations
(107) and (108) are then solved at time t and the springs are examined to identify
those that have exceeded their threshold. Such springs are broken, the time t is
advanced by t, Eqs. (107) and (108) are solved again, the next springs(s) is (are)
broken, and so on. An important parameter of the model is I = 0 /, where 0
422 8. Brittle Fracture: The Discrete Approach

Figure 8.20. Fracture pattern in the dynamic model of Mori et al. (1991) for I = 1.0083.
The arrow indicates the location of the initial microcrack.

is the initial length of the springs, and  is the initial amount of stretching that
the lattice has suffered. Increasing I is, in some sense, equivalent to increasing
the temperature of the system. Complex fracture pattern can emerge, depending
on the value of I . It was found that for I < 1.0085 the fracture pattern was tree-
like, and in fact no microcrack was formed if no initial crack was inserted in the
system. However, for I > 1.0085 the microcracks became connected and formed
a network and, moreover, even with no initial microcrack in the system, fractures
were formed “spontaneously.” Figure 8.20 shows the typical fracture pattern for
I = 1.0083.
In what follows, we describe the exact calculations carried out by Marder and
co-worker. Some of our discussions follow closely that of Fineberg and Marder
(1999), while the rest are based on the review by Sahimi (1998).

8.3.1 Dynamic Fracture in Mode I

The model consists of a triangular lattice of atoms with a lattice constant a. Let
ui be the displacement of a mass point from its equilibrium location, and assume
that the energy of the material is a sum of two-body terms, i.e., those that depend
on two particles at a time. We then linearize the energy to lowest order in particle
displacements. Translational invariance of the lattice implies that the force between
particles 0 and 1 depends only upon U1 = u1 − u0 . However, the force can be a
general linear functional of U1 . One way of writing down an expression for such
a general linear functional is to decompose the force between particles 0 and 1
 8.3. Dynamic Brittle Fracture 423

into a component along d̂ 1 , which is along the line that connects two neighboring
atoms, and a component that is along d̂⊥1 which is perpendicular to it. Hence, the
first component corresponds to central forces between atoms, whereas the second
component is a non-central force. Recall from Chapter 8 of Volume I (see also
Chapter 9) that in real materials non-central forces between atoms are the rule
rather than the exception. Suppose now that the restoring force parallel to the
direction of equilibrium bonds is proportional to F , while that perpendicular to
this direction is proportional to F⊥ . Therefore, if U1 = (U1x , U1y ), then the force
due to the displacement of the particle along U1 = ui−1,j +1 − ui,j is given by
 √  √ 
−1 3 −1 3
F d̂ 1 (U1 · d̂ 1 ) + F⊥ d̂⊥1 (U1 · d̂⊥1 ) = F⊥ , U1x , U1y
2 2 2 2
√  √ 
3 1 3 1
+F⊥ , U1x , U1y . (109)
2 2 2 2
If, by a similar method, one adds up contributions from other particles, one obtains
for the force due to neighbors


6

F(m, n) = Fq d̂qj [Uj (m, n) · d̂qj ] (110)
j =1 q= ,⊥

By varying the constants F and F⊥ , one can obtain any desired values of shear
and longitudinal wave speeds, which are given by
3a 2
cl2 = (F⊥ + 3F ), (111)
8m
3a 2
ct2 = (3F⊥ + F ), (112)
8m
where m is the mass of each particle. In addition to the forces between the neigh-
boring atoms, it is possible to take into account the effect of complex dissipative
functions that depend upon particle velocities. Marder and Gross (1995) added a
term to the equations so as to reproduce the experimentally-measured frequency
dependence of sound attenuation in Plexiglas. There is a slight technical restriction
in the calculations of Marder and Gross (1995) in that, forces right on the crack
line are required to be central. However, more detailed calculations indicate that,
when this technical limitation is removed, the results do not change significantly.
Figure 8.21 shows schematics of steady-state propagation of a crack in the lattice
loaded in Mode I.
Detailed, lengthy and difficult calculations show that, in Mode I loading, many
details of the relation between loading and fracture velocity depend upon ratios of
the sound speeds and also the frequency dependence of dissipation, implying that
there exists no universal curve that can describe Mode I fracture. When only central
forces are present (F⊥ = 0), it is difficult (but not impossible) to have steady-state
fracture propagation, implying that the range of loads for which cracks can propa-
424 8. Brittle Fracture: The Discrete Approach

Figure 8.21. Schematics of steady-state motion of fracture moving in an ideal brittle

crystalline material under Mode I loading (after Fineberg and Marder, 1999).

gate in a stable fashion is small, and depends upon the amount of dissipation. Thus,
the presence of non-central forces is essential to having stable fracture growth. This
also has important implications for MD simulations of dynamic fracture propa-
gation in materials. For example, the proper form of the interatomic potential for
representing the non-central forces, such as the bond-bending ones, is an important
issue in the MD simulations which will be discussed in detail in Chapter 9.
Given the complexities of exact calculations for Mode I fracture, it is wiser to
consider a simpler geometry for which the analytical methods can be discussed in
complete detail, and all the essential ideas that are needed for understanding Mode
I can be explored with much less elaborate calculations. This is the subject of the
next section.

8.3.2 Dynamic Fracture in Mode III

The main results described in this section are for the relation between loading
and fracture velocity, the velocity gap, and the calculation of the point at which
steady crack motion becomes unstable. Calculations of Marder and Gross (1995)
indicate that steady-state fractures, when they exist, are always linearly stable, and
therefore their stability analysis will not be discussed.
We consider a fracture moving in a strip composed of 2(N + 1) rows of mass
points. Figure 8.22 presents the schematics of the system. The bonds are linear
springs until they break at a separation of 2uf . The location of each mass point
is described by u(m, n), interpreted as the height of mass point (m, n) into or out
of the page, where m takes on integer values, while n takes values of the form
2 , 2 , · · · , N + 2 . The force between adjacent mass points is determined by the
1 3 1
 8.3. Dynamic Brittle Fracture 425

Figure 8.22. Dynamic fracture of triangular crystal in anti-plane shear (after Fineberg and
Marder, 1999).

difference in the height between them. Hence, in terms of u(m, n), the equation
of motion of the system is given by

d 2u du 1

2
= −b + F[u(m , n ) − u(m, n)], (113)
dt dt 2  
m ,n

where b is the coefficient of a small dissipative term. However, other functional

forms for the dissipation term can be used. The sum in Eq. (110) is over all nearest
neighbors (m , n ) of (m, n). Here,

F(u) = u#(2uf − | u |), (114)

describes ideal brittle springs, with # being the step function. The boundary
condition is given by
1
u[m, ±(N + )] = ±uN . (115)
2
It is important to determine the value of uN for which there is just enough stored
energy per unit length to the right of the fracture to break the pair of bonds connected
to each lattice site on the crack line. For m  0 one has
nuN
u(m, n) = , (116)
N+ 1
2

and the height difference between mass points with adjacent values of n is
nN
Uright = . (117)
N+ 1
2
426 8. Brittle Fracture: The Discrete Approach

Therefore, the energy stored per unit length in the (2N + 1) rows of bonds is
1
× (2 upper bonds/site) × (rows) × (spring constant) × Uright
2
2  2
1 1 uN
= × 2(2N + l) × × = 2N0 u2N , (118)
2 2 N+2 1

where N0 = 1/(2N + 1). The energy required to break two bonds each time the
crack propagates by a unit length is given by
1
× (2 bonds/site) × (spring constant) × (separation at fracture)2
2
1 1
= × 2 × × (2uf )2 = 2u2f . (119)
2 2
Therefore, by equating Eqs. (118) and (119) one obtains the proper dimensionless
measure of the external driving force,
√
uN N0
= , (120)
uf
which reaches 1 as soon as there is enough energy to the right of the crack to break
the bonds along the crack line. Note that  is linearly related to the displacement
uN imposed at the edges of the strip.
Slepyan (1981) and Kulakhmetova et al. (1984) developed the techniques for
solving problems of this type. However, there are some differences between details
of their solution and what is presented here, because Eq. (113) describes motion
in a strip, which is what is considered here, rather than an infinite plate that they
studied. Both geometries have certain advantages. The strip is preferable to the
infinite plate if one wishes to compare the predictions with the results of numerical
simulations. On the other hand, using the infinite plate results in certain natural
limits. Moreover, the lattice considered here is a triangular rather than a square
lattice used by Slepyan and co-workers.
We assume that a fracture moves in steady state, so that one by one, the bonds
connecting u(m, 12 ) with u(m + 1, − 12 ) or u(m, − 12 ) break because, as a result of
the driving force described by Eq. (115), the distance between these atoms exceeds
the limit set by Eq. (115). Assuming that the times at which the bonds break are
known, the original nonlinear problem is immediately transformed into a linear
problem. However, once the solution of the linear problem has been derived, one
must verify that,
(1) bonds break at the time that they are supposed to. If we impose this condition,
we obtain a relation between the crack velocity v and dimensionless loading .
(2) Conversely, no bonds break when they are not supposed to. This condition
results in a velocity gap on the low end of the velocity range, and leads to
crack tip instabilities above a critical energy flux (see below).
We note here that steady states in a perfect lattice or crystal are more complex
than those in a continuum, since in the latter case a steady state acts as u(x, vt),
 8.3. Dynamic Brittle Fracture 427

whereas the closest to such a state that one can come in a triangular lattice is by
having the symmetry
u(m, n, t) = u(m + 1, n, t + 1/v), (121)
and also
 $ % 
1
u(m, n, t) = −u m, −n, t − − gn /v , (122)
2
which implies in particular that
 
1 1
u(m, , t) = −u m, − , t − 1/(2v) . (123)
2 2
Here
⎧
⎪
⎪ 0 if n = 12 , 52 · · ·
⎨
gn = 1 if n = 32 , 72 · · · (124)
⎪
⎪
⎩
mod(n − 12 , 2) in general.
By assuming that a fracture is in steady state, one can eliminate the variable m
entirely from the equation of motion. We define
un (t) = u(0, n, t), (125)
and write the equations of motion in steady state as
⎡ ⎤
2
un+1 [t − (gn+1 − 1)/v] +un+1 (t − gn+1 /v)
d un 1⎢ ⎥ dun
= ⎣+un (t + 1/v) −6un (t) + un (t − 1/v)⎦ − b
dt 2 2 dt
+un−1 [t − (gn−1 − 1)/v] +un−1 (t − gn−1 /v)
(126)
if n > 1/2. For n = 1/2 we have
d 2 u1/2
=
dt 2
⎡ ⎤
u3/2 (t) +u3/2 (t − 1/v)
1⎢ ⎢+u1/2 (t + 1/v) − 4u1/2 (t)
⎥
⎥ − b du1/2
+u1/2 (t − 1/v)
2⎣ ⎦ dt
+[u−1/2 (t) − u1/2 (t)]#(−t) +[u−1/2 (t − 1/v) − u1/2 (t)]#[1/(2v) − t]
(127)
We have assumed that t = 0 is the time at which the bond between u(0, 12 , t)
and u(0, − 12 , t) breaks, and therefore, by symmetry, the time at which the bond
between u(0, 12 , t) and u(1, − 12 , t) breaks is 1/(2v).
Above the fracture line, Eqs. (126) and (127) are completely linear, and thus
it is not difficult to derive the solution that yields the motion of every atom with
n > 1/2 in terms of the behavior of an atom with n = 1/2. If we Fourier transform
428 8. Brittle Fracture: The Discrete Approach

Eq. (126), we obtain

⎧ 1 ⎫
⎪
⎪ un+1 (ω)[eiω(gn+1 −1)/v + eiωgn+1 /v ] ⎪
⎪
⎨ 2 ⎬
−ω2 un (ω) = + 12 un (ω)(eiω/v − 6 + e−iω/v ) + ibω (128)
⎪
⎪ ⎪
⎪
⎩ 1 ⎭
+ 2 un−1 (ω)[eiω(gn−1 −1)/v + eiωgn−1 /v ]
where ω is the Fourier transform variable conjugate to the time t. If
un (ω) = u1/2 (ω)ek(n−1/2)−iωgn /(2v) (129)
then by substituting Eq. (129) into Eq. (128), and using the fact that gn + gn+1 = 1,
we obtain
−ω2 u1/2 (ω) =
⎧ k iω(gn+1 +gn−2 )/(2v) + eiω(gn+1 +gn )/(2v) ]
⎫
2 u1/2 (ω)e [e
1
⎪
⎪ ⎪
⎪
⎨ ⎬
+ 12 u1/2 (ω)(eiω/v − 6 + e−iω/v ) + ibω u1/2 (ω)
⎪
⎪ ⎪
⎪
⎩ ⎭
+ 12 u1/2 (ω)e−k [eiω(gn−1 +gn−2 )/(2v) + eiω(gn−1 +gn )/(2v) ]
(130)
so that
ω2 + ibω + 2 cosh(k) cos[ω/(2v)] + cos(ω/v) − 3 = 0. (131)
Let
3 − cos(ω/v) − ω2 − ibω
z= , (132)
2 cos(ω/2v)
which is equivalent to
!
y =z+ z2 − 1, (133)
with y = ek .
One now constructs a solution which meets all the boundary conditions by
writing
 [N +1/2−n] 
y − y −[N +1/2−n] uN (n − 12 ) 2α
un (ω) = u1/2 (ω)e−iωgn /2v −N
+
y −y
N N α + ω2
2

(134)
This solution equals u1/2 for n = 1/2, and equals 2αuN /(α 2 + ω2 ) for n = N + 12 .
Introducing α is necessitated by the fact that for n = N + 12 , u(m, n, t) = uN .
Instead of working with the Fourier transform of this boundary condition, which
is a delta function and difficult to work with, it is better to use the following
boundary condition
uN +1/2 (t) = uN e−α|t| , (135)
and let α → 0 at the end of the calculation. In the analysis that follows, frequent
use is made of the fact that α is small.
 8.3. Dynamic Brittle Fracture 429

From a physical point of view, the most interesting variable is not u1/2 , but the
distance between the bonds which will actually break. Hence, we define
u1/2 (t) − u−1/2 (t) u1/2 (t) + u1/2 (t + 1/2v)
U (t) = = , (136)
2 2
to rewrite Eq. (127) as
⎡ ⎤
u3/2 (t) +u3/2 (t − 1/v)
2 1⎢
d u1/2 ⎥ du1/2
= ⎣+u1/2 (t + 1/v) − 4 +u1/2 (t) + u1/2 (t − 1/v) ⎦−b .
dt 2 2 dt
−2U (t)#(−t) −2U (t − 1/2v)#[1/(2v) − t]
(137)
If we Fourier transform this expression, use Eq. (134), and define
 ∞
U ± (ω) = U (t)#(±t) exp(iωt)dω, (138)
−∞

we obtain
uN 2α
u1/2 (ω)F (ω) − (1 + eiω/2v )U − (ω) = − , (139)
N ω + α2
2

where
 
y [N −1] − y −[N −1]
F (ω) = − 2z cos(ω/2v) + 1. (140)
y N − y −N
We now use Eq. (136) in the form
1
U (ω) = (1 + e−iω/2v )u1/2 (ω) (141)
2
to obtain
uN 2α
U (ω)F (ω) − 2 cos2 (ω/4v)U − (ω) = − . (142)
N ω + α2
2

If we write
U (ω) = U + (ω) + U − (ω) (143)
we finally obtain
$ %
1 1
U + (ω)Q(ω) + U − (ω) = uN N0 + , (144)
α + iω α − iω
where
F (ω)
Q(ω) = . (145)
F (ω) − 2 cos2 (ω/4v)
To derive the right-hand side of Eq. (144), we used the facts that F (0) = −1/N ,
and that α is very small, so that the right-hand side of this equation is a delta
function.
430 8. Brittle Fracture: The Discrete Approach

We now utilize the Wiener–Hopf technique (see, for example, Noble, 1958) to
write
Q− (ω)
Q(ω) = , (146)
Q+ (ω)
where Q− and Q+ are free of poles and zeroes in the lower and upper complex
ω planes, respectively. This decomposition can be carried out with the explicit
formula
  
1 ln Q(ω )
Q± (ω) = exp lim dω 
. (147)
ε→0 2π iω ∓ ε − iω
We now separate Eq. (144) into two parts, one of which has poles only in the lower
half plane, while the other part has poles only in the upper half plane:
U + (ω) u N N0 1 u N N0 1 U − (ω)
− = − . (148)
Q+ (ω) Q− (0) −iω + α Q− (0) iω + α Q− (ω)
Because the right- and left-hand sides of Eq. (148) have poles in opposite sections
of the complex plane, they must separately equal a constant. However, the constant
must be zero, because otherwise U − and U + will behave as a delta function near
t = 0, and therefore
uN N0 Q− (ω)
U − (ω) = , (149)
Q− (0)(α + iω)
and
uN N0 Q+ (ω)
U + (ω) = , (150)
Q− (0)(α − iω)
which provide an explicit solution for U (ω). Numerical evaluation of Eq. (147),
and U (t) using Eqs. (149) and (150) is fairly straightforward (using, for example,
fast Fourier transforms). However, in carrying out the numerical transforms, one
must carefully analyze the behavior of the functions for large values of ω. If these
functions decay as 1/(iω) [the inverse Fourier transform of which is #(t)], we
should subtract the 1/(iω) off before the numerical transform is performed, after
which we should add this term back with the appropriate step function. Conversely,
if the functions to be Fourier transformed have #(t) discontinuity, it is best to
subtract off the appropriate multiple of et #(t) before Fourier transforming, and
then add on the appropriate multiple of 1/(1 − iω). A solution of Eqs. (149) and
(150) constructed in this manner is given in Figure 8.23.
One can now derive a relation between the dimensionless displacement  and
the crack velocity v. Recall that making the transition from the original nonlinear
problem posed by Eq. (113) to the linear problem expressed by Eq. (126) relies on
assuming that bonds along the crack line break at time intervals of 1/2v. Because
of the symmetries expressed by Eqs. (121)–(123), it is sufficient to guarantee that
u(t = 0) = uf . (151)
 8.3. Dynamic Brittle Fracture 431

Figure 8.23. Predictions of Eqs. (149) and (150) for v = 0.5, N = 9 and b = 0.01 (after
Fineberg and Marder, 1999).

All the displacements are proportional to the boundary displacement uN , and hence
Eq. (151) fixes a unique value of uN and its dimensionless counterpart, . This
means that, once one assumes that the fracture moves in steady state at a velocity
v, there is a unique  to make it possible.
To derive Eq. (151), we must require that

1
lim U − (ω) exp(−iωt)dω = uf . (152)
t→0− 2π

This integral can be evaluated by inspection. We know that for t > 0,


U − (ω) exp(−iωt) dω = 0, (153)

and that any function that for large ω behaves as 1/(iω) has a step function
discontinuity at the origin. Therefore, Eqs. (149) and (152) become

Q− (∞)
uf = uN N0 . (154)
Q− (0)

Since Eq. (145) yields Q(∞) = 1, one obtains from Eq. (147) that,

Q− (∞) = Q+ (∞) = 1. (155)

As a result, Eq. (154) and the definition of  [Eq. (120)] yield

Q− (0)
= √ . (156)
N0
432 8. Brittle Fracture: The Discrete Approach

To make this result more explicit, we use Eq. (147) and the fact that
    
1 1 ln Q(ω ) ln Q(−ω )
Q− (0) = exp + dω 
2π 2 ε − iω ε + iω
     (157)
1 1 Q(ω ) ε 
= exp ln + ln Q(0) dω .
2π −2iω Q̄(ω ) ε 2 + ω2
Therefore,
!   
1 1 Q(ω ) 
Q− (0) = N0 exp − ln dω . (158)
2π 2iω Q(ω )
Inserting Eq. (158) into Eq. (156) yields
  
1 1 Q(ω ) 
 = exp − ln dω . (159)
2π 2iω Q(ω )
In order to obtain an expression that is correct not only for Mode III model
considered here, but also for more general cases, we rewrite Eq. (159) as
  
1 1   )}dω ,
 = C exp − {ln Q(ω ) − ln Q(ω (160)
2π 2iω
where C is a constant of order unity that is determined by the geometry of the
√ example, for the triangular lattice loaded in Mode III, C = 1, whereas
lattice. For
C = 2/ 3 for the same lattice loaded in Mode I (Marder and Gross, 1995). The
advantage of Eq. (160) is that it is suitable for numerical evaluation, since there is
no uncertainty regarding the phase of the logarithm.
When b, the coefficient of the dissipation term, becomes sufficiently small, Q
is real for real ω except in the small neighborhood of the isolated roots and poles
that are near the real ω-axis. Suppose that ri+ are the roots of Q with negative
imaginary part (since they belong to Q+ ), ri− are the roots of Q with positive
imaginary part, and pi± are the poles of Q. Equation (160) can be written as
#'
=C ri− pi+ /ri+ pi− , for b → 0. (161)
i

In deriving Eq. (161) use was made of the fact that, away from a root or pole of
Q, the integrand of Eq. (160) vanishes. Together with Eqs. (149) and Eqs. (150),
Eqs. (160) and (161) constitute the formal solution of the model. Since Q is a func-
tion of the steady-state velocity v, Eq. (159) relates the external driving force im-
posed on the crystal, , to the velocity v of the crack. The results of a typical calcu-
lation for N = 9 are presented in Figure 8.24, which show clearly the velocity gap.
Having determined the formal solution of the model, we can now investigate a
few important issues.

8.3.2.1 Phonon Emission

At  = 1 just enough energy is stored to the right of the fracture tip to break all
bonds along the crack line. However, since all steady states occur for  > 1, not
 8.3. Dynamic Brittle Fracture 433

√
Figure 8.24. Fracture velocity v, normalized by sound velocity vs = 3/2, versus the
driving force , using N = 9. The end points of the lower curve indicate the linearly-stable
lattice-trapped states (after Fineberg and Marder, 1999).

all the energy stored to the right of the fracture tip will be used for breaking bonds.
What happens to the remaining energy depends on the amount of dissipation b,
and the distance from the fracture tip at which one inspects the system. In the
limit b → 0, travelling waves leave the fracture tip and carry energy off in its
wake, with the amount of energy that they carry off being independent of b. For all
non-zero values of b, however, the travelling waves will eventually decay, and the
extra energy will have been absorbed by dissipation, but the value of b determines
whether one views the process as microscopic or macroscopic.
The frequencies of the radiation emitted by the crack have a simple physical
interpretation. Consider the motion of a particle through a lattice, in which the
phonons are described by the dispersion relation, ωα (k). If the particle moves
with a constant velocity v and interacts with the various ions according to some
function F, then, to linear order, the motions of the ions can be described by a
matrix M which describes their interactions with each other as
(l)
d 2 ui


(n) (n)
m = − M ij (R (l)
− R )u j + Fi (R (n) − vt). (162)
dt 2 n
j,n

(l)
Multiplying both sides of Eq. (162) by eik·R , summing over l, and letting K and
 be, respectively, the reciprocal lattice vectors and the volume of a unit cell, yield

d 2 ui (k)
1

m 2
= Mij (k)uj (k) + exp[iv · (k + K)t]Fi (k + K). (163)
dt 
j K
434 8. Brittle Fracture: The Discrete Approach

By inspection we can show that the lattice frequencies excited in this way are those
which in the extended zone scheme satisfy the following equation (Ashcroft and
Mermin, 1976)
ω(k) = v · k. (164)
If we pretend that the crack is a particle, we can use Eq. (164) to predict the phonons
that the crack emits.
There are actually two phonon dispersion relations to consider, one for propa-
gating radiation far behind the crack tip, and the other for propagating radiation far
ahead of the tip. Far behind the crack tip, all the bonds are broken, and therefore
to find the travelling one we must set U − = 0 in Eq. (142), since all the bonds
behind the tip are broken, and also uN = 0, because phonons can propagate with-
out any driving term, which then lead us to F (ω) = 0. Similarly, because no bond
far ahead of the crack tip is broken, U − = U , and the condition for phonons is
F (ω) − 2 cos2 (ω/4v) = 0. Therefore, according to Eq. (145), the roots and poles
of Q(ω) are the phonon frequencies behind and ahead of the fracture, which are
also the quantities that appear in Eq. (161).

8.3.2.2 Forbidden Fracture Velocities

After verifying that bonds along the crack line break when they should, we must
also verify that they have not been stretched enough to break earlier than they
should have. That is, not only must the bond between u0+ and u0− reach length
2uf at t = 0, but also must be the first time at which that bond stretches to a length
greater than 2uf . For 0 < v < c with c  0.3 (the precise value of c varies with
b and N), this condition is violated. The states have the unphysical character that
masses rise above height uf for t < 0, with the bond connecting them to the lower
line of masses remaining, however, intact, and then they descend whereupon the
bond breaks.
Since the solution of Eqs. (126) and (127) is unique, but does not in this case
solve Eq. (113), no solutions of Eq. (113) exist at all at these velocities. Once the
crack velocity has dropped below a lower critical value, all the steady states will
have this character. This argument indicates that no steady state in the sense of
Eq. (121) can exist. It is also possible to search for analytical solutions that are
periodic and travel two lattice spacings before repeating, but to our knowledge no
solutions of this type have yet been found. One can numerically verify that if a
fracture is allowed to propagate with  just above the critical threshold, and  is
then very slowly lowered through the threshold, the fracture stops propagating. It
does not slow down noticeably; rather, suddenly, the moving fracture emits a burst
of radiation that carries off its kinetic energy, and stops in the space of an atom,
which is why there exists a velocity gap.

8.3.2.3 Nonlinear Instabilities

Recall that we assumed in the calculations for predicting the steady states that the
only bonds which break are those which lie on the crack path. This assumption
 8.3. Dynamic Brittle Fracture 435

can be tested by using the numerical solutions of Eqs. (149) and (150), since the
solution fails above a critical
√ value c of . The sound speed vs (= cR ) equals,
in dimensionless units, 3/2, and thus we rescale velocities by this value. For
example, for N = 9, at a velocity, vc /vs = 0.666 · · ·, c = 1.158 · · ·, and the
bond between u(0, 12 ) and u(1, 12 ) reaches a distance of 2uf shortly after the bond
between u(0, 12 ) and u(0, − 12 ) breaks. The steady-state solutions that are obtained
when the lattice is strained with larger values of  are inconsistent; only dynamical
solutions more complex than steady states, involving the breaking of bonds off
the fracture path, are possible. To investigate these states, one must numerically
solve Eq. (113). Such numerical simulations have been carried out (Marder and
Liu, 1993). The results show that just above the threshold at which horizontal
bonds begin to break, the distance between these extra broken bonds diverges.
The reason is that breaking a horizontal bond takes energy from the fracture and
slows it down below the critical value. The fracture then tries once more to reach
the steady state, and only in the last stages of the approach does another horizontal
bond break, hence beginning the process again.
A rough estimate of the distance between broken horizontal bonds can be ob-
tained as follows. Let h (t) be the length of an endangered horizontal bond. One
must view the problem in a reference frame moving with the fracture tip, and there-
fore at every time interval 1/v one shifts attention to a bond one lattice spacing
to the right. When  is only slightly greater than c , the length of such a bond,
viewed in a moving frame, should behave, before it breaks, as
∂h
h ∼ 2uf + ( − c ) − δ exp(−bt). (165)
∂
Here, ∂h /∂ denotes the rate at which the steady-state length of h would change
with , if this bond were not allowed to break, and δ describes how much smaller
than its steady-state value the bond ends up after the breaking event occurs. Marder
and Gross (1995) showed that deviations from steady states die away at long times
as exp(−bt). From Eq. (165) we can estimate the time between breaking events
by setting h = 2uf and solving the equation for t. The result is that the frequency
f with which horizontal bonds break should scale above the critical strain c as
b
f ∼− . (166)
ln[(δ)−1 ( −  c )∂h /∂]
However, the accuracy of Eq. (166) has proven to be difficult to check.

8.3.2.4 The Connection to the Yoffe’s Criterion

The basic reason for the branching instability seen above is the lattice analogue
of the Yoffe’s instability (see Section 7.6.3), but on small scales. The Mode III
calculation finds that the critical velocity for the instability to frustrated branching
events is indeed close to the value of 0.6cR predicted by Yoffe in the continuum.
However, the critical velocity seen experimentally in amorphous materials is about
1 2
3 cR , not about 3 cR . This discrepancy can be due to some combination of three
factors.
436 8. Brittle Fracture: The Discrete Approach

(1) The force law between the atoms is actually much more complex than ideal
breaking bonds. Gao (1993) has pointed out that the Rayleigh wave speed cR
in the vicinity of a crack tip may be significantly lower than its value far away
from the tip, because material is being stretched beyond the range of validity
of linear elasticity.
(2) The experiments are at room temperature, while the calculations are at zero
temperature.
(3) The experiments are in amorphous materials, while the calculations are for a
crystal or perfect lattice.
Before closing this section, let us summarize the main results obtained from the
ideal brittle crystal.
(i) For a range of loads above the Griffith point, a fracture can be trapped by
the crystal, i.e., it neither moves nor heals, although it is energetically possible for
the fracture to move (Thomson et al., 1971; Thomson, 1986). However, molecular
dynamics simulations, to be described in Chapter 9, indicate that lattice trapping
occurs at very low temperatures and in fact disappears at room temperature.
(ii) Steady-state fracture motion exists, and is a stable attractor for a range of
energy flux.
(iii) Steadily-moving fractures emit phonons with frequencies that can be
computed from a simple conservation law.
(iv) The relation between the fracture energy and velocity can be computed.
(v) The slowest steady state runs at around 0.20cR . There is no slower-moving
steady-state fracture, and therefore there is a velocity gap.
(vi) At an upper critical energy flux, steady-state fractures become unstable and
generate frustrated branching events in a fashion reminiscent of experiments in
amorphous materials described in Chapter 7.

8.3.3 The Effect of Quenched Disorder

The robustness of the branching phenomena described in the perfect crystal model,
which contained no types of heterogeneities, has been further illustrated by numer-
ical studies of a few lattice models in which the effect of disorder was explicitly
taken into account. We briefly describe some of these models and their most impor-
tant predictions. However, before doing so, let us point out that there is a qualitative
difference between quenched disorder and thermal fluctuations which is quite im-
portant to brittle fracture. If a fracture encounters a tough spot in a material, it
can stop propagating forever. However, although thermal fluctuations might halt
a fracture temporarily, they are just as likely to give energy to a static fracture and
induce it to start moving again. Lattice-trapped fractures are not completely static
in the presence of thermal fluctuations; they creep ahead with some probability
(Marder, 1996). When the rate of creep increases to speeds that are of the order of
the sound speed, the distinction between creeping and running fractures vanishes,
and therefore the velocity gap disappears.
 8.3. Dynamic Brittle Fracture 437

Furukawa (1993) modified the model proposed by Mori et al. (1991) (see Section
8.3) by including a shear friction in the model. Thus, in this model the equation of
motion is given by
dvi


m = −D1 vi − D2 (vi − vj ) + F(Rj − Ri ), (167)
dt
j j

where the second term on the right-hand side represents shear friction or dissipa-
tion, while the last term is the force term which was taken to be F(R) = Rf (R)/R,
where R is the magnitude of R. The function f (R) was selected to be f (R) =
R − 1 for R ≤ R0 , and f (R) = (R0 − 1) exp[−κ(R − R0 )] for R ≥ R0 , where
R0 and κ are two constants. A bond breaks if R > R0 . Both square and triangular
lattices were used. Fracture formation was initiated by inserting an initial microc-
rack at the center of the lattice. In one case the lattice spacing in the direction of the
macroscopic deformation was 1, and the remaining lattice distances were Re with
1 < Re < R0 . An interesting prediction of the model was that, when κ = ∞, the
fracture velocity v = /t, where  is the distance between the central microcrack
and the most distant broken bond at time t, follows a power law:
(Re − 1)x1
v∼ , (168)
(R0 − Re )x2
where the velocity has been scaled by the 1D sound velocity. In most cases
x1 = x2 = 1, except in the square lattice without the dissipation term (D2 = 0),
in which case x1 = 1 and x2 = 1/3. In this case fracture propagation was sub-
sonic, whereas it was supersonic on the triangular lattice if (R0 − Re )/(Re − 1)
was small (see Sections 7.7.1, 7.8.15, and 9.8.3.3 for discussions of supersonic
fracture propagation). In the second case that was studied, all the lattice spacings
were equal. A variety of interesting fracture patterns were obtained, some of which
are shown in Figure 8.25. Also obtained were oscillatory fracture patterns, some
of which are also shown in Figure 8.25. As discussed above and in Chapters 6 and
7, such oscillatory fracture patterns contribute to the dynamic instability observed
during fracture propagation.
Another interesting deterministic lattice model of dynamic fracture was pro-
posed by Rautiainen et al. (1995). In their model each bond ij of a square lattice
was an elastic element with an interaction energy that was described by
α γ
Hij = [(ui − uj ) · R ]2 + [(ui − uj ) · R⊥ ]2 , (169)
2 2
where ui is the displacement of node i, and R and R⊥ are the unit vectors paral-
lel and perpendicular to the vector connecting i and j in the undeformed lattice,
respectively. To include disorder in the model, a fraction q of the bonds was re-
moved at random before fracture simulations began. To introduce dynamics into
the model, a dissipation mechanism was included in the model by incorporating in
it Maxwellian viscoelasticity which allows the description of relaxation and dis-
sipation of elastic energy as a dynamical decay of the local forces. The constitutive
438 8. Brittle Fracture: The Discrete Approach

 ! "

# $ % &

Figure 8.25. Regular and oscillatory fracture patterns obtained by the dynamic model of
Furukawa (1993). The regular patterns were obtained under isotropic tension.

equation for the forces acting at each bond at time t was taken to be
dFij dFH 1
= − Fij , (170)
dt dt tr
where Fij is the force between i and j arising out of their interaction, FH is
the elastic force derived from Eq. (169), and tr is a phenomenological parameter
which can be considered as a relaxation time scale. In effect, each bond is replaced
by a Maxwellian viscoelastic element—a spring and a dashpot in series, and was
considered as broken if its length exceeded a critical value.
The model predicts brittle fracture in the limit of slow straining. At finite strain
rates, damage development becomes ductile with increasing dissipation. For small
tr and q, the number of broken bonds increases rapidly at the initial stages of the
fracture history. However, after some time damage accumulation stops and the sys-
tem resists rupture. This is due to local viscoelastic dissipation which arrests crack
growth. For large tr the number of broken bonds increases slowly, and the manner
by which the bonds break is correlated. Thus, ductility increases with decreasing
tr . Hassold and Srolovitz (1989) had already shown, using a Born and the quasi-
static lattice model described in Section 8.2, that crack arrest can be controlled
by varying α/γ . Therefore, the role of α/γ in the quasi-static model is played by
the relaxation parameter tr . The crack velocity was found to be approximately,
 8.3. Dynamic Brittle Fracture 439

Figure 8.26. Fracture patterns in the dynamic model of Heino and Kaski (1996). The top
pattern corresponds to tr = 0, while the rest are for tr
= 0.
√
v  γ . A dynamic Born model was developed by Mart́in et al. (2000) that also
produced many of these predictions.
In another lattice model of dynamic fracture, Heino and Kaski (1996) used the
same Hamiltonian as in Eq. (169), but the lattice was more complex. Each bond of
the lattice consisted of two perpendicular springs in fixed directions representing
tensile and bending behavior. A dashpot was connected in series to each spring. In
a tensile experiment, the sites move which results in elongation or shrinking of the
springs and dashpots. Disorder was introduced into the model by assuming that
each bond has its own α and distributing it according to a uniform distribution. The
ratio α/γ for each bond was held constant though, which corresponds to varying
the Young’s modulus of the bonds. A bond was considered as broken when its
strain exceeded a threshold. Interesting fracture patterns were generated by the
model by varying its parameters, some of which are shown in Figure 8.26. As can
be seen, there is a dominating fracture that eventually spans the system and causes
it to fail. However, many daughter cracks also appear essentially symmetrically
on both sides of the main fracture. They appear periodically and advance a short
distance before dissipation damps their growth. If the network is more disordered,
the periodicity disappears and the fracture pattern becomes irregular.
However, the most interesting result emerging from this model was the behavior
of the crack tip velocity as a function of the crack width. Initially, the velocity
increased rapidly, corresponding to the emergence of a straight fracture. However,
after some time the oscillatory daughter fractures appeared, and thus the velocity
also started to oscillate with the crack width, and hence with the time. Increasing
α increased the crack speed and its oscillation frequency, but decreased the length
of the daughter fractures, although the angle that they made with the main fracture
was unaffected by α. Heino and Kaski (1997) combined a finite-element method
(for discretizing the continuum elasticity equations and obtaining a finite-element
440 8. Brittle Fracture: The Discrete Approach

mesh) with the bond-breaking process used in lattice models. The model provided
the same type of predictions as those of their earlier model.
Åström and Timonen (1996) used a square lattice of beams (see Section 8.2)
to study dynamic fracture. Their model was the dynamic analogue of the lattice
models of quasi-static brittle fracture described in 8.2. The network was strained
by an amount  in the y direction. The sites at the top and bottom edges of the
network were constrained to remain at their original positions, while the sites on
the left and right edges were free to move without constraints. The dynamics of the
model was calculated using a discrete form of the Newton’s equation of motion
including a linear viscous dissipation term (see also Section 8.4 below). A beam
was considered as broken if the strain on it exceeded a pre-assigned threshold
value. Åström and Timonen (1996) showed that, by tuning the strain and the ratio
of axial to bending stiffness of the beams, a fracture can propagate either straight,
or branch, or bifurcate. The fact that such features can be obtained by both the
Born model of Heino and Kaski (1996), described above, and by the beam model
of Åström and Timonen, indicates the universality of such features. Moreover, for
the branching fracture, Åström and Timonen (1996) found that their trajectories
follow a power law,
y ∼ x 0.7 , (171)
where x and y are, respectively, the directions parallel and perpendicular to the
direction of the main crack, with the origin being the point at which the micro-
branch begins. This result is in complete agreement with the power-law observed
in experiments in both polymethylmethacrylate (PMMA) and glass (see Section
7.10.4).

8.3.4 Comparison with the Experimental Data

The predictions of lattice models of dynamic fracture are in agreement with the re-
sults of a set of spectacular experiments by Fineberg et al. (1991, 1992) and Gross
et al. (1993). Their work and many earlier experiments have reported many inter-
esting features of dynamic crack propagation in materials, which were described in
detail in Chapters 6 and 7, and are summarized here. Recall, from Chapters 6 and
7, the main features of a fracture surface in a brittle material, namely, the mirror,
mist, and hackle sequence: An initially smooth and mirror-like fracture surface
becomes misty, and then evolves into a rough hackled region. It has also been
reported (see, for example, Döll, 1975; Kusy and Turner, 1977) that in some brittle
materials, such as PMMA, the fracture pattern exhibits characteristic wavelength,
that surface roughness increases with crack speed (see, for example, Langford et
al., 1989, and references therein), and that periodic stress waves are emitted from
the tip of the rapidly moving cracks in a wide variety of materials (see, for ex-
ample, Rosakis and Zehnder, 1985; Dally et al., 1985, and references therein). As
described in detail in Chapter 7, Fineberg et al. (1991, 1992) carried out precise
experiments to study fracture propagation in brittle plastic PMMA and showed
that, there is a critical velocity vc beyond which the velocity of fracture tip begins
to oscillate, the dynamics of the fracture changes abruptly, and a periodic fracture
 8.4. Fracture of a Brittle Material by an Impact 441

pattern is formed. For v > vc the amplitude of the oscillations depends linearly
on the mean velocity of the propagating fracture. Thus, the motion of fractures is
governed by a dynamical instability, and explains why the velocity of their tip does
not attain the limiting Rayleigh velocity cR predicted by the linear elasticity theory.
In another set of beautiful experiments, Gross et al. (1993) used two materials,
the PMMA and soda-lime glass, to show that all features of dynamics of crack
propagation in the two materials, such as acoustic emission, crack velocity, and
surface structure, exhibit quantitative similarity with each other. Thus, there exist
universal characteristics of fracture energy in most materials that are the result of
energy dissipation in a dynamical instability. Perhaps the most spectacular exper-
iments were carried out by Sharon et al. (1995) and Sharon and Fineberg (1996)
using the brittle plastic PMMA. They identified the origin of the dynamical in-
stability during fracture propagation as being the nucleation and growth of the
daughter cracks which limit the speed of the propagating crack tip. The daughter
fracture carries away a fraction of the energy concentrated at the tip of the moving
crack, thus lowering the velocity of the tip. After some time, the daughter crack
stops growing, and thus the crack tip velocity increases, until a new daughter frac-
ture starts to grow, and so on. They also observed that the branching angle for a
longer daughter fracture was smaller than that of the shorter daughter fractures.
These features are all produced by the dynamic lattice models of Heino and Kaski
(1996,1997) described above. The computations of Marder and Liu (1993) for a
perfect crystal, that were described and discussed above, also agree with these data.
As already mentioned, oscillatory fracture patterns were also observed in the
experiments of Yuse and Sano (1993). They imposed a temperature gradient along
a thin glass plate, from a hot region to a cold one. A microcrack was introduced in
the glass, and the glass was pushed. As the plate started to move the crack jumped
ahead of the thermal gradient and stayed there. It was observed that if the plate
moves slowly, the growing crack remains straight and stable. However, increasing
the velocity to a critical value vc gives rise to a transition whereby the fracture
path begins to oscillate and an instability appears. At still higher velocities crack
branching appears; see Figure 6.6. Ronsin et al. (1995) also provided experimental
data for brittle fracture propagation in thin glass strips, using a thermally induced
stress field. In their experiments the temperature field was controlled by the width
w of the plate, and induced thermal expansion in the sample. It was observed that
for widths below a critical value wc no fracture was formed. For wc < w < wo ,
where wo is a second critical width for the onset of oscillatory cracks, straight frac-
tures were formed and propagated with a constant speed. For w > wo oscillatory
fractures were generated that became more irregular as w was increased beyond wo .

8.4 Fracture of a Brittle Material by an Impact

Fracture of a brittle material by an impact, such as fracture of glass by an impact,
is an important phenomenon that most of us have seen with our own eyes. Sim-
ple and interesting experiments performed by Shinkai (1994) demonstrated this
442 8. Brittle Fracture: The Discrete Approach

Figure 8.27. Schematics of fracture pattern on impact on a thin glass plate. Hertzian fracture
is at the center (after Åström and Timonen, 1997b).

phenomenon very nicely. In his experiments thin square-shaped glass plates were
supported at the edges, and a small and heavy ball was dropped on them from a
point above the plates’ centers. The resulting fracture patterns consisted of three
types of cracks. If the impact velocity was high, only a circular hole at the point of
impact was formed, which is usually referred to as a it Hertzian fracture. However,
at lower impact velocities, radial and tangential cracks also appeared. An example
is shown in Figure 8.27. The radial cracks were fairly straight and directed outwards
from the point of impact, while the tangential ones formed a more or less circularly-
symmetric crack. At still lower impact velocities, only radial cracks were formed,
while if the impact velocity was too low, no crack was formed at all.
A simple and interesting model of dynamic fracture, in which the material cracks
as a result of an impact, was proposed by Åström and Timonen (1997b). They used
a triangular lattice of beams. The dynamics of the system were calculated using a
discrete form of Newton’s equations of motion given by
$ % $ % $ %
1 1 2 1 1
M + D u(t + t) = M − C u(t) − M − D u(t − t),
t 2 2t t 2 t 2 2t
(172)
where M is a diagonal mass matrix, C is the stiffness matrix, D is a diagonal
damping matrix, u is the vector that contains the displacements of the sites from
their equilibrium positions, and t is the length of the discrete time step. The
stiffness matrix that was used was that of slender beams (that is, a beam in which
bending is much larger than shear deformation). The boundary conditions imposed
on the lattice in the xy plane were such that the sites at the boundaries lattice are
constrained to remain at their original positions, while the sites in a circular area in
the middle of the lattice were forced to move a distance −vt in the z direction. For
the lattice in the xz plane only the sites at its left and right edges were constrained to
 8.4. Fracture of a Brittle Material by an Impact 443

Figure 8.28. Radial fractures in a lattice with an externally applied in-plane strain. The
results, from top left corner and clockwise, are for times t = 200, 350, 700 and 500 (after
Åström and Timonen, 1997b).

remain at their original positions, while a number of sites in the middle of the upper
boundary moved in the negative z-direction. A beam was considered as broken if
its elongation exceeded its pre-assigned threshold. However, beam fracture was
not instantaneous. It was assumed that, at the threshold, the Young’s modulus of a
beam begins to decrease linearly in time until it reaches zero. The rate at which the
modulus decreased was expressed by a parameter r = c/t, where c is a constant.
Figure 8.28 presents the crack patterns in a lattice with an externally applied
in-plane strain: r = 0.1/t, v = 1/600t, and the in-plane strain, xy = 0.167.
The numbers refer to the time steps at which the snapshots were taken. All the
qualitative features of the experimental crack patterns obtained by Shinkai (1994)
were reproduced by this rather simple model, and therefore it may be used for
studying other aspects of fracture of brittle materials by an impact.
444 8. Brittle Fracture: The Discrete Approach

8.5 Dynamic Fracture of Materials with Annealed

Disorder
All the lattice models that we have so far described and discussed are deterministic
in the sense that, a bond of the lattice breaks if its elongation, or the force that it
suffers, exceeds some threshold value. Therefore, at any given time, the bond to
be broken is identified deterministically, if the stress or strain field in the lattice is
given. As such, these models are appropriate for materials in which the disorder
is quenched. We now briefly describe the probabilistic lattice models of dynamic
fracture which, as discussed in Section 8.2, may be appropriate for fracture of
materials in which the disorder is annealed.As in the case of dielectric and electrical
breakdown (see Chapter 5), the probability pb that a bond of a lattice, representing
a spring or a beam, breaks is related to the force F that it suffers. Similar to the
dielectric breakdown model of Niemeyer et al. (1984) described in Section 5.4.1,
pb is assumed to have the following form
pb ∝ F η , (173)
where, as in Niemeyer et al.’s model, η is a parameter of the model. An exponential
form, pb ∝ exp(cF η ), has also been used (where c is a constant) which would
then be compatible with the theory of chemical processes according to which the
reaction rate R of a chemical process, such as the rate of breaking the interatomic
bonds, is given by
R = k(T ) exp(−Ha /kB T ), (174)
where k(T ) is a temperature-dependent pre-exponential factor, Ha is the activation
energy, and kB is the Boltzmann’s constant.
The value of η in Eq. (173) is selected based on the physics of the problem.
For example, one may argue that if an external force is exerted on the system, the
activation energy is reduced by an amount which is proportional to F 2 , and thus
one may assume that η = 2. On the other hand, one may also argue that a bond
breaks when its length has reached a critical threshold, and that for a harmonic
potential the required energy to reach this length is proportional to F , and thus
η = 1.
The process time is explicitly incorporated into this model by the following
algorithm. Each time a bond breaks, the process time is increased by an amount
t given by
1
t = (m)
, (175)
Npb
(m)
where N is the total number of unbroken bonds in the network, and pb is the
maximum probability of breaking for any bond in the network at time t. If the bond
breaking rates are actually known (i.e., if p0 is known), then Eq. (175) provides
an absolute time scale for the system; otherwise the true process time is directly
proportional to the time scale calculated by (175).
 8.5. Dynamic Fracture of Materials with Annealed Disorder 445

Figure 8.29. Fracture pattern in a viscoelastic material (after Van Damme et al., 1987a).

Such a model may be able to explain some aspects of experimental observations

of Van Damme and co-workers (Van Damme et al., 1986, 1987a,b). They displaced
clay dispersions by air or water in a Hele–Shaw cell, an essentially 2D system
consisting of two parallel glass sheets with a small gap between them. When the
clay concentration was low, the displacement pattern was similar to diffusion-
limited aggregates (DLA) described in Chapters 1 and 5. However, at high clay
concentrations a transition was seen from a DLA-like displacement to a fracture
pattern; see Figure 8.29. This is a good example of a system in which the disorder
is annealed, since as the injected water or air displaces the clay, the material must
continuously adjust itself to accommodate the fact that the displacing fluid is
pushing its way into the clay. Thus, disorder in this system changes with the time.
Curtin and Scher (1991,1992,1997) and Curtin et al. (1997) studied lattice
models of dynamic fracture in which each node of the lattice could be a site for
nucleation of defects or fractures. It was assumed that the probability pi per unit
time of nucleating a crack at site i is a monotonically-increasing function pi [σi (t)]
of the local stress σi at i, and was taken to be pi (t) = Aσi (θ )η , where η is a pa-
rameter of the model. Many interesting predictions emerge from this model either
analytically or by numerical simulations, including, (1) failure is more abrupt as
η increases; (2) failure times scale inversely with the logarithm of the system size
raised to some power, and (3) the distribution of failure times is Gumbel-like (dou-
ble exponential; see Section 8.2.1.4) and becomes broader as η increases, implying
that failure becomes less predictable as it becomes more abrupt.
Another model that had some dynamics built into it was developed by Louis and
Guinea (1987), and further developed by Fernandez et al. (1988) and Meakin et al.
(1989). In this model a triangular lattice of Hooke’s springs is used. It is assumed
that only those bonds that are on the surface of the growing fracture can break.
446 8. Brittle Fracture: The Discrete Approach

An initial microfracture is inserted into the network at its center. Three different
ways of breaking the bonds were considered. In model I only those bonds that
join pairs of sites that are both on the fracture perimeter were broken. In model II
any bond associated with a damaged node—one that had five or fewer unbroken
bonds—was broken, while in model III any of the bonds associated with any of the
sites at the fracture perimeter was broken. The probability of breaking any bond i
is given by
(i − 0 )η
p b = N , (176)
j =1 (j − 0 )
η

where i is the length of bond i, 0 is its equilibrium length, and N is the number
of the unbroken bonds on the perimeter of the fracture. The network was initially
diluted by a small amount to prevent unwanted nonlinearities, and a constant force
was applied on the boundaries of the network. The model produced fractal fracture
patterns with a fractal dimension Df that depended on the parameter η and the
boundary conditions at the perimeter of the growing fracture. The fracture patterns
were quite similar to diffusion-limited aggregates (see Chapters 1 and 5). Other
types of boundary conditions were also used in this model. For example, shear
strain and uniaxial tension were both used (Hinrichsen et al., 1989) in which case
the fracture pattern had an X-like shape. If only bonds in tension were allowed to
break, then only one arm of the X-shape grew.

8.6 Fracture of Polymeric Materials

We already described in Section 6.16.1 the general fracture properties of polymers.
Mechanical stability of polymeric materials and their resistance to fracture are
clearly of great practical importance. Commercial products made of plastics and
other types of polymeric materials must preserve their form under the allowed
external loads, and therefore one must ensure that they do not develop fracture and
break. For this reason, fracture of polymeric materials has been an active research
field for a long time.
Deformation of a polymeric solid material may include, in addition to the re-
versible part, an irreversible plastic flow or yielding which sets in when the stress
becomes large enough to surmount the yield point. Compared to metals and ce-
ramics, the yield point of polymeric materials is low, so that moderate forces or
stresses are often sufficient to trigger the yield process. Thermal effects play an
important role, and the environment can also affect the properties of polymeric
materials if fluids or gases can penetrate into these materials and degrade them.
Extensive experimental studies have shown that there are two mechanisms of
yielding in polymeric materials which have very different appearances and can be
easily determined. These two mechanisms are as follows.
(1) The first mechanism is called shear yielding. Consider, for example, polyethy-
lene. If we stretch a sample of this material with a constant rate, then its
 8.6. Fracture of Polymeric Materials 447

load-extension (stress-strain) diagram will have the following characteristics.

The stress increases at first but then, when the yield point is reached, it passes
through a maximum and a neck develops somewhere in the polymer which
can extend up to the full extent of the sample. This experiment takes place un-
der an essentially constant tensile force, and eventually elongates the material
by several times its original length. If the stretching continues, the force will
increase up to the point of break. Shear yielding is typical for partially poly-
crystalline polymers, and has also been seen in many amorphous polymers,
such as polycarbonate. Shear bands are also formed, and are oriented along
the directions of maximum shear stress.
(2) Polystyrenes exhibit a quite different behavior. If we again carry out an exper-
iment similar to what was described above and form the stress-strain diagram,
we will find that it has the following characteristics. The force increases at
first but then, after a slight bending, the material breaks before a maximum is
reached. Inspection will then indicate the formation of many void containing
microdeformation zones. As described in Chapter 6, these localized zones of
plastic flow are called crazes, and crazing is usually used to refer to this second
mechanism of yield in polymeric materials. However, we must keep in mind
that shear yielding and crazing are not mutually exclusive, and more often than
not they both are operative during deformation of polymeric materials. The
stress condition and the temperature of the system are the controlling factors
in deciding which mechanism is dominant in a polymeric material. Clearly,
the amount of flow before fracture determines the ductility of a polymeric
material, so that brittle polymers break without exhibiting much preceding
flow.
Many of the precise experiments on dynamics fracture of brittle materials that
were described in Chapters 6 and 7 were actually carried out with polymeric
materials, such as PMMA, and therefore all the continuum and discrete models of
brittle fractures that were described in Chapter 7 and this chapter are applicable
to such materials. Similarly, theories of ductile behavior of materials can also
be used for investigating this mode of deformation in polymeric materials. In
addition, we summarized in Chapter 6 the most important fracture properties of
polymeric materials (see Section 6.16.1). Given that the field of polymer fracture
is well-described and documented (see, for example, Kinloch and Young, 1983;
Kausch, 1987), there is no need here for a long discussion of fracture behavior of
polymeric materials. Instead, we restrict ourselves to a brief discussion of recent
discrete models of dynamic fracture in such materials.
There is strong evidence that applying a stress σ to a polymer reduces the
activation energy by an amount that is proportional to σ . This means that the bond
breaking rate, or the probability that a bond breaks, can be written as

pb = p0 exp[−(Ha − a σ )/kB T ], (177)

where a is the activation volume (in 3D) or surface (in 2D) of the system. In a
network model a σ is replaced with La F , where La is the activation length of the
448 8. Brittle Fracture: The Discrete Approach

bonds. If all the bonds are equivalent, then in the limit T → 0 the bond with the
largest strain will always break first, and thus in this limit the probabilistic models
reduce to the deterministic ones described earlier in this chapter. In the opposite
limit, T → ∞, the bond breaking process becomes completely random and thus
represents a percolation process. The process time is incorporated into this model
by the algorithm described above using Eq. (175).
Aside from the early work of Dobrodumov and El’yashevich (1973), the first
model of this type was developed by Termonia and Meakin (1986). In their 2D
model the probability pi that a bond i breaks is given by
pi ∝ exp[αi ei2 /(2kB T )], (178)
where αi is the elastic constant of bond i, and ei is its elongation. Fractal fracture
patterns were generated by this model with a fractal dimension Df  1.3 in 2D.
Termonia and Smith (1986) and Termonia et al. (1985, 1986) developed prob-
abilistic models of mechanical and fracture properties of polymer fibers, which
possess a very complex morphology. In their models there is a distinction between
the primary bonds—those that are parallel to the fiber axis—and the secondary
bonds—those that are perpendicular to the fiber axis. The primary bonds are strong
covalent bonds, while the secondary bonds are the much weaker van der Waals and
hydrogen bonds. Defects are also included in these models by removing a fraction
of the bonds before deformation of the system is started. The bonds break with a
probability given by Eq. (177), but the activation energies and volumes for the pri-
mary bonds were about 2 orders of magnitude larger than those for the secondary
bonds, while p0 was assumed to be the same for both types of bonds. The sec-
ondary bonds were allowed to reform between adjacent sites if their coordinates in
the direction of the primary bonds became equal, whereas primary bonds were not
allowed to do so. Figure 8.30 shows the fracture patterns generated by this model
(under isothermal condition), which are in agreement with typical experimental
observations.
Termonia and Smith (1987,1988) developed models of polymer deformation
and failure in which the effect of chain slippage and the release of entanglements

Figure 8.30. Fracture pattern in the model of Termonia et al. (1985,1986) in which the
probability of failure for the primary and secondary bonds were not the same. The top
pattern corresponds to 150% strain, while the bottom one is for 300% strain.
 8.7. Fracture of Thin Solid Films 449

was taken into account. Such effects play a prominent role in failure of polymers.
In their model the rate of failure of van der Waals bonds and that of chain slippage
were both given by Eq. (177), but with different activation energies and volumes.
For the chain slippage process, σ represents the stress difference between two parts
of a chain separated by an entanglement point. Their model used a 2D diamond-
like lattice with nodes that represented the entanglement points between pairs
of polymers. The lattice was then decorated randomly with polymer molecules
that intersect only at the entanglement points until there is an entanglement point
associated with every node. The stress σ is predicted by the classical theory of
rubber elasticity to be
$ %
−1 R
σ = βkB T L , (179)
nc 
when nc was the number of chain segments of length  between a pair of entan-
glement points separated by a distance R, L(z) = coth(z) − 1/z is the Langevin
function, and β is given by
√
N c nc
β= , (180)
3
with Nc being the number of chain strands per unit volume. The predicted fracture
patterns were found to be in good agreement with experimental observations. More
details can be found in the review by Meakin (1990).

8.7 Fracture of Thin Solid Films

The last class of materials for which dynamic lattice models of fracture have been
developed is thin solid films. A thin film is attached to a substrate with mechanical
properties that are usually very different from those of the film. As a result, the
surface layer usually suffers large stresses that are generated by various factors,
such as decohesion, buckling, spalling, and in-plane cracking, leading to fracture
patterns that resemble those in dried-up mud. Some examples are Al2 O3 sputter
deposited onto copper (see, for example, Jarvinen et al., 1984), and chromium
metal electrodeposited onto an aluminum alloy (see, for example, Namgoong and
Chun, 1984).
Meakin (1987) developed a simple model for this type of fracturing process. In
his model the thin layer is represented by a triangular network of Hooke’s springs.
In addition, each node of the network is connected to a rigid substrate by another
Hooke’s spring much weaker than those in the triangular network. Only the bonds
in the triangular network are allowed to break. The probability of their failure is
given by Eq. (178) with kB T = 1. In a typical simulation, the bond length  = 1.0
at the start of the process, and the equilibrium length 0 = 0.90. Figure 8.31 shows
the fracture pattern resulting from such a model.
Skjeltrop and Meakin (1988) modified Meakin’s original model in order to
simulate fracture patterns of polystyrene bead monolayers. The monolayers are
constructed by placing an aqueous dispersion of microspheres between two parallel
450 8. Brittle Fracture: The Discrete Approach

Figure 8.31. A typical fracture pattern in thin films, for which 5000 bonds have been broken
(after Meakin, 1987).

sheets of glass and allowing the water to evaporate slowly along one edge of the gap
between the two sheets. The system resembles a thin film. In the model of Skjeltrop
and Meakin (1988) the points of attachment of the network’s nodes to the substrate
were allowed to move. The force that a weak bond that connects a node i of the
triangular network to the substrate exerts on i is given by F = αw (Ri0 − Ri ), where
αw is the elastic constant of the weak bond, Ri is the position of the ith node, and
Ri0 is its initial position. If |Ri0 − Ri | exceeds a threshold, the point of attachment
to the substrate at Ri0 is moved towards the current position Ri of the node until
|Ri0 − Ri | becomes equal to the threshold. The resulting fracture patterns were in
very good agreement with the experimental patterns.
More recent efforts in this area include those of Crosby and Bradley (1997) and
Leung and Néda (2000). The latter group used a model in which the grains in the
thin films were represented by an array of blocks. Each pair of the neighboring
blocks was connected by a bundle of coil springs having a unit spring constant
and an equilibrium length . Initially, the blocks are randomly displaced by r  
about their mean positions on a triangular lattice. For slow cracking on a frictional
substrate, the motion of the grains is over-damped. Thus, the system evolves quasi-
statically with a driving rate which is much slower than the relaxation rate. This
implies that one does not have to solve the equation of motion, but rather update
the configuration of the system according to certain criteria.
In a typical cracking of thin films, the film often hardens and/or weakens in
time, hence tending to contract which is, however, resisted by friction from the
substrate. As a result, as mentioned above, stress is built up and relaxed slowly.
To incorporate such effects in any reasonable model, one can pre-strain the array
of the blocks. Then, two force thresholds, Fs for slipping and Fc for cracking,
are decreased systematically to induce these two modes of motion. For example,
 8.7. Fracture of Thin Solid Films 451

in a drying experiment, the narrowing and breaking of liquid bridges between

the grains, due to evaporation, provide an example of this driving. This model
generates cracking patterns that are very much similar to those that are developed
in thin films during their drying.

Summary
Lattice models offer a detailed account of most aspects of fracture, including
those that had been considered too complex to quantitatively describe, or those
that had simply been ignored. At the same time, there is no conflict between the
lattice models and conventional fracture mechanics. In addition, lattice models
of crystalline materials make the additional prediction that a forbidden band of
velocities exists for cracks. This means that a fracture can only propagate stably
above a finite minimum velocity. Molecular dynamics simulations (see Chapter 9)
of crystalline materials indicate that this forbidden band of states may disappear
at room temperature, but should be observable in low-temperature experiments.
Through precise experiments (most of which were described in Chapters 6 and
7), large-scale simulations of lattice models of quasi-static and dynamic fracture
described in this chapter, and molecular dynamics simulations to be described
in Chapter 9, we now have a deeper understanding of fracture propagation, and
in particular the structure and dynamics of mechanisms of dissipation within the
near vicinity of the tip of a propagating fracture in a brittle material. We now know
that fracture in brittle materials is governed by a dynamic instability that leads
to repeated attempts for fracture branching. Although the instability also appears
in dynamic finite-element simulations of fracture (see Chapter 7), it appears to
have no analytical explanation in a continuum framework. In fact, many classical
models of the cohesive zone have been shown to be ill-posed in that, they admit a
set of possible states under identical conditions. Theories formulated on a lattice,
on the other hand, do not exhibit such difficulties. It remains to be seen whether
a simple continuum limit exists, or whether a crucial ingredient in understanding
fracture is the discreteness of the underlying atoms.
However, despite considerable progress, our understanding of fracture propaga-
tion phenomena is still incomplete. In addition to not having a simple continuum
limit of the discrete models, we must keep in mind that the most detailed and precise
experiments on brittle fracture have so far been carried out in amorphous materials
at room temperature, whereas the most detailed theories currently available apply
mostly to very low (or zero) temperatures. A general theoretical framework for
analyzing fracture propagation over a wide range of brittle materials has not yet
been developed.
 Part III

Atomistic and Multiscale

Modeling of Materials
9
Atomistic Modeling of Materials

9.0 Introduction
In all the chapters of Volume I, as well as in the present Volume so far, we have uti-
lized continuum mechanics and lattice models to describe modeling and simulation
of morphology of heterogeneous materials and estimating their effective proper-
ties. These models are appropriate for microscopic as well as macroscopic length
scales, but cannot provide any insight into materials’ properties at the smallest
length scales, namely, the molecular scale. In this chapter we describe and discuss
modeling and simulation of materials and their properties at the molecular scale.
To achieve our goal we describe three important theoretical and computational
tools that have been developed over the past three decades, namely, the density
functional theory (DFT) and its variants, classical molecular dynamics (MD) simu-
lation, and quantum MD (QMD) technique. The advent of very fast computers and
development of massively-parallel computational strategies have made it feasible
to carry out large scale calculations at the molecular level, and in this endeavor
these three methods have become indispensable tools for predicting the properties
of materials at such length scales.
Prediction of electronic properties and morphology of a material requires, in
principle, quantum-mechanical computation of the total energy of the system and
minimization of this energy with respect to the electronic and nuclear coordinates.
To carry out such computations, one must start with the Hamiltonian of the system
which, for a system of N electrons and N  nuclei with charges Zn , is given by



N
pi2
Pn2
N
1 1

N

N
e
H= + +
2m 2Mn 2 4π ε0 |ri − rj |
i=1 n=1 i=1,i
=j j =1
N


N

N 


N 
1 Zn e 2 1 1 Zn Zn e2
− + , (1)
4π ε0 |ri − Rn | 2 4π ε0  
|Rn − Rn |
n=1 i=1 n=1,n
=n n

where pi and Pn are the momenta of the ith electron and the nth nucleus, re-
spectively (subscript i refers to the electrons and n to the nuclei), ri and Rn are
their position vectors, m is the electron mass, Mn is the mass of the nth nucleus,
e is the electron charge, and ε0 is the permittivity. The first two terms of Eq. (1)
represent the kinetic energy of the electrons and nuclei, respectively, while the
third and fourth terms are, respectively, the result of Coulomb repulsion between
456 9. Atomistic Modeling of Materials

the electrons and Coulomb attraction between electrons and nuclei. Equation (1)
is too complex for use in exact computations, especially when one must deal with
a large system, and therefore reasonable approximations must be made in order to
make the computations feasible. One obvious simplification can be made by taking
advantage of the fact that there is a large difference in mass between the electrons
and nuclei, while the forces on the particles are the same. Therefore, the elec-
trons respond essentially instantaneously to the motion of the nuclei. As a result,
electronic and nuclear coordinates in the many-body wave function can be sepa-
rated, and the nuclei can be treated adiabatically. This separation is the well-known
Born–Oppenheimer approximation which reduces the solution of the many-body
problem to that of the dynamics of the electrons in some frozen-in configura-
tions of the nuclei. Thus, the Hamiltonian of the system in the Born–Oppenheimer
approximation is given by



N
pi2 1 1

N

N
e2 1

Zn e 2
N N
H= + − . (2)
2m 2 4π ε0 |ri − rj | 4π ε0 |ri − Rn |
i=1 i=1,i
=j j =1 n=1 i=1

Based on such simplifications, other approximate schemes have been sug-

gested. A well-known example of such approximations is the Hartree–Fock theory.
Suppose that
F{k } = k k , (3)
where F is called the Fock operator defined below, k is the eigenvalue of the
operator, and k (x) is the spin-orbital (electronic state or wave function), a function
that depends on the spatial position x and spin coordinates of one electron. The
Fock operator is defined by
  N 
1 2
Zn
1
F{k } = − ∇ − k (x) + |l (x )|2  (x) dx 
| k
2 n
|r n − R n | |r − r
l=1

N 
1
− l∗ (x ) k (x )l (x) dx  (4)
|r − r |
l=1

where ∗ denotes
4 a complex
 4 conjugate, and 4 the following notation convention has
been used: dx  = s  d 3 r ; that is, dx  denotes a sum over the spin s  and
an integral over the spatial coordinate r . We have used the standard atomic units
(using the Bohr radius, and the electron mass and charge as the basic units) so
that, for example, instead of having the usual factor h̄2 /(2m)∇ 2 we have 12 ∇ 2 . We
use this convention throughout this chapter. Equations (3) and (4) constitute the
well-known Hartree–Fock theory, also known as the self-consistent-field theory.
Note that the fourth term of Eq. (4) is a non-local term, since although the operator
acts on k , its value at r is determined by the value assumed by k at all possible
positions r . Note also that the ground state electron density ρ(r) is given by


ρ(r) = |i (r)|2 . (5)
i
 9.0. Introduction 457

The Hartree–Fock equation has been a pillar of computation of the electronic

structures of atoms and molecules. It is a nonlinear equation which is solved nu-
merically by a sort of self-consistent iterative procedure, the essence of which is
as follows. Since solving Eq. (4) yields an infinite spectrum, to obtain the ground
state, one must take the lowest N eigenvalues of the spectrum as the electronic
states (spin-orbitals of the electrons). These constitute the first approximate so-
lution of Eq. (4) which are then utilized for constructing the next iteration of the
Fock operator. The operator is diagonalized again to obtain the next approximate
solution to the electronic states. The procedure is repeated until convergence has
been achieved.
However, computation of properties of solid materials based on the Hartree–
Fock equation is an extremely difficult problem, and therefore, over the past several
decades, other alternatives have been developed. Notable among them is the DFT
developed by Hohenberg and Kohn (1964) and Kohn and Sham (1965), which is
now widely used for predicting electronic properties of hard materials. (In recog-
nition of his contributions to this research field, Kohn received the 1998 Nobel
Prize in chemistry.) In words, the DFT allows one, in principle, to map exactly
the problem of a strongly interacting electron gas, in the presence of nuclei, onto
that of a single particle moving in an effective non-local potential, and provides
an expression for the total energy of the system. The effective potential is not
known exactly, but often, for reasons that are not completely understood yet, local
approximations (see below) to the non-local potential are highly accurate. One can
then minimize the total energy of the system that is provided by the DFT, often
referred to as the Kohn–Sham total-energy functional, in order to determine its
various properties. Beginning in the mid 1970s, the DFT was revitalized because
extensive computations indicated that local approximations to the effective non-
local potential can predict a variety of ground-state properties of materials that are
within a few percent of experimental data (Dreizler and Gross, 1990).
Such a computational strategy, which requires only a specification of the ions
present (by their atomic numbers), is usually referred to as an ab initio method.
Although nearly two decades ago most ab initio methods, including the DFT, were
capable only of modeling systems of a few atoms, they can now model systems
with a large number of atoms. Of all ab initio methods, the total-energy pseu-
dopotential technique (see below) based on the DFT, stands alone. This technique,
combined with a method developed by Car and Parrinello (1985) (see Section
9.4), or a direct method of minimization of the total energy, such as the conjugate-
gradient technique (see Section 9.5.2), have transformed the way in which people
view quantum-mechanical ab initio computations and hence total-energy pseu-
dopotential calculations because, in addition to their remarkable efficiency, they
allow ab initio quantum-mechanical computations at non-zero temperatures.
Computation of materials’ properties based on the DFT involves quantum-
mechanical calculations. Over the past two decades, MD simulations, in which
atoms and molecules are treated as classical particles and quantum-mechanical
effects are neglected, have also become an important tool for investigating and
predicting various static as well as dynamical properties of materials. We refer
458 9. Atomistic Modeling of Materials

to this method as the classical MD simulation. The increasing popularity of the

classical MD methods is due to the fact that, over the past decade, highly efficient
simulation techniques, based on massively-parallel and vectorized computations
(see Section 9.6), have been developed that allow one to carry out MD simulations
with billions of atoms, with the world record at the time of writing this book being
a MD simulation with 5,180,116,000 particles (Roth et al., 2000). Moreover, such
MD techniques allow us now to simulate much longer time periods than what was
believed to be feasible only a decade ago.
In order to increase the efficiency of ab initio computations, one can combine
quantum electronic structures with a MD simulation to not only calculate the
nuclear positions, but also the electronic charge cloud. This method, pioneered by
Car and Parrinello (1985), is what we refer to as the QMD. This method describes
a system in which the electronic structure does not, in general, completely relax
to the true ground state, but follows it rather closely, and has been proven to be
highly successful for describing many properties of materials.
Therefore, the purpose of this chapter is twofold.
(1) We wish to provide an overview of the essential concepts and ideas of the
ab initio quantum-mechanical computations, MD and QMD simulation, and
related problems, such as direct methods of minimization of the total energy,
and vectorized and parallelized algorithms for MD simulations. Our goal is to
describe the advances that have been made in these areas. Our discussion is not,
and cannot be, exhaustive, as comprehensive description of all aspects of each
subject would require a book by itself. For example, the review by Abraham
(1986), and the books by Allen and Tildesley (1987) and Rapaport (1995)
describe the classical MD simulations in detail. Instead, we set for ourselves
the modest goal of outlining the basic concepts, ideas, and techniques of each
of these computational tools which, together with adequate references to the
recent literature, should enable the interested reader to pursue them. These
techniques, with their ever increasing accuracy and efficiency, have gained
popularity as major tools of investigating static as well as dynamical properties
of materials, and therefore it is more than appropriate for this book to describe
them.
(2) In addition to several examples discussed throughout this chapter, we also dis-
cuss, as an application of computations of materials’properties at the molecular
scales, MD simulation of dynamic fracture of materials. At first, it may seem
strange that computer simulations that involve atoms can be used for studying
fracture propagation which is a macroscopic phenomenon involving large-
scale structural changes. Even at molecular scales the motion of dislocations
over long distances is of primary interest. So, why should computer simu-
lations in which one arranges atoms in a crystal lattice and deforms it have
anything to do with the true dynamics of fracture propagation in the crystal?
The answer is obvious: Fracture of brittle materials is a physical process which
naturally connects small and large length scales. Although stresses and strains
that cause deformation and fracture of a material are applied at macroscopic
 9.1. Density-Functional Theory 459

scales, fracture itself is the severing of the bonds at atomic scale. This im-
plies that MD simulation of dynamic fracture in materials that account for
the phenomena at molecular scales does not have to be very large, although
large-scale MD simulations of dynamic fracture with up to about 108 atoms
have been carried out (Abraham et al., 1997a,b; see below).
Our discussions in this chapter are also a prelude to Chapter 10 where we
describe how atomistic and molecular modeling of materials is integrated with the
microscopic and macroscopic methods of the previous chapters in order to develop
a multiscale approach for investigating materials’ properties over several widely
disparate length scales.

9.1 Density-Functional Theory

Comprehensive reviews of the DFT and its applications are given by Jones and
Gunnarsson (1989) and Payne et al. (1992). Some of our discussions in this sec-
tion, and in Sections 9.3 and 9.4 closely follow these reviews. Motivated by the
fundamental theorems of Hohenberg and Kohn (1964) of the DFT, Kohn and
Sham (1965) developed a set of accessible one-electron self-consistent eigenvalue
equations that have provided a practical means of realistic electronic structure
calculations on a large array of atoms, molecules and materials (see, for example,
Parr and Yang, 1989). The Hohenberg–Kohn theorems were extended to finite-
temperature quantum systems (Mermin, 1965) and to purely classical fluids (see,
for example, Hansen and McDonald, 1986). In addition, an integral formulation
of electronic structure has also been developed in which the one-electron density
is obtained directly without the introduction of orbitals (Harris and Pratt, 1985),
although to date this theory, despite its promise, has not been used extensively in
numerical studies.
The electron density ρ(r) is subject to the constraint that

ρ(r) d 3 r = N, (6)

where N is the total number of electrons in the system. The Hamiltonian of a

many-electron system is given by

 1 
1

1
H= − ∇i + Uc (ri ) +
2
, (7)
2 2 |ri − rj |
i i,i
=j j

where Uc (r) is an external potential representing the Coulomb attraction by the

frozen-in nuclei. The ground state can be determined by a variational approach
which is carried out in two steps.
(1) One minimizes, for a given electron density, the energy functional with respect
to the wave functions that are consistent with the given density.
(2) The result,
E[ρ(r)] = min |H|, (8)
460 9. Atomistic Modeling of Materials

is then minimized with respect to ρ, subjected to the constraint imposed by

Eq. (6). If we now separate the Hamiltonian, H = H0 + Uc (r), where H0
represents the Hamiltonian without the external potential (i.e., the Hamiltonian
of a homogeneous electron gas), then
 
E(ρ) = min [|H0 |] + Uc (r)ρ(r) d 3 r = E  (ρ) + Uc (r) d 3 r,
(9)
where E  (ρ) = min [|H0 |] does not depend on the external potential
Ue .
The main problem lies in the fact that E(ρ) is not known for both interacting and
non-interacting electron systems. We know, however, that in the non-interacting
case E(ρ) can be written as

E(ρ) = Ek (ρ) + Uc (r)ρ(r) d 3 r, (10)

where Ek (ρ) is the kinetic contribution to E(ρ). Variation of E with respect to ρ

yields the following equation
δEk (ρ)
+ Uc (r) = λρ(r), (11)
δρ(r)
where λ is the Lagrange multiplier associated with the constraint (6). The exact
form of Ek (ρ) is, of course, unknown, but we know that the ground state density is
given by Eq. (5), and that the spin-orbitals satisfy the single-particle Schrödinger
equation:
 
1 2
− ∇ + Uc (r) k (r) = k k (r), (12)
2
where k are the associated eigenvalues. The spin-orbitals k must be normalized
in order for the constraint (6) to be satisfied. It is clear that if Ek (ρ) can be used in
place of E  (ρ), then it must also be assumed that Ek is independent of the external
potential Ue (r) (just as E  is independent of Ue ). We now write down the energy
functional for a many-electron system with electronic interactions included:
  
1 1
E(ρ) = Ek (ρ) + Uc (r)ρ(r) d r + 3
d r ρ(r )
3
ρ(r) d 3 r  + EXc (ρ).
2 |r − r |
(13)
The first three terms of Eq. (13) represent the contribution to E(ρ) by the non-
interacting electron gas, whereas EXc (ρ), which is called the exchange-correlation
energy, represents all other contributions to E(ρ) that are not accounted for by the
first three terms. Equation (13) has the advantage that it contains no approximation,
and that all the unknown contributions have been lumped in EXc (ρ). Varying
Eq. (13) with respect to ρ, we obtain

δEk (ρ) 1 δEXc (ρ)
+ Uc (r) + 
ρ(r ) d 3 r  + = λρ(r). (14)
δρ(r) |r − r | δρ(r)
 9.1. Density-Functional Theory 461

In essence, we have an effective potential given by


1 δEXc (ρ)
Ueff (r) = Uc (r) + ρ(r )d 3 r  + . (15)
|r − r | δρ(r)
Therefore, one may write
 
1
− ∇ 2 + Ueff (r) k (r) = k k (r). (16)
2
Putting everything together, we finally obtain

N   
1 1  3  δEex (r)
E= i − d 3 r ρ(r) 
ρ(r ) d r − ρ(r) d 3 r + EXc (ρ).
2 |r − r | δρ(r)
i=1
(17)
Equations (5) and (15)–(17) constitute the DFT, first derived by Kohn and Sham
(1965).
As Eqs. (15)–(17) indicate, the Kohn–Sham equations map the interacting many-
electron system onto a system of non-interacting electrons moving in the effective
potential Ueff generated by all the other electrons. If the exchange-correlation
energy functional EXc were known exactly, then taking the functional derivative
with respect to the density would produce an exchange-correlation potential that
includes the effects of correlation and exchange exactly. However, EXc is not
known exactly, and thus the main task is developing an accurate approximation for
it. Once this approximation has been decided upon, the bulk of the work involves
finding the solution of the eigenvalue problem described by Eq. (16). Note that
the main difference between the Hartree–Fock and DFT approximations is that,
the latter replaces the Hartree–Fock exchange term by the exchange-correlation
energy. A particularly powerful method for developing accurate approximations
for EXc is the so-called local-density approximation which we will discuss shortly.
Thus, minimization of the total energy functional can be carried out directly
using the Kohn–Sham model, Eqs. (16) and (17), subjected to the constraint (6),
utilizing various minimization techniques, such as the simulated annealing method
described in Chapter 3 of Volume I, or the conjugate-gradient method that will
be described in Section 9.5. Usually, the eigenvalues k are interpreted as the
excitation energy. When such interpretation has been utilized for predicting many
properties of various atoms and molecules, the predictions have been found to be in
excellent agreement with the experimental data. For example, Table 9.1 compares
the DFT predictions for the cohesive energy and lattice constants of diamond,
silicon and germanium, and it is clear that the agreement between the predictions
and the data is very good.
Let us mention that the DFT has been extended to time-dependent problems
(Runge and Gross, 1984). Time-dependent DFT (TDDFT) makes it possible to
apply the DFT to excited states of many-body systems. In this scheme, one must
define a dynamic exchange-correlation energy EXc (ρ, r, t) which must somehow
be calculated. It has been shown that substantial improvements of excitation en-
ergies with respect to the original Kohn–Sham eigenvalues are obtained with the
462 9. Atomistic Modeling of Materials

Table 9.1. The DFT predictions of lattice constants

and cohesive energies for three materials, and their
comparison with the experimental data. Atomic
units have been used (adapted from Jones and
Gunnarsson, 1989).

Lattice Constant Cohesive Energy

Material DFT Expt. DFT Expt.
Diamond 6.807 6.740 7.58 7.37
Si 10.30 10.26 4.84 4.64
Ge 10.69 10.68 4.04 3.85

TDDFT. However, even within this extension, certain problems in solids persisted
for some time, such as the wrong Kohn-Sham band gap, regardless of the type of
approximation used for EXc . Tokatly and Pankratov (2001) showed that, at excita-
tion frequencies, EXc (r, r ) exhibits a highly non-local behavior with a range that
is as large as the system itself, and hence it diverges as the system’s size becomes
very large. They developed a perturbation technique which maintains a correct
electron density in every order of the perturbation theory, and therefore removed
this unphysical feature of the TDDFT.
However, in general, the DFT has been designed for predicting the ground-state
properties, and the Kohn–Sham eigenvalues are actually the derivatives of the
total energy with respect to the occupation numbers of these states (Janak, 1978).
Therefore, it may be appropriate to interpret k and k as only auxiliary variables
that are used for constructing the ground-state energy and density, because there
are many materials for which interpretation of k as the excitation energy is wrong
and leads to erroneous results. Examples include band gaps in semiconductors and
insulators. This should not be surprising because the DFT scheme has really been
designed for computing the ground states only.
Alternatively, one can use an indirect method for carrying out the minimization
of the total energy functional. Among the most successful such indirect methods
is the QMD method of Car and Parrinello (1985). We will describe this method in
Section 9.3 after describing the classical MD technique.

9.1.1 Local-Density Approximation

The local-density approximation assumes that the exchange-correlation energy
functional is purely local. In an inhomogeneous material, the exchange-correlation
potential at any point r depends not only on the electron density at this point, but
also on its variations in the vicinity of r. Thus, one may develop a gradient ex-
pansion in which EXc depends on ρ(r), ∇ρ(r), ∇[∇ρ(r)], · · · However, if such
an expansion is used in the DFT, the required computations will be very difficult
to carry out in a reasonable amount of time. In addition, simple gradient expansions
 9.1. Density-Functional Theory 463

are rather badly behaved, and therefore one must be careful in using a gradient
expansion, an issue that will be discussed in the next section. Such difficulties
provided the prime motivation for developing the local-density approximation
which ignores all corrections to the exchange-correlation energy at any point r
due to heterogeneities in its vicinity. Instead, one makes the ansatz that

EXc [ρ(r)] = Xc [ρ(r)]ρ(r) d 3 r, (18)

where Xc [ρ(r)] is the exchange-correlation energy per particle of a homogeneous

electron gas at density ρ(r). In other words, if the exchange-correlation potential
is assumed to be a local function, as opposed to be a functional, of the density
with the same value as for a uniform electron gas, one obtains the local-density
approximation. The non-uniform gas at r is therefore treated as if it were part of a
uniform electron gas of constant density.
The local-density approximation is very accurate if ρ(r) does not vary too
rapidly, but is also surprisingly accurate when the distribution of electrons is
strongly inhomogeneous, such as at surfaces and in molecules. To calculate Xc ,
the exchange effects are separated out from the dynamic correlations effects that
are due to the Coulomb interaction between the electrons. The exchange part,
commonly denoted by x , is given by
x [ρ(r)] ∼ −cρ 1/3 (r), (19)
where c is a constant. Equation (19) can be derived based on calculations for a
homogeneous electron gas. For open shell systems, the spin-up and -down densities
ρ↑ and ρ↓ are usually taken into account as two independent densities in the
exchange-correlation energy. In this version of the local-density approximation,
which is called the local spin density approximation, Ex , the exchange part of the
energy, is given by (Jones and Gunnarsson, 1989)

3 4/3 4/3
Ex (ρ↑ , ρ↓ ) = − [ρ↑ (r) + ρ↓ (r)] d 3 r, (20)
(3/4π )1/3
which is obtained by inserting (19) into (18). Various schemes have also been
proposed for taking into account the effect of the dynamic correlations (see, for
example, Ceperley, 1978; Perdew and Zunger, 1981, and references therein).
In general, the local-density approximation gives a single well-defined global
minimum for the energy of a non-spin-polarized system of electrons in a fixed
ionic potential, implying that any energy minimization scheme (such as simulated
annealing described in Chapter 3 of Volume I) will locate the global energy mini-
mum of the electronic system. Magnetic materials, on the other hand, are expected
to have more than one local minimum in the electronic energy. In this case, per-
forming total-energy calculations will be very difficult, because the global energy
minimum could be found only by sampling the energy functional over a large
region of phase space. Considering the nature of local-density approximation, its
success in providing quantitative predictions for many materials is remarkable.
464 9. Atomistic Modeling of Materials

9.1.2 Generalized Gradient Approximation

Another popular approximation to the exchange-correlation energy that is now
widely used is the generalized gradient approximation according to which

EXc [ρ(r)] = f (ρ, ∇ρ)d 3 r, (21)

where f is an analytic parameterized function that is fitted in such a way that

EXc satisfies several exact requirements. Many functional forms for f have been
suggested, a list of which is too long to be given here. Here, we only describe a
relatively simple one due to Perdew et al. (1996). To begin with, the total electron
density ρ is written as the sum of up and down spin densities, ρ(r) = ρ(r)↓ +
ρ(r)↑ , EXc = EX + Ec , and Xc = X + c . Perdew et al. (1996) proposed that

Ec (ρ) = ρ(r)[c (rc , ζ ) + H (rs , ζ, t)]d 3 r, (22)

where rs is called the Seitz radius (such that ρ = 3/4π rs3 ), ζ = (ρ↑ − ρ↓ )/ρ,
and t = |∇ρ|/(2ks φρ) is a dimensionless density gradient. Here, φ(ζ ) = 12 [(1 +
ζ )2/3 + (1 − ζ )2/3 ], and ks is the Thomas-Fermi screening wave number.
Moreover, c (rs , ζ ) = (e2 /a0 )φ 3 [γ ln(rs /a0 ) − ω], with γ = (1 − ln 2)/π 2 
0.031091 and ω  0.046644.
The function H is selected in such a way that EXc satisfies several rigorous
constraints. The proposed form for H is given by
$ 2%  $ %
e β 1 + At 2
H = γ φ3 ln 1 + t 2 , (23)
a0 γ 1 + At 2 + A2 t 4
with
 $ % −1
β a0 c
A= exp − 3 2 − 1 . (24)
γ γφ e
In these equations, e is the electron charge, a0 = h̄2 /me2 , and β  0.066725.
The EX portion of EXc is written as

EX = ρ(r)x (ρ)FX (s)d 3 r, (25)

where X (ρ) = −3e2 (3π 2 ρ)1/3 /(4π ), and s = (rs /a0 )1/2 φt/c is, similar to t, a
dimensionless density with c  1.2277. The function FX (s) is given by
κ
FX (s) = 1 + κ − , (26)
1 + µs 2 /κ
with κ = 0.804 and µ = βπ 2 /3  0.21951. Let us emphasize that, although we
only provided here the numerical values of the several constants that appear in
these equations, they are in fact related to fundamental physical constants, and
do not represent numerical fits to some experimental data. These equations have
been shown to provide very accurate predictions for atomization energies of many
molecules, and therefore are widely used in the DFT computations.
 9.1. Density-Functional Theory 465

Figure 9.1. A supercell for a bulk solid with a

point defect at its center. The cell is enclosed
by dashed lines, with the rest being its periodic
images (after Payne et al., 1992).

9.1.3 Nonperiodic Systems

While the DFT computations are carried out most conveniently for periodic sys-
tems, they run into difficulty when the system contains some sort of a defect. Such
systems have attracted wide attention, especially when they contain a large number
of atoms, in which case they are referred to as mesoscopic systems. Examples of
such systems are abundant and include scanning tunneling microscope tip and sur-
face, grain boundaries, quantum dots and wires, and biological macromolecules.
To study such systems with the DFT a periodic supercell is used, an example of
which for a system with one defect is shown in Figure 9.1. The supercell contains
the defect which is surrounded by a region of bulk crystal. Periodic boundary con-
ditions are applied to the supercell, so that it is replicated throughout space. Thus,
due to periodicity, one actually computes the energy per unit cell of a crystal con-
taining an array of defects, rather than the energy of a crystal containing a single
defect. To prevent the defects in the neighboring cells to interact appreciably with
each other, one must include enough bulk solid in the supercell.
Another case for which a supercell must be used in the computations is when
a surface is only partially periodic. For example, it may have periodicity in its
own plane, but not in the direction perpendicular to its plane. The supercell for
such systems is shown in Figure 9.2. It contains a crystal slab and a vacuum
region, and is repeated over all space, so that the total energy of an array of crystal
slabs is calculated. In order to ensure that the results of the computations are true
representative of an isolated surface, the vacuum region must be wide enough
that faces of adjacent crystal slabs do not interact across the vacuum. The crystal
slab must also be thick enough that the two surfaces of each crystal slab do not
interact through the bulk crystal. Even molecules can be studied in this fashion;
see Joannopoulos et al. (1991).

9.1.4 Pseudopotential Approximation

For periodic systems, one takes advantage of Bloch’s theorem to simplify the
computations.According to this theorem (Ashcraft and Mermin, 1976) in a periodic
solid material each electronic wave function can be written as the product of a
466 9. Atomistic Modeling of Materials

Figure 9.2. A supercell for a surface of a bulk solid.

The cell is enclosed by dashed lines, with the rest
being its periodic images (after Payne et al., 1992).

cell-periodic part and a wavelike part:

i (r) = exp(ik · r)fi (r). (27)
The cell-periodic part of i can be expanded using a basis set that consists of a
discrete set of plane waves with wave vectors that are reciprocal lattice vectors of
the crystal. Therefore,


fi (r) = ci,G exp(iG · r), (28)
G

where the reciprocal lattice vectors l are defined by, G · l = 2π m for all l, with l
being a lattice vector of the crystal and m an integer. Therefore,


i (r) = ci,k+G exp[i(k + G) · r]. (29)
G

In practice, this series is truncated to include only plane waves that have kinetic
energies less than a cutoff energy. This introduces some error into the computations,
but the error decreases with increasing energy cutoff. A suitable value of the energy
cutoff can be selected by an optimization technique (Rappe and Joannopoulos,
1991), since the value of the cutoff is not included in the theory itself. Substitution
of Eq. (29) in (16) converts the Kohn–Sham equations into a relatively simple
set of equations in terms of the coefficients ci,k+G and the eigenvalues k in
which the kinetic energy is diagonal. Even for systems which contain aperiodicity,
such as those with defects, use of a supercell makes the system amenable to this
type of analysis. In any event, the solution of the Kohn–Sham equations, when
written in terms of the eigenvalues of the coefficients ci,k+G , can be obtained by
diagonalization of the associated matrix. The size of the matrix is dictated by the
choice of the cutoff energy, and can be prohibitively large if the system contains
both valence and core electrons. The problem can be overcome by use of the
pseudo-potential approximation which we now describe.
 9.1. Density-Functional Theory 467

It is well known that many properties of solid materials depend on the valence
electrons much more strongly than on the core electrons. This fact is exploited
in the pseudo-potential approximation by which one removes the core electrons
and replaces them and the strong ionic potential by a weaker pseudo-potential
that acts on a set of pseudo-wave functions, rather than the true valence wave
functions (Vanderbilt, 1990). The valence wave functions oscillate strongly in
the region occupied by the core electrons, due to the strong ionic potential in this
region. Ideally, the pseudo-potential must be constructed in a way that its scattering
properties for the pseudo-wave functions are identical to those of the ion and the
core electrons for the valence wave functions. But, this must be done in a way that
the pseudo-wave functions have no radial nodes in the core region, where the total
phase shift produced by the ion and the core electrons will be greater by π , for each
node that the valence functions had in the core region, than the phase shift produced
by the ion and the valence electrons. The phase shift generated by the ion core is
different for each angular momentum component of the valence wave function,
and therefore the scattering from the pseudo-potential must depend on angular
momentum. Outside the core region, the two potentials and their scatterings are
identical. The most general form for a pseudo-potential is then given by


Ups = |lmul lm|, (30)
lm

where |lm are the spherical harmonics, and ul is the pseudo-potential for angular
momentum l.Acting on the electronic wave function with this operator decomposes
the wave function into spherical harmonics, each of which is then multiplied by
the relevant pseudo-potential ul .
A pseudo-potential that does not depend on the angular momentum components
of the wave function is called a local pseudo-potential, which is a function only
of the distance from the nucleus. While it is possible to generate arbitrary, pre-
determined phase shifts for each angular momentum state with a local potential, in
practice there are limits to the amount that the phase shifts can be adjusted for the
different angular momentum states, since one must preserve the smoothness and
weakness of the pseudo-potential, without which it becomes difficult to expand
the wave functions using a reasonable number of plane-wave basis states.
In order for the exchange-correlation energy to be represented accurately, the
pseudo and the true wave functions must be identical outside the core region. If
one adjusts the pseudo-potential to ensure that the integrals of the squared ampli-
tudes of the true and pseudo-wave functions inside the core region are identical,
the equality of the pseudo and true wave functions outside the core region is
guaranteed. Pseudo-potentials that possess this property were first constructed by
Starkloff and Joannopoulos (1977), and have been shown to be highly accurate
for heavy atoms. Moreover, Hamann et al. (1979) pointed out that, a match of
the pesudo and real wave functions outside the core region also ensures that the
first-order energy dependence of the scattering from the ion core is correct, so that
the scattering is accurately described over a wide range of energy. A technique for
468 9. Atomistic Modeling of Materials

constructing pseudo-potentials that corrects even the higher-order energy depen-

dence of the scattering was introduced by Shirley et al. (1989). In the best case
scenario, one should develop non-local pseudo-potentials that use a different po-
tential for each angular momentum component of the wave function, and efforts in
this direction (see, for example, Shirley et al., 1989, and references therein) have
also been fruitful.
The following procedure is typically used for constructing an ionic pseudo-
potential. The computations are carried out with a periodic unit cell (for non-
periodic systems, see Section 9.1.3). All-electron computations are carried out for
an isolated atom in its ground state and also some excited states, using a given
equation for the exchange-correlation density functional, which result in valence
electron eigenvalues and valence electron wave functions for the atom. A pseudo-
potential with a few parameters is then selected and the parameters are adjusted
in such a way that a pseudo-atom calculation using the same form for exchange-
correlation as in the case of the all-electron atom yields both pseudo-wave functions
that match the valence wave functions beyond some cutoff radius rc , and pseudo-
eigenvalues that are equal to the valence eigenvalues. The ionic pseudo-potential
so obtained is then utilized, without further modifications, for any environment
of the atom. The electronic density in any new environment of the atom is then
determined utilizing both the ionic pseudo-potential obtained in this way and the
same form of exchange-correlation functional employed in the construction of the
ionic pseudo-potential.
One advantage of pseudo-potential method is that, by replacing the true ionic
potential by a weaker pseudo-potential, one expands the electronic wave functions
using far fewer plane-wave basis states that would be necessary if the full ionic
potential were to be used. Moreover, the rapid oscillations of the valence wave
functions in the cores of the atoms are removed, and the small core electron states
are no longer present. Another advantage of the pseudo-potential approximation
is that fewer electronic wave functions must be calculated.
Total-energy pseudo-potential computations require significant amounts of com-
puter time, even when the number of atoms in the unit periodic cell is small.
Moreover, the computational time always increases with the number of atoms in
the unit cell, and therefore use of the most efficient computational method is crucial
to the success of this method. The quantum molecular dynamics method devised
by Car and Parrinello (1985) has improved the efficiency of these computations
dramatically, hence allowing one to simulate systems with a large number of atoms.
As already mentioned above, since the essence of the total-energy pseudo-potential
computations is finding the electronic states that minimize the Kohn–Sham energy
functional, one can also attempt to directly minimize the energy functional by a
method such as the simulated annealing method described in Chapter 3 of Volume
I, or by the conjugate-gradient technique which is used for minimization prob-
lems. After we describe the classical MD simulation, the QMD method of Car
and Parrinello and the conjugate-gradient technique for directly minimizing the
Kohn–Sham energy functional will be described.
 9.2. Classical Molecular Dynamics Simulation 469

9.2 Classical Molecular Dynamics Simulation

Historically, the first true MD simulation seems to have been carried out by Alder
and Wainwright (1957), who studied a system with only a few hundreds hard-
sphere particles and discovered a fluid-solid phase transition. At the same time,
Wood and Parker (1957) investigated the properties of simple fluids using the
Monte Carlo method. Rahman (1964) was apparently the first to carry out MD
simulations using the Lennard–Jones potential (see below for a description of
this potential). Unlike Alder and Wainwright, Rahman’s work was the first to
involve particles with smoothly varying potentials. He computed the diffusion
coefficient and pair-correlation function for liquid argon and showed them to be
in very good agreement with the experimental data. These three pioneering works
opened the way for simulation, and hence understanding, of many-body systems.
Later work by Verlet (1967), whose method helped MD simulations to become
much more efficient (see below), by Alder and Wainwright (1969) who discovered,
unexpectedly, an algebraic long-time tail in the velocity autocorrelation functions
of hard sphere (a discovery that intensified further the interest in MD simulations),
and by Rahman and Stillinger (1971), who addressed simulations of such complex
molecules as liquid water, firmly established the classical MD simulations as an
every-day tool of studying fluids and materials.
The 1970s witnessed further improvements in methodologies and algorithms for
MD simulations. For example, Evans and Murad (1977) succeeded in developing
an algorithm for computing molecular rotations [Ciccotti et al. (1982) made further
improvement to this method], and Bennett (1976) and Torrie and Valleau (1977)
developed efficient methods for measuring the free energies (see also Frenkel and
Ladd, 1984), and so on. Much more progress was made in the 1980s when An-
dersen (1980) and Parinello and Rahman (1981) developed methods for carrying
out MD simulations under constant pressure and constant temperature, and Nosé
(1984) developed equations for simulating constant-temperature MD simulations
by introducing additional degrees of freedom [his equations were simplified by
Hoover (1985); see below]. At the same time, the advent of vector computers fur-
ther motivated the search for methods that could take advantage of vectorization
techniques, especially those that could be used with the Verlet algorithm. All of
these advances took MD simulations to a stage where, by mid 1980s, they could
be used for studying non-equilibrium systems. Abraham et al. (1984) studied,
using MD simulations, the incommensurate phase of Krypton on graphite using
more than 160,000 atoms (a “revolution” for its time), and Car and Parinello (1985)
succeeded in combining MD simulations and electronic structure calculations (see
Section 9.4). The first million-particle MD simulations were carried out by Swope
and Andersen (1990) who studied homogeneous nucleation of crystals in a super-
cooled atomic liquid (i.e., below its freezing point). Their study was important, not
only because of the large number of atoms that had been used, but also because it
showed that certain physical phenomena can be reproduced in the MD simulations
only when the size of the system is large enough.
470 9. Atomistic Modeling of Materials

Generally speaking, two types of classical MD simulations can be carried out.

Equilibrium MD simulations are suited for systems that can, in principle, be treated
by statistical mechanics. This type of MD simulations can yield equilibrium prop-
erties of materials. Non-equilibrium MD techniques are appropriate for systems
that are under the influence of an external driving force, and are most suitable for
computing the transport properties of a system. We first discuss general concepts
and ideas of MD simulations that are applicable to both the equilibrium and non-
equilibrium methods, after which we describe those aspects of non-equilibrium
MD technique that are different from the equilibrium methods.

9.2.1 Basic Principles

Molecular dynamics simulation of any phenomenon consists of integration of
Newton’s equation of motion for a system of N particles that represent the material
or the system under study. Therefore, the MD method is a way of simulating the
behavior of a system as it evolves with time since, unlike the Monte Carlo (MC)
method, in the MD simulations the system moves along its physical trajectory.
The main advantage of the MD method over the MC technique is that, not only it
provides a method for computing the static properties of a system, but also allows
one to calculate and study the dynamical properties, including dynamic fracture
propagation in materials that are of interest to us in this book.
Suppose that we have a collection of N particles in a simulation cell with
dimensions Lx , Ly and Lz . The particles interact with each other, and for simplicity
we assume for now that the interaction force can be written as a sum over pair
forces F(r), the magnitude of which depends only on r, the distance between the
particle pairs. Thus, the force acting on any particle i is given by


N
Fi (rN ) = F (| ri − rj |)r̂ij , (31)
j =1, j
=i

where rN = {r1 , r2 · · · , rN } is the position coordinates of all the particles, and r̂ij
is a unit vector along ri − rj , pointing from particle i to j . Then, the equation of
motion for particle i is given by
d 2 ri (t)
mi = Fi (rN ) + Fe , (32)
dt 2
where mi is the atomic mass of particle i, and Fe represents all the external forces
that are imposed on the system, either by nature (for example, the gravitational
force) or by the experimentalist (for example, an external pressure gradient applied
to the system). Molecular dynamics consists of writing Eq. (32) for all the N
particles of the system and integrating them numerically and simultaneously (since
the equations are coupled through the force Fi ). The solution of this set of equations
describes the time evolution of the system. If the forces between the particles
depend only on their relative positions, then the system’s energy and momentum
are automatically conserved.
 9.2. Classical Molecular Dynamics Simulation 471

To check whether the system has reached a steady state, if one exists, quantities
such as the kinetic energy are computed and their variations with the time are
monitored. When these quantities no longer vary with the time appreciably, then
attainment of the steady state has been confirmed. The time to reach a steady state
depends on the initial state of the system, and how far it is from its steady state.
Although the MD approach is, in principle, a rigorous method, in practice, and
similar to almost all computational strategies for studying a phenomenon, it is only
an approximate technique. Thus, it should be used with considerable care. Some
of the problems that one must pay particular attention to are as follows.
(1) The interaction potentials between the particles are not, in almost all cases,
known, and therefore one must use approximate expressions for describing
these potentials. In principle, quantum-mechanical calculations can be used
for determining these forces, but such computations can be subject to errors.
Typically, the interaction potentials or forces are written in terms of several
parameters which are determined either by ab initio computations or by fitting
the results to experimental data (see Section 9.7).
At atomic scale, the interactions are either of intra-molecular or inter-
molecular type. We will discuss the intra-molecular interactions separately,
and for now briefly describe the inter-molecular interactions. In many MD
simulations the interaction potential between a pair of particles, the centers
of which are a distance r apart, is represented by the classical Lennard–Jones
(LJ) potential. It was thought for a long time that this potential is too simple to
mimic the behavior of real materials, particularly brittle ones. However, using
what the physics Nobel Laureate R. P. Feynman emphasized nearly 40 years
ago as a guide, the value of the LJ potential is in its universal nature, since
according to Feynman the single most important statement describing our real
world is that, all things are made of atoms, little particles that move around in
perpetual motion, attracting each other when they are a little distance apart,
but repelling upon being squeezed into one another, which is precisely how the
LJ and similar potentials have been constructed for representing real atoms,
fluids, and materials. For use in MD simulation and modeling of many proper-
ties of materials, including their dynamic fracture which is of interest to us in
this chapter (see Section 9.8), a relatively general potential can be constructed,
part of which is based on the classical LJ potential. This potential is given by
(Holian et al., 1991)
⎧   
⎪
⎪ σ 12  σ 6
⎪
⎪ 4 − , r < ri
⎨ r r
U (r) = (33)
⎪
⎪ −a1 (rc2 − r 2 )2 + a2 (rc2 − r 2 )3 , ri ≤ r ≤ rc
⎪
⎪
⎩
0, r > rc

where  is the energy parameter of the potential (the maximum energy of

attraction between a pair of molecules), or the LJ well depth, and σ is the
size parameter (or the distance at which the LJ potential passes through zero),
472 9. Atomistic Modeling of Materials

also called the collision diameter. Note that σ is not the same as molecular
diameter of the molecules, although the two quantities are usually close to
each other. The 12-6 part of U (r) is the classical LJ potential in which the
r −12 term represents a hard-core or repulsive potential, while the r −6 term is
its attractive part. The second equation (for ri ≤ r ≤ rc ) has been added for
accommodating the difference between brittle and ductile materials. Here ri
is the inflection point in the potential, and
  $ %1/2 
1 24U LJ (r i )
rc2 = ri2 5 − 5 1 − 9−  (r ) , (34)
25 ri ULJ i

5ri2 − rc2 
a1 = ULJ (ri ), (35)
8ri3 (rc2 − ri2 )

3ri2 − rc2 
a2 = ULJ (ri ). (36)
12ri3 (rc2 − ri2 )2
If the intention is to study the properties of materials via MD simulations
(without studying their fracture), then typically only the 12 − 6 part of the
potential is used. The force F(r) between the particles is then given by, F(r) =
−∇U (r) for r ≤ rc ; of course, F=0 for r > rc . If the system contains solid
walls, then the interactions between the materials’ atoms and those of the walls
must also be taken into account. A well-known potential due to Steele (1973)
has been used in many simulations (although Steel’s potential represents some
sort of a mean-field approximation, as it assumes that the wall is smooth and
structureless):
 $ % $ %4 
2 σw 10 σw σw4
Uw = 2πρw w σw  2
− − , (37)
5 z z 3(z + 0.61)3

where w and σw are the energy and size parameters that characterize the
interactions between the atoms in the system and those of the walls, z is the
vertical distance from the wall, ρw is the density of the wall’s atoms, and  is
the distance between the atomic layers within the wall.
The accuracy of the MD method depends to a large extent on the accuracy of
the interaction potentials used in the simulation. Lennard–Jones type potentials
are too simple to represent complex atoms and molecules. To remedy the
situation, one can fit the size and energy parameters of the LJ potential, so that
certain predictions of the MD simulations fit, in some sense, the experimental
data. For example, these parameters can be estimated from the properties
of fluids at the critical point, liquids at the normal boiling point, or solids
at the melting point. This method has been reasonably accurate for relatively
simple molecules and might, in some cases, provide qualitative insight into the
behavior of materials. However, many materials of technological interest have
strong and specific chemical interactions that cannot be described by simple,
pairwise additive potentials, such as the LJ potential. Thus, more sophisticated
 9.2. Classical Molecular Dynamics Simulation 473

computations must be used for determining the interaction potentials. We will

come back to this issue later in Section 9.7
(2) Molecular dynamics suffers from the same problem that every computer sim-
ulation of a physical system suffers from, namely, while the intention is to
simulate a real system which, at least in atomic units, is very large, the simu-
lated system is of finite size. The finite-size effect can be particularly severe
if there are correlations in the system (as is the case with almost any physical
system). If the correlation length is much smaller than the linear size of the
system, finite-size effects do not pose any significant problem. If the correla-
tion length is much larger than the linear size of the system, then one can use
finite-size scaling (see Chapter 2 of Volume I) for extrapolating the results for
a system of finite size to one with an infinite size (although finite-size scaling
is usually used for second-order phase transitions, and the MD methods are
not used very often for studying such transitions, as the computations would
be very intensive). The intermediate case, in which the correlation length is
not too large or too small, is usually addressed by using the periodic bound-
ary conditions, since the finiteness of the system is manifested through its
boundaries (an infinite system does not have any boundary!). Using periodic
boundary conditions means that the finite simulation cell is embedded in an
infinite system, such that it is surrounded by replicas of itself on all sides. In
that case,
 


3
F(ri − rj ) = F | ri − rj + Vk n k | , (38)
n k=1
where Vk are vectors along the edges of the rectangular simulation cell. The
first sum on the right-hand side of Eq. (38) is over all vectors n = (n1 , · · · , nk ).
The force
 F is along the line connecting particle i and the image particle
rj − 3k=1 Vk nk . In principle, calculating terms of this infinite sum until it
converges to a well-defined value is a difficult task, but methods have been
developed for computing such sums. Use of periodic boundary conditions
also has a negative side effect: In a periodic system the angular momentum
is not conserved, since the periodic boundary conditions break the spherical
symmetry of the interactions.
(3) When time-averaged properties are calculated, the averaging is clearly carried
out over a finite-time period. There is, however, a limitation in time as a result
of the finite number of integration steps that one can carry out. Finite size of
the system also limits the time, especially if the particles (in, for example,
simulation of liquids or gases) travel more than half the linear size of the
simulation cell.
(4) In any MD simulation, there is always a competition between the speed of
the computations and their accuracy. Normally, as the size of the time step
increases, so also does the inaccuracy in the simulation results. Therefore, for
any MD simulation, there is an optimal choice of the time step.
Having described the strengths and weaknesses of the classical MD simulation,
let us now describe some basic issues that arise in such computations.
474 9. Atomistic Modeling of Materials

9.2.2 Evaluation of Molecular Forces in a Periodic System

From the computational view point, the most intensive part of any MD simulation
is calculation of the forces between the particles which accounts for 70 − 90% of
the total time. Periodic boundary conditions create a problem for evaluation of the
forces between the particles, since in a periodic system not only does a particle
interact with other particles in the simulation cell, but also with those in the images
of the system that surround it, and thus, in principle, one must sum over an infinite
number of interactions. However, in many cases the force between two particles
decays rapidly as the distance between them increases, and therefore the particles
that are far from any given particle, whether they are in the simulation cell or in
its images, do not contribute significantly. If the force between two particles can
be ignored for distances that are larger than half the system’s linear size, then one
can use the minimum image convention according to which, for each particle in the
system, one takes into account only the interactions with the nearest copy of each
of the remaining particles, implying that each infinite sum over all the images is
replaced by a single term. For example, for a cubic simulation cell, the minimum
interaction distance is given by
(rij )min = min|ri − rj + Vk nk |, (39)
where the notations are the same as in Eq. (38). Of course, the potentials will no
longer be analytic, but the discontinuities will not be important if the potential is
small for distances that are larger than half the system’s linear size.
In practice, what is done in most cases is cutting the interactions off at a distance
rc . Typically, rc is set to be a multiple of the effective molecular diameter of the
largest atom in the simulations, and therefore it is usually smaller than half the
system’s linear size. If this approach is taken, then at every step of the integration
one must check, for any particle i, the distances of all other particles from i to see
whether they are at a distance larger or smaller than rc . This search constitutes one
the three important tasks in any MD simulation (the other two being computing
the forces, and integrating the equations of motion). There are efficient methods of
doing this which we will discuss below. We must point out, however, that cutting
off the interaction potentials violates energy conservation, although if rc is selected
carefully, the effect will be small. Moreover, by shifting the interaction potentials
one can avoid violation of energy conservation altogether by writing

Uo (r) − Uo (rc ) if r ≤ rc
U (r) = (40)
0 if r > rc
where Uo (r) represents the original interaction potential to be used. However,
this shift does not affect the force resulting from the shifted potential; it remains
discontinuous at rc . In order to make the force also continuous at the cutoff point,
we write
 d
Uo (r) − Uo (rc ) − [Uo (rc )(r − rc )] if r ≤ rc
U (r) = dr (41)
0 if r > rc
 9.2. Classical Molecular Dynamics Simulation 475

This algorithm was first suggested by Stoddard and Ford (1973). The actual number
of interacting particles (i.e., those that are within a sphere of radius rc , centered at
the center of a given particle) is a function of the molecular density. For example, in
the simulation of a typical liquid state using the LJ potential, the computation of a
single pair-interaction requires about 30 − 40 floating-point operations. Therefore,
a complete force calculation requires of the order of 2,000 floating-point operations
per particle, still computationally intensive, but much more efficient than a full N -
body calculation. Let us point out that, the cut-and-shift procedure cannot be used
if electric and gravitational forces are operative in the system, since they decay
only as 1/r. Such cases must be treated separately; see Section 9.2.8.

9.2.3 The Verlet and Leapfrog Algorithms

In a typical MD simulation, the time that the computer program spends for inte-
grating the equations of motion is about 2–3% of the total time. However, accurate
integration of the equations of motion is the most important part of the compu-
tations. Various methods have been proposed for achieving this goal, a detailed
discussion of which can by itself be the subject of a minireview. One heavily-used
procedure is due to Verlet (1967). According to his method, the algorithm for in-
tegrating the equation of motion for a single particle, which is subjected to a force
F that depends only on the particle’s position, is given by
t 2
r(t + t) = 2r(t) − r(t − t) + F(r), (42)
m2
where r(t) is the position of the particle at time t, and t is the integration time
step (note that t = nt, where n is an integer). The error per time step is usually
O(t 4 ), but in some cases can be as large as O(t 2 ) which is still very accurate.
As Eq. (42) indicates, the integration can proceed if the positions of the particles
at two previous times (t and t − t) are known. Thus, to begin the integration, the
positions at t = t are first computed from
1
r(t) = r(0) + tv(0) + t 2 F[r(0)] + O(t 3 ), (43)
2m
which, together with the initial positions, provide us with the two previous positions
that we need. The particles’ velocities at any time t are calculated from
r(t + t) − r(t − t)
v(t) = + O(t 2 ), (44)
2t
which is a standard finite-difference approximation to v. If periodic boundary con-
ditions are used, then one must check whether any particle has left the simulation
cell in the last integration step, in which case the particle must be translated back
over a lattice vector Vk to keep it inside the cell. Clearly, the velocity evaluation
step must be carried out before such a translation.
The Verlet algorithm, in its original form, is sometimes susceptible to error. A
modified algorithm which is the exact arithmetic equivalent of the original Verlet
476 9. Atomistic Modeling of Materials

algorithm, but is far less susceptible to numerical errors, is the leap-frog algorithm,
according to which one computes the velocity of a particle at midpoint between t
and t + t,
$ % $ %
1 1 t
v t + t = v t − t + F[r(t)], (45)
2 2 m
from which the position of the particle is calculated,
$ %
1
r(t + t) = r(t) + t v t + t . (46)
2
In another modification of the Verlet algorithm, the so-called velocity-Verlet
algorithm, one calculates the position and velocity of a particle from

t 2
r(t + t) = r(t) + tv(t) + F[r(t)], (47)
2m
t
v(t + t) = v(t) + {F[r(t + t)] + F[r(t)]}. (48)
2m
This algorithm is most stable with respect to the finite precision arithmetic, and
requires no additional computations in order to calculate the velocities.
An important property of the Verlet algorithm and its leap-frog modification is
that, the energy that one calculates when using these integration methods does not
exhibit any drift in the total energy. This important and desirable stability is due to
the fact that the Verlet algorithm is time-reversible, and therefore does not permit
steady increase or decrease of the energy for periodic systems. An important time
scale in the simulations is the so-called Poincaré time scale, which is the time
after which a system that starts out with a random configuration returns to its
initial configuration. The total time that one can integrate in any MD simulation is,
however, much smaller than the Poincaré time, and therefore there is the possibility
of having an increasing error in the calculated energy as the equations of motion
are integrated for larger times. However, the Verlet algorithm has an additional
property called symplecticity. This property gives rise to conserved quantities,
and in particular Sanz-Serna (1992) showed that, with this property present in
the integration procedure, the discrete analogue of the total energy (in numerical
integrations one can compute only discrete analogues of the properties of interest)
is rigorously conserved, and that the discrete analogue of the total energy deviates
from its continuum (that is, actual) counterpart by an amount which is of the order
of O(t k ), where k is some relatively large integer. Therefore, the Verlet algorithm
and its leap-frog version hold the deviations of the energy bounded. Such desirable
properties are the main reason for the popularity of the Verlet algorithm and its
modifications. In contrast with the Verlet algorithm, the Gear method (Gear, 1971),
which is a predictor-corrector technique, has lost its popularity, despite the fact
that it requires only one force evaluation per time step, because it is not symplectic
and thus can create major problems for those properties of the system that are
supposed to be conserved and bounded.
 9.2. Classical Molecular Dynamics Simulation 477

9.2.4 Constant-Energy Ensembles

Since a MD simulation conserves the number N of the particles and the system
volume , then, if the energy E and momentum are also conserved, the time
averages of any physical property computed from this type of simulations will be
equal to the averages in the microcanonical or the (NVE) ensemble. Let us describe
how such MD simulations are carried out. One first specifies the number of particles
and the interaction potentials between them, and assigns them their initial positions
and momenta. Since the temperature of any system can be measured, and because
one is usually interested in carrying out the simulations at a specified temperature,
rather than a specified energy, one must “push” the system toward the desired
temperature. There is an efficient way of achieving the desired temperature, which
will be discussed below.
If the LJ potential is used, then the sites of the FCC lattice are usually used as the
initial position, since the FCC lattice represents the ground state of the LJ potential,
although any other initial positions can also be utilized. The initial velocities v are
drawn from a Maxwell–Boltzmann distribution with the specified temperature T :
$ %3/2
m
f (vx , vy , vz ) = exp(−mv 2 /2kB T ), (49)
2π kB T
where v 2 = vx2 + vy2 + vz2 , and kB is the Boltzmann’s constant. This is done by
drawing the x, y and z velocity components for each particle from a Gaussian
distribution. After generating the initial momenta, the average momentum per
particle p is computed and subtracted from the individual momentum pi of each
particle. This ensures that the total initial momentum of the system is zero. Once the
initial configuration of the system has been prepared, the integration of Newton’s
equation of motion begins.

9.2.5 Constant-Temperature Ensembles

The temperature of a system is a property that can be easily measured and con-
trolled, and therefore it is often desired to perform an experiment under a fixed
temperature condition. However, in MD simulations it is often very difficult to
fix the temperature of the system at the very beginning. Since the temperature
of an infinite system is proportional to the average kinetic energy per degree of
freedom, with a proportionality constant 12 kB , this quantity is used in MD simula-
tions (of finite systems) to fix the temperature. One procedure for obtaining a fixed
temperature Tf is to calculate a rescaling factor α given by
 1/2
3kB Tf (N − 1)
α= N , (50)
2
i=1 mi vi

where mi and vi are the mass and magnitude of velocity of particle i, respectively.
The velocity vi of each particle i is then rescaled by α, i.e., vi (t) → αvi (t). The
rescaling of the velocity also necessitates rescaling of the time t, which will be
478 9. Atomistic Modeling of Materials

discussed shortly. Each time this rescaling is done, the actual temperature of the
system changes and gradually approaches the given fixed temperature Tf . This
method can actually be derived by imposing a constant kinetic energy through a
Lagrange multiplier term added to the Lagrangian of the isolated system (Haile
and Gupta, 1983). However, although the rescaling procedure can be derived, due
to certain assumptions used in the derivation, it still represents some sort of an ad
hoc method. This has motivated the development of more systematic methods for
achieving a given temperature and thus being able to perform MD simulations at
constant temperature. One particularly effective method is based on introducing
an extra force acting on the particles. The force is frictional in nature, is assumed
to be proportional to the velocity of the particles, and therefore affects the kinetic
energy of the system (and hence the system’s temperature) in a direct way. Thus,
the equation of motion for the ith particle is written as

d 2 ri dri
m 2
= Fi (R) − Cf (R, dR/dt) , (51)
dt dt

where the friction parameter Cf (R, dR/dt) is assumed to be the same for all the
particles. The sign of Cf depends on whether heat is added to or extracted from
the system; in the former case Cf < 0 while in the latter case Cf > 0. Various
equations have been suggested for Cf . The best-known equation was proposed by
Nosé (1984) and simplified by Hoover (1985), according to which

dCf i mi vi2 − 3N kB Tf
= , (52)
dt f

where f is a parameter that must be selected carefully. In effect, in order to reach

the desired temperature, the system is connected to a heat bath to exchange heat
with it so that it can reach the intended temperature, and the parameter f represents
a coupling between the system and the heat bath. The Nosé-Hoover method yields
precise canonical distribution for the particles’ positions and momenta.
Despite many desirable features, the Nosé-Hoover method does have some
shortcomings, the most serious of which is that the coupling parameter f must
be chosen. Moreover, although Cho and Joannopoulos (1992) showed that, for
LJ fluids at high temperatures, the canonical distribution is reproduced correctly,
Holian et al. (1995) demonstrated that if the temperature is lowered, it begins
to oscillate with an amplitude which is much larger than the standard deviations
expected in the canonical ensemble.
Many modifications of the Nosé-Hoover algorithm have been proposed. Chief
among them is an approach proposed by Jellinek and Berry (1988) who suggested
a generalization of the Nosé Hamiltonian involving multiplicative scaling of co-
ordinates, momenta, and time. They showed that there are infinitely many distinct
Hamiltonians (i.e., distinct dynamics) that possess all the properties of the Hamil-
tonian dynamics of Nosé-Hoover algorithm. Their model was analyzed in detail
by Brańka and Wojciechowsky (2000).
 9.2. Classical Molecular Dynamics Simulation 479

9.2.6 Constant-Pressure and Temperature Ensembles

In addition to temperature, it is often desirable to be able to carry out MD simu-
lations at a constant pressure [the (NPT) ensemble]. This is achieved by uniform
isotropic volume changes caused by rescaling the atomic coordinates. Andersen
(1980) proposed incorporating the volume  of the system into the equation of
motion by rescaling the coordinates as
ri = ri −1/3 , (53)
where ri denotes the coordinates of the particles in the rescaled system.
Furthermore, the momentum of the particle is also rescaled according to
pi = pi (s1/3 ), (54)
where s denotes a new dynamical variable that is equivalent to rescaling the real
time, dt = s(t  )dt  . In effect, the system is connected to a “piston” that can contract
or expand it. The Hamiltonian of the system is given by
1
p2i 1
pb2
H= + E p (r ij  1/3
) + + P
2 ms 2 2/3 2 2f
i ij,i
=j (55)
pp2
+ + (3N + 1)kB T ln s,
2mp
where pp and mp are the “momentum” and “mass” of the “piston,” respectively,
Ep is the potential energy, and (3N + 1) represents the total number of independent
momentum degrees of freedom of the system. The term involving pb represents
the coupling to the heat bath (so that the temperature is also held fixed), while P 
and pp2 /2mp represent the work and kinetic energy arising from the connection
of the system to the “piston.” The governing equation for the dynamical variable
s(t) is simply
ds pb ∂H
= = . (56)
dt f ∂pb
Note that if in Eq. (55) we do not rescale the lengths and momenta and delete the
“piston” terms, then the modified equation also describes the system for the case
in which only the temperature is held fixed, implying that in that case too, the time
must be rescaled in the same way as in the present case.
Martyna et al. (1992) suggested a Hamiltonian slightly different from (55), and
also equations of motion that differed from those used in the above method, in
order to remove a minor problem from the above formulation, namely, the fact
that the trajectories of the particles produced by the above method depend on the
choice of the basis lattice vector.

9.2.7 Simulation of Rigid and Semirigid Molecules

Accurate molecular simulations depend largely on realistic representation of the
molecules. Although representing atoms or molecules as simple LJ hard spheres
480 9. Atomistic Modeling of Materials

Table 9.2. The interaction parameters for CO2 .

σOO (Å) 3.027 εOO /kB (K) 74.8
σCO (Å) 2.922 εCO /kB (K) 44.8
σCC (Å) 2.824 εCC /kB (K) 26.2
qO (e) −0.332 qC (e) +0.664
M(cm2 ) −14.4 × 10−40 CO (Å) 2.324

may be adequate for predicting many qualitative features of experimental data,

more sophisticated and realistic representation of the atoms is necessary if MD
simulations are to provide quantitative predictions. As a relatively simple but
important example, consider molecular modeling of CO2 . A molecular model of
CO2 , which is much more realistic than a simple LJ hard sphere with effective
parameters σ and ε, was developed by Murthy et al. (1983), and further refined
by Hammonds et al. (1990). In their model CO2 is represented by a rigid linear
molecule with quadrupole moments and three LJ sites. Three partial charges, qO ,
qC and qO , are used on the O–C–O sites, chosen so as to preserve the quadrupole
moments of the molecule. The non-electrostatic interactions are still modeled as
site-site LJ potentials, where the interaction sites are located on the three atoms. To
fit the interaction parameters, three states of CO2 (namely, solid, liquid, and gas)
are used. The LJ parameters are fitted to the phonon frequencies and lattice energy
of the solid, the thermodynamic properties of the liquid, and the second virial
coefficient of the gas. All the parameters are given in Table 9.2, where σOO , σCO
and σCC are the LJ size parameters, εOO , εCO and εCC are the LJ energy parameters
between O–O, C–O, and C–C atoms in a CO2 molecule, respectively, OC is the
distance between the O and C atoms, while M is the quadrupole moment. If the
MD simulation involves CO2 and other molecules, then the interaction between
molecules i and j is expressed as

3

3
2 3
U (rij ) = ULJ (rim,j n ) + UC (rim,j n ) , (57)
m=1 n=1

where rim,j n is the distance between the interacting pairs (site m in molecule i and
site n in molecule j ), ULJ (r) is the-cut-and-shifted LJ potential described above,
and UC (r) is the Coulomb potential given by
qim qj n
UC (r) = . (58)
r
During the MD simulations one must keep track of the coordinates of a CO2
molecules. They can be represented by the vector Vj , with j = 1, 2, · · · , NCO2 ,
where NCO2 is the total number of CO2 molecules. The vector Vj contains three
cartesian coordinates, (rx , ry , rz ), determining the position of the molecule’s cen-
ter, and two coordinates determining its orientation. The orientation of CO2 can be
determined by a unit vector e = (ex , ey , ez ), directed along the axis of the molecule
(where only ex and ey are independent, as ez = 1 − ex − ey ) and the angle that it
makes with the surface of the system’s walls.
 9.2. Classical Molecular Dynamics Simulation 481

We now describe how MD simulations of more complex materials, such as rigid

or semi-rigid molecules, are performed. The motion of a rigid molecule consists of
translations of the center of mass and rotations around this point. The force acting
between two such molecules consists of atomic pair interactions between atoms
that belong to the two different rigid molecules. One can also consider off-center
interactions, but we neglect them here. There are several methods for treating rigid
molecules, and we describe one of them which is based on imposing constraints
on the system that depend on the spatial positions, but are independent of the
velocities. Consider the Lagrangian of the system,
⎡ ⎤
 t2
1
$ dri %2 1

L0 = ⎣ mi − Ep (ri − rj )⎦ dt, (59)
t1 2 dt 2
i i
=j

where subscript 0 indicates that constraints have not been imposed yet. A constraint
is imposed on the system through a Lagrange multiplier λ(t), which is a function
of time since the constraint should hold for all times. For example, for a rigid
molecule that consists of two atoms, we impose the constraint that the distance 
between the two particles is always fixed. When such a constraint, which is often
called the bond constraint, is imposed on the system, then the Lagrangian L of the
new system (with the constraint) is given by
 t2 7 8
L = L0 − λ(t) [r1 (t) − r2 (t)]2 − 2 dt. (60)
t1

λ(t) is determined by requiring that the solution must satisfy the constraint. We
discuss this shortly.
We are familiar with the concept of the backbone of a disordered medium. The
same concept can be applied to a material at atomic scale. In this case some of
the material’s atoms form the backbone and are fixed by the bond constraints
discussed above, while the remaining atoms are fixed by linear constraints which
we discuss shortly. A good example is provided by a branched polymer in which
the backbone is made of the multiply connected atoms. To identify the backbone of
any molecular structure, some rules of thumb may be useful. For a planar molecular
structure one can consider three non-colinear atoms as a backbone since they satisfy
the bond constraint, while the rest of the atoms in the structure are constrained
linearly. In a 3D molecular structure, four backbone atoms are subjected to six
bond constraints with the remaining ones to a linear vector constraint each. A good
example is provided by the linear molecule CS2 (Thijssen, 1999), the motion of
which is described by five positional degrees of freedom, two of which define the
orientations of the molecules and three define the position of its center of mass.
Without any constraint, the three atoms have nine degrees of freedom, but three
of them are eliminated by the bond constraints, implying that we still have six
degrees of freedom, instead of the required five. The inclusion of the un-needed
degree of freedom adds to the computations and makes them inefficient. A better
procedure is to fix only the distance between the two sulphur atoms by requiring
that, |rS (1) − rS (2) |2 = 2 , and to fix the position of the carbon atom by a linear
482 9. Atomistic Modeling of Materials

vector constraint which reads

12 3
rC = rS (1) + rS (2) , (61)
2
which adds up to the four required constraints.
Therefore, let us denote by µ the linear vector constraint. Then, from the
extended Lagrangian, the equations of motion for the three atoms are given by
d 2 rS (1) 1
mS = F1 − 2λ(t)[rS (1) − rS (2) ] − µ, (62)
dt 2 2
d 2 rS (2) 1
mS = F2 + 2λ(t)[rS (1) − rS (2) ] − µ, (63)
dt 2 2
2
d rC
mC = FC + µ, (64)
dt 2
where mS and mC denote the mass of the sulphur and carbon atoms. If we now
twice differentiate Eq. (61) with respect to time and use Eqs. (62)–(64), we obtain
mC
FC + µ = (F1 + F2 − µ), (65)
2mS
which helps us eliminate µ and rewrite the equations of motion for the sulphur
atoms as
d 2 rS (1)
mS =
 dt 2 
mC mC mS
1− F1 + F2 + FC − 2λ(t)[rS (1) − rS (2) ],
2(2mS + mC ) 2(2mS + mC ) 2mS + mC
(66)
d 2 rS (2)
mS =
 dt 2 
mC mC mS
1− F2 + F1 + FC + 2λ(t)[rS (1) − rS (2) ].
2(2mS + mC ) 2(2mS + mC ) 2mS + mC
(67)
Equations (66) and (67) govern the motion of the sulphur atoms. The position of
the carbon atom is of course fixed by the linear constraint that we have imposed
on the system.
We still have one unknown, λ(t), that must be determined. Since the bond con-
straint is quadratic, elimination of λ(t) is not as easy as eliminating µ. Therefore,
λ(t) is determined at each time step by using the constraint equation, i.e., we solve
the equations of motion iteratively until their solutions satisfy the bond constraint.
Thus, if, for example, we utilize the Verlet algorithm, we can write
$ %
mC
rS (1) (t + t) = 2rS (1) (t) − rS (1) (t − t) + t 2 1 − F1 (t)
2mS + mC
mS
+t 2 FC (t) − 2t 2 λ(t)[rS (1) (t) − rS (2) (t)], (68)
2mS + mC $ %
mC
rS (2) (t + t) = 2rS (2) (t) − rS (2) (t − t) + t 2 1 − F2 (t)
2mS + mC
mS
+t 2 FC (t) + 2t 2 λ(t)[rS (1) − rS (2) ]. (69)
2mS + mC
 9.2. Classical Molecular Dynamics Simulation 483

It can be shown that the error in the numerical values of λ(t) so obtained is of the
order of O(t 4 ).
Our discussion so far has been restricted to totally rigid molecules. We now
consider partially rigid molecules that consist of rigid clusters that can move with
respect to one another. For this purpose we describe the algorithm due to Rykaert
et al. (1977), Ciccotti et al. (1982), and Rykaert (1985). Their algorithm, which
is known as SHAKE (the author does not know why this name was given to this
algorithm), is formulated based on the notion that the forces that particles experi-
ence are the physical and constraint forces. If M is the number of the constraints,
then the constraints are written as, Ck (R) = 0, with k = 1, 2, · · · , M, where Ck
expresses the functional form of the constraint, e.g., restriction that the distance
between two particles is always fixed. The constraint forces are given by,


M
Fc = λk ∇ i Ck , (70)
k=1
where λk is the Lagrange multiplier to be determined, and subscript i signifies
the fact that the gradient of constraint Ck must be taken with respect to i. Since
we use the Verlet algorithm for integrating the equations of motion, we have the
particles’ positions at times t − t and t which satisfy the constraints imposed on
the system. Thus, an intermediate position r̃i is first calculated for particle i,
r̃i (t + t) = 2ri (t) − ri (t − t) + t 2 Fi [ri (t)], (71)
where Fi represents all the physical forces that the particle experiences. The true
position is then computed from


M
ri (t + t) = r̃i (t + t) − λk ∇ i Ck (rN ), (72)
k=1

where, as usual, rN = {r1 , r2 , · · · , rN } represents the position coordinates of all

the particles. The Lagrange multipliers are determined by an iterative process. For
iteration number j , a loop over the constraints is carried out, in each step of which
the particles’ positions and the Lagrange multipliers are updated. The positions are
updated according to the following iterative scheme,
(n) (o) (j )
ri = ri − t 2 λk ∇ i Ck (rN ), (73)
where superscripts n and o denote the new and old values, respectively. To calculate
(j )
the Lagrange multipliers λk a first-order expansion of Ck (rN ) is carried out which
is then required to vanish. Therefore,
(j )

7 8
Ck (rN )(n) = Ck (rN )(o) − t 2 λk ∇ i Ck (rN )(o) · ∇ i Ck [rN (t)] + · · · = 0,
i
(74)
(j )
where the sum is over all the particles. The final expression for λk is given by

(j ) Ck (rN )(o)
λk =  . (75)
t 2 { i ∇ i Ck (r )
N (o) · ∇ C [rN (t)]}
i k
484 9. Atomistic Modeling of Materials

The iteration continues until the constraints are satisfied numerically to within a
fixed acceptable error. This algorithm has turned out to be highly efficient and
accurate.

9.2.8 Ion–Ion Interactions

As mentioned earlier in this chapter, Coulombic and gravitational interactions are
both described by a pair-potential which is proportional to ln r in 2D and r −1
in 3D. Since these are long-range interactions, it is not clear that one can use a
finite cutoff distance in order to accurately calculate the corresponding forces. In
case of a system with local electrical charges, even the overall neutrality of the
system does not guarantee that the screening length is finite, since in most cases of
practical interest the screening length is larger than the linear size of the system,
and therefore use of a finite cutoff distance rc is not justified. Another problem
arises when periodic boundary conditions are used in simulation of a system with
a distribution of local charges. In this case the sums over the image charges in
the periodic images of the system do not converge. This problem is solved by
subtracting an offset from the potential (since adding or subtracting a constant to
the potential does not change the resulting force, as the force is the gradient of the
potential), leading to the following expression for the total configurational energy
of the system for a set of N particles with charge qi ,



N

N
qi qj


1
E= − qi qj . (76)
|ri − rj + R| R
R i<j j i<j j R
=0

The sum over 1/R is over the locations R of the periodic replicas of the system.
To discuss this further, we consider a concrete example—transport of charged
particles in a disordered material, especially one with a quenched distribution of
charge centers. This problem is relevant to many important phenomena, such as
dynamic response of non-metallic materials, e.g., ionic glasses and polymeric and
glassy conductors, highly defected crystals, and porous materials that are used for
catalytic and separation processes. Although this problem had been studied exten-
sively, until recently no consensus regarding the nature of the transport process had
emerged. In particular, if R 2 (t) is the mean square displacement of the mobile
charged particles at time t, one expects to have
R 2 (t) ∼ t α , (77)
where α = 1 for diffusive transport. However, the precise value of α was a contro-
versial subject with some researchers claiming that α > 1, while others believing
that α ≤ 1, in which case the transport process is anomalous (see Chapter 6 of Vol-
ume I for a discussion of anomalous transport). This controversy also prevented
interpretation of experimental data for diffusion of ions in random media with a
distribution of charge centers. For example, during diffusion of ions through zeo-
lites, which are porous catalytic materials with a distribution of charge center (ions
and cations), it has been observed that, upon changing the charge on the diffusing
particles (i.e., making the disorder stronger), the diffusivity decreases by orders of
 9.2. Classical Molecular Dynamics Simulation 485

magnitude. Despite its great importance, no efficient and reliable computer simu-
lation of this problem was carried out for many years, because (1) the Coulombic
interactions between the particles are long-ranged, and (2) the charge centers give
rise to deep potential wells that may capture the mobile particles for long periods
of time and slow down their motion.
Let us now describe how MD simulations of this problem can be carried and, in
particular, how the effect of long-range Coulombic interactions can be taken into
account by an efficient and reliable algorithm developed by Mehrabi and Sahimi
(1999). They used both a continuum and a lattice representation of the system. The
continuum representation was used when the fixed charge centers were distributed
randomly in the medium. The lattice, which was simple-cubic, was utilized when a
potential-potential correlation function, defined below, was utilized for generating
the potentials due to the fixed charge centers. At time t = 0 the charged mobile
particles are distributed randomly in the system, but as they move correlations
develop between them. In addition to the Coulombic interaction, a short-range,
LJ-type repulsive interaction (i.e., ∼ r −12 ; see Section 9.2.1) was also used to
prevent capture of a mobile particle by an immobile one with a charge which has
a sign opposite to that of the mobile particle.
The charge centers are either distributed explicitly throughout the medium, or
are represented by their potential distribution, generated by the potential-potential
correlation function. To make the system neutral, equal numbers of the centers
with opposite charges are inserted in the system, and the same is done with the
mobile particles. The Coulomb potential Ui acting on the ith mobile particle is
written as,
(f m) (mm)
Ui = Ui + Ui , (78)
(f m)
where Ui is due to the interaction between the mobile particle and the fixed
(mm)
centers, while Ui is contributed by the interaction between the mobile particles
(f m)
themselves. Ui can be calculated by two different methods (yielding identical
(f m) (mm)
results). In one method, Ui [and also Ui ] is computed by the multipole
expansion method described below. In the second method, one can use the fact
that diffusion of charged particles in disordered media can be viewed as a transport
process in an external potential field generated by the quenched disorder that
represents the fixed charge centers. Thus, instead of directly distributing the charge
(f m)
centers with a given density ρ(r), Ui is formally represented by the solution
of the Poisson’s equation which, for example in 3D, is given by

q f qm ρ(r )
dr
(f m)
Ui (r) = − (79)
4π ε0 |r − r |
where qf and qm are the charges for the fixed and mobile particles, respectively,
and ε0 is the permittivity. The charge density ρ(r) is represented by its correlation
function χρρ (r) which, in the case of Debye-Hückel statistics, is given by
ρ0 κ 2 e−κ|r|
χρρ (r) = ρ0 δ(r) − , (80)
4π|r|
486 9. Atomistic Modeling of Materials

where ρ0 = ρ(r), and κ −1 is the spatial correlation or the screening length. The
depth of the potential wells in which the mobile particles are captured by the
immobile ones, and thus the radius of influence of the traps, is controlled by κ −1 .
The larger κ −1 , the deeper is the potential well, and thus the larger the time spent
in such traps. In the limit κ −1 → ∞, the trapping times become infinitely large,
and therefore the effective diffusivity is zero. The power spectrum χ̂φφ (ω) for the
potential is calculated from that of the charge density χ̂ρρ (ω), since Eq. (79) is a
convolution integral of the charge density and the Green function for the potential
generated by a single charge particle, and therefore in 3D
$ %
qf qm 2 ρ0
χ̂φφ (ω) = . (81)
ε0 ω (ω2 + κ 2 )
2

Note that the 1D version of Eq. (81) is given by, χ̂φφ = ρ0 (qf qm /ε0 )2 /(ω2 + κ 2 ).
Hence, a realization of the potential
√ √ field is generated as follows. Random numbers,
distributed uniformly in [− 3, 3) (this ensures that their power spectrum is 1,
as it should be), are assigned to the sites of!the system. The resulting array is
then Fourier transformed and multiplied by χ̂φφ (ω), and then inverse Fourier
transformed.
(f m) (mm)
Ui and Ui can also be calculated by a multipole expansion method (Ding
et al., 1992; Mehrabi and Sahimi, 1999). In this method particle i interacts with
the nearby particles through the usual Coulomb potential, and with the far away
particles through their pre-calculated
 multipole expansions of the potential. The
total potential U (g) (r) = N j U j (r) produced by a group of N charges is
$ % $ % $ %
q 1 1 1 1 .. 1
U (r) = − P · ∇
(g)
+ Q : ∇∇ − O.∇∇∇ + · · · (82)
r r 2 r 6 r
where q, P, Q, and O are, respectively, the monopole, dipole, quadrupole, and
octapole moments of the group of charges around the origin. In practice, we write
⎛ ⎞
q 1
1



U (g) (r) = + 3 Pα rα + 5 ⎝ Qαα rα rα + Qαβ rα rβ − r 2 Q⎠
r r α 2r α α β
⎡ ⎛ ⎞
1 ⎣ ⎝






+ 7 15 Oααα rα rα rα + Oααβ rα rα rβ + Oαβγ rα rβ rγ ⎠
6r α α β α β γ
 


−9r 2 Oα + ···, (83)
α

with


q= qi , (84)
i


Pα = qi Riα , (85)
i


Qαα = 2
qi Riα , (86)
i
 9.2. Classical Molecular Dynamics Simulation 487


Qαβ = qi Riα Riβ , (87)
i


Q= qi Ri2 , (88)
i


Oα = qi Ri2 Riα , (89)
i


Oααα = 3
qi Riα , (90)
i


Oααβ = 2
qi Riα Riβ , (91)
i


Oαβγ = qi Riα Riβ Riγ , (92)
i

where r = |r|, Ri is the position vector of the ith charge, α, β, and γ stand for the
coordinates x, y, and z, and qi is the charge of the ith (fixed or mobile) particle.
A highly efficient simulation technique is fundamental to this study. Hence,
in addition to taking advantage of the multipole expansion, the 3D simulation
cell is divided into 8 smaller equal boxes, called children of the original box
[see, for example, Greengard and Rokhlin (1987) and references therein]. Each
child box is a parent to 8 smaller boxes, with the division continuing up to a
certain level which is called the maximum level (maxlevel) of division. The data
needed for each particle, i.e., its position and type (mobile or fixed), are stored in
a particle object. A cell object contains a list of its current particles. Each particle
is also “connected” to the next and previous particle in the list. After setting up the
entire data structure, the multipoles of each cell around its center at the maxlevel
are calculated using the above expressions. Then, the multipoles of the parent
cells are computed by translating and adding the multipoles of their children by
a displacement vector  = (x , y , z ). In terms of the old quantities, the new
translated (primed) quantities are given by,
Pα = Pα − qα , (93)

Qαα = Qαα − 2α Pα + q2α , (94)

Qαβ = Qαβ − β Pα − α Pβ + qα β , (95)

Q = Qαα + Qββ + Qγ γ , (96)


Oααα = Oααα − 3α + 22α − q3α , (97)

Oααβ = Oααβ − β Oαα + α [−2Qαβ + 2β Pα + α (Pβ − qβ )], (98)

Oαβγ = Oαβγ + γ (β Pα − Qαβ ) + β (α Pγ − Qαγ )
(99)
+α (γ Pβ − Qβγ ) − qα β γ .
488 9. Atomistic Modeling of Materials

Each particle’s potential energy is divided into Uinear and Uifar . A particle in
a cell at the maxlevel interacts with all other particles in the same cell and in
the neighboring cells by the usual Coulomb potential, thus yielding Ujnear . It also
interacts with its parent’s neighbors’ children through the corresponding multipole
expansions. Computations continue up to the entire simulation cell, hence yielding
Uifar . In this way, the number of the cells that interact with each particle is drastically
reduced as one moves away from the particle. For example, in 3D with four levels
of division the number of the interacting cells is only 415, rather than the original
4069 cells. This method is highly efficient for taking into account the effect of
Coulomb and other long-range potentials.
However, even with such an efficient algorithm, MD simulation of this problem
requires intensive computations. If the simulations are not carried out for long
enough times, one may not be able to obtain the true asymptotic (long time)
behavior of the system. As an example, consider the problem in a 1D system,
such as a highly anisotropic material so that the motion of the mobile particles is
restricted essentially to one direction. In this case, the mobile particles can only
travel in the space between themselves, since they cannot “jump” over each other.
There are many relatively fast diffusive jumps in the mean square displacements
of the particles after certain periods of time. In between the jumps one has a slow
motion that causes the overall transport to be anomalous [i.e., α < 1 in Eq. (77)],
not only in 1D but also in 2D and 3D. The mobile particles can be trapped in the
potential wells (traps) that the quenched distribution of the charge centers creates.
The traps have a finite sphere of influence, such that for any particle i within the
sphere the potential difference Ui for a displacement that can take i out of the
sphere is very large, and thus the probability of an appreciable jump is small.
However, after some time the particles are close to the boundary of the traps and
escape with a displacement that takes them out of the traps. They then resume their
diffusive motion until they are captured by another trap, and so on.

9.3 Nonequilibrium Molecular Dynamics Simulation

As mentioned in Section 9.2, equilibrium MD (EMD) simulations are applica-
ble to systems that, at least in principle, are amenable to treatment by statistical
mechanics. Although, as our discussions so far should have made it clear, many
assumptions must be made and several approximations must be used, their main
purpose is to make the computations tractable. However, if we are to compute the
effective transport properties, such as the diffusivity, shear viscosity, and thermal
conductivity, then EMD is not an effective tool. For example, as is well-known,
by calculating the velocity correlation functions for every distinct pair of species
in the system, one can obtain information about the microscopic motion of the
molecules. However, since the velocity correlation function decays as the size
of the system increases, use of EMD is not feasible for estimating the transport
properties of a mixture of molecules in a system which is under the influence of
an external potential gradient—a situation which is encountered in a very large
 9.3. Nonequilibrium Molecular Dynamics Simulation 489

number of practical problems. One can use the velocity autocorrelation function,
but this quantity can only be used for predicting the tracer or self-diffusivity Ds
of a species (i.e., when the system is very dilute) via the Green-Kubo equation:

 , -
1 ∞

N
Ds = [vi (t) · vi (0)] dt, (100)
3N 0 i=1

where vi is the velocity of particle i. The self-diffusivity is, however, completely

different from the transport diffusivity because tracer diffusion ignores the effect
of the collective motion of other molecules, especially in systems with a moderate
or high density.As such, EMD is not suitable for investigation of a transport process
in a system on which an external potential (pressure, voltage, chemical potential,
concentration, etc.) gradient has been imposed.
Non-equilibrium MD (NEMD) simulation represents a practical alternative to
EMD for those systems for which the velocity correlation function is difficult, or
meaningless, to measure. It is particularly ideal for the practical situation in which
an external driving force is applied to the system. Several such modeling efforts
have been reported. Notable among them are the Grand-Canonical Molecular Dy-
namics (GCMD) method [see, for example, Cagin and Petit (1991); Lupkowski
and van Swol (1995)] in which Monte Carlo (MC) and MD simulations are com-
bined (see below), and a dual control-volume GCMD technique (DCV-GCMD)
[see, for example, Heffelfinger and van Swol (1994); MacElroy (1994); Cracknell
et al. (1995); Xu et al. (1998)]. We describe here the DCV-GCMD method which
has become an effective tool for studying systems that are under the influence of
an external potential gradient.
To describe this method, we consider a concrete example, namely, transport of
binary gas mixtures in a carbon nanopore, a problem of considerable importance
in separation and purification processes (Xu et al., 1998, 1999). The same method
is applicable to transport of a one-component fluid in the pore which is under the
influence of an external potential gradient. Other types of carbon pores, such as
carbon nanotubes, are also the subject of intense research activity as their mechan-
ical strength appears to be much larger than what is expected of such materials at
this length scale. Transport in such nanotubes can also be studied by the method we
describe here. For simplicity, we consider a slit pore, a schematic representation of
which is shown in Figure 9.3 in which the origin of the coordinates is at the center.
The external driving force is either a chemical potential (or, equivalently, pressure)
or a concentration gradient applied in the x-direction. The system is divided into
three regions. The h- and -regions at the two ends represent two control volumes
(CVs) exposed to the bulk fluid at high and low chemical potential, pressure, or
concentration, respectively, while the middle region represents the pore in which
transport occurs. The pore’s length is nL with n being an integer. Periodic bound-
ary conditions are employed in the y-direction. The two carbon walls are located at
the top and bottom of the xy planes. We consider transport of a binary gas mixture
in such a slit pore.
490 9. Atomistic Modeling of Materials

Figure 9.3. Schematics of a slit pore, where h and

l denote the high and low pressure (or chemical
potential) control volumes (after Xu et al., 1998).

The DCV-GCMD method combines integration of the equations of motion with

GCMC insertions and deletions of the gas molecules in the two CVs. It is essential
to maintain the densities of each gas component in the two CVs at some fixed
values, which are in equilibrium with two bulk phases, each at a fixed gas pressure
and concentration. The densities, or the corresponding chemical potentials of each
component in the CVs, are maintained by carrying out a sufficient number of
GCMC insertions and deletions of the particles. The probability of inserting a
particle of component i is given by
 $ % 
+ −E/kB T
pi = min Zi Vci exp , 1 , (101)
Ni + 1
where Zi = exp(µi /kB T )/3 is the absolute activity at temperature T , i the
de Broglie wavelength, kB the Boltzmann’s constant, µi the chemical potential of
component i, E the potential energy change resulting from inserting or removing
a particle, and Vci and Ni the volume and number of atoms of component i in each
CV, respectively. The probability of deleting a particle is given by
 $ % 
−E/kB T
pi− = min Ni exp , 1 . (102)
Zi Vci
When a particle is inserted in a CV, it is assigned a thermal velocity selected
from the Maxwell–Boltzmann distribution, Eq. (49), at the given T . An important
parameter of NEMD simulations is the ratio R of the number of GCMC insertions
and deletions in each CV to the number of MD integration steps between successive
GCMC steps. This ratio must be selected appropriately in order to maintain the
correct density and chemical potentials in the CVs, and also reasonable transport
rates at the boundaries between the CVs and the transport region. Its typical value
varies anywhere from 50:1 to 400:1. During the MD calculations particles crossing
the outer boundaries of the CVs must be removed. In addition, one should allow
for a non-zero streaming velocity (the ratio of the flux to the concentration of
each component) in the entire transport region of each component, consistent
with the presence of bulk pressure/chemical potential gradients along the flow
direction. Otherwise, a zero streaming velocity in the transport region will lead
to severely underestimated fluxes. Since the two CVs are assumed to be well
mixed and in equilibrium with the two bulk phases that are in direct contact with
 9.3. Nonequilibrium Molecular Dynamics Simulation 491

Figure 9.4. Density profiles for two

gases components in the slit pore.
The pore is defined for −10 ≤ X∗ ≤
10, where X∗ = H /σ , with H be-
ing the height of the pore and σ a
Lennard–Jones molecular diameter
(after Xu et al., 1998).

them, there should be no overall non-zero streaming velocity in these regions.

However, to reduce the numerical instability caused by the discontinuity of the
streaming velocities at the boundaries between the CVs and the transport region,
a small streaming velocity can be added to the thermal velocity of all the newly
inserted molecules within each CV that are located within a distance 0.5σ1 from
the boundaries, where σ1 is the LJ size parameter (or the effective molecular size)
of the lightest of the two gases. The streaming velocity of each component in the
transport region can be obtained by linearly interpolating between its two values in
the two CVs. After a few thousands of integration steps, this procedure generates a
potential gradient across the pore, an example of which is shown in Figure 9.4. To
study the transport of a mixture in a potential gradient, the temperature of the system
must be held constant in order to eliminate any contribution of the temperature
gradient to the transport; hence special care must be taken to achieve this.
All the quantities of interest are calculated from such simulations. For example,
one can calculate the density profiles of component i along the x- and z-directions,
ρiz (x) and ρix (z), respectively. To compute ρix (x) the simulation cell is divided in
the x-direction into grids of size σ1 , and for each MD integration step the density
profiles ρiz (x) are obtained by averaging the number of particles of component i
over the distance σ1 , with a similar procedure for ρix (z). These quantities are im-
portant to understanding adsorption and transport properties of the gases between
the two pore walls. For each component i one can also calculate its flux Ji by
measuring the net number of its particles crossing a given yz plane of area Syz :
NiLR − NiRL
Ji = , (103)
ns tSyz
where NiLR and NiRL are the number of the gas molecules of type i moving from
the left to the right and vice versa, respectively, t is the MD time step, and ns
is the number of the MD steps over which the average is taken. The system is
considered to have reached a steady state when the fluxes calculated at various yz
planes are, to within an acceptable error, equal to the averaged values, after which
492 9. Atomistic Modeling of Materials

the fluxes are calculated at the center of the transport region. The permeability ki
of species i is then calculated from
Ji nLJi
ki = = , (104)
Pi /nL Pi
where Pi is the partial pressure for species i along the pore. The transport diffu-
sivity De is then obtained from the Fick’s law, Ji = −De ∂ρi /∂x, where ∂ρi /∂x is
the adsorption density gradient of component i along the x-direction. Another im-
portant property for the problem that we are describing is the dynamic separation
factor S21 defined as
k2
S21 = . (105)
k1
The NEMD method that we described here has proven to be a practical tool for
simulating transport properties of fluid mixtures in not only a carbon nanopore, but
also in nanoporous materials, such as a variety of membranes, with an intercon-
nected network of nanopores (Xu et al., 2000a). Its predictions for some properties
of interest are in good quantitative agreement with experimental data (Xu et al.,
2000b). Other NEMD methods have been discussed by Rapaport (1995).

9.4 Quantum Molecular Dynamics Simulation:

The Car–Parrinello Method
In our discussion of MD simulation presented so far, we have treated the molecular
system as a collection of classical particles for which the interaction potential is
known, and have neglected all the quantum-mechanical effects. We now discuss
quantum MD simulations in which this restriction is removed and quantum me-
chanical effects for the electronic degrees of freedom are taken into account. The
landmark paper of Car and Parrinello (1985) in which a MD technique was used
for minimizing the total energy functional and allowed one to use, with a very
high degree of efficiency, local and non-local pseudo-potentials, opened the way
for QMD computations. We now describe and discuss their method.
The key idea of Car and Parrinello was to treat the electronic wave functions
i as dynamical variables, and to define a Lagrangian for the electronic system
which is given by


L= m ˙ i |
˙ i  − E[{i }, {RI }, {n }]. (106)
i

Here mψ is a fictitious mass associated with the electronic wave functions, giving
rise to the kinetic energy term of the Lagrangian that arises due to the fictitious
dynamics of the electronic degrees of freedom, E is the Kohn–Sham energy func-
tional, described in Section 9.1, which plays the role of the potential energy, RI
is the position of the ion I , n defines the size and shape of the periodic unit cell,
and  ˙ i = di /dt. As Eq. (106) suggests, the details of the kinetic energy do not
 9.4. Quantum Molecular Dynamics Simulation: The Car–Parrinello Method 493

matter. What is more important is that the mass m should be much smaller than
the nuclear masses which would prevent transfer of energy from the classical to
the quantum degrees of freedom over long periods of the numerical simulations.
The electronic wave functions are subjected to the orthonormality constraints:

i∗ (r)j (r) d 3 r = δij , (107)

where δij is the Kronecker delta. To incorporate these constraints, Lagrange

multipliers are introduced, so that the Lagrangian of the system is rewritten as


L= m ˙ i |
˙ i  − E[{i }, {RI }, {n }]

i    (108)
+ λij i∗ (r)j (r) d r − δij .
3

i j

Mathematically, the Lagrange multipliers λjj ensure that the wave functions re-
main normalized, while λij (with i
= j ) guarantee that the wave functions remain
orthogonal. Physically, the Lagrange multipliers can be thought of as providing ad-
ditional forces acting on the wave functions, ensuring that, at any given time t, they
remain orthonormal as they propagate along their MD paths. The Lagrange mul-
tipliers are symmetric, λij = λj i , and represent 12 N (N + 1) independent values
which are determined by the 12 N (N + 1) orthonormality conditions. The iterative
algorithm SHAKE, described in Section 9.2.7, can be used for determining the
Lagrange multipliers, and was in fact utilized by Car and Parrinello.

9.4.1 The Equations of Motion

Having defined the Lagrangian of the system, the equations of motion for the
electronic states are derived from
 
d ∂L ∂L
= ∗, (109)
˙∗
dt ∂  ∂ i
i
which yield the following equations of motion
d 2 i

m 2
= −Hi + λij j , (110)
dt
j

where H is the Kohn–Sham Hamiltonian defined in Section 9.1. The similarity

between Eq. (110) and Eq. (32), the equations of motion for a classical many-
particle system solved by the MD method, is now apparent. To ensure that at any
given time during the MD integration the wave functions remain orthonormal, the
Lagrange multipliers must vary continuously with the time, and therefore they must
be estimated continuously at infinitely small time separations. However, doing so
would make the computations intractable. To make the computations tractable, it is
usually assumed that the Lagrange multipliers remain constant in each time step t
during which the equations of motion are integrated. Although this approximation
494 9. Atomistic Modeling of Materials

makes the computations tractable, it also creates a new problem: At the end of each
time step the wave functions will not be exactly orthonormal, so that a separate
orthonormalization must also be carried out at the end of each time step (see
below).
However, since a separate orthonormalization step must be carried out at the end
of each time step, one may remove the orthogonality constraints from the equations
of motion and use partially constrained equations of motion. After these equations
have been integrated, the orthogonality constraints are imposed again, and the
Lagrange multipliers λij for the normalization constraints are approximated by
the expectation values of the energies of the states given by
λi = i |H|i . (111)
With this approximation the equations of motion for the wave functions are given
by
d 2 i
m = −(H − λi )i , (112)
dt 2
which ensures that the acceleration of an electronic state is always orthogonal to
that state, and that the acceleration becomes zero when the wave function is an
exact eigenstate. However, more generally, one can start from the estimate
˙ i |
λij = 2i |H|j  − m  ˙ j , (113)
and proceed with integration of Eq. (110). The difference between (111) and (113),
aside from the approximate nature of (113), is that the latter depends on m which
itself must be somehow selected carefully.

9.4.2 The Verlet Algorithm

Equations (111) can now be integrated by a Verlet algorithm. In analogy with
Eq. (42), one writes
d 2 i
i (t + t) = 2i (t) − i (t − t) + t 2 . (114)
dt 2
If we utilize Eq. (112), the Verlet algorithm becomes
t 2
i (t + t) = 2i (t) − i (t − t) − (H − λi )i (t). (115)
m
A similar Velert algorithm can be written down for Eq. (110) if the initial estimates
(113) are to be employed. Note that, if an expansion such as (29) is used as the
solution for i (which is always the case when a pseudo-potential approximation
is made), then the coefficients ci,k+G must be time-dependent. If we substitute
such an expansion into Eq. (110) or (112), we obtain the governing equations for
ci,k+G which can be integrated by the Verlet algorithm (see also Section 9.4.5).
Generally speaking, the performance of the Verlet algorithm for this type of
computations is evaluated in terms of the rate at which it converges to the minimum-
energy state. One must carry out the integration for a certain amount of time for the
 9.4. Quantum Molecular Dynamics Simulation: The Car–Parrinello Method 495

problem to converge; the amount of computational time is controlled, of course,

by the size of the time step t used in the integration. It can be shown that for the
QMD simulations that we are describing here, the largest possible value of t is
approximately given by
&
m
t  2 , (116)
M −  m
where M and m are, respectively, the largest and smallest eigenvalues of the
problem. Use of any t which is significantly larger than the estimate (116) will
lead to instability and large errors in the numerical solution of the equations of mo-
tion. Recall that in pseudo-potential approximation the electronic wave functions
are represented by a plane-wave basis set, in which case the largest eigenvalue is
determined by the cutoff kinetic energy that is used for truncating the basis set.
This implies that if the cutoff energy is increased, the time step used in the Verlet
algorithm must decrease. Moreover, the maximum allowed value of t decreases
as the size of the system increases, limiting the maximum system size that can be
simulated by this algorithm.
We also note that, in our discussion of the integration algorithm, we have tacitly
assumed that the Kohn–Sham Hamiltonian H is constant during the time evolution
of the system. However, in addition to the instabilities that are caused by utilizing
a too large value of the time step t, another type of instability arises if H is
not allowed to vary when it must. Since the wave functions evolve under the MD
scheme, the contribution of the exchange-correlation potential to the Kohn–Sham
Hamiltonian [see Eqs. (12) and (15)] also varies with the time, leading to a new
Hamiltonian. Thus, at the end of each time step, the Kohn–Sham Hamiltonian
is updated with the new electronic density, so that the final time step leads to a
self-consistent solution of the Kohn–Sham Hamiltonian and the determination of
the minimum in the total energy.

9.4.3 The Kohn–Sham Eigenstates and Orthogonalization of the

Wave Functions
As discussed above, the wave functions that are obtained from integrating the
partially-constrained equations of motion are not orthogonal, and therefore an
orthogonalization procedure must be used. The correct application of the con-
straints of orthogonality and normalization is actually essential for the success of
the Car–Parrinello method. Car and Parrinello (1985) used an iterative method to
orthogonalize the wave functions by utilizing the following equation
(n) (o) 1
(o) (o) (o)
 i = i − j |i j , (117)
2
j
=i

where the superscripts n and o refer to the new and old estimates of the wave
functions, respectively. If Eq. (117) is applied repeatedly to the old estimates, then
the electronic wave functions can be made orthogonal to any desired accuracy. For
496 9. Atomistic Modeling of Materials

example, if algorithm (117) is applied to two wave functions, they will be exactly
orthogonal after only one application of the algorithm. Moreover, if the estimates
of the wave functions are orthonormal up to order t 4 (the accuracy of the Verlet
algorithm), then over a given time step algorithm (117) changes them to an extent
within the same order. In general, the number of times that algorithm (117) must
be applied depends on the number of the wave functions to be computed and their
initial departure from orthogonality. However, algorithm (117) does not preserve
normalization of the wave functions, and therefore they must be normalized after
each application of the algorithm.
The Kohn–Sham energy functional is minimized by any set of wave func-
tions that are a linear combination of its lowest-energy eigenstates. Under the
MD scheme, the wave functions that are obtained after orthogonalization will be
stationary, implying that in the MD method each wave function will converge to a
linear combination of the lowest-energy Kohn–Sham eigenstates. This can create
severe problems for treating metals, since the ability to converge to Kohn–Sham
eigenstates (and not to their linear combinations) is highly important for metal-
lic materials. Several methods have been proposed for addressing this problem
(see, for example, Pederson and Jackson, 1991, and references therein). The ac-
tual Kohn–Sham eigenvalues can be found by either diagonalization of the matrix
A = (Aij ), the entries of which are given by
Aij = i |H|j , (118)
or by the Gram–Schmidt algorithm (see, for example, Noble and Daniel, 1977)
which is similar to (117) but without the factor 1/2 in front of the sum. Thus, instead
of using algorithm (117) which generates wave functions that are linear combina-
tions of the Kohn–Sham eigenstates, one can use the Gram–Schmidt method which
results in orthogonal wave functions in such a way that all of the higher-energy
wave functions are forced to be orthogonal to the lowest-energy wave functions.
This in turn forces each state to converge to its lowest possible energy while sat-
isfying the constraint that it be orthogonal to all states below it. Therefore, the
set of lowest possible single-particle levels under these constraints comprises the
Kohn–Sham eigenstates.

9.4.4 Dynamics of the Ions and the Unit Cell

In our discussion of QMD simulations so far we have assumed that the ionic
positions and the shape of the unit cell remain fixed. However, in practice these
represent additional degrees of freedom that have their own dynamics, and there-
fore must be considered. Fortunately, since we are interested in the final state of
the system, which consists of ions and electrons in their minimum energy con-
figurations, the exact path for reaching this state is not very important. This fact
provides us with considerable flexibility.
The QMD technique of Car and Parrinello allows us to include in the compu-
tations the positions of the ions and the coordinates that define the size and shape
of the unit cell. These constitute dynamical variables and including them in the
 9.4. Quantum Molecular Dynamics Simulation: The Car–Parrinello Method 497

computations requires writing down a new Lagrangian, commonly referred to as

the Car–Parrinello Lagrangian, which is given by


1 $ % $ %
dRI 2
1 dn 2
L= m  ˙ i +
˙ i | mI + mc − E[{i }, {RI }, {n }],
2 dt n
2 dt
i I
(119)
which should be compared with Eq. (106) that was for a system in which the
positions of the ions and the shape and size of the unit cell were fixed. Here, mc ,
similar to m , is a fictitious mass associated with the dynamics of the coordinates
that define the unit cell, namely, {n }, and mI is the mass of the ions. The rest of
the notations is the same as those for Eq. (106). Given Lagrangian (119), we can
derive two new sets of equations of motion, one for the ions given by
d 2 RI ∂E
mI =− , (120)
dt 2 ∂RI
and a second one for the coordinates of the unit cell,
d 2 n ∂E
mc =− . (121)
dt 2 ∂n
These equations relate the rate of acceleration of the length of the lattice vectors
to the diagonal components of the stress tensor, and also relate the accelerations
of the angles between the lattice vectors to the off-diagonal components of the
stress tensor, both integrated over the unit cell. Although these equations can be
integrated at the same time that the equations of motion for the electronic states
are integrated, the matter is not as straightforward as it may seem. The reason for
this is that physical ground-state forces acting on the ions, and integrated stresses
acting on the unit cell, cannot be calculated for arbitrary electronic configurations.
The following discussion establishes this fact.

9.4.4.1 The Hellmann–Feynman Theorem

To understand why the ground-state forces acting on the ions and the inte-
grated stresses acting on the unit cell cannot be calculated for arbitrary electronic
configurations, recall that the force FI acting on an ion is given by
dE
FI = − . (122)
dRI
As ions change their positions, the wave functions must change to self-consistent
Kohn–Sham eigenstates corresponding to the new positions of the ions. These
changes contribute to FI since
∂E
∂E di
∂E d ∗
FI = − − − i
, (123)
∂RI ∂i dRI ∂i∗ dRI
i i
which follows from Eqs. (119) and (122), and should be compared with Eq. (120)
which states that, FI = −∂E/∂RI . This apparent contradiction is due to the fact
that in the Lagrange equations of motion for the ions, Eq. (120), the force acting
498 9. Atomistic Modeling of Materials

on the ions is not a physical force, rather it is a force that the ions experience
from a particular electronic configuration. However, it is not difficult to show that,
when the electronic wave functions are the eigenstates of the Hamiltonian, then
the second and third terms of Eq. (123) vanish, and therefore, in this case, ∂E/∂RI
yields the true physical force on the ions. This important result is known as the
Hellmann–Feynman theorem (Hellmann, 1937; Feynman, 1939), and in fact holds
for any derivative of the total energy.
The Hellmann–Feynman theorem greatly simplifies the task of computing the
physical forces acting on the ions and the integrated stresses that are exerted on
the unit cell, since it allows one to compute these quantities only when the wave
functions are very close to exact eigenstates. Once the forces and stresses have been
computed, the positions and the shape and size of the unit cell are also calculated
using Eqs. (120) and (121), which represent their equations of motion. Each time
the positions of the ions or the shape and size of the unit cell change, the electrons
must be close to the ground state of the new ionic configuration in order to compute
forces and stresses for the new configuration.
The simplest way that the Hellmann–Feynman forces are used is for determining
the position of a local energy minimum of the ionic system. In this scheme the ions
are moved along the directions of the Hellmann–Feynman forces until the residual
forces (i.e., the deviations from ∂E/∂RI ) on all the atoms are smaller than a given
value. These forces cause the ions to fluctuate around their equilibrium positions.
The residual forces acting on the ions are never zero. If the system is to approach
the minimum energy state, the magnitudes of the errors in the Hellmann–Feynman
forces must be reduced, implying that the electronic configuration must be relaxed
closer and closer to the instantaneous ground states as the ionic configuration
approaches the local energy minimum.

9.4.4.2 Pulay Forces and Stresses

The reader may have noticed the absence of a fourth term in Eq. (123), ∂φ/∂RI ,
the derivative of the basis set {φ} (for representing the wave functions i ) with
respect to the ionic positions. This term is called the Pulay force (Pulay, 1969). If
a plane-wave basis set, Eq. (29), is used for representing the wave functions, then
the derivative of each basis set with respect to RI vanishes and the Pulay forces
are exactly zero, in which case the Hellmann–Feynman forces are exactly equal to
∂E/∂RI , provided that the electronic wave functions are Kohn–Sham eigenstates.
If the Pulay forces do not vanish and are not computed either, then the calculated
Hellmann–Feynman forces will be in error, with the error being independent of
how close is the electronic configuration to its ground state, which means that a
local energy minimum of the ionic system cannot be computed by calculating only
the Hellmann–Feynman forces acting on the ions. This causes severe problems for
the computations and may render them inefficient.
Although when a plane-wave basis set is used, the Pulay forces acting on the
ions are exactly zero, the corresponding Pulay stresses acting on the unit cell
may not be zero. Changing the size of the unit cell while keeping the number of
 9.4. Quantum Molecular Dynamics Simulation: The Car–Parrinello Method 499

plane-wave basis states constant results in a change in the cutoff energy for the
basis set. If we increase the number of plane-wave basis states by increasing the
cutoff energy for the basis set, it will eventually result in a reduction in the total
energy of the system. However, if the cutoff energy is large enough to achieve
absolute convergence, the change in the total energy will be zero. In practice, most
pseudo-potential computations are carried out with a cutoff energy at which energy
differences have converged, but at which the total energies have not converged,
in which case the diagonal components of the Pulay stresses acting on the unit
cell will have non-zero values. It can be shown that the change in the total energy
per atom is independent of the details of the ionic configuration, provided that the
cutoff energy for the basis set is large enough for the energy differences to have
converged. This facilitates computation of the Pulay stresses, since they can be
calculated once and for all from the change in the total energy of a small unit cell
as the cutoff energy is varied.

9.4.4.3 The Structure Factor and Total Ionic Potential

The total ionic potential in a solid material is computed by placing an ionic pseudo-
potential at the position of every ion in the material. To do this, one calculates first
the structure factor which, at wave vector G for ions of species α, is given by


Sα (G) = exp(iG · RI ), (124)
I
where the sum is over the positions of all the ions of species α in a single unit cell.
The total ionic potential UI is computed from


UI (G) = Sα (G)uα (G), (125)
α
where uα is the local pseudo-potential [see Eq. (30)], and the sum is over all the
species in the unit cell.
At large distances r the pseudo-potential is purely Coulombic and is of the form
Z/r, where Z is the atom’s valence. Thus, the Fourier transform of the pseudo-
potential at large distances is ∼ Z/G2 , which diverges at G = 0 (i.e., at r = ∞),
and therefore the ion-electron energy is infinite. However, this poses no particular
difficulty as there are similar divergences in the Coulomb energies due to the
electron-electron and ion–ion interactions, which cancel the G = 0 divergence.
On the other hand, the pseudo-potential is not, in general, purely Coulombic, but
contains a constant contribution at small G which for species α is given by
  
uα,core = 4π r 2 Z/r − u0α (r) dr, (126)

where u0α is the local pseudopotential for the l = 0 angular momentum state
[see Eq. (30)]. Integral (126) is zero outside the core region since, by defini-
tion, the potentials are identical outside this region. This non-Coulomb part of the
pseudo-potential
 does contribute to the total energy at G = 0, which is given by
N −1
c α N α u α,core , where N is the total number of electrons in the system, Nα
500 9. Atomistic Modeling of Materials

is the total number of ions of species α, and c is the volume of the unit cell. We
remind the reader that the Coulomb energy of the ionic system can be calculated
by the method described in Section 9.2.8.

9.4.5 Computational Procedure for Quantum Molecular

Dynamics
Having discussed various aspects of the QMD method of Car and Parrinello, we
can now summarize the method.
(1) One starts with an initial set of trial wave functions from which the Hartree
potential and the exchange-correlation energy are computed.
The choice of a reasonable initial set of trial wave functions is crucial to the
efficiency and success of the computations, because the trial functions must
be such that the electronic configurations converge to the ground state. If one
is not careful, the convergence may not be achieved for two reasons.
(i) If the initial states do not span the ground state, then the final solutions
relax to a self-consistent Kohn–Sham eigenstates, but not to the ground
state. The most obvious initial guess for the wave functions is a set of
plane waves with the lowest kinetic energy. However, one must be careful
with using this set because such initial states may not span the ground
states of the electronic configurations. For example, if this set is used for
computation of the electronic structure of an eight-atom cubic cell of any
of the tetrahedral semiconductors, the electronic configuration does not
converge to the ground state. Germanium, silicon, and carbon each have
four valence electrons. An eight-atom cell of these materials contains 32
electrons, and therefore 16 doubly occupied electronic states are required
to accommodate the electrons.
(ii) The convergence may not be achieved if the QMD conserves any
symmetry that is shared by the Hamiltonian and the initial electronic
configuration. This symmetry can be broken when the electronic wave
functions are orthogonalized. For example, algorithm (111) for orthog-
onalization does not break this symmetry, whereas the Gram–Schmidt
algorithm does. However, the initial electronic states may not converge
to the ground state without breaking this symmetry, in which case the
Gram–Schmidt algorithm must be used. If one utilizes random initial val-
ues for the coefficients ci,k+G of the plane-wave basis electronic states
[see Eq. (29)], then the initial states span the ground state, and no sym-
metry is conserved. Since only those plane-wave basis states that have
the lowest kinetic energies contribute significantly to i , it is sensible to
assign non-zero random values to only those coefficients that have this
property.
(2) The Hamiltonian matrices for each of the points included in the computations
are constructed.
 9.4. Quantum Molecular Dynamics Simulation: The Car–Parrinello Method 501

(3) The equations of motion for the electronic states are integrated using the Ver-
let algorithm, and the resulting wave functions are orthogonalized, either by
algorithm (117) or by the Gram–Schmidt method; the wave functions are also
normalized.
There are several “tricks” that can speed up the computations. For exam-
ple, if we assume a local pseudo-potential approximation and use Eq. (29) in
Eq. (16), then it is straightforward to show that the governing equations for
the coefficients ci,k+G are given by

d 2 ci,k+G
= −ωi,k+G
2
ci,k+G − Bi,k+G , (127)
dt 2
where ωi,k+G and Bi,k+G are two sets of coefficients that arise when Eq. (29)
is inserted into Eq. (16). However, we recognize Eq. (127) as the oscillator
equation, which means that if the Verlet algorithm is to be used for integrating
this equation, the time step must be such that t ωi,k+G < 1 for all of the
plane-wave basis states. This implies that t is restricted by the plane-wave
basis states that have the largest kinetic energies. However, as we discussed
in Section 9.1.4, the largest kinetic-energy basis states contribute the least to
the solution of i , Eq. (29), and this creates an unsatisfactory situation. As
discussed by Payne et al. (1992), a way around this problem is assuming that
the coefficients Bi,k+G are constant over a time step of duration t, in which
case Eq. (127) can be easily integrated analytically over the time step, with
the solution given by

ci,k+G (t + t) = 2 cos[ωi,k+G (t + t)]ci,kG (t) − ci,k+G (t − t)

2
− 2
{1 − cos[ωi,k+G (t + t)]},
ωi,k+G
(128)
which can be used in the computations. Payne et al. (1992) provided a detailed
discussion of many other ways of speeding up integration of the equations of
motion.
In most cases, to increase the efficiency of the computations, damping is
applied to the equations of motion. For example, one can apply a damping of
the form −η ˙ i to the equations of motion for the wave functions, or use a
damping term −ηċi,k+G in (the right-hand side of) the equations of motion
for the coefficients, where η is the damping factor. The damping helps the
coefficients ci,k+G evolve to the values that minimize the Kohn–Sham total
energy functional.
(4) The electronic density generated by the new set of the wave functions is
then calculated [see Eq. (5)], and the corresponding new Hamiltonian is
constructed.
(5) A new set of the wave functions is obtained by integrating the equations of
motion and orthonormalization of the results. The iterations are repeated until
the resultant wave functions are stationary.
502 9. Atomistic Modeling of Materials

(6) The Kohn–Sham energy functional is minimized; its value gives the total
energy of the system. If the plane-wave basis set, Eq. (29), has been used (which
is always the case if pseudo-potential approximation is used), convergence
tests must be performed to ensure that the calculated total energy has converged
both as a function of the number of terms included in the set and the value
of the cutoff energy that has been used for truncating Eq. (29) after a finite
number of terms.
(7) Integration of the equations of motion for the ionic positions and the coor-
dinates of the unit cell is also done along the QMD computations for the
electronic configurations. It is sensible in the QMD methods to treat the
electronic and ionic systems independently when relaxing the ions to their
equilibrium positions, and therefore it is also possible to use different time
steps for the two systems. As the integration proceeds, the time step for the
ionic system must be progressively reduced as the ionic configuration ap-
proaches the local energy minimum. Such a procedure allows the electronic
configuration to relax closer to its instantaneous ground-state configuration as
the ions approach their equilibrium positions, hence ensuring that the errors
in the Hellmann–Feynman forces are always less than the actual forces acting
on the ions.
The QMD technique allows one to search large regions of configuration space
and locate the deeper energy minima in a very efficient manner. Since the QMD
combines a MD method with the DFT, it makes it possible to study temperature-
dependent effects by a method that is free of the common assumptions about the
nature of the interatomic forces. Figure 9.5 presents the pair correlation function

Figure 9.5. Pair correlation function

for amorphous (top) and liquid (bottom)
Si. Solid and dashed curve represent,
respectively, the theoretical predictions
and the experimental data (after Car and
Parrinello, 1985).
 9.4. Quantum Molecular Dynamics Simulation: The Car–Parrinello Method 503

Figure 9.6. Deformation of Se5 linear chain (from top to bottom)

to the calculated ground-state, stable structure. The time between
successive figures corresponds to 500 time steps of 3.4 × 10−16
sec (adapted from Jones and Gunnarsson, 1989).

C(r) for both amorphous and liquid Si, calculated by Car and Parrinello, and its
comparison with the experimental data. Given that the only piece of information
that was supplied to the simulator was the volume of the unit cell, the agreement
between the predictions and the data is truly remarkable.
Figure 9.6, adapted from Jones and Gunnarsson (1989), shows the evolution
of the structure of Se5 molecule, starting with an almost linear geometry, and
obtained by the Car–Parrinello method. The effect of the core region was repre-
sented by a pseudo-potential. The time between successive structures is almost
500 time steps of 3.4 × 10−16 seconds each. The last structure shown in the figure
is stable and agrees with the data. The important point to remember is that, the-
oretically, there are many possible structures that correspond to the local energy
minima, and the QMD method of Car and Parrinello can find the true structure
very efficiently.
Let us mention three different and successful applications of the QMD and
pseudo-potential methods that we just described. These applications represent only
a sample of a very large number of computations that have been reported so far.
Marks et al. (1996) reported the results of ab initio simulations of tetrahedral
amorphous carbon based on the Car–Parrinello method. The simulated structure
was in good agreement with the experimental data. Pickard et al. (2000) reported
the results of computations for a variety of lanthanide- and actinide-containing
compounds. The simulated structures and the associated structural parameters were
in excellent agreement with the experimental data. More significantly, they showed
that the pseudo-potential formulation allows a steady march through the Periodic
Table, in the sense of calculating reliably the structural parameters of compounds
that contain f electrons.
Yoon et al. (2001) employed ab initio pseudo-potential DFT with a linear com-
bination of atomic orbits and an exchange-correlation functional in the generalized
gradient approximation, described in Section 9.1.2, to study structural deformation
and intertube conductance of crossed carbon nanotube junctions. They reported
good agreement between the results of their simulations and the experimental data.
504 9. Atomistic Modeling of Materials

9.4.6 Linear System-Size Scaling

The QMD method becomes increasingly less efficient as the size of the system
increases, which is, of course, the problem with all molecular modeling methods.
If the size of the system becomes too large, then the choice of an appropriate time
step becomes very crucial as the charge density starts to fluctuate strongly. The
fluctuations are the result of instabilities in the Kohn–Sham energy Hamiltonian,
and reducing or eliminating them may require time steps that are too small, hence
making the computations intractable. Thus, for very large systems one must utilize
a highly efficient method.
The standard methods for solution of the electronic structure problems scale as
O(N 3 ) for large systems, where N is the number of atoms in a cluster or supercell.
This scaling holds for both iterative as well as standard eigensolution methods
since, as described above, one must keep the occupied wave functions orthonormal.
Thus, a considerable amount of effort has been devoted to development of methods
that can, in principle, provide O(N ) scaling for the computations. We mention
two of such methods that seem to achieve linear scaling. In the variational method
proposed by Li et al. (1993), one takes advantage of the locality of the density
matrix in real space to achieve O(N ) scaling. There is only one approximation
which is controlled by the real-space radius Rc that is used to truncate off-diagonal
elements of the density matrix. The method is, of course, exact when Rc → ∞. The
solution of the variational problem is obtained by an unconstrained minimization,
and can be incorporated into the Car–Parrinello method.
In the method of Ordejón et al. (1995) (see also Mauri et al., 1993) an energy
functional is used that possesses a global minimum for which, (1) the electronic
wave functions are orthonormal, and (2) the correct electronic ground-state energy
is obtained. Linear scaling is then achieved by introducing a spatially-truncated
Wannier-like representation for the electronic states.

9.4.7 Extensions of the Car–Parrinello Quantum Molecular

Dynamics Method
The QMD simulation method that was described above has become an almost
every-day tool of studying materials’properties. However, despite all of its success,
the Car–Parrinello method had certain limitations. For example, the QMD of Car
and Parrinello has the same shortcoming of the classical MD methods, namely, the
atomic nuclei are propagated according to the laws of classical mechanics on a
single potential energy surface. In addition, since nuclear forces are obtained from
the standard DFT, only the electronic ground state can be accessed.
Over the past few years, several extensions of the Car–Parrinello QMD have
been proposed that have made the method an even more powerful technique. For
example, Marx et al. (1999) incorporated the nuclear quantum effects in the method
through the use of path integrals. In addition, Doltsinis and Marx (2002) developed
a QMD that allows efficient simulations of electronically non-adiabatic processes,
by coupling the restricted Kohn–Sham excited state to the ground state using a
 9.5. Direct Minimization of Total Energy 505

surface hopping scheme. Evaluation of the non-adiabatic coupling is achieved by

an efficient method that exploits the available wave function derivatives.

9.4.8 Tight-Binding Methods

An efficient method for direct minimization of the Kohn–Sham total energy func-
tional is tight-binding MD (TBMD), a promising but, to our knowledge, not heavily
used, method for electronic structure computations. In this method, the system is
described by the following Hamiltonian

P2

occupied
n
H= + n |HTB |n  + Er , (129)
n
2Mn n
where the first term is similar to the corresponding term of Eq. (1), HTB is the
tight-binding Hamiltonian, and Er is a short-range repulsive energy. The repulsive
energy is written as, Er = i f [ j U (rij )], where U (rij ) is a pairwise repulsive
interaction and f (x) is a functional which must be specified for the computations
to proceed. In this sense, the Car–Parrinello method is superior because it requires
specification of no such terms.
In TBMD simulations, the covalent bonding of a material is incorporated into
the computations from its underlying electronic structure, rather than through an
N-body potential. As discussed above, the Car–Parrinello approach relies on the
expansion of the electronic wave functions by plane waves. In a TBMD simulation,
electronic calculations require a few atomic orbitals for each atom, hence allowing
a larger number of atoms to be used and longer simulation times to be utilized.
The second term of Eq. (129) represents the electronic-band structure energy ETB ,
which is calculated from a parameterized TB Hamiltonian. The difficulty of TBMD
simulations is that if one is to compute ETB from a direct diagonalization of the
tight-binding Hamiltonian HTB , the computations scale as (size of the system)3 ,
and hence the simulations may be limited to small systems. However, there are
approximate methods of calculating ETB without direct diagonalization of HTB ,
e.g., based on a variational approach (see, for example, Qiu et al., 1994). This
method seems to provide accurate predictions for some carbon-based materials.
Note that the parameters of the TB Hamiltonian must be estimated empirically.
However, this approach keeps some of the fundamental physics of the problem
through a quantum-mechanical description of the electronic degrees of freedom,
while the minimal basis and the sparse Hamiltonian make it very efficient. Mercer
(1996) extended this method to compounds, while Bernstein and Kaxiras (1997)
utilized it for defects and interfaces in silicon.

9.5 Direct Minimization of Total Energy

The need for direct minimization of the total-energy functional arises because when
one attempts to find the electronic states that minimize the Kohn–Sham functional
by an indirect minimization technique, certain instabilities in the evolution of the
506 9. Atomistic Modeling of Materials

electronic states may be encountered. Such instabilities will not arise in a direct
method. In this section, we describe two of such methods and their application to
minimization of the Kohn–Sham total-energy functional.

9.5.1 The Steepest-Descent Method

One of the best-known methods for minimizing a function f (x) of a 2D or 3D
variable x is the steepest-descent method, which is the oldest and most straightfor-
ward minimization technique. The algorithm proceeds by making an initial guess
x1 and then improving it by moving in the direction where the functional f (x)
appears to change most rapidly. The steepest-descent direction is aligned with the
vector
v1 = −∇f (x = x1 ). (130)
To reduce the value of f (x), one travels in the direction of v1 from x1 to
x = x1 + bv1 (where b is a scalar parameter), where the function is minimum.
Thus, one can, for example, sample f (x) at a number of points along x1 + bv1 in
order to determine that value of b which minimizes f (x). This procedure mini-
mizes only the value of the function along a particular line, and thus finds a local
minimum. To find the absolute minimum of f (x), one carries out a series of such
line minimizations by using x1 + bv1 as the starting point for the next iteration
and obtaining x2 . Thus, a series of such vectors xi is obtained such that the value
of f (xi ) decreases with increasing i, the iteration number.
However, although this method is guaranteed to converge to the true minimum,
the rate of convergence can be prohibitively slow. For example, if f (x) has narrow
valleys, successive approximations bounce off opposite sides, slowly approaching
the bottom. Moreover, after a minimization is performed along a given gradient
direction, a subsequent minimization along the new gradient re-introduces errors
that are proportional to the previous gradient. As such, the steepest-descent method
is not an attractive technique for minimization of the total energy.

9.5.2 The Conjugate-Gradient Method

Consider now the following symmetric and positive-definite functional form,
1
f (x) =x · ∇ · x. (131)
2
Suppose that we wish to minimize f (x) along a direction d1 , starting from x1 .
The minimum will be at a point x2 = x1 + b1 d1 , where b1 satisfies the following
equation
(x1 + b1 d1 ) · ∇ · d1 = 0. (132)
A subsequent minimization along a direction d2 yields x3 = x2 + b2 d2 , with b2
satisfying
(x1 + b1 d1 + b2 d2 ) · ∇ · d2 = 0. (133)
 9.5. Direct Minimization of Total Energy 507

The best choice of b1 and b2 for minimizing f (x) along d1 and d2 is obtained by
differentiating Eq. (131) with respect to b1 and b2 and evaluating the result at x3 .
This procedure yields two equations,
(x1 + b1 d1 + b2 d2 ) · ∇ · d1 = 0, (134)
(x1 + b1 d1 + b2 d2 ) · ∇ · d2 = 0. (135)
However, in order for Eqs. (132) and (133) to be consistent with (134) and (135),
one must have
d1 · ∇ · d2 = d2 · ∇ · d1 = 0, (136)
implying that the directions d1 and d2 must be conjugate to each other, hence the
name conjugate-gradient (CG) method. More generally, the directions di and dj
must be such that
di · ∇ · dj = 0, i
= j. (137)
Hence, in the CG method one takes the initial direction to be −∇f (x1 ), and
the subsequent directions are constructed from a linear combination of the new
gradient and the previous direction that minimized f (x). In practice, the new
direction di in the ith iteration is obtained from
di = vi + ωi di−1 , (138)
where
vi · vi
ωi = , (139)
vi−1 · vi−1
and ω1 = 0. Note that vi = −∇f (xi ). It has also been observed that in some cases
a better estimate of ωi is given by
(vi − vi−1 ) · vi
ωi = . (140)
vi−1 · vi−1
In the CG method a function is guaranteed to converge to its true minimum. The
reason is that since minimization along the conjugate directions are independent,
each iteration reduces the dimensionality of the vector space by 1. Thus, one
reaches the point at which the dimensionality of the function space is zero, i.e.,
there are no new directions left along which one can minimize the function, and
therefore the trial vector must be at the position of the true minimum. The number
of the iterations needed to reach the true minimum is therefore at most equal to the
dimensionality of the vector space, although in practice it usually takes far fewer
iterations to converge to the true solution.

9.5.3 Minimizing the Total Energy by the Conjugate-Gradient

Method
The CG method can be used for simultaneous updating of all the electronic wave
functions. Although, from a computational view point, this is done very efficiently,
it does have the drawback that it requires a large amount of memory for storing data
508 9. Atomistic Modeling of Materials

between the iterations in order to ensure the conjugacy of the search directions.
The ideal method is one that takes advantage of the efficiency of the CG method
without requiring much more computer memory. This can be achieved by updating
a single band at a time. The steepest-descent direction for a single band is given
by

ζim = −(H − λm m
i )i , (141)

where m denotes the iteration number and i labels the band. Here [see Eq. (111)],

i = i |H|i .
λm m m
(142)

However, since the electronic wave functions must be orthogonal, one must en-
sure that the steepest-descent vector is orthogonal to all the other bands. The
components of the steepest-descent direction vector that ensures this is given by


ζim = ζim − j |ζim j . (143)
j
=i

Note that j does not vary during iteration for band i.

Successive steps along the CG directions reduce the magnitude of the error in
the wave function. It can be shown that the components ζi of the steepest-descent
vector are only a multiple of the components i of the error vector if all the
unoccupied eigenstates of the Kohn–Sham Hamiltonian are degenerate. However,
since the Kohn–Sham Hamiltonian has a broad spectrum of eigenvalues, extending
up to the cutoff energy for the plane-wave basis set (see Section 9.1.4), it can lead
to poor convergence in the CG calculations. Each step tends to remove components
of the error vector that correspond to eigenstates in a particular energy range. The
rate of convergence can be improved significantly by using a pre-conditioning
technique. In this method, one multiplies the steepest-descent vector ζ by a pre-
conditioning matrix K to generate a pre-conditioned steepest-descent vector η
that represents more accurately the error vector . Although, there exists, in
principle, an exact pre-conditioning matrix that pre-conditions the steepest-descent
vector such that it would be parallel to the error vector, its construction is very
expensive, since it is a full matrix. Therefore, one must do the pre-conditioning in
an approximate way.
To carry out an approximate pre-conditioning, we note that the higher-energy
eigenstates of the Kohn–Sham Hamiltonian are dominated (in a pseudo-potential
calculation) by plane-wave basis states, the kinetic energies of which lie close to
the eigenvalues of the state. Therefore, to make those states, with eigenvalues that
are dominated by their kinetic energy, nearly degenerate, one must remove the
effect of the kinetic energy operator from the Hamiltonian, which can be achieved
by using a diagonal pre-conditioning matrix which is essentially the inverse of the
kinetic-energy operator. It has been found that pre-conditioned steepest-descent
vectors that accurately represent the errors in the wave functions can be obtained
by multiplying the steepest-descent vectors by a pre-conditioning matrix K with
 9.5. Direct Minimization of Total Energy 509

entries KG,G that are given by

8x 3 + 12x 2 + 18x + 27
KG,G = δG,G , (144)
16x 4 + 8x 3 + 12x 2 + 18x + 27
with
(h̄2 |k + G|2 )
x= , Tim = im |(−h̄2 /2m)∇ 2 |im ,
2mTim
where Tim is the kinetic energy of the state im . KG,G has the property that it ap-
proaches unity as x → 0, with its first, second and third derivatives all being zero,
hence ensuring that the small wave-vector components of the steepest-descent
vector remain unchanged. Moreover, for x > 1, KG,G asymptotically approach
[2(x − 1)]−1 with an asymptotic expansion correct to fourth order in 1/x. There-
fore, this factor causes all of the large wave-vector components to converge at
nearly the same rate.
The pre-conditioned steepest-descent vector ηm is now given by
ηm = Kζ  m , (145)
which is not orthogonal to all the bands. The components of the pre-conditioned
steepest-descent vector which is orthogonal to all the bands is then calculated as


ηi m = ηim − im |ηim  − j |ηim j . (146)
j
=i

The components of the pre-conditioned conjugate directions dim are now given by
[see Eq. (138)]
dim = ηi m + ωim dim−1 , (147)
with [see Eq. (139)]
ηi m |ζi m 
ωim = , (148)
ηi m−1 |ζi m−1 
with ωi1 = 0. The resulting conjugate directions will be orthogonal to all the
other bands, except the wave function of the present band. Thus, a further
orthogonalization to the present band is done via
di m = dim − im |dim im , (149)

di m
di m = . (150)
di m |dim 1/2
Having constructed the pre-conditioned conjugate direction, the search for the
minimum energy begins. A vector with the following components
im cos θ + di m sin θ (151)
is a normalized vector orthogonal to all the other bands j (with j
= i and θ real),
and thus satisfies the constraints of orthonormality required for the electronic wave
510 9. Atomistic Modeling of Materials

functions. The value of θ that minimizes the Kohn–Sham energy functional is then
required by the CG method. It has been found that the following approximation to
the Kohn–Sham energy,
E(θ ) = E + A cos(2θ ) + B sin(2θ ), (152)
is sufficient for locating its minimum. There are three constants, A, B, and E, and
therefore one needs three data points to evaluate them. One piece of information is
provided by E(θ = 0) which is already known. The second data point is supplied
by the fact that since at θ = 0
∂E 
= dim |H|im  + im |H|dim  = 2Re dim |H|im  , (153)
∂θ
and that, in order to determine the components ηim of the steepest-descent vector,
H|im  has been determined, computing E  (0) = ∂E/∂θ , evaluated at θ = 0 is
cheap. The third data point can be taken to be the value of the Kohn–Sham energy
at a point other than θ = 0. This point should be selected to be far enough from
θ = 0 to avoid round-off errors, yet not so far from θ = 0 that the estimate of
E  (0) becomes inaccurate. It has been found that θ = π/300 gives very reliable
results. With these three data points, the three constants are found to be
2E(0) − 2E(π/300) + E  (0)
A= , (154)
2[1 − cos(2π/300)]
1
B = E  (0), (155)
2
2E(π/300) − E  (0) − 2E(0) cos(2π/300)
E = . (156)
2[1 − cos(2π/300)]
We can now determine straightforwardly the value of θmin that minimizes the
Kohn–Sham energy, since the stationary point of Eq. (152),
$ %
1 −1 B
θs = tan , (157)
2 A
that lies in the range (0, π/2) is the required θmin for minimizing the Kohn–Sham
energy. To start the next iteration of the CG method, a new wave function, given
by
im+1 = im cos(θmin ) + dim sin(θmin ), (158)
is utilized. Note that each new wave function generates a new charge density, and
therefore the electronic potentials in the Kohn–Sham Hamiltonian must be updated
before starting the next iteration.
In practice, only a few iterations are performed on any given band before moving
to the next band, since there is little to gain from converging a single band exactly
while there are still errors in the estimates of the remaining bands. Once all the
bands have been updated, the CG iterations are started again on the lowest band.
One can also perform the CG iterations on one band until the total energy changes
 9.6. Vectorized and Massively-Parallel Molecular Dynamics Simulation 511

by less than a given fraction of the change of energy in the first CG iteration, and
then start iterating on the next band.

9.6 Vectorized and Massively-Parallel Molecular

Dynamics Simulation
Molecular dynamics simulations do not usually use large amounts of computer
memory because one must record only the vectors that contain information about
the atoms. Computationally, the MD simulations become very intensive and large
scale when one wishes to use a large number of atoms and simulate “long” periods
of time. Such large-scale simulations are necessitated by the fact that, since the
effective size of atoms is typically several angstroms, to approach even sub-micron
length scales, one must use millions of atoms in the simulations. Moreover, the
typical time steps that are used in the MD simulations are of the order of femtosec-
onds, and therefore the equations of motion must be integrated over hundreds of
thousands of time steps in order to simulate picosends of real time. Such length
and time scale constraints conspire together to make the MD simulations very time
consuming. They have also provided the impetus for developing efficient MD al-
gorithms that are optimized for vector supercomputers (see, for example, Grest et
al., 1989; Schöen, 1989; Morales and Nuevo, 1992), and to special-purpose com-
puters for carrying out MD computations (Auerbach et al., 1987; Bakker et al.,
1990). Let us first briefly discuss the vectorized algorithms for the MD simulations.

9.6.1 Vectorized Molecular Dynamics Algorithms

As discussed in Section 9.2, when the forces are short-ranged, one usually uses a
cutoff distance rc such that any two particles that are apart a distance larger than
rc do not interact with each other. Thus, the key to efficient MD simulations with
short-range forces is minimizing the number of neighboring atoms that must be
checked for possible interactions. An efficient algorithm for doing this was first
suggested by Verlet (1967), and is commonly referred to as the neighbor lists
method. In this method, a list is constructed for every atom in the simulation cell
that contains the nearby atoms that are at an extended distance re = rc + δ. Relative
to rc , the value of δ is small, but its optimal value depends on such parameters
as the temperature and particle density. These lists are updated after every few
integration steps, before any atom has moved from a distance r > re to r < re .
Hockney et al. (1974) proposed another method, usually referred to as the link-
cell method, in which at every time step all the atoms are binned into 3D cells
of linear size l with l = rc or slightly larger. Therefore, for every atom one must
check only 27 bins—the bin the atom is in and the 26 bins surrounding it. A very
efficient algorithm results when one combines the two methods such that the atoms
are binned only every few time steps in order to construct (or update) the neighbor
lists. The size of the cells is l > re . At intermediate time steps the neighbor lists
512 9. Atomistic Modeling of Materials

are used in the usual way in order to identify neighbors within a distance rc of each
atom. The combined algorithm is made even more efficient by taking advantage of
Newton’s third law which allows one to compute a force for each pairs of atoms,
instead of once for each atom in the pair. This reduces the necessary searches to
only half the surrounding bins of each atom to form its neighbor list. In this way
an atom j is stored in atom i’s list, but not vice versa, hence halving the number
of force computations that must be done.
Another factor that can increase the efficiency of a vectorized MD algorithm
is careful data and loop structures, without which the computer program cannot
be completely vectorized. For example, Grest et al. (1989) combined neighbor
list/link-cell method to create long lists of pairs of neighboring atoms. At each
time step, they updated the lists to keep only those particle pairs that were within
the cutoff distance rc . They also organized the lists into packets in order to prevent
any atom from appearing twice in the lists. In this way, the force computations for
each packet was completely vectorized, resulting in an algorithm that was about
one order of magnitude faster than an algorithm without such data structures.
Another way of developing highly efficient algorithms for MD simulations is
by taking advantage of the natural parallelism that exists in such simulations when
short-range forces are involved. We now discuss such algorithms in detail.

9.6.2 Massively-Parallel Molecular Dynamics Algorithms

If the forces that act on the atoms are short-ranged, then MD computations can
benefit greatly from parallel algorithms. Such algorithms began to emerge in the
late 1980s and early 1990s, and include those of Raine et al. (1989), Bruge and
Fornili (1990), Mel’cuk et al. (1991), Lin et al. (1992), Essenlink et al. (1993),
Form et al. (1993), Rapaport (1993), Lomdahl et al. (1993), Beazley and Lom-
dahl (1994), Smith and Forester (1994), Plimpton (1995), Stadler et al. (1997), and
many more, with the latest (at the time of writing this book) being that of Roth et al.
(2000). Beazley et al. (1995) reviewed and compared many of these algorithms.
Most of the work in this direction has been for single-instructor/multiple-data
(SIMD) parallel machines, or for multiple-instruction/multiple-data (MIMD) par-
allel computers. The latter machines with several thousands of processors possess
the computational power that is comparable with the fastest vector computers.
Here, we describe three parallel algorithms for MD simulations that were devel-
oped by Plimpton (1995). His algorithms and their variants have been used in
large-scale simulations of materials and have proven to be highly efficient. These
algorithms are for systems in which the forces acting on the molecules are short-
ranged. For such short-range MD simulations the computational effort per time
step scales as N, the number of atoms used in the simulations. Another feature
of these algorithms is that the atoms can undergo large displacements over the
duration of the simulation. Finally, the performance of these algorithms is optimal
with respect to both the numbers N of the atoms and Np of the processors. As
such, they are very flexible and efficient.
 9.6. Vectorized and Massively-Parallel Molecular Dynamics Simulation 513

The reason that MD simulations with short-range forces are amenable to par-
allel computations is that calculations of the forces and updating of the positions
can be done simultaneously for all the atoms. Thus, the main goal is to divide
the force computations evenly among the processors so as to achieve maximum
parallelism. There are at least three ways of achieving this, and what follows is a
brief description of each, which are patterned closely after Plimpton (1995).

9.6.2.1 Atom-Decomposition Algorithms

In the atom-decomposition (AD) algorithm, at the beginning of the MD simulations
each of the Np processors is assigned N/Np atoms. Atoms in a group do not have
to have any special relation with each other. We assume, for the sake of simplicity,
that N is a multiple of Np , although the algorithm is general. Each processor
computes the forces on its N/Np atoms and updates their positions and velocities,
regardless of where they move in the physical space. In general, one has an N × N
force matrix F, such that the entry Fij of F represents the force on atom i due to atom
j . Since the forces are short-ranged, F is sparse and, moreover, due to Newton’s
third law, Fij = −Fj i . Suppose that r and f are vectors of length N which store
the position and total force on each atom, so that for a 3D simulations ri would
store the three coordinates of atom i. The AD algorithm assigns each processor a
sub-block of F consisting of N/Np rows of the matrix. We number the processors
from 0 to Np − 1, so that processor Pm (with m = 0, 1, · · · , Np − 1) computes
matrix entries in the Fm sub-block of rows, and is also assigned the corresponding
sub-vectors rm and fm , each of length N/Np .
A most important aspect of the computations is the communication between the
processors, since to compute all the entries in Fm , processor Pm needs the posi-
tions of many atoms that belong to other processors. This gives rise to all-to-all
communication, an operation that, at every time step, supplies the updated posi-
tions of the particles to all the processors. Various algorithms have been developed
for performing this operation efficiently. We describe the algorithm due to Fox
et al. (1988), utilized by Plimpton, who referred to the all-to-all communication
procedure as an expand operation. To store the entire vector r, every processor
allocates memory of length N. At the beginning of the expand operation, proces-
sor Pm has vector rm , the updated segment of r of length N/Np . As the updated
information is supplied to the processors, they must store them in the right place
in their own copy of r. For example, such steps for an eight-processor machine are
illustrated in Figure 9.7, where the processors have been mapped consecutively
onto the sub-pieces of the vector. In the first step of the communication, each
processor exchanges sub-pieces of information. For example, processor 2 does
this with processor 3. After this step, each processor has a piece of r with length
2N/Np (factor 2 is caused by the new information that has just arrived). In the
second step, each processor exchanges its new piece with another processor two
positions away, after which each processor has a piece of r with length 4N/Np .
The last step consists of having each processor exchanging a piece of r of length 12 N
514 9. Atomistic Modeling of Materials

Figure 9.7. Expand (left) and fold (right)

operations among 8 processors, each re-
quiring three steps. During the expand,
processor 2 receives successively longer
sub-vectors from processors 3, 0, and 6. In
the fold, processor 2 receives successively
shorter sub-vectors from processors 6, 0,
and 3 (after Plimpton, 1995).

with a process or 12 Np positions away, after which each processor has the complete
vector r.
The inverse of the expansion operation is also a useful procedure and is com-
monly called a fold operation. Suppose that each processor has stored new force
values in its copy of f. Therefore, processor Pm needs to know the N/Np values
in fm , where each of the values is summed across all Np processors. This is also
shown in Figure 9.7. Thus, each processor exchanges half the vector with a pro-
cessor that is 12 Np positions away, such that it receives the half that it is a member
of, and sends the half that it is not a member of, and then sums the received values
with its corresponding retained sub-vector. This operation is repeated such that at
each step the length of the exchanged data is halved. Note that the expand and fold
operations are optimal, since each processor carries out log2 Np sends and receives
and exchanges N − N/Np additions in the fold operation. The main disadvantage
of this algorithm is that, it requires O(N ) storage on every processor.
The AD algorithm can be implemented in two different ways. In the first one,
which we refer to as the AD1 algorithm, one assumes that, at the beginning of each
time step, every processor has an updated copy of the vector r, and thus “knows”
the positions of all the N atoms. Then, the step-by-step implementation of the AD1
algorithm is as follows.
(1) This step consists of constructing the neighbor lists for all the pairwise in-
teractions that must be computed in block fm . This is not done at every time
step, but after every few steps. It is more efficient for a processor to inspect
all the N 2 /Np pairs in its fm , if the ratio of the physical size of the system to
rc is small. For large simulations, however, in which four or more bins can be
created in each dimension, it is more efficient for a processor to bin all the N
atoms and then inspect the 27 surrounding bins, the length of each is N/Np .
The overall scaling of this inspection process is about N + N/Np .
(2) The neighbor lists are now utilized for calculating the non-zero matrix entries
in Fm . Since each pairwise interaction force is calculated and the force com-
ponents are summed into fm , Fm is never actually stored as a matrix, hence
saving a lot of computer memory. At the end of this step, each processor knows
the total force fm on each of the N/Np atoms that it owns.
(3) The information in (2) is now used to update the positions and velocities of
the particles (i.e., by using them in the equations of motion and integrating
them over one time step).
 9.6. Vectorized and Massively-Parallel Molecular Dynamics Simulation 515

(4) The updated positions of the atoms obtained in the previous step are now shared
among all the Np processors, using the expand operations defined above, in
preparation for the next time step.
This algorithm does not take advantage of Newton’s third law. Algorithm AD2
does use this law and thus reduces the time of the computations by decreasing the
cost of communication between the processors. In order to do this, another matrix
G is used such that Gij = Fij , except that Gij = 0 if i > j and i + j is even,
or when i < j and i + j is odd. Thus, for example, step 1 of AD2 is similar to
that of AD1, except that only half as many neighbor list entries are constructed by
each processor because Gm has only half the non-zero entries of Fm . Similarly,
if the neighbor lists are constructed by binning, then although all the N atoms
must be binned, each processor needs to inspect only half the surrounding bins of
each of its N/Np atoms. In step 2 of the AD2 algorithm, the neighbor lists are
utilized for computing the entries of Gm . For the interaction between atoms i and
j , the forces acting on i and j are summed into both i and j locations of force
vector f, implying that each processor must store a copy of the entire force vector,
as opposed to storing only fm done in the AD1 algorithm. After all the matrix
entries are computed, f is folded across all the Np processors, hence enabling each
processor Pm to have fm , the forces acting on its atoms. Steps 3 and 4 of the AD2
algorithm are the same as that of AD1.

9.6.2.2 Force-Decomposition Algorithms

The force-decomposition (FD) algorithm is based on a block-decomposition of the
force matrix F, rather than a row-wise decomposition used in the AD algorithm.
We assume, for the sake of ease of exposition, that Np is a power of 2 and that N is
a multiple of Np , although the algorithm is general. Sub-blocks of the force matrix
F are assigned to the processors. Figure 9.8 shows the decomposition of the matrix
among 16 processors. The block-decomposition shows there is actually a permuted
force matrix F , formed by rearranging the columns of F ! in a particular way. If we
arrange the rα pieces in row-order (where α = 1, 2, · · · , Np − 1), they form the
usual position vector r which is shown as a vertical bar in Figure 9.8.! One spans
the columns with a permuted position vector rβ (where β = 1, 2, · · · , Np − 1),
shown as a horizontal bar in the figure. Thus, in the example of Figure 9.8, r stores
each processor’s piece in the usual order (0, 1, 2, · · · , 15), whereas r stores them
as (0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15). In this case, Fij is the force
acting on atom i in vector r due to atom! j in permuted! vector r .

The size of sub-block Fm is (N/ Np ) × (N/ Np ). Thus, to compute the en-
tries of Fm , processor Pm must know rα and rβ of the r and r vectors, each
!
with a length N/ Np . These elements are computed and accumulated into corre-
sponding force sub-vectors fα and fβ . Thus, for example, for processor 6 of Figure
9.8, rα consists of the r sub-vectors (4 − 7) and rβ is made of r sub-vectors
(2, 6, 10, 14). There are actually two ways, FD1 and FD2, by which this algorithm
can be implemented. Algorithm FD1 consists of the following steps.
516 9. Atomistic Modeling of Materials

Nβ

  !

Nα " # .$ %

& '  

 ! " #

Figure 9.8. The division of the permuted force matrix F  among 16 processors√ in the force-
√
decomposition algorithm. Processor 6 is assigned a sub-block F6 with size N/ P × N/ P
√
which, to compute its matrix elements, must know the corresponding N/ P −length pieces
xα and xβ and permuted position vector x (after Plimpton, 1995).

(1) Neighbor lists are constructed. If N is small, construction of the lists is most
efficiently
! done by checking all the N 2 /Np possible pairs in Fm . For large N

the N/ Np atoms in rβ are binned and the 27 surrounding bins of each atom
in rα are checked.
(2) The neighbor lists are utilized for computing the entries of Fm . As they are
computed, the entries are summed into a local copy of fα , and therefore Fm
does not have to be stored in matrix form.
(3) By a fold operation within! each of the rows, processor Pm obtains the total
force fm acting on its N/ Np atoms.
(4) Processor Pm utilizes fm to update the positions of the N/Np atoms in rm .
(5) The processors in row α perform an expand operation on their rm sub-vectors,
so that each of them obtains the entire rα . Similarly, the processors in each
column β perform an expand operation on their rm . It is in this step that using
the permuted force matrix F saves extra communications and hence computer
time.
However, the FD1 algorithm does not take advantage of Newton’s third law, and
therefore calculates each pairwise interaction force twice. In algorithm FD2, this
is avoided by using the modified force matrix G, the construction of which was
discussed for the AD2 algorithm. Then, G is permuted in the same way as F to form
G . Step 1 of FD2 is the same as in FD1 (except that half as many interactions are
stored in the neighbor lists). In step 2, for each ij entry, the computed components
of the force are now summed into two sub-vectors instead of one, such that the
force acting on atom i is summed into fα in the location corresponding to row i,
while the force acting on atom j is summed into fβ in its proper location. Step
!
3 consists of three stages. First, the Np processors in column β fold their local
 9.6. Vectorized and Massively-Parallel Molecular Dynamics Simulation 517

copies of fβ , resulting in fm , each element of which is the sum of an entire column
of G . Next, the same type of operations are performed for the row contributions,
resulting in fm , each element of which is the sum of an entire row of G . Finally,
the column and row contributions are subtracted element by element in order to
obtain the total forces fm acting on the atoms that belong to processor Pm , which
can now update the positions and velocities of its atoms. Steps 4 and 5 of FD2 are
identical to those of FD1.

9.6.2.3 Spatial-Decomposition Algorithms

In the spatial-decomposition (SD) algorithm, the simulation cell is divided into
small 3D boxes. Each box is assigned to one processor which, at each time step,
computes focres on and updates the positions and velocities of all atoms that
it contains. The atoms that are assigned to the processors can change; they are
assigned as they move through the simulation cell. To compute the forces acting on
its atoms, a processor only needs to know the positions of atoms in the nearby boxes,
and therefore, unlike the AD and FD algorithms, the communication in the SD
algorithm is local. The size and shape of the boxes depend on N, Np , and the aspect
ratio of the simulation cell (assumed to be a 3D rectangular parallelepiped). The
number of the processors in each dimension is selected so as to make the boxes as
cubic as possible, since such configurations minimize the cost of communications,
when N is very large, which is proportional to the boxes’ surface. The linear size
of the boxes can be larger or smaller than the cutoff distances rc and re defined
above.
Each processor maintains two data sets, one for the N/Np atoms that have been
assigned to it, and one for the atoms in the nearby boxes. The first data set stores all
the information, such as the positions, velocities, neighbor lists, etc. These data are
stored in a linked list to allow insertions and deletions as the atoms move to new
boxes. The second data structure maintains and updates, through communication
with other processors, only the atoms’ positions. The communication scheme for
the SD algorithm is shown in Figure 9.9. In the first step each processor exchanges
information with its neighbors in the east-west direction. Thus, processor 2 (see
Figure 9.9) fills a message buffer with atoms’ positions that it owns that are within

east/west exchanges
1 2 3 up/down exchanges

north/south exchanges
Figure 9.9. The spatial decomposition algo-
rithm. In six data exchanges all atom positions
2
in adjacent boxes in the east/west, north/south,
and up/down directions are communicated
(after Plimpton, 1995).
518 9. Atomistic Modeling of Materials

a distance re of processor 1’s box, which will be all of its atoms if its dimension l
in the east/west direction is less than re . Otherwise, it will be those that are nearest
to processor 1. Then, each processor sends its message in the westward direction
and receives a message from the eastward direction which puts it in its second data
structure. Then, the process is reversed with each processor sending a message
to the east and receiving one from the western processor. If l > re , all needed
atomic positions in the east-west direction are acquired with one such exchange.
If, however, l < re , the east-west steps are repeated, with each processor sending
more needed atomic positions to its adjacent processor. This process is repeated
until each processor knows all atoms’ positions within a distance re of its box
(dashed boxes in Figure 9.9).
The same procedure is then repeated in the north/south direction. The only
difference with the east-west step is that the messages that are now sent to the
adjacent processor contain not only the atoms that belong to the processor in its
first data structure, but also any atom positions in its second data structure that are
needed for the neighboring processor. If, for example, l = re , this has the effect
of sending three boxes worth of atom positions in one message. In the final step,
the process is repeated in the up-down direction in which the atoms’ positions
from an entire plane of boxes are sent in each message. This algorithm has several
advantages, some of which are as follows.
(1) If l ≥ re , all the needed atom positions from the 26 surrounding boxes are
obtained in only 6 data exchanges. Moreover, if the topology of the parallel
machine is that of a simple-cubic lattice, the processors can be mapped onto the
boxes in such a way that all six of these processors will be directly connected
to the center processor, as a result of which the passage of information will be
very efficient.
(2) If l < re , then atom information is needed from more distant boxes, but the
information is obtained by only a few data exchanges, all of which are still
with the 6 immediate neighbor processors.
(3) The amount of data communication is minimized, since every processor
obtains only the atom positions that are within a distance re of its box.
(4) No time is spent for rearranging the data structures. All the newly-arrived
information are placed as contiguous data directly into the processor’s second
data structure. The only time spent is for creating the buffered messages that
must be sent.
(5) Message creation is done quickly. The two data structures are scanned only
once after every few time steps, when the neighbor lists are created, in order
to decide which atom positions to send in each message.
Suppose that box number m is assigned to processor Pm which stores the po-
sitions of its N/Np atoms in rm , the first data structure, and the forces acting on
these atoms in fm . The details of the procedure for implementing the SD algorithm
is as follows.
(1) The positions, velocities, and any other information about atoms that are no
longer inside box m are moved from rm to a message buffer. These atoms are
 9.6. Vectorized and Massively-Parallel Molecular Dynamics Simulation 519

then exchanged with the six adjacent processors (see above), during which
processor Pm checks for new atoms that are now inside its box, which are
then added to rm . Then, all atom positions that are within a distance re of
box m are obtained by the communication scheme described above. Since the
different messages are buffered by scanning through the two data structures,
lists of included atoms are constructed, after which the two data structure of
the processor are updated. Then, neighbor lists of the N/Np atoms of the
processor are constructed. If two atoms i and j are in the same box, then the
pair (ij ) is stored once in the neighbor list. If i and j are in different boxes, their
corresponding processors both store them in their respective neighbor lists. If
l < 2re , it is more efficient to find neighbor interactions by checking each atom
inside box m against all the !atoms in both data structures of the processor, an
operation that scales as O( N/Np ). If l > 2re , it is more efficient to construct
the list by binning described above, by mapping all the atoms in both data
structure onto bins of size re , and checking the surrounding bins of each atom
in box m for possible neighbors.
(2) Processor Pm uses the neighbor lists to compute all the forces that act on its
atoms. If both atoms i and j are inside the same box, then the computed force
between them is stored twice in fm , once each for i and j . If the two atoms are
in different boxes, only the force acting on the processor’s own atom is stored.
(3) After computing fm , the atoms’ positions are updated.
(4) The updated positions are communicated to the surrounding processors in
preparation for MD simulation in the next time step.
The cost of these steps scales as the volume of the data exchanged. For example,
for the last step, assuming that we have uniform density in the simulation cell, the
cost is proportional to the physical volume of the shell of thickness re around box
m, which is (l + 2re )3 − l 3 . Thus, three cases must be considered. If l < re , data
from many neighboring boxes must be exchanged and the operation cost scales as
8re3 . If l  re , the data in all the 26 surrounding boxes are exchanged and the cost
of operation scales as 27N/Np (note that N/Np is roughly the number of atoms
in volume l 3 ). Finally, if l  re , only atoms’ positions near the six faces of box
m will be exchanged, and therefore the cost of the communications scales as the
surface area of the box, namely, 6re (N/Np )2/3 . As before, one can take advantage
of Newton’s third law and make the algorithm still more efficient.

9.6.2.4 Load Balance in Massively-Parallel Molecular Dynamics Simulation

We must point out that all the above algorithms have their optimal efficiency only if
there is load-balance, which means that each processor must have an equal amount
of work to perform. For the AD algorithms this will be the case if each Fm or Gm
block has roughly the same number of non-zero entries, which will be the case if
the atom density is uniform across the simulation cell. Typically, non-uniformities
do arise, but they will not pose any problem if the atoms ordering at the beginning
of the simulations is randomly permuted, which is equivalent to permuting rows
and columns of F or G.
520 9. Atomistic Modeling of Materials

There will be load-balance for the FD algorithms only if Fm and Gm  are uni-

formly sparse. Thus, if the atoms are ordered geometrically, these two matrices
will not be uniformly sparse, since geometrical ordering generates a force matrix
with diagonal bands of non-zero entries. The way to achieve uniform sparsity is
to randomly permute the atoms ordering.
Finally, the SD algorithms will be load-balanced only if there is roughly the
same number of atoms both in the boxes and in their surroundings. This will
not be the case if there is non-uniformity in the density of the particles, or if
the physical domain is not a rectangular parallelepiped. Hendrickson and Leland
(1995) developed a method for load-balancing the SD algorithms by partitioning
an irregular domain or a system with non-uniformly-distributed clusters of atoms,
although such algorithms add to the computational cost of the MD simulations.

9.6.2.5 Selecting a Massively-Parallel Molecular Dynamics Algorithm

How does one select the most efficient MD algorithm for a given problem?
Plimpton (1995) provides some guidelines.
(1) If the communications cost is expected to be negligible, which is the case when
Np is small, then algorithm AD is preferable.
(2) For all other cases algorithm FD will be faster than the AD, both of which
scale linearly with N for a fixed Np . If we double Np , the communications
cost of the AD algorithm√ remains the same, while that of the FD algorithm
decreases by a factor of 2. Thus, as Np increases and becomes large, the FD
algorithm becomes much faster than the AD algorithm.
(3) For a given Np , the scaling of the SD algorithm is not linear with N. For
large N, the efficiency of the algorithm is optimal and nearly 100%, whereas
for small N the efficiency is poor. Thus, compared to a FD algorithm, there
must be a crossover point when, with increasing N (holding Np fixed), the
efficiency of the SD algorithm exceeds that of the FD algorithm.
Plimpton (1995) tested the efficiency of these algorithms on a benchmark prob-
lem, namely, simulation of N atoms, assumed to interact through a LJ potential.
The atoms were placed in a 3D parallelepiped with periodic boundary conditions,
at a reduced density 0.8442 and reduced temperature 0.72. This represents a liquid
state near the LJ triple point. The simulations were begun by placing the atoms on
an FCC lattice. Note that the unit cell of this lattice contains 4 atoms. The particles
were given initial random velocities (see Section 9.2.4), and after some time the
solid (crystal lattice) melted quickly, as it should, and the system of atoms evolved
toward its natural liquid state. The cutoff distance was taken to be rc = 2.5σ (σ is
the size parameter of the LJ potential), the integration time step, in reduced units,
was 4.62 × 10−3 , and the simulations were carried out in the (NVE) ensemble.
Table 9.3 compares the performance of the SD1 algorithm on several machines.
In this table the dashed sign indicates that the size of the problem was so large that
the machine’s memory was not sufficient for carrying out the computations. The
108 atoms problem nearly filled the 30 Gbytes of memory on the 1904-processor
 9.7. Interatomic Interaction Potentials 521

Table 9.3. CPU seconds/time step for the SD1 algorithm. N is the number of atoms used
in the MD simulations, while Np is the number of the processors of the machine (adapted
from Plimpton, 1995).
Problem Size Cray T3D Intel iPSC/860 Intel Paragon
N Lattice Np = 256 Np = 512 Np = 32 Np = 64 Np = 1024 Np = 1904
5 x 102 5x5x5 0.00432 0.00446 0.0129 0.0106 0.00564 0.00634
5 x 104 20 x 25 x 25 0.0289 0.0167 0.420 0.224 0.0174 0.0125
5 x 106 100 x 100 x 125 1.86 0.994 − − 0.914 0.504
108 250 x 250 x 400 − − − − − 9.11

Paragon with neighbor lists consuming the majority of the space. It is clear that
the algorithm is extremely efficient. However, use of optimized version of the
same algorithm increases its efficiency by a factor 2 − 3. For example, on the Intel
Paragon machine with 1840 processors and an assembler version of the algorithm
(as opposed to, for example, the Fortran version), the CPU seconds/time step was
5.5, a factor of nearly 2 more efficient than what is listed in Table 9.3. Similar
computations were performed for the other algorithms discussed above.
Let us point out that, similar to the classical MD simulations, the electronic
structure calculations and QMD computations can also benefit tremendously from
parallelization. In particular, parallel QMD can be a powerful method for comput-
ing realistic interaction potentials for use in the classical MD simulations (see, for
example, Clarke et al., 1992). This is the subject of the next section.

9.7 Interatomic Interaction Potentials

Although quantum MD simulation has proven to be a highly successful method
for investigating and predicting materials properties at very small scales, due to
the huge computations that are necessary, they cannot be used yet for materials at
scales that involve several thousands atoms. For this reason, classical MD simula-
tions are a valuable tool for studying various properties of materials using millions
of atoms. However, as discussed in Section 9.2, accurate interaction potentials are
critical to the success of MD computations. Although simple interaction forces,
such as those derived from the LJ potentials, have provided us with much qual-
itative insight into the properties of various materials, quantitative predictions
require much more realistic representation of the interactions between the atoms.
Moreover, only for noble gases can one represent the interactions between atoms
by density-independent, pairwise-additive forces, and the repulsive and attractive
forces are due to spherical electron clouds that are close to the nuclei. The alter-
native to the LJ and similar potentials are semi-empirical expressions designed
for accurately describing small distortions from the ground state in more complex
systems, a famous example of which is the Keating potential described in Chapter
8 of Volume I for covalent bonds. Such potentials, which can be viewed as a sort of
Taylor series expansions of the energy about its minimum, are useful for describing
phonons and elastic deformations, but they are incapable of describing the energy
522 9. Atomistic Modeling of Materials

of states which differ significantly from tetrahedral ground state, or when one must
deal with large deformations. For these reasons, developing accurate representation
of the interaction forces between various atoms has been, for many years, an active
research field. We summarize in this section some of the most significant results
that have emerged over the last two decades.

9.7.1 The Embedded-Atom Model

The embedded-atom model (EAM), which is intended for metals, was developed
by Daw and Baskes (1983,1984). In metals, electrons are not all localized around
the nuclei, rather the valence electrons are often shared among many ions, similar
to a nearly free-electron gas. This implies that the energy depends upon the local
electron density, resulting in many-body, rather than pairwise, forces between
ions, hence allowing one to represent the interactions between ions in metals by
a relatively-simple approximate functional form, commonly referred to as the
embedded-atom potential. In this approximation, the total potential energy E of
N ions in an arbitrary volume  is given by
⎡ ⎤

n

E= ⎣1 Uij (rij ) + Eie (ρi )⎦ , (159)
2
i=1 j
=i

where U (rij ) is a density-independent, pairwise-additive and short-range interac-

tion potential that depends only on distance rij between particles i and j , and Eie
is the embedding energy that depends on the local embedding density ρi at atom
i. In effect, each atom is viewed as an impurity embedded in a host consisting of
all other atoms, such that the embedding energy depends on the electron density.
In this sense, the basic idea of the EAM is, on one hand, similar to the effective-
medium approximation described in the previous chapters, and, on the other hand,
similar to the DFT described in this chapter. If one makes a further simplification
by assuming that the density ρi can be approximated by


ρi = ρja (rij ), (160)
j
=i

where ρia is the atomic density of the constituents, then the energy would be a
simple function of the atoms’ positions. Note the difference between ρj and ρja :
Whereas ρja is the contribution to the density from atom j , ρj is the total electron
density at atom j .
As shown by Daw and Baskes (1984), the ground state properties of the solid
can be computed from Eq. (159). For example, consider the case of a perfect,
homonuclear crystal. In this case, Eie = E, U = Uij (rij ), and ρ = ρ a . If ρe is
the equilibrium density, then ρe = m ρ(l m ), where l m are the distances between
neighbors, and the sum is over neighbors. Moreover, one has, for every i, ρi = ρe .
Then, the lattice constant is obtained from the equilibrium condition:
  m m   m m
1
Um li lj 

ρm li lj
+ E (ρ e ) = 0, (161)
2 m lm m
lm
 9.7. Interatomic Interaction Potentials 523

where lim is the ith component of the position vector to the mth neighbor, Um =
dU/dr, and ρm  = dρ/dr, with the subscript m indicating that the derivatives are

to be evaluated at r = l m .
The elastic constants of the crystal at equilibrium are given by
1
Cij kl = [Bij kl + E  (ρe )Wij kl + E  (ρe )Aij Akl ], (162)
0
where
  

1
(Um − Um / l )li lj lk ll
m m m m m
Bij kl = , (163)
2 m (l m )2
   / l m )l m l m l m l l


(ρm − ρm i j k m
Wij kl = , (164)
m
(l m )2
  m m

ρm li lj
Aij = , (165)
m
lm

and 0 is the volume of the undeformed crystal. In particular, for a cubic crystal,
the three independent elastic constants are given by
1
C11 = [B11 + E  (ρe )W11 + E  (ρe )A211 ], (166)
0
1
C12 = [B12 + E  (ρe )W12 + E  (ρe )A211 ], (167)
0
1
C44 = [B12 + E  (ρe )W12 ]. (168)
0
These equations nicely demonstrate the effect of the interplay between the pair
potential Uij and the embedding energy E e . Clearly, if we remove Uij = U , we
obtain C11 = C12 and C44 = 0, which, for real solid materials, are wrong. On
the other hand, if we remove the embedding energy E e = E, we obtain the well-
known Cauchy relation, C12 = C44 , which does not hold for all materials, but only
for a certain class of them.
To use the EAM one must have the embedding energy, the pair potential, and the
atomic densities. In their original work, Daw and Baskes (1984) and Foiles et al.
(1986) used semi-empirical correlations for evaluating these quantities. To develop
such correlations, one takes advantage of known properties of these quantities. For
example, the embedding energy, when defined relative to the free-atom energy,
must vanish for vanishing electron density, and must have negative slope and
positive curvature (second derivative) for the background electron densities found
in typical metals. On the other hand, the pair-interaction term in Eq. (159) is purely
repulsive, and its origin is Coulombic. These observations lead to the following
equation
Zi (r)Zj (r)
Uij (r) = , (169)
r
524 9. Atomistic Modeling of Materials

where Zi (r) is the effective charge of atom i. Note that Zi (r) must be positive
and decrease monotonically with increasing r. A particularly simple, yet accurate,
expression is given by
Z(r) = Zo (1 + a1 r a2 ) exp(−a3 r), (170)
where Zo is the number of the outer electrons in the atom (for example, Zo = 10
for Ni, Pd, and Pt, and Zo = 11 for Cu, Ag, and Au). The parameters a1 , a2 and
a3 must be determined empirically, although a2 = 1 is accurate for Ni, Pd, and
Pt, while a2 = 2 leads to accurate elastic constants for Cu, Ag, and Au. Using
these observations and properties, Daw and Baskes (1984) and Foiles et al. (1986)
obtained very accurate empirical correlations for the embedding energy and the
effective charges for a variety of metals, from which they computed their various
properties, such as their elastic constants and surface energy; see also Johnson
(1988) who utilized the EAM to study FCC metals.
Holian et al. (1991) developed the following EAM for use in MD simulation of
deformation of materials under high stresses. The local embedding density ρi was
assumed to be given by a pairwise sum over all neighboring particles, weighted
by a spherical localization function w(rij ), such that


ρi = w(rij ), (171)
j
=j

where
$ %2
1 rc2 − r 2
w(r) = , (172)
ed(d + 1) rc2 − re2
where d is the dimensionality of the system, e = exp(1), rc is the cutoff distance
given by Eq. (34), and re is the equilibrium nearest-neighbor distance. The pairwise
interaction potential U (r) was taken to be that given by Eq. (33), except that the
LJ part of the potential was written as
  
σ 12  σ 6
ULJ (r) = 4χ ε − , (173)
r r
where χ is the fractional pair-potential contribution to the total cohesive energy
(χ = 1/3 is a reasonable value for many metals). Finally, the embedding energy
Eie was taken to be
1
Eie (ρi ) = eεd(d + 1)(1 − χ )ρi ln ρi . (174)
2
One can also replace the LJ potential by the more flexible Morse-like potential,
UMorse = ε{exp[−α(r/re − 1)] − 1}2 − ε, (175)
where α, the steepness of the repulsive well, is related to 0 , the volume of the
undeformed system, and the bulk modulus of the system at equilibrium.
However, the atomic densities can also be computed by the Hartree–Fock ap-
proximation, and the embedding energy and the pair potentials can be calculated
by the ab initio method. The ab initio electronic structure calculations are utilized
 9.7. Interatomic Interaction Potentials 525

in order to generate an accurate database of configurational energies, which are

then used for determining the necessary parameters of the EAM (see, for example,
Wagner et al., 1992), which take advantage of analytical functional forms obtained
from theoretical analysis. The ab initio electronic structure calculations are for-
mulated in the context of the quantum mechanical DFT described in Sections 9.1,
9.4 and 9.5.
Baskes et al. (1989) proposed a potential based on the EAM by introducing
an angular dependence in the host electron energy, which could then adequately
account for the bond-bending forces in materials such as diamond. The potential
could fit the energies of the high-density polymorphs, with the goodness of the fit
being comparable with those of the other potentials described below. It could also
describe exactly the static properties of cubic diamonds, and provide a relatively
accurate description of bulk defects and vacancies. However, it provides high
values for the energy of the intrinsic stacking faults in silicon.
Most recently, Web and Grest (2001) utilized the EAM and its modification to
compute the surface tension of liquid-vapor interface of metals. The agreement
between the predicted values and the experimental data was very good, hence
indicating that the EAM can also describe fluid interface properties of materials.

9.7.2 The Stillinger–Weber Potential

Stillinger and Weber (1985) proposed a semi-empirical potential for silicon which
appears to be relatively accurate. In their model the total interatomic potential
involves two- and three-body contributions and is written as



E= U2 (rij ) + U3 (rij , rik ), (176)
i<j i,j <k

where U2 , the two-body term, can include such effects as the steric repulsion,
charge transfer between atoms, charge-dipole and van der Waals interactions, and
therefore
Hij Z i Zj Dij Wij
U2 (rij ) = nij + exp(−rij /a) − 4 exp(−rij /b) − 6 . (177)
rij rij rij rij

The first term of Eq. (177) represents a two-parameter representation of the steric
repulsion; the second term is the Coulombic interaction due to charge transfer and
contains the effective atomic charges Zi and Zj ; the third term takes into account
the charge-dipole interaction due to large polarizability of negative ions, while the
last term corresponds to the induced dipole-dipole interactions. Covalent effects
are taken into account through three-body bond-bending and bond-stretching terms
(similar to the Keating model described in Chapter 8 of Volume I), and includes
the Si-C as well as C-Si-C bond angles. It is given by
$ %
γ γ (cos θj ik − cos θ̄j ik )2
U3 (rij , rik ) = Bij k exp − #(r0 − rij )#(r0 − rik ).
rij − r0 rik − r0 1 + C(cos θj ik − cos θ̄j ik )2
(178)
526 9. Atomistic Modeling of Materials

Table 9.4. Comparison of the MD predictions of the

properties of cubic SiC in zinc-blende structure,
obtained by using the Stillinger–Weber potential,
with the experimental data. The elastic constants are
in GPa (adapted from Shimojo et al., 2000).
Property MD Experiment
Lattice constant (Å) 4.36 4.36
C11 390 390
C12 144 142
C44 179 150–256
Bulk modulus 225 225
Melting temperature (◦ C) 3000 2830

Figure 9.10. Volume-pressure relation for

zinc-blend structure. Solid circles shows
the MD results, while open circles are the
experimental data (after Shimojo et al.,
2000).

Here, Bj ik is the strength of the interaction, #(r0 − rij ) is the step function, θj ik
is the angle formed by rij and rik , and θ̄j ik is a constant. The constant C plays an
important role if the material undergoes structural transformation. In the original
Stillinger–Weber formulation, cos θ̄j ik was assumed to have a value of −1/3, but
more generally one can treat this term as an adjustable parameter. Note that the
“ideal” tetrahedral angle θ is such that cos θ = −1/3, so that the trigonometric part
of Eq. (178) clearly discriminates in favor of pairs of bonds emanating from i with
the desired geometry. It is clear that the Stillinger–Weber potential contains many
parameters which must be estimated by fitting the predictions obtained with the
potential to certain properties of the material. Shimojo et al. (2000) employed this
potential in their MD simulations of cubic SiC under isothermal-isobaric condi-
tions (see Section 9.2), using only 1728 atoms. Table 9.4 compares the predictions
of the MD simulations with the experimental data, and it is clear that the agreement
between the two sets is excellent. In addition, the volume-pressure relation was
computed for SiC in the zinc-blende structure, i.e., in a configuration with 4-fold
coordination. Figure 9.10 compares the MD simulation results with the experimen-
tal data, and it is clear that the agreement is again excellent, hence demonstrating
the significance and utility of an accurate interatomic potential: If the interatomic
potential is accurate, then MD simulations provide quantitative predictions for
materials’ properties.
 9.7. Interatomic Interaction Potentials 527

The Stillinger–Weber potential is, by far, the most widely used potential. It
has been utilized in the study of clusters, lattice dynamics, bulk point defects, the
liquid and amorphous states, surface diffusion and reconstructions, Si(100) stepped
surfaces, the liquid-vapor and crystal-melt interfaces, pulsed melting of surfaces,
epitaxial growth from the vapor, liquid-phase epitaxy, and growth of amorphous
films via atom deposition, as well as calculation of mechanical properties that
was mentioned above. It has also been extended to Ge, sulfur, fluorine, and Si-F
materials.
Mistriotis et al. (1989) modified the Stillinger–Weber potential in order to cor-
rectly describe clusters with more than 6 atoms. The angular dependence of the
three-body term was modified, and a four-body term was also added.

9.7.3 The Tersoff Potentials

It can be shown, by simple quantum-mechanical arguments, that the more neigh-
bors an atom has, the weaker the bond to each neighbor will be. In general, the bond
strength, or bond order, depends in a complex way on the geometry of the material.
For example, even-membered rings might be favored over odd ones. However, the
most important single variable is the coordination number—the number of neigh-
bors close enough to form bonds. If the energy per bond decreases sufficiently
rapidly with increasing coordination, then the diatomic molecule will be the most
stable arrangement of atoms. Low coordination numbers are common for atoms
at the far right of the Periodic Table (especially near the top). However, if the
bond order depends only weakly on the coordination number, then close-packed
structures form so as to maximize the number of bonds. This limit corresponds,
roughly speaking, to metallic rather than covalent bonding, and is found for atoms
at the left and bottom of the Periodic Table, with a trend in between from low
coordination numbers at the upper right to high coordination number at the lower
left. Thus, bond order is a monotonically decreasing function of the coordination
number, and a trade off between this property and number of bonds determines
the equilibrium conditions.
Silicon, aside from its technological importance, is a remarkable material since,
even with modest changes of pressure, it can take on structures with a large range of
coordination. This is due to the fact that a decrease in bond strength with increasing
coordination number essentially cancels the increase in the number of bonds, over
a large range of coordination number. As such, silicon provides a stringent test of
our ability for describing the dependence of bonding upon coordination number,
and hence our ability for developing potentials that can accurately describe its
structure.
In an important paper, Tersoff (1988) suggested such a potential for silicon.
Because of the critical role of bond order and its dependence upon local geometry,
one must include an environment-dependent bond order in the potential. Thus, in
Tersoff’s formulation the total interatomic potential energy is taken to be

1


E= Ei = Eij , (179)
2
i i j
=i
528 9. Atomistic Modeling of Materials

where
Eij = Uc (rij )[aij Ur (rij ) + bij Ua (rij )]. (180)
Here Ei and Eij are, respectively, a site and a bond energy, the sums are over the
atoms of the system, and rij is the distance between atoms i and j . The function Ur
represents a repulsive pair potential, which includes the orthogonalization energy
when atomic wave functions overlap (see Section 9.4), while Ua is the attractive
pair potential associated with bonding. As already emphasized throughout this
chapter, in many applications short-ranged functions allow a tremendous reduction
in computational effort, and therefore a cutoff function Uc has been introduced to
limit the range of the potential. The function bij is a measure of the bond order,
and is assumed to be a monotonically decreasing function of the coordination of
atoms i and j . In addition, terms that act to limit the range of interaction to the
first neighbor shell are included in bij , and the function aij consists solely of such
range-limiting terms.
Ferrante et al. (1983) showed that a large number of calculated binding-energies
for solid cohesion and chemisorption can be mapped onto a single dimensionless
function using a three-parameter scaling, and Abell (1985) showed that this univer-
sal behavior can be well-explained by use of a Morse or Morse-like pair potential,
Eq. (175). Therefore, Tersoff (1988) proposed the following expressions for Ur
and Ua :
Ur (r) = A exp(−λ1 r), (181)
Ua (r) = −B exp(−λ2 r), (182)
whereas the cutoff function Uc was taken to be
⎧
⎨ 1, r <R−D
Uc (r) = 1
− 1
sin[π(r − R)/(2D)], R − d <r <R+D (183)
⎩ 2 2
0, r >R+D
The cutoff function (and its derivative) is continuous for all r, and varies between
0 and 1 in a small range around R, which is selected so as to include only the
first-neighbor shell for most structures. In effect, the potential has the form of a
Morse pair potential, ignoring the three-body and higher-order effects, but with a
bond-order parameter bij that depends on the local environment. The function bij
is given by
bij = (1 + βζijn )−1/2n , (184)
with


ζij = Uc (rik )g(θij k ) exp[λ33 (rij − rik )3 ], (185)
k
=i,j

g(θ ) = 1 + c2 {d −2 − [d 2 + (h − cos θ )2 ]−1 }, (186)

where θij k is the angle between bonds ij and ik. The function cos θij k is used to
ensure the proper analytic behavior for the dependence of bij on θij k . Note that
 9.7. Interatomic Interaction Potentials 529

bij
= bj i which, however, does not have any significant consequence. If, however,
one insists on a more symmetric form, the sum over pairs of atoms in Eq. (180)
can be replaced with a sum over bonds (i > j ), and then bij can be replaced
with a symmetric function, b̄ij = 12 (bij + bj i ). Tersoff (1988) also proposed the
following equation for the function aij ,
n −1/2n
aij = (1 + α n ηij ) , (187)


ηij = Uc (rik ) exp[λ33 (rij − rik )3 ], (188)
k
=i,j

where α is typically small enough that aij  1, unless, of course, ηij is exponen-
tially large.
Subsequently, Tersoff (1989) modified his proposed potential in order to describe
multicomponent mixtures, and more specifically C-Si and Si-Ge mixtures. Silicon
carbide, in particular, has a wide range of applications, ranging from optoelectric
devices and engineering materials to the basic substrate for membranes that must
operate at high-temperatures (Suwanmethanond et al., 2000). The reason for its
popularity is that it has excellent chemical stability, good electronic properties,
and high stiffness and hardness. In Tersoff’s generalization, Eqs. (179) and (180)
remain the same, but the remaining expressions are modified to account for the
multicomponent nature of the system. Thus,
Ur (rij ) = Aij exp(−λij rij ), (189)
Ua (rij ) = −Bij exp(−µij rij ), (190)
⎧
⎨ 1, rij < Rij
Uc (rij ) = 1
+ 1
cos[π(rij − Rij )/(Sij − Rij )], Rij < rij < Sij (191)
⎩ 2 2
0, rij > Sij
where the various parameters are given by
bij = χij (1 + βini ζijni )−1/2ni , (192)


ζij = Uc (rik )ωik g(θij k ), (193)
k
=i,j
2 −2
g(θij k ) = 1 + ci {di − [di2 + (hi − cos θij k )2 ]−1 }, (194)
1 1
λij = (λi + λj ), µij = (µi + µj ), (195)
2 2
! !
Aij = Ai Aj , Bij = Bi Bj , (196)
! !
Rij = Ri Rj , Sij = Si Sj . (197)
Note that the parameter χij strengthens or weakens the heteropolar bonds, and
therefore represents in some sense the chemistry of the mixture. Note also that
χii = 1 and χij = χj i , and that ωii = 1, although experience has indicated that
ωij = 1 is also a reasonable estimate. Compared with the case of a pure component,
530 9. Atomistic Modeling of Materials

Table 9.5. Estimates of the parameters for carbon, silicon and

germanium to be used in the Tersoff potential. Except for R and S, all
the parameters have been optimized (adapted from Tersoff, 1989).
Parameter C Si Ge
A (eV) 1393.6 1830.8 1769.0
B (eV) 346.7 471.18 419.23
R (Å) 1.8 2.7 2.8
S (Å) 2.1 3.0 3.1
c 38049.0 100390.0 106430.0
d 4.384 16.217 15.652
h −0.57058 −0.59825 −0.43884
n 0.72751 0.78734 0.75627
β 1.5724 × 10−7 1.1000 × 10−6 9.0166 × 10−7
λ (Å−1 ) 3.4879 2.4799 2.4451
µ (Å−1 ) 2.2119 1.7322 1.7047
χC−Si = 0.9776 χSi−Ge = 1.00061

Table 9.6. Comparison of the MD

predictions of properties of cubic SiC,
obtained with the Tersoff potential,
with the experimental data. The elastic
constants are in Mbar (adapted from
Tersoff, 1989).
Property MD Experiment
Lattice constant (Å) 4.32 4.36
C11 4.2 3.6
C12 1.2 1.5
C44 2.6 1.5

the potential for mixtures is somewhat simpler, as the parameter D of Eq. (183)
has been eliminated.
All the parameters of these potentials have been estimated by fitting them to heat
of formation and the properties of the respective elements. For Si the parameter D
is about 0.15Å. The rest of the parameters are listed in Table 9.5. To provide the
reader with some sense of the accuracy of these potentials, we compare in Table
9.6 the predicted lattice constant and the three elastic constants of cubic SiC with
the experimental data. It is clear that, except for C44 , the agreement between the
theoretical predictions and the data is quite good. In addition, Mura et al. (1998)
studied the properties of Si1−x Cx compounds in the range 0.125 ≤ x < 0.875 by
ab initio QMD method of Car and Parrinello (1985) (see Section 9.4) and by MD
simulations using the Tersoff potentials, and found very good agreement between
the predictions of the two methods.
However, because of the large difference in bonding characteristics between
hydrogen and carbon (recall that H is monovalent whereas C has a valency of
up to 4), a set of parameters cannot be found that can adequately describe bond
energies for a large number of hydrocarbons. Furthermore, the Tersoff potentials
 9.7. Interatomic Interaction Potentials 531

cannot describe radicals and non-conjugated double bonds. These deficiencies

motivated the development of several modified Tersoff potentials which we now
briefly mention.
Chelikowsky et al. (1989) developed an interatomic potential similar in form to
Tersoff’s, which was intended for describing the metallic to covalent transition that
occurs in clusters when their size reaches a critical size. To describe the transition,
an angular-dependent bond-bending force was incorporated in the potential. The
resulting potential provides very accurate description of perfect diamond structure,
as well as the high-density polymorphs of silicon. To model clusters, a so-called
dangling-bond vector was introduced that describes the transfer of bond strength
from a dangling bond to a backbone bond. The energies of Sin clusters with n <
10 are, however, underestimated. Moreover, the ground-state structures for such
values of n are also not well-described. The potential was not intended for bulk
point defects and for surfaces.
Following Tersoff, Kohr and Das Sarma (1989) developed a series of interatomic
potentials for tetrahedrally-bonded semiconductors. The bond-bending term was
modified to deal with the larger angular distortions from the tetrahedral angle
encountered on the various (111) surfaces. Moreover, since the bonds of a given
atom can be of different nature, the value of the effective coordination number
was fixed in an ad hoc way. The potential provides accurate description of various
(111) surfaces.
Bolding and Andersen (1990) generalized Tersoff’s potential by expressing the
attractive term of the potential as a sum of σ - and π-bonding terms. The functional
form is complex, and the potential contains over 30 parameters. Bolding and
Andersen used a very large data base for fitting the parameters, including the static
properties of the cubic diamond phase, the fact that the first pressure-induced
phase transformation from cubic diamond is the β-tin phase, and the energies
and geometries of global and local minima for clusters of 2-10 atoms. For small
clusters, the Bolding–Andersen potential generates a surface that has most of the
local and global minima predicted by the ab initio computations. Moreover, its
predictions for the ground-state structures and energies are in excellent agreement
with those predicted by quantum-mechanical computations. The static properties
of cubic diamond silicon are also well described by the potential. For bulk point
defects, only the vacancy is predicted to have a formation energy that is in good
agreement with the DFT results.

9.7.4 The Brenner Potentials

In addition to the shortcomings of the Tersoff potentials for certain materials, there
is another impetus for developing more sophisticated empirical or semi-empirical
potentials. Chemical vapor deposition (CVD) of diamond films is a process of
tremendous technological importance. It is, however, a complex process in which
diamond grows under apparently metastable conditions. Various factors, such as
addition to the surface of such species as acetylene, methyl radicals, or a mixture
of hydrocarbon molecules, the substrate temperature, and initiation of defects, all
532 9. Atomistic Modeling of Materials

may play important roles. In order to obtain a better understanding of this com-
plex process, atomic-scale simulations of this phenomenon is of course desirable.
However, its QMD simulation is still too costly (in terms of the computation time),
and thus the classical MD computations with an accurate interatomic potential is
desirable. The potential developed by Brenner (1990) is intended for this pur-
pose. Its main use is for hydrocarbons, and it can take into account the effect of
intramolecular chemical bonding (which the Tersoff potentials are incapable of
doing) in many small hydrocarbon molecules as well as graphite and diamond
lattices.
In its spirit, Brenner’s formulation is similar to Tersoff’s. The binding energy
Eb for the hydrocarbon potential is given by



Eb = [Ur (rij ) − b̄ij Ua (rij )], (198)
i j >i

where, similar to the Tersoff potentials, Ua and Ur are the attractive and repulsive
part of the energy, and are given by
 # 
Dije Sij 2  
Ua (rij ) = fij (rij ) exp − βij rij − Rij ,
e
(199)
Sij − 1 Sij

Dije Sij  !  
Ur (rij ) = fij (rij ) exp − 2Sij βij rij − Rije , (200)
Sij − 1
where Dije is the well depth, and Rije is the equilibrium distance, the value of which
is the same as re in the Morse potential, Eq. (175). In fact, for Sij = 2 the attractive
and repulsive potentials are essentially identical with the Morse potential. fij (rij )
are cutoff functions given by
⎧ (1)
⎪
⎨ 1, r ≤ Rij
(1) (2) (1) (1) (2)
fij (r) = 2 + 2 cos[π(r − Rij )/(Rij − Rij )], Rij < r < Rij
1 1
(201)
⎪
⎩ (2)
0, r ≥ Rij

It should be clear that the cutoff functions explicitly restrict the interactions to
nearest neighbors. The function b̄ij is given by
1 (t) (t) conj
b̄ij = (bij + bj i ) + Fij [Ni , Nj , Nij ], (202)
2
implying that, similar to the Tersoff potentials, b̄ij depends on the environment
around atoms i and j and implicitly contains many-body information. The function
Fij is a correction term which is used only for carbon-carbon bonds, and
⎧ ⎫−δi
⎨
  ⎬
(H) (C)
bij = 1 + g(θij k )fik (rik ) exp αij k [(rij − Rije ) − (rik − Rik
e
)] + Hij [Ni , Ni ] .
⎩ ⎭
k
=i,j

(203)
 9.7. Interatomic Interaction Potentials 533

(C) (H)
Here Ni and Ni are the number of carbon and hydrogen atoms bonded to atom
(t) (C) (H) conj
i, Ni = Ni + Ni , Nij depends on whether a bond between carbon atoms
i and j is part of a conjugated system, and θij k is the angle between bonds ij and
ik. Similar to Fij , the function Hij is a correction term, and both functions are
used only for hydrocarbons.
The cutoff functions fij (r) are used for defining the various quantities. One has
(H)


Ni = fij (rij ), (204)
j ={H}
(C)


Ni = fij (rij ), (205)
j ={C}
(t)
where, for example, {C} denotes the set of the carbon atoms. Values of Ni for
neighbors of two carbon atoms involved in a bond are used for determining whether
the bond is part of a conjugated system. For example, if the neighbors are carbon
(t)
atoms that have a coordination number of less four (i.e., Ni < 4), the bond is
defined as part of a conjugated system. For a bond between carbon atoms i and j ,
conj



Nij = 1 + fik (rik )F (xik ) + fj l (rj l )F (xj l ), (206)
k(
=i,j )={C} l(
=i,j )={C}

with
⎧
⎪
⎨ 1, xik ≤ 2
F (xik ) = 1
+ 1
cos[π(xik − 2)], 2 < xik < 3 (207)
⎪
⎩
2 2
1, xik ≥ 3
and
(t)
xik = Nk − fik (rik ). (208)
conj
The function F (xik ) yields a continuous value of Nij as bonds form and break
conj
and as second-neighbor coordinations change. If Nij = 1, a bond is not part of a
conj
conjugated system and the function yields appropriate values, while for Nij ≥ 2
the bond is part of a conjugated system and parameters fitted to conjugated bonds
are used. Finally, to make the potential continuous, 2D and 3D cubic splines are
utilized for the functions Hij and Fij , respectively, for interpolating between values
at discrete numbers of neighbors. The function g(θ ) is very similar to that in the
Tersoff potentials, Eqs. (186) and (194). For example, for carbon one has
 
c02 c02
gC (θ) = a0 1 + 2 − 2 . (209)
d0 d0 + (1 + cos θ )2
It is clear that the Brenner potential has a large number of parameters that
must be fitted to experimental data. The procedure for doing this is to first fit
to systems consisting of only carbons and hydrogen. Parameters are then se-
lected for the mixed hydrocarbon system that produce additive bond energies.
534 9. Atomistic Modeling of Materials

Because the pair terms are first fitted to solid-state carbon structures, the equilib-
rium carbon-carbon distances and the stretching force constants for hydrocarbons
are completely determined by fitting to bond energies. To determine appropriate
energies for hydrocarbons with carbon-carbon bonds, additive bond energies for
single, double, conjugated double and triple carbon-carbon bonds, and carbon-
hydrogen bonds are determined from molecular atomization energies. Values of
the parameter δi of Eq. (203) for carbon and hydrogen turn out to be identical and
equal to 0.80469. Values of the other parameters are listed by Brenner (1990), a list
too long to be given here. The Brenner potential has been utilized successfully in
studying many phenomena, including reaction of hydrocarbon species on diamond
surfaces, mechanical properties of graphite sheet and nanotubes, and CVD of di-
amond films. Moreover, excellent agreement was found (Robertson et al., 1992)
between the predictions of MD simulations of energetics of nanoscale graphitic
tubules using the Brenner potential, and those of first-principle electronic structure
calculations using local-density functional described in Sections 9.1 and 9.4.
Despite its success, the Brenner potential does have certain limitations. For ex-
ample, it does not include the resonance effect of aromatics. Most importantly, the
long-range van der Waals and Coulombic interactions are not included explicitly
in the model, although such interactions play an important role in many materials.
Che et al. (1999) modified the Brenner potential to take such effects into account.
In their formulation, the total energy of the system is written as


 
E= UijV (rij ) + Pij (rij )UijNB (rij ) . (210)
i j >i

Here superscript V denotes the valence short-range terms (for example, those in the
Brenner potential), NB indicates the long-range non-bond part of the energy (for
example, the contribution of van der Waals or Coulombic forces), while Pij = Pj i
is a screening function that properly weights the NB contributions to the total
energy, and is given by
  '  
Pij = f UijV , UijV V
f Uik V
, Ukj , (211)
k
=i,j

with

exp(−x 2 y 2 ), if x < 0 and y < 0
f (x, y) = (212)
1, otherwise.

As pointed out by Che et al. (1999), the screening function eliminates NB inter-
actions between two atoms i and j when they are directly bonded (i.e., UijV < 0)
or when they are both bonded to a common atom k (i.e., Uik V < 0 and U V < 0).
kj
In both cases the screening function Pij is negligibly small, and therefore the NB
interactions do not make improper contribution to the total energy. Using the re-
lation, Fαβ = −∂E/∂rαβ , it is now straightforward to show that the force Fαβ
 9.7. Interatomic Interaction Potentials 535

between atoms α and β is given by

Fαβ = Fαβ
V
+ Pαβ Fαβ
NB
−4 Pij UijVDW (UijV )3 Fij,αβ
V

i>j

  (213)
− V V
Uik Ukj Pij UijVDW Uik
V V
Fkj,αβ + Ukj
V V
Fik,αβ ,
ij k

with
∂UijV NB
∂Uαβ
V
Fij,αβ =− , NB
Fαβ = . (214)
∂rαβ ∂rαβ
V represents the valence forces, such as bond stretching, bending,
In Eq. (213), Fαβ
and torsion, and is contributed by the Brenner-type potential. The second term
represents the NB forces between two properly screened atoms. The third and
fourth terms are due to forces arising from correlations between screened bonds
which, in most cases, are negligible. Therefore, if atoms i and j do not form a
valence bond and do not also form a bond with the same atom k, then usually
V
Fij,αβ = 0, and either UikV = 0 or U V = 0, leading to zero contribution. However,
kj
even if both atoms i and j bind to a common atom k or if they form a valence bond
directly, these terms will still make a negligible contribution to the total energy
because of the exponential screening factor, Eq. (212), since in this case either
V < 0 or U V < 0. Note that there is no restriction on this formulation. That is,
Uiα iβ
the specifics of the NB terms do not alter the general formulation of this potential.
Che et al. (1999) used this generalized interatomic potentials to study the energetics
and structures of a variety of materials, including graphite and molecular crystals
and bucky tubes, and Lim et al. (2003) utilized it for generating porous amorphous
carbon structure and investigating transport, chemisorption and separation of gases
in such materials.

9.7.5 Other Interaction Potentials

Johnson (1964) developed a potential for representing iron which is given by,
U (rij ) = −b1 (rij − b2 )3 + b3 rij − b4 , (215)
where the bi s are parameters of the potential. The interaction force that results
from Eq. (215) decays much faster past its maximum than the LJ potential. One
advantage of the Johnson potential is that, when used in MD simulation of fracture
of a material or metal, it can support an atomically-sharp equilibrium fracture,
leaving it stable up to the critical Griffith load (see Chapters 6 and 7). Thus, one
can make a meaningful comparison between the MD results and predictions of the
continuum mechanics.
Kane (1985) listed several older four-parameter interaction potentials that are
related to the Keating model (see Chapter 8 of Volume I) and are intended for
describing diamond-structure compounds involving C, Si, Ge, and Sn, and also
the zinc-blende-structure compounds GaP, GaAs, and ZnS. These potentials also
appear to provide accurate predictions for various properties of these materials.
536 9. Atomistic Modeling of Materials

Kaxiras and Pandey (1988) constructed a potential in order to specifically sim-

ulate processes in the bulk diamond lattice. The potential was fitted to the entire
energy surface of atomic exchange obtained from an accurate DFT computation.
It correctly predicts the static properties of the perfect diamond lattice and repro-
duces the energy of the concerted exchange path to better than 0.1 eV. However,
the energies of bulk point defects in their unrelaxed configuration appear to be too
low. Because the potential describes very well a large range of local distortions
from the perfect tetrahedral configuration, it can be useful in simulations of mate-
rials such as amorphous structures where the coordination remains predominantly
fourfold.
Development of empirical or semi-empirical potentials for use in MD simula-
tions remains an active research field. It is neither possible nor necessary to list
all the interaction potentials that have been proposed in the past. Several such
empirical potentials, in addition to those discussed here, are discussed by Günes et
al. (2000). In addition, Balamane et al. (1992) and Bazant et al. (1997) compared
the performance of many interatomic potentials.
In addition, we should caution the reader that, despite their success in predicting
many properties of a wide variety of materials, none of the potentials discussed
so far is without problems. For example, many of these potentials do a relatively
poor job of modeling the energetics of small clusters, as well as the various re-
construction of the Si(111) surface. Another example of typical shortcoming of
such potentials is provided by the MD simulation of fracture propagation in sil-
icon (Holland and Marder, 1998). In these simulations the three-body part of
the Stillinger–Weber potential had to be manipulated in order to obtain physical
results, which then of course introduced some unwanted changes in the bulk prop-
erties of the material. The same type of difficulties were observed with the Tersoff
potentials.

9.8 Molecular Dynamics Simulation of Fracture

Propagation
In the last section of this chapter we describe application of the MD simulation to
fracture dynamics of materials. Our goal is threefold.
(1) We would like to describe how dynamic fracture of materials, the practical
consequences of which are at macroscopic length scales, is studied by the MD
method which considers a system at atomic scales.
(2) We aim to demonstrate that, despite the disparity between the length scales
at which MD simulations can be carried out, and the practical length scales
of interest, not only do the simulation results provide deep insight into the
fracture phenomena, but they are in fact in agreement with the experimental
observations described in Chapters 6 and 7.
(3) Finally, the discussions of this section are in fact a prelude to what we will be
describing in Chapter 10 where we consider multiscale modeling of materials.
 9.8. Molecular Dynamics Simulation of Fracture Propagation 537

The fundamental problem facing MD simulation of dynamic fracture (and, more

generally, atomistic description of any phenomenon in materials) is one of length
and time scales. Consider, as an example, fracture of silicon, a material with tremen-
dous technological significance. Its single crystals with a variety of orientations are
inexpensive, thus making numerical predictions amenable to experimental veri-
fication. Silicon is also very brittle, its crystal structure is well-known and, as
discussed in Sections 9.7.2 and 9.7.4, considerable effort has been expended in de-
veloping classical interatomic potentials suitable for use in the MD simulation of
silicon; we have already described many of such potentials. Therefore, silicon pro-
vides an ideal testing ground for MD simulation of dynamic fracture of a material
and direct comparison of its predictions with experimental observations. Suppose,
for example, that we wish to consider a silicon sample with a length and width that
are a few centimeters each and a thickness of only one millimeter. Such a sample
contains of the order of 1022 atoms. The duration of an actual fracture experiment
is around 50 µs, whereas the largest simulations that are currently feasible (see
below) allow one to follow what happens to a sample of about 108 atoms for about
10−9 seconds. Therefore, direct MD simulations of such a sample would require
eighteen orders of magnitude increase in computer power over what is currently
available, a truly daunting, if not impossible, task.
The question therefore is, how does one compare the simulation results with
the experimental observations? An appealing approach for doing this would be to
merge atomistic simulations with continuum modeling. For example, one may use
an atomistic description of the area in the vicinity of the fracture tip, where most if
not all the interesting phenomena take place, and combine that with the continuum
elasticity (see Chapter 7, and also Chapter 7 of Volume I) for describing the material
everywhere else. This approach, which can potentially solve the problem of length
scales, but not that of the time scales, will be described in Chapter 10. We describe
below what has been accomplished so far without using such a multiscale approach.
Given the enormity of the problem of computer simulation of fracture dynam-
ics of materials at the atomic scale, the goal in the MD simulations should not
be performing the largest simulation possible, but constructing the smallest one
capable of providing insightful answers to specific physical questions. In fact, cer-
tain features of brittle fracture may profitably be studied by comparatively small
simulations, involving only thousands or tens of thousands of atoms. Up until
recently, this line of thinking was the dominating factor behind most of the MD
simulations of fracture dynamics. On the other hand, there are other aspects of
dynamic fracture that, if they are to be studied by MD simulations, one must need
a very large number of atoms in the computations.
In any case, before carrying out MD simulations of dynamic fracture of any
material, three important points regarding the limitations of such simulations must
be kept in mind.
(1) Given the severely non-equilibrium nature of dynamic fracture, it is possible
that no classical interatomic potential can provide a realistic description of
the material that is investigated by the MD simulations. Moreover, in view
538 9. Atomistic Modeling of Materials

of the emission of electrons and light that is observed in the vicinity of the
fracture tip (see Chapters 6 and 7), it is entirely possible that even the DFT
would fail as well. Therefore, only a detailed and patient comparison of theory
and experiment, which, as of the time of writing this book, has not yet been
performed, will be able to settle such doubts.
(2) As discussed in Chapters 6 and 7, in steady state, the energy consumed by
the fracture per unit length must equal the energy stored, per unit length to
the left. This statement, which remains true even for strains large enough that
the applicability of linear elasticity could be called into question, relies on
symmetry rather than the matched asymptotics of fracture mechanics, and
must therefore be reproduced by any MD simulation.
(3) The MD simulation must contain a complete description of the cohesive zone.
However, as more energy is fed to the fracture tip, as temperature rises, or if
one studies heterogeneous or ductile materials, the size of the cohesive zone
increases, and therefore the size of the system used in the MD simulation
must increase accordingly. One should not expect MD simulations to pro-
vide “easy” (i.e., without much effort, patience, and efficient computational
strategy) predictions for materials where the cohesive zone is of the order
of microns, let alone millimeters, as in the silicon sample described above.
In this regard, MD simulation of fracture of amorphous polymers on which
many experiments have focused (see Chapters 6 and 7) provide a particularly
great challenge.
In what follows we first discuss the early MD simulations that involved only
a few hundred or thousand of atoms. This will give the reader an idea about the
long road that has been travelled in order to arrive at the present state-of-the-
art of the MD simulation of dynamic fracture. We then describe and discuss the
recent advances and compare the results of large-scale MD simulation of dynamic
fracture with the relevant experimental data.

9.8.1 Early Simulations

The idea that some sort of a thermodynamic approach (which can be related to
MD simulations) may be used for investigating fracture of solids was probably
first hinted in a paper by Max Born (1939), who was interested in developing a
first-principle criterion for melting. He made the observation that, “the difference
between a solid and a liquid is that the solid has elastic resistance against shearing
stress, while the liquid has not.” He developed a thermodynamic approach to this
problem, and also proposed that a generalization of his approach which includes
anisotropic stress should be capable of accounting for breaking of crystals. Over
40 years later Born’s suggestion was taken up by Nishioka et al. (1980,1981) who
developed a variational formulation for a solid under uniaxial stress, and wrote
down the free energy of the system as a function of the lattice constants parallel and
perpendicular to the loading direction. By fitting a Gaussian-type pair potential to
the zero-temperature Young’s modulus, they calculated the fracture strength as the
maximum tension allowing positive free energy curvature.
 9.8. Molecular Dynamics Simulation of Fracture Propagation 539

The statistical thermodynamic approach to fracture of solids was further devel-

oped in a series of papers by Blumberg Selinger and co-workers. In an interesting
paper, Blumberg Selinger et al. (1991a; see also Englman and Jaeger, 1990) de-
veloped such an approach in which fracture at failure threshold corresponds to a
metastability limit, or spinodal. In their formulation, the role of non-equilibrium
defects, such as macroscopic fractures, dislocations, and impurities, in lowering
the fracture strength of the material is similar to that of dust particles in lowering
the nucleation barrier. Rundle and Klein (1989) developed a similar theory using
a field-theoretic approach but, typical of such theories, theirs was a coarse-grained
theory without any reference to the structural details of the materials. Their theory
was tested by MD simulation of an ideal solid by Wang et al. (1991), who showed
that the solid remains in metastable equilibrium all the way to the critical stress
or force for its fracture, at which point it fails irreversibly by nucleation of small
defects. Building on their formulation, Blumberg Selinger et al. (1991b) proposed
that the onset of fracture in a defect-free material is associated with the loss of a
metastable minimum in its free energy at the critical stress.
We must, however, point out that the statistical thermodynamic approach to
fracture of a solid is useful only when thermal fluctuations play the dominant
role in its failure. This will be the case if the solid is perfectly periodic (without
any defect) or it contains very little disorder. As soon as the heterogeneity of a
material is even “mild,” it begins to play the dominant role in its fracture and
failure, and therefore thermal fluctuations will no longer be important. As we
have emphasized throughout this book, most real materials are at least to some
extent heterogeneous, and therefore an approach to their fracture based solely on
statistical thermodynamics is inadequate.
The first time that the word “atomistic” was used in the study of cracks was,
to our knowledge, in a paper by Sinclair and Lawn (1972). The first “molecular”
simulation of fracture dynamics was probably carried out by Weiner and Pear
(1975) who used a square lattice of atoms, inserted a crack in its middle, and
solved the equation of motion for the atoms. They assumed that if the distance
between two atoms becomes too large, they can be considered as disconnected, an
assumption similar to what has been used in the quasi-static and dynamic lattice
models of fracture described in Chapter 8, which is also typically used in the MD
simulation of dynamic fracture as well. Simulations were performed both at zero
and non-zero temperatures. Weiner and Pear found that, except at very high applied
stresses, the velocity of the crack reaches a steady subsonic and stress-dependent
value, which is in agreement with the prediction of continuum fracture mechanics
described in Chapter 7.
The first MD-like simulation of fracture was probably carried out by Ashurst
and Hoover (1976). They used a triangular lattice in which the atoms interacted
with each other by a truncated Hooke’s-law force. The most important finding
of this study was that, the velocity of the crack never reaches the Rayleigh wave
speed cR , consistent with the troubles that linear continuum mechanics of fracture
dynamics already had for explaining the experimental data on the speed of fracture
propagation (see Chapter 7).
540 9. Atomistic Modeling of Materials

The first attempt for comparing the predictions of the continuum fracture me-
chanics with the results of MD simulations was made at the beginning of the 1980s.
Thomson et al. (1971) had presented evidence for lattice trapping, a phenomenon
in which a crack neither propagates nor heals, rather it remains stable until exter-
nal loads somewhat larger than the Griffith threshold (see Chapters 6 and 7) are
imposed on the system. The magnitude of the trapping range depends strongly on
the characteristics of atomic bonding of materials. Lattice trapping may also de-
pend on the direction in which the crack tip bonds are broken, and may therefore
be different for fracture propagation along different crystallographic directions.
Moreover, as described in Section 7.12, Rice and Thomson (1974) had developed
a criterion for the degree of brittleness of a material according to which a material
can be considered as brittle if a dislocation in the neighborhood of the fracture tip
cannot escape from the tip region. These predictions were put to test in the first
truly MD simulations that were carried out by Paskin et al. (1980,1981). We call
their computations “true” MD simulations because, unlike Ashurst and Hoover
(1976), they utilized the LJ potential for representing the interactions between the
atoms in a triangular lattice. In their simulation, a crack was inserted in the middle
of the lattice in order to initiate fracture propagation. The smallest crack in a MD
simulation is represented by a pair of atoms the bond between which has been cut,
so that the two atoms do not directly interact with each other. An external force
was then applied to the lattice, and Newton’s equations of motion were solved
in order to calculate the atomic positions, velocities, and forces. The cutoff rc
for the LJ potential was assumed to be slightly smaller than two lattice bonds at
equilibrium. Paskin et al. showed that the Griffith energy criterion (see Chapters
6 and 7) is incorrect for large cracks. Their MD simulations also indicated that
lattice trapping is a negligible effect, which they attributed to the long range of
the interaction potentials. However, in general one should expect lattice trapping
to disappear at temperatures much lower than room temperature, and therefore,
in order to observe this phenomenon, experiments and MD simulations must be
carried out at very low temperatures. The necessity of a low temperature explains
why no lattice trapping has yet been observed experimentally in either crystalline
or amorphous materials. Paskin et al.’s simulations also indicated that the Rice–
Thomson criterion for brittleness is valid at low temperatures (see below for more
discussion of the Rice–Thomson theory).
In addition to Paskin et al.’s simulations, interesting MD computations were
carried out by Soules and Busbey (1983) to study fracture of sodium silicate fiber
glass. Instead of interatomic forces that result from the LJ potentials, these authors
used the following semi-empirical equation for computing the interatomic forces,
  $ %4 
e 2 Zi Z j r
Fij (r) = 3.448aij exp(−3.448r) + 1− , (216)
r2 rc

where aij is a parameter of the model in erg, Zi is the charge of atom i, e is the
electron charge, and rc is the cutoff distance at which the interatomic forces vanish.
The exponential term of this equation represents a repulsive force, while the 1/r 2
 9.8. Molecular Dynamics Simulation of Fracture Propagation 541

term obviously represents the Coulombic contribution. Simulations of Soules and

Busbey indicated that the glass breaks when it is suddenly subjected to a large
biaxial expansion. Moreover, when the temperature of the system was raised by
about one order of magnitude, the strength of the material decreased by a factor
of about 2, a result that was claimed to be in agreement with experimental data.
To our knowledge, Ray and Chakrabarti (1985a,b) and Chakrabarti et al. (1986)
were the first to carry out MD simulations of fracture of a model of materials
with quenched disorder. In their model the disorder was percolation-type, i.e., the
heterogeneity was generated by randomly breaking some of the bonds between
the atoms before the simulations were commenced. The atoms interacted with
each other through a LJ potential for which the cutoff distance rc was set to 1.6
lattice bonds. During deformation of the lattice a bond was considered broken
if the distance between its end atoms was larger than rc . Chakrabarti and co-
workers found that the stress needed for fracture vanishes, and the time to complete
fracture diverges, both at the bond percolation threshold of the lattice, pcB  0.347,
whereas the elastic moduli of the lattice vanish at the rigidity percolation threshold,
B  0.65 (see Chapter 8 of Volume I). The latter result was of course an artifact
pce
of the model, because the lattice used was effectively a central-force model and, as
discussed in Chapter 8, the percolation threshold of central-force lattices is much
larger than the connectivity threshold. Other issues of interest in dynamic fracture
using small-scale MD simulations were investigated by several research groups,
whose results up to 1987 was reviewed by Dienes and Paskin (1987).
Sieradzki et al. (1988) utilized MD simulations to study dynamics of crack ex-
tension, using a triangular lattice of atoms which interacted with each other through
the Johnson potential, Eq. (215). Their simulations indicated that the terminal frac-
ture velocity was about 1/4 of the Rayleigh wave speed cR , and that this terminal
velocity depended on the configuration of the fracture tip. This was probably the
first time that, in addition to the ample experimental evidence, reasonably-accurate
MD simulations had also indicated the breakdown of the linear continuum fracture
mechanics. Hoagland et al. (1990) employed MD simulations to investigate the
configuration of the fracture tip in aluminum, using an embedded-atom potential
described in Section 9.7.1. The pair potential used was a Morse function given by
Eq. (175), and the embedding energy was determined numerically. As described in
Section 9.7.1, the advantage of an embedded-atom potential is that it intrinsically
incorporates many-body contributions, and thus dynamic fracture may be studied
in a more realistic manner. Hoagland et al.’s MD simulations provided evidence
for several interesting phenomena, including the existence of two singular fields
for an atomically sharp fracture. One was an outer field with a strength that was
equal to the applied Griffith threshold with its origin at the fracture tip, while the
other was behind the fracture with a strength less than the threshold. This was
attributed to the nonlinearity arising out of elastic softening of the material near
the tip.
Cheung and Yip (1990), employing the embedded-atom potentials, studied the
response of a crystal containing a sharp fracture to varying stress and temperature.
Over a limited range of temperature, a transition from brittle to ductile fracture
542 9. Atomistic Modeling of Materials

was observed, caused by dislocation emission from the tip of the fracture. This
result indicated the existence of an energy barrier for nucleation of the dislocation.
Cheung and Yip (1990) showed by detailed calculations that this energy barrier
could not be predicted by the continuum theory of Rice and Thomson (1974)
mentioned above. This issue was also investigated by Zhou et al. (1994) using
MD simulations, who proposed that the Rice–Thomson theory should be modified
to include the effect of tensile broken-bonds, if it is to correctly predict dislocation
emission.

9.8.2 Large Size and Scalable Molecular Dynamics Simulation of

Fracture
As discussed above and also in Chapter 7, in order to investigate certain issues in
dynamic fracture by MD simulations, using a large number of atoms is important,
because fracture phenomena are sensitive to the sample size, and thus one needs
significant computational resources and efficient computational strategies in order
to study the size effect in the MD simulations. To this end, to our knowledge, the
first MD simulation of dynamic fracture using a massively-parallel computational
strategy was carried out by Wagner et al. (1992) who utilized 106 particles, by far
larger than all the previous MD simulations of dynamic fracture. They used a LJ
potential with a spline cutoff together with an analytic embedded-atom potential
[see Section 9.7.1 and Eqs. (33)–(36)], and investigated the phenomenon of spal-
lation which occurs at very high strain rates. They demonstrated that an adiabatic
expansion can cause spallation, and that the spall strength is proportional to the
logarithm of the applied strain rate. However, the LJ material exhibited brittle
fracture, whereas the embedded-atom material produced ductile-like fracture, so
that the effect of the potentials used in MD simulations of fracture is non-trivial.
Over the past few years, a number of large-scale MD simulations of dynamic
fracture have been performed (Abraham et al., 1994; Holian and Ravelo, 1995;
Zhou et al., 1996; Omeltchenko et al., 1997; Gumbsch et al., 1997). One of the
largest of such simulations was carried out by Zhou et al. (1997) who used 35
million atoms to study fracture of a 3D solid. They studied ductile failure and
observed dislocation loops emitting from the tip of the fracture. One important
result of this study was that, the sequence of dislocation emission events depends
strongly on the crystallographic orientation of the fracture front, a result that the
previous theories had not predicted. Other MD simulations of dynamic fracture
were performed by Zhang and Wang (1996), Machová (1996), Español et al.
(1996), and Hua et al. (1997), investigating various aspects of the problem, such
as the effect of boundary conditions, and use of an N -body potential, instead of
the usual two- or three-body potentials.
The largest MD simulations of dynamic fracture that we are aware of were car-
ried out by Abraham et al. (1997b) who studied the response of a 3D notched solid
under tension using more than 108 atoms. In their simulations, the interatomic
interactions were modeled by the LJ potential, a parallel computational strategy
 9.8. Molecular Dynamics Simulation of Fracture Propagation 543

Figure 9.11. Notched solid, a FCC

crystal, used in the simulation of frac-
ture. Due to the boundary conditions,
the front and back faces are not exte-
rior surfaces, and are thus transparent
(after Abraham et al., 1997b; courtesy
of Dr. Farid F. Abraham).

based on a spatial-decomposition algorithm described in Section 9.6.2.3 was uti-

lized, and the simulations were carried out at an initial temperature of zero. As
such, their system represented a rare-gas solid. The system studied was a slab of
atoms with Lx = Ly = 336 and Lz = 896 atoms for the three orthogonal direc-
tions. The notch was a slit beginning midway along Lx for y = 0, having a y
extension with a length of 120 atoms, extending through the entire thickness Lz .
The exposed notch faces were in the y − z planes with (110) faces, with the notch
pointing in the 11̄0 direction. Note that the (110) face does not have the lowest
surface energy, thus contradicting the conventional wisdom that would identify
the lowest energy surface, i.e., the (111) surface, as the cleavage plane for brittle
fracture. The choice of the (110) surface was, however, based on the MD result of
Abraham (1996) that indicated that the nonlinear elastic anisotropy of the crystal,
and not the anisotropy of surface energy, controls the cleavage behavior. Figure
9.11 shows the system simulated.
Abraham et al. (1997b) found that when the speed of fracture propagation ap-
proaches one third of the Rayleigh wave speed, the crack tip begins to roughen
on atomic scales, followed immediately by a dynamic ductile-to-brittle transition
where plasticity becomes dominant through prolific emission of loop dislocations
and the arrest of the crack motion. Figure 9.12 presents magnified off-diagonal
views of the cohesive zone during the time period of this transition (times 36, 43,
and 54 in the figure from second panel on the left to the right). The atomic rough-
ening is consistent with the onset of the brittle fracture instability suggested by the
experiments discussed in Chapters 6 and 7, and hence supporting the notion that
this instability is a general feature of rapid fracture of brittle materials. However,
although Kelly et al. (1967) and Rice and Thomson (1974) had proposed that rare-
gas solids are inherently ductile, Abraham et al. (1997b) found that their system
undergoes brittle fractures along the (110) plane, but fails by ductile plasticity for
a notch with (111) or (100) faces. Thus, not only their MD simulations indicated
partial break down of the Rice–Thomson and Kelly et al. theory, but also provided
new insights into fracture behavior of materials.
544 9. Atomistic Modeling of Materials

Figure 9.12. Off-diagonal views of the cohesive zone during the time of the brittle-to-
ductile transition. Darker and lighter areas indicate, respectively, whether the speed of the
atoms is less than or greater than one twelfth of the longitudinal sound speed (after Abraham
et al., 1997b; courtesy of Dr. Farid F. Abraham).

9.8.3 Comparison with the Experimental Observations

With the enormous increase in the computational power, MD simulations of dy-
namic fracture have reached that degree of accuracy that their predictions can be
directly compared with experimental observations and data. Three important as-
pects of dynamic fracture that have been reproduced by MD simulations are as
follows.

9.8.3.1 Fracture Instabilities

One of the first comparisons between the results of MD simulation of dynamic
fracture and the experimental observations was taken by Abraham et al. (1994)
in which the dynamics of a fracture was studied using a 106 atom crystal. The
interaction potential between the atoms was a LJ potential. Stress was applied to the
system by displacing its opposing boundaries at a strain rate that was approximately
equivalent to that obtained in explosive loading applied to fracture faces. These
conditions were necessary to achieve sufficient acceleration of the fracture tip so as
to achieve high enough crack velocities over the duration of the simulation, and to
be able to detect the existence of instabilities in the fracture’s motion. The results of
these MD simulations were close to the experimental observations in amorphous
materials. The fracture was observed to accelerate smoothly until it reached a
velocity of 0.32cR . The velocity of the fracture experienced large fluctuations
when its value exceeded 0.32cR , as a result of which the instantaneous velocity of
the fracture tip became erratic. These fluctuations were coupled with a ”zig-zag”
motion of the fracture tip, which formed in its wake a rough fracture surface. These
interesting simulations, which are in complete agreement with the experimental
 9.8. Molecular Dynamics Simulation of Fracture Propagation 545

observations described in Chapters 6 and 7, highlighted the robust and general

nature of the crack instability.
The robust nature of the crack instability was further highlighted by the work
of Zhou et al. (1996; see also Gumbsch et al., 1996) in which fracture propagation
was investigated in a 400, 000 atom crystal. The atoms interacted with each other
via a Morse potential, Eq. (175). By varying the applied strain rates the maximum
velocity that a fracture could achieve over the duration of the simulation was
varied between 0.18cR and 0.36cR , where the strain rates used corresponded, as
in the work of Abraham et al., to explosive loading of the system. At a velocity
of 0.36cR , instability of the crack was observed to occur by its forming several
branches. The branching process was observed to be immediately preceded by the
nucleation of a dislocation in the crystal together with a build up of the phonon
field in the vicinity of the crack tip. These features are again in agreement with
the experimental observations described in Chapter 7. In both of the above large-
scale MD simulations, the entire fracture process (from initiation to the onset
of the instability) occurred over a time of about 1 nanosecond. For this reason,
both the strain rates used and the amount of strain in the material at the onset
of fracture (approximately an order of magnitude larger than observed in real
materials) had to be extremely large. Thus, the close correspondence of the results
of the MD simulations with the laboratory results obtained in amorphous materials
is rather surprising. The short time scales used in these simulations preclude, of
course, examination of steady-state properties of the system. It should therefore be
interesting to compare the results obtained in these experiments with steady-state
results that can be obtained by smaller MD simulations.
Further insight into the limiting velocity of the crack tip was provided by the MD
simulations of Abraham (1996), whose simulations indicated that this velocity can
reach 60% of the Rayleigh wave speed by following the highest energy path. Sub-
sequent MD simulations of dynamic fracture in graphite sheets by Omeltchenko
et al. (1997) confirmed Abraham’s results. They found that, for certain crystalline
orientations, multiple fracture branches with nearly equal spacing are created as
the velocity of the fracture tip reaches 0.6cR .
Abraham et al. (1997a) carried out MD simulations of dynamic fracture of a 2D
material represented by a triangular lattice with more than 2 million atoms. The
interatomic forces were treated as central forces, modeled by a LJ spline poten-
tial [see Eqs. (33)–(36)] and the analytic embedded-atom model that represents a
many-body potential, developed by Holian et al. (1991) and described in Section
9.7.1.As discussed by Holian et al. (1991), the LJ potential can accurately represent
brittle materials, while the embedded-atom model can be used for studying ductile
solids. Under these conditions, the reduced melting temperature is kB T /ε = 0.2
for the EAM and kB T /ε = 0.4 for the LJ material. Moreover,
√ the longitudinal
wave speed √ cl [see Eqs. (7.24) and (7.25)] is about 5 ε/m for the EAM mate-
rial and 9 ε/m √ for the LJ material, where m is the atomic mass. The transverse
speed is ct = cl / 3, and the Rayleigh wave speed cR  ct . The simulations were
performed at a reduced temperature of 10−5 . Abraham et al. (1997) found that,
for rapid fracture of brittle (LJ) material, a dynamic instability of fracture growth
546 9. Atomistic Modeling of Materials

Figure 9.13. The onset of fracture instability (left), in reduced time intervals of 7 and
beginning at 85. The right panels show the fracture zipzag, beginning at reduced time 220
(after Abraham et al., 1997a; courtesy of Dr. Farid F. Abraham).

develops when the crack velocity approaches 1/3 of the Rayleigh wave speed cR .
At higher crack velocities, the fracture either follows a wavy path or branches out,
with the anisotropy that is due to the large deformation at the tip of the propagating
fracture playing the dominant role in determining the path of the fracture. This is of
course in contrast with the conventional wisdom (see Chapter 7) which associates
the lowest energy surface as the favored cleavage direction. Figure 9.13 nicely
demonstrates these results. The simulations also produced dislocations emission
from the rapidly moving fracture tip after the onset of the crack growth rough-
ening, implying that the dislocations are the consequence rather than the cause
of the dynamic instability. The number of the dislocations emitted was dependent
upon the external loading. These results are all in excellent agreement with the
experimental observations described in Chapters 6 and 7.
Let us point out that fractures at zero temperature display a clear dynamic
instability at a critical energy flux. Up to this point, phonons are able to carry
away all excess energy. This instability does not take the form of a simple micro-
branching instability, partly because atomic bonds can easily rejoin above the main
crack line in a single component solid with no environmental impurities available.
Instabilities at room temperature have not yet been explored either numerically or
analytically.

9.8.3.2 Morphology of Fracture Surface

The morphology of fracture surfaces has also been studied by MD simulations.
For example, Nakano et al. (1994) performed MD simulations of fracture surfaces
in porous silica glasses. The surface of the pores was rough, and thus particular
attention was paid to root-mean square fluctuations in the height h of surface of the
pores, averaged over a length scale . The MD results indicated that h scales with
 as in Eq. (7.130) where, as discussed in Chapters 1, 6 and 7 α is the roughness
exponent. The simulations of Nakano et al. (1994) yielded α  0.87, in agreement
with the prediction of theory of rough surfaces and the experimental data for many
fracture surfaces during rapid fracture (see Chapters 6 and 7), thus supporting the
claim that in this regime α is universal. Subsequently, Nakano et al. (1995) (see
also Tsuruta et al., 1996; Omeltchenko et al., 1996) studied, using MD simulation,
fracture dynamics in amorphous Si3 N4 films. They showed that the surface rough-
 9.8. Molecular Dynamics Simulation of Fracture Propagation 547

ness exponent α depends on the speed of fracture propagation. At the initial stages
of fracture, when the crack tip propagated slowly, α  0.44. However, once the
speed of fracture propagation exceeded a certain limit, a crossover was observed
to a higher value, α  0.8. These results are in agreement with the experiments of
Bouchaud and Navéos (1995) which were described in Chapters 6 and 7. In another
effort by this group, Li et al. (1996) studied dynamic fracture in SiSe2 nanowires,
and found that fracture is initiated in an amorphous region of the surface of the
material, while multiple fractures start at the boundaries of the amorphous region.
Finally, Kalia et al. (1997) investigated dynamic fracture in nanophase Si3 N4 ,
showing that intercluster regions of the material are amorphous, deflecting fracture
and hence giving rise to local crack branching. This implies that nanophase Si3 N4
can resist fracture much better than crystalline Si3 N4 . The roughness exponent
was found to be, α  0.84, in agreement with the experimental data.

9.8.3.3 Fracture Propagation Faster Than the Rayleigh Wave Speed

As discussed in Chapter 7, it is generally believed that a brittle crack cannot prop-
agate faster than the Rayleigh wave speed cR . Continuum mechanics predicts that
for Mode I tensile loading the forbidden velocity zone for fracture propagation is
any speed larger than cR , while for Mode II shear loading the forbidden zone exists
only for speeds between cR and the shear wave speed. However, the limiting speed
of a Mode II fracture is also cR because its forbidden zone acts as an impenetrable
barrier for the shear fracture and does not permit it to propagate faster than cR .
On the other hand, experimental evidence that was discussed in Section 7.8.15
indicates that, under certain circumstances, a Mode II shear fracture can propagate
faster than the shear wave speed.
Abraham and Gao (2000) carried out MD simulation of fracture in a 2D rect-
angular slab with Lx = 1424 and Ly = 712 atoms, thus utilizing over 106 atoms.
The system used consisted of two crystals joined by a weak interface. The inter-
atomic forces were assumed to be harmonic, excepts for those pairs of atoms with
a separation cutting the horizontal center line of the simulation slab, for which
a LJ potential was utilized. The cutoff distance was taken to be 2.5σ , where σ
is the size parameter of the LJ potential. As mentioned above, at zero tempera-
ture and pressure, the longitudinal wave
√ speed cl , in reduced units, is 9, the shear
(transverse) wave speed is ct = cl / 3  5.2, and the Rayleigh wave speed cR is
4.83, where the LJ parameters σ √ and ε are used as the basic units of length and
energy (thus, for example, cl = 9 ε/m in dimensional units). A horizontal crack
of 200 atom distance was cut midway along the left-hand vertical slab boundary.
The 2D crystal used was a triangular lattice with the initial crack being parallel
to the close packed direction. The initial temperature of the system was zero, and
the MD simulations were carried out at constant energy (see Section 9.2.4). Both
Modes I and II were simulated.
Molecular dynamics simulations of Abraham and Gao (2000) showed that, in
Mode I (cracks under tensile loading), the crack quickly approaches a constant
velocity of about 4.83, the same as the Rayleigh wave speed, after which it prop-
agated with the same constant speed. In contrast, in Mode II (cracks under shear
548 9. Atomistic Modeling of Materials

loading) the crack tip first reaches the Rayleigh wave speed cR and for some time
propagates with this speed, and then jumps to a higher constant speed with a value
of about 8.97, essentially the same as the longitudinal wave speed cl . Similar re-
sults were obtained (Gumbsch and Gao, 1999) by MD simulations of propagation
of dislocation of indentation. These results are in complete agreement with the
experimental observations described in Section 7.8.15.
Let us point out that, while effective potentials, such as the embedded-atom and
LJ potentials (with fitted parameters), may be adequate for representing metals,
they are poor representatives of non-metallic materials. In this case, one must use
a first-principle quantum mechanical description of the materials in order to calcu-
late the potentials (see Sections 9.4 and 9.5). This is a rigorous method since, unlike
the case of MD simulations with empirical or semi-empirical potentials, ab initio
quantum mechanical computations do not use any adjustable parameters. More-
over, as discussed in Sections 9.2 and 9.4, the local density approximation in its
plane-wave pseudo-potential formulation can be optimized, so that the computa-
tions will be highly efficient. Kaxiras and Duesbery (1993) presented the results of
such a study for silicon, and Spence et al. (1993) used ab initio QMD simulations to
investigate the dependence of lattice trapping energies on applied load for fractures
propagating in silicon. Pérez and Gumbsch (2000) studied the anisotropy of cleav-
age fracture in silicon and the effect of sample size on the process. This approach
has not, however, been utilized extensively, presumably because its computational
cost is much larger than MD simulations using the empirical or semi-empirical
potentials discussed above.

Summary
Ab initio quantum mechanical computations based on the density functional the-
ory in the local density approximation, together with plane-wave pseudo-potential
formulation, offer an efficient and rigorous method for computing materials prop-
erties. Quantum MD simulation method of Car and Parrinello did not change the
essentials of such computations, but offered an enormous increase in the efficiency
of the method, hence making much larger pieces of materials accessible to such
computations.
As a technique for studying materials at the atomic scale, molecular dynamics
simulation has been used for several decades. However, development of vector
computers and parallel machines, and hence vectorized and parallelized compu-
tational algorithms, together with derivation of accurate interatomic potentials,
have made MD simulations a powerful tool for studying materials at atomic scales
utilizing millions of atoms and molecules.
These two computational strategies have enabled us to investigate and accurately
predict various properties of materials. We believe that when one or both of these
methods are joined with the multiscale approach that will be described in the next
chapter, the possibilities for accurate and efficient optimal design of materials with
specific properties may be limitless.
10
Multiscale Modeling of Materials:
Joining Atomistic Models with
Continuum Mechanics

10.0 Introduction
Throughout this book we have emphasized that solid materials of industrial impor-
tance are highly heterogeneous, with the heterogeneities manifesting themselves
at several length scales, ranging from the smallest to macroscopic scales. For many
centuries, such materials were discovered, mined, and processed in a serendipi-
tous manner. However, characterization of atoms and the progress made in x-ray
diffraction during the early decades of the twentieth century provided the impetus
for the search for a theory of materials that can explain their properties in terms of
their atomic constituents. It was soon realized that developing such a theory was
not practical yet, because
(1) the disparity between the relevant length scales, i.e., those at the atomic and
macroscopic scales is huge;
(2) the existing knowledge of the principles of atomic cohesion and the basic
properties of materials was totally inadequate, and
(3) the required computational power for solving the problem was (and still
is) enormous, while the available computational power was grossly limited.
Therefore, the consensus at the time was that any theory of materials could
not have predictive power.
Later decades of the twentieth century witnessed development of many quali-
tative and semi-quantitative models of materials that could explain the principles
of atomic cohesion, and basic properties of such fundamental materials as met-
als and semiconductors. Some of these models, though relatively simple, were
surprisingly accurate and helped us make remarkable progress in understand-
ing materials’ properties. However, as discussed in Chapter 9, for most materials
of industrial importance, the interatomic interactions are complex enough to re-
quire very elaborate models. Such models must usually be accompanied by very
extensive computations.
In Volume I, as well as in the present Volume, we utilized continuum mechan-
ics and lattice models to describe modeling and simulation of the morphology,
550 10. Multiscale Modeling of Materials

and estimating the effective transport and mechanical properties of heterogeneous

materials. These models and approaches are appropriate for describing a material
at the microscopic and macroscopic length scales, but cannot provide any insight
into its properties at the smallest length scales, namely, the molecular scales. In
Chapter 9 we described modeling and simulation of materials and their properties
at the atomic and molecular scales. As discussed there, methods for computing the
properties of materials at such length scales are divided into two major classes:
Those that do not use any empirically- or experimentally-derived quantities, and
those that do. The former are the ab initio (or first-principles) methods, while the
latter are referred to as empirical or semi-empirical techniques. Several important
theoretical and computational tools of both classes, developed over the past three
decades, such as the density functional theory (DFT) and its variants, the classical
molecular dynamics (MD) simulation, the quantum MD (QMD) technique, and
the tight-binding (TB) methods, were all described in Chapter 9. The advent of
very fast computers and development of efficient strategies for massively-parallel
computations, that were also described in Chapter 9, have made it feasible to carry
out large-scale computations at the atomic and molecular length scales, making
such techniques indispensable tools for investigating and predicting materials’
properties on such length scales.
However, we should not “abuse” atomic-scale simulations in the study of
macroscopic properties of materials, since atomic-scale mechanisms are in gen-
eral separated from the macroscopic behavior that they engender by a vast array
of intervening continuum scales. These mesoscopic length scales both filter (i.e.,
average) and modulate (i.e., set the boundary conditions or driving forces for)
atomic-scale phenomena, and are therefore an essential part of the constitution of
materials.
As our discussions throughout Volume I and the present Volume should have
made it clear by now, continuum models are based on the assumption that the
relevant fields that describe the state of a material vary slowly on the atomic scale;
otherwise, continuum models that represent macroscopic behavior of a material,
and are derived by averaging the material’s properties at the smaller length scales,
lose their meaning. Therefore, if we are, for example, to describe properties of
a material with defects, continuum theories a fortiori break down in the vicinity
of the defects, or, more generally, any other entity that possesses structure on the
atomic scales. It is therefore clear that continuum theories can be “enriched” by
incorporation of additional information, and hence avert their breakdown. Thus,
atomistic and continuum models need and reinforce each other.
Given great advances in both continuum formulation and atomistic simulation
of the behavior of materials, we have reached the stage in which we may be able
to integrate models for describing materials at the atomic and molecular scales
with those at the continuum level, thus developing a methodology for making
quantitative predictions for materials’ properties that utilizes our knowledge at all
the relevant length scales. To better motivate this discussion, consider an important
phenomenon, namely, fracture of silicon, that was already described in Chapter
9. Suppose that one is to carry out a fracture experiment in samples of silicon
 10.0. Introduction 551

with lengths and widths that are several centimeters each and a thickness of about
a millimeter. Such samples would contain about 1022 atoms. The duration of the
experiment is about 50 µs whereas, as discussed in Chapter 9, the largest atomistic
simulations that are currently feasible can follow 108 atoms for around 10−9 s.
Direct atomistic simulation of fracture of such a sample of silicon will therefore
require more than eighteen orders of magnitude increase in computational power
over what we currently have. Such a breakthrough in computational power will
not be achieved any time soon, and therefore the critical question is: How can
one make a comparison between the results of the atomistic computations and the
relevant experimental data, or even the predictions of the continuum theories?
The essential problem facing such simulations is one of length and time scales.
However, before answering the above question, we should first remind ourselves
that the goal in computer simulations, both at the continuum and atomistic scales,
should not be performing the largest simulation utilizing the largest possible sys-
tem, but constructing the smallest one that is capable of answering specific physical
questions. In the example of brittle fracture of silicon, as well as in many other
phenomena that occur in materials, many important features may profitably be
studied by atomistic simulations that are comparatively small, involving only tens
of thousands of atoms. Therefore, a fruitful strategy may be based on merging
atomistic and continuum simulations. Hence, in the example of fracture of silicon,
one may describe the vicinity of the crack tip by atoms, but utilize the continuum
elasticity everywhere else.
The combined methodology is called a multiscale approach to modeling of
materials. In political jargon, multiscale modeling of materials is a divide-and-
conquer modeling paradigm. As the first step, the entire range of material behavior
is divided into a hierarchy of length scales. Next, the relevant physical processes
that are irreducible and operate independently at a given scale—sometimes referred
to as unit processes—are identified. The unit processes at one scale represent
averages of unit processes operating at the immediately lower length scale, and
this relation defines a partial ordering of the processes.
Over the past decade, multiscale modeling of materials has evolved from some-
thing that was thought of as a distant dream to a research area with intensive
methodological development. If its development continues at the current rapid
pace, it will, in the relatively near future, develop into a practical tool for indus-
trial applications involving design of materials that possess specific properties.
The goal of this chapter is to describe the basic ideas and techniques for multiscale
modeling of materials. Because the range of problems and phenomena to which
multiscale methods are applicable is very broad, we restrict the discussions to
the main theme of this book, namely, modeling of materials and predicting their
properties, and describe the essentials of the multiscale modeling approach by dis-
cussing a few recent applications. We begin our discussion by briefly describing
two main classes of multiscale modeling approach developed so far, after which
we describe some of their recent applications. More extensive discussions and
overviews pertaining to micromechanics and multiscale modeling of materials are
given by, for example, Ortiz and Phillips (1999) and Phillips (2001).
552 10. Multiscale Modeling of Materials

10.1 Multiscale Modeling

We should first point out that the multiscale paradigm is more easily stated than
carried out in practice. Currently (at the time of writing this book), the analysis of
mechanisms and the characterization of the effective properties of materials rely
on either extensive numerical computations, or an assortment of analytical tools,
ranging from mean-field and effective-medium approximations to variational ap-
proaches that have been described throughout this book. Because of this difficulty,
multiscale modeling in general, and combined atomistic-continuum modeling in
particular, cannot be easily reduced to a self-contained and unified theory. There-
fore, at this stage multiscale modeling is an art as well as science. Keeping in mind
this fact, we now begin to describe the main concepts and ideas of the multiscale
approach.
Generally speaking, there are two types of multiscale approach to modeling
of various physical phenomena that occur in materials and systems that contain
several disparate length as well as time scales. These approaches are either se-
quential or parallel methods, in the sense described below. What follows is a brief
discussion of each approach.

10.1.1 Sequential Multiscale Approach: Atomistically-Informed

Continuum Models
In this method, which has been developed and used much more extensively than the
parallel multiscale modeling (to be described below), beginning with the smallest
length/time scales of the problem, the results of one series of computations are
used as the input to the next (larger) up the length and/or time scale hierarchy.
Hence, the essential idea is to pass (as input) information (output) from finer to
coarser scales.
A good example is provided by polycrystalline plasticity, for which the main
identifiable length scales are,
(1) the nanoscale in which unit processes represent the possible behaviors of
single-crystal defects, such as individual dislocations or vacancies;
(2) the mesoscale which is characterized by the collective behavior of large
numbers of defects, as in, for example, dislocation dynamics;
(3) the subgrain scale, characterized by the formation and evolution of subgrain
dislocation structures, and
(4) the polycrystalline scale which is characterized by the collective behavior of
a large number of grains.
The atomistic scale and the continua communicate at the nanoscale through
ab initio computations of material parameters pertaining to continuum theories.
Under this scenario, the mesoscopic model sets the functional form of the response
functions of a material, while the atomistic models dictate the relevant material-
specific parameters of the mesoscopic theory. In effect, one works with continuum
models, except that the relevant parameters of the models are provided by atomistic
computations, hence the name atomistically-informed continuum models.
 10.1. Multiscale Modeling 553

A beautiful example of implementation of this strategy is the pioneering work of

Clementi (1988). He used accurate quantum-mechanical computations, of the type
that were described in Chapter 9, in order to evaluate the interaction of several water
molecules, from which he developed an accurate empirical interatomic potential
that involved two-, three-, and four-body interactions. The potential was then
utilized in a MD simulation for evaluating the viscosity of water. Subsequently,
motion of water in a channel with or without obstacles was studied, using as
input the viscosity that had been computed in the previous step. The resulting
understanding was then employed in a fluid dynamics computation for predicting
tidal circulations (which is somewhat similar to the classical Bérnard problem;
see, for example, Koschmieder, 1993) in Buzzard Bay, Massachusetts.
Another example of sequential multiscale modeling is provided by work in the
atmospheric and environmental science (see, for example, Elbern, 1997). In this
case, sophisticated computations were used to evaluate reaction barriers of as many
as 150 chemical reactions involving over 60 reactants. The results were then used in
rate equations for large scales that were coupled to computer codes for spatial grid
generation in order to determine and predict chemical meteorology. Elbern (1997)
carried out such computations for a domain that covered an area from the eastern
North Atlantic to the Black Sea, and from northern Africa to central Scandinavia.
The domain was partitioned into many subdomains, each of which was assigned to
a processor. Each processor communicated with only its neighboring processors
that used its results (output) as the input for their own computations.
A good example of sequential multiscale modeling of materials is provided by
the work of Zepeda–Ruiz et al. (1999). Their goal was designing experimental
protocols toward the development of engineering strategies for strain relaxation
of semiconductor films which are grown heteroepitaxially on semiconductor sub-
strates. The strain is caused by the lattice mismatch between the film and substrate,
and controlling it, which generates defects in the material, is the key to optimal
design of the film’s optoelectronic properties. One important design parameter is
the thickness of a compliant substrate. The thickness and elastic properties of such
substrates are comparable to those of the epitaxial film. They behave as if they
are unconstrained at their bases, and thus they can aid the film in accommodating
the lattice mismatch by either contracting or expanding parallel to the interface
between the film and the substrate. The compositional grading of the epitaxial film
is another important design parameter.
Matthews and Blakeslee (1974) and Freund and Nix (1996) have already de-
veloped a continuum elasticity theory for describing strain relaxation mechanisms
in semiconductor heteroepitaxy, and predicting the critical epitaxial film thick-
ness that marks the onset of misfit dislocation formation in heteroepitaxy on both
thick and thin substates. Their continuum elasticity theory successfully predicts
both the energetics and kinetics of some relaxation phenomena associated with
the formation of strained islands grown on semiconductor surfaces. However, the
theory does have its limitations. For example, the continuum elasticity theory in
conjunction with equilibrium bulk values of the elastic moduli are used commonly
for quantitative analyses. However, it is well-known that the elastic stiffness co-
efficients depend strongly on the stress, which may attain very large values in an
554 10. Multiscale Modeling of Materials

epitaxial thin film. Atomistic simulations can overcome such limitations. Using
either an ab initio or semi-empirical description of the interatomic interactions
between the material’s atoms, energy minimization or MD simulations, of the type
that were described in Chapter 9, can be utilized for investigating the structure,
energetics, and dynamics of a system consisting of an epitaxial film on a substrate.
Using a multiscale approach, Zepeda–Ruiz et al. (1999) studied the energetics,
strain fields, and semi-coherent interface structures in a layer-by-layer semicon-
ductor heteroepitaxy, such as InAs on GaAs(110) and InAs on GaAs(111)A, as well
as the interfacial stability with respect to misfit dislocation formation and the mor-
phology of the surface of the film grown on the substrate. The continuum theory
provided a parameterization scheme for the atomistic simulations. A Keating-type
potential (see Chapter 9), which contains the contributions of the stretching and
bond-bending forces, was utilized for representing the interatomic interactions,
and total energy minimization was used for determining the most stable configu-
ration of the system (i.e., the one with the lowest-energy state). The minimization
was done based on the conjugate-gradient method (see Section 9.5.2) with respect
to the atomic coordinates for a given strain state, which were uniaxial, biaxial, or
fully relaxed. Because of misfit dislocations a supercell (see Section 9.1.3) was
used that, depending on the state of the system, contained anywhere from 500 to
140,000 atoms. A major conclusion of this work was that the continuum theory of
elasticity can be accurate all the way down to the monolayer thickness, which is
the finest possible length scale for the theory in the context of layer-by-layer expi-
taxial growth. In addition, the theory was shown, in conjunction with the atomistic
simulations, to provide quantitative predictions for various properties of interest.
Indeed, even the linear isotropic elasticity was capable of fitting the results of the
atomistic simulations.

10.1.2 Parallel Multiscale Approach

This method is not as well-developed as the sequential approach, because it re-
quires very significant computing power which, up until very recently, was not
available. In this method, different computational methods, ranging from those
for atomic scales to continuum scales, are coupled for a simultaneous attack on
a given problem. The reason for this coupling is that many physical phenomena
are in fact inherently multiscale; that is, one must know what is happening simul-
taneously in different regions and scales of the system in order to understand and
predict its macroscopic behavior. A good example is fracture propagation in solid
materials which was described in detail in Chapters 6–9. The atoms that constitute
the material interact with each other, and nucleation of a crack and its propagation
are due to what happens at this length scale, namely, breaking of the bonds be-
tween the atoms. As the nucleated crack starts to propagate, complex phenomena,
such as plastic deformation, happen at a larger length scale which includes the tip
of the propagating fracture. At still larger length scales, which include the region
far from the tip of the crack, the material behaves as a continuum which may be
described by the classical continuum mechanics. The most important property that
 10.2. Defects in Solids: Joining Finite-Element and Atomistic Computations 555

such a multiscale approach must possess, in addition to being efficient, is accuracy.

However, accuracy in this context is somewhat different from the traditional way
that one understands this word, since accuracy in the present context means that
the dynamics of a phenomenon under study must be indistinguishable whether
they are determined from a multiscale approach or from a system that has the
same size but is studied only by a quantum-mechanical approach, i.e., at the most
fundamental length scale.
It should therefore be clear to the reader that the multiscale approach can, in
principle, solve the problem of how to deal with a heterogeneous material with
several distinct length scales. However, in order to be able to make quantitative
predictions for materials’ properties using the multiscale approach, one must also
solve the problem of the disparate time scales, namely, the wide gap between the
time scale over which the actual physical phenomena occur, and those that can,
with the current computational power, be accessed. This disparity between the two
time scales poses severe restrictions to predictive material modeling, regardless
of whether one uses sequential or parallel multiscale modeling. For example, in
parallel multiscale modeling, the time scale is determined by the finest atomic-
scale method that is used in the model, e.g., the QMD or a TB formulation, and
this is typically up to one nanosecond. Such time-scale limitations can be partially
overcome by using a dynamic Monte Carlo method, or by an accelerated MD
technique. In this book, we consider only the question of multiscale modeling of
properties of materials that contain several relevant and disparate length scales.
modeling of phenomena in which there are several relevant and widely disparate
time scales has been discussed by, for example, Voter (1997).
We now begin describing and discussing a few important examples of use of a
multiscale approach to modeling physical phenomena in materials.

10.2 Defects in Solids: Joining Finite-Element

and Atomistic Computations
The first example that we describe is the pioneering work of Tadmor et al. (1996a)
who studied how a defect, such as a dislocation, crack, or grain boundary, can
be embedded within a continuum, but without the standard assumptions that are
inherent in a simulator based on a continuum model which usually utilizes ad-
hoc assumptions about failure or fracture of a given region of a material. Their
original study was carried out before the current computational power with parallel
machines had reached its current state, and as such it had certain limitations. In
subsequent papers (see, for example, Shenoy et al., 1998, 1999, 2000; Tadmor
et al., 1999), the methodology was refined and extended to various applications.
What follows is a summary of this important work.
As pointed out by Tadmor et al. (1996a), the analysis of the structure of crystal
defects requires consideration of anharmonic effects on the scale of the lattice.
Such effects can, of course, be studied by atomistic simulations, such as ab initio
or MD computations (see, for example, Arias and Joannopoulos, 1994). However,
556 10. Multiscale Modeling of Materials

such computations alone cannot do justice to the problem, nor can they give us a
quantitative picture of what is happening during the deformation of a crystal, if
we consider the fact that macroscopic deformation of crystals involves dislocation
densities as high as 1013 m−2 , so that an area as small as 1 mm2 near the tip of a
crack can be crossed by over one hundred dislocations. In contrast, the supercell
that was used in the ab initio computations of Arias and Joannopoulos (1994)
contained only 324 atoms. Thus, it would be impossible to consider individual
dislocations at the macroscopic scale. Even the intermediate length scale, of the
order of a few hundred nanometres, is presently hardly within reach of conventional
atomistic simulation, as the number of atoms involved in such simulations would
be in excess of 108 , which makes purely atomistic simulations difficult to carry
out. One may develop phenomenological continuum models in which the defects
are treated as continuously distributed objects. However, such an approach could
not provide deep insight into the role of the defects in the properties of materials.
Tadmore et al. (1996a) were interested in studying the deformation processes
of interest at the intermediate length scale which, on one hand, involve discrete
dislocations in numbers that are too small to be described adequately by a con-
ventional continuum model of crystal plasticity, and, on the other hand, contain
too many atoms to be treatable by purely atomistic simulations. Thus, a multiscale
approach that joins the two methods, and would seem to be the only practical
solution, should have several key attributes, some of which are as follows.

(1) The theory should, at macroscopic length scales, reduce to the continuum
crystal elasticity and reproduce its important properties, such as material frame
indifference and crystal symmetry (see, for example, Milstein, 1982).
To achieve this goal, the atoms should be constrained to move in accordance
with the continuum displacement field. This would enable one to compute en-
ergies and forces from local lattice calculations. Thus, by construction, the
resulting continuum would automatically satisfy material frame indifference
and exhibit all the symmetries of the crystal. The system would also possess
lattice invariance, i.e., its energy would be invariant with respect to distor-
tions of the reference configuration. In particular, the energy density would
be periodic under crystallographic slip. An important consequence of this
periodicity is the lack of quasi-convexity of the energy functional [see, for
example, Fonseca (1988); see also Chapters 2 and 4, as well as Chapters 4,
7, and 10 of Volume I], which would make stable development of lattice de-
fects possible. The relaxation of functionals lacking quasi-convexity would
require consideration of minimizing sequences of deformations which exhibit
structure on increasingly finer scales, hence necessitating multiscale analysis
for simultaneous resolution of macroscopic and microscopic features into the
model.
(2) At the atomic scale, the theory should be built upon reliable interatomic inter-
actions, incorporate a lattice constant (or lattice parameter, as referred to by
Tadmor et al.), and possess all the invariance properties that are expected of a
crystal lattice.
 10.2. Defects in Solids: Joining Finite-Element and Atomistic Computations 557

Introduction of the lattice parameter is for preventing the displacement field

from developing unphysical sublattice-scale structure. Since the continuum
equations are solved by a finite-element (FE) method, the lattice parameter
can be set so as to impose a lower bound for the FE mesh size. In this manner,
the crystal takes on a character similar to the concept of quasi-continua devel-
oped by Kunin (1982). The elements in the model are either local or non-local
depending on their size, extent of deformation in their region, and energy. The
latter type of elements are those that are small, highly deformed and energetic.
The domains of the interaction of the two types of element are also different.
A local element interacts with deformation only within its own geometrical
domain, whereas non-local elements interact also with their neighboring ele-
ments. At large scales, the crystal becomes indistinguishable from a nonlinear
elastic crystal whereas, in the fine-mesh limit, the theory reverts to lattice stat-
ics. This implies, from an atomistic perspective, that the quasi-continua are
simply atomistic lattices that are subjected to kinematic constraints, namely,
those introduced by the FE interpolation. These constraints eliminate excess
atomistic degrees of freedom in regions where the deformation field varies
slowly on the scale of the lattices.
(3) In between, at intermediate length scales, the theory should produce a
continuous transition from the lattice to the continuum realms.
Note that the distinction between the continuum formulation and atomistic
description is not only important from a computational view point, but also
from the physical point of view. As discussed throughout this book, at the
macroscopic (continuum) level, solid materials are represented as continuous
media to which appropriate average material properties are assigned. On the
other hand, in atomistic description of a material, a solid is treated as a collec-
tion of atoms with interactions that are described by an appropriate energy or
potential function. This concept was also the foundation of the lattice models
of linear elastic properties of materials that were described and discussed in
Chapters 8 and 9 of Volume I, except that in those models the lattice was
representative of a coarser piece of materials, and therefore its sites did not
represent atoms.
(4) Finally, the theory should also enable one to obtain an accurate treatment of
lattice defects, such as dislocations, in case these defects exist or nucleate.

10.2.1 The Quasi-continuum Formulation

We first provide a qualitative description of the general ideas and concepts of the
approach, after which the details of the model are described. The theory begins
from an underlying conventional atomistic model, which is capable of delivering
the energy of the crystal as a function of the atomic positions. The configuration of
the crystal is then reduced by identifying a subset of representative atoms, which
henceforth become the sole independent degrees of freedom of the crystal. The
position of the remaining atoms are obtained by piecewise linear interpolation of
the representative coordinates, very much similar to the manner by which displace-
558 10. Multiscale Modeling of Materials

ment fields are constructed in the FE method. The selection of the representative
atoms may be based on the local variation of the deformation field. For example,
one may adapt the mesh in such a way that the variation of the displacement field
over each element of the triangulation does not exceed a fraction of the Burg-
ers vectors, hence ensuring that full atomistic resolution is achieved, for example,
near dislocation cores and on planes undergoing crystallographic slip. By contrast,
far away from any highly-stressed region, the density of the representative atoms
decreases rapidly, and the collective motion of a very large number of atoms is
governed, without significant loss of accuracy, by a small number of degrees of
freedom. In these coarse regions, the behavior of the model is ostensibly indistin-
guishable from that of a continuum. The effective equilibrium equations are then
obtained by minimizing the potential energy of the crystal over the reduced con-
figuration space. Therefore, the number of equilibrium equations that are obtained
is commensurate with the number of representative atoms. However, a direct cal-
culation of the effective force field requires, in principle, the evaluation of sums
that are extended over the full collection of atoms. Such full sums may be avoided
by the introduction of approximate summation rules, whereupon the complexity
of the computation of the effective force field becomes of the order of the reduced
model.
We now describe the details of the model. In the quasi-continuum (QC) theory
of Tadmor et al. (1996a) the continuum framework and continuum particle concept
are retained, but the macroscopic constitutive law is replaced by one based upon
direct atomistic calculations. The continuum particle is represented by a small
crystallite of radius Rc which surrounds a representative atom. This crystallite is
deformed according to the local continuum displacement field, and its energy is
computed based on an appropriate atomistic model. In order to compute the ener-
gies, one must make a correspondence between the deformation of the crystallite
and the continuum displacement field. A standard approach for doing so is based
on the Cauchy–Born rule (see, for example, Ericksen, 1984) according to which
the atomic positions are related to the continuum fields through a local deforma-
tion gradient F (which, for infinitesimal deformation, is the usual ∇u, where u is
the infinitesimal displacement) which is applied to the crystal’s undeformed lat-
tice basis. The crystal is then reconstructed from the altered base vectors. In this
manner each continuum particle is represented by an infinite crystal undergoing
homogeneous deformation. This limit is referred to as the local QC formulation.
The key idea is that, since the energy of each point is obtained directly from atom-
istic simulations, important properties of the crystal, such as its symmetries, are
automatically introduced into the description of the material. However, despite
being elegant and straightforward to implement, the local QC formulation suffers
from several shortcomings that make it necessary to develop a formulation that can
deal with non-local effects. Some of such shortcomings, as discussed by Tadmor
et al., are as follows.
(1) The most important shortcoming of the QC formulation is that, due to the
homogeneous nature of the deformation in this formulation, it is not possible
 10.2. Defects in Solids: Joining Finite-Element and Atomistic Computations 559

to model important heterogeneities, such as stacking faults, which are two

undeformed crystalline half spaces slipping over each other by a non-lattice
translation vector, and are therefore non-uniform. However, within the local
QC formulation, such structures can only be modeled via a simple shear de-
formation which, except for the two atomic layers directly adjacent to the slip
plane, results in a structure which is completely different from what is found
in real stacking faults.
(2) Another difficulty with a local formulation is that it does not allow for interface
defects, such as free surfaces, grain boundaries or other heterogeneous inter-
faces. Moreover, the lattice parameter, which serves as the crystal’s intrinsic
length scale, is lost in the Cauchy–Bohr process, hence allowing the energy
minimization methods to develop structure on sublattice length scales, which
is clearly unphysical. Such deficiencies are critical near the defects, hence
emphasizing the need for an approach that can deal with non-local effects.
Such difficulties are circumvented by a non-local QC formulation of the prob-
lem. In this approach each atom within the representative crystallite is displaced
according to the actual continuum displacement field at its position, implying that
the position rn of the atom after deformation is given by, rn = Rn + u(Rn ), where
u is the continuum displacement field, and Rn is the atom’s position before de-
formation. The local and non-local formulations are equivalent as long as the FE
is large enough to entirely contain the representative crystallite centered about its
quadrature point; see Figure 10.1. However, as the elements become smaller than
Rc , members of the representative crystallite will fall inside different elements and
experience a non-uniform displacement field; see Figure 10.2. This allows model-
ing of stacking faults in a straightforward manner. The use of non-local elements
near the stacking fault plane captures the true non-uniform deformation. The non-

Rc

Representative Atom
Figure 10.1. Local quasi-continuum/finite-element in which the triangle corners represent
the nodes, while the circles are atoms that belong to the crystallite (after Tadmor et al.,
1996a).
560 10. Multiscale Modeling of Materials

Rc

Representative Atom
Figure 10.2. Non-local quasi-continuum/finite-element (solid triangle) surrounded by
nearby elements (dashed triangles) (after Tadmor et al., 1996a).

Figure 10.3. Interfacial effects represented by a grain boundary (after Tadmor et al., 1996a).

local formulation also allows treatment of problems that involve, for example,
grain boundaries. If elements smaller than the representative crystallite radius Rc
are placed near such an interface, they will, due to non-locality, contain atoms that
are arranged in a different crystal orientation, and thus mimic a grain boundary;
see Figure 10.3.
 10.2. Defects in Solids: Joining Finite-Element and Atomistic Computations 561

10.2.2 Constitutive Models

Tadmor et al.’s use of atomistically-derived constitutive relations has the important
property that it preserves all the relevant crystal symmetries. The most important
of symmetry properties to be preserved is slip-invariance, which expresses the fact
that the energy of a solid material is invariant under crystallographic slip, and is
described by the one-parameter family of deformation mappings:
H(s ) = U + s s ⊗ n, (1)
where n is the normal to the slip plane, s is a vector in the slip direction, s is the
slip strain, and U is the identity matrix. Because of lattice invariance, the energy
density E(s ) is periodic in s with period b/d, where b = |b| is the magnitude
of the translation vector, and d is the distance between adjacent crystallographic
planes perpendicular to n. The slip invariance plays a vital role in allowing for the
presence of dislocations and other stable defects in the crystal.
A global origin in the undeformed configuration, relating the continuum FE
model to the underlying crystal structure, is now set. Then, for every quadra-
ture point in the FE mesh (where the continuum fields are sampled for numerical
integration) the nearest atom is selected as a representative atom, and a small neigh-
borhood around it is deformed either according to the local deformation gradient
tensor F (see above), if the element is local, or based on the actual displacement of
every atom according to the global continuum displacement fields, if the non-local
formulation is used (see below). The total energy of the representative atom is
then computed using an appropriate atomistic model. The computed energy and
its derivatives at the quadrature point are then supplied to the FE model. By refin-
ing the mesh and using the non-local model in highly strained regions, the effect
of nonlinear core will be taken into account, while in less strained regions far from
the core the local approximation will yield linear elastic behavior, as it should.

10.2.3 The Atomistic Model

As the atomistic model, Tadmor et al. (1996a) utilized the embedded-atom method
(EAM) (see Section 9.7.1) to compute total energies for their system. An important
flaw of EAM is that, it grossly underestimates the stacking fault energies, thus
hampering its ability for modeling dislocations. However, Ercolessi and Adams
(1993) used ab initio calculations (see Chapter 9) to fit the parameters of the EAM
to a range of material properties, thus enabling the EAM to yield substantially
larger intrinsic stacking fault energy. The Ercolessi–Adams potential was utilized
by Tadmor et al. (1996a), although, in principle, any other atomistic model could
have been used.

10.2.4 Field Equations and Their Spatial Discretization

The next step is to specify the field equations so that they can be discretized for
use in the numerical simulations. One considers a crystal that occupies a reference
configuration B0 in R 3 , represented by a material Cartesian coordinate system,
562 10. Multiscale Modeling of Materials

{XI , I = 1, 2, 3}. The crystal undergoes a motion described by a deformation

mapping (X, t). The image of B0 by (·, t) defines the deformed configuration
Bt of the crystal at time t, represented by a spatial coordinate system, {xi , i =
1, 2, 3}. The deformation at time t of an infinitesimal material neighborhood d0
around a point X of B0 is completely defined by the linear part of (·, t) at X.
This defines an affine mapping given by
dxi = FiJ (X, t)dXJ , (2)
where FiJ are the components of the deformation gradient tensor F given by
FiJ (X, t) = φi,J (X, t), (3)
where upper-case indices refer to the material frame, lower-case indices to the
spatial frame, and (·), J indicates differentiation with respect to XJ . In invariant
notation, F = ∇ 0 , where ∇ 0 denotes the material gradient operator.
The reference boundary ∂B0 is now partitioned into a Dirchlet (displacement)
component ∂B01 and a Neumann (traction) component ∂B02 . A given displacement
¯ is imposed on ∂B01 , and ∂B02 is subjected to a given traction T̄. Moreover, body
forces per unit volume ρ0 B act on the solid material, where ρ0 is the reference
mass density and B is the body force field per unit mass. Stable configurations of
the crystal are those that minimize the total potential energy given by
$   %
E[] = inf ψ Es (ψ)d0 − ρ0 B · ψd0 − T̄ · ψdS0 , (4)
B0 B0 ∂B02
where Es (ψ) is the strain energy density computed from the EAM (or any other
appropriate atomistic model), and the trial deformation mappings ψ belong to
some suitable space of functions over B0 which satisfy the boundary conditions
ψ = ¯ on ∂B01 .
If the deformation of the crystal is small, the above formulation reduces to con-
ventional anisotropic elasticity in which energy minimizers are uniquely defined
up to a rigid body motion, conditions that are realized in regions of the crystal far
from the lattice defects. However, as discussed above, the fact that the strain energy
density Es is computed from an atomistic potential implies that it is possible for
energy minimizers to develop microstructures on a fine scale, including lattice’s
defects such as dislocations. On this scale, the periodicity of the lattice, and the
resulting periodicity of the energy function with respect to crystallographic slip,
are crucial.
Because of the multiscale nature of the problem, the field equations are dis-
cretized by an adaptive FE method. First, the reference configuration B0 is
partitioned into FEs {eh , e = 1, · · · , M}, where M is the number of elements,
and h is a measure of the size of the smallest element. The deformation mapping
and deformation gradients are discretized by the standard FE method (see, for
example, Strang and Fix, 1973):


N
h (X, t) = a (t)Na (X), (5)
a=1
 10.2. Defects in Solids: Joining Finite-Element and Atomistic Computations 563

N
Fh (X, t) = a (t)∇ 0 Na (X), (6)
a=1

where a = 1, · · · , N denotes the nodes in the mesh, with N being the number of
nodes, a (t) are the nodal coordinates at time t, and Na (X) are the interpolation
functions. The main unknowns of the problem are now the nodal coordinates a (t)
which are obtained from the constrained minimization problem (4) in which the
ψ are replaced by ψ h , with the trial functions ψ h being of the following form,


N
ψ h (X) = ψ a Na (X). (7)
a=1

These trial functions must satisfy the boundary conditions identically on ∂B01 . All
integrals in Eq. (4), when written in terms of ψ h , can be conveniently computed
by numerical quadrature at the element level, which reduces all stress-strain calcu-
lations to the quadrature points of the elements. Tadmor et al. (1996a) used linear
three-noded triangular elements with one-point quadrature rule, and constructed
all the FE meshes by automatic triangulation based on the Delauney algorithm (see,
for example, Sloan, 1987). The constrained minimization problem was solved by a
conjugate-gradient approach (see Section 9.5.2), followed by a Newton–Raphson
iteration when the initial guess was too far from the solution.

10.2.5 Local Quasi-continuum Formulation

Given the continuum deformation fields at a quadrature point, one must now
compute the energy and its variations at that point. In the local quasi-continuum
formulation, each point of a solid material is represented locally by an infinite crys-
tal which is deformed homogeneously, resulting in the loss of the global origin
that links the underlying crystal lattice to the continuum. Therefore, the choice of
representative atom is unimportant since all atoms are equivalent. Thus, one can
assume that the infinite crystal surrounds a representative atom at the origin. Then,
the Cauchy–Born approximation (see, for example, Ericksen, 1984) is used so
that the infinite crystal is deformed according to the local continuum deformation
gradient. Consequently, if {AI , I = 1, 2, 3} is a crystal basis, then the coordinates
of its atoms are given by
X(m) = mI AI , m ∈ Z 3 , (8)
where Z represents the set of integers. The positions of the atoms in the deformed
configuration are then taken to be
x(m) = FX(m), m ∈ Z 3 , (9)
where F is the local deformation gradient which is constant within the element.
In practice, a region of radius Rc (taken to be about twice the cutoff radius rc of
the atomistic potential) represents the infinite crystal. The applied trial deforma-
tion F should not be so large so as to bring atoms from outside the region Rc to
564 10. Multiscale Modeling of Materials

within the cutoff radius rc . To account for this effect, Tadmor et al. introduced
the concept of an influence radius Ri associated with deformation F, which is
the radius that corresponds to the most distant point in the undeformed configura-
tion that is mapped onto the representative atom’s cut-off sphere in the deformed
configuration. Tadmor et al. showed that Ri is given by
!
Ri = rc λmax , (10)
where λmax is the largest eigenvalue of (F−1 )T F−1 . Then, every trial deformation
during the minimization process must satisfy the constraint, Ri ≤ Rc .
Given an acceptable trial deformation, the strain energy density Es , which is a
function of F, is computed using the atomistic method. The local contributions to
the out-of-balance force residual and global stiffness matrix follow from Eq. (4)
(when written in terms of ψ h ) as

∂E

local  
= (P · ∇ 0 Na )d0 − ρ0 BNa d0 − T̄Na dS0 , (11)
∂ψ a e eh B0 ∂B02

∂ 2E

local
= [C : (∇ 0 Na ⊗ ∇ 0 Nb )]d0 , (12)
∂ψ a ∂ψ b e eh

where a and b are node numbers, P = ∂W/∂F is the first Piola–Kirchhoff stress
tensor (see Section 7.11.3), and C = ∂ 2 W/∂F2 is the Lagrangian tangent stiffness
tensor. These tensors are the finite deformation analogues of the usual Cauchy
stress and elastic modulus tensors in linear elasticity described in Chapter 7 of
Volume I. Note that the first terms of Eqs. (11) and (12) are sums only over local
elements e in B0 . In terms of the components of P and C, Eqs. (11) and (12) are
rewritten as
∂E

local  
= PiJ Na,J d0 − ρ0 Bi Na d0 − T̄i Na dS0 , (13)
∂ψai e eh B0 ∂B02

∂ 2E

local
= [CiJ KL Na,J Nb,L ]d0 , (14)
∂ψai ∂ψbk e eh

where
∂Es −1
PiJ = = σij FJj , (15)
∂FiJ
∂ 2 Es −1 −1
CiJ KL = = (cij kl + δik σj l )FJj FLl , (16)
∂FiJ ∂FkL
where the relations between P and C on one hand, and their spatial counterparts,
the Kirchhoff stress σ and the spatial moduli c, on the other hand have been used,
with δik being the Kronecker delta. Tadmor et al. utilized a three-noded linear
element, which means that F was constant within each element. Thus, the integrals
in Eqs. (13) and (14) are computed by evaluating the integrands for the value of F
within each element and multiplying the result by the area of that element.
 10.2. Defects in Solids: Joining Finite-Element and Atomistic Computations 565

The treatment of the problem presented so far is completely general and inde-
pendent of the atomistic model. The appropriate expressions for the components of
the stiffness tensor C were already given in Chapter 9, Eqs. (9.162)–(9.164). The
corresponding components of the spatial moduli tensor c are then computed from
Eq. (16). For a purely local formulation, the expressions for C and c represent a
complete constitutive description of the problem. However, as discussed above, a
purely local formulation is unable to capture non-uniform effects, such as stacking
faults. Therefore, one must resort to a non-local formulation which is described in
the next section. Note, however, that for a pure infinite crystal which is deformed
homogeneously, translational invariance reduces the general expression for the
EAM to that of a single atom and all of its neighbors that are within a pre-specified
cut-off radius rc .

10.2.6 Nonlocal Quasi-continuum Formulation

In this formulation each quadrature point is represented by a single atom with
neighbors that are displaced in accordance with the continuum displacement fields.
This is shown in Figure 10.2. The global lattice is retained, and one must explicitly
account for the position of the atoms in the pre-presentative crystallite relative to the
continuum mesh. Thus, the global Cartesian coordinates R e of the representative
atom of element e are introduced. The positions of the atoms belonging to the
representative crystallite are written as
Xe (m) = R e + mI AI , m ∈ Z 3 . (17)
Equation (17) connects the Bravais lattice to the FE mesh, and establishes a one-to-
one relationship between the atomic sites and the continuum fields. The deformed
atomic positions are obtained by interpolation from the FE mesh through,
xe (m) = Xe (m) + ψ a Na [Xe (m)]. (18)
Unlike the local case, it is not possible in the non-local formulation to define
general measures of stress and stiffness, since a uniform strain field does not exist.
As a result, the out-of-balance force residual and global stiffness must be written
explicitly in terms of the EAM. We write the energy density of the system as [see
Eq. (9.159)]
⎡ ⎤
1
⎣1
1

E= Uij (rij ) + Ei (ρi )⎦ = [Ui + Ei (ρi )] , (19)
 2 
i j
=i i

where we have used a slightly different notation than what was used in Eq. (9.159)
in order to avoid confusion. As before, ρi is the local embedding density around
atom i. Then, the non-local contribution to the out-of-balance force residual is
given by
 $ %  
∂E

non−local
1
1  m ∂rem
  m
= U (ρe )ρ (re ) + E (re ) eh , (20)
∂ψai e
 m
2 ∂ψ i
a
566 10. Multiscale Modeling of Materials

where the sum over the elements e is over only non-local elements in B0 , eh is
the area of element e, and the total electron density ρe at the representative atom
of element e is given by [see Eq. (9.160)]


ρe = ρ(rem ). (21)
m

Here ρ is the electron density function, and

rem = |xe (m) − xe (0)| = |mI AI + ψ a {Na [Xe (m)] − Na (R e )}|. (22)
In Eq. (20)
m
∂rem e (ri )e
= {Na [X e
(m)] − Na (R )} , (23)
∂ψai rem
where
(rim )e = xie (m) − xie (0). (24)
One should keep in mind that for FE meshes that contain both local and non-
local elements, the total out-of-balance force residual ∂E/∂ψ a is calculated as
a superposition of the two vectors given by Eqs. (13) and (20). The non-local
contributions to the global stiffness matrix are similarly calculated:
  
∂ 2E

non−local
∂r m
∂ren
  
 j
= U (ρe ) ρ (rem ) e i ρ (ren ) j
∂ψai ∂ψb e m
∂ψa n ∂ψb
 

$ %
1  m ∂rem ∂rem
$
1  m
%
∂ 2 rem
  m   m
+ U (ρe )ρ (re ) + E (re ) + U (ρe )ρ (re ) + E (re ) .
m
2 ∂ψai ∂ψ j 2 ∂ψai ∂ψb
j
b

(25)
Similar to the total out-of-balance force residuals, FE meshes that contain both
local and non-local elements yield a stiffness matrix which is obtained by super-
position of both Eqs. (14) and (25). The inclusion of non-local elements near highly
deformed regions, such as stacking faults, completes the QC formulation.

10.2.7 The Criterion for Nonlocality of Elements

The procedure for the computations is completed once one introduces a criterion
for determining the status of an element in the FE mesh in terms of it being local
or non-local. Although it may seem that a natural criterion would be to consider,
and locally compute, elements that are larger than the local crystallite radius Rc
and treat the smaller elements non-locally, this criterion would be wasteful from
a computational point of view because the non-local QC formulation is needed
only close to defect cores and along slip planes where stacking faults develop. Far
away from such highly inhomogeneous regions the local formulation, which is less
computationally intensive and also more stable, should perform well. Therefore,
as pointed out by Tadmor et al. (1996a), it is of interest to develop a criterion which
 10.2. Defects in Solids: Joining Finite-Element and Atomistic Computations 567

is capable of identifying regions that undergo large inhomogeneous deformation.

Based on such considerations, the criterion that they developed is as follows.
Consider the second invariant I I of the Lagrangian strain tensor :
1 
I I =  :  − tr()2 = 122
+ 13
2
+ 232
− (11 22 + 22 33 + 11 33 ), (26)
2
where
1
 = (FT F − U). (27)
2
Then, an element is considered as non-local if
!
|I I | > c , (28)
where c is a critical strain, the value of which depends on the material and the
phenomenon under study. In addition to the elements that satisfy criterion (28),
Tadmor et al. imposed the additional condition that elements in their immediate
vicinity that share atoms with those elements deemed to be non-local by the strain
criterion, should be also treated as non-local. Such elements are referred to as
non-local by proximity.
The task of selecting the non-local elements is not yet complete though. Con-
sider, for example, a material that contains a dislocation, as shown in Figure 10.4,
where only elements along the slip plane are shown. In such a case, elements far
to the right of the dislocation core will be undeformed, while elements far to the
left experience perfect Burgers vector slip. In both cases, these are zero energy
modes that correspond to an undeformed crystal which can be treated by local
elements. However, with the purely kinematical criterion (28), the elements on the
far left of the slip plane will be identified as non-local. Therefore, Tadmor et al.
(1996a) added a more stringent non-locality criterion to the kinematic criterion
(28): In addition to being highly strained, the deformation within the element must
produce a non-zero strain energy Es such that
Es > E0 (29)
where E0 is typically small.
Summarizing, the procedure for determining the status of all elements in the FE
mesh is as follows.

Figure 10.4. Slip plane elements near a

dislocation core (after Tadmor et al., 1996a).
568 10. Multiscale Modeling of Materials

(1) One computes the Lagrangian strain tensor  for all elements in the FE mesh.
(2) Condition (28) is checked for all elements smaller than Rc in order to identify
non-locally strained elements.
(3) The energy of elements identified in (2) is computed by using the non-local
formulation of the problem. Only those elements that satisfy criterion (29) are
retained and considered as non-local.
(4) The elements that are non-local by proximity are located by identifying all
elements that are smaller than Rc and are within a distance Rc of an element
that satisfies both criteria (28) and (29).
(5) In addition, all surface and interfacial elements are computed non-locally.
Tadmor et al. (1996a) carried out this procedure at the start of each iteration
and integrated it into the solution process. As they pointed out, an important point
regarding this algorithm is that, to ensure convergence once an element is identified
as non-local, it must remain so from that point on even if in future iterations it no
longer satisfies the non-locality criterion, or is no longer in proximity to a non-local
element.

10.2.8 Application to Stacking Faults in FCC Crystals

An important question is whether the above theory can stably support lattice de-
fects, such as dislocations, and, if so, how similar are their core structures to those
predicted by a full atomistic simulation. To answer these questions, Tadmor et
al. (1996a) considered three FCC configurations: Stacking faults within the (111)
plane, the Lomer edge dislocation (001)[11̄0], and the primary FCC edge dislo-
cation (111)[1̄10]. In all cases the predictions of the atomistic-continuum model
were shown to be in very good agreement with the lattice statics results, under con-
ditions in which the same potentials were used in addition to equivalent boundary
conditions. Here, we describe their simulations for the stacking faults. In what fol-
lows the constitutive behavior of the crystal is modeled using the embedded-atom
potentials for aluminum due to Ercolessi and Adams (1993) with a cut-off radius,
rc = 5.56 Å, the lattice parameter, a0 = 4.032 Å, a representative crystallite ra-
dius, Rc = 9.87 Å, which for the perfect FCC lattice corresponds to 12 neighbor
shells with 249 atoms, a critical non-local strain c = 10% (which was used in the
non-locality criterion), and a zero energy tolerance, E0 = 10−3 eVÅ−3 .
Stacking faults (SFs) often arise in crystalline deformation processes. In FCC
crystals, for example, they are known to form on the dominant {111}110 slip
system, as a result of, for example, dissociation of a perfect edge dislocation into
two Shockley partial dislocations via the reaction,
1 1 1
[1̄10] → [2̄11] + [1̄21̄].
2 6 6
The two partials can then drift apart and leave a stacking fault ribbon be-
tween them. However, due to the imperfect stacking sequence in this ribbon,
 10.2. Defects in Solids: Joining Finite-Element and Atomistic Computations 569

there is an energy penalty—the SF energy—associated with its presence which

limits the indefinite drift of the partials. √
The expected out-of-plane displacement
jump across the stacking fault plane is 6a0 /12, which for aluminum is about
0.823 Å.
As the first step, the SF energy was directly calculated using lattice statics
(LS). The unrelaxed SF energy resulting from the use of the Ercolessi–Adams
potentials was 7.530 meVÅ−2 . The LS computation indicated that the atoms on
each of the four (111) planes adjacent to the slip plane (two above and two below)
contribute equally to the SF energy. This information is useful to the construction
of the FE mesh. On relaxation, the SF energy was found to reduce to about 6.5
meVÅ−2 , which is in reasonable agreement with the observed experimental values
for aluminum, which is in the range 7.5–9 meVÅ−2 .
Tadmor et al. (1996a) carried out two different analyses using the QC-FE
method. In the first analysis the FE mesh was generated in the plane that con-
tained the SF. The x-direction was taken to coincide with the Shockley partial
direction [2̄11], while the y-direction was set to the slip plane normal [111]. To
introduce the initial slip into the model, all √
nodes above the slip plane to the right
by the magnitude of the Shockley partial a0 6 were removed, and then all bound-
ary nodes were constrained to their initial positions. The resulting mesh, shown
in Figure 10.5, contained 42 elements and 32 nodes, hence resulting in 36 uncon-
strained degrees of freedom. As can be seen in this figure, the mesh contains two
different types of element. One is the large, nearly equilateral elements, which are

Figure 10.5. Finite-element mesh for initially distorted stacking faults (after Tadmor et al.,
1996a).
570 10. Multiscale Modeling of Materials

Figure 10.6. Underlying crystal

structure superimposed on finite el-
ements. Filled points are atoms with
zero depth, while open ones repre-
sent atoms with 1.426 Å depth (after
Tadmor et al., 1996a).

larger than Rc and are thus local. The elements of the second type are long and
narrow, and are considered non-local, but because their width is larger than Rc ,
they exhibit non-local effects only in the y-direction. Therefore, surface effects
play no role. The height of the narrow
√ elements was set equal to the interplanar
distance in the (111) direction, a0 / 3, and the global origin was selected such that
these elements fell between adjacent atomic planes; see Figure 10.6. Hence, the
elements that straddle the slip plane were allowed to act as a kinematical mecha-
nism for introducing slip. Thus, from an atomistic point of view, there is a jump
in displacement across the slip plane, as expected, while in the continuum there
exists a continuous linear variation in slip. However, because the energy is com-
puted atomistically, the manner in which the slip is distributed in the continuum
is unimportant. Five layers of the narrow elements were necessary to capture the
contributions of the four atomic planes adjacent to the slip plane to the SF energy;
see Figure 10.7. In this model, the SF energy is equal to the total energy per unit
thickness, divided by the width of the system. The model predicts that the unre-
laxed SF energy is identical to that obtained from the LS analysis. Following a
Newton–Raphson minimization, Tadmor et al. (1996a) found a relaxed SF energy
of about 6.2 meVÅ−2 which is smaller than, but comparable with, the LS value.
In the second analysis, the FE model was generated in the primary [1̄10]-[111]
coordinate system, and a deformation corresponding to a 1/2 Burgers vector slip
 10.2. Defects in Solids: Joining Finite-Element and Atomistic Computations 571

Figure 10.7. Representative atom

(filled circle) of central slip element
surrounded by all atoms within its
cutoff radius Rc (after Tadmor et al.,
1996a).

was used. On relaxation, Tadmor et al. (1996a) found the SF energy to be the
same as in the previous Shockley partial analysis, and out-of-plane displacements
between 0.822 and 0.824 Å were observed in all nodes above the slip plane, which
are very close to the expected value.

10.2.9 Application to Nanoindentation

Nanoindentation is a process that can be classified as a problem in small-scale
contact mechanics, but has major technological applications. Indenters that have
a radius of curvature of the order of 50–100 nm are now used routinely in experi-
ments. Recent experiments (see, for example, Zielinski et al., 1995; Gouldstone et
al., 2000) measured load-displacement curves and subsurface dislocations. One of
the important questions that arises in this setting concerns the conditions attendant
to dislocation nucleation. Upon indentation, and after an initial elastic stage, the
onset of permanent deformation is mediated by the nucleation and propagation of
dislocations. Analysis of the recent experiments indicate that the number of dis-
locations activated by such processes are typically about 100 or less, and as such
they are amenable to effective atomistic simulations (see, for example, Kelchner
et al., 1998). However, in such simulations, the indenter sizes that can be con-
sidered are much smaller than the experimentally-utilized values, which may in
572 10. Multiscale Modeling of Materials

turn cause premature nucleation of dislocations relative to experimental obser-

vations. Likewise, the size of the computational domain is necessarily limited,
and therefore the dislocations soon run up against artificial boundaries. In addi-
tion, within a strict atomistic simulations, it is difficult to account for the effect of
long-range elastic stresses that are, for example, present in a thin film-substrate
system. All of these issues indicate that the multiple length scales that are involved
must be explicitly accounted for, if a nanoindentation process is to be modeled
accurately.
Before any attempts are made for applying a mixed atomistic-continuum model
to indentation, a few key issues must be considered. For example, Sharp et al.
(1993) devised a type of “phase diagram” which divides the indentation response
into elastic, elastic-plastic, and brittle regimes. If the phase diagram is truly rep-
resentative, then the critical question to be asked is, how does this phase diagram
depend on the constituent atoms that make the solid? This is particularly impor-
tant, as one must find a way of describing the plastic deformation that takes place
beneath the indenter. Moreover, force-displacement relations that are the result of
indentation have been measured frequently. The question is whether such relations
can be predicted, and if so, whether they can be used as a way of differentiating
between different materials. Finally, indentation of a clean surface is different
from one that is coated by another material. Therefore, one must find a way of
connecting the two, if this is possible at all.
Tadmore et al. (1996b) applied their QC model to the problem of nanoinden-
tation. They considered a pseudo-2D model of indentation with both rectangular
punches and rounded indenters. The former case is, however, difficult as it leads
to nucleation of dislocations after shallow indents. A crystal orientation such that
(111) planes are perpendicular to the face being indented was considered, since
this orientation permits the development of conventional FCC edge dislocations
discussed above. A challenging problem in such simulations is the ongoing mesh
refinement that must accompany such processes as dislocation nucleation and
emission. During the indentation process, the large strains under the punch trig-
gers automatic mesh adaption, leading to refinement in areas of interest. Thus,
as dislocations nucleate, the mesh refinement follows the paths of the disloca-
tions, hence permitting the dislocations to move away from the punch into the
substrate.
Figure 10.8 shows the atomic positions obtained from the mixed atomistic-
continuum model using rectangular indenter for an indentation of 5.4 Å. As can
be seen, partial dislocations have developed that are separated by stacking faults.
Figure 10.9 shows the atomic positions when a rounded indenter was used. In
this case the response is elastic for indentation depths that are deeper than those
for which the first dislocation was nucleated in the rectangular indenter case, thus
raising the possibility of the existence of a connection between the geometry of
the indenter and the criterion for emission of dislocations. Knap and Ortiz (2001)
extended Tadmore et al.’s (1996b) and presented a full 3D QC analysis of the early
stages of nanoindentation in gold thin films.
 10.2. Defects in Solids: Joining Finite-Element and Atomistic Computations 573

Figure 10.8. Atomic positions ob- %

tained from simulation of the quasi-
continuum theory, using a rectangular
indenter for an indentation of 5.4 Å $
(after Tadmor et al., 1996b).

#

"

[I10] 

!




   ! "

[111] 

Figure 10.9. Atomic positions

obtained from the solution of
the quasi-continuum theory, for a
rounded indenter for an indenta-
tion of 4.6 Å (after Tadmor et al.,
1996b).
574 10. Multiscale Modeling of Materials

10.3 Fracture Dynamics: Joining Tight-Binding,

Molecular Dynamics, and Finite-Element
Computations
The second example that we describe is the multiscale modeling of dynamics of
fracture propagation in silicon that was developed by Broughton et al. (1999; see
also Abraham et al., 1998a,b). As they pointed out, in order to obtain meaningful
ensemble averages, or access time scales of practical use, one must be able to
simulate the system through times of the order of 1 nanosecond. A typical MD
time step is about 10−15 s. Therefore, on a typical parallel machine with about
50–100 processors, one time step is roughly 1 second of real time, implying that
106 time steps may be simulated in approximately 10 days on part of a reasonably
powerful parallel machine.
A tight-binding Hamiltonian was used for simulating the quantum-mechanical
coupling. The main justification for use of this Hamiltonian was the requirement of
computational speed, i.e., the need to simulate a non-trivial number of atoms within
1 second of real time. In addition, a TB Hamiltonian was used for its simplicity
and intuitive appeal. At the next coarser level, a molecular dynamics (MD) method
was employed for describing the material. The interatomic potential for describing
silicon in the MD simulations was the Stillinger–Weber (SW) potential described
in Section 9.7.2. As discussed there, one main advantage of the SW potential, in
addition to its accuracy, is the ease with which it can be utilized. Finally, at the most
coarsened level, FE computations were carried out assuming that linear continuum
mechanics is valid. This could be justified by the fact that the FE computations
were used only in the far-field (away from the crack tip) region where atoms are
perturbed only slightly from equilibrium, and therefore it is unnecessary to employ
nonlinear elasticity theory. Moreover, Broughton et al. studied a 2D model since
the systems of primary interest to them involved plane strain.
Figure 10.10 presents the three primary algorithms, the regions of space that
they describe, and the way they are distributed among different processors. Each
FE region was assigned to a different processor, and the MD region was domain-
decomposed (see Section 9.6.2.3) across several computer nodes. Likewise, the TB
region was spread over several processors. The computer program was written in
FORTRAN with MPI for the message passing, and hence it was portable to most
parallel architecture machines. The advantage of a pseudo-1D topology shown
in Figure 10.10 is that much message passing can be performed using the “shift”
operator. Moreover, only data within interaction range, defined by the Hamiltonian,
must be passed across boundaries, which are defined by the domain decomposition,
between processors. This is clearly a parallel multiscale approach.

10.3.1 The Overall Hamiltonian

Broughton et al. (1999) defined a Hamiltonian H for the entire system. The degrees
of freedom of this Hamiltonian are atomic positions r and their velocities ṙ for
the TB and MD regions, and displacements u and their rates of change with time, u̇,
 10.3. Fracture Dynamics 575

Figure 10.10. Domain decomposi-

Processor Number
tion of pseudo-1D system that shows
the coupling of the length scales. The 1 2 3 4 5 6 ... ... ... ... N
filled circles show the MD region
(after Broughton et al., 1999).

FE FE

TB Regions
Processors N=1; É;N+M

Domain Decomposition

for the FE regions. Conceptually, H is written as

nH = HTB ({r, ṙ} ∈ TB) + HMD/TB ({r, ṙ} ∈ MD/TB) + HMD ({r, ṙ} ∈ MD)
+HFE/MD ({u, u̇, r, ṙ} ∈ FE/MD) + HFE ({u, u̇} ∈ FE), (30)
implying that there are three separate Hamiltonians for each subsystem, as well
as Hamiltonians that govern the dynamics of variables at the interface regions
between two distinct domains belonging to distinct length scales (Broughton et
al. referred to such regions as the handshake regions). To obtain the equations
of motion for all the relevant variables in the system, the appropriate derivatives
of H are taken, and all the variables are updated in time steps using the same
integrator. Therefore, the entire time history of the system is obtained numerically,
given an appropriate set of initial conditions. Moreover, if one follows a trajectory
governed by H, the total energy of the system will be a conserved quantity, ensuring
the numerical stability of the simulations.

10.3.2 The Tight-Binding Region

Broughton et al. (1999) utilized the following TB scheme. As described in Section
9.4.6, semi-empirical TB involves, (a) an ansatz for the total energy of the system,
and (b) a parameterization of the integrals that occur in the mean-field treatments
of the electronic structure of a system. The total energy E of the system is written
as


No

ETB = n + Ur (rij ), (31)
n=1 i<j

where the sum is over all occupied states No up to the Fermi level, while the second
sum is over all pairs of atoms of the repulsive potential Ur . The eigenvalues {}
correspond to the one electron states of a first-principles Hartree–Fock or density
functional calculation (see Sections 9.0, 9.1, and 9.4), and are obtained from a
576 10. Multiscale Modeling of Materials

non-orthogonal one-electron Hamiltonian given by

H n = n S n . (32)
The one-electron wave functions {} were expanded as a linear combination of
atomic basis functions φ:


n = n
ciα φiα , (33)
iα

where the matrix elements of H and S were calculated by reducing the equiva-
lent integrals within an extensive database of first-principles calculations to the
following parametric forms, expressed in terms of the pairwise functions hαβ and
sαβ ,

Hiαjβ ≡ φiα |Ĥ|φjβ  = hαβ (rij ), (34)

Siαjβ ≡ φiα |φjβ  = sαβ (rij ). (35)
Here, n denotes the orbital number, while α and β label the basis functions which,
in the minimal basis of silicon studied by Broughton et al. (1999), represent the s,
px , py , and pz atomic orbitals, and therefore the size of the H and S matrices is
4N × 4N , where N is the number of atoms in the system. The functions hαβ and
sαβ smoothly truncate to zero near 5 Å, which is between the third- and fourth-
neighbor distances in silicon. The functions Ur , hαβ , and sαβ were obtained by
fitting to a database that involved the experimental indirect band gap of the diamond
cubic structure and the total energies of crystalline and defective diamond cubic
and β-tin silicon at different densities (Bernstein and Kaxiras, 1997).
The exact form of the basis functions φiα are not required, since the one-electron
states {} are represented (within this formalism) only by the sets of coefficients
{c}. For given set of atomic coordinates, the coefficients {c} were computed by
diagonalization. Forces were then computed from the derivative of the TB energy
with respect to displacement of the nuclei:
⎡ ⎤

No

$ %
∂Ur (rij )
∂H ∂S
fiTB = − ⎣ ⎦−
iαjβ iαjβ
n
ciα n
cjβ − n . (36)
∂ri ∂ri ∂ri
n=1 α jβ j
=i

Orthonormality forces the derivatives of the coefficients with respect to atomic

positions to vanish identically. From the knowledge of the forces, atomic coordi-
nates can be computed at any time using exactly the same algorithm as that used
for the MD system (see below) with the same time step, since the frequencies in
both cases are very similar. In fact, the parameters of the SW potential, used in the
MD simulations, are adjusted to ensure equality.

10.3.3 Molecular Dynamics Simulation

In the MD region, the interatomic forces were obtained from the SW potential,
a full description of which was given in Section 9.7.2. To integrate Newton’s
 10.3. Fracture Dynamics 577

equations of motion, Broughton et al. (1999) used the Verlet algorithm (see Section
9.2.2) since it is easily augmented to handle multiple time scale MD simulations.
A time step of t = 5 × 10−16 s was used. The mass m of the silicon atom is
m = 4.6639 × 10−26 kg. One can accelerate evaluation of the SW energy and
its corresponding forces by taking advantage of atomic neighbor tables, so that
computer time scales as O(N ), where N is the number of atoms in the system.
The two- and three- body terms in the SW potential truncate smoothly to zero
just before the second-neighbor distance in zero-pressure diamond cubic structure
silicon.

10.3.4 Finite-Element Simulation

As mentioned above, Broughton et al. (1999) used a FE method for describing the
far-field region of the system. Linear elasticity theory was used for developing the
FE equations of motion. In this formulation, the total elastic energy of a solid, in
the absence of tractions and body forces, is given by
EFE = Ep + Ek ,
⎡ ⎤

1 ⎣




3 3 3 3
Ep = ij (r)Cij kl (r)kl (r)⎦ d,
2
i=1 j =1 k=1 l=1

1
Ek = ρ(r)u̇2 (r)d, (37)
2
where  is the volume of the system, Ep is the usual Hookian potential energy
term which involves the symmetric strain tensor  and the elastic constant tensor C
(see Chapter 7 of Volume I), Ek is the kinetic energy, and ρ is the mass density. The
subscripts i, j , k, and l denote Cartesian directions. The strains and displacements
are related through the usual relation,
∂ui ∂uj
ij = + . (38)
∂xj ∂xi
The FE mesh used in this work is described shortly. From the knowledge of the
displacements and their time derivatives at the vertices of the FE cells (which were
triangular cells) and interpolation functions, one can determine the values of these
variables everywhere within each cell. Broughton et al. (1999) used linear interpo-
lation inside each cell, and therefore the displacement fields could be represented
in piecewise smooth fashion. Equation (37) is now approximated by
6  
1

Nc

EFE = p Kpq uq + u̇p Mpq u̇q ,
um m m m m m
(39)
2 m
p,q=1

where K and M are local stiffness and mass matrices, respectively, m is the cell
index, Nc is the total number of FE cells, u and u̇ defined only at the apices of each
triangle, and the sum over (p, q) is over the (3 × 2) Cartesian directions associated
578 10. Multiscale Modeling of Materials

with the same apices. The stiffness matrix K (m) associated with the mth triangle
is given by
L T (m)
K (m) = B CB , (40)
4Am
where Am is the area of the mth triangle, C is the reduced (3 × 3) elastic constant
matrix, B(m) is the matrix of coordinate differences of the apices of the FE mesh,
and L is the thickness of the material in the third dimension. C depends upon
the orientation of the system and is a function of the three basic elastic constants
of silicon, namely, C11 , C12 , and C44 , which are given for zero-temperature SW
silicon by Ray (1988) and Balamane et al. (1992) (these values are listed in Tables
9.4 and 9.6). Broughton et al. (1999) used the average of these quantities reported
in the literature. B(m) is given by
⎛ m ⎞
b1 0 a1m
⎜ ⎟
⎜ 0 a1m b1m ⎟
⎜ ⎟
⎜ bm 0 a m ⎟
⎜ 2 2 ⎟
[BT ](m) = ⎜ ⎟, (41)
⎜ 0 a2m b2m ⎟
⎜ ⎟
⎜ m ⎟
⎝ b3 0 a3m ⎠
0 a3m b3m
with
blm = yl+1
m
− yl+2
m
,
alm = xl+2
m
− xl+1
m
, (42)
where l = 1, 2 and 3 denotes the cyclic apex index, and x and y are the 2D FE
mesh coordinates with respect to which the displacements u were defined.
The mass matrix M requires some care since, in principle, the kinetic energy
density varies across any given cell. However, it is necessary to reduce the FE mesh
in the interface between FE and MD regions so as to coincide with the perfect
atomic lattice. Each atom is apportioned its kinetic energy accordingly. Thus,
for the FE mesh, Broughton et al. (1999) used the "lumped-mass" approximation
which reduces for the smallest mesh size to the atomic limit. In this approximation,
one third of the mass in each cell is apportioned to each apex, so that the kinetic
energy is given by


Nm
Ek = M t (u̇t )2 , (43)
t=1
ρL

Nc

3
Mt = δtml Am , (44)
3
m=1 l=1
where t labels the FE mesh points, of which there are Nm , and ml denotes the mesh
point index at each of the three apices of cell m. The u̇ are vectors of length two
since they relate to a mesh point.
 10.3. Fracture Dynamics 579

Forces that correspond to the displacements in Eq. (39) were computed by taking
the spatial derivatives. Displacements and their time derivatives were obtained as
functions of time, for given boundary conditions, using the same update algorithm
and time step as those used for the MD and TB computations. The force due to the
mth cell is given by
(m)
fFE = K (m) u(m) , (45)
(m)
where fFE and u(m) are of length six. The total force associated with a mesh point
is then the sum of the contributions from each of the cells with apices in common
with that point, and
t
fFE = M t üt . (46)

10.3.5 Interfacing Finite-Element and Molecular Dynamics

Regions
In order to develop the proper model for the interface between the FE and MD
regions, two principal issues must be addressed which are, (1) the overlap of the
FE mesh with the atoms, and (2) the proper form of the Hamiltonian HFE/MD .
To address the first issue, Broughton et al. (1999) generalized an idea due to
Kohlhoff et al. (1991). An imaginary surface is drawn between the FE and MD
regions. Within the range of the MD interatomic potential (i.e., the SW potential)
from this surface, FE mesh points are located at ideal lattice sites. In the absence
of diffusion, atoms or mesh points will remain on either side of this interface.
However, note that atomic motion may be viewed as displacement around a lattice
(mesh) site, and the displacement field may be viewed as motion of an atom away
from its perfect site. The same idea can also be used for amorphous materials if a
one-to-one mapping of a mesh point to an atom site is made. As one moves away
from the interface between the FE and MD regions into the FE region, the mesh
spacing may be made larger, with its size being dependent upon the physics of the
phenomenon under study.
Since Broughton et al. (1999) were interested in examining brittle fracture in
Si, they oriented their rectilinear system in such a way that it had (100) faces on all
sides. The FE region was represented as a 2D system which, nevertheless, could
handle the third dimension in plane strain, since the thickness L of the sample was
included in Eq. (44). Thus, a diamond cubic lattice was projected onto a (100)
plane. This is shown in Figure 10.11. The FE region can be made periodic in
exactly the same way as the MD region. For a periodic system, there are twice as
many cells as there are mesh points. Away from the interface region and into the
FE region, the mesh was expanded along one dimension (the long axis in Figure
10.10) while the mesh spacing was kept constant in the second dimension. The
function chosen for this expansion was of hyperbolic tangent form. Thus, near
the interface region there was no expansion of the atomic mesh, while far from
the interface region the spacing approached a constant value of ten atomic lattice
parameter. The transition region spanned about 200 Å.
580 10. Multiscale Modeling of Materials

Figure 10.11. Triangulation of the unit cell

shown by dashed lines. The lines on the left
and right are continuous, but not those on top
and bottom of the unit cell (after Broughton et
al., 1999).

To address the issue of the proper Hamiltonian HFE/MD , Broughton et al. (1999)
defined a conservative Hamiltonian so as to ensure symplectic time evolution (see
Chapter 9) of the atomic and displacement trajectories within the interface region.
To conceptualize this Hamiltonian, imagine that two different materials sit on
either side of an interface, such that on one side is FE silicon, whereas on the other
side one has SW silicon. The cross terms (i.e., the interface Hamiltonian HFE/MD )
can, to first order, be approximated by the average of the two descriptions: All FE
triangles that cross the interface contribute half their weight to the Hamiltonian,
while the triangles that are fully in the MD region contribute nothing. Similarly,
any SW interaction which crosses the interface contributes half its usual weight,
while the SW interaction between mesh points, which are fully on the FE side of the
interface, contributes nothing to HFE/MD . Since as discussed above, the atoms and
mesh points cannot be distinguished from each other, the SW energy formulation
for atomic coordinates r and the FE energy formulation for the displacements u
can be used throughout the interface. The one-to-one mapping of atoms to nodes is
no longer needed at distances (from the interface in the FE region) that are greater
than twice the SW pair cutoff, which is the distance of greatest three-body range
in the SW potential. Figure 10.12 presents such interactions and their ranges. Thus

1

Nx

6
I mI mI
EFE/MD = um
p Kpq uq
4 I
m =1 p,q=1
⎡ ⎤ (47)
1


+ ⎣ U2 (r(ij )I ) + U3 (r(ij )I , r(ik)I )⎦ ,
2 I (i<j ) [i,(j <k)]I

where U2 and U3 are the two- and three-body terms. Here, the superscript I refers
to those interactions that cross the FE/MD boundary [see Eq. (9.176)]. Indeed, as
discussed above, EFE/MD is defined only for interactions that cross the boundary.
Equation (47) allows any one atom of the triplet in the three-body terms to be on
an opposite side of the interface to the other two.
 10.3. Fracture Dynamics 581

Figure 10.12. The FE/MD interface region. Heavy

lines show the FE triangles that contribute fully
to the interface Hamiltonian, while the thin lines
are those that contribute half weight to the in-
terface Hamiltonian. Dotted lines show two- and
three-body terms of the Stillinger–Weber poten-
tial that cross the boundary and contribute half
weight, while full lines on the right represent full
SW contributions (after Broughton et al., 1999).

Since in the work of Broughton et al. (1999), the FE mesh was 2D while the MD
region was 3D, the third dimension of the FE region was treated by mean-field
approximation. Thus, in Eq. (47), x- and y-displacements of atoms on the MD
side of the FE/MD boundary that contribute to the elastic energy were obtained
by averaging over all equivalent atoms at the depth z. In a similar fashion, in
determining the SW energy contribution to HFE/MD , all x- and y-displacements in
the third dimension were replicated on the FE side of the boundary by assuming that
atoms are located at ideal lattice sites in that dimension. The overall Hamiltonian
remains conservative.
Two other issues regarding the definition of energy must still be addressed, both
of which involve a reference state. One involves the potential energy, while the
other has to do with the thermal energy. Consider first the potential energy. The
SW potential is referenced to infinitely separated atoms, whereas the FE potential
is referenced to a T = 0 unstrained lattice. Therefore, a constant offset energy that
did not affect the dynamics was added to each FE mesh point. The T = 0 energy
density for SW silicon at zero pressure is -4.33444 eV/atom. The offset energy was
computed for every FE point using an equation entirely analogous to that used to
compute mass in the “lumped-mass” approximation except that, instead of a mass
density, the SW energy density was used [see Eq. (43)]. This scheme ensured the
correct limiting behavior as the mesh spacing was reduced to atomic dimensions.
For atoms in the interface region, for systems with unusual orientation where
the offset is non-trivial to estimate atom by atom, a T = 0 calculation with zero
strain for the coupled FE/MD system can be performed. The offset may thereby
be calculated to maintain the energy/atom constant through the interface.
Rudd and Broughton (1998) showed that in the FE region (u̇)2 is related to the
temperature. However, as one moves away from the interface region and coarsens
the FE grid, atomic degrees of freedom are lost: The FE algorithm involves an
average over these degrees of freedom. Thus, to set the atomic and continuum
thermal energies on an equivalent level, the total FE thermal energy is written
again by using an offset. These corrected energies are denoted by a prime:
3 1
(E  )FE = (Na − Nm )kB T + (Ek )FE + Nm kB T , (48)
2 2
3 1
(E  )FE = (Na − Nm )kB T + EFE + Nm kB T . (49)
2 2
582 10. Multiscale Modeling of Materials

Here, Na is the number of atoms contained within an equivalent 3D volume, kB is

the Boltzmann’s constant, and equipartition has been used. Broughton et al. fur-
ther assumed that the background temperature does not vary during the simulation.
Therefore, the first terms of Eqs. (48) and (49) account for the missing degrees of
atomic freedom, while the last terms augment the 2D FE plane-strain simulation for
the missing third dimension in its degrees of freedom. Similar to the case of poten-
tial energy, these offsets do not affect the dynamics of the system and the thermal
corrections can be assigned to each mesh point in a manner similar to that described
above for the zero-temperature FE potential energy. For finite-temperature simu-
lations, the u̇ were thermalized according to the Maxwell–Boltzmann distribution
[see Eq. (9.49)]. Moreover, the appropriate elastic constants for that temperature
should be used in the FE equations of motion so as to make the MD and FE regions
seamless and compatible. Further, since this method requires a continuation of ideal
lattice sites into the FE/MD interface region so as to determine mesh coordinates,
the appropriate lattice parameter for a given temperature should be used.
The last issue to be addressed is the dissipation in the FE region. The continuum
representation of silicon used by Broughton et al. (1999) was based on linear
elasticity theory, which is a harmonic theory. Thus, vibrational modes of given
{k, ω} relationship, which depend upon the long-wavelength elastic constants of the
medium, propagate undamped. ln order to thermalize short-wavelength phonons
propagating through regions where the mesh spacing changes, and also to allow
energy to be dissipated in the FE region, Broughton et al. (1999) weakly coupled
the FE degrees of freedom to a Brownian heat bath with dynamics that was set to
the temperature at which the simulation was being performed, hence coupling the
phonon modes of the FE region. The force used in the third step of the velocity
Verlet algorithm included random and dissipative terms:
∂EFE
fFt E = + r(T , y t ) − η(y t )M t u̇t , (50)
∂ut
where r is a Gaussian random variable, and η is a friction coefficient. Through the
fluctuation-dissipation theorem, the variance σ of the Gaussian is related to η:
&
2ηM t kB T
σ = . (51)
t
ln order to minimally perturb the dynamics of the active zone (i.e., MD and TB),
η was assumed to be a function of the (time-invariant) FE mesh y-coordinate, and
was linearly increased from zero in the interface region to a finite value of about
0.1 at the extremal outer edge of the FE regions.

10.3.6 Interfacing Molecular Dynamics and Tight-Binding

Regions
In contrast to the algorithm for the FE/MD interface in which a plane between rows
of atoms was defined, the MD/TB interface is conceptually across a plane that
consists of atoms. This is necessitated due to the difficulty of assigning (localized)
energy in a computationally efficient way to specific bonds in an electronic struc-
 10.3. Fracture Dynamics 583

Figure 10.13. The MD/TB interface region.

Full lines represent two- and three-body terms
of the Stillinger–Weber potential that contribute
fully to the interface Hamiltonian, while the bro-
ken ones do not contribute at all (after Broughton
et al., 1999).

ture calculation. The MD/TB interface is shown schematically in Figure 10.13.

Since in silicon the covalent bonds are local objects, dangling bonds may be tied
off with univalent atoms (UAs), and therefore the region chosen for TB descrip-
tion is terminated with such atoms. The H and S matrices and the repulsive pair
potential Urep that couple the UAs to the silicon atoms within the interior of the
TB region were chosen to, (a) maintain electro-neutrality on both the UAs and sil-
icons; (b) locate the univalent atom potential energy minimum at a Si-Si distance,
not a Si-H distance; (c) provide a bond energy equal to a single Si-Si bond, and (d)
provide a longitudinal force constant equal to that of Si. At the perimeter of the TB
region, the UAs are constrained to stay at the Si sites of the MD region which, in
many cases, implies that more than one UA is found at a given site. Thus, there are
no matrix elements, nor Urep terms that couple any of the UAs to one another. In the
computations, a circle is drawn around an inner set of atoms and are designated TB
silicons. Then, any atom outside this circle, but within range of an inner atom, is
designated as a UA. The range that
√ was used as the criterion was the average of the
first (r0 ) and second neighbor [ 8/3r0 ] distances of the equilibrium Si lattice. In
Figure 10.13, matrix elements that couple atoms across the light gray region are of
Si-Si form. Atoms coupled across the dark gray region use Si-UA matrix elements
for which the necessary parameters are given by Bernstein and Kaxiras (1997).
The TB H and S generalized-eigenvalue problem is solved for the entire Si plus
UA system. The only remaining issue is to determine which SW two- and three-
body terms are required to couple the UAs to the MD region. Since these atoms
are not coupled to one another in the TB region, SW terms that account for such
are required. Broughton et al. (1999) included all SW pair terms between a UA
and Si and either a Si atom in the MD region or another UA-Si, as well as all SW
triplets that include at least one MD Si to one UA-Si pair. Thus, the forces that
arise on the UA-Si from these terms are added to the forces arising from the TB
Hamiltonian on these atoms. Equation (30) should more correctly be written as
H = HFE ({u, u̇} ∈ FE) + HFE/MD ({u, u̇, r, ṙ} ∈ FE/MD)
+HMD ({r, ṙ} ∈ MD) + HMD/TB ({r, ṙ} ∈ MD/TB) + HTB ({r, ṙ} ∈ MD/TB),
(52)
584 10. Multiscale Modeling of Materials

Figure 10.14. Overlapping TB regions em-

Outer Perimeter bedded in the MD region. Atomic forces are a
of Silicons MD Region function of the overlap (after Broughton et al.,
1999).

TB

TB TB

where the penultimate term involves only SW interactions crossing the boundary,
while the last term involves a TB calculation for the combined Si-UA system.
The above prescription produces a conservative Hamiltonian, if there is no
dynamic allocation of the TB region to those parts of the material where atomic
bonds are breaking. Unfortunately, for many materials and phenomena (such as
crack propagation) the 100 or so atoms, whose forces may currently be updated
using a non-orthogonal TB Hamiltonian in one second of real time, do not comprise
a large region. This problem may be partially addressed by using periodic boundary
conditions. One may also address the problem by using more than one processor
per TB region to perform the diagonalization, but such algorithms are presently
not efficient on coarse-grained scalable architecture computers. Broughton et al.
(1999) represented the region of breaking bonds by a region of TB segments, an
example of which is shown in Figure 10.14 for three overlapping TB regions. In
their simulations of fracture propagation (see below), Broughton et al. used eight
overlapping regions, with each region diagonalized separately, and also handled
by a separate processor. After forces on each atom are obtained for each TB region
separately, the force to be used in updating the velocity Verlet algorithm is obtained
as the average over the different regions; if there is no overlap of TB regions,
the same prescription as for a single TB region is used; where a UA of one TB
region overlaps a Si of another, the Si value is used. The number of atoms that are
propagated using TB forces is therefore less than the total number within all the
overlapping regions. These rules are, of course, intuitive.
In the simulation of fracture propagation, the energy and force algorithm, as
implemented in the MD and TB regions, proceeds by calculating the SW energy for
all atoms in the MD processors. The TB processors calculate not only TB energies
and forces, but also those SW forces that must be subtracted from those double
counted in MD processors. Thus, SW energies, by suitable division of two- and
three-body terms, are available for all atoms, which are then used to discriminate
different regions. The apex of a crack is found, for example, by locating the atom,
with a potential energy greater (more positive) than 60% of the bulk cohesive
potential energy, furthest to the left or right of the center of the system. The central
TB portion of the overlap region is then placed at that atom.
 10.3. Fracture Dynamics 585

10.3.7 Seamless Simulation

The description of the multiscale method indicates how the simulation is made
seamless. The TB region, the region described at the smallest length scale, de-
termines the elastic constants and the atomic force fields used elsewhere in the
system. Therefore, the procedure used by Broughton et al. (1999) is as follows.
(1) A pure TB simulation is carried out for a small number of atoms that represent
the bulk material (at given temperature and pressure). By appropriate defor-
mation of the computational cell, the elastic constants, and hence the force
constant, are computed.
(2) The SW parameters for Si are then adjusted to reproduce the same estimates.
(3) The elastic constants from the TB region are also utilized for the stiffness
matrix of the FE region.
(4) Finally, the parameters used for the Si-univalent atom matrix elements are
adjusted so that displacement of a univalent atom-Si in the coupled system
gives rise to the same force constant of the pure bulk material.

10.3.8 Multiscale Simulation of Fracture Propagation in Silicon

The foregoing multiscale algorithm was utilized by Broughton et al. (1999) to study
rapid brittle fracture of a Si slab “damaged” by a microcrack at its center and de-
formed under uniaxial tension. The MD region was spatially domain-decomposed
onto 24 processors. Each FE region was handled by its own processor. The path
of the fracture was monitored and the center of the TB region was placed at the
apex of the crack where bond breaking occurs. This region is crucial to deter-
mining the kinetics of the crack propagation. A region of eight overlapping TB
regions, each being a cylinder of radius 5.3 Å in the yz plane and distributed to
a different processor, was used. The exposed notch faces were x − z planes with
(100) faces, with the microcrack pointed in the (010) direction. Each FE region
contained 258,048 mesh points, there were 1,032,192 atoms in the MD region, and
around 280 unique atoms in the TB region. The lengths of the MD region were
10.9 Å (the slab thickness and periodic), 521 Å (before the pull, in the direction
of the pull), and 3, 649 Å (the primary direction of propagation and periodic). The
full pull length of the FE+MD system was 5, 602 Å. The entire system, including
the FE, contained 11,093,376 atoms. The time for a TB force update was 1.5 s,
1.85 s for the MD update, and 0.7 s for the FE.
The rectilinear computational cell comprised (100) faces on all sides. The re-
duced elastic constant matrix for this geometry was obtained by averaging the
results reported by Balamane et al. (1992) and Ray (1988) for the zero temperature
C11 , C12 , and C44 elastic constants of SW Si:
⎛ ⎞
1.578 × 106 0.7930 × 106 0.0
⎜ ⎟
C = ⎝ 0.7930 × 106 1.578 × 106 0.0 ⎠, (53)
0.0 0.0 0.6365 × 10 6

where the units are megabar.

586 10. Multiscale Modeling of Materials

The slab was initialially at zero temperature, and a constant strain rate was
imposed on the outermost FE boundaries defining the opposing horizontal faces
of the slab. Moreover, a linear velocity gradient was applied within the slab, which
resulted in an increasing internal strain with time. The material failed at the notch
tip after it had been stretched by about 1.5%. The propagating cracks rapidly
achieved a limiting speed (2770 m/s) equal to 85% of the Rayleigh speed cR , the
sound speed of the solid Si surface. A very accurate signature of seamless coupling,
one that represents a validation of the method, is the fact that stress waves passed
from the MD region to the FE regions with no visible reflection at the FE/MD
interface, i.e., the coupling of the MD region to the FE region appeared seamless.
Moreover, there were no obvious discontinuities at the MD/TB interface.

10.4 Other Applications of Multiscale Modeling

In addition to the above examples, over the past few years the multiscale method-
ology has been utilized for describing several other important sets of phenomena.
In this section, we briefly describe two of such applications. Complete details of
these works are given in the original papers.

10.4.1 Atomistically Induced Stress Distributions in Composite

Materials
With the rapid technological advances of the past decade, the semiconductor device
feature size has been decreasing. The feature size is now predicted to reach 70 nm by
the year 2008, if not sooner. Because materials of such sizes have a high surface-to-
volume ratios, the stress inhomogeneities that are caused by surfaces and interfaces
play an important role in the performance of the devices, since they affect the donor
distribution by trapping donors in tensile stress regions. Thus, it is important to be
able to control the growth of the materials in order for them to have the desired
properties. This can be aided by accurate models that describe the materials and
the stress and strain distributions in them. Although several continuum models
have been developed for this purpose (see, for example, Johnson and Freund,
1997), they appear to be incapable of describing atomistically-induced stresses at
the interface of two different materials, especially if one or both of the materials
are amorphous. A better approach would be based on a multiscale model that can
address the atomistic aspects of the problem near the interface, and the macroscopic
aspects deep into the substrate.
Several of such multiscale models have been developed recently. Of particular
note is the multiscale model that was developed by Lidorikis et al. (2001) who
used a hybrid FE/MD approach to study stress domains in Si nanopixels cov-
ered with amorphous Si3 N4 films. The stress domains originate from the atomic
configuration at the amorphous/crystalline interface and extend throughout the
substrate, where the semiconductor device operates. Lidorikis et al. (2001) simu-
lated a 25 × 25 nm Si(111) mesa, covered with a 25 × 25 × 5 nm Si3 N4 film, on
 10.4. Other Applications of Multiscale Modeling 587

a 50 × 50 × 15 nm Si(111) substrate. The lattice mismatch between Si and Si3 N4

is of the order of 1.2%. The Si3 N4 film, as well as its interface with Si, was treated
by the MD simulations, while most of the Si substrate was modeled by FE simula-
tions. Both crystalline (001) and amorphous Si3 N4 films were analyzed. Periodic
boundary conditions were applied on the sides of the substrate, while its bottom
was held fixed.
To model Si, the SW potential was used, while for Si3 N4 a combination of two-
and three-body interaction terms, that included electronic polarizability, charge
transfer (taken to be screened Coulomb potential), and covalent bonding effects,
was utilized. A variation of the same potential was also used for describing the
interactions across the Si/Si3 N4 interface. In the Si substrate, about 20 Å below
the interface, the deformations are small and thus were treated by linear continuum
elasticity, solved on a computational grid. The displacement field was discretized
at the nodes of the grids and, within the grid cells, was interpolated from the nodal
values.
The MD and FE regions were then merged seamlessly in the interface region
between the two. To do this, the FE mesh was refined down to atomic scale,
and was also shifted from the simple-cubic arrangement (used in the continuum
region), in order to follow the lattice structure of Si. Within the interface region, the
FE and crystalline lattices overlap, hence yielding a one-to-one correspondence
between the MD atoms and the FE nodes. For these hybrid atom/node particles,
an average Hamiltonian was defined, and the total Hamiltonian was given by an
equation similar to (52) (except that, of course, there was no TB region). The Verlet
velocity algorithm, described in Chapter 9, was used for integrating the equations
of motion in the MD region.
The hybrid simulations for the Si pixels covered with an amorphous Si3 N4 film
indicated that, the inhomogeneous stress domains are formed below the interface
with a shadow that extends deep into the substrate. These domains are caused
by atomistically-induced stresses at the lattice-mismatched amorphous/crystalline
interface, and cannot be predicted by a continuum model alone. The chemistry
of the interface and the degree of mismatch at the interface control the size of
the domains. These simulations were compared with those in which only the MD
method with several million atoms were utilized; the agreement was found to be
excellent, hence validating the multiscale model. The computational time of the
multiscale model was of course much less than that of the MD simulation alone.

10.4.2 Chemical Vapor Deposition

Chemical vapor deposition (CVD) is a complex operation that involves several
key processes which occur on widely separated length and time scales. At each of
the distinct length scales, a particular set of modeling techniques and assumptions
is appropriate. One reason for the complexity of CVD is that, frequently, processes
of different length scales interact, and thus many problems of interest cannot be
cast within the context of a single scale. An appropriate model of CVD should
therefore focus on interactions between macroscopic flow and transport, where
588 10. Multiscale Modeling of Materials

the gas phase is regarded as a continuum, and phenomena at the micron length
scale, where the discrete nature of the gas phase becomes apparent and important.
For most CVD processes, flow and transport in the reactor at large are described
by the usual macroscopic conservation equations for momentum, mass, and energy
which give rise to coupled nonlinear partial differential equations that are typically
solved with the FE method. However, these continuum models are not valid as
length scales approach and shrink below the gas phase mean-free path λ. This is
an important consideration since microelectronic device features are frequently
submicron, whereas the mean-free path λ, under low pressure CVD conditions,
can be several hundred µm. This limitation on the use of continuum models is
widely recognized and has led to the development of discrete particle transport
models.
Rodgers and Jensen (1998) developed a multiscale model for CVD on length
scales ranging from microns to meters. At the macroscopic level the problems
of fluid flow and heat and mass transfer in a single wafer, low-pressure CVD
reactor were solved using the FE method. Since on the feature scale the continuum
models are no longer valid, transport at the feature scale was linked with the
continuum model by an effective reactivity function R that included effects of
both the multiscale surface heterogeneity and microscopic transport resistance.
A Monte Carlo method was then used for computing R for any set of reaction
pathways that occur over microelectronic device features of any geometry. Feature
scale computations were then combined to yield an effective reactivity map over
the surface of the substrate, which was subsequently utilized for formulating a flux
boundary condition for the continuum model. Iteration between macroscopic and
microscopic models was then used to ensure a consistent set of conditions at the
micro-macro interface.
For a different application of multiscale modeling methodology to problems
involving transport and reaction processes in nano- and microporous materials,
see Dadvar and Sahimi (2002, 2003).

Summary
The goal in multiscale modeling is to predict the performance and behavior of
heterogeneous materials in which there are several relevant and widely disparate
length scales. Due to the great complexities that are involved, the task of devel-
oping a multiscale model is usually referred to as a grand challenge problem. The
complexities that are involved are clear: At the atomic scale electrons govern the
interactions among atoms in a material, and thus quantum-mechanical effects are
important and must be taken into account. At the same time, at the macroscopic or
engineering scale, forces that arise from macroscopic stresses and/or temperature
gradients are the factors that control the performance of materials. At the interme-
diate length scales, defects such as dislocations control the mechanical properties
of materials up to tens of micrometers, while large collections of such defects,
including grain boundaries, govern their mesoscopic properties up to length scales
 10.4. Other Applications of Multiscale Modeling 589

that are of the order of hundreds of micrometers. Coupling of all of these length
scales would not have been possible, had it not been for the tremendous advances in
the advent of ever more powerful, massively-parallel computers, and the great ad-
vances that have been made in the theoretical understanding of materials and their
properties. The emerging multiscale methodology demonstrates how coupling of
atomistic and continuum approaches results in more predictive power than either
approach offers alone.
An interesting and very useful aspect of multiscale modeling is the fact that, it
is a multidisciplinary field that brings together scientists from many disciplines.
The development of a multiscale model of any phenomenon that deals with one or
more aspect of materials’ properties should, in principle at least, involve chemists
and chemical engineers, applied physicists and mathematicians, and continuum
mechanicians.
References

Abell, G.C., “Empirical chemical pseudopotential theory of molecular and metallic

bonding,” Phys. Rev. B 31, 6184 (1985).
Abraham, F.F., “Computational statistical mechanics methodology, applications and
supercomputing,” Adv. Phys. 35, 1 (1986).
Abraham, F.F., “Dynamics of brittle fracture with variable elasticity,” Phys. Rev. Lett. 77,
869 (1996).
Abraham, F.F., D. Brodbeck, R.A. Rafey, and W.E. Rudge, “Instability dynamics of fracture:
A computer simulation investigation,” Phys. Rev. Lett. 73, 272 (1994).
Abraham, F.F., D. Brodbeck, W.E. Rudge, and X. Xu, “A molecular dynamics investigation
of rapid fracture mechanics,” J. Mech. Phys. Solids 45, 1595 (1997a).
Abraham, F.F., J.Q. Broughton, N. Bernstein, and E. Kaxiras, “Spanning the length scales
in dynamic simulation,” Comput. Phys. 12, 538 (1998a).
Abraham, F.F., J.Q. Broughton, N. Bernstein, and E. Kaxiras, “Spanning the continuum to
quantum length scales in a dynamic simulation of brittle fracture,” Europhys. Lett. 44,
783 (1998b).
Abraham, F.F., and H. Gao, “How fast can cracks propagate?” Phys. Rev. Lett. 84, 3113
(2000).
Abraham, F.F., W.E. Rudge, D.J. Auerbach, and S.W. Koch, “Molecular dynamics simu-
lations of the incommensurate phase of Krypton on graphite using more than 100,000
atoms,” Phys. Rev. Lett. 52, 445 (1984).
Abraham, F.F., D. Schneider, B. Land, D. Lifka, J. Skovira, J. Gerner, and M. Rosenkrantz,
“Instability dynamics in three-dimensional fracture: An atomistic simulation,” J. Mech.
Phys. Solids 45, 1461 (1997b).
Adda-Bedia, M., and M. Ben Amar, “Stability of Quasiequillibrium cracks under uniaxial
loading,” Phys. Rev. Lett. 76, 1497 (1996).
Adda-Bedia, M., and Y. Pomeau, “Crack instabilities in a heated glass strip,” Phys. Rev. E
52, 4105 (1995).
Aharony, A., “Universal critical amplitude ratios for percolation,” Phys. Rev. B 22, 400
(1980).
Aharony, A., “Crossover from linear to nonlinear resistance near percolation,” Phys. Rev.
Lett. 58, 2726 (1987).
Aktsipetrov, A.A., O. Keller, K. Pedersen,A.A. Nikulin, N.N. Novikova, andA.A. Fedyanin,
“Surface-enhanced second-harmonic generation in C60 -coated silver island films,” Phys.
Lett. A 179, 149 (1993).
Alder, B.J., and T.E. Wainwright, “Phase transitions for a hard sphere system,” J. Chem.
Phys. 27, 1208 (1957).
Alder, B.J., and T.E. Wainwright, “Decay of velocity autocorrelation function,” Phys. Rev.
A 1, 18 (1969).
592 References

Allen, M.P., and D.J. Tildesley, Computer Simulation of Liquids (Oxford University Press,
London, 1987).
Andersen, H.C., “Molecular dynamics simulations at constant pressure and/or temperature,”
J. Chem. Phys. 72, 2384 (1980).
Andersen, J.V., D. Sornette, and K.-t. Leung, “Tricritical behavior in rupture induced by
disorder,” Phys. Rev. Lett. 78, 2140 (1997).
Andrews, D.J., “Rupture velocity of plane strain shear cracks,” J. Geophys. Res. 81, 5679
(1976).
Anthony, J.B.C.S.R., J.B. Chubb, and J. Congleton, “The crack-branching velocity,” Philos.
Mag. 22, 1201 (1970).
Arakawa, K., and K. Takahashi, “Branching of a fast crack in polymers,” Int. J. Fract. 48,
245 (1991).
Aranson, I.S., V.A. Kalatsky, and V.M. Vinokur, “Continuum field description of crack
propagation,” Phys. Rev. Lett. 85, 118 (2000).
Arbabi, S., and M. Sahimi, “Elastic properties of three-dimensional percolation networks
with stretching and bond-bending forces,” Phys. Rev. B 38, 7173 (1988).
Arbabi, S., and M. Sahimi, “On three-dimensional elastic percolation networks with bond-
bending forces,” J. Phys. A 23, 2211 (1990a).
Arbabi, S., and M. Sahimi, “Test of universality for three-dimensional models of mechanical
breakdown in disordered solids,” Phys. Rev. B 41, 772 (1990b).
Archuleta, R.J., “Analysis og near-source static and dynamic measurements from the 1979
Imperial Valley earthquake,” Bull. Seismol. Soc. Am. 72, 1927 (1982).
Argon, A.S., “Brittle to ductile transition in cleavage fracture,” Acta Mettal. 35, 185 (1987).
Arias, T.A., and J.D. Joannopoulos, “Ab initio theory of dislocation interactions: From close-
range spontaneous annihilation to the long-range continuum limit,” Phys. Rev. Lett. 73,
680 (1994).
Ashcroft, N., and N.D. Mermin, Solid State Physics (Saunders, London, 1976).
Ashurst, W.T., and W.G. Hoover, “Microscopic fracture studies in the two-dimensional
triangular lattice,” Phys. Rev. B 14, 1465 (1976).
Åström, J.A., J.P. Mäkinen, M.J. Alva, and J. Timonen, “Elasticity of Poissonian fiber
networks,” Phys. Rev. E 61, 5550 (2000).
Åström, J.A., S. Saarinen, K.J. Niskanen, and J. Kurkijärvi, “Microscopic mechanics of
fiber networks,” J. Appl. Phys. 75, 2383 (1994).
Åström, J.A., and J. Timonen, “Crack bifurcations in a strained lattice,” Phys. Rev. B 54,
R9585 (1996).
Åström, J.A., and J. Timonen, “Fragmentation by crack branching,” Phys. Rev. Lett. 78,
3677 (1997a).
Åström, J.A., and J. Timonen, “Fracture of a brittle membrane,” Phys. Rev. Lett. 79, 3684
(1997b).
Auerbach, D.J., W. Paul, A.F. Bakker, C. Lutz, W.E. Rudge, and F.F. Abraham, “A special
purpose parallel computer for molecular dynamics: Motivation, design, implementation,
and application,” J. Phys. Chem. 91, 4881 (1987).
Baker, T.C., and F.W. Preston, “The effect of water on the strength of glasses,” J. Appl.
Phys. 17, 179 (1946).
Bakker, A.F., G.H. Gilmer, M.H. Grabow, and K. Thompson, “A special purpose computer
for molecular dynamics calculations,” J. Comput. Phys. 90, 313 (1990).
Balamane, H., T. Halicioglu and W. A. Tiller, “Comparative study of silicon empirical
interatomic potentials,” Phys. Rev. B 46, 2250 (1992).
 References 593

Ball, R.C., and H. Larralde, "Three-dimensional stability analysis of planar straight cracks
propagating quasistatically under type I loading," Int. J. Fract. 71, 365 (1995).
Bao, G., J.W. Hutchinson, and R.M. McMeeking, “Particle reinforcement of ductile matrices
against plastic flow and creep,” Acta Metall. Mater. 39, 1871 (1991).
Barabási, A.-L., and H.E. Stanley, Fractal Concepts in Surface Growth (Cambridge
University Press, Cambridge, 1995).
Barabási, A.-L., and T. Vicsek, “Multifractality of self-affine surfaces,” Phys. Rev. A 44,
2730 (1990).
Baran, G.R., C. Roques-Carmes, D. Wehbi, and M. Degrange, “Fractal characteristics of
fracture surfaces,” J. Am. Ceram. Soc. 75, 2687 (1992).
Barber, M., J. Donley, and J.S. Langer, “Steady-state propagation of a crack in a viscoelastic
strip,” Phys. Rev. A 40, 366 (1989).
Barbosa, F.F., and S.L.A. de Queiroz, “Concentration anisotropy and directionality in the
dielectric breakdown problem on a square lattice,” J. Phys.: Condens. Matter 1, 2771
(1989).
Barclay, A.L., P.J.J. Sweeney, L.A. Dissado, and G.C. Stevens, “Stochastic modelling of
electrical treeing: fractal and statistical characteristics,” J. Phys. D 23, 1536 (1990).
Bardhan, K.K., “Nonlinear conduction in composites above percolation threshold-beyond
the backbone,” Physica A 241, 267 (1997).
Barenblatt, G.I., “The formation of equilibrium cracks during brittle fracture: general ideas
and hypothesis, axially symmetric cracks,” Appl. Math. Mech. (Translation of PMM) 23,
622 (1959a).
Barenblatt, G.I., “Concerning equilibrium cracks forming during battle fracture: the stability
of isolated cracks, relationship with energetic theories,” Appl. Math. Mech. (Translation
of PMM) 23, 1273 (1959b).
Barton, C.C., “Fractal analysis of the spatial clustering of fractures,” in Fractals and
Their Use in the Earth Sciences, edited by C.C. Barton and P.R. La Pointe (American
Geophysical Union, 1992), p. 126.
Barton, C.C., and P.A. Hsieh, Physical and Hydrological-Flow Properties of Fractures,
Guidebook T385 (American Geophysical Union, Las Vegas, Nevada, 1989).
Barton, C.C., and E. Larsen, “Fractal geometry of two-dimensional fracture networks at
Yucca Mountain, southwest Nevada,” in Proceedings of the International Symposium
on Fundamentals of Rock Joints, edited by O. Stephenson (Bjorkliden, Sweden, 1985),
p. 77.
Barton, C.C., T.A. Schutter, W.R. Page, and J.K. Samuel, “Computer generation of fracture
networks for hydrologic-flow modelling,” Trans. Amer. Geophys. Union 68, 1295 (1987).
Barton, J.L., “La relaxation diélectrique de quelques verres ternaires silice oxyde alcalin
oxyde alcalin-terreux,” Verres et Refr. 20, 328 (1966).
Baskes, M.I., J.S. Nelson, and A.F. Wright, “Semiempirical modified embedded-atom
potentials for silicon and germanium,” Phys. Rev. B 40, 6085 (1989).
Batrouni, G.G., and A. Hansen, “Fracture in three-dimensional fuse model,” Phys. Rev. Lett.
80, 325 (1998).
Bazant, M.Z., E. Kaxiras, and J.F. Justo, “Environment-dependent interatomic potential for
bulk silicon,” Phys. Rev B 56, 8542, (1997).
Beale, P.D., and P.M. Duxbury, “Theory of dielectric breakdown in metal-loaded
dielectrics,” Phys. Rev. B 37, 2785 (1988).
Beale, P.D., and D.J. Srolovitz, “Elastic fracture in random materials,” Phys. Rev. B 37,
5500 (1988).
594 References

Beazley, D.M., and P.S. Lomdahl, “Message-passing multi-cell molecular dynamics on the
Connection Machine CM-5,” Parallel Computing 20, 173 (1994).
Beazley, D.M., P.S. Lomdahl, N. Grφnbech-Jensen, R. Giles, and P. Tamayo, “Parallel
algorithms for short-range molecular dynamics,” Annual Rev. Comput. Phys. 3, 116
(1995).
Benguigui, L., “Simulation of dielectric failure by means of resistor-diode random lattices,”
Phys. Rev. B 38, 7211 (1988).
Benguigui, L., and P. Ron, “Laboratory simulation of dielectric breakdown,” in Non-
Linearity and Breakdown in Soft Condensed Matter, edited by K.K. Bardhan, B.K.
Chakrabarti, and A. Hansen, Lecture Notes in Physics 437 (Springer, Heidelberg, 1994),
p. 221.
Benguigui, L., P. Ron, and D.J. Bergman, “Strain and stress at the fracture of percolative
media,” J. Physique 48, 1547 (1987).
Bennett, C.H., “Efficient estimation of free energy differences from Monte Carlo data,” J.
Comput. Phys. 22, 245 (1976).
Beran, M.J., “Use of the variational approach to determine bounds for the effective
permittivity of random media,” Nuovo Cimento 38, 771 (1965).
Beran, M.J., Statistical Continuum Theories (Wiley, New York, 1968).
Bergkvist, H., “Some experiments on crack motion and arrest in polyethylmethacrylate,”
Eng. Fract. Mech. 6, 621 (1974).
Bergman, D.J., in Fragmentation, Form and Flow in Fractured Media, Ann. Israel Phys.
Soc. 8, 266 (1986).
Bergamn, D.J., “Nonlinear behavior and 1/f noise near a conductivity threshold: effects of
local microgeometry,” Phys. Rev. B 39, 4598 (1989).
Bergman, D.J., and Y. Kantor, “Critical properties of an elastic fractal,” Phys. Rev. Lett. 53,
511 (1984).
Bergman, D.J., O. Levy, and D. Stroud, “Theory of optical bistability in a weakly nonlinear
composite medium,” Phys. Rev. B 49, 129 (1994).
Bernstein, N., and E. Kaxiras, “Nonorthogonal tight-binding Hamiltonians for defects and
interfaces in silicon,” Phys. Rev. B 56, (1997).
Berry, M.V., “Regular and irregular semiclassical wavefunctions,” J. Phys. A 10, 2083
(1977).
Berveiller, M., and A. Zaoui, “An extension of the self-consistent scheme to plastically-
flowing polycrystals,” J. Mech. Phys. Solids 26, 325 (1979).
Bhattacharya, K., and R.V. Kohn, “Elastic energy minimization and recoverable strains of
polycrystalline shape-memory materials,” Arch. Rational Mech. Anal., (1997).
Bird, R.B., R.C. Armstrong, and O. Hasseger, Dynamics of Polymeric Liquids, 2nd ed.
(Wiley, New York, 1987).
Bishop, J.F.W., and R. Hill, “A theory of the plastic distortion of a polycrystalline under
combined stresses,” Philos. Mag. 42, 414 (1951a).
Bishop, J.F.W., and R. Hill, “A theoretical derivation of the plastic properties of a
polycrystalline face-center metal,” Philos. Mag. 42, 1298 (1951b).
Blumberg Selinger, R.L., Z.-G. Wang, W.M. Gelbart, and A. Ben-Shaul, “Statistical-
theormodynamic approach to fracture,” Phys. Rev. A 43, 4396 (1991a).
Blumberg Selinger, R.L., Z.-G. Wang, and W.M. Gelbart, “Effect of temperature and small-
scale defects on the strength of solids,” J. Chem. Phys. 95, 9128 (1991b).
Blumenfeld, R., and A. Aharony, “Nonlinear resistor fractal networks, topological distances,
singly connected bonds and fluctuations,” J. Phys. A 18, L443 (1985).
 References 595

Blumenfeld, R., and D.J. Bergman, “Comment on Nonlinear susceptibilities of granular

matter,” Phys. Rev. B 43, 13682 (1991a).
Blumenfeld, R., and D.J. Bergman, “Strongly nonlinear composite dielectrics: a perturbation
method for finding the potential field and bulk effective properties,” Phys. Rev. B 44,
7378 (1991b).
Blumenfeld, R., Y. Meir, A. B. Harris, and A. Aharony, “Infinite set of exponents describing
physics on fractal networks,” J. Phys. A 19, L791 (1986).
Bolding, B.C., and H.C. Andersen, “Interatomic potential for silicon clusters, crystals and
surfaces,” Phys. Rev. B 41, 10568 (1989).
Born, M., “Thermodynamics of crystals and melting,” J. Chem. Phys. 7, 591 (1939).
Bornert, M. Morphologie Microstructurale et Comportement Mécanique: Caractérisations
Expérimentales, Approches par Bornes et Estimations Autocohérentes Généralisées.,
Ph.D. Thesis, Ecole Nationale des Ponts et Chaussées (1996).
Bornert, M., C. Stolz, and A. Zaoui, “Morphologically representative pattern-based
bounding in elasticity,” J. Mech. Phys. Solids 44, 307 (1996).
Bouchaud, E., and J.-P. Bouchaud, “Fracture surfaces: apparent roughness, relevant length
scales, and fracture toughness,” Phys. Rev. B 50, 17752 (1994).
Bouchaud, J.-P., E. Bouchaud, G. Lapasset, and J. Planes, “Models of fractal cracks,” Phys.
Rev. Lett. 71, 2240 (1993).
Bouchaud, E., J.-P. Bouchaud, J. Planes, and G. Lapasset, “The statistics of crack branching
during fast crack propagation,” Fractals 1, 1051 (1993a).
Bouchaud, E., L. de Arcangelis, G. Lapasset, and J. Planes, “Fractals breakage mater,” La
Recherche (Paris) 22, 808 (1991).
Bouchaud, E., G. Lapasset, and J. Planes, “Fractal dimension of fractured surfaces: A
universal value?” Europhys. Lett. 13, 73 (1990).
Bouchaud, E., G. Lapasset, J. Planes, and S. Navéos, “Statistics of branched fracture
surfaces,” Phys. Rev. B 48, 2917 (1993b).
Bouchaud, E., and S. Navéos, “From quasi-static to rapid fracture,” J. Phys. I France 5,
547 (1995).
Bouchitte, G., and P. Suquet, “Homogenization, plasticity and yield design,” in: Compos-
ite Media and Homogenization Theory, edited by G. Dal Maso and G. Dell’Antonio
(Birkhäuser, Basel, 1991), p. 107.
Boudet, J.F., S. Ciliberto, and V. Steinberg, “Experimental study of the instability of crack
propagation in brittle materials,” Europhys. Lett. 30, 337 (1995).
Boudet, J.F., S. Ciliberto, and V. Steinberg, “Dynamics of crack propagation in brittle
materials,” J. Phys II France 6, 1493 (1996).
Bowman, D.R., and D. Stroud, “Model for dielectric breakdown in metal-insulator
composites,” Phys. Rev. B 40, 4641 (1989).
Boyd, R.W., Nonlinear Optics (Academic Press, New York, 1992).
Brace, W.S., and A.S. Orange, “Further studies of the effects of pressure on electrical
resistivity of rocks,” J. Geophys. Res. 73, 1433 (1968).
Bradley, R.M., and K. Wu, “Studies of titanium dioxide film growth from titanium
tetraisopropoxide,” J. Phys. A 26, 327 (1993).
Brańka, A.C., and K.W. Wojciechowski, “Generalization of Nosé and Nosé-Hoover
isothermal dynamics,” Phys. Rev. E 62, 3281 (2000).
Brass, A., H.H. Jensen, and A.J. Berlinsky, “Models of flux pinning in the quasistatic limit”,
Phys. Rev. B 39,102 (1989).
Brenner, D.W., “Empirical potential for hydrocarbons for use in simulating the chemical
vapor deposition of diamond films,” Phys. Rev. B 42, 9458 (1990).
596 References

Brickstad, B., and F. Nilsson, “Numerical evaluation by FEM of crack propagation

experiments,” Int. J. Fract. 16, 71 (1980).
Broberg, K.B., “On the behaviour of the process region at a fast running crack tip,” in High
Velocity Deformation of Solids, edited by K. Kawata and J. Shiori (Springer, Berlin,
1979), pp. 182–193.
Broberg, K.B., “The near-tip field at high crack velocities,” Int. J. Fract. 39, 1 (1989).
Brockenborough, J., S. Suresh, and H. Wienecke, “Deformation of metal-matrix composites
with continuous fibers: Geometrical effects of fiber distribution and shape,” Acta Metall.
Mater. 39, 735 (1991).
Brouers, F., S. Blacher, A.N. Lagarkov, A.K. Sarychev, P. Gadenne, and V.M. Shalaev,
“Theory of giant Raman scattering from semicontinuous metal films,” Phys. Rev. B 55,
13234 (1997).
Brouers, F., S. Blacher, and A.K. Sarychev, “Giant field fluctuations and anomalous light
scattering from semicontinuous metal films,” Phys. Rev. B 58, 15897 (1998).
Brouers, F., J.P. Clerc, G. Giraud, J.M. Laugier and Z.A. Randriamantany, “Dielectric and
optical properties close to the percolation threshold. II,” Phys. Rev. B 47, 666 (1993).
Broughton, J.Q., F.F. Abraham, N. Bernstein, and E. Kaxiras, “Cocurrent coupling of length
scales: Methodology and application,” Phys. Rev. B 60, 2391 (1999).
Bruge, F., and S.L. Fornili, “A distributed dynamic load balancer and its implementation on
multi-transputer systems for molecular dynamics simulation,” Comput. Phys. Commun.
60, 39 (1990).
Budiansky, B., “A reassessment of deformation theories of plasticity,” J. Appl. Mech. 26,
259 (1959).
Budiansky, B., “On the elastic moduli of some heterogeneous materials,” J. Mech. Phys.
Solids 13, 223 (1965).
Bulatov, V.V., S. Yip, and A.S. Argon, “Atomic models of dislocation mobility in silicon,”
Philos. Mag. 72, 452 (1995).
Buresch, F.E., K. Frye, and T. Muller, in Fracture Mechanics of Ceramics, Vol. 5, edited by
R.C. Bradt, A.G. Evans, D.P.H. Hasselman, and F.F. Lange (Plenum, New York, 1983),
p. 591.
Burns, S.J., and W.W. Webb, “Fracture surface energies and dislocation processes during
dynamical cleavage of LiF. II. Experiments,” J. Appl. Phys. 41, 2086 (1970).
Burridge, R., G. Conn, and L.B. Freund, “The stability of a rapid mode II shear crack with
finite cohesive traction,” J. Geophys. Res. 84, 2210 (1979).
Cagin, T., and B.M. Pettitt, “Molecular dynamics with a variable number of molecules,”
Mol. Simul. 6, 5 (1991).
Caldarelli, G., C. Castellano, and A. Vespignani, “Fractal and topological properties of
directed fractures,” Phys. Rev. E 49, 2673 (1994).
Car, R., and M. Parrinello, “Unified approach for molecular dynamics and density-functional
theory,” Phys. Rev. Lett. 55, 2471 (1985).
Carlsson, J., L. Dahlberg, and F. Nilsson, “Experimental studies of the unstable phase of
crack propagation in metal and polymers,” in Dynamic Crack Propagation, edited by
G.C. Sih (Noordhoof, Leydon 1972), p. 3.
Ceperley, D.M., “Ground state of fermion one-component plasma—A Monte Carlo study
in two and three dimensions,” Phys. Rev. B 18, 3126 (1978).
Chakrabarti, B.K., K.K. Bardhan, and R. Ray, “Insulation breakdown in a random
nonpercolating network of conductors,” J. Phys. C 20, L57 (1987).
Chakrabarti, B.K., and L.G. Benguigui, Statistical Physics of Fracture and Breakdown in
Disordered Systems (Oxford University Press, London, 1997).
 References 597

Chakrabarti, B.K., D. Chowdhury, and D. Stauffer, “Molecular dynamic study of fracture

in 2D disordered elastic Lennard–Jones solids,” Z. Phys. B 62, 343 (1986).
Chakrabarti, B.K., A.K. Roy, and S.S. Manna, “Breakdown exponents in lattice and
continuum percolation,” J. Phys. C 21, L65 (1988).
Chan, D.Y.C., B.D. Hughes, L. Paterson, and C. Sirakoff, “Resistor networks with
distributed breakdown voltages,” Phys. Rev. A 43, 2905 (1991).
Chan, K.S., (ed.), George R. Irwin Symposium on Cleavage Fracture (The Minerals, Metals
& Materials Society, New York, 1997).
Chayes, J.T., L. Chayes, and R. Durret, “Critical behavior of the two-dimensional first
passage time,” J. Stat. Phys. 45, 933 (1986).
Che, J., T. Cagin, and W.A. Goddard III, “Generalized extended empirical bond-order
dependent force fields including nonbond interactions,” Theor. Chem. Acc. 102, 346
(1999).
Chelikowsky, J.R., J.C. Phillips, M. Kamal, and M. Strauss, “Surface and thermodynamic
interatomic force fields for silicon clusters and bulk phases,” Phys. Rev. Lett. 62, 292
(1989).
Chermant, J.L., and M. Coster, “Quantitative fractography,” J. Mater. Sci. 14, 509 (1979).
Cheung, K.S., A.S. Argon, and S. Yip, “Activation analysis of dislocation nucleation from
crack tip in α-Fe,” J. Appl. Phys. 69, 2088 (1991).
Cheung, K.S., and S. Yip, “Brittle-ductile transition in intrinsic fracture behavior of
crystals,” Phys. Rev. Lett. 65, 2804 (1990).
Ching, E.S.C., “Dynamic stresses at a moving crack tip in a model of fracture propagation,”
Phys. Rev. E 49, 3382 (1994).
Ching, E.S.C., J.S. Langer, and H. Nakanishi, “Linear stability analysis for propagating
fracture,” Phys. Rev. E 53, 2864 (1996a).
Ching, E.S.C., J.S. Langer, and H. Nakanishi, “Dynamic instabilities in fracture,” Phys.
Rev. Lett. 76, 1087 (1996b).
Ching, E.S.C., J.S. Langer, and H. Nakanishi, “Model study of fracture propagation-
solutions of steady-state propagation and their stability,” Physica A 221, 134 (1996c).
Cho, K., and J.D. Joannopoulos, “Ergodicity and dynamical properties of constant-
temperature molecular dynamics,” Phys. Rev. A 45, 7089 (1992).
Christman, T., A. Needleman, and S. Suresh, “An experimental and numerical study of
deformation in metal-ceramic composites,” Acta Metall. Mater. 37, 3029 (1989).
Chu, T., and Z. Hashin, “Plastic behavior of composites and porous media under isotropic
stress,” Inter. J. Eng. Sci. 9, 971 (1971).
Chudnovsky, A., and B. Kunin, “A probabilistic model of brittle crack formation,” J. Appl.
Phys. 62, 4124 (1987).
Chung, J.W., J.Th.M. De Hosson, and E. van der Giessen, “Failure of a disordered three-
dimensional spring network,” Phys. Rev. B 64, 064202 (2001).
Ciccotti, G., M. Ferrario, and J.P. Rykaert, “Molecular dynamics of rigid systems in cartesian
coordinates. A general formulation,” Mol. Phys. 47, 1253 (1982).
Cieplak, M., A. Maritan, and J.R. Banavar, “Optimal paths and domain walls in the strong
disorder limit,” Phys. Rev. Lett. 72, 2320 (1994).
Cieplak, M., A. Maritan, and J.R. Banavar, “Invasion percolation and Eden growth:
geometry and universality,” Phys. Rev. Lett. 76, 3754 (1996).
Clarke, L.J., I. Stich, and M.C. Payne, “Large-scale ab initio total energy calculations on
parallel computers,” Comp. Phys. Comm. 72, 14 (1992).
Clementi, E., “Global scientific and engineering simulations on scalar, vector and parallel
LCAP-type supercomputer,” Philos. Trans. R. Soc. Lond. A 326, 445 (1988).
598 References

Coleman, B.D., “On the strength of classical fibres and fibre bundles,” J. Mech. Phys. Solids
7, 60 (1958).
Coppard, R.W., L.A. Dissado, S.M. Rowland, and R. Rakowski, “Dielectric breakdown in
metal-loaded polyethylene,” J. Phys.: Condens. Matter 1, 3041 (1989).
Cotterell, B., “Velocity effects in fracture propagation,” Appl. Mater. Res. 4, 227 (1965).
Cotterell, B., and A.G. Atkins, “A review of the J and I integrals and their implications for
crack growth resistance and toughness in ductile fracture,” Int. J. Fract. 81, 357 (1996).
Cotterell, B., and J.R. Rice, “Slightly curved or kinked cracks,” Int. J. Fract. 14, 155 (1980).
Cox, H.L., “The elasticity and strength of paper and other fibrous materials,” Br. J. Appl.
Phys. 3, 72 (1952).
Cracknell, R.F., D. Nicholson, and N. Quirke, “Direct molecular dynamics simulation of
flow down a chemical potential gradient in a slit-shaped micropore,” Phys. Rev. Lett. 74,
2463 (1995).
Crosby, K.M., and R.M. Bradley, “Fragmentation of thin films bonded to solid substrates:
simulation and a mean field theory,” Phys. Rev. E 55, 6084 (1997).
Curtin, W.A., “Theory of mechanical properties of ceramic-matrix composites,” J. Am.
Ceram. Soc. 74, 2837 (1991).
Curtin, W.A., “Toughening in disordered brittle materials,” Phys. Rev. B 55, 11270 (1997).
Curtin, W.A., M. Pamel, and H. Scher, “Time-dependent damage evolution and failure in
materials. II. Simulations,” Phys. Rev. B 55, 12051 (1997).
Curtin, W.A., and H. Scher, “Brittle fracture in disordered materials: A spring network
model,” J. Mater. Res. 5, 535 (1990a).
Curtin, W.A., and H. Scher, “Mechanics modeling using a spring network,” J. Mater. Res.
5, 554 (1990b).
Curtin, W.A., and H. Scher, “Analytic model for scaling of breakdown,” Phys. Rev. Lett.
67, 2457 (1991).
Curtin, W.A., and H. Scher, “Algebraic scaling of material strength,” Phys. Rev. B 45, 2620
(1992).
Curtin, W.A., and H. Scher, “Time-dependent damage evolution and failure in materials. I.
Theory,” Phys. Rev. B 55, 12038 (1997).
Dadvar, M., and M. Sahimi, “Pore network model of deactivation of immobilized glucose
isomerase in packed-bed reactors II: three-dimensional simulation at the particle level,”
Chem. Eng. Sci. 57, 939 (2002).
Dadvar, M., and M. Sahimi, “Pore network model of deactivation of immobilized glucose
isomerase in packed-bed reactors III: simulation at the reactor level,” Chem. Eng. Sci.
(2003).
Daguier, P., E. Bouchaud, and G. Lapasset, “Roughness of a crack front pinned by
microstructural obstacles,” Europhys. Lett. 31, 367 (1995).
Daguier, P., S. Henaux, E. Bouchaud, and F. Creuzet, “Quantitative analysis of a fracture
surface by atomic force microscopy,” Phys. Rev. E 53, 5637 (1996).
Daguier, P., B. Nghiem, E. Bouchaud, and F. Creuzet, “Pinning and depinning of crack
fronts in heterogeneous materials,” Phys. Rev. Lett. 78, 1062 (1997).
Dally, J.W., “Dynamic photoelastic studies of fracture,” Exp. Mech. 19, 349 (1979).
Dally, J.W., “Dynamic photoelasticity and its application to stress wave propagation,
fracture mechanics, and fracture control,” in Static and Dynamic Photoelasticity and
Caustics, edited by A. Lagarde (Springer, Berlin, 1987), p. 247.
Dally, J.W., W.L. Fourney, and G.R. Irwin, “On the uniqueness of the stress intensity factor-
crack velocity relationship,” Int. F. Fract. 27, 169 (1985).
 References 599

Daniels, H.E., “The statistical theory of the strength of bundles of threads. I.,” Proc. R. Soc.
Lond. A 183, 404 (1945).
Dauskardt, R., F. Haubensak, and R.O. Ritchie, “On the interpretation of the fractal character
of fracture surfaces,” Acta Metall. Mater. 38, 143 (1990).
Daw, M.S., and M.I. Baskes, “Semiempirical, quantum mechanical calculation of hydrogen
embrittlement in metals,” Phys. Rev. Lett. 50, 1285 (1983).
Daw, M.S., and M.I. Baskes, “Embedded-atom method: Derivation and application to
impurities, surfaces, and other defects in metals,” Phys. Rev. B 29, 6443 (1984).
de Arcangelis, L., A. Hansen, H.J. Herrmann, and S. Roux, “Scaling laws in fracture,” Phys.
Rev. B 40, 877 (1989).
de Arcangelis, L., and H.J. Herrmann, “Scaling and multiscaling laws in random fuse
networks,” Phys. Rev. B 39, 2678 (1989).
de Arcangelis, L., S. Redner and H.J. Herrmann, “A random fuse model for breaking
processes,” J. Physique 46, L585 (1985).
deBotton, G., “The effective yield strength of fiber-reinforced composites,” lnter. J. Solids
Structures 32, 1743 (1995).
deBotton, G., and P. Ponte Castañeda, “On the ductility of laminated materials,” Inter. J.
Solids Structures 29, 2329 (1992).
deBotton, G., and P. Ponte Castañeda, “Elastoplastic constitutive relations for fiber-
reinforced solids,” Inter. J. Solids Structures 30, 1865 (1993).
deBotton, G., and P. Ponte Castañeda, “Variational estimates for the creep behavior of
polycrystals,” Proc. R. Soc. Lond. A 448, 121 (1995).
de Buhan, P., and A. Taliercio, “A homogenization approach to the yield strength of
composites,” Europ. J. Mech. A Solids 10, 129 (1991).
Dickinson, J.T., “Fracto-emission,” in Non-Destructive Testing of Fibre-Reinforced Plastics
Composites, edited by J. Summerscales, Vol. 2 (Elsevier, London, 1991).
Dienes, G.J., and A. Paskin, “Molecular dynamic simulations of crack propagation,” J. Phys.
Chem. Solids 48, 1015 (1987).
Ding, H.-Q., N. Karasawa, and W.A. Goddard III, “Atomic level simulations on a million
particles: the cell multipole method for Coulomb and London nonbond interactions,” J.
Chem. Phys. 97, 4309 (1992).
Dissado, L.A., and J.C. Fothergill, Electrical Degradation and Breakdown in Polymers
(Peregrinus, London, 1992).
Dissado, L.A., and P.J.J. Sweeney, “Physical model for breakdown structures in solid
dielectrics,” Phys. Rev. B 48, 16261 (1993).
Dobrodumov, A.M., and A.M. El’yashevich, “Simulation of brittle fracture of polymers by
a network model in the Monte Carlo method,” Sov. Phys.-Solid State 15, 1259 (1973).
Döll, W., “An experimental study of the heat generated in the plastic region of a running
crack in different polymenc materials,” Eng. Fract. Mech. 5, 259 (1973).
Döll, W., “A molecular weight dependent transition in polymethylmethacrylate,” J. Mater.
Sci. 10, 935 (1975).
Doyle, W.T., “The Clausius–Mossotti problem for cubic arrays of spheres,” J. Appl. Phys.
49, 795 (1978).
Dreizler, R.M., and E.K.U. Gross, Density-Functional Theory (Springer, Berlin, 1990).
Druker, D.C., “On minimum weight design and strength of non-homogeneous plastic bod-
ies,” in Non-homogeneity in Elasticity and Plasticity, edited by W. Olszag (Pergamon
Press, New York, 1959), p. 139.
Drucker, D.C., “The continuum theory of plasticity on the macroscale and the microscale,”
J. Mater. 1, 873 (1966).
600 References

Duarte, J.A.M.S., “The histogram characteristics of perimeter polynomials for directed

percolation,” J. Physique 47, 383 (1986).
Duarte, J.A.M.S., “Direct lattices: site-to-bond conversion and its uses for the percolation
and dilute polymer transitions,” Z. Phys. B 80, 299 (1990).
Duarte, J.A.M.S., “Series and Monte Carlo studies of 2 and 3 dimensions for axial
hyperscaling in directed percolation,” Physica A 189, 43 (1992).
Duarte, J.A.M.S., J.M. Carvalho, and H.J. Ruskin, “The direction of maximum spread in
anisotropic forest fires and its critical properties,” Physica A 183, 411 (1992).
Dubson, M.A., Y.C. Hui, M.B. Wiessman, and J.C. Garland, “Measurements of the fourth
moment of the current distribution in two-dimensional random resistor networks,” Phys.
Rev. B 39, 6807 (1989).
Dugdale, D.C., “Yielding of steel sheets containing slits,” J. Mech. Phys. Solids 8, 100
(1960).
Dulaney, E., and W. Brace, “Velocity behavior of a growing crack,” J. Appl. Phys 31, 2233
(1960).
Duva, J.M., and J.W. Hutchinson, “Constitutive potentials for dilutely voided nonlinear
materials,” Mech. Mater. 3, 41 (1984).
Duxbury, P.M., P.D. Beale, and P.L. Leath, “Size effects of electrical breakdown in quenched
random media,” Phys. Rev. Lett., 57, 1052 (1986).
Duxbury, P.M., P.D. Beale, and P.L. Leath, “Breakdown properties of quenched random
systems: the random-fuse network,” Phys. Rev. B 36, 367 (1987).
Duxbury, P.M., P.D. Beale, and C. Moukarzel, “Breakdown of two-phase random resistor
networks,” Phys. Rev. B 51, 3476 (1995).
Duxbury, P.M., and P.L. Leath, “The failure distribution in percolation models of
breakdown,” J. Phys. A 20, L411 (1987).
Duxbury, P.M., and P.L. Leath, “Exactly solvable models of material breakdown,” Phys.
Rev. B 49, 12676 (1994a).
Duxbury, P.M., and P.L. Leath, “Failure probability and average strength of disordered
systems,” Phys. Rev. Lett. 72, 2805 (1994b).
Duxbury, P.M., and Y. Li, in Disorder and Fracture, edited by C.J. Charmet, S. Roux, and
E. Guyon (Plenum, New York, 1990), p. 141.
Dvorak, G.J., “Transformation field analysis of inelastic composite materials,” Proc. R. Soc.
Lond. A 437, 311 (1992).
Dvorak, G.J., and Y.A. Bahei-El-Din, “A bimodal plasticity theory of fibrous composite
material,” Acta Mech. 69, 219 (1987).
Dvorak, G.J., Y.A. Bahei-El-Din, Y. Macheret, and C.H. Liu, “An experimental study of
elastic-plastic behavior of a fibrous boron-aluminum composite,” J. Mech. Phys. Solids
36, 655 (1988).
Edwards, S.F., and D.R. Wilkinson, “The surface statistics of a granular aggregate,” Proc.
R. Soc. Lond. A 381, 17 (1982).
Efros, A.L., and B.I. Shklovskii, “Critical behaviour of conductivity and dielectric constant
near the metal-non-metal transition threshold,” Phys. Status Solidi B 46, 475 (1976).
Elbern, H., “Parallelization and load balancing of a comprehensive atmospheric chemistry
transport model,” Atmos. Environ. 31, 3561 (1997).
Elimes, A., R.A. Romer, and M. Schreiber, Eur. Phys. J. B 1, 29 (1998).
Englman, R., and Z. Jaeger, “Third Bar-Ilan Conference on Frontiers in Condensed Matter
Physics,” Physica A 168, 655 (1990).
Ercolessi, F., and J. Adams, “Interatomic potentials from first-principles calculations: The
force-matching method,” Europhys. Lett., 26, 583 (1993).
 References 601

Ericksen, J. L., in Phase Transformations and Material Instabilities in Solids, edited by M.

Gurtin (Academic Press, New York, 1984).
Ertas, D., and M. Kardar, “Dynamic roughening of directed lines,” Phys. Rev. Lett. 69, 929
(1992).
Ertas, D., and M. Kardar, “Dynamic relaxation of drifting polymers: a phenomenological
approach,” Phys. Rev. E 48, 1228 (1993).
Ertas, D., and M. Kardar, “Anisotropic scaling in depinning of a flux line,” Phys. Rev. Lett.
73, 1703 (1994).
Ertas, D., and M. Kardar, “Anisotropic scaling in threshold critical dynamics of driven
directed lines,” Phys. Rev. B 53, 3520 (1996).
Eshelby, J.D., “The elastic field of a crack extending non-uniformly under general anti-plane
loading,” J. Mech. Phys. Solids 17, 177 (1969).
Eshelby, J.D., “Fracture mechanics,” Sci. Prog. 59, 161 (1971).
Español, P., I. Zúniga, and M.A. Rubio, “Effect of boundary conditions in mode I fracture
in brittle materials,” Physica D 96, 375 (1996).
Essenlink, K., B. Smit, and P.A.J. Hilbers, “Efficient parallel implementation of molecular
dynamics on a toroidal machine, Part I: Parallelizing strategy,” J. Comput. Phys. 106,
101 (1993).
Ewalds, H.L., and R.J.H. Wanhill, Fracture Mechanics (Arnold, London, 1986).
Falk, M.L., “Molecular-dynamics study of ductile and brittle fracture in model noncrys-
talline solids,” Phys. Rev. B 60, 7062 (1999).
Falk, M.L., and J.S. Langer, “Dynamics of viscoplastic deformation in amorphous solids,”
Phys. Rev. E 57, 7192 (1998).
Family, F., and T. Vicsek, “Scaling of the active zone in the Eden process on percolation
networks and the ballistic deposition model,” J. Phys. A 18, L75 (1985).
Family, F., and T. Vicsek (eds.), Dynamics of Fractal Surfaces (World Scientific, Singapore,
1991).
Family, F., Y.C. Zhang, and T. Vicsek, “Invasion percolation in an external field: dielectric
breakdown in random media,” J. Phys. A 19, L733 (1986).
Feigel’man, M.V., and V.M. Vinokur, “Thermal fluctuations of vortex lines, pinning, and
creep in high-Tc superconductors,” Phys. Rev. B 41, 8986 (1990).
Fernandez, L., F. Guinea, and E. Louis, “Random and Dendritic patterns in crack
propagation,” J. Phys. A 21, L301 (1988).
Ferrante, J., J.R. Smith, and J.H. Rose, “Diatomic molecules and metallic adhesion, co-
hesion, and chemisorption: a single binding-energy relation,” Phys. Rev. Lett. 50, 1385
(1983).
Feynman, R.P., “Forces in molecules,” Phys. Rev. 56, 340 (1939).
Field, J.E., “Brittle fracture: Its study and application,” Contemp. Phys. 12, 1 (1971).
Fineberg, J., S.P. Gross, M. Marder, and H.L. Swinney, “Instability in dynamic fracture,”
Phys. Rev. Lett. 67, 457 (1991).
Fineberg, J., S.P. Gross, M. Marder, and H.L. Swinney, “Instability in the propagation of
fast cracks,” Phys. Rev. B 45, 5146 (1992).
Fineberg, J., S.P. Gross, and E. Sharon, “Micro-branching as an instability in dynamic
fracture,” in Nonlinear Analysis of Fracture, edited by J.R. Willis (Kluwer Academic,
Dordrecht, 1997), p. 177.
Fineberg, J., and M. Marder, “Instability in dynamic fracture,” Phys. Rep. 313, 1 (1999).
Fisher, D.S., “Sliding charge-density waves as a dynamic critical phenomenon,” Phys. Rev.
B 31, 1396 (1985).
602 References

Fisher, D.S., M.P.A. Fisher, and D.A. Huse, “Thermal fluctuations, quenched disorder, phase
transitions, and transport in type-II superconductors,” Phys. Rev. B 43, 130 (1991).
Fleck, N.A., and J.W. Hutchinson, “Strain gradient plasticity,” Adv. Appl. Mech. 33, 295
(1997).
Fleming, R.M., and C. C. Grimes, “Sliding-mode conductivity in NbSe3 : Observation of a
threshold electric field and conduction noise,” Phys. Rev. Lett. 42, 1423 (1979).
Flytzanis, C., Prog. Opt. 29, 2539 (1992).
Foiles, S.M., M.I. Baskes, and M.S. Daw, “Embedded-atom-method functions for fcc metals
Cu, Ag, Au, Ni, Pd, Pt, and their alloys,” Phys. Rev. B 33, 7983 (1986).
Fonesca, I., J. Math. Pure. Appl. 67, 175 (1988).
Form, W., N. Ito, and G.A. Kohring, “Vectorized and parallelized algorithms for multi-
million particle MD-simulation,” Int. J. Mod. Phys. C 4, 1075 (1993).
Fox, G.C., M.A. Johnson, G.A. Lyzenga, S.W. Otto, J.K. Salmon, and D.W. Walker, Solving
Problems on Cocurrent Processors, Vol. 1 (Prentice Hall, Englewood Cliffs, NJ, 1988).
Frenkel, D., and A.J.L. Ladd, “New Monte Carlo method to compute the free energy of
solids. Application to the FCC and HCP phases of hard spheres,” J. Chem. Phys. 81,
3188 (1984).
Freund, L.B., “The mechanics of dynamic shear crack propagation,” J. Geophys. Res. 84,
2199 (1979).
Freund, L.B., Dynamic Fracture Mechanics (Cambridge University Press, Cambridge,
1990).
Freund, L.B., and W.D. Nix, “A critical thickness condition for a strained compliant
substrate/epitaxial film system,” Appl. Phys. Lett. 69, 173 (1996).
Friel, J.J., and C.S. Pande, “A direct determination of fractal dimension of fracture surfaces
using scanning electron microscopy and stereoscopy,” J. Mater. Res. 8, 100 (1993).
Fuller, K.N.G., P.G. Fox, and J.E. Field, “The temperature rise at the tip of fast-moving
cracks in glassy polymers,” Proc. R. Soc. Lond. A 341, 1213 (1975).
Furukawa, H., “Propagation and pattern of crack in two dimensional dynamical lattices,”
Prog. Theor. Phys. 90, 949 (1993).
Gadenne, P., F. Brouers, V.M. Shalaev, and A.K. Sarychev, “Giant Stokes fields on
semicontinuous metal films,” J. Opt. Soc. Am. B 15, 68 (1998).
Galambos, J., The Asymptotic Theory of Extreme Order Statistics (Wiley, New York, 1978).
Gao, H., “Surface roughening and branching instabilities in dynamic fracture,” J. Mech.
Phys. Solids 41, 457 (1993).
Gao, L., B.-C. Xie, and Z.-Y. Li, “Crossover exponents in percolating nonlinear normal
conductor-insulator mixtures,” Physica A 271, 238 (1999).
Gărăjeu, M., and P. Suquet, “Effective properties of porous ideally plastic or viscoplastic
materials containing rigid particles,” J. Mech. Phys. Solids 45, 873 (1997).
Garboczi, E.J., “Linear dielectric-breakdown electrostatics,” Phys. Rev. B 38, 9005 (1988).
Gear, C.W., Numerical Initial Value Problems in Ordinary Differential Equations (Prentice-
Hall, Englewood Cliffs, NJ, 1971).
Gefen, Y., W.-H. Shih, R.B. Laibowitz, and J.M. Viggiano, “Nonlinear behavior near the
percolation metal-insulator transition,” Phys. Rev. Lett. 57, 3097 (1986).
George, A., and H. Michot, “Dislocation loops at crack tips: nucleation and growth—an
experimental study in silicon,” Mater. Sci. Eng. A 164, 118 (1993).
Germain, P., Q.S. Nguyen, and P. Suquet, “Continuum thermodynamics,” J. Appl. Mech.
50, 1010 (1983).
Gibiansky, L.V., and S. Torquato, “New approximation for the effective energy of non-linear
conducting composites,” J. Appl. Phys. 84, 301 (1998a).
 References 603

Gibiansky, L.V., and S. Torquato, “Effective energy of nonlinear elastic and conduct-
ing composites: Approximations and cross-property bounds,” J. Appl. Phys. 84, 5969
(1998b).
Gilabert, A., S. Roux, and E. Guyon, “Current-voltage characteristic of a nonlinear resistor
network,” J. Physique 48, 1609 (1987).
Gilman, J.J., C. Knudsen, and W.P. Walsh, “Cleavage cracks and dislocations in LiF
crystals,” J. Appl. Phys. 6, 601 (1958).
Gilormini, “Insuffisance de l’extension classique du modéle auto-cohérent au comportement
non-linéaire,” C. R. Acad. Sci. Paris Ser. IIb 320, 115 (1995).
Goldstein, R.V., and R.K. Salganik, “Brittle fracture of solids with arbitrary cracks,” Int. J.
Fract. 10, 507 (1974).
Golosovsky, M., M. Tsindlekht, D. Davidov, and A.K. Sarychev, “Effective-medium ap-
proach to the microwave properties of the high-Tc superconductor-insulator composite,”
Physica C 209, 337 (1993).
Gordon, J.E., Structures or Why Things Don’t Fall Down (Penguin, London, 1978).
Gorkov, L.P., and G. Grüner (eds.), Charge Density Waves in Solids (Elsevier, Amsterdam,
1989).
Gouldstone, A., H.J. Koch, K.Y. Zeng, A.E. Ginnakopoulos, and S. Suresh, “Discrete and
continuous deformation during nanoindentation of thin films,” Acta Mater. 48, 2277
(2000).
Greengard, L., and V.I. Rokhlin, “A fast algorithm for particle simulations,” J. Comput.
Phys. 73, 325 (1987).
Grest, G.S., B. Dünweg, and K. Kremer, “Vectorized link cell Fortran code for molecular
dynamics simulations for a large number of particles,” Comput. Phys. Commun. 70, 243
(1989).
Griffith, A.A., “The phenomena of rupture and flow in solids,” Philos. Trans. R. Soc. Lond.
227, 163 (1920).
Gross, S., Dynamics of Fast Fracture, Ph.D. Thesis, University of Texas, Austin (1995).
Gross, S.P., J. Fineberg, M. Marder, W.D. McCormick, and H.L. Swinney, “Acoustic
emissions from rapidly moving cracks,” Phys. Rev. Lett. 71, 3162 (1993).
Gu, G.Q., and K.W. Yu, “Effective conductivity of nonlinear composites,” Phys. Rev. B 46,
4502 (1992).
Gumbel, E.J., Statistics of Extremes (Columbia University Press, New York, 1958).
Gumbsch, P., and H. Gao, “Dislocation faster than the speed of sound,” Science 283, 965
(1999).
Gumbsch, P., S.L. Zhou, and B.L. Holian, “Molecular dynamics investigation of dynamic
crack stability,” Phys. Rev. B 55, 3445 (1997).
Günes, P., S. Simsek, and S. Erkoc, “A comparative study of empirical potential energy
functions: Application to clusters,” Int. J. Mod. Phys. C 11, 451 (2000).
Gyure, M.F., and P.D. Beale, “Dielectric breakdown of a random array of conducting
cylinders,” Phys. Rev. B 40, 9433 (1989).
Gyure, M.F., and P.D. Beale, “Dielectric breakdown in continuous models of metal-loaded
dielectrics,” Phys. Rev. B 46, 3736 (1992).
Haile, J.M., and S. Gupta, “Extensions of molecular dynamics simulation method. II.
Isothermal systems,” J. Chem. Phys. 79, 3067 (1983).
Hamann, D.R., M. Schlüter, and C. Chiang, “Norm-conserving pseudopotentials,” Phys.
Rev. Lett. 43, 1494 (1979).
604 References

Hammonds, K.D., I.R. McDonald, and D.J. Tildesley, “Computational studies of the struc-
ture of carbon dioxide monolayers physisorbed on the basal plane of graphite,” Mol.
Phys. 70, 175 (1990).
Hansen, A., E.L. Hinrichsen, and S. Roux, “Roughness of crack interfaces,” Phys. Rev. Lett
66, 2476 (1991b).
Hansen, A., S. Roux, and E.L. Hinrichsen, “Annealed model for breakdown processes,”
Europhys. Lett. 13, 517 (1990).
Hansen, A., S. Roux, and H.J. Herrmann, “Rupture of central-force lattices,” J. Phys. France
50, 733 (1989).
Hansen, J.P., and McDonald, I.R., Theory of Simple Liquids (Academic Press, New York,
1986).
Harlow, D.G., Proc. R. Soc. Lond. A 397, 211 (1985).
Harlow, D.G., and S.L. Phoenix, “The chain-of-bundles probability model for the strength
of fibrous materials. I.Analysis and conjectures,” J. Comput. Mater. 12, 195 (1978).
Harlow, D.G., and S.L. Phoenix, “Approximations for the strength distribution and size
effect in an idealized lattice model of material breakdown,” J. Mech. Phys. Solids 39,
173 (1991).
Harris, A.B., “Field-theoretics formulation of the randomly diluted nonlinear resistor
network,” Phys. Rev. B 35, 5056 (1987).
Harris, A.B., T.C. Lubensky, W.K. Holcomb, and C. Dasgupta, “Renormalization-group
approach to percolation,” Phys. Rev. Lett. 35, 327 (1974).
Harris, R.A., and L.R. Pratt, “Discretized propagators, Hartree, Hartree–Fock equation, and
the Hohenberg–Kohn theorem,” J. Chem. Phys. 82, 856 (1985).
Hashin, Z., “The elastic moduli of heterogeneous materials,” J. Appl. Mech. 29, 143 (1962).
Hashin, Z. “Failure criteria for unidirectional fiber composites,” J. Appl. Mech. 47, 329
(1980).
Hashin, Z., and S. Shtrikman, “On some variational principles in anisotropic and
nonhomogeneous elasticity,” J. Mech. Phys. Solids 10, 335 (1962a).
Hashin, Z., and S. Shtrikman, “A variational approach to the theory of the elastic behavior
of polycrystals,” J. Mech. Phys. Solids 10, 343 (1962b).
Hashin, Z., and S. Shtrikman, “A variational approach to the theory of the elastic behavior
of multiphase materials,” J. Mech. Phys. Solids 11, 127 (1963).
Hassold, G.N., and D.J. Srolovitz, “Brittle fracture in materials with random defects,” Phys.
Rev. B 39, 9273 (1989).
Hauch, J., and M. Marder, “Energy balance in dynamic fracture, investigated by the potential
drop technique,” Int. J. Fract. (1999).
Heffelfinger, G.S., and F. van Swol, “Diffusion in Lennard–Jones fluids using dual control
volume grand canonical molecular dynamics simulation (DCV-GCMD),” J. Chem. Phys.
100, 7548 (1994).
Heiba, A.A., M. Sahimi, L.E. Scriven, and H.T. Davis, “Percolation theory of two-phase
flow in porous media,” Society of Petroleum Engineers paper 11015, New Orleans, LA
(1982).
Heiba, A.A., M. Sahimi, L. E. Scriven, and H. T. Davis, “Percolation theory of two-phase
relative permeability,” SPE Reservoir Engineering 7, 123 (1992).
Heino, P., and K. Kaski, “Mesoscopic model of crack branching,” Phys. Rev. B 54, 6150
(1996).
Heino, P., and K. Kaski, “Mesoscopic Maxwell-dissipative finite element model for crack
propagation,” Int. J. Mod. Phys. C 8, 383 (1997).
 References 605

Heinrichs, J., and N. Kumar, “Simple exact treatment of conductance in a random Bethe
lattice,” J. Phys. C 8, L510 (1975).
Hellmann, H., Einführung in die Quantumchemie (Deuticke, Leipzig, 1937).
Hendrickson, B., and R. Leland, “An improved spectral graph partitioning algorithm for
mapping parallel computations,” SIAM J. Sci. Stat. Comput. 16, 452 (1995).
Herrmann, H.J., A. Hansen, and S. Roux, “Fracture of disordered, elastic lattices in two
dimensions,” Phys. Rev. B 39, 637 (1989a).
Herrmann, H.J., J. Kertész, and L. de Arcangelis, “Fractal shapes of deterministic cracks,”
Europhys. Lett. 10, 147 (1989b).
Herrmann, H.J., and S. Roux (eds.), Statistical Models for the Fracture of Disordered Media
(North-Holland, Amsterdam, 1990).
Herrmann, H.J., and M. Sahimi, “Fluid penetration through a crack in a pressure gradient,”
J. Phys. A 26, L1145 (1993).
Herrmann, H.J., M. Sahimi, and F. Tzschichholz, “Examples of fractals in soil mechanics,”
Fractals 1, 795 (1993).
Hervé, E., C. Stolz, and A. Zaoui, “A propos de l’assemblage des sphéres composites de
Hashin,” C. R. Acad. Sci. Paris Ser. II 313, 857 (1991).
Hesselbo, B., D. Phil. Thesis, Oxford University (1994).
Hill, R., “Elastic properties of reinforced solids: Some theoretical principles,” J. Mech.
Phys. Solids 11, 357 (1963).
Hill, R., “Continuum micro-mechanics of elastoplastic polycrystals,” J. Mech. Phys. Solids
13, 89 (1965a).
Hill, R., “Self-consistent mechanics of composite materials,” J. Mech. Phys. Solids 13, 213
(1965b).
Hill, R.M., and L.A. Dissado, “Theoretical basis for the statistics of dielectric breakdown,”
J. Phys. C 16, 2145 (1983a).
Hill, R.M., and L.A. Dissado, “Examination of the statistics of dielectric breakdown,” J.
Phys. C 16, 4447 (1983b).
Hinrichsen, E.L., A. Hansen, and S. Roux, “A fracture growth model,” Europhys. Lett. 8, 1
(1989).
Hirsch, P.B., R.W. Horne, and M.J. Whelan, “Direct observation of the arrangement and
motion of dislocation in aluminum,” Philos. Mag. (series 8) 1, 677 (1956).
Hirsch, P.B., J. Samuels, and S.G. Roberts, “The brittle-ductile transition in silicon. II.
Interpretation,” Proc. R. Soc. Lond. A 421, 25 (1989).
Ho, P.S., and T. Kwok, “Electromagnetism in metals,” Rep. Prog. Phys. 52, 301 (1989).
Hoagland, R.G., M.S. Daw, S.M. Foiles, and M.I. Baskes, “An atomic model of crack tip
deformation in aluminum using an embedded atom potential,” J. Mater. Res. 5, 313
(1990).
Hockney, R.W., S.P. Goel, and J.W. Eastwood, “Quite high-resolution computer models of
a plasma,” J. Comput. Phys. 14, 48 (1974).
Hodgdon, J.A., and J.P. Sethna, “Derivation of a general three-dimensional crack-
propagation law: A generalization of the principle of local symmetry,” Phys. Rev. B
47, 4831 (1993).
Hoeing, A., “Some implications of an elastic-electrostatic analogy on certain path-
independent integrals,” Int. J. Eng. Sci. 22, 87 (1984).
Hohenberg, P., and W. Kohn, “Inhomogeneous electron gas,” Phys. Rev. 136, B864 (1964).
Holian, B.L., and R. Ravelo, “Fracture simulations using large-scale molecular dynamics,”
Phys. Rev. B 51, 11275 (1995).
606 References

Holian, B.L., A.F. Voter, and R. Ravelo, “Thermostatted molecular dynamics: How to avoid
the Toda demon hidden in Nosé-Hoover dynamics,” Phys. Rev. E 52, 2338 (1995).
Holian, B.L., A.F. Voter, N.J. Wagner, R.J. Ravelo, S.P. Chen, W.G. Hoover, C.G. Hoover,
J.E. Hammerberg, and T.D. Dontje, “Effects of pairwise versus many-bondy forces on
high-stress plastic deformation,” Phys. Rev. A 43, 2655 (1991).
Holland, D., and M. Marder, “Ideal brittle fracture of silicon studied with molecular
dynamics,” Phys. Rev. Lett. 80, 746 (1998).
Hoover, W.G., “Canonical dynamics: Equilibrium phase-space distributions,” Phys. Rev. A
31, 1695 (1985).
Hoover, W.G., and F.H. Ree, “Melting transition and communal entropy for hard spheres,”
J. Chem. Phys. 49, 3609 (1968).
Horowitz, G.E., Arch. Elektrotech. Berlin 18, 555 (1927).
Hough, S.E., “On the use of spectral methods for the determination of fractal dimension,”
Geophys. Res. Lett. 16, 673 (1989).
Hua, L., H. Rafii-Tabar, and M. Cross, Philos. Mag. Lett. 75, 237 (1997).
Hughes, B.D., Random Walks and Random Environments, Vol. 1 (Oxford University Press,
London, 1995).
Hui, P.M., “Effective nonlinear response in dilute nonlinear granular materials,” J. Appl.
Phys. 68, 3009 (1990a).
Hui, P.M., “Enhancement in nonlinear effects in percolating nonlinear resistor networks,”
Phys. Rev. B 41, 1673 (1990b).
Hui, P.M., “Crossover electric field in percolating perfect-conductor-nonlinear-normal-
metal composites,” Phys. Rev. B 49, 15344 (1994).
Hui, P.M., P. Cheung, and Y.R. Kwong, “Effective response in nonlinear random
composites,” Physica A 241, 301 (1997).
Hull, D., “Effect of Crazes on the propagation of cracks in polystyrene,” J. Mater. Sci. 5,
357 (1970).
Hull, D., Fractography (Cambridge University Press, Cambridge, 1999).
Hull, D., and P. Beardmore, “Velocity of propagation of cleavage cracks in tungsten,” Int.
J. Fract. Mech. 2, 468 (1966).
Huntington, H.B., “Electromigration in Metals,” in Diffusion in Solids-Recent Develop-
ments, edited by A.S. Novick and J.J. Burton (Academic Press, New York, 1975),
p. 120.
Imre, A., T. Pajkossy, and L. Nyikos, “Electrochemical determination of the fractal
dimension of fractured surfaces,” Acta Metall. Mater. 40, 1819 (1992).
Inglis, C.E., “Stresses in a plate due to the presence of cracks and sharp corners,” Trans.
Inst. Naval Archit. 55, 219 (1913).
Irwin, G.R., “Analysis of stresses and strains near the end of a crack traversing a plate,” J.
Appl. Mech. 24, 361 (1956).
Irwin, G.R., “Fracture,” in Handbuch der Physik, Vol. 6 (Springer, Berlin, 1958), p. 551.
Irwin, G.R., J.W. Dally, T. Kobayashi, W.L. Fourney, M.J. Etheridge, and H.P. Rossmanith,
“On the determination of the a-K relationship for birefringent polymers,” Exp. Mech.
19, 121 (1979).
Jackson, J.D., Classical Electrodynamics, 3rd Ed. (John Wiley & Sons, 1998).
Janak, J.F., “Proof that ∂E/∂ni = i in density-functional theory,” Phys. Rev. B 18, 7165
(1978).
Jarvinen, R., T. Mantyla, and P. Kettunen, “Improved adhesion between a sputtered alumina
coating and a copper substrate,” Thin Solid Films 114, 311 (1984).
 References 607

Jellinek, J., and R.S. Berry, “Generalization of Nosé’s isothermal molecular dynamics,”
Phys. Rev. A 38, 3069 (1988).
Joannopoulos, J.D., P. Bash, and A. Rappe, Chemical Design Automation News 6 (No. 8)
(1991).
Johnson, E., “Process region changes for rapidly propagating cracks,” Int. J. Fract. 55, 47
(1992).
Johnson, E., “Process region influence on energy release rate and crack tip velocity during
rapid crack propagation,” Int. J. Fract. 61, 183 (1993).
Johnson, H.T., and L.B. Freund, “Mechanics of coherent and dislocation island morpholo-
gies in strained epitaxial material systems,” J. Appl. Phys. 81, 6081 (1997).
Johnson, J.W., and D.G. Holloway, “On the shape and size of the fracture zones on glass
fracture surfaces,” Philos. Mag. 14, 731 (1966).
Johnson, J.W., and D.G. Holloway, “Microstructure of the mist zone on glass fracture
surfaces,” Philos. Mag. 17, 899 (1968).
Johnson, R.A., "Interstitials and vacancies in α-iron", Phys. Rev. 134A, 1329 (1964).
Johnson, R.A., “Analytic nearest-neighbor model for fcc metals,” Phys. Rev. B 37, 3924
(1988).
Jones, R.O., and O. Gunnarsson, “The density functional formalism, its applications and
prospects,” Rev. Mod. Phys. 61, 689 (1989).
Kahng, B., G.G. Batrouni, S. Redner, L. de Arcangelis, and H.J. Herrmann, “Electrical
breakdown in a fuse network with random, continuously distributed breaking strengths,”
Phys. Rev. B 37, 7625 (1988).
Kalia, R.K., A. Nakano, A. Omeltchenko, K. Tsuruta, and P. Vashishta, “Role of ultrafine
microstructures in dynamic fracture in nanophase silicon nitride,” Phys. Rev. Lett. 78,
2144 (1997).
Kallivayalil, J., and A.T. Zehnder, “A method for thermo-mechanical analysis of steady
state dynamic crack growth,” Int. J. Fract. 66 99 (1994).
Kalthoff, J.F., “On the propagation direction of bifurcated cracks,” in Dynamic Crack
Propagation, edited by G.C. Sih (Noordhoof, Leydon, 1972), p. 449.
Kalthoff, J.F., “The shadow optical method of caustics,” in Static and Dynamic
Photoelasticity and Caustics, edited by A. Lagarde (Springer, Berlin, 1987), p. 401.
Kane, E.O., “Phonon spectra of diamond and zinc-blende semiconductors,” Phys. Rev. B
31, 7865 (1985).
Kantor, Y., and I. Webman, “Elastic properties of random percolating systems,” Phys. Rev.
Lett. 52, 1891 (1984).
Kardar, M., “Fluctuations of interfaces and fronts,” in Disorder and Fracture, edited by
J.C. Charmet, S. Roux, and E. Guyon (Plenum, New York, 1990), p. 3.
Kardar, M., G. Parisi, and Y.C. Zhang, “Dynamic scaling of growing interfaces,” Phys. Rev.
Lett. 56, 889 (1986).
Karma, A., D.A. Kessler, and H. Levine, “Phase-field model of Mode III dynamic fracture,”
Phys. Rev. Lett. 87, 045501-1 (2001).
Kausch, H.H., Polymer Fracture (Springer, Berlin, 1987).
Kaveh, M., and N.F. Mott, “The conductivity of disordered systems and the scaling theory,”
J. Phys. C 14, L659 (1981).
Kawarabayashi, T., B. Kramer, and T. Ohtsuki, “Anderson transitions in three-dimensional
disordered systems with randomly varying magnetic flux,” Phys. Rev. B 57, 11842 (1998).
Kaxiras, E., and M.S. Duesbery, “Free energies of generalized stacking faults in Si and
implications for the brittle-ductile transition,” Phys. Rev. Lett. 70, 3752 (1993).
608 References

Kaxiras, E., and K.C. Pandey, “New classical potential for accurate simulation of atomic
processes in Si,” Phys. Rev. B 38, 12736 (1988).
Kelchner, C.J., S.J. Plimpton, and J.C. Hamilton, “Dislocation nucleation and defect
structure during surface indentation,” Phys. Rev. B 58, 11085 (1998).
Kellomäki, M., J. Åström, and J. Timonen, “Rigidity and dynamics of random spring
networks,” Phys. Rev. Lett. 77, 2730 (1996).
Kelly, A., W.R. Tyson, and A.H. Cotterell, “Ductile and brittle crystals,” Philos. Mag. 15,
567 (1967).
Kenkel, S.W., and J.P. Straley, “Percolation theory of nonlinear circuit elements,” Phys.
Rev. Lett. 49, 767 (1982).
Kent, R., and R. Rat, “Static electricity phenomena in the manufacture and handling of solid
propellants,” J. Electrostat. 17, 299 (1985).
Kerkhof, F., “Wave fractographic investigations of brittle fracture dynamics,” in Dynamic
Crack Propagation, edited by G.C. Sih (Noordhoff International Publishing, Leyden,
1973), p. 3.
Khanta, M., D.P. Pope, and V. Vitek, “Dislocation screening and the brittle-to-ductile
transition: A Kosterlitz-Thouless type instability,” Phys. Rev. Lett. 73, 684 (1994).
Kim, J.M., and J.M. Kosterlitz, “Growth in a restricted solid-on-solid model,” Phys. REv.
Lett. 62, 2289 (1989).
Kim, K.S., “Dynamic fracture under normal impact loading of the crack faces,” J. Appl.
Mech. 52, 585 (1985).
Kim, S., and S. Karrila, Microhydrodynamics. Principles and Selected Applications,
(Butterworth-Heinemann, Boston, 1991).
Kinloch, A.J., and R.J. Young, Fracture Behavior of Polymers (Applied Sciences, London,
1983).
Kinzel, W., “Directed percolation,” in Percolation Structures and Processes, edited by G.
Deutscher, R. Zallen, and J. Adler (Adam Hilger, Bristol, 1983), p. 425.
Knap, J., and M. Ortiz, “An analysis of the quasicontinuum method,” J. Mech. Phys. Solids
49, 1899 (2001).
Knaur, J.A., and P.P. Budenstein, IEEE Trans. EI-15, 313 (1980).
Kneipp, K., Y. Wang, H. Kneipp, L.T. Perelman, I. Itzkan, R. Dasari, and M.S. Feld, “Single
molecule detection using surface-enhanced Raman scattering (SERS),” Phys. Rev. Lett.
78, 1667 (1997).
Kobayashi, A., N. Ohtani, and T. Sato, “Phenomenological aspects of viscoelastic crack
propagation,” J. Appl. Poly. Sci. 18, 1625 (1974).
Kobayashi, A.S., in Handbook on Experimental Mechanics, edited by A.S. Kobayashi
(American Society for Testing and Materials, Prentice-Hall, Englewood Cliffs, NJ, 1987),
p. 231.
Kohlhoff, S., P. Gumbsch, and H. F. Fischmeister, “Crack propagation in BCC crystals
studied with a combined finite-element and atomistic model,” Philos. Mag. 64A, 851
(1991).
Kohn, R.V., and T.D. Little, “Some model problems of polycrystal plasticity with deficient
basic crystals,” (1997).
Kohn, R.V., and G.W. Milton, “On bounding the effective conductivity of anisotropic com-
posites,” in Homogenization and Effective Moduli of Materials and Media, edited by
J.L. Ericksen, D. Kinderlehrer, R.V. Kohn, and J.-L. Lions (Springer, New York, 1986),
p. 97.
Kohn, W., and L.J. Sham, “Self-consistent equations including exchange and correlation
effects,” Phys. Rev. 140, A1133 (1965).
 References 609

Kohr, K.E., and S. Das Sarma, “Model-potential study of (2n + 1) × (2n + 1) reconstruc-
tions on Si(111) surface,” Phys. Rev. B 40, 1319 (1989).
Kolsky, H., Stress Waves in Solids (Oxford University Press, London, 1953).
Koplik, J., and H. Levine, “Interface moving through a random background,” Phys. Rev. B
32, 280 (1985).
Koschmieder, E.L., Bérnard Cells and Taylor Vortices (Cambridge University Press,
London, 1993).
Kosterlitz, J.M., and D.J. Thouless, “Ordering, metastability and phase transitions in two-
dimensional systems,” J. Phys. C 6, 1181 (1973).
Kramer, B., and A. MacKinnon, “Localization: theory and experiment,” Rep. Prog. Phys.
56, 1469 (1993).
Kuang, L., and H.J. Simon, “Diffusely scattered second harmonic generation from a silver
film due to surface plasmons,” Phys. Lett. A 197, 257 (1995).
Kulakhmetova, S.A., V.A. Saraikin, and L.I. Slepyan, “Plane problem of a crack in a lattice,”
Mech. Solids 19, 102 (1984).
Kulawansa, D.M., L.C. Jensen, S.C. Langford, and J.T. Dickinson, “Scanning electron
microscopy of the mirror region of silicate glass fracture surfaces,” J. Mater. Res. 9, 476
(1993).
Kun, F., A. Zapperi, and H.J. Herrmann, “Damage in fiber bundle models,” Eur. Phys. J. B
17, 269 (2000).
Kunin, I.A., Elastic Media with Microstructure (Springer, Berlin, 1982).
Kusy, R.P., and D.T. Turner, “Influence of the molecular weight of PMMA on fracture
surface energy in notched tension,” Polymer 17, 161 (1975).
Kusy, R.P., and D.T. Turner, “Influence of the molecular weight of poly(methyl
methacrylate) on fracture morphology in notched tension,” Polymer 18, 391 (1977).
Lagarkov, A.N., L.V. Panina, and A.K. Sarychev, “Effective magnetic permeability of com-
posite materials near the percolation threshold,” Sov. Phys. JETP 66, 123 (1987) [Zh.
Eskp. Teor. Fiz. 93, 215 (1987)].
Lamaignere, L., F. Carmona, and D. Sornette, “Experimental realization of critical thermal
fuse rupture,” Phys. Rev. Lett. 77, 2738 (1996).
Landau, L.D., and E.M. Lifshitz, Statistical Physics (Pergamon Press, London, 1980).
Langer, J.S., “Models of crack propagation,” Phys. Rev. A 46, 3123 (1992).
Langer, J.S., “Dynamic model of onset and propagation of fracture,” Phys. Rev. Lett. 70,
3592 (1993).
Langer, J.S., and A.E. Lobkovsky, “Critical examination of cohesive-zone models in the
theory of dynamic fracture,” J. Mech. Phys. Solids 46, 1521 (1998).
Langer, J.S., and H. Nakanishi, “Models of crack propagation. II. Two-dimensional model
with dissipation on the fracture surface,” Phys. Rev. E 48, 439 (1993).
Langer, J.S., and C. Tang, “Rupture propagation in a model of an earthquake fault,” Phys.
Rev. Lett. 67, 1043 (1991).
Langford, S.C., M. Zhenyi, and J.T. Dickinson, “Photon emission as a probe of chaotic
processes accompanying fracture,” J. Mater. Res. 4, 1272 (1989).
Larkin, A.I., and Yu.N. Ovchinnikov, “Pinning in type II superconductors,” J. Low Temp.
Phys. 34, 409 (1979).
Larralde, H., and R.C. Ball, “The shape of slowly growing cracks,” Europhys. Lett. 30, 87
(1995).
Larson, R.G., “Derivation of generalized Darcy equations for creeping flow in porous
media,” Ind. Eng. Chem. Fund. 20, 132 (1981).
610 References

Lawn, B., Fracture of Brittle Solids, 2nd ed. (Cambridge University Press, Cambridge,
1993).
Leath, P.L., and P.M. Duxbury, “Fracture of heterogeneous materials with continuous
distributions of local breaking strengths,” Phys. Rev. B 49, 14905 (1994).
Lee, H.-C., and M.E. Mear, “Effective properties of power-law solids containing elliptical
inhomogenities. I. Rigid inclusions,” Mech. Mater. 13, 313 (1992).
Lee, H.-C., and K.W. Yu, “Effective medium theory for strongly nonlinear composites:
comparison with numerical simulations,” Phys. Lett. A 197, 341 (1995).
Lemaire, E., Y. Ould Mohamed Abdelhaye, J. Larue, R. Benoit, P. Levitz, and H. Van
Damme, “Pattern formation in noncohesive and cohesive granular media,” Fractals 1,
968 (1993).
Leung, K.-t., and Z. Néda, “Pattern formation and selection in quasistatic fracture,” Phys.
Rev. Lett. 85, 662 (2000).
Levin, V.M., “Thermal expansion coefficients of heterogeneous materials,” Mekh. Tverd.
Tela 2, 83 (1967).
Levy, O., and D.J. Bergman, “The bulk effective response of non-linear random resistor
networks: numerical study and analytic approximations,” J. Phys.: Condens. Matter 5,
7095 (1993).
Levy, O., and D.J. Bergman, “Intrinsic optical bistability and resonances in nonlinear
composites,” Physica A 207 157 (1994a).
Levy, O., and D.J. Bergman, “Critical behavior of the weakly nonlinear conductivity and
flicker noise of two-component composites,” Phys. Rev. B 50, 3652 (1994b).
Levy, O., and D.J. Bergman, and D.G. Stroud, “Harmonic generation, induced nonlinearity,
and optical bistability in nonlinear composites,” Phys. Rev. E 52, 3184 (1995).
Levy, O., and R.V. Kohn, “Duality relations for non-Ohmic composites, with application
to behavior near percolation,” J. Stat. Phys. 90, 159 (1998).
Levy-Nathansohn, R., and D.J. Bergman, “Decoupling and testing of the generalized Ohm’s
law,” Phys. Rev. B 55 5425 (1997).
Li, G., P. Ponte Castañeda, and A.S. Douglas, “Constitutive models for ductile solids
reinforced by rigid spheroidal inclusions,” Mech. Mater 15, 279 (1993).
Li, R., and K. Sieradzki, “Ductile-brittle transition in random porous Au,” Phys. Rev. Lett.
68, 1168 (1992).
Li, W., R.K. Kalia, and P. Vashishta, “Amorphization and fracture in silicon diselenide
nanowires: a molecular dynamics study,” Phys. Rev. Lett. 77, 2241 (1996).
Li, X.-P., R.W. Nunes, and D. Vanderbilt, “Density-matrix electronic-structure method with
linear system-size scaling,” Phys. Rev. B 47, 10891 (1993).
Li, Y.S., and P.M. Duxbury, “Size and location of the largest current in a random resistor
network,” Phys. Rev. B 36, 5411 (1987).
Li, Y.S., and P.M. Duxbury, “Crack arrest by residual bonding in resistor and spring
networks,” Phys. Rev. B 38, 9257 (1988).
Liao, H.B., R.F. Xiao, H. Wang, K.S. Wong, and G.K.L. Wong, “Large third-order optical
nonlinearity in Au:TiO2 composite films measured on a femtosecond time scale,” Appl.
Phys. Lett. 72, 1817 (1998).
Lidorikis, E., M.E. Bachlechner, R.K. Kalia, A. Nakano, P. Vashishta, and G.Z. Voyiadjis,
“Coupling length scales for multiscale atomistic-continuum simulations: Atomistically
induced stress distribution in Si/Si3 N4 nanopixels,” Phys. Rev. Lett. 87, 086104-1 (2001)
Lim, S., Tsotsis, T.T., and Sahimi, M., “Molecular dynamics simulation of porous amorphous
carbon and transport of fluid mixtures therein,” J. Chem. Phys. (2003).
 References 611

Lin, J.J., “Nonlinear I − V characteristics of the granular PrBa2 Cu3 O7−δ compound,” J.
Phys. Soc. Japan 61, 4125 (1992).
Lin, S.L., J. Mellor-Crumney, B.M. Pettitt, and G.N. Phillips, Jr., “Molecular dynamics on
a distributed-memory multiprocessor,” J. Comput. Phys. 13, 1022 (1992).
Lobb, C.J., P.M. Hui, and D. Stroud, “Nonuniversal breakdown behavior in superconducting
and dielectric composites,” Phys. Rev. B 36, 1956 (1987).
Lomdahl, P.S., D.M. Beazley, P. Tamayo, and N. Grφnbech-Jensen, “Multi-million particle
molecular dynamics on the CM-5,” Int. J. Mod. Phys. C 4, 1074 (1993).
López, J.M., and J. Schmittbuhl, “Anomalous scaling of fracture surfaces,” Phys. Rev. E
57, 6405 (1998).
Louis, E., and F. Guinea, “The fractal nature of fracture,” Europhys. Lett. 3, 871 (1987).
Lupkowski, M., and F. van Swol, “Ultrathin films under shear,” J. Chem. Phys. 95, 1995
(1995).
Ma, H., Xiao, R., and P. Sheng, “Third-order optical nonlinearity enhancement through
composite microstructures,” J. Opt. Soc. Am. B 15, 1022 (1998).
MacElroy, J.M.D., “Nonequilibrium molecular dynamics simulation of diffusion and flow
in thin microporous membranes,” J. Chem. Phys. 101, 5274 (1994).
Machová, A., “Dynamic microcrack initiation in alpha -iron,” Mater. Sci. Eng. A206, 279
(1996).
Machta, J., and R.A. Guyer, “Largest current in a random resistor network,” Phys. Rev. B
36, 2142 (1987).
Mahadevan, M., R.M. Bradley, and J.-M. Debierre, “Simulations of an electromigration-
induced edge instability in single-crystal metal lines,” Europhys. Lett. 45, 680 (1999).
Måløy, K.J., A. Hansen, E.L. Hinrichsen and S. Roux, “Experimental measurements of the
roughness of brittle cracks,” Phys. Rev. Lett. 68, 213 (1992).
Mandal, P.,A. Neumann,A.G. Jansen, P. Wyder, and R. Deltour, “Temperature and magnetic-
field dependence of the resistivity of carbon-black polymer composites,” Phys. Rev. B
55, 452 (1997).
Mandel, J., Plasticité Classique et Viscoplasticité CISM Udine Courses and Lectures, 97
(Springer, Berlin, 1972).
Mandelbrot, B.B., The Fractal Geometry of Nature (W.H. Freeman, San Francisco, 1982).
Mandelbrot, B.B., “Self-affine fractals and fractal dimension,” Physica Scripta 32, 257
(1985).
Mandelbrot, B.B., D.E. Passoja, and A.J. Paullay, “Fractal character of fracture surfaces of
metals,” Nature 308, 721 (1984).
Mandelbrot, B.B., and J.W. Van Ness, “Fractional Brownian motions, fractional noise and
applications,” SIAM Rev. 10, 422 (1968).
Manna, S.S., and B.K. Chakrabarti, “Dielectric breakdown in the presence of random
conductors,” Phys. Rev. B 36, 4078 (1987).
Manogg, P., “Investigation of the rupture of a Plexiglas plate by means of an optical method
involving high speed filming of the shadows originating around holes drilled in the plate,”
Int. J. Fract. 2, 604 (1966).
Mantese, J.W., W.I. Goldberg, D.H. Darling, H.G. Craighead, U.J. Gibson, R.A. Buhrman,
and W.W. Webb, “Excess low frequency conduction noise in a granular composite,” Solid
State Commun. 37, 353 (1981).
Marcellini, P., “Periodic Solutions and homogenization of nonlinear variational problems,”
Ann. Mat. Pura. Appl. 4, 139 (1978).
Marder, M., “New dynamical equation for cracks,” Phys. Rev. Lett. 66, 2484 (1991).
Marder, M., “Statistical mechanics of cracks,” Phys. Rev. E 54, 3442 (1996).
612 References

Marder, M., and J. Fineberg, “How things break,” Phys. Today 49 (No. 9), 24 (1996).
Marder, M., and S.P. Gross, “Origin of crack tip instabilities,” J. Mech. Phys. Solids 43, 1
(1995).
Marder, M., and X. Liu, “Instability in lattice fracture,” Phys. Rev. Lett. 71, 2417 (1993).
Markel, V.A., V.M. Shalaev, E.B. Stechel, W. Kim, and R.L. Armstrong, “Small-particle
composites. I. Linear optical properties,” Phys. Rev. B 53, 2425 (1996).
Markel, V.A., V.M. Shalaev, P. Zhang, W. Huynh, L. Tay, T.L. Haslett, and M. Moskovits,
“Near-field optical spectroscopy of individual surface-plasmon modes in colloid
clusters,” Phys. Rev. B 59, 10903 (1999).
Marks, N.A., D.R. McKenzie, B.A. Paithorpe, M. Bernasconi, and M. Parrinello, “Ab initio
simulations of tetrahedral amorphous carbon,” Phys. Rev. B 54, 9703 (1996).
Martin, J.E., and M.B. Heaney, “Reversible thermal fusing model of carbon black current-
limiting thermistors,” Phys. Rev. B 62, 9390 (2000).
Mart́in, T., P. Español, M.A. Rubio, and I. Zúñiga, “Dynamic fracture in a discrete model
of a brittle elastic solid,” Phys. Rev. E 61, 6120 (2000).
Martins, J.L., and A. Zunger, “Bond lengths around isovalent impurities and in
semiconductor solid solutions,” Phys. Rev. B 30, 6217 (1984).
Martyna, G.J., M.L. Klein, and M. Tuckerman, “Nose-Hoover chains: the canonical
ensemble via continuous dynamics,” J. Chem. Phys. 97, 2635 (1992).
Martyna, G.J., D.T. Tobias, and M.L. Klein, “Constant pressure molecular dynamics
algorithms,” J. Chem. Phys. 101, 4177 (1994).
Matthews, J.W., andA.E. Blakeslee, “Defects in epitaxial multilayers. I. Misfit dislocations,”
J. Cryst. Growth 32, 265 (1974).
Mauri, F., G. Galli, and R. Car, “Orbital formulation for electronic-structure calculations
with linear system-size scaling,” Phys. Rev. B, 47, 9973 (1993).
McAnulty, P., L.V. Meisel, and P.J. Cote, “Hyperbolic distributions and fractal character of
fracture surfaces,” Phys. Rev. A 46, 3523 (1992).
Meakin, P., “A simple model for elastic fracture in thin films,” Thin Solid Films 151, 165
(1987).
Meakin, P., “Simple kinetic models for material failure and deformation,” Herrmann and
Roux (1990), p. 291.
Meakin, P., Fractals, Scaling and Growth far from Equilibrium (Combridge University
Press, Cambridge, 1998).
Meakin, P., G. Li, L.M. Sander, E. Louis, and F. Guinea, “A simple two-dimensional model
for crack propagation", J. Phys. A 22, 1393 D (1989).
Mecholsky, J.J., in Strength of Inorganic Glass, edited by C.R. Kurkjian (Plenum, New
York, 1985).
Mecholsky, J.J., T.J. Mackin, and D.E. Passoja, Adv. Ceramics 22, 127 (1988).
Mecholsky, J.J., D.E. Passoja, and K.S. Feinberg-Ringel, “Quantitative analysis of brittle
fracture surfaces using fractal geometry,” J. Am. Ceram. Soc. 72, 60 (1989).
Meek, J.M., and J.D. Craggs, Electrical Breakdown of Gases (Wiley, New York, 1978).
Mehrabi, A.R., H., Rassamdana, and M., Sahimi, “Characterization of long-range
correlations in complex distributions and profiles,” Phys. Rev. E 56, 712 (1997).
Mehrabi, A.R., and M. Sahimi, “Diffusion of ionic particles in charged disordered media,”
Phys. Rev. Lett. 82, 735 (1999).
Meir, Y., R. Blumenfeld, A. Aharony, and A.B. Harris, “Series analysis of randomly diluted
nonlinear resistor networks,” Phys. Rev. B 34, 3424 (1986).
Mel’cuk, A.I., R.C. Giles, and H. Gould, “Molecular dynamics simulation of liquids on the
Connection Machine,” Computers in Physics, 311 (May/June 1991).
 References 613

Mercer, J.L., “Tight-binding models for compounds: Application to SiC,” Phys. Rev. B 54,
4650 (1996).
Mermin, N.D., “Thermal properties of the inhomogeneous electron gas,” Phys. Rev. 137,
A1441 (1965).
Michel, J.C., “A self-consistent estimate of the potential of a composite made of two power-
law phases with the same exponent,” C. R. Acad. Sci. Paris Ser. II 322, 447 (1996).
Mikitishin, S.I., Y.-N. Skhonitskii, and A.N. Tynnyi, Fiz. Khim. Mekh. Mater. 5, 69 (1969).
Miksis, M.J., “Effective dielectric constant of a nonlinear composite material,” SIAM J.
Appl. Math. 43, 1140 (1983).
Miller, S., and R. Reifenberger, “Improved method for fractal analysis using scanning probe
microscopy,” J. Vac. Sci. Technol. B10, 1203 (1992).
Milman, V.Y., “Fracture surfaces: A critical review of fractal studied and a novel morpho-
logical analysis of scanning tunneling microscopy measurements,” Prog. Mater. Sci. 38,
425 (1994).
Milman, V.Y., R. Blumenfeld, N.A. Stelmashenko, and R.C. Ball, “Comment on ’Exper-
imental measurements of the roughness of brittle cracks’,” Phys. Rev. Lett. 71, 204
(1993).
Milman, V.Y., N.A. Stelmashenko, and R. Blumenfeld, “Fracture surfaces: A critical review
of fractal studies and a novel morphological analysis of scanning tunneling microscopy
measurements,” Prog. Mater. Sci. 38, 425 (1994).
Milstein, F., in Mechanics of Solids, edited by H.G. Hopkins and M.J. Sewell (Pergamon
Press, Oxford, 1982).
Milton, G.W., “Bounds on the electromagnetic, elastic, and other properties of two-
component composites,” Phys. Rev. Lett. 46, 542 (1981a).
Milton, G.W., “Bounds on the elastic and transport properties of two-component
composites,” J. Mech. Phys. Solids 30, 177 (1981b).
Milton, G.W., “Bounds on the elastic and transport properties of two-component
composites,” J. Mech. Phys. Solids 30, 177 (1982).
Milton, G., “Modeling the properties of composites by laminates,” in Homogenization and
Effective Moduli of Materials and Media, edited by J.L. Ericksen, D. Linderlehrer, R.
Kohn, and J.-L. Lions (Springer, Berlin, 1986), p. 150.
Mistriotis, A.D., N. Flytzanis, and S.C. Farantos, “Potential model for silicon clusters,”
Phys. Rev. B 39, 1212 (1989).
Mitchell, M.W., and D.A. Bonnell, “Quantitative topographic analysis of fractal surfaces
by scanning tunneling microscopy,” J. Mater. Res. 5, 2244 (1990).
Molinari, A., G.R. Canova, and S. Ahzi, “A self-consistent approach of the large deformation
polycrystal viscoplasticity,” Acta Metall. Mater. 35, 2983 (1987).
Morales, J.J., and M.J. Nuevo, “Comparison of link-cell and neighborhood tables on a range
of computers,” Comput. Phys. Commun. 69, 223 (1992).
Morel, S., J. Schmittbuhl, E. Bouchaud, and G. Valentin, “Scaling of crack surfaces and
implications for fracture mechanics,” Phys. Rev. Lett. 85, 1678 (2000).
Morel, S., J. Schmittbuhl, J.M. López, and G. Valentin, Phys. Rev. E 58, 6999 (1998).
Moreno, Y., J.B. Gómez, and A.F. Pacheco, “Fracture and second-order phase transitions,”
Phys. Rev. Lett. 85, 2865 (2000).
Mori, Y., K. Kaneko, and M. Wadati, “Fracture dynamics by quenching. I. Crack patterns,”
J. Phys. Soc. Japan 60, 1591 (1991).
Morrissey, J.W., and J.R. Rice, “Crack front waves,” J. Mech. Phys. Solids 46, 467 (1998).
Moskovitz, M., “Surface-enhanced spectroscopy,” Rev. Mod. Phys. 57, 783 (1985).
614 References

Mosolov, A.B., “Mechanics of fractal cracks in brittle solids,” Europhys. Lett. 24, 673
(1993).
Mott, N.F., “Brittle fracture in mild steel plates,” Engineering 165, 16 (1948).
Moulinec, H., and P. Suquet, “A FTT-based numerical method for computing the mechanical
properties of composites from images of their microstructure,” in Microstructure-
Properly Interactions in Composite Materials, edited by R. Pyrz (Kluwer, The
Netherlands, 1995), p. 235.
Movchan, A.B., and J.R. Willis, “Dynamic weight functions for a moving crack. II. Shear
loading,” J. Mech. Phys. Solids 43, 1369 (1995).
Mu, Z.Q., and C.W. Lung, “Studies on the fractal dimension and fracture toughness of
steel,” J. Phys. D 21, 848 (1988).
Müller, K., B. Mehlig, F. Milde, and M. Schreiber, “Statistics of wave functions in disordered
and in classically chaotic systems,” Phys. Rev. Lett. 78, 215 (1997).
Mura, D., L. Colombo, R. Bertoncini, and G. Mula, “Structure and chemical order of bulk
Si1−x Cx amorphous alloys,” Phys. Rev. B 58, 10357 (1998).
Murray, R.T., C.J. Keily, and M. Hopkinson, “Crack formation in III-V epilayers grown
under tensile strain on InP(001) substrates,” Philos. Mag. 74, 383 (1996).
Murthy, C.S., S.F. O’Shea, and I.R. McDonald, “Electrostatic interactions in molecular
crystals. Lattice dynamics of solid nitrogen and carbon dioxide,” Mol. Phys. 50, 531
(1983).
Muskhelishvili, N.I., Some Basic Problems of the Mathematical Theory of Elasticity (P.
Noordhoff, Groningen, Holland, 1953).
Nakamura, T., and D.M. Parks, “Three-dimensional stress field near the crack front of a
thin elastic plate,” J. Appl. Mech. 55, 805 (1988).
Nakano, A., R.K. Kalia, and P. Vashishta, “Growth of pore interfaces and roughness of
fracture surfaces in porous silica: million particle molecular-dynamics simulations,”
Phys. Rev. Lett. 73, 2336 (1994).
Nakano, A., R.K. Kalia, and P. Vashishta, “Dynamics and morphology of brittle cracks: a
molecular-dynamics study of silicon nitride,” Phys. Rev. Lett. 75, 3138 (1995).
Namgoong, E., and J.S. Chun, “The effect of ultrasonic vibration on hard chromium plating
in a modified selfregulating high speed bath,” Thin Solid Films 120, 153 (1984).
Narayan, O., and D.S. Fisher, “Dynamics of sliding charge-density waves in 4 − 
dimensions,” Phys. Rev. Lett. 68, 3615 (1992).
Narayan, O., and D.S. Fisher, “Nonlinear fluid flow in random media: critical phenomena
near threshold,” Phys. Rev. B 49, 9469 (1994).
Nelson, D.R., and B.I. Halperin, “Dislocation-mediated melting in two dimensions,” Phys.
Rev. B 19, 2457 (1979).
Nie, S., and S.R. Emory, “Probing single molecules and single nanoparticles by surface-
enhanced Raman scattering,” Science 275, 1102 (1997).
Niemeyer, L., L. Pietronero, and H.J. Wiesmann, “Fractal dimension of dielectric
breakdown,” Phys. Rev. Lett. 52, 1033 (1984).
Nishioka, K., and J.K. Lee, “Temperature dependence of the ideal fracture strength of a
b.c.c. crystal,” Philos. Mag. A 44, 779 (1981).
Nishioka, K., S. Nakamura, T. Shimamoto, and H. Fujiwara, “Lattice instability theory of
fracture,” Scripta Mettal. 14, 497 (1980).
Noble, B., Methods Based on the Wiener–Hopf Technique for the Solution of Partial
Differential Equations (Pergamon, New York, 1958).
Noble, B., and J.W. Daniel, Applied Linear Algebra, 2nd ed. (Prentice-Hall, Englewood
Cliffs, NJ, 1977).
 References 615

Nosé, S., “A unified formulation of the constant temperature molecular dynamics methods,”
J. Chem. Phys. 81, 511 (1984).
Noskov, M.D., V.R. Kukhta, and V.V. Lopatin, “Simulation of the electrical discharge
development in inhomogeneous insulators,” J. Phys. D 28, 1187 (1995).
Nyikos, L., and T. Pajkossy, “Short communication—Diffusion to fractal surfaces,”
Electrochim. Acta 31, 1347 (1985).
Obukhov, S.P., “The problem of directed percolation,” Physica A 101, 145 (1980).
O’Dwyer, J.J., The Theory of Electrical Conduction and Breakdown in Solid Dielectrics
(Clarendon Press, Oxford, 1973).
Ohring, M., Reliability and Failure of Electronic Materials and Devices (Academic Press,
San Diego, 1998).
Olsen, K.B., R. Madariaga, and R.J. Archuleta, “Three-dimensional dynamic simulation of
the 1992 Landers earthquake,” Science 278, 834 (1997).
Olson, T., “Improvements on Taylor’s upper bound for rigid-plastic composites,” Mater.
Sci. Eng. A 175, 15 (1994).
Omeltchenko, A., A. Nakano, R.K. Kalia, and P. Vashishta, “Structure, mechanical
properties, and thermal transport in microporous silicon nitride-molecular-dynamics
simulations on a parallel machine,” Europhys. Lett. 33, 667 (1996).
Omeltchenko, A., J. Yu, R.K. Kalia, and P. Vashishta, “Crack front propagation and fracture
in a graphite sheet: a molecular dynamics study on parallel computers,” Phys. Rev. Lett.
78, 2148 (1997).
Ordejón, P., D.A. Drabold, R.M. Martin, and M.P. Grumbach, “Linear system-size scaling
methods for electronic-structure calculations,” Phys. Rev. B 51, 1456 (1995).
Orowan, E., “Crystal plasticity—III. On the mechanism of the glide process,” Z. Phys. 89,
605 (1934).
Orowan, E., “Energy criteria of fracture,” Weld. Res. Supp. 34, 157 (1955).
Ortiz, M., and R. Phillips, “Nanomechanics of defects in solids,” Adv. Appl. Mech. 36, 1
(1999).
Pande, C.S., L.E. Richards, N. Louat, B.D. Dempsey, and A.J. Schwoeble, “Fractal
characterization of fractured surfaces,” Acta Metall. 35, 1633 (1987).
Pannetta, C., L. Reggiani, and Gy. Trefán, “Scaling and universality in electrical failure of
thin films,” Phys. Rev. Lett. 84, 5006 (2000).
Parleton, L.G., “Determination of the growth of branched cracks by numerical methods,”
Eng. Fract. Mech. 11, 343 (1979).
Parr, R.G., and W. Yang, Density Functional Theory of Atoms and Molecules (Oxford
University Press, New York, 1989).
Parrinello, M., and A. Rahman, “Polymorphic transitions in single crystals: A new molecular
dynamics method,” J. Appl. Phys. 52, 7182 (1981).
Paskin, P., A. Gohar, and G.J. Dienes, “Computer simulation of crack propagation,” Phys.
Rev. Lett. 44, 940 (1980).
Paskin, P., D.K. Som, and G.J. Dienes, “Computer simulation of crack propagation: lattice
trapping,” J. Phys. C 14, L171 (1981).
Passoja, D.E., Adv. Ceramics 22, 101 (1988).
Passoja, D.E., and D.J. Amborski, Microstruct. Sci. 6, 143 (1978).
Payne, M.C., M.P. Teter, D. Allen, T.A. Arias, and J.D. Joannopoulos, “Iterative minimiza-
tion techniques for ab initio total-energy calculations: molecular dynamics and conjugate
gradients,” Rev. Mod. Phys. 64, 1045 (1992).
Pederson, M.R., and K.A. Jackson, “Pseudoenergies for simulations on metallic systems,”
Phys. Rev. B 43, 7312 (1991).
616 References

Perdew, J.P., K. Burke, and M. Ernzerhof, “Generalized gradient approximation made

simple,” Phys. Rev. Lett. 77, 3865 (1996).
Perdew, J.P., and A. Zunger, “Self-interaction correction to density-functional approxima-
tions for many-electron systems,” Phys. Rev. B 23, 5048 (1981).
Pérez, R., and P. Gumbsch, “Directional Anisotropy in the cleavage fracture of silicon,”
Phys. Rev. Lett. 84, 5347 (2000).
Perrin, G., and J.R. Rice, “Disordering of a dynamic planar crack in a model elastic medium
of randomly variable toughness,” J. Mech. Phys. Solids 42, 1047 (1994).
Phillips, R., Crystals, Defects, and Microstructures: Modeling across Scales (Cambridge
University Press, Cambridge, 2001).
Phoenix, S.L., and I.J. Beyerlein, “Distributions and size scalings for strength in a one-
dimensional random lattice with load distribution to nearest and next-nearest neighbors,”
Phys. Rev. E 62, 1622 (2000).
Phoenix, S.L., and R.L. Smith, “A comparison of probabilistic techniques for the strength
of fibrous materials under local load-sharing among fibers,” Int. J. Solids Struct. 19, 479
(1983).
Phoenix, S.L., and R. Raj, “Scalings in fracture probabilities for a brittle matrix fiber
composite,” Acta Metallurg. Mater. 40, 2813 (1992).
Pickard, C.J., B. Winkler, R.K. Chen, M.C. Payne, M.H. Lee, J.S. Lin, J.A. White, V.
Milman, and D. Vanderbilt, “Structural properties of Lanthanide and actinide compounds
within the plane wave pseudopotential approach,” Phys. Rev. Lett. 85, 5122 (2000).
Pietronero, L., and H.J. Wiesmann, “From physical dielectric breakdown to the stochastic
fractal model,” Z. Phys. B 70, 87 (1988).
Plimpton, S., “Fast parallel algorithms for short-range molecular dynamics,” J. Comput.
Phys. 117, 1 (1995).
Poirier, C., M. Ammi, D. Bideau, and J.-P. Troadec, “Experimental study of the geometrical
effects in the localization of deformation,” Phys. Rev. Lett. 68, 216 (1992).
Polanyi, M., Z. Phys. 89, 85 (1934).
Ponte Castañeda, P., “The effective mechanical properties of nonlinear isotropic compos-
ites,” J. Mech. Phys. Solids, 39, 45 (1991a).
Ponte Castañeda, P., “The effective properties of brittle/ductile incompressible composites,”
in Inelastic deformation of composite materials, edited by G.J. Dvorak (Springer, New
York, 1991b), p. 215.
Ponte Castañeda, P., “New variational principles in plasticity and their application to
composite materials,” J. Mech. Phys. Solids 40, 1757 (1992a).
Ponte Castañeda, P., “Bounds and estimates for the properties of nonlinear heterogeneous
systems,” Philos. Trans. R. Soc. London A 340, 531 (1992b).
Ponte Castañeda, P., “A new vanational principle and its application to nonlinear
heterogeneous systems,” SIAM J. Appl. Math. 52, 1321 (1992c).
Ponte Castañeda, P., “Exact second-order estimates for the effective mechanical properties
of nonlinear composite materials,” J. Mech. Phys. Solids 44, 827 (1996a).
Ponte Castañeda, P., “Variational methods for estimating behaviour of nonlinear compos-
ite materials,” in Continuum Models and Discrete systems (CMDS 8), edited by K.Z.
Markov, (World Scientific, Singapore, 1996b), p. 268.
Ponte Castañeda, P., “Nonlinear composite materials: Effective constitutive behavior and
microstructure evolution,” in Continuum Micromechanics, edited by P. Suquet (CISM
Courses and Lecture Notes No. 377, Springer, Wien, 1997), p. 131.
Ponte Castañeda, P., “Three-point bounds and other estimates for strongly nonlinear
composites,” Phys. Rev. B 57, 12077 (1998).
 References 617

Ponte Castañeda, P., and G. deBotton, “On the homogenized yield strength of two-phase
composites,” Proc. R. Soc. Lond. A 438, 419 (1992).
Ponte Castañeda, P., G. deBotton, and G. Li, “Effective properties of nonlinear inhomoge-
neous dielectrics,” Phys. Rev. B 46, 4387 (1992).
Ponte Castañeda, P., and M. Kailasam, “Nonlinear electrical conductivity in heterogeneous
media,” Proc. R. Soc. Lond. A 453, 793 (1997); 453, 1791(E).
Ponte Castañeda, P., and M.V. Nebozhyn, “Exact second-order estimates of the self-
consistent type for nonlinear composite materials,” Mech. Mater. (1997).
Ponte Castañeda, P., and P. Suquet, “On the effective mechanical behavior of weakly
inhomogeneous nonlinear materials,” Europ. J. Mech. A Solids 14, 205 (1995).
Ponte Castañeda, P., and P. Suquet, “Nonlinear composites,” Adv. Appl. Mech. 34 (Academic
Press, San Diego, 1998), p. 172.
Ponte Castañeda, P., and J.R. Willis, “On the overall properties of nonlinearly viscous
composites,” Proc. R. Soc. Lond. A 416, 217 (1988).
Ponte Castañeda, P., and J.R. Willis, “The effective behavior of nonlinear composites:
A comparison between two methods,” in Continuum Models and Discrete Systems
(CMDS 7), edited by K.H. Anthony and H.-H. Wagner (Trans Tech, Aedermannsdorf,
Switzerland, 1993), p. 351.
Ponte Castañeda, P., and J.R. Willis, “The effect of spatial distribution on the effective
behavior of composite materials and cracked media,” J. Mech. Phys. Solids 43, 1919
(1995).
Poon, C.Y., R.S. Sayles, and T.A. Jones, “Surface measurement and fractal characterization
of naturally fractured rocks,” J. Appl. Phys. 25, 1269 (1992).
Porto, M., S. Havlin, S. Schwarzer, and A. Bunde, “Optimal path in strong disorder and
shortest path in invasion percolation with trapping,” Phys. Rev. Lett. 79, 4060 (1997).
Prakash, V., and R.J. Clifton, “Experimental and analytical investigation of dynamic fracture
under conditions of plane strain,” in Fracture Mechanics: 22nd Symposium, vol. 1,
edited by H.A. Ernst, A. Saxena, and D.L. McDowell (American Society for Testing and
Materials, Philadelphia, 1992), p. 412.
Pratt, A.K., and P.L. Green, “Measurement of the dynamic fracture toughness of
polymethylmethacrylate by high-speed photography,” Eng. Fract. Mech. 6, 71 (1974).
Pulay, P., “Ab initio calculation of force constants and equilibrium geometries in polyatomic
molecules. I. Theory,” Mol. Phys. 17, 197 (1969).
Qiu, S.-Y., C.Z. Wang, K.M. Ho, and C.T. Chan, “Tight-binding molecular dynamics with
linear system-size scaling,” J. Phys.: Condens. Matter 6, 9153 (1994).
Qiu, Y.P., and G.J. Weng, “A theory of plasticity for porous materials and particle-reinforced
composites,” J. Appl. Mech. 59, 261 (1992).
Rahman, A., “Correlation in the motion of atoms in liquid argon,” Phys. Rev. 136, A405
(1964).
Rahman, A., and F.H. Stillinger, “Molecular dynamics study of liquid water,” J. Chem.
Phys. 55, 3336 (1971).
Raine, A.R.C., D. Fincham, and W. Smith, “Systolic loop methods for molecular dynamics
simulation using multiple transputers,” Comput. Phys. Commun. 55, 13 (1989).
Räisänen, V.I., M.J. Alava, K.J. Niskanen, and R.M. Nieminen, “Does the shear-lag model
apply to random fiber networks?” J. Mater. Res. 12, 2725 (1997).
Räisänen, V.I., E.T. Seppala, M.J.Alava, and P.M. Duxbury, “Quasistatic cracks and minimal
energy surfaces,” Phys. Rev. Lett. 80, 329 (1998).
Rambaldi, S., and O. Pinazza, “An accurate fractional Brownian motion generator,” Physica
A 208, 21 (1994).
618 References

Ramanathan, S., and D.S. Fisher, “Dynamics and instabilities of planar tensile cracks in
heterogeneous media,” Phys. Rev. Lett. 79, 877 (1997).
Ramanathan, S., and D.S. Fisher, “Onset of propagation of planar cracks in heterogeneous
media,” Phys. Rev. B 58, 6026 (1998).
Rammal, R., and A.-M.S. Tremblay, “Resistance noise in nonlinear resistor networks,”
Phys. Rev. Lett. 58, 415 (1987).
Ramulu, M., and A.S. Kobayashi, “Mechanics of crack curving and branching—a dynamic
fracture analysis,” Int. J. Fract. 27, 187 (1985).
Ramulu, M., and A.S. Kobayashi, “Strain energy density criteria for dynamic fracture and
dynamic crack branching,” Theore. Appl. Fract. Mech. 5, 117 (1986).
Rapaport, D.C., “Multi-million particle molecular dynamics. III. Design consideration for
data-parallel processing,” Comput. Phys. Commun. 76, 301 (1993).
Rapaport, D.C., The Art of Molecular Dynamics (Cambridge University Press, London,
1995).
Rappe, A., and J.D. Joannopoulos, in Computer Simulation in Materials Science, edited by
M. Meyer and V. Pontikis, NATO ASI Vol. 205, p. 409 (1991).
Rautiainen, T.T., M.J. Alava, and K. Kaski, “Dynamics of fracture in dissipative systems,”
Phys. Rev. E 51, R2727 (1998).
Ravi-Chandar, K., and W.G. Knauss, “Dynamic crack-tip stress under stress wave loading—
a comparison of theory and experiment,” Int. J. Fract. 20, 209 (1982).
Ravi-Chandar, K., and W.G. Knauss, “An experimental investigation into dynamic fracture.
I. Crack initiation and arrest,” Int. J. Fract. 25, 247 (1984a).
Ravi-Chandar, K., and W.G. Knauss, “An experimental investigation into dynamic fracture:
II. Microstructural aspects,” Int. J. Fract. 26, 65 (1984b).
Ravi-Chandar, K., and W.G. Knauss, “An experimental investigation into dynamic frac-
ture: III. On steady-state crack propagation and crack branching,” Int. J. Fract. 26, 141
(1984c).
Ravi-Chandar, K., and B. Yang, “On the role of microcracks in the dynamic fracture of
brittle materials,” J. Mech. Phys. Solids 45, 535 (1997).
Ray J.R., “Elastic constants and statistical ensembles in molecular dynamics,” Comput.
Phys. Rep. 8, 109 (1988).
Ray, P., and B.K. Chakrabarti, “The critical behaviour of fracture properties of dilute brittle
solids near the percolation threshold,” J. Phys. C 18, L185 (1985a).
Ray, P., and B.K. Chakrabarti, “A microscopic approach to the statistical fracture analysis
of disordered brittle solids,” Solid State Commun. 53, 477 (1985b).
Ray, P., and B.K. Chakrabarti, “Strength of disordered solids,” Phys. Rev. B 38, 715 (1988).
Reidle, J., P. Gumbsch, H.F. Fischmeister, V.G. Glebovsky, and V.N. Semenov, “Fracture
studies of tungsten single crystals,” Mater. Lett. 20, 311 (1994).
Reuss, A., “Calculation of the flow limits of mixed crystals on the basis of the plasticity of
the monocrystals,” Z. Angew. Math. Mech. 9, 49 (1929).
Rice, J.R., “Mathematical analysis in the mechanics of fracture,” in Fracture: An Advanced
Treatise, Vol. II, edited by H. Liebowitz (Academic Press, New York, 1968), p. 191.
Rice, J.R., “On the structure of stress-strain relations for time-dependent plastic deformation
in metals,” J. Appl. Mech. 37, 728 (1970).
Rice, J.R., and G.E. Beltz, “The activation energy for dislocation nucleation at a crack,” J.
Mech. Phys. Solids 42, 333 (1994).
Rice, J.R., Y. Ben-Zion, and K.S. Kim, “Three dimensional perturbation solution for a
dynamic planar crack moving unsteadily in a model elastic solid,” J. Mech. Phys. Solids
42, 813 (1994).
 References 619

Rice, J.R., and R. Thomson, “Ductile versus brittle behaviour of crystals,” Philos. Mag. 29,
73 (1974).
Robertson, D.H., D.W. Brenner, and J.W. Mintmire, “Energetics of nanoscale graphitic
tubules,” Phys. Rev. B 45, 12592 (1992).
Robertson, M.C., C.G. Sammis, M. Sahimi, and A.J. Martin, “Fractal analysis of three-
dimensional spatial distributions of earthquakes with a percolation interpretation,” J.
Geophys. Res. 100B, 609 (1995).
Rodbell, K.P., M.V. Rodriguez, and P.J. Ficalora, “The kinetics of electromigration,” J.
Appl. Phys. 61, 2844 (1987).
Rodgers, S.T., and K.F. Jensen, “Multiscale modeling of chemical vapor deposition,” J.
Appl. Phys. 83, 524 (1998).
Ronsin, O., F. Heslot, and B. Perrin, “Experimental study of quasistatic brittle crack
propagation,” Phys. Rev. Lett. 75, 2252 (1995).
Rosakis, A.J., J. Duffy, and L.B. Freund, “The determination of dynamic fracture toughness
of AISI 4340 steel by the shadow spot method,” J. Mech. Phys. Solids 32, 443 (1984).
Rosakis, A.J., and L.B. Freund, “The effect of crack tip plasticity on the determination of
dynamic stress intensity factors by the optical method of caustics,” J. Appl. Mech. 48,
302 (1981).
Rosakis, A.J., O. Samudrala, and D. Coker, “Cracks faster than the shear wave speed,”
Science 284, 1337 (1999).
Rosakis, A.J., and A.T. Zehnder, “On the dynamic fracture of structural metals,” Int. J.
Fract. 27, 169 (1985).
Roth, J., F. Gähler, and H.-R. Trebin, “A molecular dynamics run with 5180116000
particles,” Int. J. Mod. Phys. C 11, 317 (2000).
Roux, S., and D. Francois, “A simple model for ductile fracture of porous materials,” Scripta
Metall. 25, 1087 (1991).
Roux, S., A. Hansen, and E. Guyon, “Criticality in non-linear transport properties of
heterogeneous materials,” J. Physique 48, 2125 (1987).
Roux, S., A. Hansen, H.J. Herrmann, and E. Guyon, “Rupture of heterogeneous media in
the limit of infinite disorder,” J. Stat. Phys. 52, 251 (1988).
Roux, S., and H.J. Herrmann, “Disorder-induced nonlinear conductivity,” Europhys. Lett.
4, 1227 (1987).
Rudd, E., and J.Q. Broughton, “Coarse-grained molecular dynamics and the atomic limit
of finite elements,” Phys. Rev. B 58, 5893 (1998).
Rundle, J.B., and W. Klein, “Nonclassical nucleation and growth of cohesive tensile cracks,”
Phys. Rev. Lett. 63, 171 (1989).
Runge, E., and E.K.U. Gross, “Density-functional theory for time-dependent systems,”
Phys. Rev. Lett. 52, 997 (1984).
Ryckaert, J.P., G. Ciccotti, and H.J.C. Berendsen, “Numerical integration of the cartesian
equations of motion of a system with constraints: Molecular dynamics of n-alkane,” J.
Comput. Phys. 23, 327 (1977).
Ryckaert, J.P., “Special geometrical constraints in the molecular dynamics of chain
molecules,” Mol. Phys. 55, 549 (1985).
Sachs, G., “Zur Ableitung einer Fleissbedingun,” Z. Ver. Dtsch. Ing. 72, 734 (1928).
Safonov, V.P., V.M. Shalaev, V.A. Markel, Y.E. Danilova, N.N. Lepeshkin, W. Kim, S.G.
Rautian, and R.L. Armstrong, “Spectral dependence of selective photomodification in
fractal aggregates of colloidal particles,” Phys. Rev. Lett. 80, 1102 (1998).
Sahimi, M., “Nonlinear transport processes in disordered media,” AIChE J. 39, 369 (1993a).
620 References

Sahimi, M., “Flow phenomena in rocks: From continuum models to fractals, percolation,
cellular automata and simulated annealing,” Rev. Mod. Phys. 65, 1395 (1993b).
Sahimi, M., Applications of Percolation Theory (Taylor and Francis, London, 1994a).
Sahimi, M., “Long-range correlated percolation and flow and transport in heterogeneous
porous media,” J. Physique I France 4, 1263 (1994b).
Sahimi, M., “Effect of long-range correlations on transport phenomena in disordered
media,” AIChE J. 41, 229 (1995a).
Sahimi, M., Flow and Transport in Porous Media and Fractured Rock (VCH, Weinheim,
Germany, 1995b).
Sahimi, M., “Non-linear and non-local transport processes in heterogeneous media: From
long-range correlated percolation to fracture and materials breakdown,” Phys. Rep. 306,
295 (1998).
Sahimi, M., and S. Arbabi, “Force distribution, multiscaling, and fluctuations in disordered
elastic media,” Phys. Rev. B 40, 4975 (1989).
Sahimi, M., and S. Arbabi, “On correction to scaling for two- and three-dimensional scalar
and vector percolation,” J. Stat. Phys. 62, 453 (1991).
Sahimi, M., and S. Arbabi, “Percolation and fracture in disordered solids and granular
media: Approach to a fixed point,” Phys. Rev. Lett. 68, 608 (1992).
Sahimi, M., and S. Arbabi, “Mechanics of disordered solids. III. Fracture properties,” Phys.
Rev. B 47, 713 (1993).
Sahimi, M., and S. Arbabi, “Scaling laws for fracture of heterogeneous materials and rock,”
Phys. Rev. Lett. 77, 3689 (1996).
Sahimi, M., and J.D. Goddard, “Elastic percolation models for cohesive mechanical failure
in heterogeneous systems,” Phys. Rev. B 33, 7848 (1986).
Sahimi, M., M.C. Robertson, and C.G. Sammis, “Fractal distribution of earthquake hypocen-
ters and its relation with fault patterns and percolation,” Phys. Rev. Lett. 70, 2186
(1993).
Sammonds, P.R., P.G. Meredith, M.R. Ayling, and S.A. Murell, in Fracture of Concrete and
Rock, edited by S.P. Shah, S.E. Swartz, and B. Barr (Elsevier, London, 1989), p. 101.
Sánchez, A., F. Guinea, E. Louis, and V. Hakim, “On the fractal characteristics of the eta
model,” Physica A 191, 123 (1992).
Sanchez-Palancia, E., Non-homogeneous Media and Vibration Theory (Lecture notes in
Physics) 127 (Springer, Heidelberg, 1980), p. 12.
Sanz-Serna, J.M., “Symplectic integration for Hamiltonian systems: An overview,” Acta
Numerica 1, 243 (1992).
Sarychev, A.K., D.J. Bergman, and Y. Yagil, “Optical and microwave properties of metal-
insulator thin films: possibility of light localization,” Physica A 207, 372 (1994).
Sarychev,A.K., D.J. Bergman, and Y. Yagil, “Theory of the optical and microwave properties
of metal-dielectric films,” Phys. Rev. B 51, 5366 (1995).
Sarychev, A.K., and V.M. Shalaev, “Electromagnetic field fluctuations and optical
nonlinearities in metal-dielectric composites,” Phys. Rep. 335, 275 (2000).
Sarychev, A.K., and V.M. Shalaev, “Giant high-order field moments in metal-dielectric
composites,” Physica A 266, 115 (1999).
Sarychev, A.K., V.A. Shubin, and V.M. Shalaev, “Anderson localization of surface plasmons
and nonlinear optics of metal-dielectric composites,” Phys. Rev. B 60, 16389 (1999).
Sawada, Y., S. Ohta, M. Yamazaki, and H. Honjo, “Self-similarity and a phase-transition-
like behavior of a random growing structure governed by a nonequilibrium parameter,”
Phys. Rev. A 26, 3557 (1982).
 References 621

Schardin, H., “Velocity effects in fracture,” in Fracture, edited by B.L. Averbach (MIT
Press, Cambridge, MA, 1959), p. 297.
Schimschak, M., and J. Krug, “Electromigration-induced breakup of two-dimensional
voids,” Phys. Rev. Lett. 80, 1674 (1998).
Schmittbuhl, J., S., Gentier, and S. Roux, S., “Field measurements of the roughness of fault
surfaces,” Geophys. Res. Lett. 20, 639 (1993).
Schmittbuhl, J., S. Roux, J.-P. Vilotte, and K.J. Måløy, “Interfacial crack pinning: effect of
nonlocal interactions,” Phys. Rev. Lett. 74, 1787 (1995).
Schöen, M., “Structure of a simple molecular dynamics. II. Design considerations for
distributed processing,” Comput. Phys. Commun. 52, 175 (1989).
Schwartz, L.M., S. Feng, M.F. Thorpe, and P.N. Sen, “Behavior of depleted elastic networks:
Comparison of effective-medium and numerical calculations,” Phys. Rev. B 32, 4607
(1985).
Scott, I.G., “Basic Acoustic Emission,” in Nondestructive Testing Monographs and Tracts,
Vol. 6 (Gordon and Breach New York, 1991), p. 124.
Shalaev, V.M., R. Botet, and A.V. Butenko, “Localization of collective dipole excitations
on fractals,” Phys. Rev. B 48, 6662 (1993).
Shalaev, V.M., R. Botet, J. Mercer, and E.B. Stechel, “Optical properties of self-affine thin
films,” Phys. Rev. B 54, 8235 (1996a).
Shalaev, V.M., C. Douketis, T. Haslett, T. Stuckless, and M. Moskovits, “Two-photon elec-
tron emission from smooth and rough metal films in the threshold region,” Phys. Rev. B
53, 11193 (1996b).
Shalaev, V.M., V.A. Markel, E.Y. Poliakov, R.L. Armstrong, V.P. Safonov, A.K. Sarychev,
“Nonlinear optical phenomena in nanostructured fractal materials,” J. Nonlin. Optic.
Phys. & Mat. 7, 131 (1998).
Shalaev, V.M., and A.K. Sarychev, “Nonlinear optics of random metal-dielectric films,”
Phys. Rev. B 57, 13265 (1998).
Sharon, E., and J. Fineberg, “Microbranching instability and the dynamic fracture of brittle
materials,” Phys. Rev. B 54, 7128 (1996).
Sharon, E., and J. Fineberg, “Universal features of the micro-branching instability in
dynamic fracture,” Philos. Mag. B 78, 243 (1998).
Sharon, E., and J. Fineberg, “On massless cracks and the continuum theory of fracture,”
Nature 397, 333 (1999).
Sharon, E., S.P. Gross, and J. Fineberg, “Local crack branching as a mechanism for instability
in dynamic fracture,” Phys. Rev. Lett. 74, 5096 (1995).
Sharon, E., S.P. Gross, and J. Fineberg, “Energy dissipation in dynamic fracture,” Phys.
Rev. Lett. 76, 2117 (1996).
Sharp, S.J., M.F. Ashby, and N.A. Fleck, “Material response under static and sliding
indentation loads,” Acta Met. 41, 685 (1993).
Shenoy, V.B., R. Miller, E.B. Tadmor, R. Phillips, and M. Ortiz, “Quasicontinuum models
of interfacial structure and deformation,” Phys. Rev. Lett. 80, 742 (1998).
Shenoy, V.B., R. Miller, E.B. Tadmor, D. Rodney, and R. Phillips, “An adaptive finite
element approach to atomic-scale mechanics—the quasicontinuum method,” J. Mech.
Phys. Solids 47, 611 (1999).
Shenoy, V.B., R. Phillips, and E.B. Tadmor, “Nucleation of dislocation beneath a plane
strain indenter,” J. Mech. Phys. Solids 48, 649 (2000).
Shimojo, F., I. Ebbsjö, R.K. Kalia, A. Nakano, J.P. Rino, and P. Vashishta, “Molecular
dynamics simulation of structural transformation in silicon carbide under pressure,”
Phys. Rev. Lett. 84, 3338 (2000).
622 References

Shinkai, N., “Fracture and fractography of flat glass,” in Fractography of Glass, edited by
R.C. Bradt and R.E. Tressler (Plenum, New York, 1994), p. 253.
Shioya, T. and R. Ishida, “Microscopic fracture modes of brittle polymers in dynamic crack
propagation,” in Dynamic Failure of Materials, edited by H.P. Rossmanith and A.J.
Rosakis (Elsevier Applied Science, Essex, 1991), p. 351.
Shirley, E.L., D.C. Allan, R.M. Martin, and J.D. Joannopoulos, “Extended norm-conserving
pseudopotentials,” Phys. Rev. B 40, 3652 (1989).
Sieradzki, K., G.J. Dienes, A. Paskin, and B. Massoumzadeh, “Atomistics of crack
propagation,” Acta Mettal. 36, 651 (1988).
Sieradzki, K., and R. Li, “Fracture behavior of a solid with random porosity,” Phys. Rev.
Lett. 56, 2509 (1986).
Sieradzki, K., and R.C. Newman, “Brittle behaviour of ductile metals during stress-
corrosion cracking,” Philos. Mag. A 51, 95 (1985).
Sig, G.C., “Some basic problems in fracture mechanics and new concepts,” Eng. Fract.
Mech. 5, 365 (1973).
Sinclair, J.E., and B.R. Lawn, “An atomistic study of cracks in diamond-structure crystals,”
Proc. R. Soc. Lond. A 329, 83 (1972).
Skjeltorp, A.T., and P. Meakin, “Fracture in microsphere monolayers studied by experiment
and computer simulation,” Nature 335, 424 (1988).
Slepyan, L., “Dynamics of a crack in a lattice,” Sov. Phys. Dokl. 26, 538 (1981).
Sloan, S.W., “A fast algorithm for constructing Delaunay triangulations in the plane,” Adv.
Engng. Software, 9, 34 (1987).
Smekal, E., “Zum Bruchvorgang bei sprodem stoffrerhalten unter ein und mehrachsinen
beanspruchungen,” Osterr. Ing. Arch. 7, 49 (1953).
Smith, R.L., S.L. Phoenix, M.R. Greenfield, R.B. Henstenburg, and R.E. Pitt, “Lower-tail
approximations for the probability of failure of three-dimensional fibrous composites
with hexagonal geometry,” Proc. R. Soc. Lond. A 388, 353 (1983).
Smith, W., and T.R. Forester, “Parallel macromolecular simulations and the replicated data
strategy II: The RD-SHAKE algorithm,” Comput. Phys. Commun. 79, 63 (1994).
Smyshlyaev, V.P., and N.A. Fleck, “Bounds and estimates for the overall plastic behavior
of composites with strain gradients effects,” Proc. R. Soc. Lond. A 451, 795 (1995).
Söderberg, M., “Resistive breakdown of inhomogeneous media,” Phys. Rev. B 35, 352
(1987).
Sornette, D., “Weibull-like failure distribution induced by fluctuations in percolation,” J.
Physique 49, 889 (1988).
Sornette, D., “Elasticity and failure of a set of elements loaded in parallel,” J. Phys. A 22,
L243 (1989).
Sornette, D., and C. Vanneste, “Dynamics and memory effects in rupture of thermal fuse
networks,” Phys. Rev. Lett. 68, 612 (1992).
Sornette, D., and C. Vanneste, “Dendrites and fronts in a model of dynamical rupture with
damage,” Phys. Rev. E 50, 4327 (1994).
Soules, T.F., and R.F. Busbey, “The rheological properties and fracture of a molecular
dynamic simulation of sodium silicate glass,” J. Chem. Phys. 78, 6307 (1983).
Spence, J.C.H., Y.M. Huang, and O. Sankey, “Lattice trapping and surface reconstruction
for silicon cleavage on (111). Ab-initio quantum molecular dynamics calculations,” Acta
Mettal. Mater. 41, 2815 (1993).
Spitzig, W.A., R.E. Smelser, and O. Richmond, “The evolution of damage and fracture in
iron compacts with various initial porosities,” Acta Metall. Mater. 36, 1201 (1998).
 References 623

Sridhar, N., W. Yang, D.J. Srolovitz, and E.R. Fuller, Jr., “Microstructral mechanics model
of anisotropic-thermal-expansion-induced microcracking,” J. Am. Ceramic Soc. 77, 1123
(1994).
Srolovitz, D.J., and P.D. Beale, “Computer simulation of failure in an elastic model with
randomly distributed defects,” J. Amer. Cer. Soc. 71, 362 (1988).
Stadler, J., R. Mikulla, and H.-R. Trebin, “IMD: a software package for molecular dynamics
studies on parallel computers,” Int. J. Mod. Phys. C 8, 1131 (1997).
Stanley, H.E., and P. Meakin, “Multifractal phenomena in physics and chemistry,” Nature
335, 405 (1988).
Starkloff, T., and J.D. Joannopoulos, “Local pseudopotential theory for transition metals,”
Phys. Rev. B 16, 5212 (1977).
Stauffer, D., and A. Aharony, Introduction to Percolation Theory, 2nd ed. (Taylor and
Francis, London, 1992).
Steele, W.A., “The physical interaction of gases with crystalline solids. I. Gas-solid energies
and properties of isolated adsorbed atoms,” Surf. Sci. 36, 317 (1973).
Stephens, M.D., and M. Sahimi, “Distribution of fracture strengths in disordered continua,”
Phys. Rev. B 36, 8656 (1987).
Stinchcombe, R.B., “Conductivity and spin-wave stiffness in disordered systems-an exactly
soluble model,” J. Phys. C 7, 179 (1974).
Stinchcombe, R.B., P.M. Duxbury, and P. Shukla, “The minimum gap on diluted Cayley
trees,” J. Phys. A 19, 3903 (1986).
Stillinger, F.H., and T.A. Weber, “Computer simulation of local order in condensed phases
of silicon,” Phys. Rev. B 31, 5262 (1985).
St. John, C., “The brittle-to-ductile transition in pre-cleaved silicon single crystals,” Philos.
Mag. 32, 1193 (1975).
Stockman, M.I., “Inhomogeneous eigenmode localization, chaos, and correlations in large
disordered clusters,” Phys. Rev. E 56, 6494 (1997).
Stockman, M.I., L.N. Pandey, and T.F. George, “Inhomogeneous localization of polar
eigenmodes in fractals,” Phys. Rev. B 53, 2183 (1996).
Stockman, M.I., L.N. Pandey, L.S. Muratov, and T.F. George, “Giant fluctuations of local
optical fields in fractal clusters,” Phys. Rev. Lett. 72, 2486 (1994).
Stockman, M.I., L.N. Pandey, L.S. Muratov, and T.F. George, “Optical absorption and
localization of eigenmodes in disordered clusters,” Phys. Rev. B 51 185 (1995)
Stoddard, S.D., and J. Ford, “Numerical experiments on the stochastic behavior of a
Lennard–Jones gas system,” Phys. Rev. A 8, 1504 (1973).
Straley, J.P., “Random resistor tree in an applied field,” J. Phys. C 10, 3009 (1977).
Straley, J.P., and S.W. Kenkel, “Percolation theory for nonlinear conductors,” Phys. Rev. B
29, 6299 (1984).
Strang, G., and G.J. Fix, An Analysis of the Finite Element Method (Prentice-Hall, Englewood
Cliffs, 1973).
Stroh, A.N., “A theory of the fracture of metals,” Adv. Phys. 6, 418 (1957).
Stroud, D., and P.M. Hui, “Nonlinear susceptibilities of granular matter,” Phys. Rev. B 37,
8719 (1988).
Stroud, D., and V.E. Wood, “Decoupling approximation for the nonlinear-optical response
of composite media,” J. Opt. Soc. Am. B6, 778 (1989).
Stroud, D., and X. Zhang, “Cubic nonlinearities in small-particle composites: local-field
induced giant enhancements,” Physica A 207, 55 (1994).
Suquet, P., “Analyse limite et homogénéisation,” C. R. Acad. Sci. Paris Ser. II 296, 1355
(1983).
624 References

Suquet, P., “Elements of homogenization for inelastic solid mechanics,” in Homogenization

Techniques for Composite Media, edited by E. Sanchez-Palencia and A. Zaoui, Lecture
Notes in Physics 272. (Springer, Berlin, 1987), p. 193.
Suquet, P., “On bounds for the overall potential of power-law materials containing voids
with arbitrary shape,” Mech. Res. Comm. 19, 51 (1992).
Suquet, P., “Overall potentials and extremal surfaces of power law or ideally plastic
materials,” J. Mech. Phys. Solids 41, 981 (1993a).
Suquet, P., “Bounds and estimates for the overall properties of nonlinear composites,” in
Micromechanics of Materials, edited by J.J. Marigo and G. Rousselier (Eyrolles, Paris,
1993b), p. 361.
Suquet, P., “Overall properties of nonlinear composites: A modified secant moduli theory
and its link with Ponte Castañieda’s nonlinear variational procedure,” C. R. Acad. Sci.
Paris Ser. II 320, 563 (1995).
Suquet, P., “Effective properties of nonlinear composites,” in Continuum Micromechanics,
edited by P. Suquet, CISM Courses and Lecture Notes No. 377 (Springer, Wien, 1997),
p. 197.
Suquet, P., and P. Ponte Castañeda, “Small-contrast perturbation expansions for the effective
properties of nonlinear composites,” C. R. Acad Sci. Paris Ser. II 317, 1515 (1993).
Suwanmethanond, V., E. Goo, P.K.T. Liu, G. Johnson, M. Sahimi, and T.T. Tsotsis, “Porous
silicon carbide sintered substrates for high-temperature membranes,” Ind. Eng. Chem.
Res. 39, 3264 (2000).
Swope, W., and H.C. Andersen, “106 -Particle molecular dynamics study of homogeneous
nucleation of crystals in a supercooled atomic liquid,” Phys. Rev. B 41, 7042 (1990).
Tadmor, E.B., M. Ortiz, and R. Phillips, “Quasicontinuum analysis of defects in solids,”
Philos. Mag. A 73, 1529 (1996a).
Tadmor, E.B., R. Phillips, and M. Ortiz, “Mixed atomistic and continuum models of
deformation in solids,” Langmuir 12, 4529 (1996b).
Tadmor, E.B., R. Miller, R. Phillips, and M. Ortiz, “Nanoindentation and incipient
plasticity,” J. Mater. Res. 14, 2233 (1999).
Takayasu, H., “Simulation of electric breakdown and resulting variant of percolation
fractals,” Phys. Rev. Lett. 54, 1099 (1985).
Talbot, D.R.S., and J.R. Willis, “Variational principles for inhomogeneous nonlinear media,”
IMA J. Appl. Math. 35, 39 (1985).
Talbot, D.R.S., and J.R. Willis, “Bounds and self-consistent estimates for the overall
properties of nonlinear composites,” IMA J. Appl. Math. 39, 215 (1987).
Talbot, D.R.S., and J.R. Willis, “Some explicit bounds for the overall behaviour of nonlinear
composites,” Int. J. Solids Struct. 29, 1981 (1992).
Talbot, D.R.S. and J.R. Willis, “Upper and lower bounds for the overall properties of a
nonlinear composite dielectric I. Random microgeometry,” Proc. R. Soc. Lond. A 447,
365 (1994).
Talbot, D.R.S., and J.R. Willis, “Three-point bounds for the overall properties of a nonlinear
composite dielectric,” IMA J. Appl. Math. 57, 41 (1996).
Tandon, G.P., and G.J. Weng, “A theory of particle-reinforced plasticity,” J. Appl. Mech. 55,
126 (1988).
Taylor, G.I., “The mechanism of plastic deformation of crystals. Part I.—Theoretical,” Proc.
R. Soc. Lond. A 145, 362 (1934).
Taylor, G.I., “Plastic strains in metals,” J. Inst. Metals 62, 307 (1938).
Termonia, Y., and P. Smith, “Theoretical study of the ultimate mechanical properties of
poly(p-phenylene-terephthalamide) fibres,” Polymer 27, 1845 (1986).
 References 625

Termonia, Y., and P. Meakin, “Formation of fractal cracks in a kinetic fracture model,”
Nature 320, 429 (1986).
Termonia, Y., P. Meakin, and P. Smith, “Theoretical study of the molecular weight on the
maximum tensile strength of polymer fibre,” Macromolecules 18, 2246 (1985).
Termonia, Y., P. Meakin, and P. Smith, “Theoretical study of the influence of strain rate and
temperature on the maximum strength of perfectly ordered and oriented polyethylene,”
Macromolecules 19, 154 (1986).
Termonia, Y., and P. Smith, “Kinetic model for tensile deformation of polymers. 1. Effect
of Molecular Weight,” Macromolecules 20, 835 (1987).
Termonia, Y., and P. Smith, “Kinetic Model for tensile deformation of polymers. 2. Effect
of entanglement spacing,” Macromolecules 21, 2184 (1988).
Tersoff, J., “New empirical approach for the structure and energy of covalent systems,”
Phys. Rev. B 37, 6991 (1988).
Tersoff, J., “Modelling solid-state chemistry: Interatomic potentials for multicomponent
systems,” Phys. Rev. B 39, 5566 (1989).
Theocaris, P.S., and E.E. Gdoutos, “An optical method for determining opening-mode and
edge sliding-mode stress intensity factors,” J. Appl. Mech. 39, 91 (1972).
Theocaris, P.S., and H.G. Georgiadis, “Bifurcation predictions for moving cracks by the
T-criterion,” Int. J. Fract. 29, 181 (1985).
Thijssen, J.M., Computational Physics (Cambridge University Press, Cambridge, 1999).
Thomson, R., “The physics of fracture,” Solid State Phys. 39, 1 (1986).
Thomson, R., C. Hsieh, and V. Rana, “Lattice trapping of fracture cracks,” J. Appl. Phys.
42, 3154 (1971).
Tokatly, I.V., and O. Pankratov, “Many-body diagrammatic expansion in a Kohn–Sham
basis: Implications for time-dependent density functional theory of excited states,” Phys.
Rev. Lett. 86, 2078 (2001).
Torquato, S., “Electrical conductivity of two-phase disordered composite media,” J. Appl.
Phys. 58, 3790 (1985a).
Torquato, S., “Bulk properties of two-phase disordered media. II. Effective conductivity of
a dilute suspension of penetrable spheres,” J. Chem. Phys. 83, 4776 (1985b).
Torrie, G.M., and J.P. Valleau, “Nonphysical sampling distributions in Monte Carlo free-
energy estimation: Umbrella sampling,” J. Comput. Phys. 23, 157 (1977).
Tsuruta, K., A. Omeltchenko, R.K. Kalia, and P. Vashishta, “Early stages of sintering of sil-
icon nitride nanoclusters: a molecular-dynamics study on parallel machines,” Europhys.
Lett. 33, 441 (1996).
Tua, P.F., and J. Bernasconi, “Monte Carlo simulations of two-dimensional randomly diluted
networks of nonlinear resistors,” Phys. Rev. B 37, 1986 (1988).
Tvergaard, V., “Analysis of tensile properties for a whisker-reinforced metal-matrix
composite,” Acta Metall. Mater. 38, 185 (1990).
Tzschichholz, F., “Peeling instability in Cosserat-like media,” Phys. Rev. B 45, 12691
(1992).
Tzschichholz, F., “Fracturing of brittle homogeneous solids: finite-size scalings,” Phys. Rev.
B 52, 9270 (1995).
Tzschichholz, F., and H.J. Herrmann, “Reaction-diffusion model for the hydration and
setting of cement,” Phys. Rev. E 53, 2629 (1996).
Tzschichholz, F., H.J. Herrmann, H.E. Roman, and M. Puff, “Beam model for hydraulic
fracturing,” Phys. Rev. B 49, 7056 (1994).
Uenoyama, T., L. Esaki, and H. Kotera, “Theory of stability in a nonlinear resistive network,”
Appl. Phys. Lett. 61, 363 (1992).
626 References

Underwood, E.E., and K. Banerji, “Fractals in fractography,” Mater. Sci. Eng. 80, 1 (1986).
Van Damme, H., F. Obrecht, P. Levitz, L. Gatineau, and C. Laroche, “Fractal viscous
fingering in clay slurries,” Nature 320, 731 (1986).
Van Damme, H., C. Laroche, and L. Gatineau, “Radial fingering in viscoelastic media, an
experimental study,” Revue Phys. Appl. 22, 241 (1987a).
Van Damme, H., C. Laroche, L. Gatineau, and P. Levitz, “Viscoelastic effects in fingering
between miscible fluids,” J. Physique 48, 1121 (1987b).
van den Born, I.C., A. Santen, H.D. Hoekstra, and J.Th.M. De Hosson, “Mechanical strength
of highly porous ceramics,” Phys. Rev. B 43, 3794 (1991).
Vanderbilt, D., “Soft self-consistent pseudopotentials in a generalized eigenvalue formal-
ism,” Phys. Rev. B 41, 7892 (1990).
Vanneste, C., and D. Sornette, “The dynamical thermal fuse model,” J. Phys. France I 2,
1621 (1992).
Verges, J.A., “Localization length in a random magnetic field,” Phys. Rev. B 57, 870 (1998).
Verlet, L., “"Computer "experiments" on classical fluids. I. Thermodynamical properties of
Lennard–Jones molecules,” Phys. Rev. 159, 98 (1967).
Voigt, W., “Ueber die Bezienhung zwischen den beiden Elasticitäts-constanten isotroper,”
Ann. Physik 38, 573 (1889).
Vold, M.J., “Computer simulation of floc formation in a colloidal suspension,” J. Colloid
Inter. Sci. 18, 684 (1963).
Voss, R.F., “Random fractal forgeries,” in Fundamental Algorithms for Computer Graphics,
edited by R.A. Earnshaw, NATO ASI Series, Vol. 17 (Springer-Verlag, Heidelberg, 1985),
p. 805.
Vu, B.Q., and V.K. Kinra, “Brittle fracture of plates in tension: static field radiated by a
suddenly stopping crack,” Eng. Fract. Mech. 15 107 (1981).
Wagner, N.J., B.L. Holian, and A.F. Voter, “Molecular-dynamics simulations of two-
dimensional materials at high strain rates,” Phys. Rev. A 45, 8457 (1992).
Wan, W.M.V., H.C. Lee, P.M. Hui, and K.W. Yu, “Mean-field theory of strongly nonlinear
random composites: Strong power-law nonlinearity and scaling behavior,” Phys. Rev. B
54, 3946 (1996).
Wang, Z.G., D.L. Chen, X.X. Jiang, S.H. Ai, and C.H. Shih, “Relationship between fractal
dimension and fatigue threshold value in dual-phase steels,” Scripta Metall. 22, 827
(1988).
Wang, Z.-G., U. Landman, R.L. Blumberg Selinger, and W.A. Gelbart, “Molecular-dynamics
study of elasticity and failure of ideal solids,” Phys. Rev. B 44, 378 (1991).
Washabaugh, P.D., and W.G. Knauss, “A reconciliation of dynamic crack velocity and
Rayleigh wave speed in isotropic brittle solids,” Int. J. Fract. 65, 97 (1994).
Webb, III, E.B., and G.S. Grest, “Liquid/vapor surface tension of metals: Embedded atom
method with charge gradient corrections,” Phys. Rev. Lett. 86, 2066 (2001).
Weichert, R., and K. Schonert, “On the temperature at the tip of a fast running crack,” J.
Mech. Phys. Solids 22, 127 (1974).
Weiner, J.H., and M. Pear, “Crack and dislocation propagation in an idealized crystal model,”
J. Appl. Phys. 46, 2398 (1975).
Whitehead, S., Dielectric Breakdown of Solids (Clarendon, Oxford, 1951).
Wiener, O., “Die theorie des mischk´’orpers fr̈ das feld des stationären strömung,” Math.-
Physichen Klasse der Königl. Sächsischen Gesellschaft der Wissenschaften 32, 509
(1912).
Wiesmann, H.J., and H.R. Zeller, “A fractal model of dielectric breakdown and
prebreakdown in solid dielectrics,” J. Appl. Phys. 60, 1770 (1986).
 References 627

Williford, R.E., “Scaling similarities between fracture surfaces, energies, and a structure
parameter,” Scripta Metall. 22, 197 (1988).
Willis, J.R., “Bounds and self-consistent estimates for the overall moduli of anisotropic
composites,” J. Mech. Phys. Solids 25, 185 (1977).
Willis, J.R., “Variational principles and bounds for the overall properties of composites,”
in Continuum Models and Discrete Systems (CMDS 2), edited by J. Provan (University
of Waterloo Press, Waterloo, Canada, 1978), p. 185.
Willis, J.R., “Variational and related methods for the overall properties of composites,” Adv.
Appl. Mech. 21, l (1981).
Willis, J.R., “The overall response of composite materials,” ASME J. Appl. Mech. 50, 1202
(1983).
Willis, J.R., “Variational estimates for the overall response of an inhomogeneous nonlinear
dielectric,” in Homogenization and Effective Moduli of Materials and Media, edited by
J.L. Ericksen, D. Kinderlehrer, R. Kohn, and J.-L. Lions (Springer, New York, 1986),
p. 247.
Willis, J.R., “The structure of overall constitutive relations for a class of nonlinear
composites,” IMA J. Appl. Math. 43, 231 (1989a).
Willis, J.R., in Micromechanics and Inhomogeniety, the Toshio Mura Anniversary Volume,
edited by G.J. Weng, M. Taya, and H. Abe (Springer, New York, 1989b), p. 581.
Willis, J.R., Elasticity: Mathematical Methods and Applications (Halston Press, New York,
1990), p. 397.
Willis, J.R., “On methods for bounding the overall properties of nonlinear composites,”
Phys. Solids 39, 73 (1991).
Willis, J.R., “On method for bounding the overall properties of nonlinear composites:
Correction and addition,” J. Mech. Phys. Solids 40, 441 (1992).
Willis, J.R., and A.B. Movchan, “Dynamic weight functions for a moving crack. I. Mode I
loading,” J. Mech. Phys. Solids 43, 319 (1995).
Willis, J.R., and A.B. Movchan, “Three dimensional dynamic perturbation of a propagating
crack,” J. Mech. Phys. Solids 45, 591 (1997).
Winkler, S., D.A. Shockey, and D.R. Curran, “Crack propagation at supersonic velocities.
I,” Int. J. Fract. 6, 151 (1970).
Witten, T.A., and L.M. Sander, “Diffusion-limited aggregation, a kinetic critical phe-
nomenon,” Phys. Rev. Lett. 47, 1400 (1981).
Wood, W.W., and F.R. Parker, “Monte Carlo equation of state of molecules interacting
with the Lennard–Jones Potential. I. A supercritical isotherm at about twice the critical
temperature,” J. Chem. Phys. 27, 720 (1957).
Wright, D.C., D.J. Bergman, and Y. Kantor, “Resistance fluctuations in random resistor
networks above and below the percolation threshold,” Phys. Rev. B 33, 396 (1986).
Wu, B.Q., and P.L. Leath, “Failure probabilities and tough-brittle crossover of heteroge-
neous materials with continuous disorder,” Phys. Rev. B 59, 4002 (1999).
Wu, B.Q., and P.L. Leath, “Fracture strength of one-dimensional systems with continuous
disorder: A single-crack approximation,” Phys. Rev. B 61, 15028 (2000).
Wu, K., and R.M. Bradley, “Theory of electromigration failure in polycrystalline metal
films,” Phys. Rev. B 50, 12468 (1994).
Xu, G., A.S. Argon, and M. Ortiz, “Nucleation of dislocation from crack tips under mixed
modes of loading: Implications for brittle against ductile behaviour of crystals,” Philos.
Mag. 72, 415 (1995).
628 References

Xu, L., M. Sahimi, and T.T. Tsotsis, “Nonequilibrium molecular dynamics simulations of
transport and separation of gas mixtures in nanoporous materials,” Phys. Rev. E 62, 6942
(2000b).
Xu, L., M.G. Sedigh, M. Sahimi, and T.T. Tsotsis, “Nonequilibrium molecular dynamics
simulation of transport of gas mixtures in nanopores,” Phys. Rev. Lett. 80, 3511 (1998).
Xu, L., T.T. Tsotsis, and M. Sahimi, “Nonequilibrium molecular dynamics simulation of
transport and separation of gases in carbon nanopores. I. Basic results,” J. Chem. Phys.
111, 3252 (1999).
Xu, L., M.G. Sedigh, T.T. Tsotsis, and M. Sahimi, “Nonequilibrium molecular dynamics
simulation of transport and separation of gases in carbon nanopores. II. Binary and ternary
mixtures and comparison with the experimental data,” J. Chem. Phys. 112, 910 (2000a).
Xu, X.P., and A. Needleman, “Numerical simulations of fast crack growth in brittle solids,”
J. Mech. Phys. Solids 42 1397(1994).
Yagil, Y., G. Deutscher, and D.J. Bergman, “Electrical breakdown measurements of
semicontinuous metal films,” Phys. Rev. Lett. 69, 1423 (1992).
Yagil, Y., G. Deutscher, and D.J. Bergman, “The role of microgeometry in the electrical
breakdown of metal-insulator mixtures,” Int. J. Mod. Phys. B 7, 3353 (1993).
Yagil, Y., G. Deutscher, and D.J. Bergman, “Nonlinear electrical response and breakdown
of semicontinuous metal films,” Physica A 207, 323 (1994).
Yagil, Y., P. Gadanne, C. Julien, and G. Deutscher, “Optical properties of thin semicontinuous
gold films over a wavelength range of 2.5 to 500 µm,” Phys. Rev. B 46, 2503 (1992).
Yan, Y., G. Li, and L.M. Sander, “Fracture growth in 2d elastic networks with Born model,”
Europhys. Lett. 10, 7 (1989).
Yang, C.S., and P.M. Hui, “Effective nonlinear response in random nonlinear resistor
networks: numerical studies,” Phys. Rev. B 44, 12559 (1991).
Yoffe, E.H., “The moving Griffith crack,” Philos. Mag. (series 7) 42, 739 (1951).
Yonenaga, I., and K. Sumino, “Mechanical properties and dislocation dynamics of GaP,” J.
Mater. Res. 4, 355 (1989).
Young, A.P., “Melting and the vector Coulomb gas in two dimensions,” Phys. Rev. B 19,
1855 (1979).
Yoon, Y.-G., M.S.C. Mazzoni, H.J. Choi, J. Ihm, and S.G. Louie, “Structural deformation
and intertube conductance of crossed carbon nanotube junctions,” Phys. Rev. Lett. 86,
688 (2001).
Yu, K.W., and G.Q. Gu, “Electrostatic boundary-value problems of nonlinear media: a
perturbation approach,” Phys. Lett. A 168, 313 (1992).
Yu, K.W., and G.Q. Gu, “Effective conductivity of nonlinear composites. II. Effective-
medium approximation,” Phys. Rev. B 47, 7568 (1993).
Yu, K.W., and G.Q. Gu, “Variational calculation of strongly nonlinear composites,” Phys.
Lett. A 193, 311 (1994).
Yu, K.W., and G.Q. Gu, “Effective conductivity of strongly nonlinear composites:
Variational approach,” Phys. Lett. A 205, 295 (1995).
Yu, K.W., and P.M. Hui, “Percolation effects in two-component nonlinear composites:
Crossover from linear to nonlinear behavior,” Phys. Rev. B 50, 13327 (1994).
Yu, K.W., P.M. Hui, and D. Stroud, “Effective dielectric response of nonlinear composites,”
Phys. Rev. B 47, 14150 (1993).
Yuse, A., and M. Sano, “Transition between crack patterns in quenched glass plates,” Nature
362 329 (1993).
Zapperi, S., P. Ray, H.E. Stanley, and A. Vespignani, “First-order transition in the breakdown
of disordered media,” Phys. Rev. Lett. 78, 1408 (1997).
 References 629

Zehnder, A.T., and A.J. Rosakis, “On the temperature distribution at the vicinity of
dynamically propagating cracks in 4340 steel,” J. Mech. Phys. Solids 29, 385 (1991).
Zeng, X.C., D.J. Bergman, P.M. Hui, and D. Stroud, “Effective-medium theory for weakly
nonlinear composites,” Phys. Rev. B 38, 10970 (1988).
Zeng, X.C., P.M. Hui, D.J. Bergman, and D. Stroud, “Mean field theory for weakly nonlinear
composites,” Physica A 157, 192 (1989).
Zepeda-Ruiz, L.A., D. Maroudas, and W.H. Weinberg, “Theoretical study of the energet-
ics, strain fields, and semicoherent interface structures in layer-by-layer semiconductor
heteroepitaxy,” J. Appl. Phys. 85, 3677 (1999).
Zhang, G.M., “Cross-over components in percolation superconductor-nonlinear-conductor
mixtures,” Phys. Rev. B 53, 20 (1996a).
Zhang, G.M., “Higher order nonlinear response in random resistor networks: numerical
studies for arbitrary nonlinearity,” Z. Phys. B 99, 599 (1996b).
Zhang, X., and D. Stroud, “Numerical studies of the nonlinear properties of composites,”
Phys. Rev. B 49, 944 (1994).
Zhang, Y.M., and T.C. Wang, “Lattice Instability at a fast moving crack tip,” J. Appl. Phys.
80, 4332 (1996).
Zhenyi, M., S.C. Langford, J.T. Dickinson, M.H. Eengelhard, and D.R. Baer, “Scanning
tunneling microscope observations of MgO fracture surfaces,” J. Mater. Res. 6, 183
(1990).
Zhou, S.J., D.M. Beazley, P.S. Lomdahl, and B.L. Holian, “Large-scale molecular dynamics
simulations of three-dimensional ductile failure,” Phys. Rev. Lett. 78, 479 (1997).
Zhou, S.J., A.E. Carlsson, and R. Thomson, “Dislocation nucleation and crack stability:
lattice Green’s-function treatment of cracks in a model hexagonal lattice,” Phys. Rev. B
47, 7710 (1993).
Zhou, S.J., A.E. Carlsson, and R. Thomson, “Crack blunting effects on dislocation emission
from cracks,” Phys. Rev. Lett. 72, 852 (1994).
Zhou, S.J., and W.A. Curtin, “Failure of fiber composites: a lattice Green function model,”
Acta Metall. Matter 43, 3093 (1995).
Zhou, S.J., P.S. Lomdahl, R. Thomson, B.L. Holian, “Dynamic crack processes via
molecular dynamics,” Phys. Rev. Lett. 76, 2318 (1996).
Zielinski, W., H. Huang, S. Venkataraman, and W.W. Gerberich, “Dislocation distribution
under a microindentation into an iron-silicon single crystal,” Philos. Mag. A 72, 1221
(1995).
Zimmerman, C., W. Klemm, and K. Schonert, “Dynamic energy release rate and fracture
heat in polymethylmethacrylate (PMMA) and a high strength steel,” Eng. Fract. Mech.
20, 777 (1984).
Index

Ab initio computation 457, 503, 524, 530, Kosterlitz–Thouless transition 363

531 Rice–Thomson criterion 361, 540, 542
Absorption coefficient 134 transition temperature 361
AC field 225, 228, 238, 245
Activation energy 226, 444, 447 Car–Parrinello method (see Quantum
Anisotropic materials 33, 44, 194, 322 molecular dynamics) 492–504
Annealed disorder 389, 444 Caustics, method of 267
Atomic aspects of fracture 536 Cayley tree 64
Central forces 389, 405
Backbone fractal dimension 79, 82, 87, 97, Central-force percolation 541
402 Ceramic materials 286
Ballistic deposition 17, 20 Cleavage fracture (of crystalline materials)
Barenblatt’s cohesive zone 348 282
Beam model 391, 394 Cohesive energy 461, 462, 524
Beran bounds 32, 33, 36, 38, 39, 49, 50, Cohesive strength 250, 286
174, 192, 196 Cohesive zone 262, 296, 307, 316
Barenblatt–Dugdale model 348
Besquin law 263
dissipation of energy in 262, 268
Bethe lattice 64–69
failure of dynamic models 360
Bloch’s theorem 465
two-field continuum model 349
Bond-bending models 389, 390
Compliance tensor 166
Bond percolation 68
Composite materials
Born model 390 bounds to conductivity (see Bounds)
Born–Oppenheimer approximation 456 bounds to elastic properties (see
Boundary sliding 255 Bounds)
Bounds Conductivity (see also Random resistor
to effective conductivity 37, 45–51, networks)
53–60 effective-medium approximation 49, 55,
to dielectric constant 32–36, 38–40, 44 69, 88, 103, 142, 158, 192, 196,
to elastic moduli 170–173, 180–184, upper and lower bounds (see Bounds)
191–200 Conjugate-gradient method 506
Box-counting method 7 minimization of total energy 505
Breakdown statistics 243–245 Connectivity 230, 350
Brenner potentials 531 Constant-gap scaling 220
Brittle materials 252 Constitutive nonlinearity 1
Brittle-to-ductile transition 252, 261 Continuum percolation 406
in BCC transition metals 361, 362 Coordination number 64
in semiconductors 361 Corrections to scaling 401
632 Index

Correlation function 8, 9, 11, 15, 125, 126, scaling properties 241–243

139, 368 stochastic models 236–239
Corrosion crack 263, 392 Dielectric constant
Coulomb interaction 283, 455, 459, 463, resistor-capacitor model 108, 109
480, 484 scaling properties near percolation 241,
Covalent bonding 505, 527, 587 242
Crack nucleation 254, 255 upper and lower bounds (see Bounds)
Crack terminal velocity 263, 339, 541 Diffusion-limited aggregates 7, 8, 237
Crack velocity 260, 268–270, 334, 343, Dimple fracture, 288
352 Direct correlation function 8
Crazing 255 Directed percolation 84
Creep 165, 184, 288, 436 Dislocation 250, 252, 253, 254, 261
Critical current 87, 88, 95, 99 Dissipated heat (in fracture tip) 270, 323,
for electrical breakdown 216–218, 226 324
Critical exponents (see Percolation) Distribution of failure threshold 222–224,
Critical point 83, 472 372, 392–395
Crossover length 329 Drude dielectric function 107, 108, 117,
Cumulative failure probability 220–222, 119, 123
243–245, 403–406 Duality relation
Current distribution 81, 82, 91 for strongly nonlinear materials 75–77
Cyclic fatigue 263 for weakly nonlinear composites 97–99
Ductile materials 252, 254, 287, 361
Density-functional theory 459 Dugdale model 348
effective potential 461 Dynamic fracture, continuum models 263
crossover from three to two dimensions
exchange-correlation function 460, 461,
345
463, 464
energy dissipation 346
generalized gradient approximation 464
instability in amorphous materials
local density approximation 462
340–347
local spin density approximation 463
in three dimensions 317
non-local potential 457
mechanism of dynamical instability
non-periodic systems 465
343, 344
pseudo-potential approximation 465
onset of velocity oscillations 341
pseudo-wave function 467
relation to surface structure 342
time-dependent problems 461
universality of dynamical instability 347
with supercells 465
universality of microbranches profiles
Deviatoric components of stress tensor
345
164, 174, 235
Dynamic fracture, discrete models
Dielectric breakdown 209 419–441
computer simulation 213 connection to Yoffe’s instability 435
continuum models 209 effect of annealed disorder 444
discrete models 213 effect of quenched disorder 436–440
distribution of breakdown fields 243, faster than Rayleigh wave speed 338,
244 547
electric field-intensity factor 212 forbidden velocities 434
energy release rate 212 in Mode I 422
field multiplication factor 210 in Mode III 424
Griffith-like criterion 211, 212 molecular dynamics simulation
J -integral 212 536–548, 574–586
 Index 633

nonlinear instabilities 434 Fatigue 263

phonon emission 432–434 Fiber bundle 369–379
Dynamic stress 266 Fibrous materials 369
computer simulation 379–382
Edwards–Wilkinson model 19 deep minimum in failure probability 378
Effective-medium approximation 55, 69, effective-medium approximation
88, 103, 142, 158, 382 382–388
for conductor-insulator composites equal-load-sharing models 370–373
58–59, 69, 88 failure probability of 371–373, 375–379
for conductor-superconductor local-load-sharing models 373–379
composites 56 quasi-static fracture of 369
Elastic constant tensor 577 Field-theoretic approach 80
Elastic energy 385, 390 for power-law systems (see Nonlinear
Elastic incompatibility 253 resistor networks) 80–81
Elastic network (see also Rigidity Final breakdown 240, 242–244
percolation; Bond-bending models) Final voltage 242
410 Finite-difference methods 109, 421, 475
Elastic percolation networks 381, 410, Finite-element simulation 352
411 fracture 352–355
Electrical breakdown 213 Finite-size scaling 400, 401
computer simulation 218 Flicker noise 81–83
dilute limit 214 Fock operature 456
discrete dynamical models 225 Fractals 6–10
distribution of failure currents 220 Fractal dimension 6,7
effect of failure threshold 222 Fractional Brownian motion 10
power spectrum method 11
effect of sample size 215
successive random addition 14
in AC fields 228
Weierstrass–Mandelbrot method 14
in strongly disordered materials 216
Fractional Gaussian noise 11–13
thermal effects 228
Fractography 325
Electrical conductivity (see Conductivity;
Fracture energy, direct measurement of 268
Random resistor networks; Bounds)
dependence on crack velocity 334
Electromagnetic properties 104
Fracture mechanics (linear)
effective-medium approximation 103,
branching at microscopic scales 332
142, 158
conformal mapping 301, 302
Electromigration failure 232
continuum theory 296, 298–316, 322,
Elliptic defect or crack 210, 256
349–351, 355–361
Embedded-atom models 522–525, 541,
dynamic fracture in Mode I 302–306
542, 545, 561
equation of motion in Mode I 314–316
Endurance limit 263 equation of motion in Mode III 312–314
Energy release rate (in fracture) 212, 266, multiple fractures 322
267, 298, 308 shortcomings of 340
Extreme order statistics 371 static fracture in Mode III 301
Fracture mode
Failure current 215–222, 229 Modes I, II, and III 255
Failure life 263 mixed mode 256
Failure statistics 403–407 Fracture path 271, 316–318
Failure strength 400–410 Fracture pattern, oscillatory 271–273,
Failure voltage 217, 222–224, 228 342–344, 437–441
634 Index

Fracture properties Impact fracture 441–444

ceramics 286 Indentation 547, 571
fiber-reinforced composites 288 Inglis analysis, of elliptical crack 256–259
metals 287 Initial breakdown 240
metal-matrix composites 288 Interatomic potential 521–536
polymeric materials 284 Intergranular fracture 277, 278
Fracture propagation Inverted Swiss-cheese model 243
onset of 270, 295, 306–309 Ionic interaction 484–488
Fracture surface Irwin analysis 259–261
conic markings on 326
modeling 330
J -integral 212, 307
rib-like patterns 327
Johnson potential 535
structure of 325
Joule effect 208
topography of 325
Fracture surface roughness
mechanisms of 28, 282 Kardar–Parisi–Zhang equation 20
measurement of 277 Kerr optical nonlinearity 133–137
Fracture toughness 259
Fracture velocity, measurement of 268–270 Laminates 40, 41
faster than Rayleigh wave speed Lattice trapping 436, 540
338–340, 547 Leap-frog algorithm 475
high-speed photography 268 Legendre functions 75
measurement of resistivity 269 Legendre transformation 29, 164
ultrasonic measurements 269 Lennard–Jones potential 471
Fuse model 213–227 Light-emitting diode 245
Link-cell method (molecular dynamics)
Generalized Ohm’s law 147–155 511
Geometrical models 349 Loading condition 264–266
Global failure 371, 416 double-cantilever beam 264
Granular materials 173 infinite strip 264
Grain boundaries 250, 275, 283, 287 single-edge notched 265
Green functions 129, 310–311, 376, 486 Lower bound 33–37, 170–173, 190–194
Griffith’s criterion 256–259, 306
generalization for self-affine surfaces
Massively-parallel molecular dynamics
335–338
512–521
Growth exponent 16
atom-decomposition algorithms
Gumbel distribution, 221, 222, 375, 379
513–515
force-decomposition algorithms
Hackle 273–275, 284, 325, 326
Hartree–Fock theory 456 515–517
Hashin–Shtrikman bounds (see Bounds) load balance 519, 520
Height correlation function 15–18 selection of an algorithm 520
Hellmann–Feynman theorem (see spatial-decomposition algorithms
Quantum molecular dynamics) 497 517–519
Hertzian fracture 441 Maxwell–Boltzmann distribution 477
Hill’s lemma 185 Maxwell–Garnett approximation 54, 103
Hurst exponent 10, 16, 21, 22, 275–282, Mean-field approximation 382
327–330 Metal-insulator films 122
Hyper-Raman scattering 131–132 Metal-insulator transition 107
 Index 635

Microcrack 281 lower bounds 33–37

Minimum complementary energy criterion Maxwell–Garnett estimates 54
28, 168 second-order exact results 51–53
Minimum energy principle 28, 30–33, 170 upper bounds 38
Minimum gap 232 variational principles 27
Mirror zone (on fracture surface) 273 Nonlinear dielectric constant (see also
Mist zone (on fracture surface) 274 nonlinear conductivity) 43–45,
Molecular dynamics simulation 469–492 101–103
constant-energy ensemble 477 composites with infinite dielectric
constant-pressure and constant 44
temperature-ensemble 479 composites with zero dielectric constant
constant-temperature ensemble 477, 478 45
cut-and-shifted potentials 474 effective-medium approximation 103
evaluation of the forces 474, 475 exact results 102
ionic systems 484–488 Maxwell–Garnett approximation 103
Nosé-Hoover algorithm 478 Nonlinear mechanical properties
rigid and semi-rigid molecules 479–484 arbitrary heterogeneity 167
SHAKE algorithm 483 bounds based on linear materials
vectorized computation 511 173–180
Morse potential 524, 528 bounds with piecewise constant moduli
Multifractals 92, 220 180–184
Multiscale modeling Hashin–Shtrikman bounds 170, 191
handshake region 575 nonlinear materials with isotropic
parallel approach 554 phases 174, 181
sequential approach 552 polycrystalline materials 177, 182
simulation of chemical vapor deposition porous materials 190–193
587 power-law materials 176
simulation of defects 568 rigidly-reinforced materials 193–196
simulation of fracture 574–586 second-order exact results 186–200
Talbot–Willis method 170–173
Neighbor-lists method (molecular variational principle 168
dynamics) 511 weak heterogeneity 184
Noise (1/f ) 81 Nonlinear optical properties 104–155
Non-equilibrium molecular dynamics anomalous light scattering 123–128
488–492 computer simulations 120
dual control volume method 489 distribution of the electric field 108–114
Nonlinear conductivity enhancement in metal-dielectrics
approximate estimates of the effective 133–141
energy function 37, 38 enhancement of scattering 137
conductor-insulator composites 56 fluctuations below resonance 116
conductor-superconductor composites moments of the electric field 114
56 Rayleigh scattering 124
effective-medium approximation scaling properties of correlation
estimates 55 function 126–128
exact results 64–69, 75–77, 97–99 surface-enhanced Raman scattering
for composites with isotropic phases 128–133
48–51 Nonlinear resistor networks 62–83, 87–99
for power-law composites 54, 62–83 crossover to linear resistors 95
general two-phase materials 58–60 current distribution 81, 91
636 Index

decoupling approximation 73 Polymeric materials 155, 208, 285, 481

effective-medium approximation 69–73, fracture properties (see Fracture
88, 142, 158 properties) 285
exact solution for Bethe lattices 64–69 discrete models of fracture 446
field-theoretic approach 80 Post-failure regime 407
perturbation expansion 74 Power spectrum 11, 81, 342
resistance noise 81, 91 Propagation velocity 268
scaling properties 77, 78 Pulay forces (see Quantum molecular
series expansion 79 dynamics) 498
strongly nonlinear resistors 62–83
variational approach 74 Quantum molecular dynamics 492–506
weakly nonlinear 87–101 Car–Parrinello method 492, 504
Nosé-Hoover algorithm (see Molecular computational procedure 500
dynamics) dynamics of ions 496
Nucleation of cracks 253–255 Gram–Schmidt algorithm 496
Hellmann–Feynman theorem 497
linear system-size scaling 504
Orowan analysis 261, 296 orthogonalization of the wave functions
Outer elastic region 262, 297 495
Pulay forces 498
Pair correlation function 8 Verlet algorithm 494
Percolation Quasi-static approximation 147
backbone 79, 82, 87, 97, 217 Quasi-static fracture, continuum models
conductor-insulator composites 38, 56, planar 316
66 principle of local symmetry 316
conductor-superconductor composites three dimensional 317
37, 56 Quasi-static fracture, discrete models
correlation length 67, 77, 83, 216, 400 388–419
critical exponents 67, 77–79 dependence on extent of cracking 398
fractal dimensions 78, 79 distribution of fracture strength 403
rigid-superrigid composites 193, 195 shape of the fracture 395
size dependence 406
scaling properties 77–79
universal fixed points 414
Photoelasticity 267
versus percolation 411
Piecewise linear transport 155–161
with strong disorder 400
computer simulation 157
Quenched disorder 436
effective-medium approximation 158
scaling properties 58
Raman scattering 128–133
Pinning of a rough surface 21
for nonlinear materials (see Nonlinear
Piola–Kirchhoff stress tensor 353, 564
optical properties) 131
Plasticity 282 Random resistor network (see also
Plastic deformation 196, 253 Effective-medium approximation;
Plastic dissipation function 169 Percolation)
Plastic yield 253, 361 exact solution for Bethe lattices 64
Poincaré time step 476 field-theoretic method 80
Poisson’s ratio transfer-matrix method 160
universality of 414–419 Rayleigh wave speed 291, 338, 547
Polycrystalline materials 167, 170, 177, Red bonds 78
179, 182, 209 Residual stress 289, 392
 Index 637

Resistance fluctuations 81, 91 Swiss-cheese model 101, 217, 242, 243

Resistor-capacitor-inductor model 109 Symplecticity property 476, 580
Rotational invariance 391
Roughness exponent 16, 21, 277, 327, 330, Tensile strength 288, 392
335 Tersoff potentials 527–531
Thermal effects (in fracture), measurement
Sample-spanning cluster 67 of 270
Scattering intensity 123, 124, 131, 137 acoustic emissions 270
Self-affine fractals 9, 335, 546 Thin metal films 123
Self-consistent approach (see Third-harmonic coefficient 228, 229
Effective-medium approximation) Third-harmonic voltage 229
Self-similarity 6 Threshold nonlinearity 2
Series expansion Tight-binding molecular dynamics 505
for power-law conductors (see also Total correlation function 8, 9
nonlinear resistor networks) 79 Transfer-matrix method 160
SHAKE algorithm (see Molecular Twist hackle 275
dynamics)
Shear deformation 253 Ultrasonic emission 269
Shockley partial dislocation 568 Universal elastic region 261–263, 297
Silicon 251, 253 Upper critical dimension 85
Silicon carbide 529
Simulated annealing 461 Vacancies 287
Size effect in fracture 406 Variational principles 27, 168
Slit island method 21 Velocity hackle 275
Spherical harmonics 467 Verlet algorithm 475–477, 494
Statistical self-similarity 7 Von Miser stress 164
Steele potential 472
Stillinger–Weber potential 525 Wave function 456
Stress corrosion cracking 264, 265 Weakest bond 220, 393, 403
Stress intensity factor 259–261 Weibull distribution 221, 244, 372,
non-uniqueness 334 403–405
Superrigid composites 193, 195
Surface energy 210, 211 Yield point 285, 286
Surface-enhanced Raman scattering 128 Yield stress 262
Surface roughness 9, 281 Yoffe’s criterion (for dynamic fracture)
Surface plasmon resonance 107, 108 318, 435