Simulation

Methods for
Polymers
edited by

Michael Ko telyanskii
Rudolph Technologies, Inc.
Flanders, New Jersey, U.S.A.

Doros N. Theodorou
National Technical University of Athens
Athens, Greece

m
MARCEL

MARCELDEKKER,
INC.

DEKKER

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

NEWYORK BASEL
0

Although great care has been taken to provide accurate and current information,
neither the author(s) nor the publisher, nor anyone else associated with this publication, shall be liable for any loss, damage, or liability directly or indirectly caused or
alleged to be caused by this book. The material contained herein is not intended to
provide specific advice or recommendations for any specific situation.
Trademark notice: Product or corporate names may be trademarks or registered
trademarks and are used only for identification and explanation without intent to
infringe.
Library of Congress Cataloging-in-Publication Data
A catalog record for this book is available from the Library of Congress.
ISBN: 0-8247-0247-6
This book is printed on acid-free paper.
Headquarters
Marcel Dekker, Inc., 270 Madison Avenue, New York, NY 10016, U.S.A.
tel: 212-696-9000; fax: 212-685-4540
Distribution and Customer Service
Marcel Dekker, Inc., Cimarron Road, Monticello, New York 12701, U.S.A.
tel: 800-228-1160; fax: 845-796-1772
Eastern Hemisphere Distribution
Marcel Dekker AG, Hutgasse 4, Postfach 812, CH-4001 Basel, Switzerland
tel: 41-61-260-6300; fax: 41-61-260-6333
World Wide Web
http://www.dekker.com
The publisher offers discounts on this book when ordered in bulk quantities. For
more information, write to Special Sales/Professional Marketing at the headquarters
address above.
Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.
Neither this book nor any part may be reproduced or transmitted in any form or
by any means, electronic or mechanical, including photocopying, microfilming, and
recording, or by any information storage and retrieval system, without permission in
writing from the publisher.
Current printing (last digit):
10 9 8 7 6 5 4 3 2 1
PRINTED IN THE UNITED STATES OF AMERICA

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Preface

In the last 20 years, materials modeling and simulation has grown into an
essential component of research in the chemical and pharmaceutical industries. Because the field of process design is already quite mature, competitive
advances in chemical synthesis and separation operations occur primarily
through the development and use of material systems with tailored physicochemical characteristics (e.g., metallocene catalysts for polyolefin production, polymeric membranes offering a more favorable combination of
permeability and selectivity for separations, and environmentally acceptable
solvents with prescribed thermophysical properties). Furthermore, there is a
shift of emphasis from process to product design, which is intimately related
to materials behavior. With the current trend toward nanotechnology, the
scientist or engineer is often called on to develop new, often hierarchical
material structures with key characteristics in the 0.110-nm length scale, so
as to benefit from the unique mechanical, electronic, magnetic, optical, or
other properties that emerge at this scale. Materials that develop such structures through self-assembly processes, or modify their structure in response
to environmental conditions, are frequently sought.
Meeting these new design challenges calls for a quantitative understanding of structurepropertyprocessingperformance relations in materials. Precise development of this understanding is the main objective
of materials modeling and simulation. Along with novel experimental
techniques, which probe matter at an increasingly fine scale, and new
screening strategies, such as high-throughput experimentation, modeling has
become an indispensable tool in the development of new materials and
products. Synthetic polymers and biopolymers, either by themselves or in

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

iv

Preface

combination with ceramics and metals, are central to contemporary

materials design, thanks to the astonishing range of properties that can
be achieved through manipulation of their chemical constitution, molecular
organization, and morphology.
Polymer modeling and simulation has benefited greatly from advances in
computer hardware. Even more important, however, has been the development of new methods and algorithms, firmly based on the fundamental physical and chemical sciences, that permit addressing the wide
spectrum of length and time scales that govern structure and motion
in polymeric materials. It is by now generally accepted that the successful solution of materials design problems calls for hierarchical, or
multiscale, modeling and simulation, involving a judicious combination of
atomistic (<10 nm), mesoscopic (101000 nm), and macroscopic methods.
How to best link these methods together, strengthen their fundamental
underpinnings, and enhance their efficiency are very active problems of
current research.
This book is intended to help students and research practitioners in
academia and industry become active players in the fascinating and rapidly
expanding field of modeling and simulation of polymeric materials. Roughly
five years ago, we decided to embark on an effort to produce a how-to
book that would be coherent and comprehensive, encompass important
recent developments in the field, and be useful as a guide and reference to
people working on polymer simulation, such as our own graduate students.
Rather than attempt to write the whole book ourselves, we chose to draft a
list of chapters and then recruit world-class experts to write them. Contributors were instructed to make their chapters as didactic as possible,
incorporating samples of code, where appropriate; make reference to key
works, especially review papers and books, in the current literature; and
adhere to a more or less coherent notation. In our editing of the chapters, we
tried to enhance uniformity and avoid unnecessary repetition. The result is
gratifyingly close to what we envisioned. We hope that the reader will find
the progression of chapters logical, and that the occasional switches in style
and notation will serve as attention heighteners and perspective broadeners
rather than as sources of confusion.
Chapter 1 introduces basic elements of polymer physics (interactions
and force fields for describing polymer systems, conformational statistics
of polymer chains, Flory mixing thermodynamics, Rouse, Zimm, and
reptation dynamics, glass transition, and crystallization). It provides a brief
overview of equilibrium and nonequilibrium statistical mechanics (quantum and classical descriptions of material systems, dynamics, ergodicity,
Liouville equation, equilibrium statistical ensembles and connections
between them, calculation of pressure and chemical potential, fluctuation

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Preface

equations, pair distribution function, time correlation functions, and

transport coefficients). Finally, the basic principles of traditional
molecular simulation techniques (Monte Carlo (MC), molecular dynamics
(MD), Brownian dynamics (BD), transition state theory (TST)-based
analysis, and simulation of infrequent events) are discussed.
Part I focuses on the calculation of single-chain properties in various
environments. The chapter by E. D. Akten et al. introduces the Rotational
Isomeric State (RIS) model for calculating unperturbed chain properties
from atomistic conformational analysis and develops systematically an
illustrative example based on head-to-head, tail-to-tail polypropylene. The
chapter by Reinhard Hentschke addresses MD and BD simulations of single
chains in solution. It reviews Monte Carlo sampling of unperturbed chain
conformations and calculations of scattering from single chains and presents
a coarse-graining strategy for going from atomistic oligomer or short helix
simulations, incorporating solvent effects, to RIS-type representations and
persistence lengths.
Part II addresses lattice-based Monte Carlo simulations. The chapter by
Kurt Binder et al. provides excellent motivation for why such simulations
are worth undertaking. It then discusses the Random Walk (RW), NonReversal Random Walk (NRRW), and Self-Avoiding Random Walk (SAW)
models, presents algorithms for sampling these models, and discusses their
advantages and limitations. In the chapter by Tadeusz Pakula, algorithms
for the Monte Carlo simulation of fluids on a fully occupied lattice are
discussed. Applications to macromolecular systems of complex architecture and to symmetrical block copolymers are presented to illustrate the
power and generality of the algorithms, and implementation details are
explained.
Part III, by Vagelis Harmandaris and Vlasis Mavrantzas, addresses
molecular dynamics simulations. Following an exposition of basic integration algorithms, an example is given of how MD can be cast in
an unconventional ensemble. State-of-the-art multiple time integration
schemes, such as rRESPA, and constraint simulation algorithms are then
discussed. Examples of using MD as part of a hierarchical strategy for
predicting polymer melt viscoelastic properties are presented, followed by a
discussion of parallel MD and parallel tempering simulations.
Part IV, by Tushar Jain and Juan de Pablo, introduces the
Configurational Bias (CB) method for off-lattice MC simulation.
Orientational CB is illustrated with studies of water in clay hydrates, and
CB for articulated molecules is discussed as part of expanded grand
canonical MC schemes for the investigation of chains in slits and of the
critical behavior of polymer solutions. Topological CB is introduced and
the combined use of CB and parallel tempering is outlined.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

vi

Preface

Part V, by Andrey Dobrynin, focuses on simulations of charged polymer

systems (polyelectrolytes, polyampholytes). Chains at infinite dilution are
examined first, and how electrostatic interactions at various salt concentrations affect conformation is discussed, according to scaling theory and to
simulations. Simulation methods for solutions of charged polymers at finite
concentration, including explicitly represented ions, are then presented.
Summation methods for electrostatic interactions (Ewald, particle-particle
particle mesh, fast multipole method) are derived and discussed in detail.
Applications of simulations in understanding Manning ion condensation
and bundle formation in polyelectrolyte solutions are presented. This
chapter puts the recent simulations results, and methods used to obtain
them, in the context of the state of the art of the polyelectrolyte theory.
Part VI is devoted to methods for the calculation of free energy and
chemical potential and for the simulation of phase equilibria. The chapter
by Thanos Panagiotopoulos provides a lucid overview of the Gibbs
Ensemble and histogram reweighting grand canonical MC methods, as
well as of the NPT test particle, GibbsDuhem integration, and pseudoensemble methods. CB and expanded ensemble techniques are discussed,
and numerous application examples to phase equilibria and critical point
determination are presented. The chapter by Mike Kotelyanskii and
Reinhard Hentschke presents and explains a method for performing
Gibbs ensemble simulations using MD.
In Part VII, Greg Rutledge discusses the modeling and simulation of
polymer crystals. He uses this as an excellent opportunity to introduce
principles and techniques of solid-state physics useful in the study of
polymers. The mathematical description of polymer helices and the calculation of X-ray diffraction patterns from crystals are explained. Both
optimization (energy minimization, lattice dynamics) and sampling (MC,
MD) methods for the simulation of polymer crystals are then discussed.
Applications are presented from the calculation of thermal expansion,
elastic coefficients, and even activation energies and rate constants for defect
migration by TST methods.
Part VIII focuses on the simulation of bulk amorphous polymers and
their properties. The chapter by Jae Shick Yang et al. applies a combined
atomistic (energy minimization)-continuum (finite element) simulation
approach to study plastic deformation in glassy bisphenol-A-polycarbonate.
This is an excellent example of multiscale modeling for addressing
mechanical behavior at long time- and length-scales, also discussed in the
next section. It also serves as an introduction to the simulation of glassy
polymers and of materials under strain. The chapter by Wolfgang Paul et al.
addresses polymer dynamics in the melt state: how it can be tracked with
atomistic and coarse-grained models using MD, and what insights it can

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Preface

vii

provide about glass transition phenomena. An overview is presented of the

calculation of static scattering patterns, as well as of the observables of
nuclear magnetic resonance (NMR) relaxation and of neutron-scattering
measurements from MD simulation trajectories. Segmental dynamics
and terminal dynamics, as revealed by MD simulations of oligomeric
polyethylenes, are compared against Mode Coupling Theory and Rouse
model predictions. In the chapter by Mike Greenfield, TST-based methods for the prediction of sorption and diffusion of small molecules in
amorphous polymers are thoroughly discussed. The general approach
followed to obtain the diffusivity, based on atomistic TST-based determination of rate constants for individual jumps executed by the penetrant in the
polymer matrix and subsequent use of these rate constants within a kinetic
MC simulation to track displacement at long times, is another good example of hierarchical modeling. Three TST-based methods (frozen polymer,
average fluctuating polymer, and explicit polymer) for the calculation of
rate constants are presented, with examples. The intricacies of performing
TST analyses in generalized coordinates using the flexible, rigid, or infinitely
stiff polymer models are also explained.
Part IX is devoted to bridging length- and time-scales through multiscale
modeling, whose importance has already been stressed above and brought
forth in some of the earlier sections. The chapter by Ulrich Suter and his
colleagues discusses the key issue of coarse-graining, whereby detailed
atomistic representations can be mapped onto computationally more
manageable models with fewer degrees of freedom, without loss of
significant information. Examples of coarse-graining detailed polyethylene
models into bond fluctuation (discussed in Part I) and bead-and-spring
models are presented. Simultaneous atomistic/continuum calculations
conducted on different scales are also explained, with emphasis on combined
finite element/molecular simulation schemes for tracking inelastic deformation in polymer solids (introduced in Part VIII). The chapter by Manuel
Laso and Hans Christian Ottinger is devoted to CONNFFESSIT, a
nontraditional multiscale method for simulating polymer flows that
combines finite elements with stochastic simulation of coarse-grained
molecular models. After a dense review of the general field of computational
rheology, algorithms and codes for tracking the particles in the finite
element mesh, integrating the particle stochastic equations of motion, and
reconstructing meshes are treated thoroughly. The chapter by Wouter den
Otter and Julian Clarke introduces dissipative particle dynamics (DPD), a
mesoscopic method for tracking the temporal evolution of complex fluid
systems that fully accounts for hydrodynamic interactions. After a basic
description of DPD and the problems to which it has been applied, the
question is taken up of mapping real systems, or atomistic models thereof,

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

viii

Preface

onto the coarse-grained models employed by DPD. Applications of DPD

for the simulation of polymer solution dynamics and microphase separation
in block copolymers are presented. Last but not least, the chapter by Andre
Zvelindovsky et al. gives an account of their dynamic density functional
theory (DFT), a mesoscopic functional Langevin approach for tracking
morphology development in complex soft-matter systems. The theoretical
underpinnings of the approach are explained, and applications are presented
from pattern-formation phenomena in complex copolymer systems and
solutions of amphiphilic molecules at rest, under shear, in the presence of
chemical reactions, or confined by solid surfaces.
We are grateful to all the contributors for their dedication and
painstaking work and for sharing our vision of a book on simulation
methods for polymers. Our families are thanked for their patience and
understanding of all the time we dedicated to the book, rather than sharing
it with them. We hope that the book may prove to be of considerable value
to students and practitioners of polymer simulation in academia and
industry.
Michael Kotelyanskii
Doros N. Theodorou

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Contents

Preface
Contributors
1.

Background
Michael Kotelyanskii and Doros N. Theodorou

I.

Calculating Single-Chain Properties

2.

Rotational Isomeric State (RIS) Calculations, with an

Illustrative Application to Head-to-Head, Tail-to-Tail
Polypropylene
E. Demet Akten, Wayne L. Mattice, and Ulrich W. Suter

3.

Single Chain in Solution

Reinhard Hentschke

II.

Lattice-Chain Monte Carlo Simulations

4.

Polymer Models on the Lattice

K. Binder, J. Baschnagel, and M. Muller

5.

Simulations on the Completely Occupied Lattice

Tadeusz Pakula

III.
6.

Molecular Dynamics
Molecular Dynamics Simulations of Polymers
Vagelis A. Harmandaris and Vlasis G. Mavrantzas

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Contents

IV.

Off-Lattice Monte Carlo Methods

7.

Configurational Bias Techniques for Simulation of

Complex Fluids
T. S. Jain and J. J. de Pablo

V.

Charged Polymer Systems

8.

Molecular Simulations of Charged Polymers

Andrey V. Dobrynin

VI.
9.

Calculating Chemical Potential and Free-Energy,

Phase Equilibria
Gibbs Ensemble and Histogram Reweighting Grand
Canonical Monte Carlo Methods
Athanassios Z. Panagiotopoulos

10. Gibbs Ensemble Molecular Dynamics

Michael Kotelyanskii and Reinhard Hentschke
VII.

Polymer Crystals

11. Modeling of Polymer Crystals

Gregory C. Rutledge
VIII.

Atomistic Simulations of Amorphous Polymers

12. Plastic Deformation of Bisphenol-A-Polycarbonate:

Applying an Atomistic-Continuum Model
Jae Shick Yang, Won Ho Jo, Serge Santos, and
Ulrich W. Suter
13. Polymer Melt Dynamics
Wolfgang Paul, Mark D. Ediger, Grant D. Smith,
and Do Y. Yoon
14. Sorption and Diffusion of Small Molecules Using
Transition-State Theory
Michael L. Greenfield
IX.

Bridging Length- and Time-Scales

15. Coarse-Graining Techniques

K. Binder, Wolfgang Paul, Serge Santos,
and Ulrich W. Suter

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Contents

16.

CONNFFESSIT: Simulating Polymer Flow

Manuel Laso and Hans Christian Ottinger

17.

Simulation of Polymers by Dissipative Particle Dynamics

W. K. den Otter and J. H. R. Clarke

18.

Dynamic Mean-Field DFT Approach for Morphology

Development
A. V. Zvelindovsky, G. J. A. Sevink, and
J. G. E. M. Fraaije

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

xi

Contributors

E. Demet Akten Department of Polymer Science, The University of

Akron, Akron, Ohio, U.S.A.
J. Baschnagel

Institut Charles Sadron, Strasbourg, France

K. Binder Institut fur Physik, Johannes-Gutenberg-Universitat Mainz,

Mainz, Germany
J. H. R. Clarke Department of Chemistry, University of Manchester
Institute of Science and Technology, Manchester, United Kingdom
W. K. den Otter Department of Applied Physics, University of Twente,
Enschede, The Netherlands
J. J. de Pablo Department of Chemical Engineering, University of
WisconsinMadison, Madison, Wisconsin, U.S.A.
Andrey V. Dobrynin Institute of Materials Science and Department of
Physics, University of Connecticut, Storrs, Connecticut, U.S.A.
Mark D. Ediger Department of Chemistry, University of Wisconsin
Madison, Madison, Wisconsin, U.S.A.
J. G. E. M. Fraaije
The Netherlands

Gorlaeus Laboratoria, Leiden University, Leiden,

xiii

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

xiv

Contributors

Michael L. Greenfield Department of Chemical Engineering, University

of Rhode Island, Kingston, Rhode Island, U.S.A.
Vagelis A. Harmandaris Institute of Chemical Engineering and HighTemperature Chemical Processes, and University of Patras, Patras, Greece
Reinhard Hentschke Physics Department and Institute for Materials
Science, Bergische Universitat, Wuppertal, Germany
T. S. Jain Department of Chemical Engineering, University of Wisconsin
Madison, Madison, Wisconsin, U.S.A.
Won Ho Jo Department of Fiber and Polymer Science, Seoul National
University, Seoul, Korea
Michael Kotelyanskii
U.S.A.

Rudolph Technologies, Inc., Flanders, New Jersey,

Manuel Laso Department of Chemical

University of Madrid, Madrid, Spain

Engineering,

Polytechnic

Wayne L. Mattice Department of Polymer Science, The University of

Akron, Akron, Ohio, U.S.A.
Vlasis G. Mavrantzas Institute of Chemical Engineering and HighTemperature Chemical Processes, and University of Patras, Patras, Greece
M. Muller Institut fur Physik, Johannes-Gutenberg-Universitat Mainz,
Mainz, Germany
ttinger
Hans Christian O
Zurich, Switzerland

Eidgenossische Technische Hochschule (ETH),

Tadeusz Pakula Max-Planck-Institute for Polymer Research, Mainz,

Germany, and Technical University of Lodz, Lodz, Poland
Athanassios Z. Panagiotopoulos Department of Chemical Engineering, Princeton University, Princeton, New Jersey, U.S.A.
Wolfgang Paul Institut fur Physik, Johannes-Gutenberg-Universitat
Mainz, Mainz, Germany

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Contributors

xv

Gregory C. Rutledge Department of Chemical Engineering,

Massachusetts Institute of Technology, Cambridge, Massachusetts, U.S.A.
Serge Santos Department of Materials, Eidgenossische Technische
Hochschule (ETH), Zurich, Switzerland
G. J. A. Sevink
Netherlands

Gorlaeus Laboratoria, Leiden University, Leiden, The

Grant D. Smith

University of Utah, Salt Lake City, Utah, U.S.A.

Ulrich W. Suter Department of Materials, Eidgenossische Technische

Hochschule (ETH), Zurich, Switzerland
Doros N. Theodorou Department of Materials Science and Engineering, School of Chemical Engineering, National Technical University
of Athens, Athens, Greece
Jae Shick Yang Department of Fiber and Polymer Science, Seoul
National University, Seoul, Korea
Do Y. Yoon School of Chemistry, College of Natural Science, Seoul
National University, Seoul, Korea
A. V. Zvelindovsky
The Netherlands

Gorlaeus Laboratoria, Leiden University, Leiden,

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

1
Background
MICHAEL KOTELYANSKII
Jersey, U.S.A.
DOROS N. THEODOROU
Athens, Greece

I.

Rudolph Technologies, Inc., Flanders, New

National Technical University of Athens,

BASIC CONCEPTS OF POLYMER PHYSICS

Polymers consist of very large molecules containing thousands or millions of

atoms, with molecular weights of hundreds of thousands g/mol or more.
They can be synthesized through polymerization or copolymerization
reactions from a wide variety of monomers. These large molecules can have
the simple topology of linear chains, most common for synthetic polymers,
or they can be rings, helices, combs, stars, or large networks. A macroscopic
piece of rubber, such as an automobile tire, is a network made of crosslinked polymer molecules; it can be regarded as a single molecule built from
small monomer units, each unit containing tens to hundreds of atoms.
The large variety of chemical constitutions and molecular architectures
of polymeric materials is responsible for the wide range of properties they
exhibit. A great number of contemporary technological applications rely on
the peculiar mechanical properties of polymers. Most remarkable among
those is rubber elasticity, i.e., the ability of a material to deform to many
times its original size without breaking and to return to its original shape
when the stress is released. Also important are the toughness and ductility
exhibited by semicrystalline polymers, glassy polymers and blends, and the
strength of oriented semicrystalline and liquid crystalline polymers; these
properties, combined with light weight and processability, have formed
a basis for a great number of plastic, fiber, and structural material
applications. The rheological properties of polymers are key to the design of
processing operations such as extrusion, blow molding, and film blowing,
whereby they are shaped into the multitude of products we use in everyday
life and in advanced technologies. Permeability properties of polymers are

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Kotelyanskii and Theodorou

important in the design of membranes for blood dialysis, water desalination

by reverse osmosis, and industrial gas separations, as well as in packaging
materials with barrier properties towards atmospheric gases. The surface
and interfacial properties of polymers and copolymers are critical to their
widespread use as adhesives, stabilizers of emulsions and suspensions,
compatibilizers of blends, coatings with controlled wettability, and
biocompatible materials. The optical and electronic properties of polymers
are important to common products, such as transparent packaging film and
PlexiglasTM windows, as well as to emerging applications such as polymeric
light-emitting diodes and optical switches.
In most cases, polymer materials with unique and valuable properties are
discovered by chance, or as a result of many years of painstaking trial-anderror experimentation. The ability to design materials tailored for particular
applications is the major challenge of polymer materials science. In recent
years, computer simulations have proved to be very helpful in meeting this
challenge.
Analytical theory can only solve very simplified models of polymers,
leaving out many details of the polymer molecular structure, and even these
simplified models can be solved only using certain approximations.
Computer simulations allow study of a simplified model directly and thus
identification of whether discrepancies between theory and experiment are
due to the simplifications of the model, to the approximations used in
solving the theory, or to both. The simplified models help identify properties
that are more general and can be observed for polymers with sometimes
very different chemical structure, but having similar chain topology, or
flexibility, or charge distribution.
More realistic models that account for the details of chemical structure
of particular polymers help identify relationships between the chemical
personality of each polymer and the values of particular properties. The
use of realistic models has been popular in biological applications.
Polymer simulations started with lattice models, as exemplified by the
pioneering work of Wall and collaborators, first on single chains in the
1950s [1] and then on multichain systems [2], and also by the work of
Alexandrowicz and Accad [3]. Off-lattice Monte Carlo simulations of single
chains were presented by Lal in the late 1960s [4]. Continuum simulations of
multichain systems appeared in the 1970s, as exemplified by the work
of De Vos and Bellemans [5], by the pioneering molecular dynamics study of
liquid butane by Ryckaert and Bellemans [6], and by the work of Balabaev,
Grivtsov, and Shnol [7]. In 1980 Bishop et al. [8] presented continuum
simulations of a bulk polymer with chains consisting of freely-jointed
Lennard-Jones segments. An early molecular dynamics simulation of
polyethylene using a realistic model for both bonded and nonbonded

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

interactions was presented in 1979 by Weber and Helfand [9], while

Vacatello et al. used a similar model to explore liquid triacontane in 1980
[10]. In parallel, important theoretical investigations pertaining to the
configuration-space distribution of polymer models were conducted by
Fixman [11] and by Go and Scheraga [12,13].

A. Interactions in Polymer Systems

In molecular systems, the potential energy function V(r) includes bonded
interactions between atoms connected by chemical bonds, and nonbonded
interactions between atoms of different molecules, or between atoms of the
same molecule which are not chemically bonded, for example between the
atoms of nonadjacent monomers of the same polymer chain.
Bonded interactions depend on the deviations of the chemical bond
lengths and bond angles from their equilibrium values, as well as on the
values of the dihedral (torsion) angles. The simplest and most often used
approach is to represent the bonded energy as a sum of three separate
contributions, described below:
VB

X
b

Vb

X

V

V

The chemical bonds are very stiff, deviations from the equilibrium bond
length usually being much less than a tenth of an A at room temperature.
Thus, the energy due to bond stretching is described by a harmonic
potential, proportional to the square of the deviation of the bond length
from its equilibrium value:
1
V b kb l
l0 2
2

The bond angles are less stiff than the bond lengths; nevertheless, at usual
temperatures they normally do not deviate by more than a few degrees from
their equilibrium values. The bond angle potential is also fairly well
approximated by a harmonic potential
1
V
k



0 2
2

Torsional potentials describe the energy change due to rotation around a

bond. This energy originates from interactions between the atoms connected

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Kotelyanskii and Theodorou

FIG. 1 Bond angles (
1,
2), and torsion angle () in the butane molecule. For
simplicity, only the carbon atoms are shown.

to the atoms linked by the bond, and possibly spinspin interactions

between the binding electrons from the adjacent bonds. The torsion angles
define the conformation of the polymer molecule. The torsional potential is
a periodic function* of the torsion angle . Its minima correspond to the
conformations where the molecule likes to spend most of its time. For
butane, shown in Fig. 1, the minima correspond to the trans-( p) and the
two gauche-( p/3) conformations.y
Typically, the torsional potential is represented as a Fourier series
V

An cosn

Nonbonded interactions between skeletal or pendant atoms connected to

the two ends of the bond, whose relative position changes when  is
changed, may be considered as separate nonbonded components of the
energy or may be incorporated as part of V .
The constants kb, l0,
0, An differ for different kinds of bonds. They are
usually adjusted to fit experimental data or results of quantum chemical

*Full 360 degrees rotation around the bond does not change the energy. If some of the pendant
atoms happen to be the same, the period can be less than 360 degrees.
y
The reader should be careful about the conventions used in measuring torsion angles. The old
convention places
0 at trans, the newer convention places
0 in the energetically most
unfavorable conformation where all four carbon atoms are on the same plane, atoms 1 and 4
being on the same side of the middle bond between 2 and 3 [14]. In both conventions,
is taken
as positive when it corresponds to rotation in a sense that would lead to unscrewing if the
middle bond were a right-handed screw.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

calculations, or both. More sophisticated representations of the bonded

interactions, designed to better fit experimental data and ab initio
calculation results, may include cross-terms that simultaneously depend
on bond lengths, bond angles and torsions [15].
Nonbonded interactions are typically modeled as electrostatic interactions between partial charges on the atoms, London dispersion forces due to
correlated fluctuations of the electronic clouds of the atoms, and exclusion
forces at short distances. They depend on the distance between the atoms
rij |ri
rj|, and are represented as a sum of Coulomb and Lennard-Jones
potentials.
X
X
V NB
V LJ rij
V q rij
5
ij

ij

12
6 !


V LJ rij 4 


rij
rij

1 qi qj
4p0 rij

V q rij

Here  and  are parameters dependent on the type of atoms.

p is the well
depth of the Lennard-Jones potential achieved at rij r0 6 2. The r
6
attractive term in the Lennard-Jones (LJ) potential finds theoretical justification in the 1930 quantum mechanical perturbation theory calculation of
London. Recent modeling work has shown that an exponential term,
A exp(
kr), for excluded volume interactions provides a more satisfactory
representation of thermodynamic properties than the r
12 term of the LJ
potential. Such an exponential term for the repulsions also has better
theoretical justification.
0 is the dielectric permittivity of free space, equal to 8.854 
10
12 C/(m V) in SI units and to 1/(4p) in cgs units. Potential parameters
are not just different for different elements, they also change depending on
what molecule or group the atoms belong to. For example, the values
typically used for the Lennard-Jones  for a carbon in an aromatic ring are
higher by almost a factor of two than the corresponding values for an
aliphatic carbon in the polyvinyl chain backbone. Partial charges on atoms
are determined by the electron density distribution within the molecule.
They are usually obtained from quantum ab initio or semiempirical
calculations and sometimes adjusted to reproduce experimentally measured
multipole moments of low-molecular weight analogs.
The set of parameters describing the interatomic interactions is often
referred to as a force field. Most popular force fields [1517] are adjusted
to fit the experimental and quantum calculation results over a wide range of

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Kotelyanskii and Theodorou

organic compounds. Parameters for the same atoms may be quite different
between different force fields.
The way the potential energy function is split into various kinds of
interactions is quite arbitrary and may vary from one force field to another.
Two different force fields may give very similar energy for the same
molecule in the same conformation, but the individual contributions:
torsional, nonbonded, bond angles, etc. may be quite different in different
force fields. This is why it is generally dangerous to mix force fields, i.e.,
take, say, the Lennard-Jones parameters from one force field and bonded
parameters from another.
In polymer simulations energy is usually expressed in kcal/mol or kJ/mol,
distances in A and angles in radians or degrees, so the dimensions of
the constants are the followingkb: kcal/(mol A2), k
: kcal/(mol rad2),
An: kcal/mol, : kcal/mol, rij, , l, and l0: A, charges qi: multiples of the
elementary charge e 1.6022  10
19 C.

B. Simplified Polymer Chain Models

From typical values of bonded and nonbonded interaction parameters one
can conclude that the fastest motions in a polymer system are the chemical
bond stretching vibrations, their frequencies typically being on the order of
1014 Hz; bond angle bending vibrations and torsional librations are about 10
times slower. Conformational (e.g., gauche $ trans) isomerizations over the
free energy barriers separating torsional states occur with rates on the order
of 1011 Hz or slower at room temperature. The characteristic time for the
end-to-end distance of a sequence of monomeric units to lose memory of its
orientation through these elementary motions grows rapidly with the length
of the sequence. The characteristic time for a chain in a melt of moderate
molecular weight (say, C10,000 linear polyethylene) to diffuse by a length
commensurate to its size and thus forget its previous conformation can well
exceed 1 ms.
Most interesting polymer properties are observed at frequency scales of
106 Hz and lower* This means that, in order to compare atomistic
simulation results to experiment, one should be able to reproduce the
model behavior at time scales spanning more than 10 orders of magnitude!
This task would be daunting even for a computer 10,000 times faster than
the most powerful supercomputers available today.

*These are the typical working frequency ranges for routine dielectric relaxation and rheological
experiments. More sophisticated techniques, such as NMR relaxation and inelastic neutron
scattering, are sensitive to higher frequencies up to 1012 Hz.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

Very large model systems, which are often necessary to track the
morphological characteristics responsible for the peculiar properties of
polymers, also present a great challenge in polymer simulations. Detailed
atomistic multichain polymer models used today seldom contain more than
a few thousands of atoms, although domain decomposition strategies on
parallel machines offer the possibility of going up to millions of atoms.
General computer simulation techniques useful for both low-molecular
weight molecules and polymers have been covered in some excellent
textbooks. We will briefly go over the fundamentals of these techniques in
this chapter. The major focus of the book, however, is on specific
approaches developed to handle the long time- and length scale challenges
of polymer problems. By necessity, these techniques must be hierarchical
ones, involving several levels of description, each designed to address a
specific window of time- and length scales. Going from one level to another
should entail a systematic coarse-graining procedure, wherein the detailed
information from more fundamental (shorter time- and length scale) levels
of modeling is built into some key parameters invoked by the more
macroscopic levels.
Analytical polymer theories usually address simplified models of polymer
chains, which capture universal features such as the chain topology,
flexibility, etc. Despite lacking many fine details, such models still manage to
predict, sometimes even quantitatively, many physical properties of polymer
networks, solutions, and melts. When a simplified general model turns out
to be capable of describing a particular polymer property or phenomenon,
this means that it successfully captures the relevant physics. Such results
provide valuable understanding of which particular features (e.g., chain
length, architecture, stiffness) are mainly responsible for a particular
property. Many polymer-specific effects and properties, such as rubber
elasticity, the viscoelastic rheological response of melts in the terminal
region, and overall molecular shapes in dilute solutions and melts in the
bulk and next to nonadsorbing surfaces, are similar for polymers of different
chemical structures. They can be rationalized with relatively simple arguments based on the decrease of entropy associated with chain extension and
on environment-dependent excluded volume interactions between segments.
Figure 2 shows the most popular simplified models of polymer chains
[18,19].
The freely-jointed chain model (Fig. 2a) is the simplest; it has been
described as the ideal gas of polymer physics, as all interactions between
the chain segments, except the chemical bonds connecting them, are
neglected. The model represents a chain as a sequence of steps, or rectilinear
statistical (Kuhn) segments, connected together at their ends. Each
statistical segment represents a sequence of several chemical bonds,

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Kotelyanskii and Theodorou

FIG. 2 The most common simplified models of polymer chains: freely-jointed chain
(a), bead-spring chain (b), rotational isomer model (c), and the typical torsional
potential V() for butane molecule (d).

depending on the conformational stiffness of the chain. The steps can point
in any direction, the directions of different steps being completely
independent, and the chain can intersect itself. Clearly, this model is
identical to the random flight model. It can easily be shown that the
probability to observe a certain value of the vector R connecting the two
ends of a sufficiently long freely-jointed chain obeys a Gaussian distribution
with zero average and mean squared value
hR2 i  hR2 i NK b2K

for a chain of NK statistical segments, each of length bK. This is one of the
most fundamental results of polymer science; its derivation and corollaries
can be found in polymer textbooks [1820]. The radius of gyration tensor,
characterizing the average shape of the chain, and its properties can also
be calculated. Once the distribution is known, all the statistical and
thermodynamic properties of the ideal chain can be obtained.
A real chain whose conformation is governed only by local (effective)
interactions along its contour can be mapped onto a freely-jointed chain by
requiring that Eq. (8) is satisfied, and also that the contour length of the
freely-jointed chain, NKbK, matches that of the real chain at full extension.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

The distribution of the distance between two points in the freely-jointed

chain is also well approximated by a Gaussian distribution, but with the
chain length NK in Eq. (8) replaced by the number of statistical segments
between the points. This makes a freely-jointed chain self-similar on various
length scales. If we look at it with different spatial resolutions it looks the
same, as long as the resolution is much larger than bK and less than NK1=2 bK .
The bead-spring model (Fig. 2b) represents a polymer chain as a
collection of beads connected by elastic springs. It, too, is a coarse-grained
model. The coarse-graining is based on the polymer chain self-similarity,
with a single bead corresponding to a chain fragment containing several
monomers. Springs reproduce the Gaussian distribution of separations
between monomers connected through a large number of chemical bonds.
The spring constant is given by 3kB T=hR2sp i, where hR2sp i is the mean square
end-to-end distance of the actual chain strand represented by the spring. The
spring reproduces the entropic free energy rise associated with the reduction
of conformations of a strand as its two ends are pulled apart.
The rotational isomer model (Fig. 2c) makes a step from the freelyjointed chain towards the geometry of the real polymer. It invokes fixed,
realistic values for the chemical bond lengths and the chemical bond angles

, and a potential energy function V2 ,  , . . .. Figure 2d schematically
shows the torsional potential for a butane molecule, containing only three
backbone CC bonds, where it is a function of only one angle. The
rotational isomer model addresses unperturbed chains, i.e., chains whose
conformation remains unaffected by nonlocal interactions between topologically distant atoms along the backbone. Under unperturbed conditions,
V(2, 3, . . .) can be written as a sum of terms, each depending on a small
number (usually two) of adjacent torsion angles. A rotational isomer chain
of sufficient length also follows the scaling relation of Eq. (8), which can be
written more specifically in terms of the number of chemical bonds N along
the backbone and the length of a bond as [20,18]:
2

hR2 i C1 N2

The characteristic ratio C1 characterizes chain flexibility. It depends on

the
and torsional potential and is determined by the chemical structure of
the monomers [20]. The rotational isomeric state (RIS) model, introduced
by P.J. Flory [20] is essentially an adaptation of the one-dimensional Ising
model of statistical physics to chain conformations. This model restricts
each torsion angle  to a discrete set of states (e.g., trans, gauche,
gauche
), usually defined around the minima of the torsional potential
of a single bond, V() (see Fig. 2d). This discretization, coupled with the
locality of interactions, permits calculations of the conformational partition

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

10

Kotelyanskii and Theodorou

function, the C1, and conformational averages of many properties by

matrix algebra, once a set of statistical weights is available for all pairs of
torsional states assumed by successive bonds. The latter weights can be
determined by atomistic conformational analysis of oligomers. Extensions
to chains containing several types of chemical bonds and to nonlinear
chemical architectures have been made. The RIS model is described in detail
in chapter 2 by E.D. Akten, W.L. Mattice, and U.W. Suter, and in the
book [14].
Unperturbed conditions are realized experimentally in bulk amorphous
polymers and in dilute solutions in specific (rather bad) solvents at specific
temperatures, defining the so-called
point. For a chain dissolved in a good
solvent, nonlocal interactions cannot be neglected; polymer segments strive
to maximize their favorable contacts with solvent molecules, leading to a
repulsive intersegment potential of mean force. These repulsions prohibit
chain self-intersections. The simplest mathematical model describing this
effect is the self-avoiding random walk (SAW). The SAW problem can
only be solved numerically using computer simulations. The end-to-end
distance of a SAW follows the scaling hR2i / N with  (the excluded volume
exponent) close to 0.6, i.e., significantly different from the value 0.5
characteristic of unperturbed chains. This is easy to understand, as the
repulsion between monomers leads to an increase in coil size.
All major simulation techniques described in the rest of this book are
applicable to these simplified models, as well as to the detailed atomistic
models. Simulations of the simplified models, however, require much less
computer power. They are helpful in studying phenomena where universal
chain characteristics, such as chain length, flexibility, and topology are of
interest.

C. Unperturbed Polymer Chain

As already mentioned, a popular descriptor of the overall size of a polymer
chain is the mean squared end-to-end distance hR2i. For a chain of N 1
identical monomers, linked by N bonds of length , the end-to-end vector is
a sum of bond vectors i ri
ri
1 , and its mean square is:
*
hR2 i

N
X

!2 +
2

i1

N

1
X

N
X

hi  j i

i1 ji1

N
X
hi  i i
i1

N

1 X
N
i
X

hi  ik i N2

i1 k1

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

10

Background

11

For very long chains, ignoring end effects, we can write

hR2 i N2 22

N
1
X

N
kh1  k1 i=2

k1

N2 2N2

X
1
k=NCk

11

Here C(k) is a bond direction correlation function. As the number of monomers separating two bonds increases, their directions become uncorrelated,
and hence C(k) is a decreasing function of k. For a random walk, C(k) 0
for all k  1. For an unperturbed P
chain, C(k) falls exponentially with
increasing k at large k and C 1 2 kC(k).
The mean squared end-to-end distance of a linear unperturbed chain of N
monomers is proportional to N. As R is a sum of a large number of
independent random vectors, the probability density for the end-to-end
separation vector to have a certain value R is given by a Gaussian
distribution:

3
N R
2phR2 i

3=2

3 R2
exp

2 hR2 i


12

The quantity
kBT ln N(R) is a conformational free energy associated

with the fact that a given end-to-end vector R can be realized through a large
number of conformations.
When the chain is deformed, the distribution of end-to-end distance
vectors will change towards less probable values, thus reducing the polymer
chain entropy and causing an increase in free energy.
A rubber is a network built of cross-linked chains; the above reasoning
explains, at least qualitatively, the mechanism of rubber elasticity.
The conformation of a chain of N 1 segments, numbered from 0 to N,
can also be characterized by the mean square radius of gyration:
hR2g i

*
+
N

1 X
N
X
1
2

ri
rj
N 12 i0 ji1

13

or by the hydrodynamic radius

"
*
+#
1
N

1 X
N
X
1
1
1
Rh
2 N 12 i0 ji1 jri
rj j

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

14

12

Kotelyanskii and Theodorou

measured by elastic and inelastic light scattering techniques, respectively.

The former is related to the coil size and shape, while the latter describes a
coils hydrodynamic properties, as inelastic light scattering measures the
self-diffusion coefficient of polymer coils in dilute solution.
It can be shown, that, for large chain lengths, both hydrodynamic and
gyration radii of the polymer coil can be expressed in terms of the average
end-to-end distance [19]:
1
hR2g i hR2 i
6



3p 1=2 2 1=2
hR i
Rh
128

15

D. Mixing Thermodynamics in PolymerSolvent

and PolymerPolymer Systems
The fact that a polymer molecule consists of a large number of monomers
linked together in a chain is responsible for various properties, which are
specific to polymers in solution and in the melt. Polymer solutions have very
different properties, depending on concentration. As the concentration
increases, a polymer solution changes from the dilute to the semidilute, to
the concentrated regimes.
When the concentration is very small, polymer coils do not contact each
other. The effect of the polymer on the solution properties is additive,
consisting of single-coil contributions.
In the semidilute solution the monomer volume fraction is still very low,
but the polymer coils begin to overlap and entangle with each other. This is
possible because the monomer concentration in the Gaussian coil is
relatively low. Indeed, N monomers in a good solvent occupy a volume of
the order of magnitude of R3 / 3 N 9=5 , so that the average monomer
density in a coil is of the order of
3 N
4=5 . The coils will begin to touch
when the polymer volume fraction in the solution becomes * / N
4/5. This
concentration, which can be very low for large N, marks the boundary
between dilute and semidilute regimes.
Concentrated solutions with polymer fraction on the order of 1 behave
pretty much like polymer melts.
Polymers are in general much less soluble and miscible with each other,
compared to the corresponding monomers under the same conditions.
Miscibility or solubility is defined by the balance of the entropy gain and
energy loss or gain upon mixing. Consider mixing substance A with

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

13

substance B. If the interactions between molecules of different components

(AB interactions) are stronger than between molecules of the same kind (AA
and BB interactions), mixing is favorable at all temperatures. Such a system
is always miscible. Usually this case is realized when there are some specific
interactions (e.g., acidbase, hydrogen-bonding) between A and B. More
common is the case where AB interactions are energetically less favorable
than AA and BB interactions. This is typically the case for nonpolar
substances held together by London dispersion forces. In this case, mixing
occurs above a certain temperature, when the entropic contribution
TS
to the free energy of mixing overcomes the unfavorable positive energy of
mixing. The entropy gain upon mixing is due mainly to translational
entropy of the molecules. In the mixture, each A and B molecule can occupy
the whole volume occupied by the mixture, while in the phase-separated case
each molecule is localized in the volume occupied by the particular phase.
In polymer chains large numbers of monomers are linked together,
therefore, the translational entropy gain upon mixing per unit mass or
volume is much smaller than in the case of mixing unlinked monomers. The
entropy gained by a monomer in a liquid of unlinked monomers is now
gained by each polymer molecule, and thus the entropy gained per mole of
monomers is many times less in the polymer case. One of the most
fundamental results of polymer thermodynamics, first derived (independently) by Flory and Huggins [20,21] through a lattice-based mean field
theory, states that the free energy change per mole of segments upon mixing
polymer chains made up of A-type segments with chains made up of B-type
segments of lengths NA and NB, respectively is:
Gmix
1
1

A lnA
B lnB A B
NA
NB
RT

16

In the theory, A and B segments are envisioned as equal in volume, occupying sites on a lattice of coordination number Z. A is the volume fraction of
A-type segments in the system, B 1
A is the volume fraction of B-type
segments.  (Z/kBT)("AB
1/2"AA
1/2"BB) is the Flory interaction
parameter, describing the interaction energy difference between unlikeand like- segment interactions occupying adjacent sites on the lattice. The
parameter  is positive for most polymer mixtures. From Eq. (16) it is seen
that the entropy gain per mole of segments upon mixing polymers is smaller
than the corresponding gain for mixing monomers (NA NB 1) by a factor
equal to the inverse chain length. This explains the difficulty in mixing
different polymers. For NA NB N, Eq. (16) predicts an upper critical
solution temperature which scales proportionally with the chain length,
N assuming a value of 2 at the critical point. Traditionally, a strategy

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

14

Kotelyanskii and Theodorou

for achieving miscible blends has been to use systems capable of developing
specific interactions, which can make  negative.
As we discussed above, polymer chains are not rigid bodies, and there is
an entropy associated with the different polymer chain conformations.
Polymer chains usually contain very large numbers of monomers (thousands
or more), and, provided that environmental conditions ensure the same
excluded volume exponent , all polymers with the same chain topology are
similar at the length scales comparable to the chain radius of gyration.
Florys model is the simplest; it assumes that chain conformation does not
change upon mixing, and therefore there is no conformational entropy
contribution in Eq. (16). This contribution is, however, very important in
many cases and has to be taken into account to explain the thermodynamics
and phase behavior of mixtures and solutions where chain conformation is
different in different phases, for example in systems containing block
copolymers, chains next to surfaces or in restricted geometries, in the
swelling of polymer gels etc. [19,22,23].

E. Polymer Chain Dynamics

The motion of a large polymer molecule is quite complex. Even though
individual atoms move about with the same equilibrium distribution of
speeds as if they were disconnected, their motion is constrained by the
chemical bonds keeping the chain together. Longer and longer parts of the
chain need longer and longer times to rearrange, and quantities depending
on overall conformation, such as the end-to-end distance or the radius of
gyration, take a very long time to forget their original values.
A general characteristic of polymer liquids is their viscoelastic response
to flow. Because of the long relaxation times associated with large-scale
conformational rearrangements, chains subjected to a fast flow field are
oriented, or even unraveled by the flow. Part of the energy imparted by the
flow is stored as elastic free energy and is released upon cessation of the
flow, when chains spring back to their unperturbed conformations causing
macroscopic recoil phenomena. Polymeric fluids have memory, and this is
of paramount importance to their processing and applications.
The dynamics of a coil in a low-molecular weight melt can be described
well by an ingeniously simple model, the Rouse model [18]. This model
represents a polymer chain as a set of beads connected by harmonic springs
(compare Fig. 2b). The beads move as (tethered) Brownian particles subject
to random forces and to frictional forces proportional to their velocity
exerted from their environment (see Section I.F). For a linear polymer chain
the Rouse model predicts that the longest relaxation time (time needed for

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

15

the chain center of mass to diffuse by a length commensurate to hR2i1/2,

approximately equal to the time needed for the end-to-end vector R to lose
memory of its original orientation) is proportional to the chain length
squared N2. The viscosity of a melt in the molecular weight range described
by the Rouse model is also proportional to N and the selfdiffusivity of
a chain is proportional to N
1.
The Rouse model has been extended to deal with the dynamics of chains
in dilute solution [18]. In solution a moving bead perturbs the solvent
flow around another bead, leading to effective, so-called hydrodynamic,
interactions. The Zimm model generalizes the Rouse model by taking
hydrodynamic interactions into account.
As the molecular weight of a polymer melt increases, it is envisioned that
entanglement constraints arise between different polymer coils, making
relaxation even slower. Experimentally, it is found that the viscosity of the
melt rises with increasing chain length as N3.4. The reptation theory of
polymer dynamics [18] postulates that long chains move in a Rouse-like
fashion along their contour (primitive path), while motion normal to the
primitive path is constrained by entanglements. The reptation model in its
original formulation predicts a relaxation time and a melt viscosity which
scale as N3, and a chain self-diffusivity which scales as N
2. Several
extensions and refinements of the reptation model have appeared.
The slow dynamics of polymers presents a formidable challenge for
computer simulations, as on the one hand the most interesting polymerspecific phenomena occur at time scales comparable with the relaxation
times of the whole chain, but on the other hand to reproduce the motion
correctly one has to solve the equations of motion with a time step smaller
than the period of the fastest monomer motions. Using coarse-grained
models in the simulation is one way in which one may try to overcome this
problem.

F. Glass Transition Versus Crystallization

Low-molecular weight liquids usually crystallize upon cooling. Very fast
cooling rates are usually required to form a glass (amorphous solid).
Contrary to that, because of their very slow relaxation, polymer melts are
much more viscous, and the glassy state is the most common for solid
polymers. Many polymers, such as atactic vinyl polymers, cannot crystallize
due to their irregular stereostructure. In polymers crystallized from the
melt, the crystalline phase is usually only around 1050%. The degree of
crystallinity and the morphology of semicrystalline polymers are highly
dependent on the conditions of temperature and processing flow under

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

16

Kotelyanskii and Theodorou

which they are formed. A high degree of crystallinity is usually obtained

by cooling the polymer from a highly oriented liquid crystal state, or by
stretching.
Semicrystalline polymers also exhibit very different behavior under
deformation. When stretched, crystals of low-molecular weight substances
usually break at an elongation of several percent. Semicrystalline polymers
can deform up to 100% without breaking; on the contrary, the elastic
modulus increases with increasing elongation (strain-hardening effect). This
is because, at high deformations, polymer chains in the amorphous regions,
which may grow at the expense of the crystalline regions, begin to stretch
and align along the direction of deformation, leading to a higher elastic
modulus.
Our fundamental understanding of glass formation, even in lowmolecular weight fluids, is still incomplete. The mode coupling theory
[24,25] describes the falling out of ergodicity at a critical temperature which
is substantially higher than the glass temperature Tg, where macroscopic
manifestations of solidification are seen in volume, enthalpy, and elastic
constant measurements.
From a practical viewpoint, one can say that glass formation occurs
because the characteristic time of segmental motions (motions associated with the -mode of relaxation measurements) is a very strongly
increasing function of temperature. As temperature is decreased, segmental
motions become slower and slower, until there comes a point where the rate
of segmental motion cannot keep up with the rate of change in temperature.
The system is no longer capable of exploring its entire configuration space
over ordinary time scales and is confined within a small region of that
space, containing at least one local minimum of the potential energy V.
Macroscopically, a change in slope is seen in the specific volume vs. temperature and specific enthalpy vs. temperature curves. The glass temperature, Tg, depends, to a certain extent, on cooling rate. It is typically reported
for cooling rates on the order of 1 K/min.
A glassy material is not in thermodynamic equilibrium. Its physical
properties change gradually with time (physical ageing). The time scales
of these changes, however, are enormous a few decades of degrees below Tg,
so that the material can be regarded as a dimensionally stable solid for all
practical purposes.
The temperature dependence of the characteristic time of molecular
motions responsible for the glass transition is strongly non-Arrhenius.
Above Tg it is described by the VogelFulcher law
T 0 exp

TA
T
TV

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

17

Background

17

with TA (activation temperature) and TV (Vogel temperature) being

constant.
The ratio of characteristic times at two temperatures T  To,
aT (T )/ (To) , (shift factor) consequently follows the WilliamsLandel
Ferry (WLF) equation
log aT
C1

T
To
T
To C2

18

When the reference temperature To is chosen as the calorimetric glass temperature, Tg, the constants fall in relatively narrow ranges for all polymers:
C1 14 to 18 and C2 30 to 70K [26]. At the glass temperature, is on the
order of 1 min.

II. STATISTICAL MECHANICS

A. Trajectories in Phase Space
In classical mechanics, the state of a molecular system is described by
positions of atomic nuclei and their time derivativesvelocities (or
momenta). Each state represents a point in the multidimensional space
spanned by positions and momenta, which is termed the phase space of
the system. The position vectors of the atoms, or the set of generalized
coordinates providing the same information as atomic positions, are called
degrees of freedom; they span the configuration space of the system. The
space spanned by the momenta (or generalized momenta) of the degrees
of freedom is called momentum space. The evolution of the systems
microscopic state with time can be represented as a set of state points,
corresponding to successive moments in time. This set defines a line in phase
space, which constitutes the systems dynamical trajectory. The following
Fig. 3 shows the simplest example of state points and trajectories in phase
space. Notice that not just any line can represent a trajectory. The line
crossed out in Fig. 3 cannot be a trajectory. It corresponds to the impossible
situation where the coordinate is increasing with time, while the velocity is
negative.

B. Classical and Quantum Mechanics

In both quantum and classical mechanics, a system is defined by its degrees
of freedom and by its potential and kinetic energy functions. In a quantum
mechanical description, electrons, in addition to nuclei, are included among
the degrees of freedom. Given the potential and kinetic energies as functions

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

18

Kotelyanskii and Theodorou

FIG. 3 State points at different moments in time t1<t2<t3 and trajectory in the
phase space {x, p} for two free particles traveling along the x axis with different
velocities x_ i vi and momenta pi mivi. Particle 2 goes to the left ( p2<0). The
crossed out trajectory is impossible.

of the degrees of freedom, we can write the equations of motion and, in

principle, we should be able to predict the systems motion under any given
initial conditions by solving these equations. Unfortunately, the situation is
not so simple for at least two reasons. First, in the majority of cases, it turns
out that the equations are too complex to be solved, even numerically.
Imagine that to describe the motion of 1 cm3 of water classically we need to
store positions and velocities for about 1022 molecules and solve the same
number of equations. Just the storage is beyond the capacity of the most
powerful computer available today.
A second difficulty is that equations of motion are highly nonlinear
and therefore lead to instability with respect to initial conditions. Small
uncertainties in the initial conditions may lead to large deviations in the
trajectories. Two trajectories starting at almost the same state point will
diverge exponentially with time, a small uncertainty in the initial coordinates
or velocities making the motion completely unpredictable. If we are trying
to describe a molecular system in a cubic container of side L, when the
uncertainty in the calculated particle coordinates becomes of the same order
of magnitude as the characteristic model size L, the prediction becomes
totally worthless.
The latter, however, can be turned into an advantage. The inherent
mechanical instability and complexity of the motion justifies a statistical
description of the system. In contrast to the deterministic approach, i.e.,

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

19

trying to follow the motion of every molecule as closely as possible, the

statistical approach concentrates on the average properties calculated over
the systems trajectory or over phase space.y As was demonstrated by the
founding fathers of statistical mechanics, L. Boltzmann and J.W. Gibbs in
the late 19th century, these trajectory or phase space averages are directly
related to the experimentally observed thermodynamic properties.
In general, molecular motion should be described using the laws of
quantum mechanics. In quantum mechanics dynamical trajectories themselves are probabilistically defined entities. The state of the system is
described by a probability amplitude function, , which depends on
coordinates and, possibly, spin states of all nuclei and electrons present in
the system. * is the probability density for observing the system in
a particular point in phase space. Motion of the system, or in other words
its change in state with time, is described by the time-dependence of the
-function. It is determined by solving the Schrodinger equation:
X
i

h2 2
^  @
rri  V  H
@t
2mi

19

Here, mi is the mass of particle i, V is the potential energy, a function of the

^ is the Hamilton operator.z
positions of electrons and nuclei, and H
Equation (19) is a partial differential equation. The number of its
variables equals the total number of electron and nuclear coordinates and
spin states, or in other words the total number of degrees of freedom.
Solving this equation is a very difficult, virtually impossible task, but
fortunately it can be simplified by the following two approximations.
First, the nuclei of almost all atoms, except maybe hydrogen and helium,
are heavy enough for the classical approximation to describe their motion
sufficiently well at normal temperatures.
Electrons, on the other hand, being several thousand times lighter than
nuclei, move much faster. The second approximation is to assume that the
electrons adjust practically instantaneously to the current positions of the
nuclei. Given a particular set of nuclear coordinates, the electron state is
adjusted to make the total potential energy of the system minimal with
respect to the electron state. This approximation is called adiabatic, or

y
The equivalence of these two averages is the essence of a very important system property called
ergodicity, which deserves a separate subsection further in this chapter.
zA hat is used here to emphasize that, in quantum mechanics, the Hamiltonian is an operator
acting on the function and not just a function of particle coordinates and momenta, as in
classical mechanics.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

20

Kotelyanskii and Theodorou

the BornOppenheimer approximation; in engineering language, it is a

quasi-steady state approximation for the electrons. The adiabatic
approximation may not be applicable when electron excitation processes
are important, for example in the case of chemical reactions or conformational changes induced by the absorption of light.
The adiabatic approximation is very helpful because, if the electron states
are adjusted to the positions of the nuclei, the systems state can be described
by the coordinates of the nuclei alone. The number of degrees of freedom is
then significantly reduced. Accordingly, in the Schrodinger equation, V is
replaced by the effective potential V(rn) V(rn, e0(rn)), where e0(rn)
describes the electron state providing the lowest potential energy for the
given set of nuclear coordinates rn, and is thus also a function of the nuclear
coordinates. The accuracy of U(rn) determines the accuracy of quantities
calculated from computer simulations.
An approach for tracking electronic degrees of freedom in parallel with a
numerical integration of the classical equations of motion for the nuclei, and
therefore determining V(rn) on the fly, has been devised by Car and
Parrinello [27]. This extended ensemble molecular dynamics method, termed
ab initio molecular dynamics, solves the electronic problem approximately
using the KohnSham formulation of Density Functional Theory. This
approach proved useful for covalent systems; it still has to be applied to
the systems where the properties of interest are defined by Lennard-Jones
interactions.

C. Classical Equations of Motion

When the interactions in the molecular system are known, the classical
description can be cast in one of three different forms: Hamilton, Newton,
or Lagrange equations of motion. Consider a molecular system with
potential energy Vq1 , q2 , . . . , qNf and kinetic energy Kq1 , q1 , . . . , qNf ,
q_ 1 , q_ 2 , . . . , q_ Nf or Kq1 , q2 , . . . , qNf , p1 , p2 , . . . , pNf , where qi are generalized
coordinates with q_ i being their time derivatives, and pi are generalized
momenta. The configuration space and momentum space are both
Nf -dimensional, where Nf is the total number of degrees of freedom. The
phase space is 2Nf -dimensional. When Cartesian coordinates of the atoms
are used as degrees of freedom, a triplet of qi stands for the position vector
of an atom, rk, and a triplet of pi stands for an atomic momentum vector
pk mk r_ k mk vk . In this case, the kinetic energy can simply be written as
K

X p2
X p2
X1
X1
i
k
mi q_ 2i
mk v2k

2
2
2m
2m
i
k
i
i
k
k

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

21

The three different forms of the equations describing the motion of

atoms are:
.

Hamilton equations



H q1 , q 2 , . . . , qN f , p1 , p2 , . . . , pN f K V


@qi @H q1 , q2 , . . . , qNf , p1 , p2 , . . . , pNf

@pi
@t


@H q1 , q2 , . . . , qNf , p1 , p2 , . . . , pNf
@pi


@t
@qi
20

Newton equations
dri
vi
dt
dvi
fi
mi
dt



@V r1 , r2 , . . . , rNf =3
fi

@ri

21

Lagrange equations



L q1 , q2 , . . . , qNf , q_ 1 , q_ 2 , . . . , q_ Nf K
V
dqi
dt 

d @L @L q1 , q2 , . . . , qN , q_ 1 , q_ 2 , . . . , q_N

dt @ q_ i
@qi
q_ i

22

In the above, Hq1 , q2 , . . . , qNf , p1 , p2 , . . . , pNf and L q1 , q2 , . . . , qNf , q_ 1 ,
q_ 2 , . . . , q_ Nf stand for the Hamiltonian and Lagrangian functions of the
system, respectively. As an illustration, let us consider these equations for
two examples. The first is a monatomic fluid, and the second is a harmonic
oscillator.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

22

Kotelyanskii and Theodorou

A monatomic fluid is governed by a potential energy function of the form

V(r1, r2 , . . . , rN). The degrees of freedom are the 3N Cartesian coordinates
{r1, r2 , . . . , rN}. The Newton equations are:
dri
vi
dt
dvi
fi
mi
dt
fi

@V

rri V
@ri

23

If we use {r1, r2 , . . . , rN} as the generalized coordinates, the Lagrangian is

a function Lr1 , r2 , . . . , rN , r_ 1 , r_ 2 , . . . , r_ N , and the Lagrange equations of
motion are:
r_ i
Lr1 , r2 , . . . , rN , r_ i , r_ 2 , . . . , r_ N K
V

dri
dt
N
X
1
i1

mi r_ 2i
Vr1 , r2 , . . . , rN

d @L
d 2 ri @L
mi 2

rri V
dt @ r_ i
@ri
dt

24

The Hamiltonian function and Hamilton equations are:



H r1 , r2 , . . . , rN , p1 , p2 , . . . , pN K V
H

N
X
p2i
Vr1 , r2 , . . . , rN
2mi
i1

@ri @H pi

@pi mi
@t
@pi
dvi @H
 mi

rri V
@ri
@t
dt

25

In the special case where V(r1, r2 , . . . , rN) 0 for all configurations, the
above equations describe an ideal monatomic gas. The ideal gas particles
travel with their velocities not changing with time. This is because they do
not interact with each other or with anything else. The ideal gas is a limiting
case of a system with no interactions; it is a simple system, whose statistical
mechanics can be solved exactly. It is used as a reference to build more
complex systems, where interactions are non-negligible.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

23

Another very important analytically solvable case is the harmonic

oscillator. This term is used for a mechanical system in which potential
energy depends quadratically on displacement from the equilibrium
position. The harmonic oscillator is very important, as it is an interacting
system (i.e., a system with nonzero potential energy), which admits an
analytical solution. A diatomic molecule, linked by a chemical bond with
potential energy described by Eq. (2), is a typical example that is reasonably
well described by the harmonic oscillator model. A chain with harmonic
potentials along its bonds (bead-spring model), often invoked in polymer
theories such as the Rouse theory of viscoelasticity, can be described as a set
of coupled harmonic oscillators.
The harmonic oscillator is particularly important, because any mechanical system in the vicinity of stable equilibrium can be approximated by a
harmonic oscillator. If the deviations from equilibrium are small, one can set
the origin at the equilibrium point and expand the potential energy in
powers of the displacement. To take a simple example, for a pair of identical
atoms interacting via the Lennard-Jones potential [Eq. (6)], if p
the distance
between the atoms r is close to the equilibrium value r0  6 2, we can
expand the potential in powers of u r
r0
12

6 !


4


r0 1 u=r0
r0 1 u=r0

2

 3 !!
u
u

1 36
O
r0
r0

V LJ

 V0
V 0
,

ku2
2
k

72
r20

26

If u x1
x2
r0, assuming that the y and z coordinates of both atoms

are zero (Fig. 4), Newtons equations of motion become:
mx 1

@V

kx1
x2
r0
@x1

mx 2

@V
kx1
x2
r0
@x2

my i 0,

i 1, 2

mzi 0,

i 1, 2

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

27

24

Kotelyanskii and Theodorou

FIG. 4 Harmonic approximation (dashed line) for the Lennard-Jones potential

(solid line).

Adding and subtracting the first and second equations gives the following
two equations:
mx 1 mx 2 0
mu
2ku

28

The first equation describes the motion of the dimer center of mass
xc (mx1 mx2)/2m. Its left hand side, mx 1 mx 2 , equals the total mass
2m times the center of mass acceleration. As long as there is no net external
force applied, the acceleration is zero. However, if the initial velocities of
the atoms v1(t 0) and v2(t 0) are such that the center of mass velocity
vc [mv1(t 0) mv2(t 0)]/2m is not zero, the value of vc does not change
with time, and the dimer travels as a whole with constant speed. The
second equation describes the motion of the atoms relative to each other,
as u measures the deviation of their separation from its equilibrium value
r0. The solution of the second equation,
ut A cos!t B sin!t
2k
!2
m

29

with A and B defined from the initial conditions:

A ut 0;

B! u_ t 0 v1 t 0
v2 t 0

30

describes oscillations around the equilibrium value u 0 with period 2p/!.

A typical period of the vibrations due to Lennard-Jones interactions is on
the order of several picoseconds.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

25

To derive the Lagrange and Hamilton equations of motion we need

expressions for the kinetic and potential energies:
m1 r_ 21 m2 r_ 22

2
2
k
V jr1
r2 j
r0 2
2

31

Here we used the harmonic approximation for potential energy (26) and set
the zero energy level at V 0
. Choosing the x axis along the line connecting the atoms, following the Lagrangian or Hamiltonian formalism, given
by the Eqs. (20) or (22), and remembering that, for simplicity, we consider
here the case m1 m2 m, one can arrive at the Newton equations of
motion (27). When Cartesian coordinates are used, Newton, Lagrange,
and Hamilton equations result in exactly the same differential equations
for the coordinates.
Newton equations, however, do require use of Cartesian coordinates,
while Lagrange and Hamilton equations do not. Let us see how the
Lagrange and Hamilton formulations can help us derive the more convenient set of Eqs. (28) from the very beginning. All we have to do is choose
the center of mass coordinate xc and the oscillation amplitude u as
generalized coordinates, describing the motion along the x direction in
which we are interested. We leave the Cartesian coordinates y1, z1 and y2, z2
for the other four degrees of freedom. The potential energy is simply
V ku2/2. The kinetic energy is a little more tricky:
K

m1 y_ 21 z_21 m2 y_22 z_22

2
2



2
m1 m2 vc 1 1
1
1 2

u_

2 m1 m2
2

32

Again, using the above equations in the Lagrange (20) or Hamilton (22)
formalism, working out through all the necessary derivatives, and using
m1 m2 m, one arrives directly at Eqs. (28) for u and xc. Figure 5 shows
the harmonic oscillators trajectories, corresponding to different initial
conditions.
The harmonic oscillator model is particularly useful in the study of
crystalline solids. Potential energy of the crystal can be expanded around the
state of mechanical equilibrium under given macroscopic dimensions
(quasiharmonic approximation). By diagonalizing the Hessian matrix
of second derivatives of the potential energy with respect to atomic

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

26

Kotelyanskii and Theodorou

FIG. 5 Trajectories of the harmonic oscillator in its two-dimensional phase space,

and positions at different moments of time t2>t1, but t2
t1<T.

displacements, vibrational motion can be analyzed as a set of 3N

independent harmonic oscillators characterized by a spectrum of frequencies. Lower frequencies correspond to longer-range, collective motions. The
Debye model is a good model for this spectrum in the region of low
frequencies [28].
Observing that Lagrange, Hamilton, and Newton equations of motion
lead to exactly the same differential equations for particle coordinates, one
might conclude that the three different forms were invented by arrogant
mathematicians to create more confusion and show off their sophistication.
This is not at all the case, however. Hamilton and Lagrange equations do
reduce to the simple Newtons law, known from high-school physics, only
when the systems motion is described using Cartesian coordinates, and
when potential energy is a function of coordinates only, but not momenta or
velocities, and the kinetic energy does not depend on coordinates, but is a
function of only momenta or velocities. This choice is not always the most
convenient, as we tried to demonstrate above for the harmonic oscillator
approximation to the Lennard-Jones dimer. The general theorem
concerning the equivalence of Lagrange and Hamilton equations is
beyond the scope of this book; the interested reader is referred to the
excellent books by Arnold [29], Goldstein [30] or any other classical
mechanics text. The power of Lagrange or Hamilton equations is realized in
cases where the generalized, not necessarily Cartesian coordinates, are more
convenient to use. As the chemical bond lengths are usually fixed in
simulations of macromolecules, the bond angles and torsion angles (Fig. 1)
are a natural choice as generalized coordinates. Fixing all or some bond
angles in addition to the bond lengths is also practiced in some cases.
Torsion angles are often used as sole coordinates in biological
applications [12,3133] in which one is interested in refining conformations
of a single protein or DNA molecule, or their complexes, and where large
pieces of the molecules, containing a hundred or more atoms, can be
considered moving as rigid bodies. Movements of these rigid fragments can

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

27

be described by changing a few torsion angles instead of hundreds of

Cartesian coordinates. This approach is sometimes used to simulate
atomistic models of polymer chains [34,35]. If bond lengths and bond
angles are considered constant, the configuration of the system can be
described by the torsion angles, the position of the first atom and the
orientation of the first bond of each chain. It is possible to describe
polyvinyl or polycarbonate chains of a hundred or so monomers and more
than a thousand atoms with just a few hundred variables instead of several
thousand Cartesian coordinates. Such fixed bond length and bond angle
models are particularly effective in energy minimization calculations
(molecular mechanics) used to generate representative configurations for
glassy polymers. In dynamic simulations it is advisable to fix the bond
lengths but let the skeletal bond angles be flexible, to allow for cooperativity
in the motion of bond angles and torsion angles.
As good things never come for free, the reduction in the number of
degrees of freedom when generalized coordinates are used is traded for more
complex equations of motion. The kinetic energy depends on both angles
and angular velocities, and solving the Lagrange equations in generalized
coordinates requires computationally expensive matrix inversions. Besides,
the nonbonded potential energy is usually a function of Cartesian coordinates of the atoms, and it is still necessary to perform transformations
between the generalized and Cartesian atom coordinates every time the
energy and force are evaluated.
The Lagrange equations of motion allow another useful alternative. For
instance, if the freely-jointed chain is described in Cartesian coordinates, the
monomer motions are constrained by the constant bond lengths bK. For a
polymer chain of NK 1 segments and NK bonds there is a set of constraint
equations,
gi ri1
ri 2
b2K 0

33

for every bond i. The Lagrangian is modified by including additional terms

for the constraints:*

LK
V

NK
X

i gi

34

i1

*This technique is usually called the Method of Lagrange Multipliers in physical and mathematical literature.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

28

Kotelyanskii and Theodorou

The set of Lagrange multipliers

i is added to the set of 3(NK 1) Cartesian
coordinates tracked during the numerical solution of the equations of
motion. Accordingly, NK equations for the
is should be derived from
(22) in addition to the 3(NK 1) equations for the Cartesian coordinates.
These additional equations are equivalent to the constraint equations (33).
The physical meaning of the Lagrange multiplier associated with a bond
length constraint is the magnitude of the force that is exerted on the particles
connected by the bond in order to maintain the bond length fixed. Bond
length constraint equations are solved iteratively in the SHAKE and
RATTLE molecular dynamics algorithms, popular in polymer simulations.
These algorithms will be described in detail in the following chapters.
Naturally, the Lagrange multiplier approach can be generalized to
constrain bond angles, or any other geometrical characteristics, e.g.,
distances between particular atoms, as in protein or DNA structure refinement, based on the experimental information obtained from X-ray or NMR
experiments.
Solving the Lagrange equations of motion in the presence of holonomic
constraints for bond lengths and bond angles amounts to sampling the
rigid polymer model; here, the constraints are considered as being
imposed from the beginning. Alternatively, one can consider the bonds and
bond angles as being subject to harmonic potentials and take the limit of the
properties as the force constants of these potentials are taken to infinity
(flexible model in the limit of infinite stiffness); this model is sampled by
Monte Carlo simulations with constant bond lengths and bond angles. The
two models differ in their kinetic energy function and have nonequivalent
statistical mechanics. One can make the rigid model sample the configuration space of the flexible model in the limit of infinite stiffness by adding to
the potential energy function the Fixman potential [13,11].

D. Mechanical Equilibrium, Stability

Equation (21) tells us that force equals negative potential energy gradient,
and therefore points in configuration space tend to move in the direction
of decreasing potential energy. The position of (mechanical) equilibrium
is defined as a point where all forces equal zero, hence the potential energy
gradient must be zero at equilibrium. This, in turn, implies that the equilibrium point must be an extremum of the potential energy as a function of
the generalized (or Cartesian) coordinates.
Stability of the equilibrium point requires that the system returns back to
the equilibrium state in response to a small perturbation. This excludes
maxima and saddle points, leaving only potential energy minima as

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

29

candidates for stable equilibrium points. Indeed, if we make a small

displacement from a maximum or from a saddle point, we will find a point
with potential energy lower than its value at the extremum.
As the force is directed in the direction of decreasing potential energy, it
points away from the extremum near the maximum or saddle point, pulling
the system out of a state of unstable equilibrium. Only in the vicinity of the
potential energy minima is the potential energy always higher than at the
extremum itself, and the force is directed back towards equilibrium.
Let us consider for example a system with one degree of freedom, q. If the
system has more than one degree of freedom, its Hessian matrix of second
derivatives can be diagonalized, and the problem is reduced to several
independent one-dimensional problems. The kinetic energy being K 12m q_ 2 ,
the potential energy can be expanded in Taylor series in terms of the
deviation u from the equilibrium point q0:
V Vq0



@V 
1 @2 V  2
u

u Ou3
@q q0
2 @q2 q0

35

As q0 is an equilibrium point, the first derivative is zero and the leading term
is quadratic. Setting k @2 V=@q2 q0 , the equation of motion becomes
mu
ku
ut A exp
! t B exp! t

36

with coefficients A and B

by the initial conditions u_ t 0 v0 ,
pdetermined

ut 0 u0 , and !
a=m. If a<0, ! is real, and the solution for u

contains a term that is exponentially increasing with time, so that, given a
small initial deviation from equilibrium, u0, the deviation will increase exponentially with time. The equilibrium is unstable in this case.
If, on the contrary, a is positive, ! is imaginary (! i!. As in the case of
the harmonic oscillator (29), the solution becomes
ut A0 sin!t B0 cos!t

37

which describes bounded oscillations around the equilibrium configuration.

The equilibrium is stable. These results are in agreement with what we see in
the two-dimensional example of Fig. 6.
When there is more than one degree of freedom, in the vicinity of the
equilibrium, the Hessian matrix of second derivatives of the potential energy
can be diagonalized and the problem can essentially be reduced to a set of

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

30

Kotelyanskii and Theodorou

FIG. 6 Different potential energy extrema. Minimum (a), saddle point (b), and
maximum (c). Force f
rV points along the direction of the fastest potential
energy decrease.

independent one-dimensional problems by a linear coordinate transformation u ! u . In terms of the new coordinates,
in the vicinity of the stationary
P
point, the potential energy V 12 i ki u 2i . When all ki are positive, the
potential energy has a minimum and the equilibrium is stable. When at least
one of ki is negative the equilibrium is unstable, as any small initial
perturbation along the direction u i will grow exponentially with time, taking
the system away from equilibrium. A maximum of the potential energy
occurs when all ki are negative.
Notice that, if we consider two trajectories starting from two points that
are close to each other in phase space in the vicinity of the stable equilibrium
point, the difference between the two trajectories will always remain of the
same order of magnitude as the initial difference; the two trajectories will
remain about as close to each other as their initial points are.
In contrast to this, if the two points are in the vicinity of an unstable
equilibrium point, the difference between the trajectories will grow
exponentially with time, and the trajectories will diverge. Thus, in the
vicinity of an unstable equilibrium point, the small uncertainty in the initial
condition will grow exponentially, with characteristic time of !
1 . Any
attempt to predict the system motion for a time much longer than that will
fail. Notice, also, that the same argument applies to any two trajectories, as
soon as they are not confined to the immediate vicinity of the stable
equilibrium. If the system is unharmonic, as almost all systems are, and its
trajectories are not confined to the vicinity of a stable equilibrium, the
trajectories are exponentially divergent.

E. Statistical Description, Ergodicity

In principle, in computer simulations we could specify the initial coordinates
and velocities for all atoms in the system, follow the trajectory by solving the
equations of motion, and calculate some properties of interest from this

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

31

particular realization of the systems trajectory. A real experiment is very

different; we have no control over the positions, velocities or spin states of
each atom or electron in the molecules. So, it is natural to question if any
relationship exists between the results of such calculations and the values
observed in real experiments. Material properties, such as the equation of
state, the elastic moduli, thermal conductivity, viscosity, etc., originate from
molecular motions and interactions. The relationship between molecular
geometry and interactions and macroscopic properties is established by
statistical mechanics. Statistical mechanics tells us that it is really not
necessary, and moreover it does not even make any sense, to know the exact
system state at any moment of time to predict its observable properties. The
observable properties, such as pressure, temperature, strain, heat conductivity, diffusion coefficients, polarization, etc., are related to the average
values of different functions, which in turn depend on the system
microscopic state, calculated along many different trajectories with different
initial conditions. Therefore, the relevant quantity to look for is the
probability to observe a particular value of the energy, polarization, or any
other observable of interest.
The probabilistic approach of statistical mechanics is justified by the fact
that, except for some special cases, such as the harmonic oscillator or the
ideal gas, system trajectories in phase space are very complex and virtually
unpredictable, as discussed in the previous section.
The very important property of ergodicity states that the time averages
along dynamical trajectories of an equilibrium system are equivalent to
phase space averages. In other words, if we are interested in macroscopic
thermodynamic properties, such as pressure, temperature, stress, average
polarization, etc., it is not necessary to follow the system dynamics exactly.
It is sufficient to sample enough points in phase space and to calculate the
proper average. Ergodicity is based on the assumption (provable for some
Hamiltonians) that any dynamical trajectory, given sufficient time, will visit
all representative regions in phase space, the density distribution of points
in phase space traversed by the trajectory converging to a stationary
distribution.
Observed equilibrium properties are time averages over long dynamical
trajectories:
1
t!1 t

lim

Aq , p d hAiobs

38

By virtue of ergodicity, they can also be calculated as averages over phase

space, with respect to an appropriate equilibrium probability density
distribution.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

32

Kotelyanskii and Theodorou

Ergodicity is very important for the statistical mechanical and

thermodynamic descriptions to be valid for particular systems; it has to
be checked for each particular case. It holds for systems where the
characteristic molecular relaxation times are small in comparison to the
observation time scale. Polymeric and other glasses represent a classical
example of nonergodic systems. They can be trapped, or configurationally arrested, within small regions of their phase space over very long
times.
In the context of equilibrium simulations, it is always important to
make sure that the algorithm used in the simulation is ergodic. In other
words, that no particular region in phase space is excluded from sampling
by the algorithm. Such an exclusion would render the simulation wrong,
even if the simulated object itself is ergodic. As an extreme example
consider an algorithm to simulate various conformations of the freelyjointed chain in three dimensions, which, for some reason, such as a
programming error, never selects bonds parallel to the z axis. Evidently,
many representative conformations will be erroneously excluded. Of
course, such a programming error is easy to find and fix. Often the
situation can be more complex, however. It is quite common for lattice
simulation algorithms to sample only the odd or even numbered lattice
sites, or to be unable to find their way out of particular configurations,
especially in two dimensions [36].
In simulation practice, ergodicity of the system can and should be
checked through reproducibility of the calculated thermodynamic properties
(pressure, temperature, etc.) in runs with different initial conditions.

F. Microscopic and Macroscopic States

The microscopic state of the system defines coordinates, momenta, spins for
every particle in the system. Each point in phase space corresponds to a
microscopic state. There are, however, many microscopic states, in which
the states of particular molecules or bonds are different, but values of the
macroscopic observables are the same. For example, a very large number of
molecular configurations and associated momenta in a fluid can correspond
to the same number of molecules, volume, and energy. All points of the
harmonic oscillator phase space that are on the same ellipse in Fig. 5 have
the same total energy.
The set of values of the macroscopic observables, such as temperature,
pressure, average polarization or magnetization, average chain end-to-end
distance, etc., describes the systems macroscopic state. One macroscopic
state combines all the microscopic states that provide the same values of the
macroscopic observables, defined by the macroscopic state.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

33

The probability to observe a certain macroscopic state of the system (or,

in other words, certain values for the observables) equals the sum of the
probabilities of all the corresponding microscopic states.

G. Probability Distribution of the Microscopic States.

Statistical Ensembles
We use the abbreviation  to denote the set of coordinates and momenta
corresponding to one microscopic state.* If a Cartesian description is used,
 (r1, r2 , . . . , rN, p1, p2 , . . . , pN). The probability density for finding a
system in the vicinity of  will be denoted as (). () depends on the
macroscopic state of the system, i.e., on the macroscopic constraints
defining the systems size, spatial extent, and interactions with its environment. A set of microscopic states distributed in phase space according to a
certain probability density is called and ensemble. According to the
ergodicity hypothesis we can calculate the observables of a system in
equilibrium as averages over phase space with respect to the probability
density of an equilibrium ensemble. Equation (38) for the average can be
rewritten as
1
hAobs i lim
t!1 t

Aq , p d
Z
hAji A d
0

39

H. Liouville Equation
Imagine that we start with a set of identical systems, whose states are
distributed in phase space according to a density distribution () at time
t 0, and let the systems move according to their equations of motion. The
ensemble constituted by the systems (points in phase space) evolves in time.
As the systems evolve, the density distribution () should, in general,
change with time. However, systems just move, no new systems are created,
and none of the systems is destroyed. Therefore, there should be a
conservation law for the probability density, similar to the continuity
equation (mass conservation) of hydrodynamics. The conservation law in
phase space is called the Liouville theorem. It states that the total time

*Or to one point in phase space.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

34

Kotelyanskii and Theodorou

derivative of the probability density is zero. For a classical system of Nf

degrees of freedom qi and momenta pi, the Liouville equation is:

@, t X
@
@

, t 0
p_ i
q_i
@t
@qi
@pi
i

40

Combining with the Hamiltonian equations of motion, (20), one obtains:

@, t X @H @
@H @

, t 0


@t
@pi @qi @qi @pi
i

41

which is the Liouville equation.

For an equilibrium system there should be no explicit time dependence
for , i.e., a stationary distribution should be reached by the ensemble.
The main postulate of statistical mechanics states that, for an equilibrium
system of given mass, composition, and spatial extent, all microstates with
the same energy are equally probable [37,38]. This postulate, along with the
ergodic hypothesis, can be justified on the basis of the mixing flow in phase
space exhibited by the dynamical trajectories of real systems [28]. It means
that all microstates of an isolated system, which does not exchange energy
and material with its environment, should occur equally often.
The equilibrium probability density in phase space for a system with total
energy E0 is therefore given by:
  H
E0

42

The (x) is a Dirac -function, which is nonzero only when its argument
equals zero. Here it selects those microscopic states  that have total energy*
H equal to E0.
The above ensemble of systems with constant number of particles N,
occupying constant volume V, with the total energy E conserved, is called
the microcanonical, or (NVE) ensemble.

I.

Partition Function, Entropy, Temperature

A very important measure of the probability distribution of an equilibrium

ensemble is the partition function Q. This appears as a normalizing factor in
the probability distribution defined by the ensemble.

*Remember the denition of the Hamiltonian, Eq. (20).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

35

For the microcanonical ensemble:

X
QNVE
H
E0


 Q
1
NVE H
E0
X
H
E0 A
hAi Q
1
NVE

43

The summation over states  here and in the previous equations is used for
the quantum case, where microscopic states are discrete and () has the
meaning of a probability. For one-component classical systems, on which
we have mainly focused so far, the sum should be replaced by the integral:
Z
X
1 1
d
!
N! h3N

d

N
Y

d 3 r i d 3 pi

44

i1

N! here takes care of the indistinguishability of particles of the same species.

For a multicomponent system of N1 particles of type 1, N2 particles of type
2 , . . . , N! should be replaced by N1!N2! . . . . h is Plancks constant. It
describes the phase space volume occupied by one state and renders the
product dr dp dimensionless.
The partition function defines the thermodynamic potential. The
partition function is a function of the thermodynamic state variables that
are kept constant in the definition of the equilibrium ensemble. The
expression of the thermodynamic potential in terms of these state variables
constitutes a fundamental equation of thermodynamics.
The proper thermodynamic potential for the microcanonical ensemble is
the entropy:

S=kB
ln QNVE

45

where kB R/NA is the Boltzmann constant. The argument of the logarithm

on the right hand side of Eq. (45) is just the number of states with the energy
E0. We therefore have a statistical thermodynamic definition of entropy as a
quantity proportional to the logarithm of the number of microscopic states
under given N, V, E [37,38]. The proper way to simulate the microcanonical
ensemble is to numerically solve the equations of motion (20) or (21) or (22)
for the closed system. This is done in the simplest versions of molecular
dynamics simulation.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

36

Kotelyanskii and Theodorou

If the system of N particles is contained in the constant volume V but

is allowed to exchange energy with its environment through diathermal
walls, its energy is not constant any more. The energy is fluctuating, but
the temperature is constant [37,38]. Temperature describes the probability
distribution of energy fluctuations. Such a system is represented by the
canonical or NVT ensemble. The probability density is given by the Gibbs
distribution:



H

1
NVT  QNVT exp

kB T



Z
1 1
H
d exp

QNVT
N! h3N
kB T
A

lnQNVT
46
kB T
A, the Helmholtz energy, is a thermodynamic potential for the canonical
ensemble. QNVT, often symbolized simply as Q, is the canonical partition
function. The last of Eqs. (46) defines a fundamental equation in the
Helmholtz energy representation by expressing A as a function of N, V, T.
Often in classical systems it is possible to separate the energy contributions
that depend on the momenta only (kinetic energy K) from the potential
energy V, which depends only on the coordinates. When Cartesian coordinates are used as degrees of freedom, for example, the partition function can
be factorized as:
QNVT

1 1
N! h3N

Z Y



N
K
V
d 3 pi exp

d 3 ri exp

kB T
kB T
i1
i1

Z Y
N

ex
Qig
NVT QNVT

47

into two contributionsthe partition function of the ideal gas:

VN
N!3N


1=2
h2

2mkB T

Qig
NVT

48

( being the thermal or de Broglie wavelength), and the excess part:

N
Qex
NVT V


V
d 3 ri exp

kB T
i1

Z Y
N

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

49

Background

37

It is also customary to use the configurational integral, defined as

ZNVT

Z Y
N

V
d ri exp

k
BT
i1
3


50

instead of the Qex

NVT .
As a consequence, all the thermodynamic properties can be expressed as a
sum of an ideal gas part and an excess part. All specifics of systems are
included in the latter, and more attention is usually focused on it. In fact, in
Monte Carlo simulations the momentum part of the phase space is usually
omitted, and all calculations are performed in configuration space. The ideal
gas contribution is added after the simulations to compare to experiments or
to other simulations.
Another important consequence of Eq. (47) is that the total average
kinetic energy is a universal quantity, independent
interactions in the
P of the
2
p
=2m
with respect to
system. Indeed, computing the average of K N
i1 i
the probability distribution of Eq. (46) and using the factorization of Eq.
(47) we obtain that hKi 3=2 NkB T or, more generally, hKi 1/2 Nf kBT
for a system of Nf degrees of freedom. If the kinetic energy can be separated
into a sum of terms, each of which is quadratic in only one momentum
component, the average kinetic energy per degree of freedom is 1/2 kBT.
The last result is a special case of the equipartition theorem which characterizes classical systems [38,37]. It is often used in simulation practice to test for
system equilibration.
When, in addition to the energy, a system is allowed to exchange volume
with the environment by being contained in (partly) movable boundaries, as
for instance is the gas in a vessel covered by a piston, both the volume and
energy are fluctuating, but the temperature and pressure are constant. Such
a system is represented by the isothermalisobaric or NPT ensemble. The
probability density is given by the distribution:



H PV
NPT , V Q
1
exp


NPT
kB T



Z
PV
QNPT dV=V0 QNVT exp

kB T



Z
Z
1 1 1
H PV
dV d exp

QNPT
N! h3N V0
kB T
G

lnQNPT
kB T

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

51

38

Kotelyanskii and Theodorou

The last of Eqs. (51) defines a thermodynamic fundamental equation for

G G(N, P, T) in the Gibbs energy representation. Note that passing from
one ensemble to the other amounts to a Legendre transformation in macroscopic thermodynamics [39]. V0 is just an arbitrary volume used to keep the
partition function dimensionless. Its choice is not important, as it just adds
an arbitrary constant to the free energy. The NPT partition function
can also be factorized into the ideal gas and excess contributions. The
configurational integral in this case is:
Z
ZNPT

Z Y



N
PV
V
3
dV exp

d ri exp

kB T
kB T
i1

52

Some authors also include the N! and V0 factors in the configurational

integral. A simulation in the isothermalisobaric ensemble should be conducted with the volume allowed to change, but the way these changes are
implemented must provide for the proper probability density distribution
given by Eq. (52).
The grand canonical ensemble describes a system of constant volume, but
capable of exchanging both energy and particles with its environment.
Simulations of open systems under these conditions are particularly useful
in the study of adsorption equilibria, surface segregation effects, and
nanoscopically confined fluids and polymers. Under these conditions, the
temperature and the chemical potentials i of the freely exchanged species
are specified, while the system energy and composition are variable. This
ensemble is also called the VT ensemble. In the case of a one-component
system it is described by the equilibrium probability density

H
N
exp

 VT , N
kB T



X
N
QNVT exp
Q VT
kB T
N


Z



X 1 1
N
H
exp
Q VT
d
exp


N! h3N
kB T
kB T
N
Q
1
VT

=kB T
lnQ VT

53

Q VT, often symbolized as VT, is the grand partition function.  is the

grand potential. For bulk systems, in the absence of interfaces, it can be
shown that 
PV [37]. Simulations in the grand canonical ensemble
should allow the total number of particles in the system and its energy

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

39

to fluctuate. These fluctuations should occur according to the  VT(, N)

probability distribution.
Notice that, to formulate the NPT and VT ensembles starting from the
NVT ensemble, we had to add one more product term to the numerator of
the exponent (
PV and N, respectively). This product contains the
variable of the original ensemble that is allowed to fluctuate in the new
ensemble (V and N for the NPT and VT, respectively) times its conjugate
thermodynamic variable or field, which is kept constant in the new
ensemble (
P and , respectively). The partition function of the new
ensemble is derived by integration or summation of the partition function of
the original (NVT) ensemble over all values of the variable allowed to
fluctuate. Of course, these observations are by no means just a coincidence.*
Remember that an ensemble is equivalent to a particular phase space
probability density distribution. Each of these ensembles can be considered
as a superset of another ensemble: the canonical ensemble is a set of
microcanonical ensembles with different energies weighted by the factor
exp(
H/(kBT)); the NPT ensemble is a set of NVT ensembles with different
volumes weighted by exp(
PV/(kBT)); the grand canonical ensemble is a set
of canonical ensembles with different numbers of particles weighted with
exp( N/(kBT)). In the thermodynamic limit (system increasing in size to
become macroscopic, while all intensive variables are kept constant), the
integration or summation operation by which one passes from one ensemble
to the other is equivalent to a Legendre transformation.
To appreciate the generality of this approach, consider, for instance, a
system of N molecular dipoles at constant volume V. We can introduce an
ensemble to describe this system under constant external electric field E,
at constant temperature T. The energy contribution due to the electric field
equals the negative product of the field and the total dipole moment of the
system,
VP  E, where P is the electric polarization of a configuration. The
probability density and the partition function for this ENVT-ensemble is:

H
VP  E
ENVT 
exp

kB T






Z
VP  E
H
QENVT d exp
exp

kB T
kB T
Q
1
ENVT

54

The macroscopic polarization P hP i resulting from this field will no

longer be zero for nonzero E.
*More detailed and rigorous discussion can be found in [38,37,40], or any other general
thermodynamics or statistical mechanics textbook.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

40

Kotelyanskii and Theodorou

Many more different ensembles can be introduced; we shall see some of

them in the following chapters. When designing a new simulation technique,
it is always important to understand which ensemble it actually simulates, as
this determines how to interpret the simulation results.
With all the above said, it is natural to expect, that equations of motion,
phase space, statistical ensemble, and other abstract mathematical concepts
we spend so much effort describing in this and previous sections, do indeed
have something to do with computer simulations. In fact, various simulation
techniques are nothing else but methods for the numerical solution of
the statistical mechanics given a Hamiltonian H(). They allow for
the realization of these abstract concepts. It is through the principles
of statistical mechanics, which we have briefly described above, that the
numbers produced by the computer simulation program are linked to the
results of real-life experiments and to the properties of real materials.

III. PROPERTIES AS OBTAINED FROM SIMULATIONS.

AVERAGES AND FLUCTUATIONS
According to statistical mechanics, physically meaningful results, comparable to experiment, are obtained by calculating phase or configuration space
averages, with different microscopic states weighted according to the proper
ensemble probability density distribution. Different simulation techniques
are essentially different ways of sampling the systems phase space with the
proper probability density in order to calculate the thermodynamic averages.
Given a way to generate microscopic states according to the equilibrium
ensemble distribution, we can calculate any thermodynamic average simply
as an average over these states. This is how thermodynamic averages are
calculated in Monte Carlo simulations. The simplest example is of course
the configurational part of the internal energy in thermodynamics, which
equals the average potential energy over the set of configurations generated
in the Monte Carlo simulation run. Kinetic energy is trivial as soon as the
equipartition theorem holds; its average is just 1/2 kBT for each degree of
freedom, and can be added after the run to compare the total average energy
to the experiments.
Any equilibrium property that has a microscopic quantity associated with
it, such as polarization, stress, average orientation, average specific volume,
average polymer chain conformation, or density in the isobaric or grand
canonical ensembles, can be calculated as a simple average of values for the
corresponding microscopic parameter over the random configurations
generated by the Monte Carlo algorithm, according to the desired ensemble
probability distribution.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

41

A. Pressure
Pressure is one of the most important thermodynamic properties. According
to the previous subsection, to calculate pressure from a simulation we need a
function of the coordinates and momenta of the particles, whose average
equals the thermodynamic pressure.
In equilibrium thermodynamics, pressure is defined as a negative
derivative of the Helmholtz energy with respect to volume:

@A 
1 @Q
1 @Z
kB T
kB T
@V N, T
Q @V
Z @V
Z
Z
@Z
@

. . . exp
V r1 , . . . , rN =kB T d 3 r1    d 3 rN
@V @V V
V
P

55

Here we consider, for simplicity, a classical system described in terms of

Cartesian coordinates. In this case the kinetic energy contribution to the
partition function is independent of the coordinates and volume, and
Eq. (55) allows us to consider only the configurational integral Z [Eq. (50)]
instead of the full partition function Q. To calculate the derivative of the
integral with respect to volume, we also assume, for simplicity, that
particles are in a cubic container* with edge length V1/3. Now let us make
a transformation to the normalized coordinates si: {r1 , . . . , rN}
{V1/3s1 , . . . , V1/3sN}. This transformation changes the integration over the
volume to an integration over a cube with sides of unit length.



Z1
Z1
@Z
@
N
3
3

V

exp
V r1 , . . . , rN =kB T d s1 . . . d sN
@V @V
0
0
Z1
Z1

exp
V r1 , . . . , rN =kB T d 3 s1 . . . d 3 sN
NV N
1
0

56

Z
Z1
VN 1

...
exp
V r1 , . . . , rN =kB T
kB T 0
0
"
#
N
X
@V
1
1 3


 rk V

d s1    d 3 sN
@r
3
k
k1
N

P
Z

@V=@rk  rk
N
1 k1
Z

VkB T
V
3
*The following argument is obviously valid for any complex container shape, but the calculations are much longer. In the case of two dimensions the factor 1/3 is to be replaced by 1/2.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

42

Kotelyanskii and Theodorou

So, from Eq. (55),

*
+
N
1 X
@V


 rk
PV NkB T

3 k1 @rk

57

Usually, V depends on distance vectors ri
rj between sites. Then [41] one

can rewrite Eq. (57) as
*
+

1 X
N
X
1 N
@V
 ri
rj


PV NkB T

3 i1 ji1 @ri
rj

58

The quantity @V=@ri
rj can be taken as a definition of the force fkl on site
k due to site l, even in cases where V is not merely a sum of pairwise-additive
sitesite potentials [41]. Then,
*
+

1 X
N
X
1 N
f ij  ri
rj
PV NkB T

3 i1 ji1

59

Equation (59) is useful in simulations where periodic boundary equations

are employed. In the presence of periodic boundary conditions Eq. (57)
should not be used [41]. The product of the force acting on a particle
times its position vector is called virial, so Eqs. (57)(59) are forms of the
(atomic) virial theorem for the pressure.
Equation (59) remains valid in the presence of constraints, such as fixed
bond lengths and/or bond angles. In this case, if i and j are involved in a
constraint, fij is the Lagrange multiplier force acting on i from j due to the
constraint. If the rigid model is sampled in the simulation and
properties of the flexible model in the limit of infinite stiffness are
desired, fij must additionally incorporate contributions from the Fixman
potential.
In molecular systems one can write an equation similar to (59), with fij
being the total force exerted on molecule i from molecule j (i.e., the sum
of all sitesite interactions between the molecules) and ri
rj being
the difference between the position vectors of the centers of mass of the
two molecules (molecular virial equation [41]). The molecular virial
equation is particularly convenient to use in systems of chain molecules,
because it does not require knowledge of intramolecular bonded or
constraint forces.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

43

B. Chemical Potential
A very important theorem due to Widom [42] allows expression of the excess
chemical potential of a component in a fluid (i.e., the chemical potential of
the component minus the chemical potential it would have if it were an ideal
gas under the temperature and molar density with which it is present in the
fluid) as an ensemble average, computable through simulation. The Widom
theorem considers the virtual addition of a test particle (test molecule) of
the component of interest in the fluid at a random position, orientation, and
conformation. If V test is the potential energy change that would result from
this addition,
ex
i








V test
V ig

kB T ln exp


ln exp

kB T
kB T

60

where the first average is taken over all configurations of the fluid and all
positions, orientations, and conformations of the added test molecule, while
the second average is a weighted average of the intramolecular energy of one
molecule over all conformations adopted in the ideal gas state. Equation
(60) holds in the canonical ensemble. Analogous expressions for the NVE
and NPT ensembles have been derived [43]. Also, bias schemes that can deal
with the difficulties of randomly inserting long flexible molecules in the
simulated phase of interest have been developed [44].
The estimation of the chemical potential is also possible through virtual
deletion of a real particle (inverse Widom scheme), provided the bias
associated with creating a hole in the fluid as a result of the removal is
correctly accounted for [45].

C. Fluctuation Equations
The heat capacity, isothermal compressibility, and other thermodynamic
properties that correspond to second derivatives of the free energy, can be
calculated from the fluctuations in different ensembles [43,38,37]. Let us
consider the heat capacity in the NVT ensemble, as the simplest example.
With E hHi the internal energy, computed as an ensemble average of the
Hamiltonian, we have


@E 
@hHi

cV 
@T N, V
@T N, V
Z
1 1
H exp
H=kB T d
hHi
Q h3N N!

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

61

44

Kotelyanskii and Theodorou

Z
@hHi
1 @Q 1

2
H exp
H=kB T d
@T
Q @T h3N N!

2 
Z
1 1
H

exp
H=k
T
d
B
Q h3N N!
kB T 2



Z
@Q
1
H
1

hHiQ
exp
H=kB T
d
@T h3N N!
kB T 2
kB T 2
cV

hH2
hHi2 i H2

kB T 2
kB T 2

Similar manipulations in the NPT ensemble lead to equations for the

isobaric heat capacity cp and isothermal compressibility T
hH PV 2 iNPT kB T 2 cP

62

hV iNPT VkB TT

More fluctuation expressions can be derived using similar manipulations
[38,37,43].
As the expressions for averages are different in different ensembles, so are
the relationships between the free energy derivatives and fluctuations. The
fluctuation equation for cV, (61), is not valid in the NPT or any other
ensemble. Likewise, Eqs. (62) are valid in the NPT ensemble only.

D. Structural Properties
One of the main goals of molecular simulations since their very early days
was to obtain information about structure together with the thermodynamic
properties, given the interaction potentials between the atoms.
The most important structural characteristics are the density distribution
and correlation functions. The n-particle distribution function g(n)(r1 , . . . , rn)
is defined so that ng(n)(r1 , . . . , rn), with  being the mean density of particles
in a system, is the probability density for finding a particle at position r1, a
particle at position r2, . . . , a particle at position rn.
Particularly important is the pair distribution function, g(2)(r1, r2), or
simply g(r1, r2). In a homogeneous system this is only a function of the
interparticle distance vector r r2
r1. The quantity g(r) can be interpreted
as the mean local density of particles at position r relative to a particle that is
taken as reference. The quantity h(r) g(r)
1 is called the (total) pair
correlation function. If the system is also spherically symmetric, as in the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

45

case of an unperturbed liquid or glass, g and h depend only on the magnitude

of the separation, r |r|. This of course does not apply to oriented polymers,
or to a polymer melt under flow. In this case g is also called the radial
distribution function, g(r).
The structure of low density gas phases and crystalline solid states is easy
to describe, because the correlation functions for them are easy to obtain. In
low density gases, intermolecular interactions are negligible and there are no
correlations between the particles. g(r) is 1 for all r. At the other extreme, in
crystalline solids correlations are very strong but, due to the lattice
symmetry, the structure and correlation functions can be easily determined.
Correlation functions are particularly important for dense disordered
systems such as liquids and glasses, and hence for polymers, for which the
crystalline state is rather the exception than the norm. Different models used
in classical liquid state theory are essentially different approximations to
calculate the correlation functions. All thermodynamic properties can be
expressed in terms of the structure correlation functions and interaction
potentials [37,46].
The Fourier transform of the pair correlation function, the structure
factor, can be measured experimentally by X-ray, neutron, or light
scattering techniques [37,38,43,47]. Moreover, in the simple and often
used approximation, whereby all the potential energy is expressed as a sum
of pair interaction u(r) over all pairs of atoms, knowing g(r)  g(2)(r) allows
one to calculate all thermodynamic properties.
Energy and pressure, in an isotropic system for instance, are:
Z1
3
1
E NkB T N
urgr4r2 dr
2
2
0
Z1
1
dur
gr4r2 dr
PV NkB T
N
r
6
dr
0

63

Essentially, integration with the g(r) replaces sums over all pairs, and similar
expressions can be derived for averages of any pairwise function.
Calculating the pair distribution function in a simulation is straightforward; all we need to do is count the number of atom pairs separated by a
distance in the range from r to r r, and then normalize it. Usually, g(r) is
normalized by the number of pairs Nig(r, r r) that would be observed
in an ideal gas of the same density, so that, in the limit of large distances,
r ! 1, where correlations disappear, g(r) ! 1. A typical FORTRAN code
to calculate g(r) from 0 to rmax with the resolution deltar is given below.
The separation distance histogram is calculated for nbins rmax/
deltar bins, for a model of N particles in a box with sides boxx,

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

46

Kotelyanskii and Theodorou

boxy, and boxz along the x, y, and z directions, respectively, with the
atomic coordinates stored in the arrays x(i),y(i),z(i).*
integer nbins, ibin, N
double precision hist(0:nbins)
double precision g(0:nbins)
double precision x(N),y(N),z(N)
double precision boxx,boxy,boxz,invbx,invby,invbz
double precision anorm, pi
double precision volume, rmax, deltar, dist
pidacos(
1.0d0)
.........................................
volumeboxx*boxy*boxz
anorm N*(N
1)/volume * 4.d0/3.d0 * pi
invbx1.d0/boxx
invby1.d0/boxy
invbz1.d0/boxz
nbinsrmax/deltar
do i1,N
1
do ji1,N
dx x(i)
x(j)
dx dx
boxx*anint(dx*invbx)
dy y(i)
y(j)
dy dy
boxy*anint(dy*invby)
dz z(i)
z(j)
dz dz
boxz*anint(dz*invbz)
dist sqrt(dx*dxdy*dydz*dz)
ibin int(dist/deltar)
hist(ibin) hist(ibin)2.d0
enddo
enddo
do ibin1,nbins
r ibin*deltar
g(ibin)hist(ibin)/anorm/(r**3
(r-deltar)**3)
enddo
.................

The above code fragment would be run for each configuration of the
simulation run, and the corresponding arrays would have to be averaged
over the configurations.

*This code is not optimized for performance. The periodic boundary conditions are taken into
account using the intrinsic function anint().

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

47

E. Time Correlation Functions. Kinetic Properties

Transport coefficients, such as the diffusion coefficient, the viscosity, and
the thermal conductivity describe a systems response in the time domain to
time-dependent perturbations. If the perturbation is not very strong, the
response depends linearly on the perturbation and is described by a
corresponding transport coefficient. For instance, if a concentration
gradient is created in the system, the system will respond by developing
a mass flux proportional to the magnitude of the concentration gradient,
which tends to equalize the concentrations everywhere in the systemthis
is the well known Ficks law, and the proportionality coefficient is the
diffusion coefficient. If a velocity gradient is imposed on a liquid, the
liquid responds with a stress (momentum flux) which is proportional to
the velocity gradient, the proportionality coefficient being the viscosity.
Similarly, when an electric field is applied to a conductor, the conductor
responds with an electric current density (charge flux) proportional to the
electric field, and the proportionality coefficient is the conductivity. This
approach is applicable to a wide variety of different phenomena; it is called
linear response theory, and is described in detail in [40,37,38]. The
coefficients of proportionality between the relaxation rate and the applied
driving force are called kinetic coefficients.
Of course, one way to determine kinetic coefficients from simulation is to
simulate the perturbation and measure the systems response directly. This is
a feasible approach in some systems, which will be discussed in one of the
following chapters.
One of the most important principles of linear response theory relates the
systems response to an externally imposed perturbation, which causes it to
depart from equilibrium, to its equilibrium fluctuations. Indeed, the system
response to a small perturbation should not depend on whether this
perturbation is a result of some external force, or whether it is just a random
thermal fluctuation. Spontaneous concentration fluctuations, for instance,
occur all the time in equilibrium systems at finite temperatures. If the
concentration c at some point of a liquid at time zero is hci c(r, t), where
hci is an average concentration, concentration values at time t t at r and
other points in its vicinity will be affected by this. The relaxation of the
spontaneous concentration fluctuation is governed by the same diffusion
equation that describes the evolution of concentration in response to the
external imposition of a compositional heterogeneity. The relationship
between kinetic coefficients and correlations of the fluctuations is derived in
the framework of linear response theory. In general, a kinetic coefficient is
related to the integral of the time correlation function of some relevant
microscopic quantity.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

48

Kotelyanskii and Theodorou

The time correlation function of two quantities A and B, which are

functions of the phase space point , is defined as:
CAB t hA0Bti

64

Here A A
hAi and B B
hBi are the deviations of A and B from

their average values. The brackets indicate averaging over an ensemble of
systems prepared under the same macroscopic initial conditions and subject
to the same macroscopic constraints; here, equilibrium averages will be
considered. When A and B are the same function, cAA is called the autocorrelation function of A. A normalized correlation function cAB can be
defined as
CAB
cAB q
hA2 ihB2 i

65

The absolute value of this function is bounded between zero and one.
The self-diffusion coefficient D, for instance, is related to the velocity
autocorrelation function. In three dimensions
D

1
3

dthvi 0  vi ti

66

where vi is a center of mass velocity of a single molecule. The shear viscosity

is related to the correlations of the off-diagonal ( 6 ) components of the
instantaneous stress tensor [compare Eq. (57)] P :
X
1 X
pi pi =mi
ri fi
P 
V
i
i
Z1
V

dthP  0P  ti
kB T 0

67

Here, ,  stand for the Cartesian components x, y, z of the corresponding quantityforce acting on the particle fi, particle coordinate ri, or
momentum pi.
Equations (66) and (67) are known as GreenKubo relations for the
transport coefficients.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

49

Alternative, mathematically equivalent expressions, known as Einstein

relations, may be obtained by carrying out the integration by parts:
2 E
1 D
ri t
ri 0
t!1 6t
1 V
hL t
L 02 i
 lim
t!1 2t kB T
1X
L
ri pi
V i
D lim

68
69
70

It is easy to see that P  is a time derivative of L.

The above equations provide two alternative routes for calculating
kinetic coefficients from simulations of a system at equilibrium. Averages in
the above equations are ensemble averages, hence the results are ensemblesensitive. The time correlation functions contain more information than just
the kinetic coefficients. The Fourier transforms of time correlation functions
can be related to experimental spectra. Nuclear magnetic resonance (NMR)
measures the time correlation functions of magnetization, which is related to
the reorientation of particular bonds in the polymer molecule; inelastic
neutron scattering experiments measure the time correlation functions of the
atom positions; infrared and Raman scattering spectroscopies measure
the time correlation function of dipole moments and polarizabilities of the
molecules.
Using the Einstein relations is generally a more robust way to calculate
the kinetic coefficients. One just has to calculate the average on the right
hand side of the Eqs. (68), (69) as a function of t for long enough times,
when it becomes approximately linear in time, and to determine the slope; in
contrast, application of the GreenKubo relations requires long dynamic
simulations to accumulate the tails of time correlation functions with
sufficient precision.
Usually, in computer simulations the time correlation functions are
calculated after the run is finished, using the information saved during the
run (i.e., in a postprocessing stage). One typically saves the instantaneous
values of the quantities of interest during the run in one or several files on
the hard disk, which are processed after the simulation run.
A direct way to calculate the time correlation function from the values
saved during the simulation run* is just to literally implement its definition.
Suppose we have M 1 values of A(t) and B(t), obtained at the regular time
intervals mt, where m is an integer running from 1 to M, stored in the
*Or calculated from the saved congurations.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

50

Kotelyanskii and Theodorou

arrays a and b. To calculate the time correlation function CAB(t) one would
have to scan the array, picking up pairs of values a(l) and b(n),
calculating their product, and saving it in the appropriate location in the
array c(j), j|n
l|, where the values of CAB( jt) are to be stored.
Afterwards, the accumulated quantities have to be normalized by the
number of a and b pairs used for each value of l. The following code
fragment implements the subroutine that takes the values of A(t) and B(t)
stored in the arrays a, b, and returns array c, containing the normalized
correlation function values for times from 0 to mcor t. It also calculates the
averages of A(t) and B(t) a_av and b_av and their root mean squared
deviations delta2a, delta2b, to obtain normalized correlation functions.
A and B are available for time slices 0 to m.
subroutine correl(a,b,c,a_av,b_av,delta2a,delta2b,mcor,m)
double precision a(0:m),b(0:m),c(0:mcor)
double precision a_av, b_av,a2av, b2av, delta2a,delta2b
double precision aux1,aux2, ai
integer mcor, count (0:mcor), maxt, m, i
a_av0.0d0
b_av0.0d0
a2av0.0d0
b2av0.0d0
c zero c and count
do i0,m
c(i)0.0d0
count(i)0.0d0
end do
c double loop through the arrays
do i0,m
aia(i)
a_ava_avai
b_avb_avb(i)
maxtmin(imcor,m)
do ji,maxt
1j
i
count(1)count(1)1
c(1)c(1)ai*b(j)
end do
end do
c normalize averages
aux 1.0d0/dble(m1)
a_ava_av*aux
b_ava_av*aux

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

51

a2ava2av*aux
b2avb2av*aux
delta2asqrt(a2av-a_av*a_av)
delta2bsqrt(a2av-b_av*b_av)
c normalize corr. function
aux1a_av*b_av
aux21.0d0/(delta2a*delta2b)
do i0,m
c(i)(c(i)/count(i) - a_av*b_av)*aux2
end do
return
end

For most of the modern high-speed CPUs, multiplication takes significantly fewer cycles than division. This is why division is replaced by
multiplication whenever possible. Also, the values unchanged inside the
loops are calculated in advance outside of the loop.
An alternative method for calculating the time correlation function,
especially useful when its spectrum is also required, involves the fast Fourier
transformation (FFT) algorithm and is based on the convolution theorem,
which is a general property of the Fourier transformation. According to the
convolution theorem, the Fourier transform of the correlation function CAB
equals the product of the Fourier transforms of the correlated functions:
^ 
B^ 
C^ AB  A

71

^  and B^  should be calculated by a

The direct Fourier transforms A
Fourier transform routine that can be found in many scientific program
libraries. Then the Fourier transform of the correlation function is calculated using (71). In the case where the autocorrelation function is calculated
(A B), this Fourier transform represents a frequency spectrum of the
system associated with the property A and can be related to experimental
data, as discussed above. The correlation function in the time domain is
then obtained by an inverse Fourier transformation. Fast Fourier transformation routines optimized for particular computer architectures are usually
provided by computer manufacturers, especially for the parallel or vector
multiprocessor systems.

IV. MONTE CARLO SIMULATIONS

In a Monte Carlo simulation many microscopic states are sampled using the
generation of random numbers, and averages are calculated over these

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

52

Kotelyanskii and Theodorou

states. It is this random nature of the technique which is responsible for its
name. The Monte Carlo method was developed in the Los Alamos National
Laboratory in New Mexico in the late 1940searly 1950s and its inventors
named it after Monte Carlo, capital of Monaco, a small country in Southern
Europe famous for its casinos.*
Let us first consider a very simple examplecalculation of the area of a
complex shape, such as the polygon  in Fig. 7. The simplest way would be
to set up a mesh that is fine enough to provide the required accuracy, and to
calculate the number of mesh points inside the polygon. If the complex
shape happens to be a hole in the fence, this can be done by covering the
fence with chicken wire and counting the number of knots inside the hole
and the number of knots in a rectangular section ABCD of the fence which
completely surrounds the hole. The ratio of these two numbers times the
area of the rectangular section gives the area of the polygon.
Alternatively, one might draw a rectangle ABCD of known size
around the polygon and start throwing points inside ABCD at random

FIG. 7

Area calculations.

*Maybe, if the technique had been developed several years later, it could have been called Las
Vegas.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

53

(i.e., according to a uniform distribution in ABCD). Clearly, the probability

for a randomly chosen point to fall inside the polygon equals the ratio of the
hatched area inside the polygon to the known area of the rectangle. Again,
in the spirit of Fig. 7, one might equally well do this by randomly shooting,
or throwing darts, at the brick fence ABCD and counting the number of
shots that went through the hole . It is important to note that the shooter
must be blind, or at least very inexperienced, to guarantee the randomness
of the shots. Naturally, the accuracy of the area estimate increases with the
number of trial points. According to statistics, the error is inversely
proportional to the square root of the total number of trials. This technique
is the simplest variant of the Monte Carlo method. It can easily be
generalized to calculate an integral of some arbitrary function f(x, y) over a
complex-shaped region. Indeed, if we were to calculate not just the area of
the hatched polygon but an integral of function f(x, y) over the polygon ,
we could do this by first extending the integration region to the whole
rectangular region ABCD and, secondly, by redefining the function F(x, y)
to be equal to f(x, y) inside and on the border of  and F(x, y) 0
everywhere else. Now by throwing M random points in the rectangle (x, y)i
we can estimate:
Z
dx dy f x, y


M
1 X
Fx, yi
M i1

72

Of course, if  accounts for a very small fraction of ABCD, the random

technique becomes increasingly wasteful. Imagine that the function F(x, y) is
nonzero over less than 1% of the total area ABCD. Then, on the average, 99
out of 100 shots will contribute zeroes to the average, and only one out of a
hundred will deliver a nontrivial contribution.
To take a polymer example, consider simulations of the bead-spring
model of the linear polymer chain from Fig. 2, at temperature T. The
probability density to find it in a particular conformation , specified by the
set of coordinates of its N monomers,  {r1, r2, . . ., rN}, is a product of
probability densities for the bonds [compare Eq. (12)]; it can be considered
as being a canonical distribution of the form of Eq. (46) (here we use  to
denote a point in configuration space, not in phase space).

3

2

3=2

"
#
N
1
3 X
2
exp
2
ri
ri
1
b3
2b i2

73

distance between successive beads. The quantity

with b2 the mean square
P
2
V 3kB T=2b2 N

r
i2 i
ri
1 can be thought of as an elastic energy

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

54

Kotelyanskii and Theodorou

of the springs, which, as we have pointed out, is entirely of entropic origin.

The mean squared end-to-end distance R2 |rN
r1|2 should be calculated
as:
2

hR i

djrN
r1 j2

74

The simplest Monte Carlo scheme for simulating this chain, as described
above, would generate random conformations , scattered with uniform
probability throughout configuration space; each time it would select the N
bead positions randomly within a three-dimensional domain whose
dimensions are much larger than b. Thus, it would calculate an average of
the product ()|rN
r1|2 over all realizations of the chain.
Most of the randomly generated configurations, corresponding to
random positions of the monomers, however, will have a very small
weight, as the distances between connected beads will in most cases be much
longer than b. The probability density of each of these states () will then
be negligibly small, and so will be its contribution to the right hand side of
Eq. (74). This simplest Monte Carlo scheme will then be spending most of
the time calculating zeroes, almost never finding a configuration in which all
the distances between connected monomers are on the order of b, which
would contribute substantially to the average. This simple approach is very
inefficient.
There is a solution, though. Imagine that we have an algorithm that will
generate microscopic states using pseudorandom numbers, but with a
nonuniform probability density, visiting the relevant lower-energy states
more often and almost avoiding the high-energy states. The perfect choice
would be to sample the phase space with the probability density ()
directly. Then, ensemble averages can be calculated simply as averages over
the generated conformations:

M

1 X
R2
jrN
r1 j2i
M i1

75

The core part of the Metropolis Monte Carlo simulation technique is the
algorithm to sample the microscopic states according to the required
ensemble probability distribution. The Monte Carlo algorithm simulates a
stochastic process (Markov chain) producing a sequence of points or states
in configuration space, in which the choice of the next state depends only on
the current state. The original Metropolis (or MR2T2) Monte Carlo
algorithm, proposed by Metropolis et al. [48], was for sampling the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

55

canonical (NVT) ensemble. The algorithm generates a sequence of random

phase space points according to the following rules: (a) if the state generated
at step k is k with energy value V(k), the candidate for the new state new
is generated randomly with some distribution of attempt probabilities
(k ! new). In the case of an atomic or molecular system this is usually
realized by adding a random displacement to one or several randomly
chosen atoms. The average displacement is of course zero, and its
components are most often just uniformly distributed inside the interval
between
max and max in each coordinate direction. The new state is
accepted with the probability Pacc max(1, exp[
(V(new)
V(k))/(kBT)]).
If the new has lower energy than the current V(k), it is always accepted;
if it has higher energy, it can still be accepted with probability exponentially
decreasing with the energy difference. If new is accepted, k1 new.
Otherwise the old state is counted once again as a new state, k1 k.
Given this description of the Metropolis algorithm, how do we make sure
that it does reproduce a canonical ensemble distribution?

A. Microreversibility
Let us assume that the probability density for being at the phase space point
 after completion of step k in the Metropolis algorithm is k
M . Let the
probability to go from one point  to another 0 in one Monte Carlo step be
W( ! 0 ). This function W, describing how to proceed from one state to
the next, completely defines the algorithm. The following equation describes
the evolution of the probability density during our random process:*
X
0
0
kM 
k
1
k
1
M 
M  W ! 
0

0
k
1
M W ! 

76

0

The equation is actually quite simple; it is similar to the Liouville equation

(41). As we are looking for a stationary algorithm, producing results independent of the initial conditions, we want M to depend only on , not time,
and hence we will set the left hand side to zero. The first sum on the right
hand side describes the influx to , the probability to come there from
other states in step k. It is a sumy over all other states 0 of the product of

*This equation is very important in the theory of random processes; it has many dierent
names, of which master equation is most appropriate. We will not cover this subject in depth
but rather refer the interested reader to the excellent book by van Kampen [49].
y
Or an integral, if the phase space is continuous.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

56

Kotelyanskii and Theodorou

the probability to be at 0 at step k
1, times the probability of transition

0 !  at the kth step. The second sum is very similar to the first; it equals
the total probability to leave . Again it is expressed as a sum of the
probabilities to go from  to any other point 0 at step k over all other
points 0 . How can we satisfy Eq. (76) with zero left hand side? Notice that
the right hand side is essentially a sum over different pairs (, 0 ). One
possibility is to require the contribution from each pair to be zero. This
leads to the following equation, which describes the detailed balance, also
called microreversibility condition:
M 0 W0 !  M W ! 0
W0 !  M 

W ! 0 M 0

77

Equation (77) requires that the flux between any two points in phase space
be the same in both forward and reverse directions. We have to emphasize
that Eq. (77) is not equivalent to (76). It provides a sufficient but not necessary condition for the latter to hold. In simple terms, this means that,
if detailed balance is satisfied, Eq. (76) is always satisfied, too. However,
it may also be possible to make Eq. (76) true with M and W not
satisfying (77).
It is easy to check that detailed balance holds for the Metropolis
algorithm described above with M() equal to the canonical ensemble
distribution in configuration space. In the Metropolis algorithm,
W( ! 0 ) ( ! 0 )min(1, M(0 )/M()), where ( ! 0 ) is the probability of attempting a move from  to 0 and Pacc( ! 0 )
min(1, M(0 )/M()) is the probability of accepting that move. In the
classical Metropolis algorithm, the attempt step is designed such that
( ! 0 ) (0 ! ). In the NVT ensemble, the acceptance probability is
min(1, exp[
(V 0
V)/(kBT)]), as described above. One can easily verify that
the detailed balance equation (77) is satisfied under these conditions.
In recent years it has been realized that the sampling of configuration
space may be more efficient if a nonsymmetric attempt probability matrix
( ! 0 ) is utilized. In that case, the correct acceptance criterion for
ensuring microscopic reversibility is



0 ! M 0
Pacc  ! 0 min 1,
 ! 0 M 

78

Furthermore, it may be convenient to conduct the moves by introducing

random changes not in the original set of coordinates  (e.g., Cartesian

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

57

coordinates) with respect to which the probability density M() is known,

but rather in a set of generalized coordinates ~ . The correct acceptance
criterion is then
~ 0 ! ~ M 0 J~ 0 ! 0
Pacc  ! 0 min 1,
~ ! ~ 0 M J~ ! 

!
79

where J~ !  is the Jacobian of transformation from the set of coordinates ~ to the set of coordinates ; J~ !  times a differential volume in
~ -space gives the corresponding volume in -space. The presence of the
Jacobian takes care of the fact that equal volumes in ~ -space may not
correspond to equal volumes in -space.
Since both MC and molecular dynamics methods are available for
predicting the equilibrium properties of material systems, and since
molecular dynamics can provide kinetic in addition to thermodynamic
properties, one may ask why MC is used at all. The answer is that, by
appropriate selection of the moves employed, MC can achieve a sampling of
configurations many orders of magnitude more efficient than that provided
by MD. This means that the MC simulation can converge to wellequilibrated averages in a small fraction of the time that would be required
by MD. Another advantage of MC is that it can readily be adapted to
simulations with variable numbers of particles (e.g., grand canonical and
Gibbs ensemble simulations) and is therefore very convenient for phase
equilibrium calculations. Such simulation schemes will be described in the
remainder of this book.

V. MOLECULAR DYNAMICS (MD)

At first glance Molecular Dynamics looks like a simplistic brute force
attempt to literally reproduce what we believe is happening in the real world.
Given a set of the initial velocities and coordinates, it tries to integrate the
equations of motion numerically. Following the trajectory, MD provides
information necessary to calculate various time correlation functions, the
frequency spectra, diffusion coefficients, viscosity, and other dynamic
properties. It also calculates thermodynamic properties (P, T, . . .) as time
averages at equilibrium.
In the simplest version of MD, the Newton (21) or Lagrange equations
(22) are integrated for a closed system, in which the volume, total energy,
and number of particles are conserved. This simulates the microcanonical,
NVE ensemble. Both kinetic and potential energies fluctuate in the
microcanonical ensemble but their sum remains constant.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

58

Kotelyanskii and Theodorou

For many problems, however, it is more convenient to keep the

temperature, pressure, or chemical potential constant, instead of the total
energy, volume, and number of particles. Generalizations of the molecular
dynamics technique to virtually any ensemble have been developed, and
they will be discussed in the following chapters. Of course, constant
temperature MD does not conserve the total system energy, allowing it to
fluctuate as is required at constant temperature. Similarly, volume is allowed
to fluctuate in constant pressure molecular dynamics. The trick is to make
these quantities fluctuate in a manner consistent with the probability
distribution of the desired ensemble.
There are two different ways of accomplishing this. The first is to
introduce some random perturbations to the system, which emulate its
interaction with the environment. For constant temperature, for instance,
this can be done by randomly [50] changing momenta by some random
increments drawn from the Gaussian distribution. This mimics collisions
between the molecules in the system and the virtual thermostat particles.
The choice of parameters for the Gaussian distribution of momentum
increments is determined by the required temperature.
Alternative techniques were proposed by Nose [51] and Hoover [52].
Interaction with an external thermostat is described by an extra degree of
freedom s, with associated momentum variable ps. Extra potential energy
V s ( f 1)kBT ln(s) and kinetic energy Ks 1=2 Qs s_2 are added. Here f is
the number of the degrees of freedom in the original system and Qs is a
thermal inertia parameter, with dimensions of energy times time squared.
The degree of freedom s is a scaling factor between real time and simulation
time, s dtsimulation/dtreal. A simulation time clock is introduced, in
addition to the real time clock, s being an instantaneous ratio of simulation
time to real time intervals; for a well-designed simulation, s should fluctuate
around 1. In the Nose version the extended system is conservative, and its
equations of motion can be derived following either a Lagrangian or a
Hamiltonian formalism:
Qs

mi v2i s
f 1kB T=s

r f=ms2
2_sr_ =s

80

The dots denote derivatives with respect to simulation time. The total
energy, including the kinetic and potential energy associated with the additional degree of freedom, is conserved. It can be shown that the original
system samples the canonical ensemble. Hoover [52] proposed a mathematically equivalent scheme in which the coupling parameter is a friction

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

59

factor, whose dynamics is governed by the difference between the current

kinetic energy and its average at the desired temperature T.
r_ p=m
p_ f
sp
1 X
s_
mi v2i =2
fkB T=2
Qs
i

!
81

In the Hoover scheme, the distinction between real time and simulation time
disappears. Again here, f is the number of degrees of freedom and Qs
is a parameter that determines how strongly the system interacts with the
thermostat. It is chosen by trial and error for a particular system. If it
is too large, the system dynamics is strongly perturbed and is dominated by
the coupling to the thermostat. If Qs is too small, the coupling is weak,
and equilibration takes too long.
For the above techniques it is possible to prove that they reproduce the
canonical ensemble distribution.
There are, however, some not so rigorous but still very practical methods,
based on velocity rescaling. One obvious and the most crude method to
simulate constant temperature in MD would be to just rescale all the
velocities, so that the kinetic energy corresponds to the desired temperature
according to K 1/2 Nf kBT, where Nf is the number of degrees of freedom.
This, of course, is a caricature of the real NVT ensemble, where kinetic
energy is allowed to fluctuate and is not always constant. Even though this
approach is totally unsuitable for production runs, it is quite practical and is
often used to equilibrate the model at a required temperature, starting from
an initial configuration. Some gentler variations on this theme have been
suggested in [53]. There, the kinetic energy is not kept constant at every step.
Velocities are rescaled by a factor chi, which is dependent on the difference
between the current and desired kinetic energy values. Deviations of the
kinetic energy from the desired value relax to that value with a predefined
characteristic time:




1=2
t T

1
 1
tT T
1 X
T
mi v2i
f =kB i

82

Here t is the time step and tT is a parameter that determines how strong the
perturbation is. Similar to the parameter Qs in the Nose and Hoover

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

60

Kotelyanskii and Theodorou

schemes, it has to be chosen by trial and error, so that the perturbation of

the original system dynamics is not too strong, but at the same time is strong
enough to equilibrate the system. This scheme is reminiscent of the method
of Hoover (81). The important difference between them is that here it is the
rescaling factor itself which depends on the difference between the instant
and desired temperatures, while in Eq. (81) it is its time derivative. Unlike
the stochastic and NoseHoover methods, velocity rescaling with the factor
described by Eq. (82) does not rigorously sample the canonical ensemble.
Nevertheless, this method is widely used in practice for equilibration purposes. It was also shown to provide results similar to Hoovers thermostat
(81) [54].
Every technique that involves velocity rescaling contains a parameter,
adjustments in which determine how strong is the perturbation to the
system dynamics. An inappropriate choice can lead to very interesting
phenomena, which unfortunately can be passed unnoticed if special care is
not taken. Rescaling particle velocities makes the average kinetic energy
correspond to the required temperature, but still relies on the systems
internal mechanisms to redistribute the kinetic energy between different
degrees of freedom to provide equipartitioning. Usually the energy is
redistributed due to the interactions between the particles and due to the
anharmonicity of the system. Such mechanisms are absent in the extreme
cases of the ideal gas and of the system of coupled harmonic oscillators,
discussed earlier in this chapter. As there are no interactions between the
particles in the former case and no interactions between the vibrational
modes in the latter, the energy of each degree of freedom remains at its
initial value. Imagine that some initial coordinates and velocities are
assigned to the particles in the ideal gas. If the kinetic energy happens to
be less than K0 3/2 NkBTset, the scaling factor for the velocities will be
greater than one, and hot faster particles will get more energy than
slower particles. Similarly, in the harmonic oscillator, hotter vibrational
modes will gain more energy than the colder ones. If errors in the time
integration scheme are such that the total energy decreases with time,*
together with velocity rescaling, they may lead to an unequal kinetic
energy distribution between different degrees of freedom. Of course, the
modeled systems are never ideal, and there is always interaction between
different degrees of freedom, which redistributes energy between them.
However, care must be taken as, if these channels are not fast enough,

*This is common for many stable integration algorithms. It is a well known fact that when the
most popular Verlet algorithm is used to numerically integrate the equations of motion, in very
long MD runs the system gets colder, due to the integration errors.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

61

undesired effects may occur [55]. One of the most common is the so-called
flying ice cube effect. The translational motion of the system as a whole
is often decoupled from other degrees of freedom, as potential energy does
not depend on it. There is no way for energy exchange between this degree
of freedom and the rest. Velocity rescaling can lead to the energy being
pumped into the overall translational motion, while being taken away
from other degrees of freedom. Nominally the kinetic energy of the system
is 1/2NkBT, but most of it is in the translational motion of the systems
center of mass. With no kinetic energy left in the other modes, the system
looks like a frozen cube flying in space at a rather high velocity. This can
be taken care of by correcting for the center of mass motion during the
run. If the initial velocities are chosen at random, there is always some
nonzero center of mass velocity component, which has to be subtracted in
the beginning of the run.* Overall center of mass motion can also
accumulate in very long MD runs due to the rounding errors.
With this said we see that, when using any one of the thermostats
described by Eqs. (80), (81), and (82), the time constant chosen should not
be too small to allow the energy to redistribute equally between the different
degrees of freedom after it is artificially pumped into the system or taken
from it by velocity rescaling.

VI. BROWNIAN DYNAMICS

The Brownian Dynamics technique is based on the Langevin equation of
motion, originally proposed to describe the Brownian motion of a heavy, for
instance colloidal, particle in a solvent. The Brownian particle is subjected to
a random force from many collisions with solvent molecules. In addition,
the solvent exerts a friction force which is assumed to be proportional to
the particles velocity.
p_
p f r t

83

Here  is a friction coefficient and fr is the random force, which is also

assumed to exhibit no directional preference: hf r ti 0, to be uncorrelated
with the particles momentum: hf r t  p0i 0 for all t, and to be uncorrelated with previous and future values of itself, hfr(t)fr(t0 )i Ro(t
t0 ).
The Langevin equation (83) is a stochastic differential equation. It can be
derived formally from the full set of dynamical equations describing the

*Unless it aects the goal of the simulation, of course.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

62

Kotelyanskii and Theodorou

particle and solvent system by projecting out the motion of the fast solvent
particles [56,57].
Multiplying both sides of Eq. (83) by p(0) and averaging over all
stochastic trajectories consistent with the same initial conditions, taking into
account the delta-correlation of the random force, we obtain:
d
hp0  pti
hp0  pti
dt
hp0  pti hp2 i exp
t

84

We see that the velocity autocorrelation function of a free Brownian particle

falls exponentially with time. The time required for an arbitrary initial
velocity distribution of the particle to settle down to the Maxwell
Boltzmann form corresponding to the temperature of the bath (thermalization of velocities) is on the order of 
1.
For the diffusion coefficient we calculate the mean square displacement
of the particle:
hrt
r02 i

1
m2

dt0
0

hp2 i
m2
2

hp i
2 2
m
2

dt00 hpt00  pt0 i

0

dt0
0

t0

dt00 exp
t0
t00

0

t
0

dt exp
t
0

t0

expt00 dt00

hp2 i t
hp2 i 1


2
1
exp
t
m2 
m2  2

85

In the limit of long times (t  1), invoking the equipartition theorem for
kinetic energy,
lim hrt
r02 i 2

t!1

hp2 i t
kB T t
6
6Dt
m2 
m 

86

[compare Eq. (68)]. So, the diffusion coefficient of the particle D is related
to the friction constant  as
D

kB T
m

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

87

Background

63

Equation (87) is an Einstein relation for Brownian motion. There is a connection between the friction factor  and the quantity Ro, measuring the
mean squared magnitude of the random force. Solving Eq. (83) just like a
deterministic ordinary differential equation gives
Z

expt0 f r t0 dt0

pt p0 exp
t exp
t

88

Squaring and taking the ensemble average,

2

hp ti hp 0i exp
2t 2 exp
2t


expt0 f r t0  p0 dt0

dt0

exp
2t

Using the properties

we obtain


dt00 expt0 t00 f r t0  f r t00

89




f r t0  p0 0 and f r t0  f r t00 Ro t0
t00 ,

hp2 ti hp2 0i exp
2t

Ro
1
exp
2t
2

90

As t ! 1, the particle will become thermally equilibrated with the bath and
so, by equipartition, limt!1 hp2 i 3mkB T. We have already used
 this in
deriving (86). On the other hand, Eq. (90) gives limt!1 p2 Ro =2.
Equating,
Ro 6mkB Thf r t  f r t0 i 6mkB Tt
t0

91

This is the fluctuation-dissipation theorem. It requires that the mean energy

given to the particle by the random force be equal to the mean energy taken
out of the particle by the frictional forces with its environment. For a more
rigorous formulation, the reader is referred to [58].
If the particle is in an external potential, or if there are many interacting
particles, another deterministic force term describing this interaction is
added to the right hand side of the Langevin equation:
p_
p
rV f r t
p r_ =m

92

Let us try to derive an equation for the evolution in time of the phase
space probability density (p, r, t) for a random process generated by

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

64

Kotelyanskii and Theodorou

numerical integration of the Langevin Eq. (92) with some time step t. As
in the Monte Carlo master Eq. (76), we should write down the balance of
the incoming and outgoing fluxes for the phase space point  {p, r} after
the time increment t, taking into account that the coordinate increment
equals r (p/m)t [59] and the momentum increment equals p
 p, with 
 p the momentum increment due to the random

p
rVt 
force. P(p) is the probability density of the momentum increment p
Z
p, r, t t  p, r, t
d 3 p  p, r, t Pp
Z
d 3 p p
p, r
r, t Pp

93

The first two terms on the right handR side cancel each other, as (p, r, t) can
be pulled out from the integral and d3pP(p) 1. Expanding  around
{p, r} under the integral of the third term on the right hand side,
Z
p, r, t t

d 3 pPpp, r, t

X
i

X @ Z
@pi
X

pi @
m @ri

d 3 pPp

d 3 pPppi

1
2

i, j

@2 
@pi @pj

d 3 pPppi pj

Z
@ 2  pi pj
d 3 pPp
@ri @rj m2

Z
@ 2  pi
3
t
d pPppj
@ri @pj m

t2

94

The indices i, j run over all coordinate directions. Taking into account the
 pi 0 and the lack
following relationships, which are a consequence of h
of correlation between the random force and momentum or position:
Z
Z

d 3 pPppi t
pi
ri V

 p2 iij
d 3 pPppi pj t2
pi
ri V
pj
rj V 1=3h
95

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

65

we obtain
2
p, r, t t
p, r, t t4

X pi @
i

m @ri

X
i

3
X @
@
ri V

pi 5
@pi
@pi
i

1 X @2   2
h p i
6 i @p2i

X pi @ @
1
@ pj
pj
rj V
t2
2
@rj m
m @ri @pj
ij

96
 p2 i 6tmkB T fluctuationIn the limit of t ! 0, using the fact that h
dissipation theorem, compare Eq. (91), Eq. (96) reduces to:
X pi @ X
X @2 
@p, r, t
@ X @

ri V

pi  mkB T
@t
@pi
@pi
m @ri
@p2i
i
i
i
i
97
which is a FokkerPlanck equation [59]. By direct substitution one can
check that the Boltzmann distribution


2

p
V =kB T
p, r, t Const: exp

2m
is its stationary solution. This means that a Brownian dynamics method
solving the Langevin equation reproduces the canonical ensemble. Notice,
however, that such numerical schemes are approximate, with accuracy on
the order of (t2). This error is called discretization error. To obtain exact
canonical ensemble averages from a Brownian dynamics simulation it is
necessary to run simulations with different time steps and to extrapolate
the results to t ! 0 [58].
Another interesting and important fact is that it is really not necessary
for the random momentum increments to be drawn from a Gaussian
distribution. Any distribution with zero mean and appropriate mean square
 p2 i 2mkB Tt along each coordinate
deviation satisfying the condition h
i
direction will be appropriate.
Brownian dynamics was first introduced by Ermak et al. [60,61] and is
commonly used to simulate polymers, ions, or colloidal particles in solution.
The monomer dynamics is modeled by the Langevin equation, with the
solvent modeled as the medium providing the frictional and random forces.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

66

Kotelyanskii and Theodorou

If the evolution of the system over times t  1/ is of interest

(thermalization of velocities in each time stephigh friction limit), the
friction term is dominant and the inertia term on the left hand side of the
Langevin equation can be neglected. Then the equation becomes first order:
r_

D

rV f r t
kB T

98

The configuration-space probability distribution in this case is defined by

the Smoluchowski equation, which can be derived in the same way as the
FokkerPlanck equation above.
@
D
r, t

rrVr, t Dr2 r, t
@t
kB T

99

This equation describes diffusion in the external potential V(r). If the potential is constant (or zero), it reduces to the usual diffusion equation.

VII.

TECHNIQUES FOR THE ANALYSIS AND

SIMULATION OF INFREQUENT EVENTS

Molecular dynamics is a useful technique for probing transport and

relaxation processes in materials, provided the relevant time correlation
functions decay appreciably over times shorter than 100 ns. Unfortunately,
many dynamical processes in real-life materials, especially polymers, are
governed by time scales appreciably longer than this. Techniques alternative
to brute force MD must be developed to predict the kinetics of such
processes from molecular constitution.
Many important dynamical processes in materials occur as successions of
infrequent events. The material system spends most of its time confined
within relatively small regions or states in its configuration space, which
are surrounded by high (relative to kBT) energy barriers. Only infrequently
does the system jump from one state to another through a fluctuation that
allows it to overcome a barrier. Once initiated, the jump process occurs quite
quickly (on a time scale that can be followed by MD). The mean waiting
time between jumps, however, is very long. A MD, or even BD, simulation
would exhaust itself tracking the relatively uninteresting motion of the
system as long as it is confined in a few states, but would be unable to
sample a sufficient number of the jumps. Examples of infrequent event
processes include conformational transitions of chains in solution and in the
melt, diffusion of small molecules in glassy polymers, structural relaxation

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

67

in the glassy state, the formation of stable nuclei at the onset of

crystallization, and the conversion from reactants to products in a chemical
reaction.
By shifting attention towards the energy barriers that must be overcome
for a transition between states to occur, special techniques for the analysis
and simulation of infrequent events manage to calculate a rate constant for
the transition utilizing the machinery of common MD and MC simulation.
These techniques are based on the principles of Transition-State Theory.
Envision a system whose free energy as a function of a generalized
coordinate q looks as shown in Fig. 8. Defining an appropriate reaction
coordinate in an Nf-dimensional configuration space is in itself an
interesting problem, which we will address briefly below. The free energy
plotted in Fig. 8 incorporates the effects of thermal fluctuations in directions
orthogonal to the reaction coordinate. The barrier at qz defines two states,
centered around local free energy minima: A state A with q<qz and a state
B with q>qz. Macroscopically, if the barrier is high relative to kBT, there
will be a wide range of times that are long relative to the correlation times

FIG. 8 Free energy profile as a function of the reaction coordinate q showing two
states separated by a barrier. The lines in the upper part of the figure show different
ways in which dynamical trajectories can cross the barrier region.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

68

Kotelyanskii and Theodorou

governing the motion of the system in each of the states, but still very short
compared to the residence time of the system in either state (time scale
separation). Over this range of times, the probabilities of occupancy pA, pB
of the states evolve according to the master equations:
dpA

kA!B pA kB!A pB
dt
dpB

kB!A pB kA!B pA
dt

100

A microscopic expression for the rate constant kA ! B is [62]

kA!B t

q_ 0qz
q0
qt
qz



qz
q


101

where  and
stand for a Dirac delta function and a Heaviside function,
respectively. The numerator in Eq. (101) is an average rate of motion along
the reaction coordinate taken over all dynamical trajectories that cross the
barrier at time 0 and end up in state B after time t. Trajectories crossing the
barrier in the direction from A to B at time 0, such as (a) and (c) in Fig. 8,
contribute positively to the average in the numerator, while trajectories
crossing in the opposite direction, such as (b), contribute negatively. The
denominator is simply the equilibrium probability of occupancy of state A.
When time scale separation holds, the kA ! B calculated through Eq. (101)
turns out to exhibit a plateau value, independent of t, over a wide range of
(not too short) times. Eyrings Transition-State Theory (TST) rests on an
approximation: It assumes that any trajectory crossing the barrier in the
direction from A to B will eventually lead the system to state B. That is, it
ignores barrier recrossing trajectories along which the system ultimately
thermalizes in the state in which it originated [such as trajectory (b) in
Fig. 8]. In more formal terms, TST replaces q(t) in the numerator of
Eq. (101) with q(0), i.e., with its limit for t ! 0, at which the system is
bound to be in the state towards which q_ 0 is pointing. The TST estimate
of the rate constant thus reduces to the ratio of two equilibrium ensemble
averages:

kTST
A!B

q_ 0
q0
qz qz
q0



qz
q


q_
q_ qz
q0



qz
q

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

102

Background

69

kTST
A!B emerges essentially as an average velocity in the direction from A to B
along the reaction coordinate, which can readily be computed in most cases
from the momentum-space distribution of the system, times the ratio of
probabilities of being at the barrier and of residing in the origin state A.
The latter ratio can be calculated by MC or constraint MD methods
designed to sample the barrier region.
It is customary to express the rate constant kA ! B of Eq. (101) in terms
of its TST estimate, kTST
A!B , as
kA!B t kTST
A!B t

103

where the dynamical correction factor or transmission coefficient (t)

is given by:


q_ 0 qz
q0
qt
qz

t
q_ 0 qz
q0
q0
qz

104

In a two-state system with time scale separation, (t) quickly settles to a

plateau value within a time comparable to the correlation time of molecular
velocities and much shorter than 1/kA ! B. This value is less than unity, for
TST overestimates the rate constant by neglecting dynamical recrossing
events. k(t) can be calculated straightforwardly by generating a number of
MD trajectories initiated on the barrier with an equilibrium distribution
of velocities resulting in a q_ value pointing from A to B. Each trajectory
quickly thermalizes in state A or state B. The fraction of trajectories which
thermalize in state B provides the plateau value of . A variant of the
method is to initiate the trajectories with random velocities, integrate
them both forward and backward in time, and count the fraction of them
that are effective in bringing about an A ! B or B ! A transition [63,64].
The possibility of calculating kTST
A!B and  with modest computational
means offers a way out of the problem of long times in the case of dynamical
processes occurring as sequences of infrequent events. Note, however, that,
to apply TST-based approaches, one must have a good idea of the states and
reaction coordinates, i.e., one must already know something about the (free)
energy landscape of the system.
To illustrate these ideas with a simple example, consider a onedimensional particle of mass m moving in an external field V(q) of the
form depicted by the curve in Fig. 8. In this case, the ordinate is merely
potential energy, rather than free energy. The field V(q) could be generated
by the atoms of a medium, in which the particle is diffusing. We will
implicitly assume that there is a mechanism for weak energy exchange

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

70

Kotelyanskii and Theodorou

between the particle and the medium (heat reservoir), as a result of which
the particle velocity q_ is distributed according to the requirements of
equilibrium at temperature T. By virtue of the separability of momentum
and configuration-space distributions, and realizing that positive and
negative values of q_ are equally probable, we can recast Eq. (102) for this
problem as

kTST
A!B


 
1 q_  qz
q



2
qz
q

kB T

2m

1=2



exp
Vqz =kB T
R
q2A exp
Vq=kB T dq

105

The integral in the denominator of Eq. (105) is taken over the entire state A,
to the left of qz. At low temperatures, the major contribution to this integral
comes from the immediate vicinity of the energy minimum qA, where the
function V(q) can be approximated by its Taylor expansion around qA
truncated at the second order term [compare Eq. (35)]. Setting
kA @2 V=@q2 jqA , we obtain
Z

Z 1

Vq
VqA
kA q
qA 2
exp

exp

dq exp

dq
kB T
kB T
2kB T
q2A

1

2kB T
kA

1=2

VqA
exp

kB T

106

Using Eq. (106) in Eq. (105), we are led to the harmonic approximation for
HA
:
the transition rate constant, kTST,
A!B



1 kA 1=2
Vqz
VqA
exp

2 m
kB T

Vqz
VqA
A
 exp

kB T

, HA

kTST
A!B

107

The harmonic approximation to the rate constant emerges as a product of

the natural frequency A of oscillation of the system within the well of the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

71

origin state A times the Boltzmann factor of the barrier height measured
from the minimum of state A.
An ingenious analysis of the transition rate in systems such as the one
of Fig. 8, which takes explicitly into account energy exchange with the
degrees of freedom not participating in the reaction coordinate, was
carried out by Kramers [59]. In Kramers analysis, these degrees of
freedom are envisioned as comprising a Brownian bath which exerts
Langevin and frictional forces on the degree of freedom q. The motion of
q is described by a one-dimensional Langevin equation of the form (92).
This, of course, presupposes that the motion of the bath degrees of
freedom is fast relative to the evolution of q. The potential (or, more
generally, free energy) function V(q) is approximated by its Taylor
expansion truncated at the second order term both around the bottom of
the well of the origin state (qA) and around the top of the barrier qz.
Symbolizing by  the friction coefficient, as in Eq. (92), by A the
curvature of the potential at qA, as above, and by kz>0 the opposite of
the curvature of the potential at the barrier, kz
@2 V=@q2 jqz , Kramers
result for the rate constant is:
kA!B


p

Vqz
VqA
2
 1
 exp


kB T
A

108

p
where A 1=2 A =m [natural frequency of oscillation
p in the harmonic
region of the well, as in Eq. (107), and  =2 m=z is a dimensionless
parameter which increases with the friction coefficient and decreases with
the curvature of the potential at the barrier. When the coupling between
the reaction coordinate and the other degrees of freedom is weak,  1
and the Kramers result, Eq. (108), reduces to the harmonic approxiHA
. Note that the transmission coefficient
mation TST estimate, kTST,
pA!B

TST
 kA!B =kA!B 2 1
 predicted by Kramers theory is always
smaller than 1,
as it should be. In the case   1 (strong friction limit),
p
the quantity 2 1
 is well approximated by 1/2 and the Kramers
result becomes
kA!B

A

r



kz
Vqz
VqA
exp

kB T
m

109

In this limit (strong coupling between q and the bath and/or flat barrier)
kA ! B can be orders of magnitude lower than the TST estimate.
The exact location where qz is chosen affects the values of kTST
A!B and ,
but not their product kA ! B. Nevertheless, in practice every effort should be

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

72

Kotelyanskii and Theodorou

made to define the reaction path so that the transition-state theory estimate
of the rate constant is minimized and the dynamical correction factor is
maximized.
How does one define the reaction path and the transition state given the
potential energy function of a system with many degrees of freedom? We
will assume here that the configuration is described in terms of the Nf massweighted Cartesian coordinates xi m1=2
i ri for all atoms. We will use the
symbol x to refer to all these degrees of freedom collectively. States, such
as A and B in the examples above, are constructed around local minima of
the potential energy, at which the gradient vector gx rVxx is 0 and
the Hessian matrix Hx @2 V=@x@xT jx is positive definite. Between two
neighboring local minima xA and xB there will be at least one saddle point
xz, at which g(xz) 0 and the Hessian H(xz) has one negative eigenvalue
with associated unit eigenvector nz. This saddle point is the highest energy
point on the lowest energy passage between xA and xB; it is usually a good
choice for the transition state z between A and B. The reaction path between
A and B, along which the reaction coordinate is measured, is a line in
Nf-dimensional space connecting xA and xB. To construct it, one initiates
two steepest descent trajectories at xz, one in the direction nz and the
other in the direction
nz. Each such trajectory consists of small steps
dx
(g(x)/|g(x)|) dq parallel to the direction of the local gradient vector
and terminates in one of the two minima connected through z (see Fig. 9).
The dividing surface between states A and B is defined as an (Nf
1)dimensional hypersurface with equation C(x) 0, with the following
properties: (a) it passes through the saddle point, i.e., C(xz) 0; (b) at the
saddle point it is normal to the eigenvector nz corresponding to the negative
eigenvalue of the Hessian (and, therefore, to the reaction path); (c) at all
points other than xz, it is tangent to the gradient vector. Conditions (b) and
(c) can be expressed mathematically as:

rCx 


rCx

nz

110

xz

x 6 xz

rCx  gx 0,

111

With this definition of the dividing surface, the TST estimate of the transition rate constant becomes [65]
kTST
A!B

d Nf x
x2A

d Nf p nx  p Cx jrCxjNVT x, p
nxp>0

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

112

Background

73

FIG. 9 Reaction paths and dividing surfaces in a two-dimensional configuration

space. Hypersurfaces (lines in this case) of constant V are shown as thin closed
curves. There are three minima with corresponding states A, B, and C, labeled from
left to right. (Top) Reaction paths from A to B and from B to C are drawn as a thick
line. Note that the direction of the reaction path may curve away from the straight
line connecting two minima. (Bottom) The three minima and the two saddle points
between them are shown as bold dots. The dividing surface (line) between states A
and B is drawn as a solid line of medium thickness running from bottom left to top
right. The hyperplane (line) approximation to the dividing surface is shown as a
broken straight line. The thick curve traces an approximation to the dividing surface
constructed through a local criterion. (After Ref. [65].)

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

74

Kotelyanskii and Theodorou

where p is the vector of mass-weighted momenta conjugate to x, NVT(x) is

the canonical ensemble probability density in phase space, and the Dirac
delta function selects configurations on the dividing surface. Upon performing all momentum-space integrations, one obtains from Eq. (112)

kTST
A!B






R

kB T 1=2 x2A d Nf x Cx rCx exp
Vx=kB T
R

Nf
2
x2A d x exp
Vx=kB T

113

As in Eq. (105), kTST

A!B emerges as a velocity associated with moving
across the dividing surface times a ratio of configurational integrals, one
taken over the dividing surface and the other taken over the origin state.
[To reconcile the dimensions in Eq. (105), remember that the coordinates x
are mass-weighted.]
In practice, the definition of the dividing surface C(x) 0 that we
introduced above may be difficult to use in systems with large Nf, because it
is nonlocal. [If an analytical expression for C(x) is not available, one cannot
judge whether a point x belongs on the dividing surface, unless one initiates
a steepest descent trajectory at that point and sees where it ends up].
Approximate local criteria have been devised [65]. When the pass between
states A and B is narrow, most of the contribution to the integral in the
numerator of Eq. (113) comes from the immediate vicinity of xz, and the
dividing surface can be approximated by a hyperplane tangent to it at xz.
That is, one may use for the dividing surface the approximate equation
Cx nz  x
xz 0

114

The ratio of configurational integrals appearing in Eq. (113) can then be

computed through a free energy perturbation technique which involves
calculating free energy differences for the system being confined on a
succession of hyperplanes normal to the reaction path [66].
If, in addition, a harmonic approximation is invoked for V(x), Eq. (113)
leads to:
QNf A

Vxz
VxA
i1 i

exp


kTST
QNf z
A!B
kB T
i2 i

115

where the eigenfrequencies A

i are the square roots of the eigenvalues of the
Hessian matrix of second derivatives of V with respect to the mass-weighted

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

75

coordinates x in the origin state A and, similarly, zi are the square roots
of the Nf
1 positive eigenvalues of the Hessian matrix at the saddle
point. Equation (115), first derived by Vineyard [67], is a useful generalz
ization of Eq. (107). In case the eigenfrequencies A
i or i are too high for
the corresponding vibrational motion to be describable satisfactorily by
means of classical statistical mechanics, a more appropriate form of
Eq. (115) is

kTST
A!B



QNf 

kB T i1 1
exp
hA
Vxz
VxA
i =kB T
h


i
exp


h QNf 1
exp
hz =k T
kB T
i2

116

(In the latter equation, classical partition functions for the harmonic
oscillators representing the modes have been replaced with quantum
mechanical ones. Zero point energy contributions are considered as being
part of the Vs.) The latter equation is a special case of the more general TST
expression

kTST
A!B


kB T
Gz
GA
exp

h
kB T

117

where Gz is the Gibbs energy of the system confined to the dividing surface
and GA is the Gibbs energy of the system allowed to sample the entire origin
state.
The above discussion concerned ways of calculating the rate constant
for a transition between two states, A and B. The evolution of real
material systems (e.g., by diffusion or relaxation) often involves long
sequences of transitions between different states. Once the rate constants
are known between all pairs in a network of connected states, ensembles of
dynamical trajectories for the system can readily be generated by Kinetic
Monte Carlo simulation [64]. In such a simulation, the times for the next
transition to occur are chosen from the exponential distribution of waiting
times governing a Poisson process. If the rate constants are small, these
times are long; the evolution of the system can be tracked over time
periods many orders of magnitude longer than can be accessed by MD.
Thus, the long-time problem of MD is eliminated. The challenge in such
an approach is to identify the states and saddle points between them and
to include all relevant degrees of freedom in the transition rate constant
calculations.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

76

Kotelyanskii and Theodorou

VIII. SIMULATING INFINITE SYSTEMS, PERIODIC

BOUNDARY CONDITIONS
Even the most powerful supercomputers available today can only handle up
to about a million atoms. If we are interested in the properties of single
molecules or small drops or clusters of diameter 100 A or less, this is not a
problem. Most of the time, however, one is interested in the properties of
bulk materials.
When the simulated system is bounded by walls or by free surfaces, a
substantial fraction of the atoms is located next to the surface, in an
environment different from the bulk. Interactions with about half of the
neighbors are replaced by interactions with the bounding wall, or are just
absent. The effect of the surface is roughly proportional to the fraction of
atoms in its vicinity, relative to the total number of atoms in the model. An
obvious way to reduce the effect of the surface is to increase the system size,
which is however limited by computational resources. This limitation was a
lot more severe in the early days, when simulations were run on mainframes
capable of handling only about a hundred particles. A very good solution to
the problem was found back then and is still in wide use today.
System size effects are substantially reduced by using periodic boundary
conditions [68,43,69]. A primary box, containing the particles, is
replicated in space in all directions to form an infinite lattice of boxes
with lattice parameters equal to the box lengths in the corresponding
directions, Lx, Ly, and Lz. Each particle with coordinates ri in the primary
simulation box has an infinite number of periodic images with coordinates
ri nxLxex nyLyey nzLzez, where (nx, ny, nz) is a set of three integers
ranging from minus to plus infinity and ei are the unit vectors along the
three coordinate directions.
Thus, a bulk material is simulated by an infinite periodic array of the
simulation cells. Interactions are calculated taking into account the periodic
images of particles. When a particle moves, all its periodic images move
by exactly the same displacement. A two-dimensional example of periodic
boundary conditions is shown in Fig. 10.
Unfortunately, periodic boundary conditions do not eliminate size effects
completely. Imposing periodicity on the system cuts off its fluctuation
spectrum. In the solid state the longest wavelength with which density
fluctuations (sound) can propagate equals the box size. Density or
composition fluctuations with characteristic lengths larger than the box
size are impossible. Such long-range fluctuations play an important role in
phase transitions. Thus, accurate simulations of phase transitions or phase
coexistence, especially near critical points, require large system sizes and the
use of finite size scaling techniques [70].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

FIG. 10

77

Two-dimensional system with periodic boundary conditions.

If interparticle interactions are long range, the interactions between

a particle and its own periodic images are substantial, and the symmetry of
the lattice, artificially imposed by the periodic boundary conditions, affects
properties of the otherwise isotropic system.
It is absolutely necessary to check for size effects on the simulation results
by performing simulations with different model system sizes.
Periodic boundary conditions can be implemented in two different ways.
One can keep track of the particles in the central (or primary) box, switching
attention to the particles periodic image entering the primary box from the
opposite side whenever the particle crosses the box boundary. Alternatively,
one can follow the coordinates of particles initially located inside the
primary box and calculate coordinates of their images inside the primary
box when necessary. The computational cost is about the same in both
cases. The latter choice is more convenient when calculating diffusion
coefficients, as the particle trajectory remains unperturbed, while one box
length is added or subtracted to a particle coordinate when switching
between the periodic images.

A. Calculating Energy and Forces with Periodic

Boundary Conditions
The core part of any molecular dynamics, Brownian dynamics, or Monte
Carlo simulation is the calculation of the interaction energy and forces.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

78

Kotelyanskii and Theodorou

In general, the energy in the periodic system encompasses interactions

between all periodic images, and thus constitutes an infinite sum over all
periodic images. The potential energy contribution from pairwise interactions u (ri
rj) is given by:
V

XX X

uri
rj nx Lx ex ny Ly ey nz Lz ex

118

j>i nx , ny , nz

nx 60, ny 60, nz 60

unx Lx ex ny Ly ey nz Lz ez

The sum over pairs of particles includes summation over the periodic images
of the second particle. The second term describes the contribution from a
particle interacting with its own periodic images.
If the interaction potential u is long range, all terms in the infinite sum
should be taken into account. The brute force approach here is to
truncate the summation at some large enough values of nx, ny, nz. Efficient
ways to do this for Coulombic interactionsthe Ewald summation, fast
multipole, and particlemesh methodswill be described in the following
chapters.
A much simpler technique is used for Lennard-Jones potentials or other
short-range interactions. The Lennard-Jones potential, Eq. (6), drops down
to only
0.016  at rc 2.5. If we just cut the potential off at this distance,
assuming that u 0 at r>rc, the error we make is less than 2%. We will see
later, how it can be accounted for. If such a cutoff is used, the sums in Eq.
(118) are limited only to the pairs of images that are closer to each other
than the cutoff distance rc. If rc is smaller than one half of the smallest of Lx,
Ly, or Lz, only the closest of images are taken into account in the first sum,
and there is no second summation over the self-images. This approximation
is called the minimum image convention. The intermolecular potential
energy is then given by:
V

XX
i

uri
rj n0x Lx ex n0y Ly ey n0z Lz ez

119

j>i

In this equation, n0x n0y n0z stands for the set of integers that correspond to the
pair of closest images of particles i and j. Notice that the circle centered
at particle A (Fig. 10) with radius rc may include a periodic image BN of
particle B, but not B itself.
The following fragment of FORTRAN code illustrates the implementation of the minimum image convention. It calculates the x, y, and z

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

79

components of the separation vector ri
rj, stored in the variables dx, dy,
dz, given that particle coordinates are stored in the arrays x, y, z, and box
sizes Lx, Ly, Lz are stored in the variables boxx, boxy, boxz.
..............................
invbx1.d0/boxx
invby1.d0/boxy
invbz1.d0/boxz
...............................
dx x(i)
x(j)
dx dxboxx*anint(dx*invbx)
dy y(i)
y(j)
dy dy
boxy*anint(dy*invby)
dz z(i)
z(j)
dz dz
boxz*anint(dz*invbz)
............................

This piece of code is used throughout the modeling program with the
periodic boundary conditions whenever the distance between two particles is
needed: in the energy and force calculation, in the calculation of the pair
correlation function, etc. The code uses the standard Fortran function*
anint(a), which returns the nearest integer to the real or double
precision number a. Note that anint(0.51)1, anint(0.49)0,
anint(
0.51)
1, anint(
0.49)0.
To fold all the particles inside the primary simulation cube,

Lx/2<xi<Lx/2,
Ly/2<yi<Ly/2,
Lz/2<zi<Lz/2, we can use similar
code:
..................................
invbx1.d0/boxx
invby1.d0/boxy
invbz1.d0/boxz
...............................
x(i) x(i)
boxx*anint(x(i)*invbx)
y(i) y(i)
boxy*anint(y(i)*invby)
z(i) z(i)
boxz*anint(z(i)*invbz)
............................

When a chemical bond crosses the box boundary, its periodic image
reenters the box from the opposite side, as shown in Fig. 10. Unlike
intermolecular interactions, the intramolecular interactions should be
calculated between atoms of the same image of a molecule and not between

*anint(a) is also available in some C compiler libraries.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

80

Kotelyanskii and Theodorou

their periodic images. Nevertheless, it is also safe to use the minimum image
convention for the intramolecular interactions. As chemical bonds are
always shorter than half the box length, the convention will automatically
provide the right separation vectors, no matter whether the particle
coordinates are always folded back in the primary box, or whether the
particle trajectory in space is preserved.
When the potential energy is cut off at distance rc, interactions between
particles separated by a distance larger than rc are neglected. If these
interactions are described by spherically symmetric pairwise potentials, the
energy can be expressed in terms of the pair distribution function, as in
Eq. (63). If the rc is large enough, such that we can assume that there are
no structural correlations beyond this distance (gr 1), then the
long-distance part of the integration can be replaced by the integration of
the potential for r>rc:
Z

grur4r2 dr

V V sum N

120

r
Z c1

ur4r2 dr
Z1
1
dur
4r2 dr
PV Psum V
N
gr r
6
dr
rc
Z1
1
dur
r4r2 dr
Psum V
N
r
6
dr
rc
V sum N

rc

V sum and Psum are the values obtained from explicit summation over the
pairs within the cutoff radius rc. Similar equations can be derived for any
other property obtained via the summation over pairs of particles. The
integrals on the right hand sides of the above equations can easily be calculated for any pair potential u(r).*
Setting g(r) 1 corresponds to replacing the material at long distances
from the particle by a uniform background, with the average interaction
properties. Thus, on the one hand one would like the cutoff radius to be not
very large, to reduce the amount of calculations necessary for obtaining
the energy, on the other hand rc should be chosen large enough, so that
structural correlations can be negligible at distances r>rc. The latter
assumption has to be checked by calculating g(r).
Even though the interaction potential value u(rc) is quite small, the effect
of the cutoff can be quite substantial. For instance, using the Lennard-Jones

*They might be nonconverging for the long-range potentials or in fewer than three dimensions.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

81

(6-12) potential with cutoff rc 2.5 in the simple monatomic model of the
noble gas lowers the critical temperature by about 30% [54] compared to the
value obtained with a much longer cutoff or with the long-range corrections
given by Eq. (120).
If the interaction is not pairwise, similar expressions can be obtained using
the corresponding multiparticle correlation functions. If the interaction
potential u depends on the direction of the separation vector r as well as on
its length r, a multidimensional integration over the components of r is to be
performed instead of the spherical averaging leading to integration over
r |r|.

IX. ERRORS IN SIMULATION RESULTS

Although simulations are often associated with theory, they have a lot in
common with experiments. In particular, simulations are subject to errors.
There are systematic and statistical errors. The sources of the systematic
errors are system size effects, inaccuracy of the interaction potentials, poor
equilibration, deficiency in the random number generator, etc. Systematic
errors are estimated by performing simulations with different system sizes,
using different random number generators, and varying the interaction
potentials, equilibration conditions, and starting configurations.
Properties calculated in computer simulations are obtained as statistical
averages, and therefore are inherently subject to statistical errors. The
average is calculated as a sum over many steps. According to the central
limit theorem of probability theory, the average value obtained in the
simulation is sampled from a Gaussian distribution, with average at the true
mean. The variance of this distribution provides an estimate of the
difference between the average estimate obtained in the simulation and
the true mean value. According to statistics, given N uncorrelated
PNvalues of
Ai obtained during the simulation,
with
average
hAi

1=N
run
i1 Ai and
P
2
variance hA2 irun 1=N N
i1 Ai
hAirun , the variance of the estimated
average hAi,  2(hAirun), equals h(A)2irun)/N. The probability to observe a
certain value of hAirun in a sample of N uncorrelated configurations is
described by a Gaussian distribution with average at the true mean and
variance h(A)2irun/N. The expected error in the simulation estimate of A is
1=2
.
thus hA2 i1=2
run =N
The configurations obtained in successive steps of a molecular dynamics,
Monte Carlo, or Brownian dynamics simulation are usually not very
different from each other. Hence, one should expect the energy, virial, or
any other function of the microscopic state obtained from the successive
configurations to be correlated. These correlations disappear after some

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

82

Kotelyanskii and Theodorou

time, or number of steps. The correlation time depends on the modeled

system and on the quantity evaluated. For instance, in solids or liquids,
where collisions between particles are very frequent, velocities usually have
much shorter correlation times than any structural characteristics. The
correlation time is usually defined
R 1 as an integral of the normalized time
correlation function [43]: 0 cAA t. Values separated by at least
one correlation time must be used to estimate the error in the average. If
A-values are saved every time interval t, each kth value, with k /t,
should be used. Thus, instead of N we have only N/k uncorrelated values,
and b N/k must replace N in the expression for the expected error in the
average. A common practice is to take averages over blocks of data of
length k, hAik
i , and use them as b independent values to estimate the
variance (squared expected error in the simulation estimate) [43]:
hhAirun 2 i

b
kX
hAiik
hAirun 2
N i1

121

The correct block size k that ensures statistical independence of the hAik
i
can be determined through systematic analysis for different k [43].
To estimate errors in structural or other properties, expressed not just
as numbers but as functions of either distance (e.g., g(r)), or time (e.g.,
time correlation functions), a similar analysis should be carried out for
different values of the argument. For instance, when the time correlation
function is calculated, much better statistics is obtained for short times,
as there are much more data available. The time correlation function
values for times comparable to the length of the simulation run are
obtained with larger errors, as the average is taken over only a few pairs
of values.
When the property calculated is a single particle property, such as the
velocity autocorrelation function, or the particles mean squared displacement, averaging over the particles does help a lot. Averaging over even a
moderate number of particles of about 1000 decreases the error by more
than an order of magnitude. Unfortunately, this is not possible for all
properties. For instance, the calculation of viscosity calls for the stress
correlation function [Eq. (67)], which is not a single particle property. This
is why self-diffusion coefficients are usually estimated with much better
accuracy than viscosity.
A major challenge in polymer simulations is to produce a large number
of uncorrelated configurations, to sample various chain conformations. In
molecular dynamics, each time step involves small changes in the monomer
positions. Significantly changing the chain conformation involves large

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

83

monomer displacements and requires many steps. The same thing happens
in Monte Carlo simulations involving only small monomer displacements
at each step. Global updates, involving large fragments of the polymer
molecules, which are used, e.g., in Configurational Bias Monte Carlo
techniques, help overcome this problem. Monte Carlo strategies for the
rapid equilibration of polymer systems will be discussed in subsequent
chapters of this book.

X. GENERAL STRUCTURE OF A SIMULATION

PROGRAM
A typical computer simulation experiment consists of the following
stages:
.
.

Initial Configuration Generation

Main Simulation Run
Read the initial configuration and the run parameters: total
number of simulation steps, frequency of saving configurations,
cutoff parameters, temperature, pressure, etc.
For the required number of simulation steps:
* Advance coordinates (according to the chosen MD, MC, or BD
algorithm)
* Increment volume or number of particles if NPT or Grand
Canonical ensemble simulation
* Calculate averages monitored during the run (energy, atomic
and molecular pressure, total momentum, atomic and molecular temperature)
* Save configuration on disk for further analysis.
Analysis of the obtained configurations. Calculate g(r), time correlation functions, diffusion coefficients, etc.

Different simulation packages may have these stages implemented as

one large program, or as separate modules, controlled by a common user
interface.
Sophisticated commercial simulation packages usually also include a
database of force field parameters, and prebuilt monomer units.
Generating an initial structure is fairly easy for the crystalline state, as
one just has to arrange molecules according to the structure of the crystal.
Liquid or glassy configurations of small molecules are usually generated
by melting a crystalline structure. This happens quite fast when the
molecules are small. For dense polymers it is impossible, due to the very
long relaxation times. Polymer chains have a very specific conformation in

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

84

Kotelyanskii and Theodorou

the crystal, which is far from the random coil conformation prevailing in
the polymer glass or melt. Therefore, special techniques are used to generate
realistic models of polymer melts or glasses [34,71].

REFERENCES
1. Hiller, A.; Wall, F.T.; Wheeler, D.J. Statistical computation of mean dimensions of macromolecules. J. Chem. Phys. 1954, 22, 1036.
2. Wall, W.A.; Seitz, F.T. Simulation of polymers by self-avoiding, non-intersecting random chains at various concentrations. J. Chem. Phys. 1977, 67,
37223726.
3. Alexandrowicz, Z.; Accad, Y. Monte Carlo of chains with excluded
volume: distribution of intersegmental distances. J. Chem. Phys. 1971, 54,
53385345.
4. Lal, M. Monte Carlo simulation of chain molecules I. Molec. Phys. 1969, 17, 57.
5. De Vos, E.; Bellemans, A. Concentration dependence of the mean dimension of
a polymer chain. Macromolecules 1974, 7, 812814.
6. Ryckaert, J.P.; Bellemans, A. Molecular dynamics of liquid butane near its
boiling point. Chem. Phys. Lett. 1975, 30, 123125.
7. Balabaev, N.K.; Grivtsov, A.G.; Shnol, E.E. Doklady Akad. Nauk SSSR
1975, 220, 10961098, in Russian.
8. Ceperley, H.L.; Frisch, H.L.; Bishop, M.; Kalos, M.H. Investigation of
static properties of model bulk polymer uids. J. Chem. Phys. 1980, 72,
32283235.
9. Weber, T.A.; Helfand, E. Molecular dynamics simulation of polymers. I.
Structure. J. Chem. Phys. 1979, 71, 47604762.
10. Vacatello, M.; Avitabile, G.; Corradini, P.; Tuzi, A. A computer model of
molecular arrangement in an n-paranic liquid. J. Chem. Phys. 1980, 73,
543552.
11. Fixman, M. Classical statistical mechanics of constraints. A theorem and application to polymers. Proc. Nat. Acad. Sci. USA 1974, 71, 30503053.
12. Go, N.; Scheraga, H. Analysis of the contribution of internal vibrations to the
statistical weights of equilibrium conformations of macromolecules. J. Chem.
Phys. 1969, 51, 47514767.
13. Go, N.; Scheraga, H. On the use of classical statistical mechanics in the treatment of polymer chain conformation. Macromolecules 1976, 9, 535542.
14. Mattice, W.; Suter, U.W. Conformational Theory of Large Molecules, the
Rotational Isomeric State Model in Macromolecular Systems; J. Wiley: New
York, 1994.
15. Insight II User Guide, version 2.3.0. Biosym Technologies: San Diego, 1993;
http://www.accelrys.com.
16. van Gunsteren, W.; Berendsen, H.J.C. Computer simulation of molecular
dynamics: methodology, applications and perspectives in chemistry. Angew.
Chem. Int. Ed. Engl. 1990, 29, 9921023.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

85

17. Weiner, S.J.; Kollman, P.A.; Nguyen, D.T.; Case, D.A. An all atom force eld
for simulations of proteins and nucleic acids. J. Comput. Chem. 1986, 7, 230,
http://www.amber.ucsf.edu/amber/amber.html.
18. Doi, M.; Edwards, S.F. The Theory of Polymer Dynamics; Clarendon Press:
Oxford, 1994.
19. Khokhlov, A.R.; Grosberg, A.Yu. Statistical Physics of Macromolecules; AIP
Press: New York, 1994.
20. Flory, P.J. Statistical Mechanics of Chain Molecules; Hanser Publishers:
Munich, 1989.
21. Ted Davis, H. Statistical Mechanics of Phases, Interfaces, and Thin Films; VCH:
New York, 1996.
22. deGennes, P.-G. Scaling Concepts in Polymer Physics; Cornell University Press:
Ithaca, 1979.
23. Fleer, G.J.; Cohen Stuart, M.A.; Scheutjens, J.M.H.M.; Cosgrove, T.; Vincent,
B. Polymers at Interfaces; Chapman and Hall: London, 1993.
24. Gotze, W.G. Recent tests of the mode-coupling theory for glassy dynamics.
J. Phys.: Condens. Matter 1999, 11, A1A45.
25. Cummins, H.Z. The liquid-glass transition: a mode coupling perspective.
J. Phys.: Condens. Matter 1999, 11, A95A117.
26. Strobl, G.R.; The Physics of Polymers: Concepts for Understanding Their
Structures and Behavior; Springer-Verlag: Berlin, 1997.
27. Car, R.; Parrinello, M. Unied approach for molecular dynamics and density
functional theory. Phys. Rev. Lett. 1985, 55, 24712474.
28. Reichl, L.E. A Modern Course in Statistical Physics; University of Texas Press:
Austin, 1980.
29. Arnold, V.I. Mathematical Methods of Classical Mechanics; Springer-Verlag:
2nd Ed., 1989.
30. Goldstein, H. Classical Mechanics; AddisonWesley: 2nd Ed., 1980.
31. Kneller, G.R.; Hinsen, K. Generalized Euler equations for linked rigid bodies.
Phys. Rev. 1994.
32. Hinsen, K.; Kneller, G.R. Inuence of constraints on the dynamics of polypeptide chains. Phys. Rev. 1995, E52, 6868.
33. Rice, L.M.; Brunger, A.T. Torsion angle dynamicsreduced variable conformational sampling enhances crystallographic structure renement. Proteins
Aug 1994, 19, 277290.
34. Theodorou, D.N.; Suter, U.W. Detailed molecular structure of a vinyl polymer
glass. Macromolecules 1985, 18, 14671478.
35. Ludovice, P.J.; Suter, U.W. Detailed molecular structure of a polar vinyl polymer glass. In Computational Modeling of Polymers; Bicerano, J., Ed.; Marcel
Dekker: New York, 1992; page 401.
36. Madras, N.; Sokal, A. Nonergodicity of local, length conserving MonteCarlo algorithms for the self-avoiding walk. J. Statist. Phys. 1987, 47(2),
573595.
37. McQuarrie, D.A. Statistical Mechanics; Harper and Row: New York,
1976.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

86

Kotelyanskii and Theodorou

38. Landau, L.D.; Lifshitz, E.M. Statistical Physics, volume 5 of Course of

Theoretical Physics. Pergamon Press: Oxford, 3rd edition, 1980.
39. Modell, M.; Reid, R.C. Thermodynamics and Its Applications; PrenticeHall:
Englewood Clis, 1983.
40. Chandler, D. Introduction to Modern Statistical Mechanics; Oxford University
Press: New York, 1987.
41. Theodorou, D.N.; Dodd, L.R.; Boone, T.D.; Manseld, K.F. Stress tensor in
model polymer systems with periodic boundaries. Makromol. Chem. Theory
Simul. 1993, 2, 191238.
42. Widom, B.; Some topics in the theory of uids. J. Chem. Phys. 1963, 39,
28082812.
43. Allen, M.P.; Tildesley, D.J. Computer Simulation of Liquids; Clarendon Press:
Oxford, 1993.
44. Spyriouni, T.; Economou, I.G.; Theodorou, D.N. Thermodynamics of chain
uids from atomistic simulation: a test of the chain increment method for
chemical potential. Macromolecules 1997, 30, 47444755.
45. Boulougouris, G.C.; Economou, I.G.; Theodorou, D.N. On the calculation of
the chemical potential using the particle deletion scheme. Molec. Phys. 1999,
96, 905913.
46. Hansen, J.-P.; McDonald, I.R. Theory of Simple Liquids; Academic Press:
New York, 2nd edition, 1986.
47. Higgins, J.S.; Benoit, H.C. Polymers and Neutron Scattering; Clarendon Press:
Oxford, 1994.
48. Metropolis, N.; Rosenbluth, A.W.; Rosenbluth, M.N.; Teller, A.H.; Teller, E.
Equation of state calculations by fast computing machines. J. Chem. Phys.
1953, 21, 10871092.
49. van Kampen, N.G.; Stochastic Processes in Physics and Chemistry; Amsterdam,
North Holland, 1992.
50. Andersen, H.-C. Molecular dynamics simulations at constant pressure and/or
temperature. J. Chem. Phys. 1980, 72, 23842393.
51. Nose, S. A molecular dynamics method for simulations in the canonical
ensemble. Molec. Phys. 1984, 52, 255268.
52. Hoover, W.G. Canonical dynamics: equilibrium phase space distribution. Phys.
Rev. 1985, A31, 16951697.
53. Berendsen, H.J.C.; Postma, J.P.M.; van Gunsteren, W.F.; Di Nola, A.; Haak,
J.R. Molecular dynamics with coupling to an external bath. J. Chem. Phys.
1984, 81, 36843690.
54. Kotelyanskii, M.J.; Hentschke, R. Gibbs-ensemble molecular dynamics: liquidgas equilibria for Lennard-Jones spheres and n-hexane. Molec. Simul. 1996, 17,
95112.
55. Lemak, A.S.; Balabaev, N.K. A comparison between collisional dynamics and
brownian dynamics. Molec. Simul. 1995, 15, 223231.
56. Mazur, P.; Oppenheim, I. Molecular theory of Brownian motion. Physica
1970, 50, 241258.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Background

87

57. Deutsch, J.M.; Oppenheim, I. The concept of Brownian motion in modern

statistical mechanics. Faraday discuss. Chem. Soc. 1987, 83, 120.
58. H.-C. Ottinger. Stochastic Processes in Polymer Fluids: Tools and Examples for
Developing Simulation Algorithms; Springer: Berlin, 1996.
59. Kramers, H.A. Brownian motion in a eld of force and the diusion model of
chemical reactions. Physica 1940, 7, 284305.
60. Ermak, D.L.; Yeh, Y. Equilibrium electrostatic eects on behavior of polyions
in solution: polyionmobile ion interaction. Chem. Phys. Lett. 1974, 24,
243248.
61. Ermak, D.L.; Buckholtz, H. Numerical integration of the Langevin equation:
Monte Carlo simulation. J. Comput. Phys. 1980, 35, 169182.
62. Chandler, D. Statistical mechanics of isomerization dynamics in liquids and the
transition state approximation. J. Chem. Phys. 1978, 68, 29592970.
63. Voter, A.; Doll, J. Dynamical corrections to transition state theory for multistate systems: surface self-diusion in the rare-event regime. J. Chem. Phys.
1985, 82, 8087.
64. June, R.L.; Bell, A.T.; Theodorou, D.N. Transition-state studies of xenon and
SF6 diusion in silicalite. J. Phys. Chem. 1991, 95, 88668878.
65. Bell, T.A.; Sevick, E.M.; Theodorou, D.N. A chain of states method for investigating infrequent event processes in multistate, multidimensional systems.
J. Chem. Phys. 1993, 98, 31963212.
66. Elber, R. Enhanced sampling in molecular dynamicsuse of time dependent
hartree approximation for a simulation of carbon-monoxide diusion through
myoglobin. J. Chem. Phys. 1990, 93, 43124321.
67. Vineyard, G.H. Frequency factors and isotope eects in solid state rate processes. J. Phys. Chem. Solids, 1957, 3, 121127.
68. Born, M.; Von Karman, Th. Uber schwingungen in raumgittern. Z. Physik,
1912, 13, 297309.
69. Binder, K. editor. The Monte Carlo Method in Condensed Matter Physics;
Springer: Berlin, 2nd edition, 1995.
70. Wilding, N.R.; Binder, K. Finite-size scaling for near-critical continuum uids
at constant pressure. Physica A 1996, 231, 439447.
71. Kotelyanskii, M. Simulation methods for modeling amorphous polymers.
Trends Polymer Sci. 1997, 5, 192198.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

2
Rotational Isomeric State (RIS)
Calculations, with an Illustrative
Application to Head-to-Head,
Tail-to-Tail Polypropylene
E. DEMET AKTEN and WAYNE L. MATTICE
Akron, Ohio, U.S.A.
ULRICH W. SUTER
Zurich, Switzerland

I.

The University of Akron,

Eidgenossische Technische Hochschule (ETH),

INTRODUCTION

The Rotational Isomeric State (RIS) model is excellent for the analysis of
conformation-dependent physical properties of chain molecules in their
unperturbed state. Its special strength is the incorporation of detailed
information about the covalent structure (bond lengths, bond angles,
torsion angles, and torsion potential energy functions) in a formalism that
can be evaluated quickly by the smallest computer. The answer is the exact
result for the specific model, as defined by the geometry and energies of the
short-range intramolecular interactions.
The most frequently calculated property is the mean square unperturbed
end-to-end distance, hr2i0. Other properties susceptible to rapid computation include the average of the end-to-end vector, hri0, and the mean square
unperturbed radius of gyration, hs2i0. The viscosity of a dilute solution in a

solvent can be estimated from hs2i0 or hr2i0 via the equivalent sphere
model for hydrodynamic properties. Several higher even moments, hr2pi0
and hs2pi0, p 2, 3, . . . , provide information about the shape (width,
skewness) of the distribution functions for r2 and s2. When combined with
information about the electronic charge distribution within individual rigid
units, the RIS model can calculate the mean square dipole moment, h 2i0.
Also accessible are optical properties that depend on the anisotropy of the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

90

Akten et al.

polarizability tensor (optical anisotropy) in conjunction with r (stressoptical coefficient), k (molar Kerr constant), or the magnetic susceptibility
(CottonMouton constant).
The speed of the calculations derives from the use of efficient generator
matrix techniques that were introduced six decades ago [1]. These techniques
were adapted to the study of macromolecules a decade later [2]. Therefore
the description of conformation-dependent physical properties of macromolecules with the RIS model predates the widespread availability of
computers and simulation methods that require enormous computational
power for their implementation. A strong increase in the use of the RIS
model over the decade from 1959 to 1969 culminated in the publication of
Florys classic book on the subject [3]. Some of the subsequent developments are contained in two books published in the 1990s. One book has an
emphasis on self-instruction in the use of the RIS model [4]. The other book
is a compilation, in standard format, of RIS models for over 400 polymers
that have appeared in the literature through the mid-1990s [5].

II. THREE FUNDAMENTAL EQUATIONS

IN THE RIS MODEL
A. The First Equation: Conformational Energy
This equation describes the conformation partition function, Z, as a serial
product of statistical weight matrices, Ui, that incorporate the energies of all
of the important short-range intramolecular interactions.
Z U1 U2 . . . Un

There is one statistical weight matrix for each of the n bonds in the chain.
The dimensions of the statistical weight matrices depend on the number of
rotational isomers assigned to each bond. The continuous range for the
torsion angle, , is replaced by a finite set of torsion angles, i.e., a continuous conformational integral is replaced by a discrete summation. The sum
should be designed so that the terms include all regions of the local conformational space that have a significant population. Many unperturbed
chains have pairwise interdependence of their bonds, causing each Ui to
have a number of columns given by the number of rotational isomers at
bond i (denoted i), and a number of rows given by the number of rotational
isomers at the preceding bond. Each element in Ui is the product of a
statistical weight for a first-order interaction (which depends only on
the state at bond i) and the statistical weight for a second-order interaction (which depends simultaneously on the states at bonds i
1 and i).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

RIS Calculations: Polypropylene

91

Each statistical weight is formulated as a Boltzmann factor with the appropriate energy and temperature. Sometimes there is inclusion of a preexponential factor that accounts for the differences in conformational entropy of
the various rotational isomeric states.
The first equation incorporates all of the energetic information that will
be employed in the calculation. It incorporates structural information only
indirectly, via the assignment of the numbers of rotational isomers at the
various bonds, and the assignment of numerical values for the statistical
weights.

B. The Second Equation: Structure

The second equation describes a conformation-dependent physical property
of a chain in a specific conformation as a serial product of n matrices, one
matrix for each bond in the chain.
f F1 F2 . . . Fn

There is no concern here about the probability of observation of this conformation; we are ignoring, for the moment, the energetic information contained in Z. Each Fi includes the information about the manner in which
bond i affects f. The details, of course, depend on how we have chosen f. If f
is the squared end-to-end distance, we need to know the length of bond i, the
angle between bonds i and i 1, and the torsion angle at bond i. These three
properties are denoted li,
i, and i, respectively. The square of the end-toend distance is written in terms of these local variables as

r2 2

n
1 X
n
X
k1 jk1

lTk Tk Tk1 . . . Tj
1 lj

n
X

lk2

k1

where li denotes the bond vector, expressed in a local coordinate system for
bond i,
2 3
li
li 4 0 5
0

and Ti is a transformation matrix. A vector expressed in the local coordinate

system of bond i 1 is transformed into its representation in the local

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

92

Akten et al.

coordinate system of bond i through premultiplication by Ti,

2

cos

Ti 4
sin
cos 

sin
sin 

3
sin

0

cos
cos 
sin  5

cos
sin  cos  i

The single and double sums in Eq. (3) are evaluated exactly via Eq. (2) using


F1 1 2lT1 T1 l12
3
2
1 2lTi Ti li2
7
6
Fi 6
Ti
li 7
5,
40
0
2

1<i<n

3
2

ln
6 7
7
Fn 6
4 ln 5

C. The Third Equation: Conformational Energy

Combined with Structure
The third equation combines the energetic information from Eq. (1) with the
structural information in Eq. (2), in order to average f over all of the
conformations in the RIS model. The combination is achieved in a
generator matrix, G.
h f i0

1
G1G2 . . . Gn
Z

G i Ui  IF diagF , F , . . . , Fi

9
10

The generator matrix contains the energetic information in an expansion of

the statistical weight matrix (via its direct product with an identity matrix of
order equal to the number of columns in F). This energetic information is
multiplied onto the structural information (in the form of a block-diagonal
matrix of the i expressions for F for the rotational isomers at bond i). The
efficiency of the calculation arises from the fact that the conformational
average of the property of interest is obtained as a serial product of matrices
via Eq. (9).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

RIS Calculations: Polypropylene

93

The remainder of this chapter is devoted to a case study. Readers who

seek more detailed or general information on the origin, structure, and
usage of the equations for Z, f, and hfi0 are referred to books on the RIS
model [3,4].

III. CASE STUDY: MEAN SQUARE UNPERTURBED

DIMENSIONS OF HEAD-TO-HEAD, TAIL-TO-TAIL
POLYPROPYLENE
Polypropylene is an important polymer that is usually assembled with the
monomer units arranged in head-to-tail sequence, producing -CH(CH3)CH2- as the repeat unit. Several RIS models, based on three [612], four [13],
five [14,15], or seven [16] states per rotatable bond, have been described for
this common form of polypropylene. Head-to-head, tail-to-tail units have
been considered briefly [11] in order to assess how the incorporation of a few
such units, as defects, alters the unperturbed dimensions of the conventional
head-to-tail polypropylene.
Two experimental investigations of head-to-head, tail-to-tail polypropylene, obtained by reduction of poly(2,3-dimethylbutadiene), give discordant
results [17,18]. Values of 4.5 and 6.16.4, respectively, are obtained for the
characteristic ratio, Cn, of samples of high molecular weight.
Cn 

hr2 i0
nl 2

11

Can an analysis with the RIS model assist in the resolution of this
discrepancy in the experimental results? A three-state RIS model for the
head-to-head, tail-to-tail form of polypropylene, with a repeating sequence
given by -CH2-CH(CH3)-CH(CH3)-CH2-, is constructed and analyzed here.
The mean square unperturbed dimensions of exclusively head-to-head, tailto-tail polypropylene are determined for chains with various stereochemical
compositions and stereochemical sequences. The critical parameters in the
model are identified, and the predicted mean square unperturbed dimensions are compared with experiment.

A. Construction of the RIS Model

Rotational Isomeric States. The model assumes three states (trans and
gauche, abbreviated t and g) for each rotatable bond. The torsion angle,
as defined on page 112 of Mattice and Suter [4], is 180 for a t placement.
It increases with a clockwise rotation.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

94

Akten et al.

Short-Range Interactions. The model is limited to first- and second-order

interactions. The repulsive first-order interactions are variants of the one that
occurs in the g conformations of n-butane. The second-order interaction
included is the one commonly described as the pentane effect, which is
generated by the repulsive interaction of the terminal methyl groups in the
conformations of n-pentane with g placements of opposite sign. The
statistical weight matrices are constructed under the simplifying assumption
that methyl, methylene, and methine groups are equivalent insofar as firstand second-order interactions are concerned. This assumption reduces the
number of distinguishable parameters in the model.
Description of the Stereochemical Sequence. The stereochemical sequence
is described using dl pseudoasymmetric centers, as defined on page 175 of
Mattice and Suter [4]. The CC bonds in the chain are indexed sequentially
from 1 to n. A local Cartesian coordinate system is associated with each
bond. Axis xi lies along bond i, with a positive projection on that bond. Axis
yi is in the plane of bonds i
1 and i, with a positive projection on bond
i
1. Axis zi completes a right-handed Cartesian coordinate system. The
chain atom at the junction of bonds i
1 and i is a d pseudoasymmetric
center if it bears a methyl group with a positive coordinate along zi; it is
an l pseudoasymmetric center if this coordinate is negative.
General Form of the Statistical Weight Matrices. First the statistical weight
matrices are formulated in symbolic form. Later numerical values will be
assigned to the various symbols.
The 3  3 statistical weight matrix for bond i has rows indexed by the
state of bond i
1, columns indexed by the state of bond i, and the order
of indexing is t, g, g
in both cases. Each statistical weight matrix is
assembled as Ui ViDi, where Vi is a 3  3 matrix that incorporates the
statistical weights for second-order interactions, and Di is a diagonal matrix
that incorporates the statistical weights for first-order interactions. Efficient
description of the interrelationships between selected pairs of matrices
employs a matrix Q with the property Q2 I3.
2
3
1 0 0
Q 40 0 15
12
0 1 0
Statistical Weight Matrices for the CH2CH2 Bond. This statistical weight
matrix incorporates short-range interactions that occur in the fragment
-CH-CH(CH3)-CH2CH2-CH-, where the longest dash denotes bond i.
The matrix is denoted by UCC;x, where x denotes the stereochemistry (d or l )
of the attachment of the methyl group. Figure 1 depicts this fragment when
the methyl group produces an l pseudoasymmetric center. Internal rotation

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

RIS Calculations: Polypropylene

FIG. 1

95

Fragment considered in the construction of UCC;l.

about bond i establishes the position of the terminal CH group with

respect to the remainder of the fragment. A diagonal matrix incorporates
the first-order interactions between the pair of bold CH groups in
-CH-CH(CH3)-CH2CH2-CH-.
DCC diag1, , 

13

This matrix is independent of the stereochemistry of the attachment of the

methyl group, because it only involves carbon atoms that are in the main
chain. In Eq. (13),  is the statistical weight for a g placement at the
CH2CH2 bond, the t placement being assigned a statistical weight of 1.
The second-order interactions in UCC;x take place between the terminal
bold CH group in -CH-CH(CH3)-CH2CH2-CH- and the two groups in
this fragment that are bold in the previous paragraph. Involvement of the
methyl group in the second-order interactions demands two forms of V,
depending on whether the fragment contains a d or l pseudoasymmetric
center.
2

VCC;l

1 1
6
41 !
1 !

3
!
7
!5

14

VCC;d QVCC;l Q

15

Here ! is the statistical weight for the second-order interaction of the terminal chain atoms when the conformation has two g placements of opposite
sign (the ! in the 2,3 and 3,2 elements), or the equivalent interaction of the
methyl group with the terminal CH (the remaining two !, which occur in the
1,3 and 2,2 elements of VCC;l, but in the 1,2 and 3,3 elements of VCC;d). The
full statistical weight matrices, incorporating both first- and second-order
interactions, are
2

6
UCC;l VCC;l DCC 4 1
1

!

7
! ! 5
! 

UCC;d QUCC;l Q

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

16
17

96

Akten et al.

FIG. 2

Fragment considered in the construction of UCM;l.

Statistical Weight Matrices for the CH2CH Bond. This statistical weight
matrix, denoted by UCM;x, incorporates statistical weights for short-range
interactions that occur in the fragment -CH-CH2-CH2CH(CH3)-CH-,
where the longest dash denotes the bond that is now considered as bond i.
An example of this fragment is depicted in Fig. 2. Rotation about bond i
establishes the position of the bold CH and bold CH3 in -CH-CH2-CH2
CH(CH3)-CH- with respect to the remainder of the fragment. The firstorder interactions now occur between the bold CH2 and the other two bold
groups. Involvement of the methyl group demands two forms of the matrix
for the first-order interactions.
DCM;l diag1, 1,

18

DCM;d QDCM;l Q

19

Here is the statistical weight for two simultaneous first-order interactions,

relative to the case where there is a single first-order interaction.
The second-order interactions occur between the initial CH and the other
two bold groups in the fragment -CH-CH2-CH2CH(CH3)-CH-. Rotation
about bond i alters the position of the methyl group, which was not true in
the prior consideration of the second-order interactions that are incorporated in UCC;x. The matrices of second-order interactions differ for the two
bonds.
2

6
VCM;l 4 !
1

7
!5

20

! !

VCM;d QVCM;l Q

21

The statistical weight matrices are

2

UCM;l

1
6
4!
1

1
1

3

7
! 5

22

! !

UCM;d QUCM;l Q

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

23

RIS Calculations: Polypropylene

FIG. 3

97

Fragment considered in the construction of UMM;ll.

Statistical Weight Matrices for the CHCH Bond. This statistical weight
matrix incorporates short-range interactions in the fragment -CH2-CH2CH(CH3)CH(CH3)-CH2-, depicted in Fig. 3. The matrix is denoted by
UMM;xy, where x and y define the stereochemistry of the attachments of the
two methyl groups, in the order that they occur in the fragment. The firstorder interactions occur between pairs of bold groups in -CH2-CH2CH(CH3)CH(CH3)-CH2-, with a member of the pair being selected from
among the bold groups on either side of bond i. Involvement of two methyl
groups demands four forms of the diagonal matrix for the first-order
interactions. Two of these forms are identical.
DMM;dd DMM;ll diag1, , 

24

DMM;dl diag, 1, 

25

DMM;ld QDMM;dl Q

26

Here  is the statistical weight for three simultaneous first-order interactions, relative to the case where there are two first-order interactions.
The second-order interactions occur between the initial methylene and
the other two bold groups in -CH2-CH2-CH(CH3)CH(CH3)-CH2-.
Involvement of only one of the methyl groups means that there are only
two forms of the matrix of second-order interactions. These two matrices
were encountered in the consideration of the previous bond.
VMM;ll VMM;dl VCM;l

27

VMM;ld VMM;dd VCM;d

28

The statistical weight matrices are

2

6
UMM;ll 6
4!

!

7
! 7
5

29

!

UMM;dd QUMM;ll Q

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

30

98

Akten et al.

6
UMM;dl 6
4 !


7
! 7
5

!

UMM;ld QUMM;dl Q

31

32

Statistical Weight Matrices for the CHCH2 Bond. This matrix incorporates statistical weights for short-range interactions in the fragment -CH2CH(CH3)-CH(CH3)CH2-CH2-, depicted in Fig. 4. It is denoted by
UMC;xy, where x and y define the stereochemistry of the attachments of the
two methyl groups, in the order that they occur in the fragment. The firstorder interactions involve the bold groups in -CH2-CH(CH3)-CH(CH3)
CH2-CH2-, and the second-order interactions involve the bold groups in
-CH2-CH(CH3)-CH(CH3)CH2-CH2-. All of the short-range interactions
have been encountered previously, as have all of the matrices of first- and
second-order interactions.
DMC;ll DMC;dl DCM;d

33

DMC;dd DMC;ld DCM;l

34

VMC;ll VMC;ld VCC;l

35

VMC;dd VMC;dl VCC;d

36

These matrices combine to give statistical weight matrices distinct from

those obtained with the other bonds.
2

6
UMC;ll 6
41
1

7
! ! 7
5
! 1

UMC;dd QUMC;ll Q

FIG. 4

Fragment considered in the construction of UMC;ll.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

37

38

RIS Calculations: Polypropylene

6
UMC;ld 6
41 !
1 !

99

7
! 7
5

39

UMC;dl QUMC;ld Q

40

None of the statistical weight matrices for head-to-head, tail-to-tail

polypropylene is found also in a three-state RIS model for the conventional
head-to-tail polypropylene. This result follows immediately from the fact
that none of the fragments in Figs. 14 is also a fragment found in head-totail polypropylene.
Conformational Partition Function. The conformation partition function
for a chain of n bonds (see, e.g., pp. 7783 of Ref. [4]) is given by Eq. (1)
where U1, U2, and Un adopt special forms.

U1 1
2

41

6
6
U2 6 1
4

7
7
1 7D2
5

42

2 3
1
6 7
6 7
Un 6 1 7
4 5

43

1
Any ! that appears in U2 is replaced by 1, because the chain has been
initiated in a manner that eliminates any possibility of second-order interactions at bonds 1 and 2.
For long chains in which all of the pseudoasymmetric centers have the
same chirality (isotactic chains), Z is dominated by the terms


UCC;l UCM;l UMM;ll UMC;ll

a

44

or


UCC;d UCM;d UMM;dd UMC;dd

a

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

45

100

Akten et al.

where a is the number of successive configurational repeat units in the

chain. Using Eqs. (17), (23), (30), and (38), along with Q2 I3,


a

a
UCC;d UCM;d UMM;dd UMC;dd Q UCC;l UCM;l UMM;ll UMC;ll Q

46

which demonstrates the equivalence of the Zs for these two chains (they are
enantiomers). The two chains with perfectly alternating (dl or ld) pseudoasymmetric centers also have the same Z [which, however, is different from
the Z in Eq. (46)] because the two structures are identical; they constitute
different representations of a mesoform, which is syndiotactic.

B. Behavior of the RIS Model

1. Preliminary Estimates of the Values
of the Statistical Weights
The foregoing matrices contain four distinct statistical weights. Three (, ,
and ) are statistical weights for first-order interactions that appear in the
diagonal D matrices. The V matrices contain another statistical weight, !,
for a second-order interaction. Two of the statistical weights,  and !, are
for interactions that are similar to the short-range interactions considered
many years ago in RIS models for polyethylene. Model B of Abe et al.
[19] estimates the corresponding energies as E 1.82.5 kJ/mol and
E! 7.18.0 kJ/mol. The first-order interaction with statistical weight
occurs also in the head-to-tail variety of polypropylene. The corresponding
energy has been estimated as E 3.8  1.7 kJ/mol [9]. The statistical weight
denoted here by  was previously estimated as 0.6, from a CNDO
investigation of the conformations of 2,3-dimethylbutane [11].
A provisional set of values of statistical weights is defined at this point in
the development of the model. Then the sensitivity of the model to each of
the initial estimates can be determined. Our provisional initial set is  0.43,
0.22,  0.6, ! 0.034. These values are Boltzmann factors computed
from the corresponding energies at 300K. For the geometry, we adopt a
bond length (l) of 1.53 A, a bond angle (
) of 109.5 for all C-C-C angles in
the backbone, and torsion angles () of 180 and 60 for the trans and
gauche states at each rotatable bond. The geometry enters the calculations
via the matrices in Eqs. (6)(8) [20].

2. Unperturbed Dimensions of Chains with

Simple Stereochemical Sequences
The two simplest stereochemical sequences are those in which all
pseudoasymmetric centers have the same chirality (either ll. . .ll or

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

RIS Calculations: Polypropylene

101

dd. . .dd), or where there is a strict alternation in chirality (either ld. . .ld or
dl. . .dl). The values of C1 can be calculated using the program in Appendix
C of Mattice and Suter [4]. This short program computes the mean square
unperturbed end-to-end distance, hr2i0, for a repeating unit of m bonds
embedded in a chain of n bonds, using a special case of Eq. (VI-66) of
Ref. [4], which for present purposes can be written as
2
1
r 0 G 1 G 2 G 3 . . . G m G 1 G 2 n=m
1 G 3 . . . G m
1 G m
Z

47

Z U1 U2 U3 . . . Um U1 U2 n=m
1 U3 . . . Um
1 Um

48

The internal bold italic G matrices in Eq. (47) are of dimensions 15  15,
formulated as described in Eq. (10) (see, e.g., pp. 122126 and 128130 of
Ref. [4]). The first and last matrices in Eqs. (47) and (48) take special forms
of a row and column, respectively [20]. The hr2i0 are discussed here using the
asymptotic limit for the dimensionless characteristic ratio, which is determined by linear extrapolation of Cn vs. 1/n (see pp. 1920 of Ref. [4]).
C1  lim Cn

49

n!1

Table 1 shows that identical results are obtained for a chain and its
mirror image, as expected, but different C1 can be obtained for pairs of
chains in which the stereochemical sequences are not mirror images. The
relationship between the two distinct results in Table 1 can be rationalized
qualitatively by identification of the preferred conformation of the sequence
of four rotatable bonds. Identification can be achieved by inspection of the
pattern of the appearance of , , , and ! in the U matrices, realizing that
each statistical weight has a value smaller than 1.
Equations (16), (22), (29), and (37), or, equivalently, Eqs. (17), (23), (30),
and (38), show that the preferred conformation is tttt when a single chirality
is propagated along the chain. The all-trans conformation has a statistical
weight of 1, and all other conformations have a statistical weight of less

TABLE 1 Characteristic Ratios with the Provisional Set of Statistical Weights

Sequence
ll . . . ll dd . . . dd
ld . . . ld
dl . . . dl

Us (equation numbers)

C1

16, 22, 29, 37 (17, 23, 30, 38)

16, 23, 31, 39
17, 22, 32, 40

7.94
7.30
7.30

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

102

Akten et al.

than 1. The preferred conformation becomes either ttgt or ttg
t when there
is a strict alternation in chirality of the pseudoasymmetric centers. The g
conformation is preferred at the CH2CH(CH3)CH(CH3)CH2 bond, due to
the pattern of the statistical weights denoted by 1 and  in the D matrices in
Eqs. (24)(26), and hence in the U matrices in Eqs. (29)(32). Therefore, on
the basis of the preferred conformations, one would expect the first entry
in Table 1 to have a higher C1 than the last two entries.

3. Sensitivity Tests for the Statistical Weights

The sensitivity of the unperturbed dimensions to each of the statistical
weights was assessed using chains with the two distinct types of
stereochemical sequences presented in Table 1. The values of C1 were
recalculated by variation of one of the statistical weights, the other three
statistical weights being retained at their provisional values. This information can be summarized in a tabular form as



@ ln C1
C1, x0 x
C1, x0
x x0

@ ln x x0
2x
C1, x0

50

Here x is the statistical weight being varied, x0 is its value in the provisional
set, and all other parameters are held constant at their values in the provisional set. The values of this partial derivative are summarized in Table 2.
An increase in the value of any of the statistical weights produces a
decrease in C1 if all pseudoasymmetric centers are either d or l. Changes in 
affect C1 relatively more than comparable changes in any of the other three
statistical weights. In contrast, a mixed pattern is seen with the chain
in which there is strict alternation in the pseudoasymmetric centers. C1
increases with an increase in  or , or a decrease in or !, with the greatest
sensitivity seen upon a change in . Table 2 also shows how C1 is affected
by an increase in all bond angles, or by increasing only that half of the bond

TABLE 2 Sensitivity to the Statistical Weights and Bond Angle

Parameter



!
All


at methylene

Single chirality

Strictly alternating chirality

0.14

0.069

0.22

0.11
3.8
0.65

0.028

0.106
0.055

0.067
4.6
1.3

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

RIS Calculations: Polypropylene

103

angles that is centered on a methylene group. Opening up the bond angles

produces an increase in C1, as expected from the behavior of the simple
freely rotating chain.
It is worthwhile to also consider the consequences of minor adjustments
in the torsion angles assigned to the rotational isomers. The preferred
conformations of n-butane have torsion angles of 180 and (60 i),
where i is positive [19]. Displacement of the gauche states arises because
of the weak onset at  60 of the repulsive interaction characteristic of
the cis state of n-butane, thereby displacing the dihedral angle for the gauche
states slightly toward the trans state. Therefore the gauche states at the
CH2CH2 bond in head-to-head, tail-to-tail polypropylene can be adjusted
to torsion angles of (60 i). A displacement of similar origin occurs
for the trans and one of the gauche states at the bonds between CH2 and
CH. For example, when all of the pseudoasymmetric centers are l, torsion
angles associated with the columns of UCM;l are 180
i, 60 i, and

60 , and those associated with the columns of UMC;ll are 180 i, 60 ,
and
60
i. An increase in all of the i by 1 increases C1 by about
1.5% when all pseudoasymmetric centers are of the same chirality, but the
increase is only about 0.3% when the pseudoasymmetric centers alternate
between d and l.

C. Comparison with Experiment

Arichi et al. [17] reported an estimate of 4.5 for the characteristic ratio of
predominantly head-to-head, tail-to-tail polypropylene in isoamylacetate
at 316K, which is the
temperature for this polymersolvent system. This
number is significantly smaller than the result of 5.9 (at 311K) reported
for atactic head-to-tail polypropylene [21]. The sample studied by Arichi
et al. was prepared by reduction of poly(2,3-dimethyl-1,3-butadiene). The
parent polymer was composed predominantly (94%) of units arising from
1,4 addition, with the remaining 6% of the units arising from 1,2 addition.
Upon reduction, the latter units yield chain atoms that bear both an
isopropyl group and a methyl group.
Recently neutron scattering has been used to determine the mean square
unperturbed dimensions for the polymer in its melt [18]. The samples were
hydrogenated poly(2,3-dimethyl-1,3-butadiene), with deuterium introduced
during the reduction. The parent polymers had 3% 1,2-units. The
polymer had a characteristic ratio of 6.16.4 when studied over the range
27167 C [18].
A characteristic ratio of 4.5 is lower than either number in Table 1. The
values in Table 1 might be too high for an atactic polymer. The strictly

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

104

Akten et al.

alternating chirality imposes preferred conformations that are either

ttgt . . . ttgt or ttg
t . . . ttg
t, depending on whether the sequence is
written as ld . . . ld or dl . . . dl. In both of these preferred sequences, the
gauche placement at every fourth bond is always of the same sign. In
contrast, a random sequence of l and d, with equal amounts of each, would
not perpetuate a gauche placement of a single sign at every fourth bond.
The consequences for the unperturbed dimensions are brought out in
Table 3, which contains results for six distinguishable chains with equal
numbers of l and d pseudoasymmetric centers, arranged in different
repeating sequences. The first entry, taken from Table 1, is for the shortest
such repeating sequence, where there is a strict alternation of l and d along
the chain. The next two entries have a strict alternation of a pair of ls and a
pair of ds. This sequence can be embedded in the chain in two distinct ways,
which differ in whether two bonded carbon atoms with methyl substituents
have the same or opposite chirality. The fourth entry has a repeating pattern
of three ls followed by three ds. The fifth and sixth entries have a homopair
(e.g., ll) followed by the opposite homopair (e.g., dd), and then a heteropair
(e.g., ld), which can be embedded in the chain in two ways.
The results in Table 3 show that the strictly alternating dl polymer does
indeed overestimate the C1 expected for a truly atactic polymer, but the size
of the overestimate is small. If one averages over the first three entries in
Table 3, the result is 6.77. The average is only slightly smaller (6.73) if one
considers all repeating sequences of three -CH2-CH(CH3)-CH(CH3)-CH2units, assuming equal numbers of d and l pseudoasymmetric centers in each
repeat unit, and equal probability that any pseudoasymmetric center is d
or l. We infer that the C1 considering all possible sequences with equal
numbers of d and l pseudoasymmetric centers cannot be smaller than 6, and

TABLE 3 Characteristic Ratios for Six Repeating

Sequences, Each of Which Contains the Same Number
of l and d Pseudoasymmetric Centers
Repeating sequence of pseudoasymmetric centersa

C1

(ld)
(ld)(dl)
(ll)(dd)
(ll)(ld)(dd)
(ll)(dd)(ld)
(ld)(dl)(dl)

7.30
6.03
6.97
6.94
6.74
6.32

a
Parentheses enclose symbols for two bonded carbon atoms
that each bear a methyl group.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

RIS Calculations: Polypropylene

105

is probably in the range 6.06.7. The predictions from the model are
definitely larger than the experimental estimate of 4.5 [17], but are consistent
with the experimental estimate of 6.16.4 [18].

1. Plausible Adjustments in Parameters

We now inquire whether reasonable adjustment in any of the parameters in
the model reduces C1 for the atactic chain from 6.06.7 down to 4.5. First
consider reasonable adjustments in the statistical weights. Table 2 shows
that C1 for the chain with alternating chirality is more sensitive to than to
any of the other statistical weights. An increase in will produce a decrease
in C1. However, while it is conceivable that might be nearly as large as 
[22], there are strong physical arguments for assuming that  , because
two simultaneous repulsive interactions are unlikely to be weaker than two
times a single interaction. An increase in to  reduces C1 by only
about 10%, which is insufficient to explain the difference between the model
and experimental result of 4.5. We conclude that reasonable adjustments in
the statistical weights cannot account for the difference.
Reasonable adjustment in the bond angles would increase them above
their provisional values of 109.5 , which would lead to an increase in C1.
Similarly, any reasonable change in the torsion angles would assign i  0 ,
which would also increase C1. It appears that no reasonable adjustment in
the model, either in its energetics (, , , !) or in its geometry (
, i), will
bring the calculated C1 into the range reported in the experiments that find
a characteristic ratio of 4.5. No adjustment is necessary in order to
rationalize the experiment that reports a characteristic ratio in the range
of 6.16.4.
The experiments were performed with a predominantly head-to-head,
tail-to-tail chain that contained defects, arising from the 1,2 addition of 3 or
6% of the monomer units in the parent poly(2,3-dimethyl-1,3-butadiene).
When short branches are present as defects in a polyethylene chain, they
produce a decrease in C1, by disrupting the preferred conformation
consisting of short sequences of trans placements [23]. The defects present
in the chains examined in the experiment, illustrated in Fig. 5, may play

FIG. 5 A fragment showing the defect arising from the 1,2 addition in the parent
polymer.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

106

Akten et al.

a similar role. However, it does not seem likely that an increase in defects
from 3 to 6% could depress the characteristic ratio from 6.16.4 to 4.5.

D. Conclusion
The construction and analysis of a RIS model for a homopolymer have been
illustrated by consideration of head-to-head, tail-to-tail polypropylene.
Statistical weight matrices, incorporating first- and second-order interactions, are formulated for all bonds, and for all configurations of the
attachments of the methyl groups to the chain. Preliminary estimates of the
values of the statistical weights in these matrices could be obtained by
reference to prior work, which in turn is based on conformational energy
calculations for small hydrocarbons. Along with preliminary estimates of
the geometry in the three rotational isomeric states for each rotatable bond,
this information permits calculation of C1 for chains with simple repeating
sequences.
The influence of the individual energetic and structural parameters on
C1 is determined in the sensitivity tests.
The values of C1 specified by the model for an atactic homopolymer of
head-to-head, tail-to-tail polypropylene are in excellent agreement with
those reported by Graessley et al. [18]. Reasonable adjustments of the
parameters in the RIS model cannot reduce Cn into the range reported by
Arichi et al. [17].

ACKNOWLEDGMENT
The construction and analysis of the RIS model was supported by the
Edison Polymer Innovation Corporation.

REFERENCES
1. Kramers, H.A.; Wannier, G.H. Phys. Rev. 1941, 60, 252.
2. Volkenstein, M.V. Dokl. Akad. Nauk SSSR 1951, 78, 879.
3. Flory, P.J. Statistical Mechanics of Chain Molecules; Wiley-Interscience:
New York, 1969; reprinted with the same title by Hanser, Munchen (Munich),
1989.
4. Mattice, W.L.; Suter, U.W. Conformational theory of large molecules.
The Rotational Isomeric State Model in Macromolecular Systems; Wiley:
New York, 1994.
5. Rehahn, M.; Mattice, W.L.; Suter, U.W. Adv. Polym. Sci. 1997, 131/132.
6. Abe, Y.; Tonelli, A.E.; Flory, P.J. Macromolecules 1970, 3, 294.
7. Biskup, U.; Cantow, H.J. Macromolecules 1972, 5, 546.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

RIS Calculations: Polypropylene

8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.

107

Tonelli, A.E. Macromolecules 1972, 5, 563.

Suter, U.W.; Pucci, S.; Pino, P.J. Am. Chem. Soc. 1975, 97, 1018.
Asakura, T.; Ando, I.; Nishioka, A. Makromol. Chem. 1976, 177, 523.
Asakura, T.; Ando, I.; Nishioka, A. Makromol. Chem. 1976, 177, 1493.
Alfonso, G.C.; Yan, D.; Zhou, A. Polymer 1993, 34, 2830.
Flory, P.J. J. Polym. Sci. Polym. Phys. 1973, 11, 621.
Heatley, F. Polymer 1972, 13, 218.
Suter, U.W.; Flory, P.J. Macromolecules 1975, 8, 765.
Boyd, R.H.; Breitling, S.M. Macromolecules 1972, 5, 279.
Arichi, S.; Pedram, M.Y.; Cowie, J.M.G. Eur. Polym. J. 1979, 15, 113.
Graessley, W.W.; Krishnamoorti, R.; Reichart, G.C.; Balsara, N.P.; Fetters,
L.J.; Lohse, D.J. Macromolecules 1995, 28, 1260.
Abe, A.; Jernigan, R.L.; Flory, P.J. J. Am. Chem. Soc. 1966, 88, 631.
Flory, P.J. Macromolecules 1974, 7, 381.
Zhongde, X.; Mays, J.; Xuexin, C.; Hadjichristidis, N.; Schilling, F.C.; Bair,
H.E.; Pearson, D.S.; Fetters, L.J. Macromolecules 1985, 18, 2560.
Mathur, S.C.; Mattice, W.L. Makromol. Chem. 1988, 189, 2893.
Mattice, W.L. Macromolecules 1986, 19, 2303.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

3
Single Chain in Solution
REINHARD HENTSCHKE

I.

Bergische Universitat, Wuppertal, Germany

PHENOMENOLOGICAL FORCE FIELDS

AND POLYMER MODELING

Perhaps the simplest theoretical approach to polymers in solution is

molecular dynamics, i.e., the numerical integration of Newtons equations
~ r~ U~r1 , . . . , r~n , where i extends over all n atoms in the
of motion, mi r~i
r
i
simulation box. Common integrators like the Verlet or predictor corrector
algorithms [1], yield trajectories consisting of positions and velocities of
the interaction centers (usually the nuclei) stored at regular time intervals
on the picosecond time scale. Positions and velocities can be tied to the
thermodynamic quantities like pressure and temperature via the equipartition theorem. The forces in the above equations of motion may be derived
from a potential of the form
U

fb b
bo 2

bonds

p
X X

f 
o 2

valence
angles

f#, k 1 cosmk #
k

torsion k1
angles

X Aij Bij qi qj

6
rij
r12
rij
ij
i<j

!
1

The first three terms describe potential energy variations due to bond (b),
valence angle (), and bond torsion angle (#) deformations. The remaining
(nonbonding) terms are Lennard-Jones and Coulomb interactions between
interaction sites separated by a distance rij j~ri
r~j j. Nonbonded interactions usually exclude pairs of sites belonging to the same bond or valence

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

110

Hentschke

angle. A model of this type is appropriate to study the dynamics of a system

with maybe 10,000 interaction centers in a time window of about 10 ns, on
current workstation computers. Figure 1 gives an impression of the typical
size of such a system. Figure 2 on the other hand shows how slow (global)
conformation changes may be, even for a short oligomer, compared to the
accessible time window.
Clearly, Eq. (1) is a crude approximation to molecular interactions. It
basically constitutes the simplest level on which the chemical structure of a
specific polymer can still be recognized. The development of more complex
molecular and macromolecular force fields is still ongoing, and the reader is
referred to the extensive literature (a good starting point is Ref. [2]).
Fully atomistic molecular dynamics simulations quickly require excessive
computer time as systems become large. This is despite the fact that the
computational effort for large n, governed by the calculation of the
nonbonded forces or interactions, scales as O(n) for short-range interactions
and as O(n ln n) for long-range interactions. This improvement over the
naively expected O(n2) behavior (assuming simple pairwise additive
interactions) is achieved by the use of various cell techniques, which are
discussed in this book. Actually, a more severe restriction than the limited

FIG. 1 Snapshot taken during the initial stage of a molecular dynamics simulation
of oligo(vinylpyrrolidone) (20mer) in an ionic solution. The box indicates the cubic
simulation volume. Periodic boundary conditions are used to extend the system to
(quasi-)infinity. The arrows indicate (from left to right) Na, C6H5SO

3 , Oligomer,
and water.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Single Chain in Solution

111

FIG. 2 A sequence of conformations obtained for the oligomer in Fig. 1 in pure

water at room temperature. The water molecules are omitted.

number of interaction centers is the time step in the integration algorithm,

which is governed by the highest frequency in the system. Here multipletime-step concepts [35] as well as the mapping of the real polymer onto a
much simpler model with less complicated interactions (coarse graining) are
used to speed up the calculations. Coarse graining reduces high frequencies
effectively but the reverse procedure, i.e., going from the simplified
model back to the specific polymer, is anything but straightforward [6,7].
Again, these techniques as well as sampling of conformation space with

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

112

Hentschke

sophisticated Monte Carlo techniques, are discussed in this book, and the
reader is referred to the pertinent chapters.
Force fields contain numerous parameters (here: fb, bo, f, o, f#, k , mk,  k,
Aij, Bij, and qi). Parameterization procedures employ training sets, e.g.,
amino acids in the case of proteins or relevant monomer-analogous
molecules in the case of technical polymers. The parameters are adjusted
using thermodynamic, spectroscopic, or structural data, and increasingly
quantum chemical calculations (most importantly in the case of the torsion
potential in the valence part of a force field) available for the training set.
Probably the most difficult parameters are the partial charges, qi, usually
located on the nuclei, since they are influenced by a comparatively large
number of atoms in their vicinity. Let us consider the case of an oligomer
in explicit solution (cf. Fig. 1). One may start with the vacuum values for
the qi determined for the training molecules via an ESP procedure
(ElectroStatic Potential fitting) based on fast semiempirical methods like
AM1 (Austin Model 1) [8]. Polarization effects due to solvent are included
roughly in a mean field sense via a scaling factor multiplying the charges.
The scaling factor may be determined by comparison with suitable
experimental data (e.g., the density of a solution consisting of the solvent
of interest and a monomer-analogous solute as function of temperature and
solute concentration [9]). The same scaling factor is subsequently applied to
the charges on the oligomer, which also can be determined in vacuum using
AM1 (possibly including conformational averaging, which, if it is done
correctly, of course requires information not available at this point). A more
systematic and maybe in the long run more promising approach is the
fluctuating charge method [10,11]. Here the qi are dynamic variables just like
the positions, r~i . Their equations of motion follow from the Lagrangian
L

1XX
1XX
2
mim r~_im
mq q_ 2im
Ufrg, fqg
2 m im
2 m im

The index m labels all molecules in the system, and im indicates the ith atom
or charge center in the mth molecule. The quantity mq is a mass parameter,
which, for simplicity, is the same for all charges. The potential energy is
Ufrg, fqg U non-Coulomb frg Efrg, fqg

Here {r} means all positions, and {q} means all partial charges. The first
term includes all non-Coulomb contributions [cf. Eq. (1)]. The second term,
the Coulomb part of the potential,
P is approximated
P in terms of a truncated
expansion in qi, E({r}, {q}) i Eio oi qi 12 i, j qi qj Jij . The quantities

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Single Chain in Solution

113

Eio and oi are single atom parameters, and Jij Jij rij; i , j is a Coulomb
integral (note: Jij ! r
1
ij ) computed from two charge distributions modrij !1

eled in terms of single Slater orbitals with the exponents i and j. The
equations of motion, i.e.,
~ r~ Ufrg, fqg
mim r~im
r
im

and
mq q im
im

m

Rt
P
P
P

1
follow from  t12 L
m lm im qim dt 0, where
im im
P the lm
Nm
are Lagrange parameters, i.e., the conditions im qim 0 ensure that the
total charge on a molecule is zero. The quantity im @E=@qim assumes
the role of the chemical potential of the charges, and Nm is the total
number of charge sites in the mth molecule.
The additional computational effort, compared to fixed partial charges, is
rather minor, i.e., about 15% for simulations of neat water [10,11] (the
authors consider the TIP4P and SPC/E water models). Figure 3 shows an
application of this technique to oligo(ethyleneoxide) in water [12]. This
simulation contained a 7mer immersed in 1000 water molecules at ambient

FIG. 3 Simulation of oligo(ethyleneoxide) in water using the fluctuating charge

model. Solid line: pair distribution function, g2 rOp
Ow , where rOp
Ow denotes the
distance between oxygen sites in the oligomer and in water. Broken line:
corresponding magnitude of the water dipole moment, H2 O .

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

114

Hentschke

conditions. The figure exemplifies the variation of the magnitude of the

water dipole moment corresponding to the structure in the Op
Ow pair
correlations. Here rOp
Ow is the separation between oligomer oxygen atoms
and water oxygen atoms. Note that the bulk dipole moment of this
polarizable water model is 2.7 D. (The calculation included long-range
interactions in terms of the particle mesh Ewald method.)

II. SOLVENT SPECIFIC POLYMER CONFORMATIONS

IN SOLUTION BASED ON OLIGOMER
SIMULATIONS
Probably the most powerful technique for sampling the conformation space
of real polymers is the transfer matrix approach in combination with the
RIS approximation discussed in the previous chapter. The approach was
pioneered by Flory (e.g., [13]), but its potential for generating configurational averages was recognized very early (cf. the reference to E. Montrol in
the seminal paper by Kramers and Wannier [14]). In the transfer matrix
approach the partition function of a polymer is expressed in terms of a
product of matrices, whose elements in the simplest case, e.g., polyethylene,
have the form
tij

 

 1



1 1




UM #i UMM #i , #j
UM #j
 exp

kB T 2
2

The lower indices indicate neighboring monomers along the chain, and the
upper indices indicate RIS values of the corresponding torsion angles #i .
Here the total potential energy of the polymer is
U

X
i

UM #i

UMM #i , #i1

where UM #i is the potential energy of the isolated monomer (assumed to

depend on #i only) and UMM #i , #i1 is the coupling between adjacent
monomers. For long chains the conformation free energy becomes
Fconf 
kBTNln l(max), where l(max) is the largest eigenvalue of the transfer
matrix, and N is the chain length.
This approach also allows the computation of correlation functions along
a polymer chain, e.g., the conditional probability, p#j#0 p##0 =p#0 ,
that a torsion angle #0 is followed by a torsion angle #. Here p#0 is the
probability of occurrence for the main chain torsion angle #0 . Similarly,
p##0 is the probability of occurrence for the neighbor pair ##0 . Assuming

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Single Chain in Solution

115

 m
the general case of m distinct rotational isomeric states # 1 this allows
the following simple construction procedure for polymer conformations:
1.
2.

3.

4.

Compute a random number z 2 [0,1].

Following the torsion angle #k extend the chain with
#1
#2
#3
..
.

if
if
if
..
.

z  P1jk
P1jk < z  P1jk P2jk
P1jk P2jk < z  P1jk P2jk P3jk
..
.

#m

if

P1jk    Pm
1jk < z

If #l is selected in (2) then generate the coordinates of the new bond

vector, b~ bb^, via b^i1
b^i cos  b^i  b^i
1 sin #l
b^i  b^i
1 
b^i cos #l . Note that all valence angles  are assumed to be identical,
and that the cis conformation defines # 0. Starting vectors may be
arbitrary but not parallel.
Continue from (1).

It should be noted that in general the conformation of a polymer backbone is characterized by a periodic sequence of nonequivalent torsion
angles, #A , #B , #C , . . . , #A , #B , . . . , rather than by merely one (type of)
torsion angle, #, as in the present example. Nevertheless, the generalization
is straightforward.
A Mathematica implementation of the above algorithm for m 3
(polyethylene) is:
CP[i_,j_]:PP[[i,j]]/P[[j]];
k1;
b1{1,0,0};
b2{0,1,0};
R{0,0,0};
f112/180 Pi;
tab[n_]:Line[Table[zRandom[ ];
If[z<CP[1,k],l1,If[z<CP[1,k]CP[2,k],l2,l3]];
kl;
t(2 l-1) Pi/3;
b3N[-b2 Cos[f] CrossProduct[b2,b1] Sin[t] CrossProduct[CrossProduct[b2,b1],b2] Cos[t]];
b3b3/Sqrt[b3.b3];
RRb3;
b1b2;
b2b3;
R,{n}]];
Show[Graphics3D[tab[10000]]]

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

116

Hentschke

CP is the above conditional probability computed from the normal probabilities P and PP. The indices i and j may assume the values 1, 2, and 3,
here corresponding to the rotational isomeric states gauche(
), trans, and
gauche().
An example for polyethylene, using the Jorgensen potential [15] to model
the potential in Eq. (7), is illustrated in the upper panel of Fig. 4. Here
n(r, r) is the average number of united atom carbon pairs, i.e., pairs of
effective methylene groups, divided by N, whose separation is between
r
r/2 and r r/2. Here N is the total number of methylene groups in
the chain. For small r the distribution is discrete. The first pronounced peak

FIG. 4 (Top) Monomermonomer distribution, n(r, r), here truncated at 1,

averaged over 100 conformations generated for N 1000 at T 140 C. The inset
shows a single conformation constructed using the Mathematica program in the text.
(Bottom) Corresponding Kratky plot of the scattering intensity, q2 I=Io N, vs. q for
different chain lengths, i.e., N 20 (a), 50 (b), 200 (c), 1000 (d). The crosses are
experimental data taken from [34].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Single Chain in Solution

117

occurs at r 1 corresponding to the bond length in the units used here [note:
n(1, r) 2]. As r increases the distribution becomes continuous. An
important quantity based on n(r, r) is the scattering intensity
imax
X
I
sinqri
nri, r
N 1
Io
qri
i1

!
8

allowing the direct comparison with X-ray and neutron scattering experiments. Here Io is a constant independent of the polymer structure, r r i,
and q is the magnitude of the scattering vector. For simplicity we have set
the form factors for each scattering site equal to one. The scattering intensity
corresponding to the above n(r, r) is shown in the lower panel of Fig. 4. At
large q all curves come together, because the relevant lengths ( 2q
1 ) are
small compared to any of the chain lengths considered. The effect of the
chain length is most pronounced for q values around 0.15, where a plateau
develops as N is increased. It is worth emphasizing that in general scattering
curves are more complex due to the more complex chemical architecture of
most polymers in comparison to polyethylene [16]. With some experience,
however, it is possible to relate many features of the scattering curve to
underlying conformational features, e.g., helical conformations. In the present very simple example there are two easily distinguishable q-regimes (for
large N), i.e., the linear increase beyond q  0.3, which reflects the simple
chemical structure of our model chain, and the initial increase with a
subsequent development of a plateau below q  0.3, which reflects the
statistical chain behavior. In this q-regime the scattering intensity of a
simple Gaussian chain is a good approximation to the shown scattering
intensity (for large N). The Gaussian chain in particular yields the plateau
q2 I=Io N ! 12=CN for large q (referring to the continuum limit, i.e., there
is no atomic structure no matter how large q is). The quantity


R2N
CN
Nb2

is the characteristic ratio, and RN is the end-to-end distance of the polymer

consisting of N monomers. Note that RN is a useful quantity to monitor
during an oligomer simulation if one wants to decide whether configurational equilibration is possible or not. CN, or rather its asymptotic value for
N ! 1, C1, is used commonly to characterize the shape or unperturbed
dimension of polymers in different solvents in terms of a single number.
Here the 1000mers come close to yielding a plateau at C1  C1000  7. One
may also obtain this value by direct evaluation of Eq. (9) using the above

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

118

Hentschke

construction procedure to generate a large number of single chain conformations. The above value actually corresponds to polyethylene in a melt,
an example which we have chosen for its simplicity. For most dilute polymer
solutions one obtains larger values usually between 10 and 20. Notice that
C1 obtained in this fashion describes the shape of a polymer on an intermediate length scale, on which self-avoidance has no significant effect yet.
Because the above model includes short-range interactions only, a chain
may intersect itself. Polymers modeled with this method eventually obey
hR2N i / N 2v with v 1/2.
Powerful as the transfer matrix approach is, it suffers from certain
deficiencies. Including coupling beyond nearest neighbors is difficult.
Accounting for solvent effects also is difficult. In a crude approximation
solvent effects may be included via an effective dielectric constant in a
Coulomb term integrated into the coupling in Eq. (7) [17], which itself could
be improved to include local solventpolymer interactions in terms of
effective interactions (e.g., [18]). On the other hand, it is of course possible to
extract the probabilities p# and p##0 directly from oligomer simulations
in explicit solvent, and feed them into a construction procedure like the one
outlined above [19]. Applications of this general idea to poly(isobutene)
in vacuum (!) or to poly(vinylpyrrolidone) in water are discussed in [20]
and [9]. Thus, other than in the usual transfer matrix calculations of these
probabilities, molecular solvent effects are included on the length scale of
the oligomer. Some care must be taken to diminish end effects, however.
Moreover it is possible to include couplings beyond nearest neighbors, i.e.,
next-nearest neighbor effects may be introduced analogously via conditional
probabilities of the type p#j#0 #00 p##0 #00 =p#0 #00 . An example can be
found again in [9].
Before leaving the subject we briefly mention one common approximate
calculation of CN. The central assumption is the uncoupling of adjacent
monomers along the chain, i.e., UMM 0, which allows us to derive the
expression




 

2

1

1 2
N
CN I hTi  I
hTi
hTi  I
hTi  I
hTi
10
N
1, 1
which is discussed in many polymer textbooks. Here I is a unit matrix, and
T, in cases where the polymer backbone contains one type of torsion angle
only (e.g., polyethylene), is given by
0
1

cos 
sin 
0
T @
cos # sin 
cos # cos 
sin # A
11

sin # sin 
sin # cos  cos #

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Single Chain in Solution

119

using the above convention. The index 1,1 indicates that CN is the 1,1element of the matrix. More complex polymer architectures may be accommodated by dividing the polymer into repeating sequences of effective bonds
(chemical units), again assumed to be independent, where T is the product
of the respective T-matrices for each effective bond. The conformation
averages (e.g., hcos i or hcos # sin i) again may be computed directly
from the corresponding oligomer simulations. Despite its crude approximations this method appears to yield reasonable first guesses for CN [21].
Frequently the molecular structure and dynamics of the polymersolvent
interface also are of interest. Structural information may be derived via pair
correlation functions of the general type considered in Fig. 3. For instance,
g2(r) may be calculated via g2(r) (4pr2r)
1n(r, r), where n(r, r) is the
average number of solvent atoms of a certain type within a spherical shell of
radius r and thickness r centered on a polymer atom, and  is the density
of the same solvent atoms in the bulk solvent. A dynamic quantity of
interest is the average residence time of a solvent atom or molecule in a
certain distance and/or orientation relative to a polymer atom or atom
group, e.g., hydrogen bonds between certain polar groups in a polymer and
water. Here n(t, te; r, r) is the average number of solvent atoms or molecules
again within a spherical shell (omitting the orientation dependence) at time
t, under the condition that the same solvent molecules were present in the
shell at t 0. The quantity te is an excursion time allowing a solvent
molecule to leave the shell and return within the excursion P
time without
max
n e
t=  ,
being discarded. The result may be fitted via nt, te ; r, r 1
which is sensible for max  2 (cf. Ref. [9]). For large max, i.e., a broad
distribution of residence times, , the dynamics assumes glasslike behavior.
There are important systems, where one is interested in the intrinsic
stiffness of polymers, which in solution possess a regular superstructure
most commonly helical conformations. Here the above approaches are
less useful, and a method called the segment method, is more applicable.
A periodic section of the helix of interest is simulated in explicit solvent
using molecular dynamics. The top and bottom of the segment are
connected via suitable periodic boundary conditions. These are different
from the usual periodic boundary conditions, because they include a
covalent bond of the backbone. The simulation box should be large enough
(a) to ensure that the solvent attains its bulk structure at large perpendicular
distances from the helix, and (b) that the important undulatory wavelengths
along the helix backbone should be included. Provided the simulation run is
long enough one may extract sufficient independent segment (backbone)
conformations to Build up realistic and solvent specific helix conformations of a desired molecular weight. Build up means that an existing piece
of the helix may be extended by an extracted segment provided that a certain

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

120

Hentschke

smoothness condition is fulfilled. The smoothness condition may be that the

last torsion angle in the helix matched the first corresponding torsion angle
in the segment fulfilling a certain accuracy condition. More complex
conditions involving additional constraints can be employed also of course.
Ideally there should be no correlation between k, where k is the kth segment
conformation extracted from the simulation, and l, where l is the position
of the segment along the constructed helix. Figure 5 shows high molecular
weight fragments of poly(-benzyl-L-glutamate) (PBLG), a helical polypeptide, which were constructed in this fashion from the trajectory of a
27 A-helix segment immersed in about 1000 dimethylformamide (DMF)
molecules [22].
The contour flexibility of molecular helices, such as in the case of PBLG,
is best described as persistent flexibility or homogeneous bend-elasticity.
The measure of this type of flexibility is the persistence length, P, defined via


u~0  u~s exp
s=P

12

FIG. 5 Four PBLG backbone conformations (IIV) constructed via the above
segment method. The inset illustrates the transformation of the explicit local
chemical structure into a persistent flexible rod.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Single Chain in Solution

121

where the u~ are unit vectors tangential to the backbone contour, and s is
their separation along the contour [23] (cf. the inset in Fig. 5). Note that P in
certain cases may be related to CN [24]. Via direct application of Eq. (12)
to contours like those shown in Fig. 5 one obtains P  1000 A for PBLG
in DMF at T 313K, in good accord with the experiments [22].

III. POLYMER CONFORMATIONS IN SOLUTION

VIA DIRECT SIMULATION
Langevin dynamics is a method to simulate molecules in contact with a heat
bath or solvent without considering the explicit structure of the solvent.
An important application of this approach is conformation sampling for
macromolecules in solution, because the expensive integration of the solvent
trajectories is omittedas well as solvent specific effects of course. Here the
equations of motion are given by
~
~
~_i
r~i m
1
i
ri Ueff Zi
i r

13

Ueff is a potential of mean force of the solvent molecules including the

explicit intramolecular interactions of the polymer. Z~i is a stochastic force
simulating collisions of polymer site i with the solvent molecules. It is
assumed that Z~i is uncorrelated with the positions and the velocity of the
sites i. Moreover Z~i has a Gaussian distribution centered on zero with
variance hZ~i t  Z~j t0 i 6mi i kB TB t
t0 ij , where TB is the temperature
of the solvent, and i is a friction parameter. In practice the components 
of Z~i are extracted from a Gaussian distribution with hZi,  tZi,  t0 i
2mi i kB TB =t, where t is the integration time step. A simple leapfrog
Verlet implementation of the equations of motion is the following






~ i Ueff Z~i t

r
1
1
u~ i t t u~i t
t 1
i t
t
2
2
mi
Ot2



1
r~i t t r~i t u~ i t t t Ot3
2

14

A detailed analysis of different algorithms in the context of Langevin

dynamics may be found in [25] (see also [26]).
A nice application of Eq. (14), which compares simulations for
poly(ethyleneoxide) in the explicit solvents water and benzene with
Langevin dynamics simulations, can be found in [27,28]. The authors use
a modified Stokes relation, i l 4p Ri,eff /mi, where l is a parameter

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

122

Hentschke

(0.01<l<0.1) related to the transition rates between the rotational isomeric

states, Ri,eff is an effective radius corresponding to the solvent accessible
surface of the atom or atom type, and  is the experimental viscosity
coefficient of the solvent. It is probably fair to say that Langevin dynamics
works for unspecific solvent interactions, whereas specific solvent interactions, e.g., hydrogen bonding, require explicit inclusion of molecular
solvent.
Langevin dynamics in its above form is useful for sampling the
conformations of a polymer, but it does not yield the correct dynamics.
To see this we consider a simple polymer chain consisting of N monomers
with masses m and total mass M Nm. Using i  for simplicity the
equation
of the P
polymer center of mass is r~cm
PN of motion
N ~
_
_

1

1
~
~cm or u~ cm M
ucm . Note that the net force
M
i1 Zi
 r
i1 Zi
~
due to the gradient term
in
Eq.
(13)
vanishes.
The analytic solution is
Rt
P
0
~ 0
~cm 0 0. Subsequent
u~cm t M
1 exp
t o dt0 N
i1 Zi t expt with u
partial integration yields r~cm t, and using the above expression for
hZ~i t  Z~j t0 i we obtain h~rcm t2 i 6Dcm t for large t. The diffusion
coefficient, Dcm kB TM
1 , is proportional to M
1. Experimentally,
however, one observes exponents in the range from 0.5 to 0.6 instead
of 1 [29]. The reason for this discrepancy is the neglect of hydrodynamic
interactions. The latter may be included via the Oseen tensor [30,31], leading
to a modified dynamics [32]. For a discussion of hydrodynamic effects in the
context of single chain dynamics see for instance [33].

REFERENCES
1. Allen, M.P.; Tildesley, D.J. Computer Simulation of Liquids; Oxford: Clarendon
Press, 1990.
2. Lipkowitz, K.B.; Boyd, D.B. Reviews in Computational Chemistry; Weinheim:
Wiley/VCH, 1990.
3. Street, W.B.; Tildesley, D.J.; Saville, G. Mol. Phys. 1978, 35, 639648.
4. Tuckerman, M.E.; Berne, B.J. J. Chem. Phys. 1991, 95, 83628364.
5. Scully, J.L.; Hermans, J. Mol. Simul. 1993, 11, 6777.
6. Tschop, W.; Kremer, K.; Hahn, O.; Batoulis, J.; Burger, T. Acta Polymerica
1998, 49, 7579.
7. Tschop, W.; Kremer, K.; Batoulis, J.; Burger, T.; Hahn, O. Acta Polymerica
1998, 49, 6174.
8. Leach, A.R. Molecular Modeling; Harlow: Addison Wesley Longman Limited,
1996.
9. Flebbe, T.; Hentschke, R.; Hadicke, E.; Schade, C. Macromol Theory Simul
1998, 7, 567577.
10. Rick, S.W.; Stuart, S.J.; Berne, B.J. J. Chem. Phys. 1994, 101, 61416156.
11. Rick, S.W.; Berne, B.J. J. Am. Chem. Soc. 1996, 118, 672679.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Single Chain in Solution

123

12. Stockelmann, E. Molekulardynamische Simulationen an ionischen

Grenzachen. Ph.D. dissertation, Universitat Mainz, Mainz, 1999.
13. Flory, P.J. Macromolecules 1974, 7, 381392.
14. Kramers, H.A.; Wannier, G.H. Phys. Rev. 1941, 60, 252262.
15. Jorgensen, W.L.; Madura, J.D.; Swenson, C.J. J. Am. Chem. Soc. 1984, 106,
66386646.
16. Kirste, R.G.; Oberthur, R.C. In X-Ray Diraction; Glatter, O., Kratky, O.,
Eds.; Academic Press: New York, 1982.
17. Tarazona, M.P.; Saiz, E.; Gargallo, L.; Radic, D. Makromol. Chem. Theory
Simul. 1993, 2, 697710.
18. Chandler, D.; Pratt, L.R. J. Chem. Phys. 1976, 65, 29252940.
19. Mattice, W.L.; Suter, U.W. Conformational Theory of Large Molecules;
John Wiley & Sons: New York, 1994.
20. Vacatello, M.; Yoon, D.Y. Macromolecules 1992, 25, 25022508.
21. Jung, B. Simulation der Kettenkonformation von Polymeren mit Hilfe
der Konzepte der Molekulardynamik-Rechnungen. Ph.D. dissertation,
Universitat Mainz, Mainz, 1989.
22. Helfrich, J.; Hentschke, R.; Apel, U.M. Macromolecules 1994, 27, 472482.
23. Landau, L.D.; Lifschitz, E.M. Statistische PhysikTeil 1; Akademie-Verlag:
Berlin, 1979.
24. Cantor, C.R.; Schimmel, P.R. Biophysical Chemistry. Part III: The Behavior of
Biological Macromolecules; W.H. Freeman and Company: New York, 1980.
25. Pastor, R.W.; Brooks, B.R.; Szabo, A. Mol. Phys. 1988, 65, 14091419.
26. Hunenberger, P.H.; vanGunsteren, W.F. In Computer Simulation of
Biomolecular
SystemsTheoretical
and
Experimental
Applications;
vanGunsteren, W.F., Weiner, P.K., Wilkinson, A.J., Eds.; Kluwer Academic
Publishers: Dordrecht, 1997.
27. Depner, M. Computersimulation von Modellen einzelner, realistischer
Polymerketten. Ph.D. dissertation, Universitat Mainz, Mainz, 1991.
28. Depner, M.; Schurmann, B.; Auriemma, F. Mol. Phys. 1991, 74, 715733.
29. Strobl, G. The Physics of Polymers; Heidelberg: Springer, 1997.
30. deGennes, P-G. Scaling Concepts in Polymer Physics. Ithaca: Cornell
University Press, 1979.
31. Doi, M.; Edwards, S.F. The Theory of Polymer Dynamics; Oxford: Clarendon
Press, 1986.
32. Ermak, D.L.; McCammon, J.A. J. Chem. Phys. 1978, 69, 13521360.
33. Dunweg, B.; Kremer, K. J. Chem. Phys. 1993, 99, 69836997.
34. Lieser, G.; Fischer, E.W.; Ibel, K. J. Polym. Sci., Polym. Lett. 1975, 13, 3943.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

4
Polymer Models on the Lattice
LLER
K. BINDER and M. MU
Mainz, Germany
J. BASCHNAGEL

I.

Johannes-Gutenberg-Universitat Mainz,

Institut Charles Sadron, Strasbourg, France

INTRODUCTION

Simple random walks (RWs) and self-avoiding random walks (SAWs) on

lattices have been used as models for the study of statistical properties of
long flexible polymer chains, since shortly after the important sampling
Monte Carlo method was devised [1]. Figure 1 explains the meaning of
RWs, of SAWs, and of a further variant, the nonreversal random walk
(NRRW), by giving examples of such walks on the square lattice [2]. The
idea is that each site of the lattice which is occupied by the walk is
interpreted as an (effective) monomer, and the lattice constant (of length a)
connecting two subsequent steps of the random walk may be taken as an
effective segment connecting two effective monomers. Thus, the lattice
parameter a would correspond physically to the Kuhnian step length bK.
For a RW, each further step is completely independent of the previous step
and, hence, the mean square end-to-end distance hR2i after N steps simply is


R2 a2 N

and the number of configurations (i.e., the partition function of the chain
ZN) is (z is the coordination number of the lattice)
ZN zN

Although the RW clearly is an extremely unrealistic model of a polymer

chain, it is a useful starting point for analytic calculations since any desired
property can be calculated exactly. For example, for N  1 the distribution

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

126

Binder et al.

FIG. 1 (a) Construction of a 22-step random walk (RW) on the square lattice.
Lattice sites are labeled in the order in which they are visited, starting out from the
origin (0). Each step consists in adding at random an elementary lattice vector
[(1, 0)a, (0, 1)a, where a is the lattice spacing] to the walk, as denoted by the
arrows. (b) Same as (a) but for a non-reversal random walk (NRRW), where
immediate reversals are forbidden. (c) Two examples of self-avoiding walks (SAWs),
where visiting any lattice site more than once is not allowed: trials where this
happens in a simple random sampling construction of the walk are discarded. From
Kremer and Binder [2].

of the end-to-end vector is Gaussian, as it is for the freely-jointed chain

(see Chapter 1), i.e., (in d 3 dimensions),
3
2
PR
2 R

!3=2

3 R2
exp
2 
2 R

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Polymer Models on the Lattice

127

which implies that the relative mean square fluctuation of R2 is a constant,

2R


R2 2
hR2 i2
hR2 i2

2
3

Equation (4) shows that one encounters a serious problem in computer

simulations of polymers, which is not present for many quantities in simple
systems (fluids of small molecules, Ising spin models, etc.; see Binder and
Heermann [3]), namely the lack of selfaveraging [4]. By selfaveraging
we mean that the relative mean square fluctuation of a property decreases to
zero when the number of degrees of freedom (i.e., the number of steps N
here) increases towards infinity. This property holds, e.g., for the internal
energy E in a simulation of a fluid consisting of N molecules or an Ising
magnet with N spins; the average energy is calculated as the average of the
Hamiltonian, E hHi, 2E  hH2 i
hHi2 =hHi2 / 1=N ! 0 as N ! 1,
at least for thermodynamic states away from phase transitions. This
selfaveraging does not hold for the mean square end-to-end distance of a
single polymer chain, as Eq. (4) shows for the RW (actually this is true for
more p
complicated
polymer models as well, only the constant R may differ

from 2=3, e.g., R  0.70 for the SAW in d 3 [3]).

This lack of selfaveraging is one of the reasons why one is forced to
use such extremely simplified models, as shown in Fig. 1, under certain
circumstances. If it is necessary for the physical property under study to
work with large chain lengths N or if a good accuracy is to be reached in the
estimation of hR2i (and/or the gyration radius Rg, single-chain scattering
function S(k) at wavevectors k of the order kRg  1, etc.), a huge number
of (statistically independent!) configurations of single chains need to be
generated in the simulation. This is because the relative error of R2 simply
is R/(number of generated chains)1/2. Correspondingly, high precision
estimates of R2 for very long chains exist only for simple lattice models
(see Sokal [5], also for a tutorial overview of programming techniques and
a comprehensive comparison of different algorithms).
Another motivation why lattice models are useful for polymer simulation
is the fact that analytical theories have often been based on lattice models
and the resulting concepts are well defined in the lattice context. For
instance, the FloryHuggins theory of polymer mixtures [6] constructs the
thermodynamic excess energy and the entropy of mixing precisely on the
basis of a (many chain-) lattice model of the type of Fig. 1. Experimentally
widely used quantities like the FloryHuggins parameter  have in the
lattice context a well-defined meaning, and thus a critical test of this theory
via simulations could be performed rather straightforwardly by Monte
Carlo simulation of a corresponding lattice model [7]. Similarly, the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

128

Binder et al.

GibbsDi Marzio entropy theory of the glass transition [8,9], also based on a
lattice description, has been critically examined by a corresponding
simulation recently [10]. The key point is that not only has the
configurational entropy a well-defined meaning for a lattice model, but
also algorithms can be devised to accurately sample the entropy S of manychain systems [10], while in general the entropy is not a straightforward
output of computer simulations [3]. In addition, the simulation can then
extract precisely the quantities used in the theoretical formulation from the
simulated model as well, and hence a stringent test of the theory with no
adjustable parameters whatsoever becomes possible.
A third motivation is that with present day computers even for medium
chain length N of the polymers it is very difficult to study collective longwavelength phenomena (associated with phase transitions, phase coexistence, etc.), since huge linear dimensions L of the simulation box and a huge
number Nm of (effective) monomers in the simulation are required. For
example, for a finite size scaling [11] study of the nematicisotropic phase
transition in a model for a melt of semiflexible polymers a lattice with
L 130 was used, and with the bond fluctuation model [12,13] at volume
fraction of about   1/2 occupied sites this choice allowed for 6865 chains
of length N 20, i.e., 137,300 effective monomers, to be simulated [14].
Similarly, for a study of interfaces between unmixed phases of polymer
blends a lattice model with 512  512  64 sites was used, i.e., more than 16
million sites, and more than a million effective monomers (32,768 chains
of length N 32) [15]. Dealing with such large systems for more realistic
off-lattice models would be still very difficult.
Finally, we emphasize that lattice models are a very useful testing ground
for new algorithms that allow more efficient simulations of polymer
configurations under suitable circumstances. Many of the algorithms that
now are in standard use also for off-lattice models of polymerssuch as the
slithering snake algorithm [16,17], the pivot algorithm [18,19], the
configurational bias algorithm [20] (see also Chapter 7 by T.S. Jain and
J.J. de Pablo), and the chain breaking algorithm [21,22]were first
invented and validated for lattice models.
Why is there a need for these many different algorithms at all? Of course,
there would be no need for these algorithms if one wanted to simulate only
the simple random walk of Fig. 1a. But in fact, this is not a useful model for
polymer chains in most cases, since it allows arbitrarily many effective
monomers to sit on top of each other, ignoring excluded volume interactions
completely. A slightly better variant of the RW is the NRRW, Fig. 1b,
where immediate reversals are forbidden, but it is possible that loops are
formed where the NRRW crosses itself. The NRRW is still straightforward
to programat each step there are z
1 choices to proceed, from which one

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Polymer Models on the Lattice

129

selects a choice at random. The partition function is hence simply

NRRW
z
1N , and also hR2i can be worked out exactly. The problem
ZN
is reminiscent of the RIS model of a single chain [23], and Chapter 2 by E.D.
Akten, W.L. Mattice, and U.W. Suter, where three choices occur at each
step, the long-range excluded volume along the backbone of a chain also
being ignored. One believes that the Gaussian statistics, Eq. (3), which holds
for both the RW and the NRRW, is true both in dense polymer melts, where
the excluded volume interaction between monomers of the same chain is
effectively screened by monomers of other chains [24] and in dilute solutions
at the so-called Theta temperature
, where the repulsive excluded volume
interaction is effectively compensated by an attractive interaction between
monomers. However, this does not mean that effective monomers can sit on
top of each other at any point of the lattice to faithfully model these
situations. In the cases of dense melts or theta solutions, Eq. (1) is not
directly valid, but rather involves a nontrivial constant C1,


R2 C1 a2 N,

N!1

Therefore, the construction of appropriate equilibrium configurations of

polymer chains in a lattice model is always a problem. For a dilute solution
under good solvent conditions the excluded volume interaction between all
monomers of the chain can no longer be ignored. It leads to a swelling of the
chain with respect to its size in the melt or in theta solution. Thus, Eq. (5)
does not hold, but has to be replaced by


0
R2 C1
a2 N 2 ,

N!1

where C1 is another constant, and the exponent  differs from 1/2,  0.59
in d 3 dimensions and  3/4 in d 2 dimensions [24,5]. The behavior of
Eq. (6) is reproduced by the SAW model (Fig. 1c), and one can also predict
the corresponding chain partition function,
ZN / N 
1 zN
eff ,

N!1

where zeff<z
1 is a kind of effective coordination number that depends

on the type of lattice, and the exponent  takes the values  43/32 in d 2
dimensions [5] and  1.16 in d 3 dimensions [25].
Equation (7) is the reason why the sampling of SAWs is so difficult. Note
that all configurations of chains with N steps on the lattice that obey the
excluded volume condition should have equal a priori probability to occur
in the generated sample. A straightforward way to realize this is simple
random sampling: One carries out a construction of a RW, as in Fig. 1a,

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

130

Binder et al.

or better of a NRRW (see Fig. 2 for pseudo-codes), and whenever the walk
intersects, thus violating the SAW constraint, the resulting configuration
has to be discarded, and a new trial configuration is generated (see Fig. 3 for
pseudo-code). Obviously, for large N the fraction of trials that is successful
becomes very small, since it is just given by the ratio of the respective
partition functions,

fraction of successful constructions

SAW
zeff N
ZN

1
/
N
NRRW
z
1
ZN

Since zeff<z
1 (e.g., zeff  2.6385 for the square lattice, zeff  4.6835 for
the simple cubic lattice, see Kremer and Binder [2] for various other lattices),
the fraction of successful constructions decreases exponentially with increasing N, namely as exp(

N) with

ln[zeff/(z
1)]  0.1284 (square lattice)
and 0.08539 (simple cubic lattice). This problem that the success rate
becomes exponentially small for large N is called the attrition problem,

is the so-called attrition constant.

Of course, the problem gets even worse when considering many chains on
a lattice: for any nonzero fraction  of occupied lattice sites, the success rate
of this simple sampling construction of self- and mutually avoiding walks
will decrease exponentially with increase in the volume of the lattice. As a
consequence, other methods for generating configurations of lattice models
of polymers are needed. We discuss some of the algorithms that have been
proposed in the following sections.

II. STATIC METHODS

A very interesting approach to overcoming the attrition problem was
already proposed by Rosenbluth and Rosenbluth in 1955 [26] (inversely
restricted sampling). The idea is to avoid failure of the simple sampling
construction of the SAW by not choosing at random blindly out of z
1
choices at each step, but by choosing the step in a biased way only out of
those choices that avoid failure of the construction at this step. Consider a
SAW of i steps on a lattice with coordination number z. To add the (i 1)th
step one first checks which of the z0 z
1 sites are actually empty. If ki
(with 0<ki  z0) sites are empty, one takes one of those with equal
probability 1/ki to continue the simple sampling construction. Only for
ki 0 is the walk terminated and one has to start from
Q the beginning. The
probability of each N-step walk then is PN fri g N
i1 1=ki , ri being the
site of the ith monomer. One immediately sees that dense configurations of
SAWs are more probable than less dense ones. To generate a sample of

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Polymer Models on the Lattice

131

FIG. 2 Pseudo-code for the simulation of a random walk (RW; code above the line)
and a non-reversal random walk (NRRW; code below the line) via simple sampling.
In both cases, the walks consist of N steps (innermost for loop), and the construction
of the walks is repeated M times (outer for loop) to improve statistics. Whereas these
repeated constructions always lead to a walk of N steps for the RW, they are not
always completed successfully for the NRRW because immediate back-folding can
occur. Then, the hitherto generated walk has to be discarded and the construction
must be resumed from the beginning (jump to start). In both cases, a new step is
appended to the walk by the subroutine add-next-step, in which ranf denotes a
random number uniformly distributed between 0 and 1 i.e., 0  ranf<1.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

132

Binder et al.

FIG. 3 Pseudo-code for the simulation of a self-avoiding random walk (SAW) by

simple sampling. For a SAW, it is necessary to keep track of all lattice sites that have
already been occupied by a monomer during the construction of the walk, since one
has to discard the hitherto obtained walk and restart the construction if the present
step attempts to place a monomer on an already occupied lattice site (jump to
start). A possibility to take into account the history of the walk is to define a
(large) (2N 1)(2N 1) lattice whose sites are initialized by 0 and updated to 1 if a
monomer is successfully added. For more details and further algorithms see Binder
and Heermann [3], Sokal [5], or the excellent textbook of Kinzel and Reents [62].

equally probable walks, one has to correct for this bias by not counting each
chain with the weight WN 1 in the sampling, as one would do for simple
sampling, but rather with a weight,
W N fr i g

N
Y

ki =z0

i1

For large N, this weight varies over many orders of magnitude, however,
and hence the analysis of the accuracy reached is rather difficult in practice.
In fact, Batoulis and Kremer [27] demonstrated that the expected errors
increase exponentially with increasing N.
In principle, this method can also be applied to multichain systems, but
the problem of correcting for the bias becomes even more severe. In practice,
one therefore has to resort to the configurational bias method which is
an extension of the Rosenbluth sampling (see Chapter 7). But, inversely
restricted sampling is still one of the possible options for not too large

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Polymer Models on the Lattice

133

chain length N and volume fraction  in order to generate a configuration of

a multichain system that can be used as a starting point for a dynamic
simulation method.
There are many variants of biased sampling methods for the SAW, for
instance the idea of scanning future steps [28] but these techniques will
remain outside of consideration here.
A different approach to overcoming attrition is the dimerization
method [29], i.e., the idea of assembling long chains out of successfully
generated short chains. Assume two (uncorrelated) walks of length N/2.
Both walks are taken out of the ZN/2 / (zeff)N/2(N/2)
1 different configurations. The probability that they form one of the ZN SAWs of N steps is
simply [2,5]
ZN
N 
1
22
1 N

1
P
2 /
N=22
1
ZN=2

10

Thus the acceptance rate decreases only with a (small) power of N rather
than exponentially.
Still another useful technique is the enrichment method [30]. This early
work presented the first and compelling evidence that the exponent 
[Eq. (6)] has the value   0.59 (d 3) and   0.75 (d 2), and hence is a
nice example of important discoveries made by Monte Carlo simulation.
The simple idea of the enrichment method is to overcome attrition by
using successfully generated short walks of length s not only once, but p
times: one tries to continue a chain of s steps to 2s steps by p different simple
sampling constructions. Then, the number N2s of 2s step chains is
N2s pNs exp

s

11

where
is the attrition constant. This process is continued for blocks of s
steps up to the desired chain length. One must fix p ahead of time to avoid a
bias. The best choice is p exp(
s), since taking p exp(
s) does not reduce
the attrition problem enough, while p> exp(
s) would lead to an exploding
size of the sample and all walks are highly correlated. The fact that the
generated chains are not fully statistically independent and that therefore
the judgement of the accuracy of the procedure is a subtle problem is the
main disadvantage of the method.
There exist many extensions and variants for all these algorithms. For
example, for a star polymer with s arms one can use an enrichment
technique where one tries to add one step at each arm, and at each size of the
star this attempt is repeated p times [31]. It is also possible to combine
various methods suitably together. For example, Rapaport [32] combined

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

134

Binder et al.

the dimerization method with the enrichment technique, while Grassberger

[33] combined the RosenbluthRosenbluth method with enrichment
techniques. With this so-called Pruned-Enriched Rosenbluth Method
(PERM) simulations of polymers at the Theta temperature up to a chain
length of a million were performed, in order to clarify the nature of
logarithmic correction factors to Eq. (1) predicted by renormalization group
theory [34]. Note that the SAW, modeling a swollen chain in good solvent,
can be extended from an athermal situation to describe temperaturedependent properties by adding, for instance, a nearest neighbor energy "
that occurs if two (nonbonded) effective monomers occupy adjacent sites at
the lattice. If a chain has n such nearest neighbor contacts in a sample
generated by simple sampling, one takes the statistical weight of this chain
configuration as proportional to exp(n"/kBT).
In PERM, enrichment is implemented in the RosenbluthRosenbluth
framework by monitoring the weight Wn of partially grown chains of n
steps. If Wn exceeds some preselected upper threshold Wn> , we make two or
more copies of the chain, divide Wn by the number of copies made, place all
except one onto a stack and continue with the last copy. In this way the total
weight is exactly preserved, but it is more evenly spread on several
configurations. This is done at every chain length n.
The last entry to the stack is fetched if the current chain has reached its
maximal length N, or if we prune it . Pruning (the opposite of enrichment)
is done when the current weight has dropped below some lower threshold
Wn< . If this happens, a random number rn with prob {rn 0}prob {rn 1}
1/2 is drawn. If rn 1, we keep the chain but double its weight. If not, it is
discarded (pruned), and one continues with the last entry on the stack. If
the stack is empty, a new chain is started. When the latter happens, one says
a new tour is started. Chains within one tour are correlated, but chains
from different tours are uncorrelated.

III. DYNAMIC METHODS

Dynamic Monte Carlo methods are based on stochastic Markov processes
where subsequent configurations X are generated from the previous one
(X1 ! X2 ! X3    ) with some transition probability W(X1 ! X2). To some
extent, the choice of the basic move X1 ! X2 is arbitrary. Various methods,
as shown in Fig. 4, just differ in the choice of the basic unit of motion.
Furthermore, W is not uniquely defined: We only require the principle of
detailed balance with the equilibrium distribution Peq(X),
Peq XW X1 ! X2 Peq X2 W X2 ! X1

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

12

Polymer Models on the Lattice

135

FIG. 4 Various examples of dynamic Monte Carlo algorithms for SAWs: bonds
indicated as wavy lines are moved to new positions (broken lines), other bonds are
not moved. (a) Generalized VerdierStockmayer algorithms on the simple cubic
lattice showing three types of motion: end-bond motion, kink-jump motion,
crankshaft motion; (b) slithering snake (reptation) algorithm; (c) pivot
(wiggle) algorithm.

In the athermal case (the standard SAW problem) each configuration has
exactly the same weight (for single chains the normalized probability simply
is Peq(X) 1/ZN). Then, Eq. (12) says that the probability to select a motion
X1 ! X2 must be the same as the probabilty for the inverse motion,
X01 ! X1 . One has to be very careful to preserve this symmetry in the
actual realization of the algorithm, in particular if different types of move
are carried out in the simulation (e.g., Fig. 4). This can be achieved by
randomly choosing one of the possible moves and the new lattice sites to
which it should shift the monomer (or monomers), and by rejecting this
choice if it violates the excluded volume constraint. It would be completely
wrong if one first checked for empty sites and made a random choice only
between such moves that lead the monomers to these sites.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

136

Binder et al.

If there is an additional energy H(X) in the problem depending on the

configuration
P X, the equilibrium distribution is Peq(X) exp(
H(X)/kBT)/Z,
where Z x exp(
H(X)/kBT). Hence, Eq. (12) leads to (H  H(X2)

H(X1) is the energy change caused by the move)



WX1 ! X2
H
exp

WX2 ! X1
kB T

13

Following Metropolis et al. [1] one can take the same transition probability
as in the athermal case if H  0, but multiplied by a factor exp(
H/kBT)
if H > 0. Then, Eq. (13) is automatically satisfied at finite temperature,
if it was fulfilled in the athermal case.
Thus, at every step of the algorithm one performs a trial move X1 ! X2.
If W(X1 ! X2) is zero (excluded volume restriction being violated), the move
is not carried out, and the old configuration is counted once more in the
averaging. If W(X1 ! X2) is unity, the new configuration is accepted,
counted in the averaging, and becomes the old configuration for the next
step. If 0<W<1 we need a (pseudo-) random number x uniformly
distributed between zero and one. We compare x with W: If W  x,
we accept the new configuration and count it, while we reject the trial
configuration and count the old configuration once more if W<x.
In the limit where the number of configurations M generated tends to
infinity, the distribution of states X obtained by this procedure is
proportional to the equilibrium distribution Peq(X), provided there is no
problem with the ergodicity of the algorithm (this point will be discussed
later). Then, the canonical average of any observable A(X) is approximated
by a simple arithmetic average,
hA i  A

M
X
1
AX
M
M0 M

14

where we have anticipated that the first M0 configurations, which are not yet
characteristic of the thermal equilibrium state that one wishes to simulate,
are eliminated from the average. Both the judgements of how large M0
should be taken and how large M needs to be chosen to reach some desired
statistical accuracy of the result, are hard to make in many cases. Some
guidance for this judgement comes from the dynamic interpretation of the
Metropolis algorithm [5,35], to which we turn next. This interpretation also
is very useful to evaluate the efficiency of algorithms.
e with the label  of
We associate a (pseudo-) time variable t0  =N
e is the total number of
successively generated configurations, where N

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Polymer Models on the Lattice

137

e (chain length N 1)  (number of chains nc).

monomers in the system, N
e
e, and Eq. (14) becomes a time average,
Then, t0 M0 =N , t M=N
1
A
t
t0

At0 dt0

15

t0

Since a move X1 ! X2 typically involves a motion of a single monomer or

of a few monomers only (see Fig. 4), we have chosen one Monte Carlo step
(MCS) per monomer as a time unit, i.e., every monomer has on average one
chance to move.
The precise interpretation of the dynamics associated with the Monte
Carlo procedure is that it is a numerical realization of a Markov process
described by a master equation for the probabilty P(X, t) that a
configuration X occurs at time t,
X
X
d
PX, t

W X1 ! X2 PX, t
W X2 ! X1 PX2 , t
dt
X
X
2

16

Obviously, the principle of detailed balance, Eq. (13), suffices to guarantee

that Peq(X) is the steady-state solution of Eq. (16). If all states are mutually
accessible, P(X, t) must relax towards Peq(X) as t ! 1 irrespective of the
initial condition.
This dynamic interpretation is useful in two respects: (i) One can deal to
some extent with the dynamics of polymeric systems [36]. For example, a
basic model is the Rouse model [37] describing the Brownian motion of a
chain in a heat bath. The heat bath is believed to induce locally stochastic
changes of the configuration of a chain. In a lattice model, this is
qualitatively taken into account by motions of the types shown in Fig. 4a.
Since the Rouse model is believed to describe the dynamics of not too long
real chains in melts, there is real interest in investigating the dynamic
properties of models such as that shown in Fig. 4a, using lattices filled
rather densely with many chains. In fact, the related bond fluctuation
model [12,13,38,39], see Fig. 5, has been used successfully to study the
dynamics of glass-forming melts [40,41], the Rouse to reptation crossover
[4244] etc. (ii) One can understand the magnitude of statistical errors (see
below).
The fact that the algorithm in Fig. 4a should be compatible with the
Rouse model can be understood from a simple heuristic argument: As a
result of only local motion of beads in a chain, we expect that the center of
gravity moves a distance of the order of a/N, where a is a vector connecting
two nearest neighbor sites on the lattice, and whose direction is random.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

138

Binder et al.

FIG. 5 Schematic illustration of the bond fluctuation model in three dimensions.

An effective monomer blocks a cube containing eight lattice sites for occupation by
other monomers. The length of the bonds connecting
cubes along
p
p p two neighboring
the chain must be taken from the set 2, 5, 6, 3, and 10 lattice spacings.
Chain configurations relax by random diffusive hops of the effective monomers by
one lattice spacing in a randomly chosen lattice direction. For the set of bond vectors
possible for the bond lengths , quoted above, the excluded volume constraint that
no lattice site can belong to more than one cube automatically ensures that bonds
can never cross each other in the course of their random motions. From Deutsch and
Binder [13].

These random displacements add up diffusively. If we define the relaxation

time N so that after NN such motions a mean square displacement of the
order of the end-to-end distance is reached, we have from Eq. (6)

a=N 2 N N R2 C1 a2 N 2 ) N / N 21

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

17

Polymer Models on the Lattice

139

Note that the unit of time has been chosen such that each monomer on
average attempts to move once, and we have implied that the slowest modes
involve long distance properties of the order of the chains linear dimension.
Thus, the chain configuration should be fully
relaxed when the center of
p
mass has traveled a distance of the order of hR2 i. This argument implies
that the diffusion constant of the chain scales with chain length as
DN / N
1

18

irrespective of excluded volume restrictions.

However, one must be rather careful with such arguments and with
the use of algorithms as shown in Fig. 4a in general. For instance, if only
end-bond and kink-jump motions are permitted, as in the original
VerdierStockmayer algorithm [45], Eq. (17) does not hold for the SAW,
rather one finds reptation-like behavior N / N3 [46]. As pointed out by
Hilhorst and Deutsch [47], the kink-jump move only exchanges neighboring
bond vectors in a chain and does not create any new onesnew bond
vectors have to diffuse in from the chain ends, explaining that one then finds
a reptation-like law [24]. In fact, if one simulated a ring polymer with an
algorithm containing only the kink-jump move, the bond vectors contained
in the initial configuration would persist throughout the simulation, and
clearly good equilibrium could not be established! Thus, when using a new
algorithm one has to check carefully whether it has some undesirable
conservation laws.
When considering the slithering snake algorithm of Fig. 4b, one can
argue by an analogous argument as presented above that the center of mass
vector moves a distance of kRk/N at each attempted move. If we require
again that the mean square displacement is of order hR2i at time N ( NN
attempted moves), we find (hR2i/N2) NN hR2i, i.e., N / N and DN is of
order unity. This argument suggests that the slithering snake algorithm
is faster by a factor of N compared to any algorithm that updates the
monomer positions only locally (kink-jump, etc.).
While such a faster relaxation is desirable when one is just interested
in the simulation of static equilibrium properties, the slithering snake
algorithm obviously is not a faithful representation of any real polymer
dynamics even in a coarse-grained sense. This argument holds for the pivot
algorithm (Fig. 4c) a fortiori: Due to the global moves (i.e., large parts of
the chain move together in a single step) the relaxation of long-wavelength
properties of the coil is very fast, but the dynamics does not resemble the
real dynamics of macromolecules at all.
Nevertheless, the pivot algorithm has one distinctive advantage. It
exhibits very good ergodicity properties [5], while slithering snake and

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

140

Binder et al.

FIG. 6 (a) Example of a SAW on the square lattice that cannot move if the
slithering snake algorithm is used. (b) Example of a SAW on the square lattice
that cannot move if a combination of the slithering snake algorithm and the
VerdierStockmayer algorithm is used.

VerdierStockmayer type algorithms are manifestly nonergodic [48].

One can show this failure of ergodicity by constructing counterexample
configurations that cannot relax at all by the chosen moves (Fig. 6). Such
configurations also cannot be reached by the algorithms shown in Fig. 4a
and b, and thus represent pockets of phase space simply left out
completely from the sampling. There is good evidence [2,7,49] that in
practice the systematic errors induced by this lack of ergodicity are rather
small and are even negligible in comparison with the statistical errors of
typical calculations. Nevertheless, in principle the lack of full ergodicity is
a serious concern. While the pivot algorithm for single-chain simulations
is the method of choice, both because of this problem and because it yields a
very fast relaxation, it clearly cannot be used for simulations of dense
polymer systems, and therefore algorithms such as those of Fig. 4a and b are
still in use.
Irrespective of whether the chosen moves (Fig. 4) resemble the physically
occurring dynamics (case (a)) or not (cases (b), (c)), Eqs. (15), (16) always
apply and imply that observations of observables A, B are dynamically
correlated. A dynamic correlation function, defined as


A0Bt
hAihBi
AB t
hABi
hAihBi

19

is implemented in a Monte Carlo simulation using the estimate (tobs total

observation time)


A0Bt  A0Bt

1
tobs
t
t0

tobs
t

At Bt t dt
t0

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

20

Polymer Models on the Lattice

141

Even if one is not interested in the study of these correlations per se, they are
important because they control the statistical error of static observables.
Thus, if we estimate the error A of an average A [Eqs. (14), (15)] as follows
(there are n observations of A, i.e., A(t ) A , 1, . . . , n)

!2
n 

1X
A  A
A
A
n 1
2

21

one can to show that [35,3]

 1  2 
n
AA o
A
hA i2 1 2
A2 
,
n
t

for

tobs nt  AA

22

where t is the time interval between subsequent observables included in the

sum in Eq. (21) and the relaxation time is defined as follows
Z

AA

AA t dt

23

Applying these concepts to the sampling of the end-to-end vector R, for

instance, we recognize that statistically independent observations for R are
only obtained if t> N, a time that can be very large for very long chains
{cf. Eq. (17)}. In the simulation of models for dense polymer systems with
local algorithms of the type of Fig. 2(a) or Fig. 3, the reptation model [24,50]
implies that N / N3, and in addition, the prefactor in this law can be rather
large, because the acceptance rate of all moves decreases very strongly with
increasing density of occupied sites in the lattice. Thus, in practice one has to
work with a volume fraction  of occupied lattices that does not exceed 0.8
for the simple SAW model on the simple cubic lattice [7] or 0.6 (d 3) or 0.8
(d 2) for the bond fluctuation model [12,13]. If one wishes to work with
higher volume fractions (or with the fully occupied lattice,  1, in the
extreme case), no motion at all would be possible with any of the algorithms
shown in Figs. 4 and 5. Then the only alternative is to use either bondbreaking algorithms [21,22], which have the disadvantage that one does not
work with strictly monodisperse chain length, or the cooperative motion
algorithm (CMA), see the following Chapter 5 by T. Pakula. Neither of
these algorithms has any correspondence with the actual chain dynamics,
of course.
An intricate problem is how to obtain initial configurations of the system
of chains for the dynamic Monte Carlo algorithms. Several recipes have

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

142

Binder et al.

been advocated in the literature. For a system of many short chains, where
the linear dimension of the box L is much larger than the length Rmax bN
of a chain fully stretched to its maximal extension, one can put all chains in
this fully stretched state parallel to each other in a lattice direction onto the
lattice. Excluded volume interactions can then be trivially respected. The
time t0 necessary to forget the initial state that is not characteristic of
equilibrium can then simply be measured via the decay of the orientational
order of the chains, which is initially present in the system. Although in
principle this is a good method, it is clear that for large N and  close to
volume fractions that correspond to melt densities (i.e.,  0 0.5) the
resulting time t0 is rather large. But the most important drawback is that it
cannot be used at all for systems wherepbN
exceeds L (note that in the bond
fluctuation model b is between 2 and 10 lattice spacings, and simulations
up to N 2048 for semidilute solutions using lattice linear dimensions
L  400 have been carried out [51]). In this case one has the following
alternatives: Either the chains in the initial state are put on the lattice with
the RosenbluthRosenbluth method or the configurational bias method, or
one uses NRRW chains, ignoring excluded volume initially. Then during the
initial equilibrium run only moves are accepted that do not increase the
number N of lattice sites for which the excluded volume constraint is
violated, and one has to run the simulation until N 0. There is no
guarantee, of course, that this latter algorithm works at all for every initial
configuration. An alternative would be to replace the hard excluded volume
constraint by a soft one, introducing an energy penalty U for every double
occupancy of a lattice site, and gradually increase U ! 1 during the
initial equilibration run. Despite the fundamental importance of producing
well-equilibrated initial configurations of many-chain (athermal) systems,
we are not aware of any systematic comparative evaluation of the efficiency
of the various approaches outlined above.
It should also be emphasized that the best method of choosing an initial
state for the simulation of a dense polymer system may depend on the
physical state of the system that one wishes to simulate. For example, the
configurational bias method is a good choice for creating initial configurations of long flexible chains in semidilute solutions [51] or melts [43], but not
for melts of stiff chains in a nematic state [14]: In this case, it is much better
to start with a monodomain sample of a densely filled lattice of fully
stretched chains, remove at random chains such that the desired volume
fraction of occupied lattice sites is reached, and then start a relaxation run
with slithering snake and local moves. Conversely, if we start from a
configuration with no initial nematic order, it already takes a very long time
to create small nematic domains out of chain configurations that are initially
random-walk like, and relaxing the system further so that the multidomain

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Polymer Models on the Lattice

143

configuration turns into a nematic monodomain configuration would be

prohibitively difficult [14].
As the last point of this section, we briefly comment on several variations
of lattice models, such as the SAW (Fig. 4c) or the bond fluctuation model
(Fig. 5). What are the respective merits of these models?
The simple SAW has the advantage that it has been under study for more
than 40 years, and thus a large body of work exists, to which new studies can
be compared. For single short chains, exact enumerations have been carried
out, and for some Monte Carlo techniques, e.g., the PERM method, one can
proceed to particularly long chains. However, there are some disadvantages:
(i) In d 2 dimensions, the moves of Fig. 4a do not create new bond vectors,
and hence the local algorithm does not yield Rouse dynamics in d 2. (ii)
Branched chains (stars, polymer networks) cannot be simulated with the
dynamic Monte Carlo methods of Fig. 4, since the junction points cannot
move. (iii) Since there is a single bond length (one lattice spacing) and only a
few bond angles (0 and 90 on square or cubic lattices, respectively), there is
little variability in the model.
The bond fluctuation model has been used for about 12 years only, but it
has numerous advantages: (i) There is a single type of move in the random
hopping algorithm, namely an effective monomer is moved by a lattice
spacing in a randomly chosen lattice direction. It is easy to write a very fast
code that executes this move. This move almost always creates new bond
vectors, also in d 2 dimensions. (ii) If the set of bond vectors is suitably
restricted, the noncrossability constraint of bonds is automatically taken
into account. (iii) In simulations of branched chains, junction points can
move. (iv) Since there exist (in d 3 dimensions) 108 bond angles
, the
model in many respects mimics the more realistic behavior of off-lattice
models in continuous space. (v) The different choices of bond lengths b
allow introduction of an energy function U(b), and this opens the way to
modeling the glass transition of polymers [10], or carrying out a mapping
of specific polymers to bond fluctuation models with particular choices of
potentials U(b), V(
) [52]. (vi) While the random hopping algorithm of
Fig. 5 is not strictly ergodic, the class of configurations that are not sampled
is much smaller than that of the algorithms in Fig. 4a, b.
Finally, we also mention models that are in some way intermediate
between the SAW and the bond fluctuation model. For example, Shaffer
[53,54] has used a model where the monomer
takes
p
p only a lattice site like the
SAW, but the bond length can be 1, 2, and 3 lattice spacings, and then
moves can be allowed where bonds cross each other, junctions can move,
etc. This model is useful to separate the effect of excluded volume (site
occupancy) from entanglement constraints (noncrossability of chains) in
the dynamics [53,54]. Since because of the chain crossability a rather fast

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

144

Binder et al.

relaxation is provided, this diagonal bond model is also advocated for the
study of ordered phases in models of block copolymers [55].

IV. CONCLUDING REMARKS

Although lattice models of polymers are crudely simplified, they have the
big advantage that their study is often much less demanding in computer
resources than corresponding off-lattice models. Therefore, a huge activity
exists to study the phase behavior of polymer mixtures and block copolymer
melts [56,57], while corresponding studies using off-lattice models are scarce.
In this context, a variety of specialized techniques have been developed, to
extract quantities such as chemical potential, pressure, entropy, etc. from
lattice models, see e.g., [5860]. An important consideration in such
problems also is the best choice of statistical ensemble (e.g., semi-grand
canonical ensemble for mixtures, etc.) [7,56]. We refer the reader to the more
specialized literature for details.
The lattice models of polymers reach their limits when one wants to study
phenomena related to hydrodynamic flow. Although study of how chains
in polymer brushes are deformed by shear flow has been attempted, by
modeling the effect of this simply by assuming a smaller monomer jump rate
against the velocity field rather than along it [61], the validity of such
nonequilibrium Monte Carlo procedures is still uncertain. However, for
problems regarding chain configurations in equilibrium, thermodynamics
of polymers with various chemical architectures, and even the diffusive
relaxation in melts, lattice models still find useful applications.

ACKNOWLEDGMENTS
This chapter has benefited from many fruitful interactions with K. Kremer
and W. Paul.

REFERENCES
1. Metropolis, N.; Rosenbluth, A.W.; Rosenbluth, M.N.; Teller, A.H.; Teller, E.
J. Chem. Phys. 1953, 21, 1087.
2. Kremer, K.; Binder, K. Monte Carlo Simulations of Lattice Models for
Macromolecules. Comput. Phys. Rep. 1988, 7, 259.
3. Binder, K.; Heermann, D.W.; Monte Carlo Simulation in Statistical Physics. An
Introduction, Ed.; 3rd Ed, Springer: Berlin, 1997.
4. Milchev, A.; Binder, K.; Heermann, D.W. Z. Phys. 1986, B63, 521.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Polymer Models on the Lattice

145

5. Sokal, A.D. Monte Carlo Methods for the Self-Avoiding Walk. In Monte Carlo
and Molecular Dynamics Simulations in Polymer Science; Binder, K., Ed.;
Oxford University Press: New York, 1995; 47 pp.
6. Flory, P.J. Principles of Polymer Chemistry; Cornell University Press: Ithaca,
New York, 1953.
7. Sariban, A.; Binder, K. Macromolecules 1988, 21, 711.
8. Gibbs, J.H.; Di Marzio. J. Chem. Phys. 1958, 28, 373.
9. Gibbs, J.H.; Di Marzio, J.D. J. Chem. Phys. 1958, 28, 807.
10. Wolfgardt, M.; Baschnagel, J.; Paul, W.; Binder, K. Phys. Rev. 1996, E54,
1535.
11. Binder, K. In Computational Methods in Field Theory; Gausterer, H., Lang,
C.B., Eds.; Springer: Berlin-New York, 1992; 59 pp.
12. Carmesin, I.; Kremer, K. Macromolecules 1988, 21, 2819.
13. Deutsch, H.-P.; Binder, K. J. Chem. Phys. 1991, 94, 2294.
14. Weber, H.; Paul, W.; Binder, K.; Phys. Rev. 1999, E59, 2168.
15. Muller, M.; Binder, K.; Oed, W. J. Chem. Soc. Faraday Trans. 1995, 91, 2369.
16. Kron, A.K. Polymer Sci. USSR. 1965, 7, 1361.
17. Wall, F.T.; Mandel, F. J. Chem. Phys. 1975, 63, 4592.
18. Lal, M. Mol. Phys. 1969, 17, 57.
19. Madras, N.; Sokal, A.D. J. Stat. Phys. 1988, 50, 109.
20. Siepmann, I. Mol. Phys. 1990, 70, 1145.
21. Olaj, O.F.; Lantschbauer, W. Makromol. Chem., Rapid Commun. 1982,
3, 847.
22. Manseld, M.L. J. Chem. Phys. 1982, 77, 1554.
23. Mattice, W.L.; Suter, U.W. Conformational Theory of Large Molecules, the
Rotational Isomeric State Model in Macromolecular Systems; Wiley: New
York, 1994.
24. de Gennes, P.G. Scaling Concepts in Polymer Physics; Cornell University Press:
Ithaca, 1979.
25. Caracciolo, S.; Serena Causo, M.; Pelissetto, A. Phys. Rev. 1998, E57, R1215.
26. Rosenbluth, M.N.; Rosenbluth, A.W. J. Chem. Phys. 1955, 23, 356.
27. Batoulis, J.; Kremer, K. J. Phys. 1988, A21, 127.
28. Meirovitch, H. J. Phys. 1982, A15, 735.
29. Alexandrovicz, Z. J. Chem. Phys. 1969, 51, 561.
30. Wall, F.T.; Erpenbeck, J.J. J. Chem. Phys. 1959, 30, 634.
31. Ohno, K.; Binder, K. J. Stat. Phys. 1991, 64, 781.
32. Rapaport, D.C. J. Phys. 1985, A18, 113.
33. Grassberger, P. Phys. Rev. 1997, E56, 3682.
34. des Cloizeaux J.; Jannink, G. Polymers in Solution: Their Modeling and
Structure; Oxford University Press: Oxford, 1990.
35. Muller-Krumbhaar, H.; Binder, K. J. Stat. Phys. 1973, 8, 1.
36. Binder, K.; Paul, W. J. Polym. Sci. 1997, B35, 1.
37. Rouse, P.E. J. Chem. Phys. 1953, 21, 1272.
38. Wittmann, H.-P.; Kremer, K. Comp. Phys. Comm. 1990, 61, 309.
39. Wittmann, H.-P.; Kremer, K. Comp. Phys. Comm. 1991, 71, 343.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

146

Binder et al.

40. Paul, W.; Baschnagel, J. Monte Carlo Simulations of the Glass Transition of
Polymers. In: Monte Carlo and Molecular Dynamics Simulations in Polymer
Science; Binder, K., Ed.; Oxford Univeristy Press: New York, 1995; 307 pp.
41. Baschnagel, J.; Dynamic Properties of Polymer Melts above the Glass
Transition: Monte Carlo Simulation Results. In Structure and Properties of
Glassy Polymers, ACS Symposium Series No. 710; Tant, M.R., Hi, A.J.,
Eds.; American Chemical Society: Washington, DC, 1998; 53 pp.
42. Paul, W.; Binder, K.; Heermann, D.W.; Kremer, K. J. Phys. 1991, II (France)
1, 37.
43. Muller, M.; Wittmer, J.; Barrat, J.-L. Europhys. Lett. 2000, 52, 406412.
44. Kreer, T.; Baschnagel, J.; Muller, M.; Binder, K. Macromolecules 2001, 34,
11051117.
45. Verdier, P.H.; Stockmayer, W.H. J. Chem. Phys. 1962, 36, 227.
46. Verdier, P.H. J. Chem. Phys. 1966, 45, 2122.
47. Hilhorst, H.J.; Deutsch, J.M. J. Chem. Phys. 1975, 63, 5153.
48. Madras, N.; Sokal, A.D. J. Stat.Phys. 1987, 47, 573.
49. Wolfgardt, M.; Baschnagel, J.; Binder, K. J. Phys. 1995, II (France) 5, 1035.
50. Doi, M.; Edwards, S.F. Theory of Polymer Dynamics; Clarendon: Oxford 1986.
51. Muller, M.; Binder, K.; Schafer, L. Macromolecules 2000, 33, 4568.
52. Paul, W.; Binder, K.; Kremer, K.; Heermann, D.W. Macromolecules. 1991, 24,
6332.
53. Shaer, J.S. J. Chem. Phys. 1994, 101, 4205.
54. Shaer, J.S. J. Chem. Phys. 1995, 103, 761.
55. Dotera, T.; Hatano, A. J. Chem. Phys. 1996, 105, 8413.
56. Binder, K. Monte Carlo Studies of Polymer Blends and Block Copolymer
Thermodynamics. In Monte Carlo and Molecular Dynamics Simulations in
Polymer Science; Binder, K., Ed.; Oxford University Press: New York, 1995;
356 pp.
57. Binder, K.; Muller, M. Monte Carlo Simulations of Block Copolymers. Curr.
Opinion Colloid Interface Sci. 2000, 5, 315.
58. Dickman, R. J. Chem. Phys. 1987, 86, 2246.
59. Muller, M.; Paul, W. J. Chem. Phys. 1994, 100, 719.
60. Kumar, S.K.; Szleifer, I.; Panagiotopoulos, A. Phys. Rev. Lett. 1991, 66, 2935.
61. Lai, P.-Y.; Binder, K. J. Chem. Phys. 1993, 98, 2366.
62. Kinzel, W.; Reents, G. Physics by Computer; Springer: Berlin, 1998.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

5
Simulations on the Completely
Occupied Lattice
TADEUSZ PAKULA Max-Planck-Institute for Polymer Research, Mainz,
Germany, and Technical University of Lodz, Lodz, Poland

I.

INTRODUCTION

There is a large number of different methods (algorithms) used for simulation of polymers on the coarse grained molecular scale [16]. Models of
polymers considered in this range usually disregard the details of the chemical constitution of macromolecules and represent them as assemblies of
beads connected by nonbreakable bonds. In order to speed up recognition
of neighboring beads, the simplified polymers are often considered to be on
lattices with beads occupying lattice sites and bonds coinciding with lines
connecting neighboring sites. The methods used for simulation of the lattice
polymers can be considered within two groups. The first group includes
algorithms that can operate only in systems with a relatively large fraction
of lattice sites left free and the second group includes algorithms suitable for
lattice systems in which all lattice sites are occupied by molecular elements.
Whereas, the systems considered within the first group should be regarded
as lattice gases, the systems treated within the second group of methods can
be considered as lattice liquids. This reflects the differences in the mechanisms of molecular rearrangements used within these two groups to move the
systems through the phase space in order to reach equilibrium. The latter
problem concerns the physical nature of molecular rearrangements in dense
polymer systems and is related to the microscopic mechanism of motion in
molecular liquids. Unfortunately, this is not yet solved entirely. The most
popular picture is that a molecule, or a molecular segment in the case of
polymers, needs a free space in its neighborhood in order to make a translational step beyond the position usually occupied for quite a long time. Most
simulation methods assume this picture and consequently a relatively large
portion of the space in the form of free lattice sites has to be left free to allow

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

148

Pakula

a reasonable mobility [15]. On the other hand, there is only one simulation
method that assumes a cooperative nature of molecular rearrangements on
the local scale and which does not require such a reserve space to allow the
molecular mobility. The method, which uses the mechanism of cooperative
rearrangements for polymer systems is the Cooperative Motion Algorithm
(CMA) suggested originally in [6] and presented in improved form in subsequent publications [710]. A mechanism of this kind has been formulated
recently also for low molecular weight liquids, based on assumptions taking
into account both a dense packing of molecules interacting strongly due to
excluded volume and a condition of preservation of local continuity of the
system. The later version of the microscopic mechanism, called the Dynamic
Lattice Liquid (DLL) model, has been described in detail [1114].
The aim of this chapter is to present a background to the simulations of
molecular lattice systems on a completely filled lattice (the DLL model) as
well as to show some examples of application of the CMA method for
simulation of static and dynamic properties of various polymers.

II. THE DYNAMIC LATTICE LIQUID MODEL

Macroscopically, liquids differ from solids by the absence of rigidity and
from gases by having a tensile strength. On a microscopic level, this means
that the molecules in a liquid can move more easily than in a solid but they
remain condensed due to attractive interactions, which are almost negligible
in a gaseous phase. Owing to the dense packing of molecules, the dynamic
properties of liquids become complex and the relaxations extend over various, usually well distinguishable, time scales. On the short time scale, V,
the molecules oscillate around some quasi-fixed positions, being temporarily
caged by neighbors. It is believed that more extensive translational
motions of molecules take place on a much longer time scale, , due to
the breaking down of the cages by cooperative processes. Trajectories of
molecules consist, therefore, both of oscillatory components and of occasional longer range translational movements between subsequent quasifixed states, as illustrated in Fig. 1a. The macroscopic flow of the liquid is
related to the longer time scale ( ), in which the arrangement of molecules
becomes unstable because each molecule wanders through the material
changing neighbors during the translational motion steps.
Although the picture of motion in a liquid shown above is documented
by computer simulations of dense Lennard-Jones systems [15,16], it is not
quite clear under which conditions single diffusional steps can occur.
Theories of transport phenomena in liquids do not consider this problem
explicitly [1722].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Simulations on Completely Occupied Lattice

149

FIG. 1 Schematic illustration of a molecule trajectory in a liquid. (a) A trajectory

consisting of vibrations around quasi-localized states and occasional translational
steps. (b) Local correlated motions of neighboring molecules contributing to the
translational step of the single molecule by a cooperative rearrangement.

The model that has been proposed to answer this question is based on a
lattice structure and is called the Dynamic Lattice Liquid model [11,12]. The
positions of the molecules are regarded as coinciding with the lattice sites.
The lattice is, however, considered only as a coordination skeleton defining
the presence of nearest neighbors but not precisely the distances between
them. Under the condition of uniform and constant coordination (z), all
lattice sites are assumed to be occupied. It is assumed that the system has
some excess volume so that molecules have enough space to vibrate around
their positions defined by lattice sites but can hardly move over larger distances because all neighboring lattice sites are occupied. The vibrations are
assumed to take place with a mean frequency V 1/ V, related to the short
time scale ( V). Each large enough displacement of a molecule from the
mean position defined by the lattice is considered as an attempt to move
to the neighboring lattice site. For simplicity, the attempts are assumed to
take place only along the coordination lines but are independent and randomly distributed among z directions. Most of the attempts remain unsuccessful, because it is assumed that all the time the system remains quasicontinuous, which means that no holes of molecular sizes are generated
and multiple occupations of lattice sites are excluded (excluded volume
condition). The continuity condition of the system, for the vector field of

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

150

Pakula

attempted displacements r, can be written as follows

rJ

@
"
@t

with

"!0

where J is the current of displacing molecules and  is the density and should
not deviate considerably from 1 if all lattice sites remain occupied; " is
introduced as a small number allowing density fluctuations but excluding
generation of holes. Consequently, most of the attempted displacements
have to be compensated by a return to the initial position within the
period V. Only those attempts can be successful that coincide in such a
way, that along a path including more than two molecules the sum of the
displacements is close to zero. In the system considered, only the paths in a
form of closed loops can satisfy this condition (Fig. 1b). The probability
to find an element taking part in such coincidences will determine the
probability of longer range rearrangements and the slower time scale ( ).
A determination of this probability reduces to counting the number of
self-avoiding n-circuits on a given lattice, i.e., to a problem that has
already been extensively considered in the literature [23]. The generally
accepted result describing the probability to find the self-avoiding circuits
is given by
pn Bn
h n

where B is a lattice dependent constant, plays the role of an effective

coordination number (called the connective constant) of the lattice, and
the exponent h is positive and dependent on the dimensionality d of the
lattice, but is presumably largely independent of the detailed structure of
the lattice. Related theories predict only bounds for the exponent h (h  d/2).
More detailed information is obtained from enumerations performed for
various lattices [23].
The model discussed can easily be implemented as a dynamic Monte
Carlo algorithm. A system of beads on the fcc lattice is considered. The
beads occupy all lattice sites. It is assumed that the beads vibrating with a
certain frequency around the lattice sites attempt periodically to change
their positions towards nearest neighboring sites. The attempts are represented by a field of unit vectors, assigned to beads and pointing in a direction of attempted motion, chosen randomly. Attempts of all beads are
considered simultaneously.
An example of such an assignment of attempted directions of motion is
shown in Fig. 2, for a system of beads on a triangular lattice, taken here as a
two-dimensional illustration only. From the field of attempts represented in

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Simulations on Completely Occupied Lattice

151

FIG. 2 An illustration of the vector field representing attempts of molecular

displacements towards neighboring lattice sites in the lattice liquid model. Shaded
areas represent various local situations: (1) unsuccessful attempt when neighboring
elements try to move in the opposite direction, (2) unsuccessful attempt because the
element in the center will not be replaced by any of the neighbors, and (3) successful
attempts, in which each element replaces one of its neighbors. System elements
taking part in cooperative translational rearrangements are shown in black.

this way, all vectors that do not contribute to correlated sequences (circuits)
satisfying the continuity condition Eq. (1) are set to 0. This concerns, for
example, such situations as attempts of displacements in opposite directions
(e.g., area 1 in Fig. 2) or attempts of motion starting from lattice sites,
towards which any other bead is not trying to move at the same time
(e.g., area 2 in Fig. 2). What remains after this operation are vectors contributing to a number of closed loop traces, which are considered as paths of
possible rearrangements (e.g., areas 3 in Fig. 2). If the system is considered
as athermal, all possible rearrangements found are performed by shifting
beads along closed loop traces, each bead to the neighboring lattice site.
The above procedure, consisting of (1) generation of motion attempts, (2)
elimination of unsuccessful attempts, and (3) displacing beads within closed
loop paths, is considered as a Monte Carlo step and assumed to take place
within the time scale V. The procedure is exactly repeated in subsequent
time steps always with a new set of randomly chosen directions of attempted
displacements.
The system treated in this way can be regarded as provided with the
dynamics consisting of local vibrations and occasional diffusional steps
resulting from the coincidence of attempts of neighboring elements to
displace beyond the occupied positions. Within a longer time interval, this
kind of dynamics leads to displacements of individual beads along random
walk trajectories with steps distributed randomly in time. Small displacements related to vibrations of beads could be considered explicitly but are
neglected here, for simplicity.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

152

Pakula

Models of macromolecular melts are considered, in analogy to models of

simple liquids, as systems of structureless beads occupying lattice sites. In
the case of polymers, the beads are connected by bonds to form linear chains
or other more complex assemblies regarded as macromolecules. An extension of the algorithm to simulation of systems representing polymer melts
appears to be straightforward. The presence of bonds between beads influences only the second step of the described procedure. Assuming that the
bonds are not breakable, only attempts that would not lead to bond breaking can be considered as possible. Therefore, the selection of rearrangement
possibilities is made under this additional condition, i.e., after having found
all possible rearrangement paths, in parallel, as for nonbonded beads, all
those ones are rejected that would lead to a disruption of polymer chains.
All other steps in the algorithm are identical with those described for systems representing simple liquids. The procedure results in local cooperative
conformational rearrangements of polymers, which can take place simultaneously at various places of the system; however, within a single time step
one bead can be moved only once and only to a neighboring lattice site. It is
a characteristic feature of this algorithm that types of local conformational
changes are not precisely specified. All changes of chain conformations that
satisfy the assumed conditions of system continuity and nonbreakability of
bonds are allowed. This makes the algorithm nonspecific for any type of
polymer architecture. During motion, the identities of polymers given by
numbers and the sequences of beads within chains are preserved. Local
chain moves in this algorithm are probably similar to those in the CMA
but some other conformational rearrangements are probably also possible,
when two or more displacement loops influence neighboring parts of a chain
simultaneously (within a single time step).
Ergodicity has not been shown for any polymer algorithm. But for
dimers, it was shown, for instance, for the CMA [24,25]. Therefore, in the
DLL algorithm, which leads to the same types of cooperative moves,
the ergodicity can be assumed for dimers, as well. The additional types
of moves mentioned above can only improve the accessibility of various
conformational states.
The requirement of a detailed balance in the athermal polymer system,
as considered here, reduces to showing that the transition probabilities
between two neighboring states A and B are equal. In this algorithm,
two such states are always reversible and are separated by cooperative
rearrangements along loops of the same size and form but different
motion directions. Because loops consist of vectors that are pointing equally
probably at any direction, this condition is satisfied. Moreover, it remains
valid for any polymer system because the loops are independent of the
structure.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Simulations on Completely Occupied Lattice

153

Here, only the athermal version of the DLL model has been described. It
has been shown elsewhere [1113] that the model is able to represent the
temperature dependent dynamics of systems as well. This, however, requires
additional assumptions concerning free volume distribution and related
potential barriers for displacements of individual elements. It has been
demonstrated [1113] that in such a version the model can describe a broad
class of temperature dependencies of relaxation rates in supercooled liquids
with the variety of behavior ranging between the extremes described by the
free volume model on one side and the Arrhenius model on the other side.

III. THE COOPERATIVE MOTION ALGORITHM

Cooperative rearrangements based on displacements of system elements
(beads) along closed loops have been introduced originally for dense lattice
polymer models within the Cooperative Motion Algorithm (CMA). In this
algorithm rearrangements are performed along random self-avoiding closed
trajectories generated sequentially in randomly chosen places of the simulated systems. Each attempt to perform a rearrangement consists of a
random choice of an element in the system and of searching for a selfavoiding non-reversal random walk in the form of a closed loop, which
starts and ends at that element. Subsequent attempts are counted as time.
Any distortion in searching loops, such as a reversal attempt or a cross point
in the searching trajectory, breaks the attempt and a new attempt is started.
If a loop is found, all elements lying on the loop trajectory are translated by
one lattice site so that each element replaces its nearest neighbor along the
loop. In systems representing polymers, only rearrangements that do not
break bonds and do not change sequences of segments along individual
chains are accepted. Chains can rearrange only by displacements involving
conformational changes. As a result of a large number of such rearrangements, elements of the system move along random walk trajectories. The
motion takes place in the system in which all lattice sites are occupied but the
system remains continuous and the excluded volume condition is satisfied.
In this algorithm, both the time step definition and the distribution of
rearrangement sizes are different than in the previously described DLL
model. The time step in the CMA has usually been chosen as corresponding
to the number of attempts to perform cooperative rearrangements resulting
on average in one attempt per system element (bead). This can be considered
as comparable with the time step definition in the DLL algorithm in which
each element attempts to move within each time step. This involves, however, a difference in distributions of waiting times. A more distinct difference
between these two algorithms is in the distribution of sizes of performed

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

154

Pakula

rearrangements, which has only been tested precisely for an athermal nonpolymeric system. The distribution of rearrangement sizes within the DLL
algorithm is well described by Eq. (2) with parameters 0.847, h 3, and
B 0.4, whereas the distribution of rearrangement sizes in the CMA is
considerably broader. It has been established that the two distributions
differ by the factor n2. This difference results in much higher efficiency of
simulation by means of CMA, achieved however at the expense of some
simplifications concerning short time dynamics of the system with respect to
that in the DLL model.
On the other hand, it has been proven for melts of linear polymers that
switching between DLL and CMA does not influence static properties of the
system. The CMA, similarly to the DLL algorithm, involves only conformational changes within polymers which all the time preserve their identities
given by the number of elements and by their sequences in individual chains.
In the CMA as in the DLL, a variety of conformational rearrangements
results from the algorithm and cannot be precisely specified. An important
advantage of both DLL and CMA methods with respect to other simulation algorithms of polymers on lattices, consists in allowed moves of chain
fragments both along contours of chains and transversally to the chain
contours. This probably makes both algorithms ergodic.
The CMA has been applied successfully for simulation of static and
dynamic properties of a variety of dense complex polymeric systems. Some
examples of such applications are presented in subsequent sections.

IV. EXAMPLES OF APPLICATION

A. Melts of Linear Polymers
Systems representing melts of linear polymers have been simulated using
both DLL and CMA methods. Linear polymers of various lengths N have
as usual been represented as chains of beads connected by nonbreakable
bonds. The face-centered cubic lattice has been used in both cases with
chains occupying all lattice sites. Analysis of static properties of such systems has shown that, using both methods, Gaussian chain conformations
are generated as indicated by the characteristic scaling laws for chain dimensions represented by the mean square values of the end-to-end distances or
by radii of gyration (e.g., [26,27]).
Dynamic properties of such systems have been analyzed using the DLL
method in the range of short chains (up to N 32) [1114] and the CMA
for larger chains (up to N 800) [26,27]. In the DLL method the presence of bonds between beads imposes additional limits on rearrangement
possibilities, which can be considered as an additional reduction of the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Simulations on Completely Occupied Lattice

155

connectivity constant of the lattice. This results in a strong effect of the

number of bonds and consequently chain length on bead relaxation rates.
It has been shown that the chain length dependence of the end-to-end vector
relaxation time is much stronger, especially for short chains, than the Rouse
model and other simulations predict. On the other hand, when the chain
relaxation times are normalized by the segmental mobility described by S
they nearly satisfy the N2 dependence, as expected for the Rouse chains.
This suggests that the S vs. N dependence represents the effect of chain
length on local dynamics, which in the Rouse model is included in the
friction coefficient and is not considered explicitly. It has been tested experimentally that the effects observed in model systems are also seen in real
polymer melts of polyisobutylene samples with chain lengths in a range
comparable with that of the simulated systems [11,12]. The simulation
results used for the comparison with experiments have been obtained for
systems at athermal states. This means that the effects of the influence of
chain lengths on rates of segmental relaxation are obtained without any
additional assumptions, such as, for example, additional free volume effects
at chain ends [28]. In the model described, they are only caused by a reduction of the connectivity constant within the lattice involved by introduction
of nonbreakable bonds between lattice elements. Nevertheless, the range of
changes of segmental relaxation times with the chain length is comparable
with the corresponding range of changes observed in experiments.
In application to polymers, the DLL algorithm has many similarities with
the CMA [1114,26]. There is, however, an important difference between
them, consisting in a parallel treatment of all system elements in the case
of DLL in contrast to the sequential treatment in the CMA. This difference
results in differences in distributions of rearrangement sizes and consequently in some specific effects related to the local dynamics. The effects of
chain length on the local relaxation rates observed in systems simulated using
DLL are, therefore, not seen in systems simulated by means of the CMA. On
the other hand the higher computational efficiency of the CMA allows the
study of systems with longer chains and within a broader time range.
The dynamical properties of the model systems, with n chains of length
N, are usually characterized by the following quantities:
1.

the autocorrelation function of a vector representing bond

orientation
b t

N
1 XX
bi tbi 0
Nn n i

where bi are unit vectors representing bond orientation;

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

156

2.

Pakula

the autocorrelation function of the end-to-end vector of chains

R t

3.

1X
R0  Rt
n n

with end-to-end vectors R(0) and R(t) at time t 0 and t, respectively; and
the mean squared displacements of monomers and of the centers of
mass of chains
1 XX
2
rm t
rm 0
Nn n N
1X
hr2cm i
rcm t
rcm 0 2
n n

hr2m i

5
6

where rm(t) and rcm(t) are monomer and chain center of mass coordinates at time t, respectively.
The first two quantities allow determination of relaxation times of
corresponding objects in the model (bonds and chains) and the last two
allow determination of diffusion constants. Chain length dependencies of
the self-diffusion constant and of the relaxation times of bonds and chains
are presented in Fig. 3. These results show that the dynamic behavior of

FIG. 3 Chain length dependencies of (a) self-diffusion constants of chains and (b)
relaxation times of bonds and end-to-end vectors in model systems of linear chain
melts.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Simulations on Completely Occupied Lattice

157

simulated systems corresponds well to the behavior of real polymer melts, as

detected experimentally. In the simulation, the diffusion constants for long
chains reach the well known scaling dependence D  N
2 and the relaxation
times of chains reach the scaling law  N!, with !>3, as observed in chain
length dependencies of melt viscosities for long chains. Simulation of models
with chain lengths much longer than these presented here becomes unrealistic at the moment because of limits imposed by the calculation speed.
More detailed analysis of the static and dynamic behavior of polymer
melts simulated by means of the CMA has already been presented in
other publications [26,27], but this problem still remains not completely
understood and further results concerning the mechanism of motion of
polymer chains will be published elsewhere. Models considered as polymer
melts have been simulated by many other methods [4,29,30], in which, however, a dense packing of molecules under the excluded volume condition is
hard to achieve.

B. Melts of Macromolecules with Complex Topology

The advantages of the CMAits high efficiency and high flexibility in the
representation of complex molecular objectscan be well demonstrated in
simulations of melts of stars and microgels. Experimental observations of
melts of multiarm polymer stars [31,32] and microgels [33] generate questions concerning the dynamics of molecular rearrangements in such systems.
As an illustration of the application of the CMA to such systems, fragmentary results concerning the structure and dynamics of three systems illustrated in Fig. 4 are presented here. All three systems consist of nearly the
same number of monomer units (N 500) but the units are joined together
to form different molecular objects. In the first case they form a linear chain,
in the second a multiarm star (24 arms), and in the third case the linear chain
is crosslinked internally by additional bonds (25 per chain) to a kind of
microgel particle. The structure and dynamics of such systems in the melt
have been studied [34,35]. Figure 4 shows a comparison of mass distributions, (r), in such objects around their centers of mass as well as a comparison of the pair correlation functions, g(r), of the centers of mass of different
chains. The first quantity, (r), describes the averaged fraction of space
occupied by elements of a given molecule at the distance r from the mass
center of that molecule, and the second, g(r), describes the probability that
at the distance r from the mass center of one molecule a mass center of
another molecule will be found. A considerable difference in the structure of
stars and microgels in comparison to linear chains can be seen when these
characteristics are considered. Whereas, the linear chains are loosely coiled

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

158

Pakula

FIG. 4 Illustration of molecular objects of the same total chain length but different
topology and a comparison of intrachain segment densities (a) and center of mass
pair correlation functions (b) for three computer simulated melts with different
molecular objects: linear chains, multiarm stars, and microgels.

and interpenetrate each other, the stars and microgels constitute compact
objects with intramolecular density almost reaching the density of the
system. This means that the latter molecules do not interpenetrate each
other and remain at well correlated distances as indicated by the center
of mass pair correlation functions with distinct maxima at preferred
intermolecular distances (Fig. 4b).
The dynamics of the three systems is characterized in Fig. 5, where
various correlation functions are compared. For the linear chain system
the segment position (s) and the end-to-end vector (R) [see Eq. (4)]

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Simulations on Completely Occupied Lattice

159

FIG. 5 Dynamical properties of three simulated polymer melts with various molecules: (a) linear chains, (b) stars, and (c) microgels. Various correlation functions
are shown: s segment position autocorrelation, R end-to-end vector autocorrelation for linear chains, A arm end-to-end autocorrelation for stars, B
autocorrelation of microgel chain segments of length N 20, pos position
autocorrelation for all types of molecules.

autocorrelation functions represent the local and the longest relaxation

rates, respectively. The third correlation function (pos) describing a correlation of chains with their initial positions (at t 0) characterizes the
translational motion. The position correlation of a structural element, a
chain segment, or the whole molecule, is defined as

pos t

N
1 XX
Ci 0Ci t
Nn n i

where Ci 1 at all places occupied by a given structural element and Ci 0

everywhere else. For linear chains, this correlation with the initial position

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

160

Pakula

decays much faster than the orientation correlation of chains described by

the end-to-end vector relaxation. Quite different behavior than for linear
chains is observed both for the stars and microgels with the same local
mobility indicated by the bond autocorrelation. In the case of stars, the
orientation relaxation of arms [A, the autocorrelation function of the
vector connecting star center with arm end, defined analogously to R,
Eq. (4)] is much faster than the position correlation of the whole star. The
position correlation shows two well distinguishable modes, the first probably related to the shape relaxation and the second, the slower one, related
to the translational motion of stars. This indicates that the flow of melts of
stars should be dominated by rearrangements involving displacements of
stars, which have to be cooperative because of their well correlated positions. Microgels behave similarly to stars when the position autocorrelation
is considered. They differ, however, in the intramolecular flexibility.
Whereas the arms of stars relax relatively fast, chain elements of microgels
of corresponding length relax only when the whole molecule can relax by
changing position and orientation.
The results concerning multiarm stars and microgels have been presented
in detail elsewhere and the influence of parameters of their molecular structure, such as arm number and arm length in stars [34,35] or size and crosslink density in microgels, on the mechanism of motion are discussed. The
results presented here demonstrate, however, the possibilities of the simulation method used in studies of structure and dynamics of melts of these
complex molecules. The high efficiency of the CMA method has allowed
the first simulations of melts of such molecules. Other simulations of
dynamic properties of star molecules [36] have been performed for single
stars and in a rather narrow time range.

C. Block Copolymers
In order to illustrate the application of the CMA to simulate heterogeneous
systems, we present here results concerning properties of a diblock copolymer melt considered in a broad temperature range including both the homogeneous and the microphase separated states. Only symmetric diblock
copolymers, of composition f 0.5 of repulsively interacting comonomers
A and B, are shown here. The simulation, in this case, allows information
about the structure, dynamics, and thermodynamic properties of systems
[9,3740] to be obtained.
In the case of copolymers, two types of partially incompatible monomers
(A and B) are considered, characterized by direct interaction parameters "ij,
and often it is assumed that the energy of mixing is given only by the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Simulations on Completely Occupied Lattice

161

interaction of monomers of different types, i.e., that "AA "BB 0 and

"AB 1. The effective energy of a monomer Em is given by the sum of
"AB over z nearest neighbors and depends in this way on the local structure.
The moving of chain elements alters the local energy because the monomers contact new neighbors. In order to get an equilibrium state at a given
temperature the probability of motion is related to the interaction energy of
the monomers in the attempted position in each Monte Carlo step.
This means that at a given temperature, the Boltzmann factor p
exp(
Em, final/kBT ) is compared with a random number r, 0<r<1. If
p>r, the move is performed, and another motion is attempted. Under
such conditions, at low temperatures, the different types of monomers
tend to separate from each other in order to reduce the number of AB
contacts and consequently to reduce the energy. The simulation performed
in this way allows information about the structure, dynamics, and thermodynamic properties of systems to be obtained. It is worthwhile to notice that
the frequently used Metropolis method of equilibration of systems can lead
to a nonrealistic dynamics, therefore it should not be used in cases when the
dynamics is of interest.
Examples of temperature dependencies of the thermodynamic quantities
recorded during heating of the initially microphase separated system of a
symmetric diblock copolymer are shown in Figs. 6a and 6b. The following
quantities have been determined: (1) the energy of interaction of a monomer, Em, determined as the average of interactions of all monomer pairs at a
given temperature
Em

z
X

"kl i=z

i1

and (2) the specific heat calculated via the fluctuation-dissipation theorem
cV

hE 2 i
hEi2
kB T 2

where the brackets denote averages over energy of subsequent states

sampled during simulation of the system at constant temperature. The temperature at which a stepwise change in the energy and the corresponding
peak in the specific heat are observed is regarded as the temperature of the
order-to-disorder transition, TODT.
The nature of the transitions corresponding to structural changes
in copolymers can be well established from an analysis of distributions
of local concentrations, which are directly related to the free energy. An

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

162

Pakula

FIG. 6 Temperature dependencies of: (a) the average interaction energy per
monomer, (b) the specific heat, and (c) concentration distributions in small volume
elements consisting of the nearest neighbors of each chain segment. Characteristic
structures corresponding to various temperature ranges are illustrated.

example of such distributions for a symmetric diblock copolymer, in a broad

temperature range, is shown in Fig. 6c, by means of contour lines of equal
composition probability projected on the compositiontemperature plane.
Such contour plots reflect many details of the thermodynamics and structure of the system. It is easily seen that, at high temperatures, the system can
be considered as homogeneous because locally the most probable concentration corresponds to the nominal composition in the diblock. This is

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Simulations on Completely Occupied Lattice

163

changed at temperatures close to TODT where at first a plateau and later two
maxima corresponding to two coexisting phases are detected. At TODT, a
sudden change transforms the system to a state with well defined microphases indicated by the most probable local concentrations corresponding
to pure components. These results indicate three characteristic ranges of
thermodynamic behavior of the system assigned as (1) disordered, (2)
weakly segregated, and (3) strongly segregated regimes appearing with
decreasing temperature. Structures of simulated systems corresponding to
these regimes are illustrated in Fig. 6 by assuming different colors for different copolymer constituents.
The structure of the simulated block copolymer systems has been characterized in detail [3840]. Temperature dependencies of various structural
parameters have shown that all of them change in a characteristic way in
correspondence to TODT. The microphase separation in the diblock copolymer system is accompanied by chain extension. The chains of the diblock
copolymer start to extend at a temperature well above that of the transition
to the strongly segregated regime. This extension of chains is related also to
an increase of local orientation correlations, which appear well above the
transition temperature. On the other hand, the global orientation correlation factor remains zero at temperature above the microphase separation
transition and jumps to a finite value at the transition.
In order to get information about dynamic properties of the system
various quantities have been monitored with time at equilibrium states corresponding to various temperatures [3840]: the mean squared displacement
of monomers, hr2m i, the mean squared displacement of the center of mass of
chains, hr2cm i, the bond autocorrelation function, b(t)), the end-to-end
vector autocorrelation function, R(t), and the autocorrelation of the endto-end vector of the block, bl(t). On the basis of these correlation functions,
various quantities characterizing the dynamic properties of the systems can
be determined, i.e., the diffusion constant of chains and various relaxation
times corresponding to considered correlations.
Examples of various correlation functions for the diblock copolymer
system at high and at low temperatures are shown in Fig. 7. It has been
observed that at high temperatures (T/N 1), the systems behave like a
homogeneous melt. All correlation functions show a single step relaxation.
The fastest is the bond relaxation and the slowest is the chain relaxation
described by the end-to-end vector autocorrelation function. The relaxation
of the block is faster than the whole chain relaxation by a factor of approximately two. Such relations between various relaxation times in the disordered state of the copolymer can be regarded as confirmed experimentally
for some real systems, in which the dielectric spectroscopy allows distinction
of the different relaxation modes [41]. At low temperatures, drastic changes

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

164

Pakula

FIG. 7 Various correlation functions determined at various temperatures for the

symmetric diblock copolymer melt. The two extreme temperatures correspond to
the edges of the temperature range within which the system has been simulated. The
high temperature (T/N 1.0) corresponds to the homogeneous regime and the low
temperature (T/N 0.3) to the strongly segregated limit. The intermediate temperature (T/N 0.5) is only slightly higher than the temperature of the orderto-disorder transition for this system.

can be noticed for the dynamics of the block copolymer. At temperatures

T/N<0.45 (see Fig. 6) the diblock system is in the microphase separated
regime and most of the correlation functions determined show bifurcation
of the relaxation processes into fast and slow components. The fast components of chain, block, and concentration relaxations are attributed to the
almost unchanged in rate, but limited, relaxation of chains when fixed at the
AB interface and the slow components indicate the part of relaxation
coupled to the relaxation of the interface within uniformly ordered grains
with the lamellar morphology. The concentration relaxation becomes the
slowest one in such a state of the system. The dynamic behavior of diblock
copolymers is presented in detail and discussed in [3840], where the spectra
of various relaxation modes have been determined in order to compare
simulation results with dielectric spectra determined for real copolymer
systems in the vicinity of the microphase separation transition [41].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Simulations on Completely Occupied Lattice

165

The diffusion in the systems studied has been detected by monitoring in

time the mean squared displacements of monomers and centers of mass of
chains. Typical results for the diblock copolymer system are shown in Fig. 8.
They indicate that the short time displacement rates are not sensitive to
temperature but the long time displacements are influenced slightly by the
microphase separation. The self-diffusion constants of chains determined at
the long time limit are shown in the inset of Fig. 8a, where the effects of the
microphase separation in the diblock can be clearly noticed. The slowing
down observed at the microphase separation of the system is, however,
rather small and indicates a considerable mobility of chains left even
when the chains are confined at interfaces. The nature of this mobility has
been analyzed by monitoring the correlation between orientation of chain
axes and directions of chain displacements (Fig. 8b). It is established that

FIG. 8 (a) Mean square displacements of monomers and mean square displacements of chain centers of mass vs. time for the diblock copolymer system at various
temperatures. Temperature dependence of the self-diffusion constant of block
copolymer chains is shown in the inset. (b) Orientation correlation factor between
the end-to-end vector and the center of mass displacement vector of copolymer
chains at various temperatures above and below the ODT.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

166

Pakula

in the phase separated regime the chains diffuse easily in directions parallel
to the interface. This is indicated by negative values of the orientation
correlation factor and means that the limitations in mobility of diblock
chains imposed by the morphology concern only the diffusion through the
interface.
Besides the above presented example, the CMA has been applied for
simulation of various other copolymer systems with more complex topology
[42] of macromolecules and other distributions of comonomers along chains
[43,44]. In all cases, the high computational efficiency of the method has
enabled detailed information about the structure and dynamics to be
obtained, including also the microphase separated states, in which the
dynamics becomes considerably slower.

V. IMPLEMENTATION DETAILS
There are essentially several versions of algorithms that are based on similar
ideas of performing cooperative rearrangements in dense molecular systems.
All versions are equally suitable both for simulation of assemblies of nonbonded beads representing small molecules and for simulation of assemblies
of lattice structures, which mimic polymer skeletons with various complexities. Some details essential for implementation of these methods for complex polymers on the fcc lattice, taken as an example, will be described
here. The efficiency of an algorithm strongly depends on details concerning
methods used for description of system elements, methods of recognition
and description of systems states, methods of rearranging the elements,
and finally on the programming methods, which allow fast accessibility to
large data arrays.

A. Description and Generation of Model Systems

The architecture of complex polymers can be represented by simplified
models consisting of beads connected by nonbreakable bonds in a way
that corresponds to backbone contours of the macromolecules. Such molecules consist usually of a large number of beads assuming specific positions
within a complex bond skeleton characteristic for each type of macromolecule. In this simplified representation of macromolecular structures, sizes of
monomers are not distinguishable. With this approximation, the macromolecules can, however, be represented on lattices. The lattice plays the role of
a topological skeleton of space and allows fast identification of neighbors.
An example of such a representation of a linear macromolecule on the facecentered cubic (fcc) lattice is shown in Fig. 9a.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Simulations on Completely Occupied Lattice

167

FIG. 9 Illustration of details concerning description of polymer chains and coding

directions in the fcc lattice: (a) Fragment of a simplified linear polymer chain on the
fcc lattice with beads occupying lattice sites and bonds of constant length connecting
only nearest neighbors. The size of the model system is given by numbers nx, ny, and
nz of length units in the x, y, and z directions, respectively. (b) The coordination cell
of the fcc lattice illustrating the orientation code of the sitesite vectors. The vector
shown has the direction code cd 4. (c) Description of a linear chain of length nw by
means of numbering of beads and using two sets of vectors a and b pointing in
opposite directions for defining of the chain contour.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

168

Pakula

In a model system with all lattice sites occupied, each bead at a position
(x, y, z) is described by a sequence of numbers defining its position in the
chain, m(x, y, z), the chain to which it belongs, k(x, y, z), and the connectivity
with other beads by means of up to four vectors, a(x, y, z), b(x, y, z),
c(x, y, z), and d(x, y, z), which point towards the bonded neighbors. Vector
orientation is given by a code number, cd, assuming values from 1 to 13,
which describe the 12 possible orientations of the sitesite vectors in the 12
coordinated fcc lattice, as illustrated in Fig. 9b. The value d 13 is used as a
code for the vector of length equal to 0, which is used to describe chain ends
and single beads representing elements of simple liquid (e.g., solvent). An
example of a linear chain of length nw is illustrated in Fig. 9c. It is seen that
each bond is described by two vectors (a and b) pointing in opposite directions. This has the advantage that, at each position, (x, y, z), the connectivity
and consequently the local conformation can immediately be established.
Table 1 shows the conformation code, con(b,a), used for the description of
various angles between the neighboring bonds along the chain contour.
Positions of beads are described by the coordinates (x, y, z). It has appeared
useful to introduce matrices xnp, ynp, and znp by means of which coordinates (xn, yn, zn) of neighboring lattice sites in direction cd can be found as
xn xnp(x, cd ), yn ynp( y, cd ), and zn znp(z, cd ). These matrices can
consider system sizes and the type of boundary conditions (e.g., periodic).

TABLE 1

Conformation Code used in the DLL and CMA Simulations

Angle between
neighboring
bonds
Illustration
Code (con)


Range of application

9
>
>
>
>
>
>
>
=

180

120

90

60

b 13, m 1

a 13, m nw

a 13, b 13; free bead (e.g., solvent)

Chain interior
1<m<nw
a<13 and b<13

>
>
>
>
>
>
>
;

Ends of chains

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Simulations on Completely Occupied Lattice

169

The initial states of model systems can generally be obtained in two

ways: (1) by generating ordered structures of uniform polymers and
subsequent melting of the ordered systems or (2) by polymerization in
bulk, which can be performed according to various polymerization mechanisms leading in this way to systems with realistic nonuniformities of
molecular parameters. The first method has mainly been used in published papers [26,27,34,35,37], whereas the second method has been used
recently [45].

B. Implementation of the DLL Model

This is the simplest version of the cooperative algorithm. It has minimum
assumptions concerning types of moves and has no limits in complexity of
the macromolecular structures. Moreover, it represents probably the most
plausible dynamics on the local scale.
An implementation of the DLL model for nonbonded beads is extremely
simple. In an athermal case, it consists in periodic repetition of the following
sequence of steps: (1) generation of the vectors (one per bead) representing
attempts of bead motion to neighboring lattice sites, (2) recognition of
attempts forming circuits for which the continuity condition in the
simplest form applies, and (3) performing rearrangements by displacing
elements along the found circuits, each to a neighboring position. In the
non-athermal case and in the case when beads are bounded to more complex
structures (macromolecules), additional conditions immobilizing some
system elements should be taken into account before the second step
[1114]. One period including the above described steps is considered in
this algorithm as a unit of time. The second step of the procedure is the
step that controls the computational efficiency. It has been recognized that
the vectors of attempted moves form highly branched structures sporadically including closed loops. When the branches, which usually end with
situation 2 illustrated in Fig. 2, are declared as nonmobile, only the loops
remain. Further details of implementation are dependent on the type of
hardware and software used. Large fragments of the algorithm based on
this model (steps 1 and 3) can be vectorized, therefore an implementation on
a vector computer may be efficient as well.
It is worthwhile to mention that the DLL model can be considered as
defining a special purpose parallel computer. A large number of microprocessors, each controlling the logic of a single lattice site, when arranged
in an architecture of a spatial, highly coordinated lattice system could constitute a machine, the size of which and consequently the size of the simulated system will be limited less by the computation time than by the cost of

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

170

Pakula

corresponding investment. A microprocessor with the single lattice site logic

has already been designed [46] and further steps of corresponding development depend on availability of financial support.

C. The CMA (Cooperative Motion Algorithm)

In contrast to the parallel DLL algorithm the CMA, as most other known
lattice simulation methods, uses a sequential searching for motion possibilities in randomly chosen fragments of the system. A virtual point, called
the searching point is introduced to search for the mobile loops. The
searching procedure will be described here for a simple, two-dimensional
example (Fig. 10) of nonbonded beads on a triangular lattice. This procedure is applied periodically and consists of the following steps: (1) random
choice of the initial position of the searching point (1 in Fig. 10), (2) random
choice of the direction to one of the nearest neighbors (2 in Fig. 10), (3)
moving of a bead from position 2 to 1 (a temporary double occupation
defect is created in this way in position 1 and the site denoted as 2 becomes
temporarily empty), (4) moving the searching point to position 2, (5)
random choice of a new direction (e.g., vector pointing to position 3 in
Fig. 10), (6) bead from position 3 moves to position 2, (7) searching point
moves to position 3, (8) random choice of a new direction from position 3
(e.g., position 1), and (9) the bead from position 1 moves to position 3. With
the last step the loop is closed, each element along the loop has been moved
by one lattice sitesite distance, and the temporary defects are relaxed. The
procedure can be repeated again starting with the random choice of the new
searching point position. Not all attempted loops are successful as in the
illustrated example. The acceptability depends on further conditions
imposed on the type of random walk allowed. Various assumptions here
lead to various versions of the algorithm. The following possibilities have
been considered: (1) random walk, (2) non-reversal random walk, and (3)
self-avoiding walk. The case (3) leads to the dynamics which is the closest to
that of the DLL model, whereas the case (2) can be very efficiently used for
studies of the static properties of systems or for dynamic effects much slower
than the single local rearrangements.
For motion of chains, the searching becomes more complicated because
of the demands imposed on the system such as nonbreakability of chains,
conservation of sequences of beads along chains, conservation of chain
lengths and architecture, etc. There are two possibilities (illustrated in Fig.
11) to move chain parts satisfying the above conditions: (i) the searching
path is crossing the chain contour (Fig. 11a) and a displacement of the bead
involved is possible just by rotation of bonds connecting that bead to the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Simulations on Completely Occupied Lattice

171

FIG. 10 An example of a simple rearrangement loop as searched using the CMA

for a system of nonbonded beads (shown as rings) on a two-dimensional lattice
(thick points represent lattice sites). The virtual searching point is denoted by
means of a gray circle. The initial state is shown in the upper left figure. The
searching point is located at position 1 by a random choice and the direction to
position 2 is chosen randomly as well. Shifting the bead from position 2 to 1 creates
two temporary defects: a double occupancy at position 1 and a vacancy at position 2,
as illustrated in the upper right figure. In this state, the searching point is shifted to
position 2 and by chance the position 3 is chosen. As the lower right part illustrates
the bead from position 3 moves to position 2 and the searching point moves in the
opposite direction. Finally, again by chance from position 3, the position 1 is found
and the bead located originally at this position is moved to position 3. This closes the
loop and relaxes the temporary defects. As a result of such a single rearrangement the
state shown in the lower left figure is obtained, in which all lattice sites are occupied
but at positions 1, 2, and 3 the elements (beads) changed their positions with respect
to the original state replacing each other along the closed loop.

remaining chain structure and (ii) the searching path enters the chain contour and leaves it after a short walk along the chain (Fig. 11b). In the first
case, the motion can be performed without extending any bonds, whereas, in
the second case, in each step of the searching point, one and only one bond
is attempted to be extended but this extension must immediately be relaxed
by motion of other beads along the chain as long as the searching point
leaves the chain at places with suitable local conformations or through chain
ends. The latter forces the searching point to go along the chain contour for
some time. In fact, the first possibility can also be considered as a special

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

172

Pakula

FIG. 11 Illustration of two types of rearrangements, which are considered in a

system of beads bonded to chains: (a) the path of the searching point (gray arrows)
crosses the chain contour and the resulting rearrangement takes place by rotation of
bonds connecting the moved bead with the remaining parts of the chain and (b) the
path of the searching point enters the chain contour, goes locally along the chain,
and leaves it after shifting several beads. In both cases, the chain is neither broken
nor extended and the sequence of beads along the chain is preserved. The only result
is a local conformational change incorporating one or several beads.

case of the second possibility, in which the path of the searching point along
the chain is reduced to zero.
A very high efficiency of the implementation of the CMA for polymers
has been achieved, when it was recognized that there is a limited number of
situations in which the searching point can meet while making a walk
through a system containing beads bonded to polymers. A specific situation
code, pat cr(cd, a, b), is used in order to recognize these situations for
various orientations of bonds, various directions of chains, and various
local conformations. There are 15 situations distinguished; these are listed
in Table 2. From these 15 cases, one corresponds to the case, when propagation of motion of the searching point becomes impossible ( pat 1),
because the attempted movement would extend two bonds simultaneously.
Two cases, pat 2 and pat 3, describe the motion of the searching point
along a chain in two different directions. In cases pat 4, 5 and pat 10, 11,
the searching point can enter the chain contour through a chain end or
through a suitable local conformation, respectively. When pat 6, 7 or
pat 13, 14, the searching point leaves the chain contour. The other cases
correspond to situations of the type (i) illustrated in Fig. 11a. In all cases,

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Simulations on Completely Occupied Lattice

TABLE 2
Code (pat)

173

The Situation Code [pat cr(cd, a, b)] used in the CMA Procedures
Situation before ! after

Description

Motion impossiblethe searching point

attempts to stretch two bonds

Motion of the searching point along the

chain in direction of vector a

Motion of the searching point along the

chain in direction of vector b

Searching point enters the chain contour

in direction of a via the end of chain
(b 13)

Searching point enters the chain contour

in direction of b via the end of chain
(a 13)

Searching point leaves the chain contour

via the end of chain (b 13)

Searching point leaves the chain contour

via the end of chain (a 13)

Rotation of bond with the chain end

(b 13)

Rotation of bond with the chain end

(a 13)

10

Searching point enters the chain contour

along the direction of b

11

Searching point enters the chain contour

along the direction of a

12

Searching point
contour

13

Searching point leaves the chain in

direction a

14

Searching point leaves the chain in

direction b

15

Searching point meets a single bead

Temporarily free lattice site.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

crosses

the

chain

174

Pakula

when the searching point is not forced to move along the chain, the direction
of motion is chosen randomly. Coding of situations and resulting rearrangements can considerably speed up the simulation. The described codes
should be considered as suitable examples. Other, maybe more effective,
solutions are certainly possible.
The various versions of the CMA are suitable for implementation and
usage on personal computers. With recently available technology, dense
polymeric systems of up to one million beads organized in various macromolecular architectures can be simulated in a reasonably inexpensive way.

VI. CONCLUDING REMARKS

It has been demonstrated that a lot of problems concerning the behavior of
complex polymer systems can be analyzed successfully by the methods discussed. Difficulties in simulations of dense systems and systems with complex molecular topologies seem to be overcome in the DLL and CMA
algorithms. Both can be very efficiently applied to the study of static and
dynamic properties of polymer melts with consideration of various topologies of molecules and various interactions between molecular elements.
This success has been achieved by an application of a reasonably simplified dynamic model of cooperative rearrangements, which cuts off details of
the dynamics characteristic for shorter time ranges. Such treatment, consisting in a simplification of the dynamics to features characteristic to a given
time range, should be considered as analogous to simplifications of the
structure in relation to various size scales. Unfortunately, a scheme of
dynamic simplifications over a broad time range corresponding to the
scheme of structures considered with variable size resolution is generally
not as complete as the latter.
The DLL model can be considered as defining a special purpose parallel
computer, the development of which could solve the problem of strong size
limits for simulated systems caused usually by strong dependencies of
computation times on system sizes.

REFERENCES
1.
2.
3.
4.
5.
6.
7.

Verdier, P.H.; Stockmeyer, W.H. J. Chem. Phys. 1962, 36, 227.

Lal, M. Mol. Phys. 1969, 17, 57.
Wall, F.T.; Mandel, F. J. Chem. Phys. 1975, 63, 4592.
Carnesin, I.; Kremer, K. Macromolecules 1988, 21, 2819.
Binder, K.; Kremer, K. Comput. Phys. Rep. 1988, 7, 261.
Pakula, T. Macromolecules 1987, 20, 679.
Geyler, S.; Pakula, T.; Reiter, J. J. Chem. Phys. 1990, 92, 2676.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Simulations on Completely Occupied Lattice

175

8. Reiter, J.; Edling, T.; Pakula, T. J. Chem. Phys. 1990, 93, 837.
9. Gauger, A.; Weyersberg, A.; Pakula, T. Macromol. Chem. Theory Simul. 1993,
2, 531.
10. Gauger, A.; Pakula, T. Macromolecules 1995, 28, 190.
11. Pakula, T.; Teichmann, J. Mat. Res. Soc. Symp. Proc. 1997, 495, 211.
12. Teichmann, J. Ph.D. thesis, University of Mainz, 1996.
13. Pakula, T.J. Mol. Liquids 2000, 86, 109.
14. Polanowski, P.; Pakula, T. J. Chem. Phys. 2002, 117, 4022.
15. Barker, J.A.; Henderson, D. Rev. Mod. Phys. 1976, 48, 587.
16. Alder, B.J.; Wainwright, T.E. J. Chem. Phys. 1969, 31, 459.
17. Widom, A. Phys. Rev. A 1971, 3, 1394.
18. Kubo, R. Rept. Progr. Phys. 1966, 29, 255.
19. London, R.E. J. Chem. Phys. 1977, 66, 471.
20. Gotze, W.; Sjogren, L. Rep. Prog. Phys. 1992, 55, 241.
21. Cohen, M.H.; Grest, G.S. Phys. Rev. B 1979, 20, 1077.
22. Adam, G.; Gibbs, J.H. J. Chem. Phys. 1965, 43, 139.
23. Hughes, B.D. Random Walks and Random Environments; Clarendon Press:
Oxford, 1995.
24. Reiter, J. Macromolecules 1990, 23, 3811.
25. Reiter, J. Physica A 1993, 196, 149.
26. Pakula, T.; Geyler, S.; Edling, T.; Boese, D. Rheol. Acta 1996, 35, 631.
27. Pakula, T.; Harre, K. Comp. Theor. Polym. Sci. 2000, 10, 197.
28. Fox, T.G.; Flory, P.J. J. Appl. Phys. 1950, 21, 581.
29. Paul, W.; Binder, K.; Herrmann, D.W.; Kremer, K. J. Chem. Phys. 1991, 95,
7726.
30. Kolinski, A.; Skolnick, J.; Yaris, R. J. Chem. Phys. 1987, 86, 7164.
31. Fetters, L.J.; Kiss, A.D.; Pearson, D.S.; Quack, G.F.; Vitus, F.J.
Macromolecules 1993, 26, 647.
32. Roovers, J.; Lin-Lin Zbou, Toporowski, P.M.; van der Zwan, M.; Iatron, H.;
Hadjichristidis, N. Macromolecules 1993, 26, 4324.
33. Antonietti, M.; Pakula, T.; Bremser, W. Macromolecules 1995, 28, 4227.
34. Pakula, T. Comp. Theor. Polym. Sci. 1998, 8, 21.
35. Pakula, T.; Vlassopoulos, D.; Fytas, G.; Roovers, J. Macromolecules 1998, 31,
8931.
36. Grest, G.S.; Kremer, K.; Milner, S.T.; Witten, T.A. Macromolecules 1989, 22,
1904.
37. Pakula, T. J. Comp. Aided Mat. Des. 1996, 3, 329.
38. Pakula, T.; Karatasos, K.; Anastasiadis, S.H.; Fytas, G. Macromolecules 1997,
30, 8463.
39. Pakula, T.; Floudas, G. In Block Copolymers, Balta Calleja, F.J., Roslaniec, Z.,
Eds.; Marcel Dekker Inc.: New York, 2000; 123177.
40. Pakula, T. J. Macromol. Sci.Phys. 1998, B37, 181.
41. Karatasos, K.; Anastasiadis, S.H.; Floudas, G.; Fytas, G.; Pispas, S.;
Hadjichristidis, N.; Pakula, T. Macromolecules 1996, 29, 1326.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

176

Pakula

42. Floudas, G.; Pispas, S.; Hadjichristidis, N.; Pakula, T.; Erukhimovich, I.
Macromolecules 1996, 29, 4142.
43. Vilesov, A.D.; Floudas, G.; Pakula, T.; Melenevskaya, E. Yu.; Birstein, T.M.;
Lyatskaya, Y.V. Macromol. Chem. Phys. 1994, 195, 2317.
44. Pakula, T.; Matyjaszewski, K. Macromol. Theory Simul. 1996, 5, 987.
45. Miller, P.J.; Matyjaszewski, K.; Shukla, N.; Immaraporn, B.; Gelman, A.;
Luokala, B.B.; Garo, S.; Siclovan, T.M.; Kickelbick, G.; Vallant, T.;
Homann, H.; Pakula, T. Macromolecules 1999, 32, 8716.
46. Polanowski, P. Ph.D. thesis, Technical University of Lodz, 2002.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

6
Molecular Dynamics Simulations
of Polymers
VAGELIS A. HARMANDARIS and VLASIS G. MAVRANTZAS Institute
of Chemical Engineering and High-Temperature Chemical Processes, and
University of Patras, Patras, Greece

I.

THE MOLECULAR DYNAMICS TECHNIQUE

Molecular dynamics (MD) is a powerful technique for computing the

equilibrium and dynamical properties of classical many-body systems.
Over the last fifteen years, owing to the rapid development of computers, polymeric systems have been the subject of intense study with MD
simulations [1].
At the heart of this technique is the solution of the classical equations
of motion, which are integrated numerically to give information on the
positions and velocities of atoms in the system [24]. The description of a
physical system with the classical equations of motion rather than quantummechanically is a satisfactory approximation as long as the spacing h
between the successive energy levels described is h<kBT. For a typical
system at room temperature this holds for <0.6  1013 Hz, i.e., for
motions of time periods of about t > 1.6  10
13 sec or 0.16 ps.
A simple flow diagram of a standard MD algorithm is shown in Fig. 1
and includes the following steps:
1.

2.

First, a model configuration representing a molecular-level snapshot

of the corresponding physical system is chosen or constructed, and
is initialized (initial positions, velocities of each particle within the
system).
Then the total force acting on each particle within the system
is computed. For polymer systems such a force has two

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

178

FIG. 1

Harmandaris and Mavrantzas

A simple flow diagram of a standard MD algorithm.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

3.

4.

5.

179

components: intermolecular (from atoms belonging to different polymer chains) and intramolecular (from atoms belonging to the same
chain).
The integration of the equations of motion follows with an appropriate method. The most popular of these will be described in detail
in the next section.
Actual measurements are performed (positions, velocities, energies,
etc., are stored) after the system has reached equilibration, periodically every Nk steps.
After completion of the central loop (N steps), averages of the measured quantities and of the desired properties are calculated and
printed.

II. CLASSICAL EQUATIONS OF MOTION

As stated above, at the heart of an MD simulation is the solution of the
classical equations of motion. Let us consider a system consisting of N
interacting molecules described by a potential energy function V. Let us also
denote as qk and q_ k the generalized coordinates describing the molecular
configuration and their time derivatives, respectively. The classical equations of motion for this system can be formulated in various ways [5]. In the
Lagrangian formulation, the trajectory q(t) (q1(t), q2(t), . . . , qk(t), . . .)
satisfies the following set of differential equations:


@L
d @L

@qk dt @q_ k

where L is the Lagrangian of the system. This is defined in terms of the

kinetic energy, K, and potential energy, V, as L Lq, q_ , t  K
V. The
generalized momenta pk conjugate to the generalized coordinates qk are
defined as
pk

@L
@q_k

Alternatively, one can adopt the Hamiltonian formalism, which is cast in

terms of the generalized coordinates and momenta. These obey Hamiltons
equations
q_ k

@H
,
@pk

p_k

@H
@qk

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

180

Harmandaris and Mavrantzas

where H is the Hamiltonian of the system, defined through the equation

H p, q

q_ k pk
L

If the potential V is independent of velocities and time, then H becomes

equal to the total energy of the system: H(p, q) K(p) V(q) [5]. In
Cartesian coordinates, Hamiltons equations of motion read:
r_ i  vi

pi
,
mi

p_ i
rri V 

@V
Fi
@ri

hence
mi r i  mi q_ i Fi

where Fi is the force acting on atom i. Solving the equations of motion then
involves the integration of the 3N second-order differential equations (6)
(Newtons equations).
The classical equations of motion possess some interesting properties, the
most important one being the conservation law. If we assume that K and V
do not depend explicitly on time, then it is straightforward to verify that
H_ dH=dt is zero, i.e., the Hamiltonian is a constant of the motion. In
actual calculations this conservation law is satisfied if there exist no
explicitly time- or velocity-dependent forces acting on the system.
A second important property is that Hamiltons equations of motion
are reversible in time. This means that, if we change the signs of all
the velocities, we will cause the molecules to retrace their trajectories
backwards. The computer-generated trajectories should also possess this
property.
There are many different methods for solving ordinary differential
equations of the form of Eq. (6). Criteria for the proper choice of an
algorithm include the following:
.

The algorithm must not require an expensively large number of force

evaluations per integration time step. Many common techniques for
the solution of ordinary differential equations (such as the fourthorder RungeKutta method) become inappropriate, since they do
not fulfill this criterion.
The algorithm should satisfy the energy conservation law. It is also
desirable that it be time reversible and conserve volume in phase space
(be symplectic).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

.
.

181

The algorithm should permit the use of a large time step dt.
The algorithm should be fast and require little memory.

Concerning the solution of equations of motion for very long times,

it is clear that no algorithm provides an essentially exact solution. But
this turns out to be not a serious problem, because the main objective
of an MD simulation is not to trace the exact configuration of a system
after a long time, but rather to predict thermodynamic properties as
time averages and calculate time correlation functions descriptive of the
dynamics.
In the following we briefly describe the two most popular families of
algorithms used in MD simulations for the solution of classical equations of
motion: the higher-order methods and the Verlet algorithms.

A. Higher-Order (Gear) Methods

The basic idea in the higher-order methods is to use information
about positions and their first, second, . . . , nth time derivatives at time t
in order to estimate positions and their first, second, . . . , nth time derivatives at time t dt [2]. If we consider the Taylor expansion of the position
vectors of a given particle at time t dt including terms up to fourth-order
we have
dt2
dt3 
dt4 
r t
r t
r t   
2
6
24
dt2 
dt3 
v p t dt vt dt r t
rt
rt   
2
6
2
dt 
r p t dt r t dt rt
rt   
2

r p t dt rt dt 
rt   
r p t dt rt dt vt

7
8
9
10

In the above equations, the superscript p is used to denote predicted

values and the dots, time derivatives. Equations (7)(10) do not generate
classical trajectories, since we have not as yet introduced the equations of
motion. To do this we estimate the size of the error incurred by the expansion, x, by calculating the forces (or, equivalently, the accelerations) at
the predicted positions
x  r r p t dt
r p t dt

diag1=m1 , 1=m2 , . . . , 1=mN rr Vr p t dt
r p t dt

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

11

182

Harmandaris and Mavrantzas

The error is accounted for and corrected in a corrector step, that is

rc t dt r p t dt c0 x

12

vc t dt v p t dt c1 x
r c t dt r p t dt c2 x

13
14


r c t dt r p t dt c3 x

15

where ci, i 1, . . . , n are constants. The values of ci are such that they yield
an optimal compromise between desired level of accuracy and algorithm
stability [2].
The general scheme of an algorithm based on the predictor-corrector
method goes as follows:
1.
2.

3.

Predict positions and their first, second, . . . , nth time derivatives at

time t dt using their values at time t.
Compute forces using the predicted positions and then the corresponding error x from the differences between accelerations as
calculated from forces and accelerations as predicted by the prediction scheme.
Correct the predicted positions and their first, second, . . . , nth time
derivatives guided by x.

B. Verlet Methods
Algorithms in this family are simple, accurate, and, as we will see below,
time reversible. They are the most widely used methods for integrating the
classical equations of motion. The initial form of the Verlet equations [3] is
obtained by utilizing a Taylor expansion at times t
dt and t dt
dt2
dt3 
r t
rt Odt4
2
6
dt2
dt3 
r t

rt
dt rt
dt vt
rt Odt4
2
6

rt dt rt dt vt

16
17

Summing the two equations gives

rt dt 2rt
rt
dt dt2 r t Odt4
with r t calculated from the forces at the current positions.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

18

Molecular Dynamics Simulations

183

Two modifications of the Verlet scheme are of wide use. The first is the
leapfrog algorithm [3] where positions and velocities are not calculated at
the same time; velocities are evaluated at half-integer time steps:



dt
rt dt rt dt v t
2






dt
dt
v t
v t

dt r t
2
2

19
20

In order to calculate the Hamiltonian H at time t, the velocities at time t are

also calculated as averages of the values at t dt/2 and t
dt/2:







1
dt
dt
v t
v t

vt
2
2
2

21

The problem of defining the positions and velocities at the same time can
be overcome by casting the Verlet algorithm in a different way. This is the
velocity-Verlet algorithm [3,6], according to which positions are obtained
through the usual Taylor expansion
rt dt rt dt vt

dt2
r t
2

22

whereas velocities are calculated through

vt dt vt

dt
r t r t dt
2

23

with all accelerations computed from the forces at the configuration corresponding to the considered time. To see how the velocity-Verlet algorithm is
connected to the original Verlet method we note that, by Eq. (22),
rt 2dt rt dt dt vt dt

dt2
r t dt
2

24

If Eq. (22) is written as

rt rt dt
dt vt

dt2
r t
2

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

25

184

Harmandaris and Mavrantzas

then, by addition, we get

rt 2dt rt 2rt dt dtvt dt
vt

dt2
rt dt
r t
2
26

Substitution of Eq. (23) into Eq. (26) gives

rt 2dt rt 2rt dt dt2 r t dt

27

which is indeed the coordinate version of the Verlet algorithm. The calculations involved in one step of the velocity algorithm are schematically shown
in [2, figure 3.2, page 80].
A sample code of the velocity-Verlet integrator is shown in Algorithm 1.
In this algorithm, N is the total number of atoms in the system and the
subroutine get_ forces calculates the total force on every atom within the
system.
Algorithm 1: Velocity-Verlet Integration Method
......
do i 1, N
r(i) r(i) dt * v(i) dt * dt/2 * F(i)
v(i) v(i) dt/2 * F(i)

! update positions at t dt
using velocities and forces at t
! update velocities at t dt using
forces at t

end do
call get_ forces (F)
do i 1, N
v(i) v(i) dt/2 * F(i)

! calculate forces at t dt
! update velocities at t dt
using forces at t dt

end do
......

In general, higher-order methods are characterized by a much better

accuracy than the Verlet algorithms, particularly at small times. Their
biggest drawback is that they are not reversible in time, which results in
other problems, such as insufficient energy conservation, especially in very
long-time MD simulations. On the other hand, the Verlet methods are not
essentially exact for small times but their inherent time reversibility
guarantees that the energy conservation law is satisfied even for very long
times [4]. This feature renders the Verlet methods, and particularly the
velocity-Verlet algorithm, the most appropriate ones to use in long atomistic
MD simulations.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

185

III. MD IN OTHER STATISTICAL ENSEMBLES

The methods described above address the solution to Newtons equations
of motion in the microcanonical (NVE) ensemble. In practice, there is
usually the need to perform MD simulations under specified conditions of
temperature and/or pressure. Thus, in the literature there exist a variety
of methodologies for performing MD simulations under isochoric or
isothermal conditions [2,3]. Most of these constitute a reformulation of the
Lagrangian equations of motion to include the constraints of constant T
and/or P. The most widely used among them is the NoseHoover method.

A. The NoseHoover Thermostat

To constrain temperature, Nose [7] introduced an additional degree of
freedom, s, in the Lagrangian. The parameter s plays the role of a heat bath
whose aim is to damp out temperature deviations from the desired level.
This necessitates adding to the total energy an additional potential term of
the form
Vs gkB T ln s

28

and an additional kinetic energy term of the form


Q s_ 2 p2s
Ks

2 s
2Q

29

In the above equations, g is the total number of degrees of freedom.

In a system with constrained bond lengths, for example, g 3 Natoms

Nbonds
3, with Natoms and Nbonds standing for the total numbers of atoms
and bonds, respectively; the value of 3 subtracted in calculating g takes care
of the fact that the total momentum of the simulation box is constrained to
be zero by the periodic boundary conditions. Q and ps represent the effective mass and momentum, respectively, associated with the new degree of
freedom s. Equations of motion are derived from the Lagrangian of the
extended ensemble, including the degree of freedom s. Their final form,
according to Hoovers analysis [8], is
pi
mi
@V s_
p_ i


p
@ri s i
!
N
X
p2i

gkB T =Q,
p_ s
mi
i1
r_ i

30
31
ps Q

s_
s

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

32

186

Harmandaris and Mavrantzas

An important result in Hoovers analysis is that the set of equations of

motion is unique, in the sense that no other equations of the same form can
lead to a canonical distribution.
The total Hamiltonian of the system, which should be conserved during
the MD simulation, is

HNose

Hoover

N
X
p2

 
p2
V rN gkB T ln s s
mi
2Q
i

i1

33

To construct MD simulations under constant P, an analogous reformulation of the Lagrangian was proposed by Andersen [9]. The constantpressure method of Andersen allows for isotropic changes in the volume
of the simulation box. Later, Hoover [8] combined this method with the
isothermal MD method described above to provide a set of equations for
MD simulations in the NPT ensemble. Parrinello and Rahman [10] extended
Andersens method to allow for changes not only in the size, but also in the
shape of the simulation box. This is particularly important in the simulation
of solids (e.g., crystalline polymers) since it allows for phase changes in the
simulation involving changes in the dimensions and angles of the unit cell.

B. The Berendsen ThermostatBarostat

Berendsen proposed a simpler way for performing isothermal and/or isobaric MD simulations without the need to use an extended Lagrangian, by
coupling the system into a temperature and/or pressure bath [11]. To achieve
this, the system is forced to obey the following equations
dT
T
Text = T
dt

34

and
dP
P
Pext = P
dt

35

where Text and Pext are the desired temperature and pressure values and T
and P are time constants characterizing the frequency of the system coupling to temperature and pressure baths. T and P are the instantaneous
values of temperature and pressure, calculated from the momenta and configuration of the system [2]. The solution of these equations forces velocities

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

187

and positions to be scaled at every time step by factors T and xP, respectively, with




1=2
dt T
xT 1

1
T Text
dt
xP 1
T P
Pext
P

36
37

and T being the isothermal compressibility of the system.

The method proposed by Berendsen is much simpler and easier to
program than that proposed by Nose and Hoover. It suffers, however, from
the fact that the phase-space probability density it defines does not
conform to a specific statistical ensemble (e.g., NVT, NPT). Consequently,
there exists no Hamiltonian that should be conserved during the MD
simulation.

C. MD in the NTLxryyrzz Ensemble

To further illustrate how extended ensembles can be designed to conduct
MD simulations under various macroscopic constraints, we discuss here the
NTLx yy zz ensemble. NTLx yy zz is an appropriate statistical ensemble for
the simulation of uniaxial tension experiments on solid polymers [12] or
relaxation experiments in uniaxially oriented polymer melts [13]. This
ensemble is illustrated in Fig. 2. The quantities that are kept constant during
a molecular simulation in this ensemble are the following:
.
.
.
.

FIG. 2

the
the
the
the

total number of atoms in the system N,

temperature T,
box length in the direction of elongation Lx, and
time average values of the two normal stresses  yy and  zz.

The NTLx yy zz statistical ensemble.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

188

Harmandaris and Mavrantzas

The NTLx yy zz ensemble can be viewed as a hybrid between the NVT
ensemble in the x direction and the isothermalisobaric (NPT) ensemble in
the y and z directions. The temperature T is kept fixed at a prescribed value
by employing the NoseHoover thermostat; the latter introduces an additional dynamical variable in the system, the parameter s, for which an
evolution equation is derived. Also kept constant during an MD simulation
in this ensemble is the box length Lx in the x direction; on the contrary, the
box lengths in the other two directions, Ly and Lz, although always kept
equal, are allowed to fluctuate. This is achieved by making use in the
simulation of an additional dynamical variable, the cross-sectional area
A( LyLz) of the simulation cell in the yz plane, which obeys an extra
equation of motion involving the instantaneous average normal stress
( yy  zz)/2 in the two lateral directions y and z, respectively; ( yy  zz)/2
remains constant on average and equal to
Pext throughout the simulation.
The derivation of the equations of motion in the NTLx yy zz ensemble has
been carried out in detail by Yang et al. [12], and goes as follows: Consider a
system consisting of N atoms with rik being the position of atom i belonging
to polymer chain k. The bond lengths are kept fixed, with gik denoting the
constraint forces on atom i. The Lagrangian is written as a function of
~ k is the scaled (with respect
~ k , xik, A, s} where R
the extended variables {R
to the box edge lengths) position of the center of mass of every chain k, and
xik is the position of atom i in chain k measured relative to the chain center of
mass. This ensemble is extended in the sense that it invokes the additional
variables A and s, makes use of a scaled coordinate system, and is formulated
with respect to a virtual time t0 . The equations of motion are derived from
the extended Lagrangian by exactly the same procedure as for the other
statistical ensembles. The final equations are further recast in terms of real
coordinates and real time and have the following form:
s_
mi rxik Fxik gxik
pxik
s
!
mi Ryk A_ 2
s_
A

mi ryik Fyik gyik
pyik
2A
2A
s
!
mi Rzk A_ 2
s_
A

mi rzik Fzik gzik
pzik
2A
2A
s
"
#
X X p2xik p2yik p2zik
s_2

g 1kB T
Qs Q s
s
mi
i
k



1 
s_A_
2

yy
zz
Pext

WA W
s Lx
2
s

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

38
39

40

41
42

Molecular Dynamics Simulations

189

where the forces with two indices indicate center of mass forces, while those
with three indices are forces on atoms within a particular polymer chain. Rk
denotes the center of mass of molecule k, while Q and W are inertial constants governing fluctuations in the temperature and the two normal stresses
 yy and  zz, respectively. The total Hamiltonian of the extended system,
derived from the Lagrangian, has the form:

HNTLx yy zz

!2


X p2
Q s_ 2
ln s W A_
i

V r
g 1
Pext Lx A
2 s

2 s
2mi
i
43

The first term on the right hand side represents the kinetic energy, the
second term is the potential energy, and the last four terms are the contributions due to the thermostat and the fluctuating box cross-sectional area
in the plane yz. Conservation of HNTL x yy zz is a good test for the
simulation.
For the solution of equations of motion, a modification of the velocityVerlet algorithm proposed by Palmer [14] can be followed.

IV. LIOUVILLE FORMULATION OF EQUATIONS OF

MOTIONMULTIPLE TIME STEP ALGORITHMS
In Sections II.A and II.B we presented the most popular algorithms for
integrating Newtons equations of motion, some of which are not reversible
in time. Recently, Tuckerman et al. [15] and Martyna et al. [16] have shown
how one can systematically derive time reversible MD algorithms from the
Liouville formulation of classical mechanics.
The Liouville operator L of a system of N degrees of freedom is defined in
Cartesian coordinates as

iL

N
X
@
@
r_ i
Fi
@ri
@pi
i1

44

If we consider the phase-space of a system,  {r, p}, the evolution of the

system from time 0 to time t, can be found by applying the evolution
operator
t expiLt0

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

45

190

Harmandaris and Mavrantzas

The next step is to decompose the evolution operator into two parts
such that

iL iL1 iL2

with

N
X
@
,
iL1
Fi
@pi
i1

N
X
@
_
iL2
ri
@ri
i1

46

For this decomposition, a short-time approximation to the evolution

operator can be generated via the Trotter theorem [16] as
expiLt expiL1 L2 t=PP



expiL1 dt=2 expiL2 dt expiL1 dt=2P O t3 =P2

47

where d t t/P. Thus, the evolution operator becomes





 
dt
dt
expiL2 dt exp iL1
O dt3
expiLdt exp iL1
2
2

48

The evolution of the system at time t using the above factorization, Eq.
(48), is described through the following scheme [16]
rdt r0 dt v0
vdt v0

dt2
Fr0
2m

dt2
Fr0 Frdt
2m

49
50

which can be derived using the identity expa@=@gx x g
1 gx a .

The result is the well-known velocity-Verlet integration scheme, described
before, which is now derived in a different way.
Based on the previous factorization a very efficient algorithm can be
developed, through the use of different time steps for integrating the
different parts of the Liouville operator. This is the so-called reversible
REference System Propagator Algorithm (rRESPA).

A. The rRESPA Algorithm

In the rRESPA algorithm, the above factorization is employed together
with an integration of each part of the Liouville operator with a different time step. In addition, the forces F are also decomposed into fast
(short-range) forces Ff, and slow (long-range) forces Fs, according to

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

191

F(r) Ff(r) Fs(r). The total evolution operator is broken up into

iL iL1 iL2 iL3 with

N
N
N
X
X
X
@
@
@
r_ i
,
iL2
,
iL3
51
iL1
Ffi r
Fsi r
@pi
@ri
@pi
i1
i1
i1
The heart of the rRESPA algorithm is that the equations of motion are
integrated by using two different time steps, it is therefore a Multiple
Time Step (MTS) method: the slow modes (slow forces, iL3) are integrated
with a larger time step, t, whereas the fast modes (fast forces and
velocities, iL1, iL2) with a smaller time step, t(t t/n). In this case the
evolution operator becomes [16]










t
t
t n
t
exp iL3
expiLtexp iL3
exp iL1
expiL2 texp iL1
2
2
2
2
 3
52
O t
The force calculated n times (fast force) is called the reference force. A
sample code of an MTS integrator is given in Algorithm 2.
Algorithm 2: rRESPA Integration
. . .. . ..
do i 1, N
v(i) v(i) t/2 * Fs(i)

! update velocities using

! slow forces at t

end do
do j 1, n
do i 1, N
v(i) v(i) t/2 * Ff(( j
1)t)
r(i) r(i) t * v(i)
end do
call fast_ forces (Ff )

! update velocities using

! fast forces at t ( j
1)t
! update positions at t jt
! get fast forces at t jt

do i 1, N
v(i) v(i) t/2 * F f( jt)
end do
end do
call slow_ forces (Fs)
do i 1, N
v(i) v(i) t/2 * Fs(i)

! get slow forces at t t

! update velocities using
! slow forces at t t

end do
. . .. . ..

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

192

Harmandaris and Mavrantzas

B. rRESPA in the NVT Ensemble

For MD simulations in NVT ensemble, a modification of the rRESPA
algorithm has been proposed. The method uses a modification of the
Lagrangian of the system based on the NoseHoover approach, described in
Section III.A. The difference from the standard rRESPA scheme described
before is that now the total Liouville operator is decomposed as
iL iL1 iL2 iL3 iLNH

53

with

N
X
@
f
iL1
F i r
,
@pi
i1

N
X
@
r_ i
iL2
,
@ri
i1

N
X
@
s
Fi r
iL3
@pi
i1
54

Also,
iLNH

N
X
i1

  vi 

@
@
@
 G
@vi
@
@

55

where
G

N
X
p2
i

i1

mi

!

gkB T =Q,

 

56

and  is a transformation of the additional degree of freedom s, log s N.

Two modifications of the standard RESPA method exist, depending on
the application of the extended operator exp(iLNHt). The first variant of
RESPA is useful when the evolution prescribed by the operator exp(iLNHt)
is slow compared to the time scale associated with the reference force. It is
formed by writing






t
t
expiLt exp iLNH
exp iL3
2
2







t
t n
 exp iL1
expiL2 t exp iL1
2
2






 3
t
t
 exp iL3
exp iLNH
O t
2
2

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

57

Molecular Dynamics Simulations

193

and is named XO-RESPA (eXtended system Outside-REference System

Propagator Algorithm). In general, XO-RESPA can be applied to systems
characterized by fast vibrations, as the time scale associated with the
extended system variable is usually chosen to be quite slow compared
with these motions.
If the motion prescribed by the operator exp(iLNHt) occurs on the same
time scale as that generated by the fast forces, then a useful RESPA
algorithm includes the application of this operator for the small time step dt.
The evolution operator takes then the form
expiLt









t
t
t
exp iL3
exp
iLNH
exp iLNH
2
2
2













t
t
t
t n
exp iL1
expiL2 t exp iL1
exp iLNH
 exp iLNH
2
2
2
2









 3
t
t
t
 exp
iLNH
exp iL3
exp iLNH
O t
58
2
2
2
The resulting integrator is named XI-RESPA (eXtended system InsideREference System Propagator Algorithm).
Modifications of the RESPA method for MD simulations in the NPT
statistical ensemble have also been formulated in an analogous manner.
More details can be found in the original papers [15,16].

V. CONSTRAINT DYNAMICS IN POLYMERIC

SYSTEMS
One of the most important considerations in choosing the best algorithm for
the solution of the classical equations of motion is, as we saw above, the
time step of integration. This should be chosen appreciably shorter than the
shortest relevant time scale in the simulation. For long-chain polymeric
systems, in particular, where one explicitly simulates the intramolecular
dynamics of polymers, this implies that the time step should be shorter than
the period of the highest-frequency intramolecular motion. This renders the
simulation of long polymers very expensive. One solution to this problem
is provided by the MTS algorithm discussed above. Another technique
developed to tackle this problem is to treat bonds between atoms,
characterized by the highest-frequency intramolecular vibrations, as rigid.
The MD equations of motion are then solved under the constraint that bond
lengths are kept constant during the simulation. The motion associated with

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

194

Harmandaris and Mavrantzas

the remaining degrees of freedom is presumably slower, permitting the use

of a longer time step in the simulation.
In general, system dynamics should satisfy many constraints (e.g., many
bond lengths should be constant) simultaneously. Let us denote the
functions describing the constraints by  k 0 with k r2ij
dij2 , meaning
that atoms i and j are held at a fixed distance dij. A new system Lagrangian is
introduced that contains all constraints
Lc L

k k r

59

where k denotes the set of constraints and

k the corresponding set of
Lagrange multipliers. The equations of motion corresponding to the new
Lagrangian are
mi r i Fi

@k
F i
gi
@ri

60

where the second term on the right hand side of Eq. (60) denotes the constraint forces. The question then is how to calculate the set of Lagrange
multipliers
k. Two methods that have widely been used in our MD simulations are discussed here: The EdbergEvansMorriss method and the
SHAKE method.

A. The EdbergEvansMorriss Algorithm

This algorithm [17] starts by considering a set of a linear system of equations
in {
ij}, which are formulated by taking the second derivatives of the
constraint equations in time:
 2
r2ij
dij2 0 ) 2rij  r_ ij 0 ) rij  r ij r_ ij 0

61

One then solves the following set of algebraic and differential equations
simultaneously:
mi r i Fi gi
X @k
gi

k
@ri
k
 2
rij  r ij r_ ij 0

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

62
63
64

Molecular Dynamics Simulations

195

Note that site velocities enter this formulation explicitly. Upon

substitution of the site accelerations from Eq. (60) into Eq. (64) one obtains
a system of linear equations in {
ij}; thus, the determination of the
ijs
reduces to the solution of a linear matrix equation which should be
addressed at each time step.

B. The SHAKERATTLE Algorithm

The approach described above suffers from the problem that it is
computationally expensive, since it requires a matrix inversion at every
time step. The problem gets worse with increasing chain length, with the
algorithm becoming practically inappropriate for chains of more than about
100 atoms long. Ryckaert et al. [18] developed a simpler scheme, named
SHAKE, to satisfy the constraints in this case.
If one considers the classical form of the Verlet algorithm, then in the
presence of constraints
rci t dt rui t dti

dt2 X
@k t

k
k
@ri
mi

65

where rci are the constrained and rui the unconstrained positions. If the
constraints are satisfied at time t dt, then kc t dt 0. But if the
system moved along the unconstrained trajectory, the constraints would
not be satisfied at t dt. In this case, by performing a Taylor expansion
around the unconstrained positions, we get
kc t dt ku t dt


N

X
@k
i1

@ri

rui tdt



 rci t dt
rui t dt Odt4
66

and by using Eq. (65)

ku t

dt

N
X
dt2 X
i1

mi

k0

k0

@k
@ri




@k0
@ri

67

The above equation has the structure of a matrix equation

ku t dt dt2 M,

68

By inverting the matrix, one can solve for the vector ,. However, since
the Taylor expansion in Eq. (66) has been truncated, the s should be

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

196

Harmandaris and Mavrantzas

computed at the corrected positions, and the preceding equations should be

iterated until convergence is reached.
This procedure is also computationally expensive, because it requires a
matrix inversion at every iteration, as does the EdbergEvansMorriss
algorithm. Ryckaert proposed a new method, SHAKE, where the iterative
scheme is not applied to all constraints simultaneously but to each
constraint in succession. Thus the need to invert a large matrix is avoided.
The key point is that rci
rui is approximated as
rci t dt
rui t 

dt2
k @k t
mi @ri

69

By inserting the above equation into Eq. (66), one gets

ku t dt dt2
k

N
X
1 @k t dt @k t
m
@ri
@ri
i
i1

70

from which

k dt2 PN

i1

ku t dt
1=mi @k t dt=@ri @k t=@ri

71

In an MD simulation, the constraints are treated in succession during one

cycle of the iteration and the process is repeated until all constraints have
converged to the desired accuracy. An improvement of the SHAKE
algorithm is the RATTLE algorithm, which was proposed by Andersen [19].
In RATTLE, the velocity-Verlet algorithm is employed to integrate the
dynamical equations.
As was also stated above, there are several applications of MD
simulations in polymer science. An example taken from a recent study of
polymer melt viscoelasticity is presented in the following section.

VI. MD APPLICATIONS TO POLYMER MELT

VISCOELASTICITY
Despite its simplicity and unquestionable utility, a brute-force application of
the atomistic MD technique to polymeric systems is problematic, due to the
enormously large computation time needed to track the evolution of such
systems for times comparable to their longest relaxation times [1,2]. This is

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

197

the well-known problem of long relaxation times. To overcome this problem,

a number of approaches have been proposed over the years. The first is to
develop new, more efficient, clever algorithms, such as the multiple time
step algorithms for the integration of equations of motion described in
Section IV, which have allowed extension of the simulation to times almost
an order of magnitude longer than what is usually achieved with
conventional algorithms. The second is to increase the range of length
scales simulated with MD by using a number of processors (nodes) and
special parallelization techniques; such techniques are described in detail
in the next section.
Alternatively, a hierarchical approach can be adopted, which uses
information from many different levels of abstraction, ultimately connecting
to the atomistic level, and a combination of different molecular simulation
methods and theoretical approaches. Such a methodology can be followed,
for example, for the atomistic MD simulation of the viscoelastic properties
of polymer melts. The methodology has been described in a number of
publications [13,20] and includes two variants. In the first, equilibrium
atomistic MD simulations are conducted on model polymer melt configurations preequilibrated with the powerful connectivity-altering endbridging Monte Carlo (MC) algorithm; the latter algorithm is not subject to
the limitations associated with long relaxation times faced by MD.
Dynamical as well as viscoelastic properties are then extracted either
directly from the MD simulations or indirectly through a mapping of the
atomistic trajectories accumulated in the course of the MD simulation onto
an analytical coarse-grained model [2022], such as the Rouse model for
unentangled melts or the reptation model for entangled melts. In the second
variant, nonequilibrium MD simulations are conducted on model polymer
melt configurations which have been preoriented and thoroughly equilibrated with a field-on MC method [23,24] which generates configurations
representative of a melt under conditions of steady-state uniaxial elongational flow. The MD tracks the relaxation of the preoriented chains back to
equilibrium upon cessation of the flow. In this case, again, the linear
viscoelastic properties of the melt are extracted either directly by the
simulation or indirectly by utilizing a mapping onto an analytical coarsegrained model [13].

A. Study of Polymer Viscoelasticity Through

Equilibrium MD Simulations
In detail, the methodology followed for the prediction of the viscoelastic
properties of polymer melts under equilibrium conditions is a three-stage

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

198

Harmandaris and Mavrantzas

hierarchical approach, whereby the dynamical properties of polymer melts

are calculated through the following procedure:
1. First exhaustive end-bridging Monte Carlo (EBMC) simulations
[23,24] are conducted to equilibrate the melts at all length scales.
The EBMC algorithm employs moves that modify the connectivity
among polymer segments, while preserving a prescribed (narrow) molecular
weight distribution. It can thus equilibrate the long-length scale features of a
polymer melt orders of magnitude more efficiently than MD or other MC
methods, its relative efficiency increasing rapidly with increasing chain
length.
2. Relaxed configurations thus obtained are subjected to equilibrium
MD simulations to monitor their evolution in time and extract dynamic
properties. During the atomistic MD simulations, a large number of dynamical trajectories are accumulated.
3. Finally, the trajectories accumulated are mapped onto theoretical
mesoscopic (coarse-grained) models to extract the values of the parameters
invoked in the mesoscopic model description of the same systems.
With the above methodology, atomistic MD simulations were performed
on united-atom model linear polyethylene (PE) melts with molecular length
ranging from N 24 up to N 250 in the canonical NVT ensemble
(T 450K, P 1 atm). To speed-up the MD simulations the multiple time
step rRESPA algorithm, presented in Section IV, was used. The overall
simulation time ranged from 100 ns to 300 ns, depending on the chain
lengths of the systems studied. Many of the dynamical properties (such as
the self-diffusion coefficient D) were calculated directly from the MD
simulations. Others, however, such as the zero-shear rate viscosity 0,
required mapping atomistic MD data upon a mesoscopic theoretical model.
As such one can choose the Rouse model for relatively short PE melts and
the reptation model for the longer-chain melts.
Figure 3 shows the mean square displacement of the chain center of mass,
hRcm t
Rcm 02 i, in time for the longer-chain systems, C156, C200, and
C250. From the linear part of these curves the self-diffusion coefficient D can
be obtained using the Einstein relation,


Rcm t
Rcm 02
D lim
6t
t!1

72

Figure 4 presents predictions for the self-diffusion coefficient D

as a function of mean chain length N. For comparison, also shown in

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

199

FIG. 3 Mean square displacement of the center of mass for the C156 (solid line),
C200 (dotted line), and C250 (dashed line) systems.

FIG. 4 Predicted and experimental [25] self-diffusion coefficients D vs. chain length
N in a loglog plot (T 450K, P 1 atm).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

200

Harmandaris and Mavrantzas

the figure are experimental data [25]. Three distinct regions appear in
the figure:
1. A region of a small-MW, alkane-like behavior (N<60), where D
follows a power-law dependence D  M
b, with b>1. In this regime chain
end effects, which can be described through a free volume theory, dominate
system dynamics [21].
2. An intermediate, Rouse-like regime (from N 6070 up to N 156)
where b  1. System dynamics in this regime is found to obey the Rouse
model, at least for the overall relaxation of the chain [20].
3. A long chain-length, reptation-like regime (N>156), where chain
diffusivity exhibits a dramatic slow down, b  2.4. According to the original
formulation of reptation theory, the latter exponent should be 2.
Phenomena such as contour length fluctuations (CLF) and constraint
release (CR) typically accelerate the escape of the chain from the tube,
causing an increase in D and a decrease in 0 [26]. A recently proposed
theory that incorporates CLF and CR phenomena predicts a stronger exponent, between
2.2 and
2.3 [27]. These values agree with recent experimental results for concentrated polymer solutions and melts, which suggest
an exponent between
2.2 and
2.4 for a variety of polymer systems [25].
In contrast to D, the prediction of other viscoelastic properties, such as
the friction coefficient  or the zero-shear rate viscosity 0, requires that the
atomistic MD data be mapped upon a mesoscopic theoretical model. For
unentangled polymer melts, such a model is the Rouse model, wherein a
chain is envisioned as a set of Brownian particles connected by harmonic
springs [25,28]. For entangled polymer melts, a better model that describes
more accurately their dynamics is the tube or reptation model [26].
According to this model, the motion of an individual chain is restricted by
the surrounding chains within a tube defined by the overall chain contour or
primitive path. During the lifetime of this tube, any lateral motion of the
chain is quenched.
The Rouse model is formulated in terms of three parameters: the number
of beads N, the length of the Kuhn segment b, and the monomeric friction
coefficient . The friction coefficient  can be calculated directly from the
diffusion coefficient D through
kB T
73

ND
while the zero-shear rate viscosity 0 can be calculated from the density ,
the end-to-end distance hR2 i, and the diffusion coefficient D through
0

RThR2 i
36MD

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

74

Molecular Dynamics Simulations

201

Reptation theory is formulated in terms of four parameters: N, b, , and

the entanglement spacing (or, alternatively, the tube diameter) . If  were
known,  and 0 could be calculated through:


kB T2
3NhR2 iD

75

and
0

RT hR2 i hR2 i

36M D 2

76

The calculation of the tube diameter  is a formidable task and can be

addressed either through a topological analysis of accumulated polymer
melt configurations thoroughly equilibrated with an efficient MC algorithm
[29] or by utilizing a geometric mapping of atomistically represented chain
configurations onto primitive paths [22,30]. The latter mapping is realized
through a projection operation involving a single parameter , which governs the stiffness of the chain in the coarse-grained (primitive path) representation. The parameter  is mathematically defined as the ratio of the
constants of two types of Hookean springs: The first type connects adjacent
beads within the projected primitive path, and the second type connects the
projected beads of the primitive path with the corresponding atomistic units
[30]. Different values of  lead to different parameterizations, i.e., to different primitive paths and, consequently, to different values of the contour
length L. Once a value for  has been chosen, the primitive path is fully
defined which allows the calculation of the tube diameter  through the
following equation of reptation theory
L hR2 i

77

To find the proper value of the projection parameter , one can follow a selfconsistent scheme based on the mean square displacement of the primitive
path points (s, s; t) [22]. (s, s; t) is defined as


s, s; t  Rs, t
Rs, 02
78
where R(s, t) is the position vector of the primitive segment at contour length
s at time t, and R(s, 0) is the position vector of the primitive segment at
contour length s at time 0. According to reptation theory
s, s; t 6Dt

1
X
4hR2 i
p1

p2  2

cos



ps
tp2
1
exp

L
d

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

79

202

Harmandaris and Mavrantzas

where the sum is over all normal modes p and d denotes the longest relaxation or disengagement time. For small times (t< d), (s, s; t) is dominated
by the terms with large p and the above equation becomes





Z1
4La 1
tp2
dp 2 2 1
exp

s, s; t 6Dt
p 2
d
0


1=2
3 2
R D
6Dt 2
t1=2
80

Equation (80) offers a nice way of mapping atomistic MD trajectories
uniquely onto the reptation model, through a self-consistent calculation of
the parameter . First, a value of  is chosen and the mapping from the
atomistic chain onto its primitive path is carried out by following the
procedure described by Kroger [30]. Then, Eq. (78) is used to calculate
(s, s; t) for the primitive path points, averaged over all s values. For times
t< d, the resulting curve is compared to that obtained from Eq. (80), using
the values of hR2 i and D (long-time diffusivity of the centers of mass)
calculated directly from the atomistic MD simulations. The procedure is
repeated until convergence is achieved, that is until a  value is found for
which the two curves coincide. This mapping is performed self-consistently,
without any additional adjustable parameters or any experimental input. It
allows a reliable estimation of the tube diameter , by utilizing atomistically
collected MD data only for times shorter than d. Thus, the total duration of
the MD simulations required is governed solely by the time needed reliably
to calculate the center-of-mass diffusion coefficient D. With the values of
hR2 i, D, and , those of  and 0 can be calculated using Eqs. (75) and (76).
With the above procedure the tube diameter  was calculated to be
  60 A for the longer-chain systems C200 and C250, whereas for the shorter
systems, N<200, no proper value of the parameter  could be identified [22].
Figure 5 shows results for the monomeric friction factor  as a function of
mean chain length N, over the entire range of molecular lengths studied.
Filled squares depict results obtained by mapping atomistic MD data onto
the Rouse model, whereas open circles depict results obtained from mapping the atomistic MD data onto the reptation model. According to its
definition,  should be independent of chain length, its value determined
solely by the chemical constitution of the melt. The figure shows clearly that,
at around C156, a change in the mechanism of the dynamics takes place,
which cannot be accommodated by the Rouse model unless a chain-length
dependent  is assumed. In contrast, in this regime (N>156), the reptation
model provides a consistent description of the system dynamics characterized by a constant (0.4  10
9 dyn s/cm) chain-length independent  value
per methylene or methyl segment.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

203

FIG. 5 Monomer friction coefficient  vs. chain length N, obtained from mapping
the atomistic MD data onto the Rouse model (squares) or the reptation model
(circles).

Figure 6 presents the zero-shear rate viscosity 0 as a function of

molecular weight for all systems studied here. For systems of chain length
less than C156 the Rouse model, Eq. (74), was used, whereas for the longer
systems the reptation equation, Eq. (76), was used. Also reported in the
figure are experimental 0 values from measurements [22] conducted in
a series of linear monodisperse PE melts. The 0 predictions from the
reptation model were obtained using the value of  60 A for
the entanglement spacing. The agreement of the simulation results with
the experimental ones is remarkable.

B. Study of Polymer Viscoelasticity Through

Nonequilibrium MD SimulationsSimulation
of the Stress Relaxation Experiment
An alternative way to learn about viscoelastic properties of polymer melts in
the linear regime is to conduct MD simulations of preoriented polymer melt

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

204

Harmandaris and Mavrantzas

FIG. 6 Zero-shear rate viscosity 0 vs. molecular weight M, obtained from the MD
simulation and the Rouse model for small M (circles) or the reptation model for high
M (squares). Also shown in the figure are experimentally obtained 0 values
(triangles).

configurations generated by a field-on MC algorithm. It involves three

stages:
1. First, a coarse-grained description of the polymer melt is invoked
through the definition of the conformation tensor, c, which is a global
descriptor of the long-length scale conformation of polymer chains. The
conformation tensor c is defined as the second moment tensor of the endto-end distance vector of a polymer chain reduced by one-third of the
unperturbed end-to-end distance and averaged over all chains in the system:

ct 3

RtRt
hR2 i0


81

In the above equation, R stands for the end-to-end vector of a macromolecule and hR2 i0 is the mean-squared magnitude of that vector in the equilibrium, quiescent state, where chains are unperturbed to an excellent

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

205

approximation. With the above definition, a series of detailed atomistic MC

simulations can be initiated on model melt systems at various values of the
orienting thermodynamic field xx, starting from the zero value (xx 0,
equilibrium, quiescent, field-free state) that drive the tensor c away from
equilibrium [23,24].
2. In the second stage, the isothermal relaxation of these configurations
to thermodynamic equilibrium is monitored, keeping their dimension along
the x direction constant and the average normal pressure in the y and z
directions equal to the atmospheric pressure. The experiment simulated is
that of stress relaxation upon cessation of a steady-state uniaxial elongational flow. The MD simulation takes place in the NTLx yy zz statistical
ensemble discussed in Section III. The macroscopic variables kept constant
in this ensemble are the typical macroscopic constraints encountered in the
process of fiber spinning at the end of the spinning operation, when the
fibers are under constant extension and the stress  xx in the direction of
pulling is allowed to relax from its initial value to the equilibrium, field-free
value, equal to
Pext [i.e.,  xx(t ! 1)
Pext]. In addition to monitoring the
temporal evolution of the stress component  xx(t) during the NTLx yy zz
MD simulation, also recorded is the evolution of certain ensemble-averaged
descriptors of the chain long-length scale configuration. These descriptors
include the diagonal components of the chain conformation tensor (cxx, cyy,
and czz) and the chain mean square end-to-end distance hR2 i.
3. The third stage includes the development of expressions describing
analytically the time evolution of these quantities by solving the Rouse
model under the initial and boundary conditions corresponding to our
atomistic computer experiment.
Results are presented here from averaging over about 100 NTLx yy zz
MD trajectories for each stress relaxation experiment, initiated at ensembles
of strained configurations of two PE melt systems: a 32-chain C24 and a
40-chain C78 PE melt.
Figures 7a and 7b show the time evolution of the diagonal components
cxx, cyy, and czz of the conformation tensor for the C24 and C78 melts,
respectively. For both systems, the initial value of cxx is significantly higher
than 1, whereas those of cyy and czz are a little less than 1, indicative of the
oriented conformations induced by the imposed steady-state elongational
structure of flow field xx. As time evolves, cxx decreases whereas cyy and czz
increase continuously, approaching the steady-state, field-free value of 1,
indicative of fully equilibrated, isotropic structures in the absence of any
deforming or orienting field.
Figure 8 shows the time evolution of the stress tensor component  xx for
the C24 PE melt systems studied. The stress tensor is calculated in two ways.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

206

Harmandaris and Mavrantzas

FIG. 7 Evolution of the diagonal components cxx, cyy and czz of the conformation
tensor c with time t for (a) the C24 and (b) the C78 PE melt systems. Results are
averaged over all NTLx yy zz trajectories (T 450K, Pext 1 atm).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

207

FIG. 8 Evolution of the component  xx of the stress tensor with time t for the C24
system. The results at every time t have been obtained either by applying the virial
theorem and averaging over all dynamical trajectories (broken line) or by using a
thermodynamic expression based on the free energy as a function of the
conformation tensor (thick solid line) (T 450K, Pext 1 atm).

The first (thin broken line) tracks the evolution of  xx as obtained from
applying the molecular virial theorem on the relaxing configurations and
averaging over all dynamical trajectories. The second way (thick solid line)
uses the Helmholtz energy function and an affine deformation assumption
for chain ends [23,24]. According to the latter approach, the stress tensor at
every time t is calculated from the ensemble-averaged values of mass density
, conformation tensor cxx, and partial derivative of the Helmholtz energy
function with respect to cxx at time t, through
"
#

R
@A=Nch 
xx t
Pext 2 Ttcxx
M
@cxx T, , cjxxj

82
cxx cxx t

where Pext denotes the equilibrium (atmospheric) pressure and M the

number average MW of the system. This approach tracks the evolution of
 xx as obtained from applying the thermodynamic stress equation, Eq. (82),

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

208

Harmandaris and Mavrantzas

FIG. 9 Evolution of the components  yy and  zz of the stress tensor with time t for
the C24 system (T 450K, Pext 1 atm).

based on the current values of cxx and @A=Nch =@cxx , the latter taken from
the melt elasticity simulations presented in [23,24]. As expected, in the
estimates based on the virial theorem are subject to much higher statistical
uncertainty, owing to the fluctuations in the instantaneous configurations.
Clearly, averaging over many configurations is needed in order to improve
the statistical quality of the virial theorem results. Apart from highfrequency noise, the virial theorem results display an oscillatory character.
When the noise and oscillations are smoothed out, the ensemble-averaged
stress  xx(t) from the virial theorem is in very good agreement with the
thermodynamic estimate obtained from cxx and @A=Nch =@cxx . This is an
important result, as it opens up the possibility of calculating stress with high
precision directly from ensemble average conformational properties, based
on a free energy function accumulated via efficient MC runs. The transverse
components  yy and  zz are displayed in Fig. 9. Both  yy and  zz fluctuate
continuously around the constant value
Patm, as required by the macroscopic restrictions placed on the NTLx yy zz ensemble. The plots for the C78
system are similar. Relaxation times extracted from these stress relaxation
computer experiments are identical to those determined from equilibrium
MD simulations (Section VI.A).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

VII.

209

PARALLEL MD SIMULATIONS OF
POLYMER SYSTEMS

In the previous sections, we presented two different approaches for

addressing the problem of long relaxation times, which plagues the
conventional MD method: the multiple time step algorithm and a
hierarchical methodology that leads to the prediction of the dynamical
and rheological properties of polymer melts by mapping simulation data
onto analytical theories. Enhanced MD simulation algorithms can further
be developed by resorting to special parallelization techniques that allow
sharing the total simulation load over a number of processors or nodes. The
key idea in all these techniques is to compute forces on each atom or
molecule and to perform corresponding velocity/position updates independently but simultaneously for all atoms. It is also desired that the force
computations be evenly divided across the processors so as to achieve the
maximum parallelism.
MD simulations of polymer systems, in particular, require computation
of two kinds of interactions: bonded forces (bond length stretching, bond
angle bending, torsional) and nonbonded forces (van der Waals and
Coulombic). Parallel techniques developed [3133] include the atom-decomposition (or replicated-data) method, the force-decomposition method, and
the spatial (domain)-decomposition method. The three methods differ only
in the way atom coordinates are distributed among the processors to
perform the necessary computations. Although all methods scale optimally
with respect to computation, their different data layouts incur different
interprocessor communication costs which affect the overall performance
of each method.

A. Parallel MD Algorithms
Here we will focus on a discussion of algorithms for parallelizing MD
simulations with short-range interactions where the nonbonded forces
are truncated, so that each atom interacts only with other atoms within a
specified cutoff distance. More accurate MD models with long-range forces
are more expensive to deal with. Thus, special techniques are required for
parallelizing MD simulations with long-range interactions.

1. Atom-Decomposition (Replicated-Data) Method

The most commonly used technique for parallelizing MD simulations of
molecular systems is the replicated-data (RD) method [33]. In the literature
there are numerous parallel algorithms and simulations that have been

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

210

Harmandaris and Mavrantzas

FIG. 10 Division of the force matrix among P processors in the atomdecomposition method.

developed based on this approach [34]. The key idea of this method is that
each processor is assigned a subset of atoms and updates their positions and
velocities during the entire simulation, regardless of where they move in the
physical domain.
Let us consider a polymeric system with a total number of atoms N,
where both intramolecular and intermolecular interactions are present. The
system is simulated in a parallel system of P processors. We first define x
and f as vectors of length N, which store the position and total force on
each atom, respectively. We also define the force matrix F, of dimensions
N  N, with Fij denoting the force on atom i due to atom j. In the RD
method each processor is assigned a group of N/P atoms at the beginning of
the simulation. Each processor is also assigned a subblock of the force
matrix F which consists of N/P rows of the matrix, as shown in Fig. 10. If z
indexes the processors from 0 to P
1, then processor z computes forces in
the Fz subblock of rows. It also is assigned the corresponding position and
force subvectors of length N/P denoted as xz and fz.The computation of the
nonbonded force Fij requires only the two atom positions, xi and xj. But to
compute all forces in Fz, processor Pz will need the positions of many atoms
owned by other processors. In Fig. 10 this is represented by having the
horizontal vector x at the top of the figure span all the columns of F,
implying that each processor must store a copy of the entire x vector, hence
the name replicated-data. This also implies that, at each time step, each
processor must receive updated atom positions from all other processors
and send the positions of its own atoms to all other processors, an
operation called all-to-all communication.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

211

A single time step of an RD algorithm comprises the following steps:

1.

2.

3.

4.

First every processor z computes its own part of the nonbonded

forces Fz. In MD simulations this is typically done using neighbor
lists to calculate nonbonded interactions only with the neighboring
atoms. In an analogous manner with the serial algorithm, each processor would construct lists for its subblock Fz once every few time
steps. To take advantage of Newtons third law (that Fij
Fji), each
processor also stores a copy of the entire force vector f. As each
pairwise nonbonded force between atoms i and j is calculated, the
force component is summed for atom i and negated for atom j. Next,
the bonded forces are computed. Since each processor z knows the
positions of all atoms, it can compute the nonbonded forces for its
subvector xz and sum the resulting forces into its local copy of f.
Calculation of both bonded and nonbonded forces scales as N/P,
i.e., with the number of nonbonded interactions computed by each
processor.
In this step, the local force vectors are summed across all processors
in such a way that each processor ends up with the total force on each
of its N/P atoms. This is the subvector fz. This force summation is a
parallel communication over all processors, an operation known as
fold [35]. In the literature there are various algorithms that have been
developed for optimizing this operation. The key characteristic is that
each processor must receive N/P values from every other processor to
sum the total force on its atoms. This requires total communication
of P times N/P; i.e., the fold operation scales as N.
Once each processor has the total force on its subvector xz in step
(3), it can update the positions and the velocities of each atom
(integration step) with no communication at all. Thus, this operation
scales as N/P.
Finally the updated positions of each processor xx should be shared
among all P processors. Each processor must send N/P positions to
every other processor. This operation is known as expand [35] and
scales as N.

A crucial aspect in any parallel algorithm is the issue of load-balance.

This concerns the amount of work performed by each processor during the
entire simulation, which ideally should be the same for all processors. As we
saw before, the RD algorithm divides the MD force computation (the most
time consuming part in a typical MD simulation) and integration evenly
across all processors. This means that steps (1) and (3) scale optimally as
N/P. Load-balance will be good so long as each processors subset of atoms
interacts with roughly the same number of neighbor atoms. This usually

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

212

Harmandaris and Mavrantzas

occurs naturally if the atom density is uniform across the simulation domain
(e.g., bulk simulations). In a different case (e.g., polymer chains at interfaces
or adsorbed polymers) load-balance can be achieved by randomizing the
order of the atoms at the start of the simulation or by adjusting the size of
the subset of each processor dynamically during the simulation to tune the
load-balance; these are called dynamical load-balancing techniques [36].
In summary, the RD algorithm divides the force computation evenly
across all processors. At the same time, its simplicity makes it easy to
implement in existing codes. However, the algorithm requires global
communication in steps (2) and (4), as each processor must acquire
information from all other processors. This communication scales as N,
independently of the number of processors P. Practically this limits the
number of processors that can be used effectively.

2. Force-Decomposition Method
The next parallel MD algorithm discussed here is based on a blockdecomposition of the force matrix rather than the row-wise decomposition
used in the RD algorithm. The partitioning of the force matrix F is shown in
Fig. 11 and the algorithm is called the force-decomposition (FD) algorithm
[37]. The method has its origin in block-decomposition of matrices, which is
commonly encountered in linear algebra algorithms for parallel machines.
The block-decomposition, shown in Fig. 11, is actually applied on a
permuted force matrix F0 , which is formed by rearranging the columns of the
original F in a particular way. The (ij) element of F is the force acting on
atom i in vector x due to atom j in the permuted vector x0 . Now the Fz0
subblock owned by each processor z is of size (N/P1/2)  (N/P1/2). As shown

FIG. 11 The division of the force matrix among P processors in the forcedecomposition method.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

213

in the figure, to compute the nonbonded forces in Fz0 , processor z must know
one N/P1/2-length piece of each of the x and x0 vectors, i.e., the subvectors xa
and x0b . As these elements are computed they will be accumulated into
the corresponding subblocks fa and fb0 . The subscripts a and b each run
from 0 to P1/2 and reference the row and the column position occupied by
processor z.
As in the RD algorithm, each processor has updated copies of the atom
positions xa and x0b needed at the beginning of the time step. A single time
step of the FD algorithm consists of the following steps:
1.

2.

3.
4.

The first step is the same as that of the RD algorithm. First the
nonbonded forces Fz0 are computed. The result is summed into
both fa and fb0 . Next, each processor computes a fraction N/P of
the bonded interactions. A critical point here is that in a preprocessing step of the run, we should guarantee that each processor knows
all the atom positions needed for the bonded (intramolecular) interactions. This step again scales as N/P.
Step (2) is also the same as that of the RD algorithm. The key
difference is that now the total force on atom i is the sum of the
elements in row i of the force matrix minus the sum of the elements in
column i0 , where i0 is the permuted position of column i. Thus this
step performs a fold of fa ( fb0 ) within each row (column) of processors
to sum-up these contributions. The important point is that now the
vector fa ( fb0 ) being folded is only of length (N/P1/2) and only the P1/2
elements in one row (column) are participating in the fold. Thus, this
operation scales as N/P1/2 instead of N as in the RD algorithm.
Finally, the two contributions are jointed to yield the total forces fz
( fz0 ) on the atoms owned by processor Pz.
The processor can now perform the integration for its own atoms,
as in the RD case. This operation scales as N/P.
Step (4) shares these updated positions with the processors that
need them for the next time step. As with the fold operation the
processors in each row (column) expand their xa (x0b ) subvectors
within the row (column) so that each acquires the entire xa (x0b ).
This operation scales again as N/P1/2 instead of N, as in the RD
algorithm.

The FD algorithm, as the RD algorithm, divides the force computation

evenly across all processors. The key point is that the communication and
memory costs in steps (2) and (4) scale as N/P1/2, rather than as N as in the
RD algorithm. When large number of processors are used, this can be very
important. At the same time, although more steps are needed, the FD
algorithm retains the overall simplicity and structure of the RD method.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

214

Harmandaris and Mavrantzas

3. Domain-Decomposition Method
The third parallel method discussed here for MD simulations of systems
with short-range interactions is domain-decomposition (DD) method
[31,3840]. In this method the physical simulation box is divided into
small 3D boxes, one for each processor. The partitioning of a simulation
box of length L in a DD algorithm is shown in Fig. 12 (2D projection). Each
processor z owns a sub-box labeled Bz with edge length Lz (Lz L/P) and
updates the positions of all atoms within its own box, xz, at each time step.
Atoms are reassigned to new processors as they move through the physical
domain. In order to compute the forces on its atoms, a processor must
know the positions of atoms in nearby boxes (processors), yz. Thus the
communication required is local in contrast to global in the AD and FD
algorithms. As it computes the force fz on its atoms, the processor also
computes the components of forces fzn on the nearby atoms.
A single time step of a DD algorithm consists of the following steps:
1.

2.

3.

For each processor Pz, the first step concerns the calculation of
bonded and nonbonded forces for atoms within box Bz. This step
scales with the numbers of atoms N/P per processor.
In step (2) the forces gz are shared with the processors owning neighboring boxes. The received forces are summed with the previously
computed fz to create the total force on the atoms owned by the
processor. The amount of data exchanged in this operation (and
consequently the scaling of this step) is a function of the force
cutoff distance and box length.
After computing fz, the atomic positions xz are updated. This
operation also scales as N/P.

FIG. 12

Partitioning of the simulation domain in a DD algorithm.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

4.

5.

215

Next the updated positions are communicated to processors owning

neighboring boxes so that all processors can update their yz list of
nearby atoms.
Finally, periodically, atoms that have left box Bz are moved into the
appropriate new processor.

The scaling of steps (1) and (3) in the DD algorithm is again the optimal
N/P. The communication cost involved in steps (2), (4), and (5) is more
complicated. It proves to be dependent on the relative value of the cutoff
distance rc in comparison with the subdomain edge length Lz [31]. More
specifically, if Lz>rc the scaling goes as the surface-to-volume ratio
(N/P)2/3. If Lz  rc, then the communication scales as N/P. In practice,
however, in the MD simulations of polymer systems there are several
obstacles to minimizing communication costs in the DD algorithm.
.

If Coulombic interactions are present then, because of the 1/r dependence, long cutoffs should be used. Thus Lz<rc and extra communication is needed in steps (2) and (4). Data from many neighboring
boxes must be exchanged and the communication operation scales as
r3c . Special techniques such as Ewald summation, particleparticle, or
particlemesh methods can be implemented in parallel [41].
As atoms move to new processors in step (5) the molecular connectivity information should be exchanged and updated between processors. This requires extra communication cost, depending on the type
of the bonded interactions.
If macromolecular systems are simulated uniformly in a simulation
domain, then all boxes have a roughly equal number of atoms (and
surrounding atoms); load-balance occurs. This will not be the case if
the physical domain is nonuniform (e.g., for polymers in vacuum or
with surrounding solvent). In this case it is not trivial to divide the
simulation domain so that each processor has an equal number of
atoms. Sophisticated load-balancing algorithms have been developed
[36] to partition an irregular or nonuniformly dense physical domain,
but the result is subdomains which are irregular in shape, or connected in an irregular fashion to their neighboring boxes. In both
cases the communication operation between processors becomes
more costly. If the physical atom density changes with time, then
the subject of load-balance becomes more problematic. Dynamical
load-balancing schemes are again needed, which require additional
computational time and data transfer.

In general, the DD algorithm is more difficult to integrate to existing

serial codes than the RD and FD ones. This fact, coupled with the specific

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

216

Harmandaris and Mavrantzas

problems for macromolecular systems, has made DD implementations less

common than RD in the simulations of polymer systems. For simulations
with very large systems, however, and in terms of optimal communication
scaling, DD is more advantageous. It allows distribution of the computational work over a large number of processors, and this explains why it is
nowadays preferred in commercial MD packages such as LAMMPS [41,42],
NAMD [43], and AMBER.

B. EfficiencyExamples
Examples of applications of the above techniques can be found in various
articles that describe parallel MD applications for systems consisting of
many thousands (or even millions) of atoms performed on a large number of
processors [31,40]. Here we focus on the implementation of the parallel
algorithm for the united-atom polyethylene melt MD simulations described
in the previous section [44].
A critical point in any parallel implementation of an existing serial code
is the percentage of the code that can be parallelized. To measure the
performance of parallel implementations of existing sequential algorithms,
the speed-up parameter S is used. This is defined as the ratio of the execution
time of the serial (sequential) algorithm on a single processor to the
execution time of the parallel algorithm running on P processors
S P

s
p

83

where s and p denote the execution time of the algorithms on one and P
processors, respectively. For a fully parallel code running on P processors,
p s/P and S(P) P (linear speed-up).
A typical example where parallel implementation of atomistic MD
algorithms can substantially speed-up code execution is in NTLx yy zz
simulations described in Section VI.B (stress relaxation simulations). In this
case the simulated systems are independent and parallelism is straightforward to achieve by assigning each relaxing configuration to a different node
(processor). Since no data communication between different systems is
required in this case, excellent speed-up should be achieved.
This has been verified with trial runs on a Cray T3E 900 machine at the
Edinburgh Parallel Computing Center (EPCC) using the standard MPI
approach [45]. Figure 13 presents a graph of the speed-up of the (parallel)
code for a number of simulations, each one being executed with a model
system containing 40 chains of C78 PE melt. As was expected, the speed-up
is practically perfect (linear).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

217

FIG. 13 Speed-up graph of the parallelization of the NTLx yy zz MD simulation

runs and the optimal (linear) speed-up.

Figure 14 presents the corresponding speed-up graph for a system

containing 24 chains of a C1000 PE melt (a total of 24,000 atoms). The runs
were executed on a shared memory machine (Sun Enterprise 3000/3500
cluster at EPCC) using OpenMP Fortran [46], which is suitable for
developing parallel MD applications in machines with shared memory
architecture. Results are shown from parallel runs with different numbers of
processors, ranging from 1 to 8 (dotted line) [44]. For comparison, the
optimal (linear) speed-up is also shown (solid line). A speed-up of 5.5 is
easily reached by using just eight processors.

C. Parallel Tempering
A trend in modern MD (and MC) simulations is to enhance system
equilibration at low temperatures by the use of novel parallel techniques.
One such technique which has recently attracted considerable attention in
different variations is parallel tempering (PT) [47,48]. PT was introduced
in the context of spin glass simulations, but the real efficiency of the method
was demonstrated in a variety of cases, such as in the study of the
conformational properties of complex biological molecules [49,50]. Sugita

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

218

Harmandaris and Mavrantzas

FIG. 14 Speed-up graph of the parallelization of the NVT MD simulation runs

(dotted line). Also shown is the optimal speed-up (solid line).

and Okamoto [51] used a similar method called replica exchange molecular
dynamics to simulate protein folding and overcome the multiple-minima
problem. They also used multiple-histogram reweighting techniques to
calculate thermodynamic quantities in a wide temperature range. Recently,
Yamamoto and Kob [52] used the same method to equilibrate a twocomponent Lennard-Jones mixture in its supercooled state. They found that
the replica-exchange MC method is 10100 times more efficient than the
usual canonical molecular dynamics simulation.
PT has also been used successfully to perform ergodic simulations with
Lennard-Jones clusters in the canonical and microcanonical ensembles
[53,54]. Using simulated tempering as well as PT, Irback and Sandelin
studied the phase behavior of single homopolymers in a simple hydrophobic/hydrophilic off-lattice model [55]. Yan and de Pablo [56] used
multidimensional PT in the context of an expanded grand canonical
ensemble to simulate polymer solutions and blends on a cubic lattice. They
indicated that the new algorithm, which results from the combination of a
biased, open ensemble and PT, performs more efficiently than previously
available techniques. In the context of atomistic simulations PT has been
employed in a recent study by Bedrov and Smith [57] who report parallel

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

219

tempering molecular dynamics simulations of atomistic 1,4 polybutadiene

polymer melts, in the 323473K temperature domain at atmospheric
pressure. It has also been employed in the context of MC strategies in
order to access the regime of low temperatures in the simulations of cis-1,4
polyisoprene melts [58]. In both of these works, it was shown that for a
polymer melt well above the glass transition temperature, PT ensures a
thorough sampling of configuration space, which is much more effective
than afforded by conventional MD or MC simulation methods.
In PT, not a single system but an extended ensemble of n systems, labeled
as i 1, . . . ,n, is considered. Each such system is viewed as a copy of the
original system, equilibrated, however, at a different temperature Ti,
i 1, . . . ,n, with T1<T2<  <Tn. The partition function of this extended
ensemble is given by
Q

n
Y

Qi

84

i1

where Qi denotes the individual partition function of the ith system in its
relevant statistical ensemble. The strategy of simultaneously equilibrating
not a single but a number of systems at different temperatures accelerates the
equilibration of the lowest temperature systems by accepting configurations
from the higher temperature systems, for which the rate of the system equilibration is higher. Thus, in the PT method, configurations are swapped
between systems being equilibrated at adjacent temperatures. The acceptance probability of a swapping move between system configurations
i and j i 1 within a parallel tempering series of NPT simulations is
given by:
acci, j ! j, i min1, expfi Uj Pj Vj
j Ui Pi Vi


j Ui Pi Vi j Uj Pj Vj g
min1, expfU PV g

85

with U, P, and V symbolizing the potential energy function, pressure, and

volume, respectively [58].
For swapping to occur between two systems, it is important that their
U PV instantaneous enthalpy histograms overlap, since only then can
there be a nonzero probability to exchange configurations. In this case, the
success rate of configuration swapping can increase by performing the
simulation with smaller-size systems characterized by larger (U PV)
fluctuations or with systems whose temperatures are not too wide apart.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

220

Harmandaris and Mavrantzas

PT is ideal to implement in parallel architectures by assigning each system

to a single node, and by using standard MPI libraries [45] for inter-node
communication.

REFERENCES
1. Binder, K., Ed. Monte Carlo and Molecular Dynamics Simulations in Polymer
Science; Oxford University Press: New York, 1995.
2. Allen, M.P.; Tildesley, D.J. Computer Simulation of Liquids; Oxford University
Press: Oxford, 1987.
3. Frenkel, D.; Smith, B. Understanding Molecular Simulations: From Algorithms
to Applications; Academic Press: New York, 1996.
4. Ciccoti, G.; Hoover, W.G. In Molecular Dynamics Simulations of Statistical
Mechanics Systems, Proceedings of the 97th International Enrico Fermi
School of Physics, North Holland, Amsterdam, 1986.
5. Goldstein, H. Classical Mechanics; Addison-Wesley: Reading, MA, 1980.
6. Swope, W.C.; Andersen, H.C.; Berens, P.H.; Wilson, K.R. J. Chem. Phys.
1982, 76, 637649.
7. Nose, S. Mol. Phys. 1984, 52, 255268.
8. Hoover, W.G. Phys. Rev. A 1985, 31, 16951697.
9. Andersen, H.C. J. Chem. Phys. 1980, 72, 23842393.
10. Parrinello, M.; Rahman, A. Phys. Rev. Lett. 1980, 45, 11961199.
11. Berendsen, H.J.C.; Postma, J. P.M.; van Gunsteren, W.F.; DiNola, A.;
Haak, J.R. J. Chem. Phys. 1984, 81, 36843690.
12. Yang, L.; Srolovitz, D.J.; Yee, A.F. J. Chem. Phys. 1997, 107, 43964407.
13. Harmandaris, V.A.; Mavrantzas, V.G.; Theodorou, D.N. Macromolecules
2000, 33, 80628076.
14. Palmer, B.J. J. Comp. Phys. 1993, 104, 470472.
15. Tuckerman, M.; Berne, B.J.; Martyna, G.J. J. Chem. Phys. 1992, 97,
19902001.
16. Martyna, G.J.; Tuckerman, M.E.; Tobias, D.J.; Klein, M.L. Mol. Phys. 1996,
87, 11171157.
17. Edberg, R.; Evans, D.J.; Morriss, G.P. J. Chem. Phys. 1986, 84, 69336939.
18. Ryckaert, J.P.; Ciccoti, G.; Berendsen, H.J.C. J. Comp. Phys. 1977, 23,
327341.
19. Andersen, H.C. J. Comp. Phys. 1983, 52, 2434.
20. Harmandaris, V.A.; Mavrantzas, V.G.; Theodorou, D.N. Macromolecules
1998, 31, 79347943.
21. Harmandaris, V.A.; Doxastakis, M.; Mavrantzas, V.G.; Theodorou, D.N.
J. Chem. Phys. 2002, 116, 436446.
22. Harmandaris, V.A.; Mavrantzas, V.G.; Theodorou, D.N.; Kroger, M.;
Ram rez, J.; Ottinger, H.C.; Vlassopoulos, D. Macromolecules 2003,
36,13761387.
23. Mavrantzas, V.G.; Theodorou, D.N. Macromolecules 1998, 31, 63106332.
24. Mavrantzas, V.G.; Ottinger, H.C. Macromolecules 2002, 35, 960975.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Dynamics Simulations

221

25. Doi, M.; Edwards, S.F. The Theory of Polymer Dynamics; Clarendon: Oxford,
1986.
26. Milner, S.T.; McLeish, T.C.B. Phys. Rev. Lett. 1998, 81, 725728.
27. Lodge, T.P.; Phys. Rev. Lett. 1999, 83, 32183221.
28. Ferry, J.D. Viscoelastic Properties of Polymers; J. Wiley & Sons: New York,
1980.
29. Uhlherr, A.; Doxastakis, M.; Mavrantzas, V.G.; Theodorou, D.N.; Leak, S.J.;
Adam, N.E.; Nyberg, P.E. Atomic structure of a high polymer melt. Europhys.
Lett. 2002, 57, 506511.
30. Kroger, M.; Ram rez, J.; Ottinger, H.C. Polymer 2002, 43, 477487.
31. Plimpton, S. J. Comp. Phys. 1995, 117, 119.
32. Gupta, S. Comp. Phys. Comm. 1992, 70, 243270.
33. Smith, W. Comp. Phys. Comm. 1991, 62, 229248.
34. Smith, W.; Forester, T.R. Comp. Phys. Comm. 1994, 79, 5262.
35. Fox, G.C.; Johnson, M.A.; Lyzenga, G.A.; Otto, S.W.; Salmon, J.K.; Walker,
D.W. Solving Problems on Concurrent Processors; Prentice Hall: Englewood
Clis, NJ, 1988; Vol. 1.
36. Deng, Y.; Peiers, R.; Rivera, C. J. Comp. Phys. 2000, 161, 250263.
37. Hendrickson, B.; Plimpton, S. J. Parr. Distr. Comp. 1995, 27, 1525.
38. Clark, T.W.; Hanxleden, R.V.; McCammon, J.A.; Scott, L.R. In Proc.
Scalable High Performance Computing Conference94, IEEE Computer
Society Press: Los Alamitos, CA, 1994; 95102 pp.
39. Brown, D.; Clarke, J.H.R.; Okuda, M.; Yamazaki, T. Comp. Phys. Commun.
1994, 83, 113.
40. Putz, M.; Kolb, A. Comp. Phys. Commun. 1998, 113, 145167.
41. Plimpton, S.J.; Pollock, R.; Stevens, M. Particle-Mesh Ewald and rRESPA
for Parallel Molecular Dynamics Simulations, In Proceedings of the Eighth
SIAM Conference on Parallel Processing for Scientic Computing, SIAM,
Minneapolis, MN, 1997.
42. http://www.cs.sandia.gov/sjplimp/lammps.html
43. http://www.ks.uiuc.edu/Research/namd
44. Unpublished data collected at the Edinburgh Parallel Computing Centre
(EPCC) under the TRACS programme during the period 10 January
200020 February 2000.
45. http://www.mpi-forum.org/
46. http://www.openmp.org/
47. Newman, M.E.J.; Barkema, G.T. Monte Carlo in Statistical Physics;
Clarendon: Oxford, 1999.
48. Geyer, C.J. Computing Science and Statistics, Proceedings of the 23rd
Symposium on the Interface, Interface Foundation of North America: Fairfax
Station, VA, 1991, p. 156.
49. Hansmann, U.H.E. Chem. Phys. Lett. 1997, 281, 140150.
50. Wu, M.G.; Deem, M.W. Mol. Phys. 1999, 97, 559580.
51. Sugita, Y.; Okamoto, Y. Chem. Phys. Lett. 1999, 314, 141151.
52. Yamamoto, R.; Kob, W. Phys. Rev. E 2000, 61, 54735476.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

222

Harmandaris and Mavrantzas

53. Neirotti, J.P.; Calvo, F.; Freeman, D.L.; Doll, J.D. J. Chem. Phys. 2000, 112,
1034010349.
54. Calvo, F.; Neirotti, J.P.; Freeman, D.L.; Doll, J.D. J. Chem. Phys. 2000, 112,
1035010357.
55. Irback, R.; Sandelin, E. J. Chem. Phys. 1999, 110, 1225612262.
56. Yan, Q.; de Pablo, J.J. J. Chem. Phys. 2000, 113, 12761282.
57. Bedrov, D.; Smith, G.D. J. Chem. Phys. 2001, 115, 11211124.
58. Doxastakis, M.; Mavrantzas, V.G.; Theodorou, D.N. J. Chem. Phys. 2001,
115, 1135211361.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

7
Configurational Bias Techniques for
Simulation of Complex Fluids
T. S. JAIN and J. J. DE PABLO
Madison, Wisconsin, U.S.A.

I.

University of WisconsinMadison,

INTRODUCTION

Monte Carlo methods offer a useful alternative to Molecular Dynamics

techniques for the study of the equilibrium structure and properties,
including phase behavior, of complex fluids. This is especially true of
systems that exhibit a broad spectrum of characteristic relaxation times; in
such systems, the computational demands required to generate a long
trajectory using Molecular Dynamics methods can be prohibitively large. In
a fluid consisting of long chain molecules, for example, Monte Carlo
techniques can now be used with confidence to determine thermodynamic
properties, provided appropriate techniques are employed.
A generic algorithm for carrying out a Monte Carlo simulation
constructs a weighted walk in the configuration space of the system, and
samples different states according to their equilibrium probability distribution function. This is typically achieved by proposing a small, random
change (about some original state) in one or more degrees of freedom of the
system, which is subsequently accepted or rejected according to specific
probability criteria [1] pertaining to the statistical ensemble of interest.
In a canonical ensemble, for example, random trial displacements of the
particles positions would be accepted according to
Pacc min1, exp
U

where U is the change in energy in going from one state of the system to
another, and  is the inverse of kBT, with kB being Boltzmanns constant
and T the temperature.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

224

Jain and de Pablo

For many complex systems, blind, or naive random displacements

of the particles coordinates can lead to inefficient algorithms. Poor
sampling can in turn yield simulated properties that are not representative
of the system under study. Configurational bias techniques provide a
fairly general means of biasing trial moves in such a way as to guide the
system to favorable states, thereby sampling configuration space more
effectively.
We begin this chapter by providing a few illustrative examples and by
discussing some of the typical problems that can be encountered in a
simulation. We then explain, in general terms, how an arbitrary bias can
be incorporated in a Monte Carlo simulation by introducing the concept
of detailed balance and that of rejection algorithms. We then present a
number of different techniques, all based on biased sampling ideas, that
can facilitate considerably the simulation of complex systems. This is
done in the context of specific examples taken from our own recent work
and that of others. Here we note that the goal of this chapter is not to
provide a review of configurational bias techniques, but rather to introduce these methods to the reader and to present a broad view of how
these ideas can be used in a wide variety of systems to improve sampling
efficiency.

II. SHORTCOMINGS OF METROPOLIS SAMPLING

As mentioned above, the main drawback of a naive, Metropolis-sampling
based simulation is that trial moves are proposed at random, without ever
taking into account the specific nature of the system. Unless the random
trial moves are small, most of them will be rejected. If the trial displacements
are small, trial moves are more likely to be accepted, but successive
configurations end up being highly correlated; the averages obtained from
the resulting trajectories may not provide an accurate description of the
system as a whole. This problem can be particularly acute in molecular
systems, where topological constraints (e.g., as in long polymeric molecules)
cause most random displacements of an entire molecule to be rejected. In the
case of polar systems (e.g., water), a molecule in a new, trial position (and
orientation) must not only conform to its new geometrical environment but,
if the move is to have a reasonable probability of being accepted, the
energetic contributions arising from polar or electrostatic effects must be
comparable to those for the original state.
The above examples serve to illustrate that for complex systems, or
even for simple systems at elevated densities, a naive random sampling of
configuration space does not provide an efficient simulation algorithm. In

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Configurational Bias Techniques

225

order for the system to evolve to different states, it will have to be guided
out of local energy minima by biasing proposed trial moves. These biases
will in general be system specific; they can be designed in such a way
as to influence the position of a molecule, its orientation, or its internal
configuration. Whatever the bias is, however, it must be subsequently
removed in order to generate the desired, correct distribution of states
dictated by the ensemble of interest. The next section discusses how this can
be achieved.

III. DETAILED BALANCE AND CONFIGURATIONAL

BIAS
The condition of detailed balance ensures that the states of the system are
sampled according to the correct probability distribution in the asymptotic
limit, i.e., in the limit of a large number of proposed transitions. The
condition of detailed balance [2] simply ensures that the probability of going
from one state of the system to another is the same in the forward and
reverse directions. To introduce the concept of detailed balance, we denote
the overall state of a system (e.g., the coordinates of all of the particles
for the canonical, NVT ensemble or the coordinates of all of the particles
and the volume of the system for the NPT ensemble) collectively by X. This
probability can be constructed as the product of two probabilities: the
probability of being in a particular state X, and the conditional probability
of moving from that state to another state Y. This can be written as
KXjY f X KYjX f Y

where f(X) is the probability of being in state X and K(X|Y) is the probability of moving to Y given that the system is at X. The latter probability
can be further decomposed into the product of two factors: the probability
of proposing the transition, T(X|Y), and that of accepting the transition,
A(X|Y). With this separation, Eq. (2) becomes:
AXjYTXjY f X AYjXTYjX f Y

The way in which the moves are proposed is arbitrary; any function T(X|Y)
can be used as long as it is properly normalized, i.e.,
Z
TXjY dY 1

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

226

Jain and de Pablo

In order to satisfy detailed balance, the function A(X|Y) must have a specific
form. In the prescription of Metropolis et al. [1], moves are accepted with
probability



TYjX f Y
AXjY min 1,
TXjY f X

where min(x, y) is used to denote the smaller of its two arguments. It can
be shown that in the limit of a large number of trial moves, the sequence of
states generated by accepting trial moves according to Eq. (5) follows the
desired probability distribution function f.
In the particular case of a naive algorithm, i.e., the case in which
displacements are proposed at random, one can see that the transition
probabilities T(X|Y) and T(Y|X) are the same. For a canonical, NVT
ensemble, the function f takes the form
f X

1
exp
UX
Q

where Q is the partition function of the system. For the canonical ensemble
Eq. (5) therefore reduces to Eq. (1).
In a Configurational Bias algorithm, a bias is introduced in the forward
or reverse direction; as a result of the bias, the forward and reverse
transition probabilities are no longer equal to each other. This leads to
additional flexibility in the construction of a trial move, and can result in
considerable improvements of sampling efficiency. This bias is removed in
the acceptance criteria according to Eq. (5), thereby leading to sampling
according to the correct, original equilibrium probability distribution f.
A few illustrative examples on the use of Eq. (5) are discussed in the next
section. In each case, the correct acceptance criteria are arrived at by
establishing the form of the functions T(X|Y) and T(Y|X).

IV. CASE STUDIES

A. Orientational Configurational Bias
Orientational bias moves are useful for the study of fluids that exhibit highly
directional forces. A prime example of such a fluid is water, where strong
hydrogen bonding interactions lead to well defined, relatively stable lowenergy molecular structures. Cracknell et al. [3] were among the first to use
orientational bias for the study of water. A typical orientational bias

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Configurational Bias Techniques

227

algorithm consists of the following steps:

1.
2.

3.

4.

Displace the center of mass of the molecule at random to a new

position
If the potential energy can be separated into orientation-dependent
and -independent parts, calculate the orientation independent part,
0
.
Unew
Generate a set of k orientations OF {o1, o2, . . . ,ok} about the new
center of mass for the forward move. Calculate the orientation
dependent potential energy Uo(l ) for each orientation in OF. Note
that henceforth, superscripts F and R will be used to denote corresponding quantities for the forward and reverse moves respectively.
Evaluate the so-called Rosenbluth weight for the new position as
follows
RFW

k
X

exp
U o l

l1

5.

Select one orientation j, from the set OF, according to the probability
distribution
p j

exp
U o j
RFW

This implies that

TXjY p j
6.

For the reverse move construct a new set of k
1 orientations

OR {o1, o2, . . . ,ok
1} and calculate the orientation dependent
energy of the molecule in its old, or original orientation Uo(k).
Calculate the Rosenbluth weight for the old position:
RR
W

k
1
X

exp
U o l exp
U o k

10

l1

For the reverse move we have

TYjX

exp
U o k
RR
W

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

11

228

7.

Jain and de Pablo

This leads to the following acceptance criteria for this move




RF
0
0
exp

U


U

AXjY min 1, W
new
old
RR
W

12

where subscripts old and new are used to denote quantities evaluated
in the original (old) and the trial (new) configurations, respectively.
The above algorithm can be applied to both lattice and continuum
simulations. One important difference, however, must be noted. In the
case of a lattice, all possible trial orientations of the molecule can be
considered explicitly; the sets OF and OR are therefore the same (on a
cubic lattice, a simple dimer or a bond can adopt an orientation along the
coordinate axes, and n 6). In a continuum, a molecule can adopt one of
an infinite number of trial orientations. By restricting the trial orientations
to the finite sets OF and OR, the probability of generating these sets also
enters the transition matrix along with the probability of choosing
a direction from the probability distribution (which itself is a function
of these sets [4]). The acceptance criteria are therefore also a function of
these two sets:

TYjXPOR f Y
AXjY, OF , OR min 1,
TXjYPOF f X

13

The a priori probabilities, P(OF) and P(OR), of generating these sets of trial
orientations are the same and drop out of the final expression. In order to
satisfy detailed balance, the condition of super-detailed balance is
imposed, i.e., detailed balance should be obeyed for every pair of sets
chosen. Hence the acceptance criteria of Eq. (12) generate the true equilibrium distribution.

1. Continuum Example: Simulation of Water Clay Hydrates

Bulk water has been the subject of a large number of simulation studies.
Water in confined environments has been studied less extensively and, more
importantly to the goals of this chapter, it represents a case where
orientational bias moves are particularly helpful.
The question at hand is to determine the structure of water in the
presence of geometrical constraints, at elevated densities. For concreteness,
we consider an example taken from our studies of clay hydrates. At small
length scales, these hydrates can be viewed as a stack of closely spaced disklike particles, having a thickness of a few nanometers. The disks are

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Configurational Bias Techniques

229

negatively charged, and electroneutrality is maintained by the presence of

compensating cations such as sodium or calcium. The interstitial spaces
between different disks (also called galleries) are occupied by water;
depending on pressure, temperature, and ionic strength, more or less water
can enter the galleries and swell the hydrate. Configurational bias is helpful
because the presence of charged species and strong confinement restrict
considerably the freedom of water molecules.
Figure 1 depicts a typical configuration of a hydrated sodium
montmorillonite clay [5]. The structure and partial charges of the model

FIG. 1 Representative equilibrium configuration of a system comprising two clay

sheets, twelve sodium ions, and interlayer water. Sodium ions are not shown for
clarity.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

230

Jain and de Pablo

FIG. 2 Swelling curve for two clay sheets of sodium montmorillonite from NPzzT
simulations.

clay were taken from Skipper et al. [6]. The interaction parameters for the
claywater system were based on the TIP4P and MCY models of water
[79]. Simulations suggest that the swelling of clays occurs through the
formation of discrete layers of water. The resulting basal spacings as a
function of number of water molecules (Fig. 2) agree with experiment. The
use of a grand canonical formalism, in which water is inserted into the clay
galleries via a configurational bias algorithm, permits calculation of the
water content of the clay (Fig. 3b) as a function of thermodynamic
conditions. The disjoining pressure can also be calculated as a function of
basal distance to provide reliable estimates of the mechanically stable basal
spacings (Fig. 3a), which is important in many applications of clays. In
addition to providing information on the swelling of the clay, these simulations reveal that, in general, the compensating cations (e.g., sodium or
calcium) are fully hydrated (Fig. 4). Arriving at such conclusions on the
basis of conventional Monte Carlo simulations would require considerable
computational resources; using an orientational bias increases efficiency
considerably and lends credence to the results.

B. Configurational Bias (CB) for Articulated or

Polymeric Molecules
Polymer molecules can be studied at various levels of detail, from highly
coarse-grained lattice representations to fully atomistic models. Whatever
the level of detail, however, the connectivity of a complex molecule imposes

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Configurational Bias Techniques

231

FIG. 3 Grand canonical ensemble simulation results at T 300 K and chemical

potential 
17.408, which is the value that corresponds to TIP4P water in the
bulk (at T 300 K and P 1 bar). (a) Normal component of the pressure tensor. (b)
Average number of water molecules per clay.

severe constraints on the types of trial moves that can be used to generate
distinct configurations of the system. At liquid-like densities, it is clear that a
random displacement of the center of mass of a molecule (while maintaining
the internal configuration) would lead to overlaps with other molecules, and
a rejection of the trial move.
This problem can be partially alleviated by conducting highly localized
trial moves. On a simple cubic lattice, for example, the conformation of a
linear polymeric molecule can be explored by resorting to local trial
moves such as kink-jumps, end-rotations, or crankshafts[10]. These
moves typically displace one or two sites of the chain. The so-called
reptation [11] moves involve displacing an entire chain in a slithering,
snake-like motion along its contour. In a continuum, simple random moves
of individual sites can be used to sample the configuration space of the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

232

Jain and de Pablo

FIG. 4 Distribution of oxygens, hydrogens, and sodium ions inferred from an

NPzzT simulation at Pzz 1 bar and T 300K, using two clay sheets: (a) Nw 40, (b)
Nw 64.

system. These moves, however, are generally insufficient to achieve

appreciable relaxation of the intramolecular conformation of long chain
molecules [12]. Configurational bias ideas can be used to probe energetically
favorable configurations over length scales of several sites, thereby
accelerating the relaxation of configurations.
In its simplest form, a configurational bias trial move for a linear polymer chain involves cutting off a terminal part of the chain, and regrowing
the end sites using an energetic bias (Fig. 5). Such a move is inspired by
the seminal work of Rosenbluth and Rosenbluth [13]. In the canonical
ensemble, the algorithm for this move is as follows [14,15]:
1.
2.
3.

Select one end of the chain of length L at random and remove n

interaction sites.
Use i to denote the current site to be added back to the chain.
Calculate the energy of site i, in k trial positions.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Configurational Bias Techniques

FIG. 5

4.

233

Illustration of configurational bias move.

Select one of the trial positions, j, from the probability distribution

given by
j

e
Ui
pi j Pk

Uil
l1 e
5.

14

After all the sites have been grown calculate the Rosenbluth weight
of the chain as follows:
RFW

L
Y

k
X

e
Ui

15

iL
n1 l1

The construction of the move was such that

TXjY

L
Y

pi j

16

iL
n1

6.

7.

Similarly, to construct the reverse move, a new set of k
1 orientations is generated about the old positions, and the Rosenbluth factor
of the original configuration of the molecule, RR
W , is calculated.
Substitution into Eq. (5) gives acceptance criteria of the form:



RF
AXjY min 1, W
RR
W

17

A number of remarks are in order. First, we note that there is no

restriction on the number of sites that can be cut and regrown for this move.
In fact, an entire chain can be removed and regrown at another point if so

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

234

Jain and de Pablo

desired (such a large trial move, however, would be computationally

demanding and inefficient at elevated densities). Second, we note that a
piece of the chain could be cut from one side of the chain and appended
to the other end of the molecule. Third, it is important to point out that
to arrive at Eq. (17) care must be exercised in the construction of a
hypothetical, reverse move that would lead to the original configuration of
the molecule (the calculation of RR
W ). The question to ask at a given state of
the system is: If we had used a (hypothetical) configurational bias move to
arrive at the precise, current state of the system, what would the Rosenbluth
weight of the molecule in question have been? The answer is of course RR
W,
but the concept of conducting a hypothetical process to recreate the original
state of the molecule can be difficult to grasp.
The algorithm outlined above is equally applicable to chains on a lattice
or chains in a continuum. In the particular case of models having strong
intramolecular interactions (e.g., bond-stretching or bond-bending potential-energy functions), minor alterations of this algorithm can be implemented to improve sampling. The k trial orientations for i can be
generated using the Boltzmann distribution arising from the intramolecular
potential [16]
bondedl

pli Ce
Ui

l 1, 2, . . . , k

18

where C is a normalization constant. For each of these trial positions, the

external Boltzmann factor arising from intermolecular interactions, external
fields, and intramolecular nonbonded interactions is calculated. The subsequent procedure to be followed is the same as above, but the Rosenbluth
weight is calculated from the sum of the external Boltzmann factors. For
the reverse move, a new set of k
1 orientations is generated and the
Rosenbluth weight is calculated. The arguments of super-detailed balance
presented for the Orientational Bias algorithm in the continuum are applicable to the growth of each site. The acceptance criteria reduce to Eq. (17).

1. Expanded Grand Canonical Ensemble Simulation of

Polymer Chains Using Configurational Bias
In the case of long polymer chains, the insertion or deletion of entire
molecules required for grand canonical ensemble simulations is difficult,
even when configurational bias techniques are employed. In that case it is
beneficial to implement configurational bias moves in the context of an
expanded ensemble formalism [17], which essentially allows one to create or
delete a smaller number of sites of a molecule (as opposed to an entire chain)
in each trial move, thereby increasing the likelihood of acceptance. The

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Configurational Bias Techniques

235

combination of an expanded grand canonical technique with configurational bias can lead to highly effective insertion schemes, even for long chain
molecules at elevated densities [18].
Formally, an expanded grand canonical ensemble can be defined through
the partition function:


1 X
M
X

QN, y, V, T exp

y exp N

19

N0 y1

where is the chemical potential, and Q is the canonical partition function

for a system of volume V, temperature T, and N full molecules and one
tagged molecule in state y. States 1 and M correspond to a fully decoupled
and a fully coupled molecule, respectively. The ys are arbitrary weighting
factors whose relevance is discussed below. The probability of finding the
system in a configuration described by a particular set of V, T, N, y is given by
p y

1
exp N exp
UN, y exp


20

The ensemble average probability of finding the system in a state given

by N, y is

 QN, V, T, y
exp N exp
p y


21

The goal of the algorithm is to sample fluctuations of the number of molecules effectively. If the molecules of interest consist of L sites, this can be
done by allowing a tagged chain to fluctuate in length; at the two extremes
of the spectrum these fluctuations can cause the tagged chain to comprise L
interaction sites and become a real, full chain, or they can cause the tagged
chain to have zero sites and disappear. In order for these sequential insertion
or deletion processes to be effective, the probability with which terminal
states of the tagged chain are visited should be appreciable. This can
be controlled by judicious assignment of the weighting factors appearing
in Eq. (19). The ratio of the probabilities for two states x and y of the system
is given by


px

 exp
p y

QN, V, T, x
QN, V, T, y

22

The logarithm of the ratio of the partition functions, under identical values
of N, V, and T, is equal to the difference in the chemical potential for a

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

236

Jain and de Pablo

transition from state x to state y. In order to avoid a bottleneck at any

one state, we can prescribe that the probabilities of visiting different states of
the expanded ensemble be equal, which leads to
x

 ex x ! y  ex x, N
 ex y, N

23

where ex( y, N) denotes the excess chemical potential of a system having N

regular molecules and a tagged chain in state y.
A natural choice for the weights y is therefore:
y

3 

 Ny
 y, N y  N y 
ln
V
ex

ex

24

where ex(N) is the chemical potential required to insert an entire chain in a

system with N full chains, and Ny is the number of chains in the system
including the tagged chain if it is of nonzero length. The physical significance of the weights y is now apparent: They are chemical potentials for
particular states of the growing chains. The prefactors y should satisfy
1 0 and M 1. These factors can be approximated as y Ly/L for
homonuclear, monodisperse polymer chains, where Ly is the length of the
polymer molecule in state y. This approximation is based on the result that,
in a simple polymer melt, the incremental chemical potential (chemical
potential required to insert one additional site in the chain) of a polymer
chain is relatively insensitive to chain length [19,20]. In other cases, they can
be determined from a preliminary run so that eventually each state is uniformly sampled. If the number of states of the system is reduced to two, then
the method reduces to the simple Grand Canonical method.
The algorithm for performing a trial move in an expanded grand
canonical ensemble is as follows:
1.
2.

3.
4.

Given that the system is in state x, propose a move to a neighboring

state y.
If the transition involves
a. Adding n sites to the chain, then calculate the Rosenbluth weight
of the added sites as discussed previously.
b. Removing n sites from the chain, then calculate the Rosenbluth
weight of the existing sites.
Normalize the Rosenbluth weight by dividing by nk, where k is the
number of trial positions proposed.
The modified acceptance criteria for the move are

AXjY min 1, R
W exp

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

25

Configurational Bias Techniques

237

where  1 if the move involves a growth process and 
1 if

the move involves a deletion.
The choice of intermediate states for the expanded ensemble can vary for
different systems. In certain cases, the interaction of the entire tagged
molecule with the system can be turned on through a gradual progression
from an ideal noninteracting state to a fully interacting one [2123],
characterized by the parameter y.

2. Adsorption of Hard-Core Flexible Chain Polymers

in a Slit-Like Pore
As with the previous example concerning simulations of clay hydrates, a
striking demonstration of the efficiency of an expanded ensemble/configurational bias algorithm is provided by the study of long chain molecules
confined to narrow slit pores. One of the goals of that work would be to
investigate how polymeric molecules can segregate between a bulk solution
or melt and a porous system. The use of EGCMC can alleviate the problems
associated with the insertion and relaxation of molecules in such a porous
system.
Simulations of hard-core flexible tetramers, hexadecamers, and 100-mers
have been performed by Escobedo and de Pablo [18] to study the equilibrium segregation of large molecules between a pore and bulk fluid.
The width of the pore was kept constant at 5. The pore walls were
impenetrable. Figure 6 shows some of the results. It can be seen that, in spite
of the fact that the packing fractions in the pore are relatively high (above
0.4), EGCMC is able to provide reliable estimates of the concentration of
long chains in the pore. For small chains (tetramers), simple molecular
dynamics simulations can be used to follow the actual diffusion of the
molecules into the pore, and in that case good agreement is found between
EGCMC and MD [24].

3. Critical Behavior in Polymer Solutions

The EGCMC formalism has been used extensively to investigate the phase
behavior of chain molecules. One recent application which is particularly
instructive is concerned with the critical scalings of large-molecular weight
polymer solutions [25]. Polymer solutions (e.g., polystyrene in methylcyclohexane) often exhibit an upper critical solution temperature, whose precise
location depends on molecular weight [26]; the nature of that dependence is
of interest for both theoretical and practical considerations. Using EGCMC
simulations of lattice polymers, it has now been possible to show that the
scaling of the critical density with molecular weight follows a scaling relation
of the form c  MW, with  0.5 in the limit of high molecular weights.
Figure 7a shows simulated and experimental phase diagrams for several

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

238

Jain and de Pablo

FIG. 6 Simulated adsorption isotherms for hard-sphere chains of 4, 16, and 100
sites in a slit-like pore. Reference refers to [24].

polystyrene solutions. It can be seen that simulations are consistent with

experiment, and that a good mapping of the actual data onto a cubic lattice
model can be achieved. The chain lengths examined in simulations are found
to be comparable with, and in many cases longer than, those considered
in experiments. Most experimental studies had indicated that  0.38;
simulations of long chains were necessary to show that a gradual crossover
occurs from 0.38 to 0.5 (Fig. 7b), and that it occurs at large molecular weights (polystyrene of about MW  1,000,000). These findings are
consistent with theoretical arguments [27], and serve to illustrate that even
for large molecules Monte Carlo simulations can be a useful complement to
experimental data, provided appropriate techniques are employed.

C. Topological Configurational Bias

The methods described so far for study of polymeric molecules rely on
the presence of end sites. For long chain molecules or for cyclic or ring
molecules, they are of little use. It turns out that in such cases it is also
possible to construct configurational-bias based algorithms in which the
source of the bias is not only the potential energy of individual sites in
the new, trial position, but also a connectivity constraint. Topological

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Configurational Bias Techniques

239

configurational bias moves are similar in spirit, but different in implementation, to concerted rotation (CONROT) moves [28,29], which are not
elaborated upon here.
In a topological configurational bias move, an arbitrary number of
inner sites of a chain are excised and regrown by taking into account
energetic biases, as well as geometric (topology) considerations that ensure
that connectivity is preserved. Detailed balance is satisfied by determining
the number of fixed-end random walks that exist between two sites of the
molecule. While different schemes may vary in the way in which the number
of random walks, N( j), is estimated, the general idea remains the same.
We begin by discussing the implementation on a lattice, and then describe
several more recent methods suitable for a continuum.

1. Lattice Case
An algorithm to perform such a move on a cubic lattice [30] could be:
1.
2.
3.

A piece of the chain consisting of n sites is cut starting from site s of

the chain. The sites to be regrown are numbered from 1 to n.
The current site to be regrown is labelled i.
Each of the possible k trial orientations for site i are visited and the
following quantity is calculated:
l

N l i, n 1e
Ui

4.

26

where N(l )(i, n 1) is the number of random walks (this quantity is

discussed below) from site i in trial position l to site n 1, and Uil is
the energy of site i in trial position l.
A position, j, is selected for site i from the following probability
distribution:
j

N j i, n 1e
Ui
pi j Pk

Uil
l
l1 N i, n 1e
5.

27

The modified Rosenbluth weight and Random Walk weight for the
chain are constructed as follows:
RFW

n X
k
Y

N l i, n 1e
Ui

i1 l1
n
Y

GFRW

N j i, n 1

i1

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

28
29

240

Jain and de Pablo

6.

To estimate the weight of the original configuration, the same procedure is repeated for the original positions of the sites that were
excised.
The modified acceptance criteria for this move then take the form



RF GR
RW
30
AXjY min 1, W
F
RR
W GRW

7.

In the case of a lattice model, the expression for N( j) is given by [30]:

j

N i,s n


NX

y X
N

x 0

Ns !
x !x x!y !y y!N

x
y !N
x
y !
y0
31

Ns
x
y
z
2
Ns
x
y z

N
2
N

32
33

where x x
x , y y
y, z z
z are the difference in the coordinates between site i and site n 1, Ns is the number of steps, n
i 1, along
the chain between sites i and n 1, and x, y, z are the number of steps along
the positive coordinate axes and x , y , z are the steps along the negative
coordinate axes.
In order to assess the performance of this algorithm, we have studied the
decay of the bond autocorrelation function as a function of CPU time. The
bond autocorrelation function is defined by
bt

hB~t  B~0i
hB~0  B~0i

34

where B~t denotes the vector along a bond of the molecule at time t.
Figure 8 shows the bond autocorrelation function corresponding to a simulation for cyclic chains consisting of 100 sites, in which only local moves
(kink-jump and crankshaft) were employed, and one in which topological
configurational bias moves were implemented. One can clearly see that the
decay in the latter case is much faster, thereby resulting in a much more
efficient simulation algorithm. In the case of ring molecules, care must
be exercised in the selection of the number of sites to be cut and regrown.
This is because the topological nature of the molecule can change (e.g., from
unknotted to knotted) if a large number of sites are selected. In the case
of linear chains this restriction is irrelevant and, in principle, any number

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Configurational Bias Techniques

241

FIG. 7 (a) Phase diagram for solutions of polystyrene in methylcyclohexane. The

lines correspond to results of expanded ensemble simulations, and the symbols refer
to experimental data [26]. (b) Scaling of the critical volume fraction with chain
length. The squares and the circles correspond to literature simulation data [49,50];
the diamonds and the triangles correspond to expanded ensemble simulations for a
simple cubic lattice and for a bond fluctuation model, respectively [25].

of sites can be used in an individual trial move. It turns out that an optimum
number of sites can be identified because the computational expense of the
trial move increases with the number of cut sites. This point is illustrated by
Fig. 9, which depicts the decay of the bond autocorrelation function for
linear chains of 500 sites on a lattice. For a fixed amount of CPU time, the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

242

Jain and de Pablo

FIG. 8 Rate of decay of the bond autocorrelation function (b(t)) for cyclic (ring)
chains of 100 sites, using local moves and topological configurational bias moves, as
a function of CPU time.

decay of the bond autocorrelation function is plotted as a function of the

number of sites used in the topological configurational bias move. As indicated above, there is a clear optimum number of sites for the move. These
simulations correspond to thin polymer films at relatively high volume fractions (  0.95); they are of relevance for the study of glass transition phenomena in polymeric films, where the presence of free interfaces introduces
pronounced departures from bulk behavior [31]. Recently, configurationalbias techniques have also been combined with a density of states formalism
to calculate thermodynamic properties over a wide range of temperature
from a single simulation [32].

2. Continuum Case
In the continuum case, various strategies can be used to achieve closure. In
the case of Extended continuum configurational bias (ECCB) Monte Carlo
[33], a simple geometric constraint is employed to ensure that the chain is
correctly closed. Since the geometric constraint does not take intramolecular

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Configurational Bias Techniques

243

FIG. 9 Decay of the bond autocorrelation function (b(t)) for linear chains of
length 500 as a function of number of sites cut and regrown, for a given amount of
CPU time.

interactions into account, the ECCB method is efficient for molecules with
relatively flexible backbones. This method is described in some detail below.
1.
2.

3.

Select a chain at random to be moved.

Select two sites, a and b, at random within the chain such that the
number of sites between them (including the two chosen sites), n, is
less than or equal to some specified maximum Nmax. The number of
sites to be cut and regrown for a particular move is n.
If one of the sites happens to be an end site, then the usual configurational bias move is performed.
a. If n 1, then a simple crankshaft move is carried out.
b. If the current site to be appended is i, then lower and upper
bounds exist for the angles,
min and
max. These can be determined from the line joining sites i and i
1 with the line joining

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

244

Jain and de Pablo

FIG. 10 Limiting cases for the bounds of the
angle for the two regrowing sites
(1 and 2) involved in an extended configurational bias move (ECCB). (a) Neighbor
sites a and b in the configuration of maximum separation, (b) and (c) neighbor sites a
and b in the configuration of closest approach.

4.

5.

i
1 and b, where b is the first uncut site on the side of the cut
opposite to the growing end (see Fig. 10).
Once the bounds for
have been established, several orientations of
the bond vector bi, i
1 are sampled according to the usual configurational bias method, and one is selected from the probability distribution given by Eq. (14).
The same procedure is repeated for the reverse move. There is a
temporary change in degrees of freedom used to generate the new
configuration because of the restricted sampling of
(Fig. 11). In
order to correct for this, the Jacobian determinants must be incorporated into the acceptance criteria. The acceptance criteria for this
move become



RF J F
35
AXjY min 1, RW R
RW J
where JF and JR are the Jacobian determinants for the forward and
reverse moves, respectively. For a detailed explanation on the calculation of the bound of
and the Jacobians readers are referred to [33].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Configurational Bias Techniques

245

FIG. 11 The entire spherical sampling space is not accessible for selection of the
trial orientation of site 1, because the
angle must lie within the bounds [
min,
max]
(see Fig. 10). Note that there are no bounds on the spherical angle  for the bond
vector.

This algorithm can also be extended to simulation of branched chains and

cross-linked structures [34]. In some cases, it is useful to target or restrict
ECCB moves to a particular region of space. If such a region is spherical, for
example, sites within that sphere can be regrown into new positions by using
ECCB (if they are inner sites) or configurational bias (if they are terminal or
end sites). This approach has been referred to as the HOLE method in the
literature (Fig. 12). Figure 13 shows the bond and end-to-end relaxation
functions for a system of 16-mer hard spheres.
Another strategy for topological bias moves involves the use of an
analytical probability distribution function (for ideal, non-self-avoiding
random walks) to bias the choice of trial orientations [4,35]. Again, this
algorithm is applicable to relatively flexible molecules.
For molecules with strong intramolecular interactions, inner sites can be
regrown using the concerted rotation (CONROT) method [28,29]. This
method is particularly useful for molecular models that exhibit rigid
constraints (e.g., rigid bonds or rigid angles). It involves cutting a trimer
from the inner sites of the chain. Before the trimer is regrown, either one
or both of the sites adjacent to the ends of the trimer are displaced.
The problem of regrowing the trimer is handled by solving a system of
simultaneous equations [28,29,36].
Configurational Bias methods have now also been used to simulate
molecules with strong intramolecular potentials [37,38]; in that case, it is
necessary to incorporate such potentials in the biasing factors used to
generate trial orientations. If the trial orientations are proposed at random,
the efficiency of the trial move is low, due to high intramolecular energies

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

246

Jain and de Pablo

FIG. 12 Illustration of the HOLE method. (a) Local configuration before the
move; (b) configuration after the move.

for a significant number of the proposed positions. The use of intramolecular potentials along with the probability distribution for distances
between sites has been successful in biasing the random walk and achieving
closure [37]. A preliminary run is necessary in order to generate a
probability distribution for site separations as a function of site spacing
along the chain. A brief outline of the implementation as proposed by
Escobedo and Chen is provided below:
1.

2.
3.

A chain is selected at random, and part of the chain consisting of n

sites is cut, starting from site s of the chain. The sites to be regrown
are numbered from 1 to n.
Assume that the current site to be regrown is i.
Select k trial orientations for i, using the Boltzmann distribution
arising from the intramolecular potential
bondedl

pli Ce
Ui

l 1, 2, . . . , k

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

36

Configurational Bias Techniques

247

FIG. 13 Correlation functions for crankshaft, Configurational bias, ECCB-3,

and HOLE moves for a system of 16-mer hard spheres at  0.3. (a) Bond
autocorrelation function (b(t)); (b) end-to-end autocorrelation function (e(t)).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

248

4.

Jain and de Pablo

where C is a normalization constant. Contributions to Uibondedl can

result from bond-stretching, bond-bending or bond-torsion.
The contribution to a sites energy arising from external interactions
and intramolecular nonbonded interactions is calculated for each
trial position; one position, j, is selected from the following probability distribution:
ext

Poi rj, n1 e
Ui j

wi
k
X
ext
Poi rl, n1 e
Ui l
wi

pi j

37
38

l1

5.

where rl, n 1 is the distance from site i in trial position l to site n 1,

and Poi rl, n1 is the separation probability distribution for two sites
that are n
i 1 bonds apart along the chain.
This process is repeated for all but the last two sites, n
1 and n,
which are grown using a look-ahead strategy (Fig. 14). Two such
strategies, RCB1 and RCB2, have been considered by Escobedo
and Chen [37].
a. RCB1
i.

Since the bending angle,
n (formed by n
1, n, and n 1),

and the Jacobian factor, Jn
1, of the crankshaft rotation of
site n are known as soon as the trial position for n
1 is
known, they can be used to bias the selection of the position
for site n
1 out of kn
1 trial positions according to

wn
1

j
e
Un
1 j e
U
n Jn
1

wn
1
kn
1
X
ext
bend l
l

e
Un
1 l e
U
n Jn
1

ext

j
pn
1

bend

39
40

l1

ii.

iii.

Having grown site n
1, site n can be selected by the usual

configurational bias technique out of a set of kn trial positions according to Eq. (14).
After all the sites have been grown, the Rosenbluth weight
for the entire chain is constructed:
RFW

n
Y

wi

i1

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

41

Configurational Bias Techniques

249

FIG. 14 Illustration and notation of the reconstruction process of the last two sites
(n 1 and n) of an ECCB move for a chain that exhibits stiff intramolecular energy
functions [37].

iv.

b.

The modified acceptance criteria for this move are

2
Q
R 3
n
2 o
F
RW
i1 Pi rjn1
6
7
AXjY min 41,
Q
F 5
n
2 o
R
RW
i1 Pi rjn1

42

RCB2
i.

In this variation, instead of selecting one trial position for

n
1, a total of m final trial positions are selected from
mkn
1 trial choices, using the following probability distribution:
j
pn
1,
t

wn
1, t

e
U

bend

j
j

n,
tJ
n
1, t

wn
1, t
kn
1
X

e
U

bend

t 1, 2, . . . , m

l
l

n,
tJ
n
1, t

43
44

l1

ii.

where subscript index t runs over the m sets of trial

positions.
For each final trial position of n
1, jn
1, t , one trial
position of n is chosen from kn trial choices, just as in (a).
This gives a set of Rosenbluth weights, wn,t, and a set of final
trial orientations, jn,t.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

250

Jain and de Pablo

iii.

Out of the m position pairs of sites n
1 and n, one pair, j,

is finally selected from the probability distribution:
ext

ext

e
Un
1 j e
Un j wn
1, j wn, j

wn, n
1
m
X
ext
ext

e
Un
1 t e
Un t wn
1, t wn, t

j
pn
1,
n

wn, n
1

45
46

t1

iv.

After all the sites have been grown, the Rosenbluth weight
for the entire chain is constructed:
RFW wn, n
1

n
2
Y

wi

47

i1

v. The acceptance criteria are the same as given by Eq. (42).

3. Simulation of Linear and Cyclic Alkanes Using

Configurational Bias Approach
Escobedo and Chen [37] have used this technique to study linear and cyclic
alkanes (using a united-atom force field [39]). In order to test the efficiency
of these algorithms, a model system consisting of one isolated linear alkane
molecule at T 400K was simulated. A half-chain end-to-end vector
autocorrelation function was used as a measure of the relaxation of the
chain. This function was measured in the simulation as a function of number
of sites regrowing in every RCB2 move (Fig. 15). The dependence of the
autocorrelation function on the choice of Poi rl, n1 was considered (Fig. 16).
A comparison between RCB1, RCB2, and simple crankshaft moves was also
considered (Fig. 17). One can see that the improvement over the simple
crankshaft move is significant for both the RCB1 and RCB2 moves. The
optimum number of sites to be regrown depends on the system conditions,
but in general the improvements tend to be large for the first sites that are
excised, and they taper off gradually.

D. Parallel Tempering and Configurational Bias

A problem that pervades simulations of complex fluids, particularly at
elevated densities and low temperatures (e.g., near the glass transition
temperature), is that the system gets trapped in local energy minima from
which it is unable to escape. In such cases Configurational Bias moves
are often insufficient, and additional tricks are necessary to improve
sampling. One class of methods that is particularly helpful and easy to
implement is provided by Parallel Tempering techniques.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Configurational Bias Techniques

251

FIG. 15 Half-chain end-to-end correlation function (h(t)) for RCB2, regrowing

exclusively 2, 4, 6, or 8 sites in a system consisting of one isolated C30 chain at 400K.

FIG. 16 Half-chain end-to-end correlation function (h(t)) for RCB1 for an isolated
C30 molecule at 250K for 4 and 8 site moves. Three different forms of Poi rl, n1 are
used: the Continuous Unperturbed Chain (CUC) Poi rl, n1 (dash-dot line), the
uniform Poi rl, n1 implied by conventional ECCB moves (dashed line), and Poi rl, n1
for a fully flexible Guassian chain (solid line).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

252

Jain and de Pablo

FIG. 17 Half-chain end-to-end correlation function (h(t)) for crankshaft, RCB1,

and RCB2 moves for an isolated linear C70 molecule at 400K. The number of sites
regrown ranged from 2 to 10.

Just as configurational bias moves, parallel tempering methods come in a

number of flavors. We begin by discussing briefly Parallel Tempering in one
dimension, and we then proceed to discuss its implementation in several
dimensions or in combination with expanded ensemble ideas.
In parallel tempering, a series of noninteracting replicas of the system
are simulated simultaneously. The temperature of each replica of the system
is different. The partition function of the overall system is given by
Q

Nr
Y

Qi N, V, Ti

48

i1

where Nr denotes the number of replicas and Ti denotes the temperature of

replica i. In addition to the equilibration moves carried out in the canonical
ensemble (configurational bias moves, local moves, etc.), the identity of two
replicas is also allowed to mutate along the T axis (Fig. 18). For any two
replicas, i and j, the probability of accepting such a mutation (or swap) is
given by
AXjY min1, expi, j Ui, j

49

where i, j is the difference in the inverse temperatures of the two replicas
and Ui, j is the difference in the potential energies. It can be seen from
Eq. (49) that the trial move will have a reasonable chance of being accepted
only if the energy distributions for the two replicas overlap to some extent.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Configurational Bias Techniques

FIG. 18

253

Illustration of Parallel Tempering Monte Carlo.

A preliminary run may therefore be necessary to determine appropriate

temperature spacings between replicas. Parallel tempering in temperature
has been used to study simple glasses and isolated biological molecules
[4043]. It has also been used in conjunction with topological configurational bias to study linear and cyclic alkanes [37].

1. Multidimensional and Hyperparallel Parallel Tempering

In many cases Parallel Tempering in temperature does not provide
significant improvements over a conventional, one-replica simulation. It
turns out that the performance of parallel tempering can be improved
considerably by considering several dimensions, e.g., replicas having
different T and or different T and P. We refer to such methods as
Multidimensional Parallel Tempering. In a grand canonical ensemble, for
example, different replicas i can have different values of chemical potential
i and temperature Ti; configuration swaps between any two replicas are
accepted with probability:
Pacc min1, expi, j Ui, j
Ni, j  ij

50

where i, j i
j, Ui, j Ui
Uj, Ni, j Ni
Nj and ( )i, j ( )i

( )j. Multidimensional Parallel Tempering has been employed successfully
to study the phase behavior of electrolytes at low temperatures [44,45]. It has
also been used to investigate the behavior of glass-forming disaccharides
near the glass transition [46].
In the particular case of polymeric molecules, it is also possible to
combine multidimensional parallel tempering and expanded ensemble
ideas into a powerful technique, which we call hyperparallel tempering
(HPTMC) [47]. In that method, a mutation along the number of sites of a
tagged chain is also considered (see Section IV.B.1); Fig. 19 provides a
schematic representation of the algorithm. Two of the axes refer to replicas
having different chemical potentials or different temperatures. A third axis
or dimension is used to denote the fact that, in each replica of the system,

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

254

Jain and de Pablo

FIG. 19 Schematic representation of the implementation of hyperparallel Monte

Carlo. Each box in the figure represents a distinct replica of the system; these replicas
are simulated simultaneously in a single run. In addition to traditional Monte Carlo
trial moves within each replica, distinct replicas can (1) change their state variables
in the expanded dimension and (2) exchange or swap configuration with each other,
thereby visiting different values of T and .

one can have Ni full molecules and one tagged molecule of length y (recall
Section IV.B.1). When two replicas exchange configuration, they also swap
a tagged chain. Within each of the replicas, however, the tagged chain is
allowed to grow or shrink according to a conventional expanded ensemble
formalism.
This technique has been applied to the study of phase behavior of long
polymeric molecules. Figure 20 compares the performance of various
simulation techniques to that of Hyperparallel Tempering Monte Carlo. It
can be seen that the decay of the end-to-end autocorrelation function for
HPTMC is considerably faster than that for conventional canonical or
grand canonical simulations. However, it is important to point out that
much of the performance of that method is due to the efficiency of the
underlying moves, namely the expanded ensemble configurational bias
gradual growth or deletion of sections of the molecules.

V. FUTURE DIRECTIONS
Configurational bias trial moves offer significant improvements in efficiency
over conventional, blind trial moves. As discussed throughout this chapter,
local random moves usually provide small displacements of the system,
whereas biased moves can efficiently displace several sites per move while

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Configurational Bias Techniques

255

FIG. 20 End-to-end vector autocorrelation function (e(t)) for polymer chains

obtained by different simulation methods: (1) short dashed line: canonical ensemble;
(2) dashed-dotted-dotted line: grand canonical ensemble (3) dashed line: multidimensional parallel tempering; (4) dashed-dotted line expanded grand canonical
ensemble; (5) solid line: hyperparallel tempering.

preserving a high acceptance rate. As illustrated by the variety of systems

and moves discussed here, configurational bias techniques offer considerable flexibility; new implementations will surely be proposed as researchers
continue to explore increasingly complex systems.
In the particular context of polymeric molecules, recent developments
with end-bridging and double-bridging moves [48] indicate that a powerful
algorithm could be arrived at by using topological configurational bias
to cut simultaneously several chains, and to reconnect them in different
ways, thereby achieving considerable molecular relaxation in just a few
moves. The details of such moves remain to be worked out, particularly
for the more complex case of highly branched molecules. But it is a
direction in which we are likely to see interesting new developments in the
near future.

REFERENCES
1. Metropolis, N.; Rosenbluth, A.W.; Rosenbluth, M.N.; Teller, A.H.; Teller, E.
Equation of state calculations by fast computing machines. J. Chem. Phys. 1953,
21, 10871092.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

256

Jain and de Pablo

2. Kalos, M.H.; Whitlock, P.A. Monte Carlo Methods; Wiley: New York, 1986.
3. Cracknell, R.F.; Nicholson, D.; Parsonage, N.G.; Evans, H. Rotational insertion bias: a novel method for simulating dense phases of structured particles,
with particular application to water. Mol. Phys. 1990, 71 (5), 931943.
4. Frenkel, D.; Smit, B. Understanding Molecular Simulation; Academic:
New York, 1996.
5. Chavez-Paez, M.; Van Workum, K.; de Pablo, L.; de Pablo, J.J. Monte Carlo
simulations of wyoming sodium montmorillonite hydrates. J. Chem. Phys.
2001, 114 (3), 14051413.
6. Skipper, N.T.; Chang, F.R.C.; Sposito, G. Monte-Carlo simulation of interlayer molecular-structure in swelling clay-minerals. 1. methodology. Clay Clay
Miner. 1995, 43 (3), 285293.
7. Boek, E.S.; Convey, P.V.; Skipper, N.T. Molecular modeling of clay hydration:
a study of hysteresis loops in the swelling curves of sodium montmorillonites.
Langmuir 1995, 11 (7), 46294631.
8. Jorgensen, W.L.; Chandrasekhar, J.; Madura, J.D. Comparison of simple
potential functions for simulating liquid water. J. Chem. Phys. 1983, 79 (2),
926935.
9. Lie, G.C.; Clementi, E.; Yoshimine, M. Study of the structure of molecular
complexes. XIII. Monte Carlo simulation of liquid water with a conguration
interaction pair potential. J. Chem. Phys. 1976, 64 (6), 23142323.
10. Lax, M.; Brender, K. Monte Carlo studies of lattice polymer dynamics.
J. Chem. Phys. 1977, 67 (4), 17851787.
11. Mandel, F. Macromolecular dimensions obtained by an ecient Monte Carlo
method: the mean square end-to-end separation. J. Chem. Phys. 1979, 70 (8),
39843988.
12. de Pablo, J.J.; Yan, Q.L.; Escobedo, F.A. Simulation of phase transitions in
uids. Annu. Rev. Phys. Chem. 1999, 50, 377411.
13. Rosenbluth, M.N.; Rosenbluth, A.W. Monte Carlo calculation of the average
extension of molecular chains. J. Chem. Phys. 1955, 23 (2), 356359.
14. de Pablo, J.J.; Laso, M.; Suter, U.W. Simulation of polyethylene above and
below the melting-point. J. Chem. Phys. 1992, 96 (3), 23952403.
15. Siepmann, J.I.; Frenkel, D. Congurational bias Monte-CarloA new sampling scheme for exible chains. Mol. Phys. 1992, 75 (1), 5970.
16. Smit, B.; Karaborni, S.; Siepmann, J.I. Computer simulations of vapour-liquid
phase equilibria of n-alkanes. J. Chem. Phys. 1995, 102 (5), 21262140.
17. Lyubartsev, A.P.; Martinovski, A.A.; Shevkunov, S.V.; VorontsovVelyaminov, P.N. New approach to Monte Carlo calculation of the free energy:
Method of expanded ensembles. J. Chem. Phys. 1991, 96 (3), 17761783.
18. Depablo, J.J.; Escobedo, F.A. Expanded grand canonical and Gibbs
ensemble Monte Carlo simulation of polymers. J. Chem. Phys. 1996, 105 (10),
43914394.
19. Kumar, S.K.; Szleifer, I.; Panagiotopoulos, A.Z. Determination of the
chemical-potentials of polymeric systems from Monte-Carlo simulations.
Phys. Rev. Lett. 1991, 66 (23), 23952398.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Configurational Bias Techniques

257

20. de Pablo, J.J.; Escobedo, F.A. Chemical-potential and equations of state of

hard-core chain molecules. J. Chem. Phys. 1995, 103 (9), 19461956.
21. Attard, P. Simulation of the chemical potential and the cavity free energy of
dense hard-sphere uids. J. Chem. Phys. 1993, 98 (3), 22252231.
22. Wilding, N.B.; Muller, M. Accurate measurements of the chemical-potential of
polymeric systems by Monte-Carlo simulation. J. Chem. Phys. 1994, 101 (5),
43244330.
23. Rutledge, G.C.; Khare, A.A. Chemical potential of aromatic compounds in
pure n-alkanes using expanded ensemble Monte Carlo simulations. J. Phys.
Chem. B 2000, 104 (15), 36393644.
24. Yethiraj, A.; Hall, C.K. Monte-Carlo simulation of the equilibrium partitioning of chain uids between a bulk and a pore. Mol. Phys. 1991, 73 (3),
503515.
25. Yan, Q.L.; de Pablo, J.J. Critical behavior of lattice polymers studied by Monte
Carlo simulations. J. Chem. Phys. 2000, 113 (14), 59545957.
26. Dobashi, T.; Nakata, M.; Kaneko, M. Coexistence curve of polystyrene
in methylcyclohexane. 1. Range of simple scaling and critical exponents.
J. Chem. Phys. 1980, 72 (12), 66856697.
27. Frauenkron, H.; Grassberger, P. Critical unmixing of polymer solutions.
J. Chem. Phys. 1997, 107 (22), 95999608.
28. Dodd, L.R.; Boone, T.D.; Theodorou, D.N. A concerted rotation algorithm
for atomistic Monte-Carlo simulation of polymer melts and glasses. Mol. Phys.
1993, 78 (4), 961996.
29. Pant, P.V.K.; Theodorou, D.N. Variable connectivity method for the atomistic
Monte-Carlo simulation of polydisperse polymer melts. Macromolecules 1995,
28 (21), 72247234.
30. Dijsktra, M.; Frenkel, D.; Hansen, J.P. Phase separation in binary hard-core
mixtures. J. Chem. Phys. 1994, 101 (4), 31793189.
31. Jain, T.S.; de Pablo, J.J. Monte Carlo simulation of free-standing polymer lms
near the glass transition temperature. Macromolecules 2002, 35 (6), 21672176.
32. Jain, T.S.; de Pablo, J.J. A biased Monte Carlo technique for calculation of the
density of states of polymer lms. J. Chem. Phys. 2002, 116 (16), 72387243.
33. Escobedo, F.A.; de Pablo, J.J. Extended continuum congurational bias Monte
Carlo methods for simulation of exible molecules. J. Chem. Phys. 1994, 102
(6), 26362652.
34. Escobedo, F.A.; de Pablo, J.J. Molecular simulation of polymeric networks and
gels: phase behavior and swelling. Phys. Rep. 1999, 318 (3), 86112.
35. Vendruscolo, M. Modied congurational bias Monte Carlo method for
simulation of polymer systems. J. Chem. Phys. 1997, 106 (7), 29702976.
36. Wu, M.G.; Deem, M.W. Analytical rebridging Monte Carlo: application to
cis/trans isomerization in proline-containing, cyclic peptides. J. Chem. Phys.
1999, 111 (14), 66256632.
37. Escobedo, F.A.; Chen, Z. A congurational-bias approach for the simulation
of inner sections of linear and cyclic molecules. J. Chem. Phys. 2000, 113 (24),
1138211392.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

258

Jain and de Pablo

38. Siepmann, J.I.; Wick, C.D. Self-adapting xed-end-point congurational-bias

Monte Carlo method for the regrowth of interior segments of chain molecules
with strong intramolecular interactions. Macromolecules 2000, 33 (19),
72077218.
39. Nath, S.K.; Escobedo, F.A.; de Pablo, J.J. On the simulation of vapor-liquid
equilibria for alkanes. J. Chem. Phys. 1998, 108 (23), 99059911.
40. Hansmann, U.H.E. Parallel tempering algorithm for conformational studies of
biological molecules. Chem. Phys. Lett. 1997, 281 (13), 140150.
41. Deem, M.W.; Wu, M.G. Ecient Monte Carlo methods for cyclic peptides.
Mol. Phys. 1999, 97 (4), 559580.
42. Okamoto, Y. Tackling the multiple-minima problem in protein folding by
Monte Carlo simulated annealing and generalized-ensemble algorithms. Int.
J. Mod. Phys. C 1999, 10 (8), 15711582.
43. Marinari, G.; Parisi, E.; Ricci-Tersenghi, F.; Zuliani, F. The use of optimized
Monte Carlo methods for studying spin glasses. J. Phys. A-Math. Gen. 2001,
34 (3), 383390.
44. Yan, Q.L.; de Pablo, J.J. Phase equilibria of size-asymmetric primitive model
electrolytes. Phys. Rev. Lett. 2001, 86 (10), 20542057.
45. Yan, Q.L.; de Pablo, J.J. Phase equilibria and clustering in size-asymmetric
primitive model electrolytes. J. Chem. Phys. 2001, 114 (4), 17271731.
46. Ekdawi-Sever, N.C.; Conrad, P.B.; de Pablo, J.J. Molecular simulation of
sucrose solutions near the glass transition temperature. J. Phys. Chem. A
2001, 105 (4), 734742.
47. Yan, Q.L.; de Pablo, J.J. Hyperparallel tempering Monte Carlo simulation of
polymeric systems. J. Chem. Phys. 2000, 113 (3), 12761282.
48. Theodorou, D.N. Variable-connectivity Monte Carlo algorithms for the atomistic simulation of long-chain polymer systems. Bridging Time Scales:
Molecular Simulations for the Next Decades, Nielaba, P., Mareschal, M.,
Cicotti, G., Eds.; Springer Verlag: Berlin, 2002, 67127.
49. Panagiotopoulos, A.Z.; Wang, V. Phase equilibria of lattice polymers from
histogram reweighting Monte Carlo simulations. Macromolecules 1998,
31 (3), 912918.
50. Wilding, N.B.; Muller, M.; Binder, K. Chain length dependence of the polymersolvent critical point parameters. J. Chem. Phys. 1996, 105 (2), 802809.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

8
Molecular Simulations
of Charged Polymers
ANDREY V. DOBRYNIN
Connecticut, U.S.A.

I.

University of Connecticut, Storrs,

INTRODUCTION

Considerable theoretical and experimental work during the past half century
has been devoted to charged polymers [19]macromolecules with ionizable groups. Under appropriate conditions, such as in aqueous solutions,
these groups dissociate, leaving ions on chains and counterions in solutions.
If the charges on the polymers are all positive or all negative, these polymers
are called polyelectrolytes. Common polyelectrolytes are polyacrylic and
methacrylic acids and their salts, cellulose derivatives, sulfonated polystyrene, DNA and other polyacids and polybases. If after dissociation of the
charged groups the polymers carry both positive and negative charges, they
are called polyampholytes. Examples of polyampholytes include proteins,
for example gelatin, and synthetic copolymers made of monomers with
acidic and basic groups. If these groups are weak acids or bases, the net
charge of polyampholytes can be changed by varying the pH of aqueous
solutions and at high charge asymmetry these polymers demonstrate
polyelectrolyte-like behavior.
Despite these extensive efforts we are still far away from a complete
understanding of the behavior of polymeric systems with electrostatic
interactions. The main factor hindering our progress is the long-range
nature of the Coulombic forces between charged species. At this stage
computer simulations have proven to be a valuable tool for the elucidation
of structural and physical properties of charged systems as well as for
verification of old and creation of new theoretical models. The explosive

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

260

Dobrynin

FIG. 1 Illustration of the mapping procedure of sodium polystyrene sulfonate

(NaPSS) in water where the repeat units of the NaPSS chain are replaced by
spherical monomers connected by springs and the solvent molecules are represented
by a continuum with dielectric constant ".

growth of computer power over the last few years has led to development of
large scale simulation techniques whose goal is to reproduce and simulate
processes on the molecular level. It is possible now to simulate thousands of
atoms over nanosecond time scales. But such detailed molecular simulations
are still impractical for the simulation of real polymeric systems, for which
typical time scales to reach thermal equilibrium vary from microseconds to
seconds.
The coarse-grained models of polymers (see [10] for a review) allow the
extension of molecular simulations beyond nanosecond time scales by
leaving aside the atomistic details of the solvent and polymers and
concentrating on their macroscopic properties. In this approach the monomers are not chemical monomers, but rather groups of chemical units, and
solvent molecules are represented by a continuum with macroscopic physical properties. Figure 1 shows an example of such mapping by presenting
the snapshots of sodium polystyrene sulfonate in water and its representation as a coarse-grained chain with counterions in a continuum with
macroscopic physical properties. Of course, the challenging part is to relate
the parameters of the coarse-grained models with the real atomistic ones by
calculating the parameters of intermolecular and intramolecular coarsegrained potentials [10].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

261

II. COMPUTER SIMULATIONS OF

SINGLE CHAIN PROPERTIES
A. Models and Methods
We begin our discussion by considering the conformation of a single chain
in solution under the influence of the long-range electrostatic and shortrange Ush (|ri
rj|) monomermonomer interactions. The potential energy
U({ri}) of the charged polymer chain of the degree of polymerization N in
a solvent with dielectric permittivity " with monomers located at positions
r1, r2, . . . , rN and carrying the charges eq1, eq2, . . . , eqN is*
U fri g

U bond fri g

bond

XX
i

i<j

!


 

qi qj




 exp
 ri
rj U sh ri
rj
kB TlB 
ri
rj 

where U bond ({ri}) is the bond potential describing the effect of the connectivity of monomers into the polymer chain, kB is the Boltzmann constant, T
is the absolute temperature, and lB is the Bjerrum length (lB e2/"kBT,y the
length scale at which the electrostatic interaction between two elementary
charges e in the medium with dielectric permittivity " is of the order of the
thermal energy kBT). The electrostatic interactions between free ions and
charged monomers are not explicitly included in the chain potential energy;
instead, their effect is treated through the dependence of the inverse
DebyeHuckel screening length  on the electrolyte concentration
2 4lB

q2s cs

where cs is the concentration of small ions of type s and qs is their valence.

As discussed in previous chapters, the choice of the bond U bond ({ri}) and
short-range U sh (|ri
rj|) potential varies from simulation to simulation. Offlattice models, for example, have used the harmonic-spring potential, the
FENE (finitely extendable, nonlinear elastic) potential, the rigid bond with
fixed valence angles, and the freely-jointed chain model to represent the
bonding interaction between adjacent monomers. For the short-range

*CGS units are used in this chapter, unless specically indicated.

y
In SI units it is lB e2/(4""0kBT).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

262

Dobrynin

monomermonomer interactions the Lennard-Jones and hard-core type

potentials are commonly used in simulations of charged polymers.
To sample the chain configurational space one can perform canonical
Monte Carlo (MC) [1116] or molecular dynamics (MD) simulations
[1517]; see also the discussions in previous chapters in this book. The
Monte Carlo method involves generation of successive trial chain
configurations which are then accepted or rejected with the correct statistical
weight. The trial configurations are generated by randomly modifying the
monomer positions. For example, a new chain conformation may be
obtained by reptation [18,19] and pivot [2022] moves or by local moves
randomly displacing a single monomer. In reptation or the slithering
snake algorithm one chooses an end of the chain at random and transfers
a monomer from this end to other end. Applied to an N-bead chain this
method leads to a statistically independent configuration after N2 attempts
to exchange the ends of a chain. In the pivot algorithm a new chain
configuration is obtained by rotating a part of the chain around a randomly
selected bond by a randomly selected angle. The pivot moves are very
radical: only after a few accepted moves may a chain reach an essentially
new configuration. (For lattice models of a chain the moves have to be
modified to preserve the lattice structure.)
To decide whether or not a trial conformation should be accepted or
rejected, the change in energy, U, associated with the move is calculated,
and the move is accepted or rejected according to the Metropolis algorithm
with probability





U
accold ! new min 1, exp

kB T

If the new conformation is rejected then the current configuration is

recounted in the averaging process.
The choice of one or another type of move or combination of moves is
dictated by the most efficient sampling of the configurational space of the
chain during the course of simulations. The art is to find their optimal
combination. Below we give a theoretical estimate of the computational
efficiency for different types of moves defined as the CPU time required
to generate a new statistically independent chain configuration. The less
computationally efficient are the local moves, when a new chain configuration is generated by random displacement of a single monomer. This
MC scheme describes the VerdierStockmayer dynamics [25] of a polymer
chain, for which a chain renews its configuration after N2 successful moves
per monomer. Since the monomers are chosen at random, each monomer
moves on average after N elementary steps. This leads to an increase of

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

263

relaxation time with the number of monomers on a chain N as N 3 .

After each move one has to recalculate the electrostatic interactions with
N
1 monomers, assuming that each monomer is charged and Coulomb
interactions are unscreened. The total CPU time required for chain
relaxation will grow with the number of monomers N as flip N N 4 .
For pivot moves with acceptance probability pacc, the new configuration is
reached after p
1
acc attempts. After each pivot move, a part of a chain
with, say, n monomers is displaced with respect to the remaining N
n
monomers. And (N
n) n new pairs are created for which the electrostatic
interactions have to be reevaluated. Since the number of new pairs cannot
exceed N2/4, the CPU time for pivot moves cannot increase faster than
pivot  N2/pacc. It is important to point out that the efficiency of the pivot
moves is strongly dependent on the acceptance probability pacc. In a system
with a high acceptance rate, pacc 1, the CPU time may grow as N2.
However, for low acceptance probability it can grow faster than N3. For
reptation moves the CPU time grows as rep N3, because a chain renews its
configuration after N2 steps and at each step one has to recalculate the
electrostatic interactions with N
1 monomers. Examples of computational
efficiency of different types of moves for a polyelectrolyte chain are shown
in Fig. 2.
Molecular dynamics simulations [16,17] are based on numerical integration of the equations of motion of the system. During the course of
simulation the system moves in phase space generating new chain

FIG. 2 Variation of the total energy of the system Etot versus the number of MC
steps for different types of moves. (From Ref. [69].)

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

264

Dobrynin

conformations along its physical trajectory determined by the equations of

motion. The equation of motion of the ith monomer with mass m is


4
mri t
ri U frj tg
m_ri t Fi t
where the first term describes the deterministic forces between monomers
and the last two terms implicitly take into account the effect of the solvent
by coupling the system to a Langevin thermostat [23,24] which maintains a
constant average temperature of the system. In Eq. (4) the parameter  is
the
coefficient and Fi(t) is a random force with zero average value
friction

Fi t and delta-functional time correlations
hFi tFj t0 i 6kB Tmij t
t0

The MD simulation coupled with the Langevin thermostat simulates

Rouse dynamics of a polymer chain. The Rouse relaxation time scales with
the number of monomers on a chain as N2 and it is necessary to perform at
least cN2 (where constant c depends on the value of the integration time step
t) integrations of the equation of motion for a chain to completely renew
its configuration. During each time step, t, N(N
1)/2 calculations of
forces between monomers are performed. The CPU time required to do cN2
integrations of the equations of motion will grow with the number of
monomers on a chain as MD N 4 . Thus, the computational efficiency of
MD simulation has the same N dependence as a MC simulation with only
local moves.
In a salt solution the electrostatic interactions are exponentially screened
over distances larger than the Debye screening length. In this case the
simulations may be significantly speed up by considering electrostatic
interactions between only those monomers that are within a cutoff distance
rcut. The choice of the cutoff distance is dictated by the desired accuracy.
After generating a set of chain conformations during the course of MC or
MD simulations, the analysis of the effect of electrostatic and short-range
interactions on global properties of a polymer chain can be achieved by
2
2
looking at the mean square end-to-end distance
P hRe i hr2N
r1 i and the
2

2
mean square radius of gyration hRg i N
i< j hri
rj i.
The internal structure of a polyelectrolyte chain can be obtained from the
chain structure factor, defined as

2 +
N
X


exp iq  rj 
Sq N 

 j1
*

1 

where q is the wave vector.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

265

B. Polyelectrolyte Chain in h and Good Solvents

1. Chain Conformation in Dilute Salt-Free Solutions
In a
-solvent the short-range interactions cancel to zero and only bonded
and electrostatic potentials are present in the chain potential energy U({ri}),
Eq. (1). In spite of its simplicity, this model already contains two important
ingredients responsible for the unique behavior of polyelectrolyteschain
connectivity and the long-range electrostatic interactions. In salt-free
solutions (!0) the size L0 of a polyelectrolyte chain with fN charged
monomers can be estimated from a simple Flory argument by balancing the
chain elastic energy kB TL20 =b2 N (here b is the bond length) and
electrostatic energy kBTlB f 2N2 ln (L0) /L0; the logarithm in the electrostatic
energy is accounting for chain elongation. This leads to the well known
result that in salt-free solutions the chain size
 1=3
L0  b uf 2
N ln1=3 N

is proportional to the number of monomers N with weak logarithmic correction. In Eq. (7) u is the ratio of the Bjerrum length lB to the bond length b.
The parameter uf 2 is sometimes called the coupling parameter. In fact,
polyelectrolytes are classified according to the strength of this parameter.
Polyelectrolyte chains with small values of the coupling parameter uf 2 1
are referred to as weakly charged polyelectrolytes, while ones with uf 2 01
are considered to be strongly charged. The larger the value of the coupling
parameter uf 2, the more elongated the polyelectrolyte chain is.
To better understand the different length scales involved in the problem,
it is useful to introduce the concept of the electrostatic blob [2631]. The
conformation of a chain inside the electrostatic blob is almost unperturbed
by electrostatic interactions, with the number of monomers in it being
ge  (De/b)2 in a
-solvent. The size of the electrostatic blob De containing ge
monomers can be found by comparison of the electrostatic energy of a blob
e2 g2e f 2 ="De with the thermal energy kBT. This leads to the electrostatic
blob size
 
1=3
De  b uf 2
and the number of monomers in it
 
2=3
ge  uf 2

For weakly charged polyelectrolytes (uf 2 1) there is more than one

monomer in an electrostatic blob (ge>1). At length scales larger than the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

266

Dobrynin

electrostatic blob, electrostatic interactions are much stronger than the

thermal energy kBT and the chain configuration is that of a fully extended
array of N/ge electrostatic blobs of length L0
L0 

 1=3
N
De  bN uf 2
ge

10

However, for a polyelectrolyte chain to be elongated, the number of monomers in a chain N should be larger than the number of monomers in an
electrostatic blob ge N > uf 2
2=3 . Thus, short chains with number of
monomers N smaller than NG  (uf 2)
2/3 will still be Gaussian.
Of course, this simple scaling picture does not account for the
nonuniform stretching of a polyelectrolyte chain. In reality the chain is
more strongly stretched in the middle than at the ends. Logarithmic
corrections to the chain size may be obtained by allowing the blob size to
vary along the chain.
The generalization of this scaling approach to the case of the good
solvent is straightforward by replacing the relation between the blob size De
and the number of monomers in it ge from random to self-avoiding random
walk statistics [2631].
The conformational properties of an isolated polyelectrolyte chain in a
salt-free solution have been investigated by lattice Monte Carlo simulations
[3235], and by off-lattice Monte Carlo simulation of a chain made of hard
spheres connected by rigid bonds with fixed valance angles [36], of a freelyjointed chain [37], and of a bead-spring chain [3842]. These simulations
support the trivial scaling prediction that the chain crosses over from a coil
to a rod-like conformation with increasing the strength of the electrostatic
interactions (increasing value of the coupling parameter uf 2). This was done
either by adding ionized groups to a chain or by increasing the Bjerrum
length lB.
Quantitative analysis of the scaling model of a polyelectrolyte chain as a
chain of electrostatic blobs was done by Higgs and Orland [37] and by
Barrat and Boyer [38]. Figure 3 shows the mean square end-to-end distance
as a function of the degree of polymerization N for different values of the
coupling parameter uf 2 (in references [37,38] the parameter f was kept
constant and equal to unity). For each curve there is a crossover between
two regimes. For small N the electrostatic interactions between charges have
little or no effect at all on the conformation of the polyelectrolyte chain
and the chain remains Gaussian hR2e i N. For large N the electrostatic
interactions dominate and the chain adopts an elongated conformation with
size hR2e i N 2 .

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

267

FIG. 3
Mean square end-to-end distance hR2e i as a function of number of
monomers N for a polyelectrolyte chain in a
-solvent. Curves are for three different
values of the parameter u 0.02, 0.05, and 0.2. The dasheddotted lines have
gradients 1 and 2. (From Ref. [37].)

In Fig. 4 the dependence of the mean square end-to-end distance hR2e i is

shown as a function of the coupling parameter uf 2 for several values of N.
As expected from the blob model, the chain size hR2e i demonstrates only
weak dependence on the coupling parameter for small values of uf 2 and
increases with uf 2 for larger values of the coupling parameter. For large
values of the coupling parameter the increase in the mean square end-to-end
distance hR2e i follows the scaling law with the exponent 2/3 [see Eq. (10)].
In their simulation Barrat and Boyer [38] gave a quantitative definition of
the electrostatic blob. Using the model of the chain under tension they
established the quantitative relation between the number of monomers n in a
section and its size R(n). The electrostatic blob size De was directly related to
the prefactor of n2 in the expression for the mean square section size
hRn2 i nb2 n2 b2 =ge . To verify the dependence of the blob size on the
value of the coupling parameter uf 2 Barrat and Boyer also calculated the
ratio De uf 2
1=3 =b. This ratio varies between 1.45 and 0.92 when the value
of the coupling parameter uf 2 changes between 0.01 and 1, providing
reasonable agreement with the blob model [2631].
Systematic MC studies of the chain size dependence on its degree of
polymerization, performed by different groups [3842], have shown that the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

268

Dobrynin

FIG. 4 Scaling of hR2e i with the parameter u for a polyelectrolyte in a
-solvent.

Chain lengths between 10 and 100 are shown. The dasheddotted line shows the
theoretical prediction hR2e i u2=3 . (From Ref. [37].)

chain size grows faster than linearly, L0 Nln N1=3 , with the degree
of polymerization N. Figure 5 shows the p
dependence
of the normalized

root-mean-square end-to-end distance

hR2e i=N on the parameter
uf 2 1=3 lnN=e1=3 for the chains with N ranging from 20 to 2000. The
nature of this deviation from the scaling law, L0 N, is due to nonuniform
(logarithmic) stretching of the polyelectrolyte chain.

2. Effects of Added Salt on Chain Conformation and

Electrostatic Persistence Length
When a finite concentration of salt is present in a solution the electrostatic
interactions between charged monomers are screened by salt ions and fall
off exponentially with distance. However, at distances smaller than the
Debye screening length 
1 the charges still interact through a bare
Coulomb potential. A polyelectrolyte chain will not feel the presence of the
salt until the Debye screening length is larger than the chain size L0. On the
other hand, at very high salt concentrations such that the Debye screening
length 
1 is smaller than the electrostatic blob size De, the electrostatic
interactions can be viewed as short-range ones with the effective monomeric
second virial coefficient el  f 2 lB 
2 . At these salt concentrations a
polyelectrolyte chain has the same structure as a neutral polymer in a
good solvent with chain size depending on the Debye screening length and
degree of polymerization as L 
2=5 N 3=5 . But the question is what

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

269

p
FIG. 5
Root-mean-square end-to-end distance
hR2e i divided by N vs.
1=3
2 1=3
uf ln N=e , for chains with N 20, 80, 320, 1000, and 2000 (filled circles,
open squares, filled triangles, open circles, and filled squares, respectively). (From
Ref. [39].)

conformation is adopted by the polyelectrolyte chain at intermediate salt

concentrations De<
1<L0.
It was first shown independently by Odijk [43] and by Skolnick and
Fixman [44] (OSF) that for intrinsically stiff chains, the intrachain
electrostatic repulsion can induce additional chain stiffening beyond the
Debye screening length 
1. To understand this result let us consider the
variation in the electrostatic energy between two consecutive chain sections
of length 
1 due to bending by small angle
. This energy variation is
E
 kB TlB Z2 
2

11

where Z is a charge carried by a section of size 
1, which for weakly

charged polyelectrolytes is Z  fge 
1 =De . The typical mean square value
of the bending angle
caused by the thermal fluctuations (E
 kBT ) is
h
2 i  De . The bending of each pair of sections along a polyelectrolyte
chain can be considered to be statistically independent, because electrostatic
interactions are exponentially weak at these length scales. A chain section
containing l units of size 
1 will deviate from its initial direction by an angle
h
2 li  lh
2 i  lDe  and forgets its initial orientation after 1/De steps
when the value of h
2 li becomes of the order of unity. Multiplying the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

270

Dobrynin

number of sections 1/De by their length 
1 one finds the following expression for the electrostatic persistence length [27,4547]
Le  
2 =De

12

Indeed for salt concentrations such that 
1>De the chain is stiffened at
the length scales larger than the Debye screening length Le>
1. The size
of a chain with persistence length Le and a contour length of a chain of
electrostatic blobs L0 is
L2  Le L0  
2 N=ge

13

and crossover fromp

a
rod-like
chain to a Gaussian chain with bond length Le

occurs at 
1  De N=ge  bN 1=2 . For even higher salt concentrations the
number of the persistence segments increases, making it more probable for
interactions between segments separated by larger distances along the chain
contour to occur. The second virial coefficient  for these interactions can
be estimated as that between two rods of length Le and thickness

1   L2e 
1 . The excluded volume effect becomes important, when the
interaction parameter z of a polyelectrolyte chain

L0
z
Le

1=2


 
2 De bN 1=2
L3e

14

becomes
larger than unity. In this range of salt concentrations
p
(
1 < De bN 1=2 ) the chain size scales as L / 
3=5 N 3=5 . This regime continues until the Debye radius is larger than the electrostatic blob size De.
However, the predictions of this model are still challenged in the
literature. Computer simulations of weakly charged polyelectrolyte chains
[38,5861] and some experiments [6,4854] as well as analytical calculations
[5557] indicate that the exponent for the dependence of the electrostatic
persistence length Le on the Debye screening length is closer to 1 rather than
to 2 or even shows sublinear [62] dependence.
The most complete analysis to date of the electrostatic persistence length
dependence on the Debye screening length was performed by the Kremer
group [62]. Micka and Kremer have performed hybrid MC and MD
simulations of a polyelectrolyte chain with DebyeHuckel interaction
bonded by harmonic springs. For different bond lengths (b 2, 4, 8, 10, 16)
and chain lengths (N 16, 32, 64, 128, 256, 512) they investigated the
dependence of the chain dimensions and electrostatic persistence length on
the Debye screening length in the interval 0.001  
1  0.48. The data for

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

271

the mean square end-to-end distance can be collapsed into universal curves
reasonably well by either assuming linear or quadratic dependence of the
electrostatic persistence length Le on the Debye radius 
1. Both of these
plots show crossover from a rod-like chain hR2e i N 2 to a chain with
excluded volume interactions hR2e i N 6=5 as the value of the Debye radius
decreases. In order to differentiate between the models, the persistence
length of the chain was calculated from the bond angle correlation function
G(n). This function is defined by the scalar product of two normalized bond
vectors b(k) and b(k n)


Gn bk  bk n

15

where the brackets hi denote the ensemble average over all chain conformations. The averaging procedure was improved by moving reference point k
along the chain. The persistence length was estimated from the exponential
decay of the function G(n). This method lead to sublinear dependence of the
electrostatic persistence length on the Debye radius (see Fig. 6). Similar
sublinear dependence was obtained by analyzing the chain structure
factor S(q).
The effect of internal chain stiffness on the dependence of the
electrostatic persistence length on the Debye radius was studied in [63,64].

FIG. 6 Persistence length Le dependence on the Debye screening length 
1. The
two lines in the upper left corner indicate the slopes predicted by OSF (dashed) and
the variational method (solid). (From Ref. [62].)

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

272

Dobrynin

It was shown that with increasing internal chain stiffness the effective
exponent y for Le 
y crosses over from a value of one toward two as the
internal stiffness of a chain increases. The quadratic dependence of the
electrostatic persistence length on the Debye radius for the discrete
KratkyPorod model of the polyelectrolyte chain was recently obtained in
[65]. It seems that the concept of electrostatic persistence length works better
for intrinsically stiff chains rather than for flexible ones. Further computer
simulations are required to exactly pinpoint the reason for its failure for
weakly charged flexible polyelectrolytes.

C. Polyelectrolyte Chain in a Poor Solvent

In a poor solvent, the chains have a negative second virial coefficient,
corresponding to an effective attraction between monomers. This attraction
causes a neutral polymer chain without charged groups to collapse into
dense spherical globules in order to maximize the number of favorable
monomermonomer contacts. The size R of such a globule scales with the
number of monomers on a chain as N1/3 [31]. By charging the polymer chain
one forces the polymeric globule to change its shape. The shape of a
polyelectrolyte in a poor solvent is determined by the interplay between
long-range electrostatic and short-range attractive interactions [68,71,72]. It
is interesting to point out that the problem of the shape of a charged globule
bears similarities with the classical problem of the instability of a charged
droplet, considered by Lord Rayleigh over a hundred years ago [66].
Rayleigh showed that a charged droplet is unstable and breaks into smaller
droplets if the electric charge exceeds some critical value. The value of the
critical charge is controlled by the electrostatic energy of the charged droplet
of size R and carrying charge Q, Q2/R, and its surface energy R2, where  is
the surface tension. Balancing those two energies one finds that the critical
charge Qcrit scales with the size of a droplet as R3/2. [For a charged globule
with size R bN 1=3 the value of the critical charge Qcrit is proportional
pto
the square root of the number of monomers on a chain N Qcrit  N :]
The equilibrium state of the charged droplet with Q>Qcrit is a set of smaller
droplets with charge on each of them smaller than the critical charge and
placed at infinite distance from each other. This final state is impossible for
a charged polymer because it consists of monomers connected into a chain
by chemical bonds. In this case the system reduces its energy by splitting into
a set of smaller charged globules connected by strings of monomersthe
necklace globule. The number of monomers mb in each bead is determined
by Rayleighs stability conditionthe surface energy of a bead D2b is of the
order of its electrostatic energy kB TlB f 2 m2b =Db . For dense beads with

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

273

  kB T=b2 and Db  bm1=3

this results in the number of monomers in a
b
bead mb  1=uf 2 . Those beads are separated by strings whose length lstr is
obtained by equating the surface energy of a string blstr with the
electrostatic repulsion between two consecutive beads kB TlB f 2 m2b =lstr . The
string length
p lstr is proportional to the critical charge on a bead fmb
(lstr  b= uf 2 ).
The qualitative features of an abrupt conformational transition in a
polyelectrolyte chain in a poor solvent were first studied by Hooper et al.
[67] and by Higgs and Orland [37] by performing lattice Monte Carlo
simulations of a polyelectrolyte chain with attractive segmentsegment
interactions. These authors find that with increasing charge, the chain
undergoes an abrupt transition from collapsed to extended conformations.
As the strength of the segmentsegment attraction decreases, the transition
becomes less pronounced [67].
The details of the transition and conformations of the polyelectrolyte
chain above the transition were studied by Dobrynin, Rubinstein, and
Obukhov [68]. Performing Monte Carlo simulations of a freely-jointed
uniformly charged polyelectrolyte chain with fractional charge on each
chain
monomer f, they have shown that the critical charge Qcrit on a p
at
which the charged globule becomes unstable is proportional to N (see
Fig. 7). For charge on the chain Q above the critical value Qcrit, the
polyelectrolyte chain first assumes a dumbbell configuration (see Fig. 8). At
still higher charge, the polymer forms a necklace with three beads joined by
two strings (see Fig. 8). These simulations have shown that in fact there is a

FIG. 7 Reduced mean square radius of gyration hR2g i=N 2=3 as a function of reduced
valence fN1/2 for chains with degrees of polymerization N 16, 32, 64, 128, and 200.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

274

Dobrynin

FIG. 8 Typical configurations of a freely-jointed uniformly charged chain with

N 200 interacting via Coulomb and Lennard-Jones potentials (with "LJ 1.5 and
u 2) for three different charge densities: (a) spherical globule for f 0; (b) dumbbell
for f 0.125; (c) necklace with three beads for f 0.15.

cascade of transitions between necklaces with different numbers of beads as

charge on the chain increases.
The effect of solvent quality and salt concentrations on the cascade of
transitions between different pearl-necklace structures was investigated by
Monte Carlo [69] and by molecular dynamics [70] simulations. In these
simulations the effect of the salt has been taken into account through the
DebyeHuckel potential. In [69] the authors have investigated the full
domain of stability of pearl-necklace globules in the solvent quality/salt
concentration plane. They found necklaces with up to twelve beads for a
polyelectrolyte chain with degree of polymerization N 200. With increasing
salt concentration the necklace transforms into a cylindrical globule and at
very high salt concentrations, when electrostatic interactions are completely
screened, a polyelectrolyte globule takes once again a spherical shape.
These results of computer simulations are in a good qualitative agreement
with theoretical models [68,71,72] of a polyelectrolyte chain in a poor
solvent.

D. Conformational Properties of a
Polyampholyte Chain
Polyampholytes are charged polymers with both positively and negatively
charged monomers. In this case the Coulombic interactions between charges
are not only repulsive, as in the case of polyelectrolytes, but are also

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

275

attractive in nature. The overall size and shape of polyampholytes is

determined by the balance of four factors [7378].
1.
2.

3.
4.

Chain entropy tends to keep the polymer configuration as close to

the Gaussian statistics as possible.
Fluctuation-induced attractions between charges (similar to those
observed in electrolyte solutions [79]) tend to collapse the chain
into a globule.
Excluded volume interactions between monomers stabilize the size
of globule.
If the chain has a nonzero total charge (either positive or negative)
the overall Coulomb repulsion between excess charges tends to
elongate the chain.

The relative importance of these factors depends on the number of

positive N and negative N
charges on the chain, on the degree of
polymerization N, and on the ratio of the Bjerrum length lB to the bond size
b, defining the strength of the electrostatic interactions.
P
The effect of the overall excess charge Q ej qi j ejN
N
j on the
i
ensemble averaged properties of polyampholyte chains
was the central point
of lattice Monte Carlo simulations of polyampholytes [8084], of off-lattice
Monte Carlo simulations of a bead-spring model of a polyampholyte chain
[85], and of the MD simulations [86,87]. These simulations have shown that
nearly symmetric p
random
polyampholytes with small charge asymmetry,

Q smaller than N N
, collapse into a globule as the temperature

decreases. This collapse is caused by fluctuation-induced attraction between
charged monomers. Assuming that the charges of both signs are distributed
with average concentration cch N N
=R30 within the volume of
a polymer coil, the fluctuation-induced attraction energy Watt
3=2 3
R0 [7375] of a Gaussian polyampholyte chain with size

kB Tlp
B c
ch
R0  b N . A polyampholyte chain collapses when the fluctuation-induced
attraction energy Watt becomes stronger than the thermal energy kBT. This
happens
for a value of the interaction parameter u larger than
p

N =N N
.
p
Polyampholytes with charge imbalance, Q larger than N N
, form
a necklace globule at low temperatures. There is a striking similarity
between
the instability of a polyampholyte globule with excess charge
p
Q  N N
and the necklace instability of a polyelectrolyte chain in a
poor solvent [68]. But for polyampholytes the factors responsible for the
shape of a globule are all electrostatic in nature. The main difference
between polyampholytes and uniformly charged polyelectrolytes is the
randomness in the charge sequence. For polyampholytes the structure of the
necklace is predetermined by the initial charge distribution. Monte Carlo

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

276

Dobrynin

studies by Kantor and Kardar [82,83] showed that the necklace may consist
of a few almost neutral globules connected by charged necks, or even of one
big neutral globule with a tail sticking out of it. The necklace instability was
also observed in MC simulation of circular polyampholytes by Lee and
Obukhov [88].
The effect of charge sequences on the collapse of polyampholytes was
studied by lattice MC simulations [89,90]. Polyampholytes with alternating
distribution of charges behave like polymers with short-range interactions
and can be characterized by an effective second virial coefficient [91]. The
collapse transition of these polyampholytes is similar to the coilglobule
transition of a neutral polymer [89]. A qualitatively different picture of the
collapse transition was discovered for diblock polyampholytes. The collapse
of these polymers happens in two stages. First, at high temperatures, a
zipping transition occurs that corresponds to the formation of dipolar pairs
between oppositely charged monomers. Second, at lower temperatures this
preassembled zipped structure undergoes an ordinary coilglobule transition. The possibility of a freezing transition in random polyampholyte
globules was investigated in computer simulations [92].
The influence of a uniform external electric field E on the conformation
and dynamics of polyampholyte chains has been studied by the MD
simulations of Soddemann et al. [87]. These simulations have shown that
the polyampholyte globule becomes unstable above the critical external
electric field Ec1 and breaks up forming a random necklace structure. A
polyampholyte collapses back into a globule as the strength of the external
electric field E is lowered below Ec2. The strength of the electric field Ec2 at
which a polyampholyte collapses back into a globule is weaker than that
required to break up a globule Ec1 (Ec2<Ec1). There is a hysteresis in
chain size dependence on the external electric field. This hysteresis is
a manifestation of the coexistence of two states of a chaincollapsed and
elongatedseparated by a barrier.

III. SIMULATION METHODS FOR SOLUTIONS

OF CHARGED POLYMERS
Molecular simulations of polyelectrolyte solutions at finite concentrations
with explicit counterions and salt ions require special handling of the
Coulombic interaction between the charges. These simulations are
commonly performed under periodic boundary conditions to ensure that
small sample surface effects are suppressed and that the results obtained for
systems with various numbers of particles can be considered as approximations of properties of the hypothetical infinite system. The earliest computer

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

277

simulations of polyelectrolyte solutions [9397] were done by using the

minimum image convention for electrostatic interactions, which includes
Coulombic interactions only between charges within the first periodic
image. The minimum image convention is usually sufficient for short-range
interactions, but the long-range nature of Coulombic interactions causes a
pair of charges to interact far beyond their first periodic image. The Ewald
summation method [98] allows us to overcome this problem and properly
account for contributions from all periodic images. We begin this section
by describing methods such as the Ewald summation method, particle
particle particlemesh method (P3M), particlemesh Ewald method, and
Fast Multipole Method. These methods have become the standard tool
for evaluation of long-range electrostatic interactions in simulations of
molecular systems [102,103].

A. Lattice-Sum Methods for Calculation

of Electrostatic Interactions
1. Ewald Summation
Consider an electroneutral system of Np charges eq1, eq2, . . . , eqNP such that
q1 q2    qNP 0, located at positions r1, r2, . . . , rNP within the unit cell
of size L. For simplicity we will use a cubic unit cell with primitive vectors
a, 1, 2, 3 forming the edges of the unit cell, however the methods
described below are valid for any Bravais lattice. [For the cubic lattice these
vectors are a1 (L, 0, 0), a2 (0, L, 0) and a3 (0, 0,L).] The charges {eqj}
interact according to Coulombs law with each other and with their infinite
number of replicas in periodic boundary conditions (see Fig. 9). The point

FIG. 9

2D periodic system of 3  3 periodic lattice built from unit cells.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

278

Dobrynin

charge eqk located at point rk interacts with all other charges eqj at positions
rj (k 6 j ) within the cell as well as with all of their periodic images located at
rj n1a1 n2a2 n3a3 where n1, n2, n3 are all integer numbers between
M
and M. The electrostatic energy of the unit cell interacting with an infinite
number of periodic images can be written as follows
 

X
Uel frj g
lB X
q q
 k j 
lim
rkj n
M!1 2
kB T
n k, j

16

where we introduced rkj rk
rj. The first sum in Eq. (16) is over all vectors
n n1a1 n2a2 n3a3, and the asterisk indicates that the terms with k j
for n 0 are excluded from the summation.
The sum (16) is only slowly and conditionally convergentthe result
depends on the order in which the terms are added. The original method for
summation of the lattice sum was introduced by Ewald [98], who replaced
the sum in Eq. (16) by the sum of two absolutely convergent seriesa direct
sum in Cartesian space and a reciprocal sum in Fourier space. There is a
simple physical interpretation of this decomposition of the lattice sum
(see Fig. 10). In the direct sum each charge in the system qi is viewed as
being surrounded by a Gaussian charge distribution of the opposite sign
[r
qi 3 exp
2 r2 =3=2 , where the parameter  determines the width
of the distribution], such that the net charge of this cloud exactly cancels qi.
The electrostatic interaction between screened charges is a rapidly decaying
function of distance, and is effectively reduced to short-range interactions in
the direct space. To counteract this diffuse cloud around each charge, a
second Gaussian charge distribution of the same sign as the original charge
is added for each point charge. This second distribution varies smoothly in
space and acts as a source term in a Poisson equation for the reciprocal
potential with periodic boundary conditions.
However, these most commonly used Ewald transformation formulas are
correct for centrosymmetric crystal structures but give results [98101] that

FIG. 10 The Ewald sum components of a one-dimensional point charge system.

The vertical lines are unit charges.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

279

disagree with the explicit evaluation of sum (16) in the case where the unit
cell has a net dipole moment [104107]. The discrepancy is due to uncritical
use of the Fourier transformation converting the conditionally convergent
lattice sum in direct Cartesian space into a sum in reciprocal Fourier space.
The evaluation of the lattice sum (16) is cumbersome and details of
this calculation are given in the Appendix. The final expression for the
electrostatic energy of the cell interacting with all its periodic images (16)
can be represented as a sum of four terms: sums in the direct Udir ({rj}) and
reciprocal Urec ({rj}) space, self-energy term, and shape term that depends
on the net dipole moment of the cell
 
 


X
Uel frj g
erfc rkj n
lB X



qk qj
rkj n
2 n k, j
kB T


lB X exp
2 m2 =L2 2

SmS
m
m2
2L m60

2

lB X 2 2lB X


p
qj
q
r
17


j
j

3L3  j
 j
p R 1
error function
where erfcx 2=  x dt exp
t2 is the complementary
P
which tends to zero as x ! 1 and S(m) qj exp i2m  rj =L is the
j
charge structure factor.
The Coulombic forces on charge k are obtained by differentiating the
electrostatic energy of the system Uel({rj}) with respect to rk. The resulting
force
shape
rec
Fk Fdir
k Fk Fk

18

has three contributions: the real space part

Fdir
k

kB TlB qk

X
j

qj

X
n

 !

 
2 2  erfc rkj n
2




p exp
rkj n 
rkj n


rkj n


rkj n2

19

the reciprocal (Fourier) space part

Frec
k






2lB qk X im exp
2 m2 =L2 2
2im  rk
kB T
exp
S
m
m2
L2 m60
L

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

20

280

Dobrynin

and the dipolar contribution

Fshape

kB T
k

4lB qk X
qj r j
3L3 j

21

The value of the parameter  controls the rate of convergence of the

direct Udir ({rj}) and reciprocal Urec ({rj}) sums. Increasing  causes the
direct sum Udir ({rj}) to converge more rapidly, at the expense of slower
convergence in the reciprocal sum Urec ({rj}). The usual method is to take a
value of  such that only one term with n 0 is necessary in the direct sum.
Thus, the first sum reduces to the normal minimal image conversion. To
achieve the same accuracy in the reciprocal sum, a subset of m0 m-vectors
has to be used. Usually, in practice the parameter  is set to 5/L and it is
required to use on the order of 100200 wave vectors for reasonable
convergence of the reciprocal sum. This choice of parameter leads to a linear
increase of the computational time for evaluation of the reciprocal sum with
the number of particles in the system Np. The major computational
overhead comes from evaluation of the direct sum, which requires
evaluation of Np2 =2 error functions. For large systems, Np>104, this
method becomes computationally very inefficient.
A more efficient algorithm for evaluation of the lattice sum was designed
by Perram et al. [108]. In this algorithm each side of the unit cell is divided
into K equal segments, each of length L/K. This leads to tessellation of the
original cell into MB K3 identical cubic subboxes, each containing on
average Np/MB particles. If we apply the minimum image convention for
each subbox for evaluation of the direct space contribution to the lattice
sum, such that the interactions between particles separated by distances
larger than L/K ( K=L) are neglected, the number of operations required
for this evaluation will decrease with the number of subboxes MB as
MB Np =MB 2 Np2 =MB . To maintain the same accuracy in evaluation of
the reciprocal sum, the number of wave vectors in the reciprocal space
should be proportional to the volume of a sphere of size . (The
contributions from the vectors with modulus jmj > L K are exponentially suppressed.) The number of operations involved in the evaluation of
the reciprocal sum will be of the order of NpK3 NpMB and will increase
for
with the number of boxes MB. By determining the number of boxes
p

N
optimal computational efficiency of this method one finds
that
M
B
p.
p
If the boxing is implemented with the optimum value Np the computer
time for lattice sum evaluation increases only as Np3=2 . This is a considerable
improvement over the N2 rate for traditional methods with fixed value of .
Before closing this section, let us comment on the parallel implementation of the Perram et al. algorithm [108]. In practice the computation of the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

281

direct and reciprocal space contributions to the Ewald sum are done
separately. Since the direct sum represents short-ranged interactions
between particles, the parallelization of the summation in the direct sum
is handled by three-dimensional spacial domain decomposition based
upon blocks of chaining cells, considering the interactions only between
neighboring cells. The computation of the sum in the reciprocal (Fourier)
space is ideal for parallel implementation. The required charge structure
factor S(m) is rewritten as follows
X

 XX

 X
qj exp i2m  rj =L
qj exp i2m  rj =L
SP m
Sm
j

j2P

22
where P denotes processors. Particles are distributed to the processors and
the partial charge structure factor SP (m) is computed on each processor for
all m vectors. These partial charge structure factors SP (m) are then summed
across the processors to obtain the total charge structure factor S(m).
Computation of the reciprocal forces, where summation of factors
2imS(m) is involved, does not require further computer communications.
The factors 2imSP(m) can be summed on each processor for the whole
range of m vectors.

2. Particle Mesh Ewald Method (PME)

The particle mesh Ewald (PME) method [109113] is based on approximations that turn the structure factor S(m) into discrete Fourier transforms,
making the evaluation of the reciprocal sum more computationally efficient.
This is achieved by substituting the real charge distribution by a weighted
charge distribution on a regular three-dimensional grid.
Consider a grid of length L with K subintervals along L with the lattice
spacing of the grid being h L/K. Let us rescale all lengths in the system by
the grid spacing h, such that u r/h. Due to periodic boundary conditions
the rescaled coordinates u vary in the interval 0  u<K for  1, 2, 3. After
rescaling, the charge structure factor is



X
2m  uj
Sm
23
qj exp i
K
j
In the PME algorithm, the complex exponentials












2m  u
2m1 u1
2m2 u2
2m3 u3
exp i
exp i
exp i
exp i
K
K
K
K

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

24

282

Dobrynin

are replaced by the linear combination of their values at the nearby grid
points. For example, let [u] denote the integer part of u. Then, using linear
interpolation, we can approximate the individual exponents in the r.h.s. of
Eq. (24) by their linear combination at the nearest grid points [u] and
[u] 1




2m u
2m
exp i
 1
u
u exp i
u
K
K



2m
u
u exp i
u 1
K

25

Similar interpolation can be done for each term in the Eq. (24), with
 1, 2, 3. Let us define W2(x), the linear hat function given by
W2(x) 1
|x| for |x|1, and W2(x) 0 for |x|>1. Then Eq. (25) can be
rewritten as






1
X
2m u
2m
k

W2 u
k exp i
exp i
K
K
k
1

26

The sum in the last equation is finite, because the function W2(x) is nonzero
only within the interval |x|  1. The accuracy of the approximation can be
improved by including more grid points in the interpolation scheme.
In the original implementation of the PME algorithm [109,111],
Lagrangian weight functions W2p(x) [116] were used for interpolation of
the complex exponents, by using values at 2p points in the interval |x|  p.
Unfortunately, these functions W2p(x) are only piecewise differentiable,
so the approximate reciprocal lattice sum cannot be differentiated to arrive
at the reciprocal part of the Coulomb forces. Thus, the forces were
interpolated as well.
Instead of the Lagrangian weight functions W2p(x), the Cardinal
B-splines Mn(x) [117,118] were utilized in later versions of the PME
method [110,112,113]. These weight functions are continuously differentiable and allow the forces to be obtained from analytical differentiation of the
approximation of the reciprocal lattice sum. The Cardinal B-spline of
the second order M2(x) gives the linear hat function M2(x) 1
|x
1| in
the interval 0  x  2 and M2(x) 0 for x<0 and x>2. The nth order
B-spline satisfies the following properties:
1.
2.
3.

Mn(x)>0 for 0  x  n and Mn(x) 0 for x<0 and x>n.

Mn(x) Mn(n
x).
P1
j
1 Mn x
j 1.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

4.
5.

283

Mn x x=n
1Mn
1 x n
x=n
1Mn
1 x
1.

d=dxMn x Mn
1 x
Mn
1 x
1.

The interpolation of the complex exponential, called the Euler

exponential spline, has a simple solution that, for even n, has the form






1
X
2m u
2m
exp i
k
Mn u
k exp i
 b m 
K
K
k
1

27

where the coefficient b(m) is given by

"X


#
1
n
2
2m
2m
n
1
k
Mn k 1 exp i
b m exp i
K
K
k0

28

This approximation for the complex exponential in the charge structure

factor leads to the following approximation for S(m)





2m
k
Mn uj, 
k exp i
K
j1
1
k
1



3
X
Y
2

b m 
Qk exp i m  k
K
k2V
1

Sm 

NP
X

qj

3
Y

b m 

1
X

bm
b1 m1 b2 m2 b3 m3 Q

29

where k is the vector with components (k1, k2, k3), VK {0  k  K
1,

bm is the discrete Fourier transform of the charge
for  1, 2, 3}, and Q
array Q, defined as

Qk

NP
X
j1

qj

3 X
Y



Mn uj, 
k
s K

30

1 s

where the inner sum is over all integers s. Using this spline representation
of the charge structure factor S(m), the approximate reciprocal energy can
now be rewritten as
 


Urec fuj g
lB X exp
2 m2 =L2 2
bmQ
b
m

BmQ
kB T
m2
2L m60

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

31

284

Dobrynin

where we introduce
Bm

3 
Y

b m 2

32

1

The last expression can be simplified further by introducing the reciprocal

pair potential





exp
2 m2 =L2 2
1X
2
1

m
Bm
exp
i
rec 1
 m60
K
m2
b1
C

33

where C is defined as
8


exp
2 m2 =L2 2
<1
, for m 6 0
34
Cm  Bm
m2
:
0,
for m 0
 0 0 0
with vector m, defined as m m1 , m2 , m3 , where m0 m for 0  m  K/2
and m0 m
K=2 otherwise. It is important to point out that
rec
1 , where
rec )
1 is the inverse discrete Fourier transform defined
C d
P (d

3
b
as Am K
k2V Ak exp
i2=Km  k. Using the property of the
b
m K 3 Am, after some algebra, the reciproFourier transformation A
cal lattice sum reduces to
 
Urec fuj g
lB X

Qmrec
Qm
35
kB T
2L m2V
where rec
Q denotes the convolution of rec and Q defined as
rec
Qm

K

1 K

1 K

1
X
X
X

rec m1
k1 , m2
k2 , m3
k3 Qk1 , k2 , k3

k1 0 k2 0 k3 0

36
To obtain the reciprocal part of the Coulomb forces, Eq. (35) has to be
differentiated with respect to rj.
Frec
j



@Urec fuj g
kB TlB X @Qm


rec
Qm
@rj
L m2V @rj

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

37

Molecular Simulations of Charged Polymers

285

The numerical implementation of this algorithm involves the following steps

[110,111]:
1.
2.

3.

4.

5.
6.

Determine B-spline coefficients b(m).

Fill the charge grid array Q, using coefficients Mn(uj,
k) for
j 1, . . . , Np,  1, 2, 3, and j 1, . . . , n, computed from scaled
fractional coordinates of the particles uj,. The cost of this step is
O(Nn3), where n is the order of Cardinal B-spline.
Calculate the inverse 3DFFT of the charge grid array Q. The
cost of this step is O(K3 log K3). For dense systems K3 is of the
order of N.
The approximate expression of Urec ({uj}) is computed by using
Eq. (35). At the same time the transformed Q array is overwritten
by the product of itself with the arrays C and B.
Forward 3DFFT on the new Q array to evaluate the convolution
3
3
rec*Q. The cost of this step is again O(K log K ).
Generate forces by using convolution and Eq. (37). The cost of this
step is O(N).

Thus, the overall cost of the PME algorithm is O(NlogN). Direct

comparison of the Ewald [108] and PME methods has shown that the PME
method is significantly faster than the Ewald method for Np>104 [110,111].
The PME method with Cardinal B-spline interpolation scheme is
included as a standard routine in the AMBER [115] and DL_POLY [114]
simulation packages.

3. ParticleParticle ParticleMesh Method (P3M)

The key to the P3M method lies in splitting the interparticle Coulombic
forces into a smoothly varying long-range component and a shortrange component that is nonzero only for particles whose separation
is less than the cutoff distance [119124]. These two forces must be properly
matched for the total force to be correctly reproduced. The total shortrange force on the particle is computed by the direct summation of
the particleparticle (PP) forces, while the smoothly varying longrange component is approximated by a particlemesh (PM) force calculation. A key feature of the PM calculation is the use of the FFT convolution
method to solve for the electrostatic potential at a given particle
distribution.
The traditional Ewald representation of the Coulombic lattice sum as a
sum of the two fast converging series is a perfect example of such splitting
into the short-range and smoothly varying long-range components. In
the Ewald sum the direct sum is due to the point charge and Gaussian

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

286

Dobrynin

counter-ion cloud and is short ranged, while the reciprocal sum is due to the
Gaussian charge cloud and is a smooth function, with its Fourier transform
rapidly convergent. Below we outline the general steps of particlemesh
calculations of the long-range component.
In the P3M method the Gaussian charge clouds surrounding each charge
within the unit cells are substituted by finite range weight functions Wn(x)
[123,124] that interpolate the original charge density over the n grid points,
with spacing between them hL/K.

Qr

X qj  3
j

3=2


X

2 


) Qg k
exp
2 r
rj
qj W n uj
k

38

where k is the radius vector of the grid point and vector k(k1, k2, k3) with
all 0  k  K
1 being integers. As before, we rescale all lengths in the
system by the grid spacing h, such that uj rj/h. Due to periodic boundary
conditions, the rescaled coordinates u vary in the interval 0  u<K for
 1, 2, 3.
The simplest one-dimensional weight function W1(x) [W1(x) 1, for
|x|  1/2 and W1(x) 0, for |x|>1/2] assigns the charge densities to the
nearest grid point. The full 3D weight function is a product of weight
functions in the x, y, and z directions. HigherR order weight functions are
1
convolutions of W1 x,Wn x W1
Wn
1 x
1 W1 x
x1 Wn
1 x1 dx1 ,
and span n grid points in each direction. One can also use the Cardinal
B-splines, introduced for interpolation of the exponentials in the PME
method, as the weight functions for charge assignment. The higher order
Cardinal B-splines can also be generated by convolutions of the hat
function M1(x) [M1(x) 1, for 0  x  1 and M1(x) 0 otherwise]. In fact,
the charge assignment functions Wn(x) and Mn(x) are identical up to
translation.
Since the original charge density Q(u) has been replaced by charge
density over the grid points Qg(k), Hockney and Eastwood [119] suggested
to minimize the effect of such substitution by replacing the Coulomb
Greens function by function Gn(q), which will minimize the mean square
error in forces due to the new assignment function Wn and finite size grid
errors.
Consider two charges q1 and q2 placed at random positions u1 and
u2 in the unit cell. The mean square error in force due to charge assignment is
Z


du1
K3

2
du2  EW
F u2 , u1
FPM u2 , u1 
3
K

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

39

Molecular Simulations of Charged Polymers

287

where FEW (u2, u1) is the exact reciprocal force and FPM (u2, u1) is the
bn q that minimizes
approximated force. The optimal Greens function G
the functional (39) is [119,122124]
bn q lB
G
L

2

bqW
bn b q
q


A
b

2
2
P 

b
W
q2
b

q

b n

40

bq 1=q2 exp
2 q2 =2 and W

bn q is the Fourier transform of the
where A
3D assignment function. The sum in Eq. (40) is taken over a Brillouin zone
vector b (b1, b2, b3) with b1, b2, b3 all integers and the vector q belongs to
the first Brillouin zone with components
K/2  q<K/2,  1, 2, 3. Since
this expression does not depend on the particle positions, it has to be evaluated only once at the beginning of the simulation.
The usual PM calculations of the long-range forces between particles in
the P3M consist of the following steps (see Fig. 11):
1.

Form an effective grid charge density Qg(k) by assigning the charges

over the grid points. This charge density is Fourier transformed by
using the forward discrete Fourier transform
bg q
Q

Qg k expik  q

41

FIG. 11 A 2D schematic of the particlemesh technique. (i) System of charged

particles. (ii) Charges are interpolated on a 2D grid. (iii) Using FFT, to solve for the
potential and forces at grid points. (iv) Interpolation of forces back to particles.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

288

2.

Dobrynin

where vector q is the vector with components (q1, q2, q3) with each
q 2m/K and 0  m  K
1.
bn q and charge distribution
Using the modified Greens function G
b
g q over the grid points
Qg q solve for the electrostatic potential b
bn qQ
bg q
b
g q G

42

and find the Fourier transform of the electric field

b
bn qQ
bg q
Eg q iqb
q iqG
3.

Using the inverse Fourier transform find the electric field at the grid
points Eg(k)
Eg k

4.

43

1 Xb
Eg q exp
ik  q
K3 q

44

Note, that three inverse three-dimensional Fourier transforms have

to be calculated to obtain the electric field distribution over the grid
points because b
Eg q is a vector.
Finally, the electric field Eg(k) is interpolated back to the positions of
the particles giving the long-range component of the forces on them
X
 
Wn uj
kEg k
FPM uj qj

45

The interpretation of the back interpolation is very simple, since each

charge was replaced by several subcharges located at surrounding
grid points. The force acting on each subcharge is given by its
charge qjWn(uj
k) times the electric field at its grid point Eg(k).
Thus, the net force on the original charge qj is the simple sum of
all forces acting on its subcharges.
As one can see from the description of the algorithm for the P3M method,
it is very close to the one presented for the PME method in the previous
section. The PME differs in the final steps, in which the potential is first
interpolated onto the particles and the resulting potential is differentiated
with respect to the particle positions to give the electric field and forces. The
PME uses half as many FFTs as P3M; it is more efficient where the cost of
the FFT becomes an issue, such as in the parallel implementations. The
number of FFT transformations in the P3M method can be reduced by first

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

289

calculating the electrostatic potential over the mesh points by applying the
inverse FFT to the electrostatic potential g(q). The electric field at the mesh
points is then obtained by numerical differentiation of the electrostatic
potential at the neighboring grid points. However, this implementation of
the P3M is less accurate than ones that use direct differentiation in the
Fourier space [125,126]. Extensive comparisons and analysis of both
methods can be found in [112,113,125,126].

B. Fast Multipole Method for Ewald Summation

The Fast Multipole Method (FMM) was originally introduced by
Greengard and Rokhlin [127130] for efficient simulation of Np particles
interacting through a Coulomb-like potential confined in a nonperiodic cell.
The FMM relies on the standard multipole expansion for the electrostatic
potential (forces) by separating the pairwise interactions into two
components: one due to nearby particles, computed directly, and another
due interaction with distant particles, approximated by their multipole
expansions.
Consider a set of point charges qj ( j 1, . . . , k), see Fig. 12, within a
sphere of radius R, whose center is at distance r far away, r |r|>R, from a
test charge located at point Q. The electrostatic potential (r) at point Q due
to all charges qj at locations (j, j, j) is given by the infinite multipole
expansion if, for all j, j<r
r kB TlB

1 X
n
X
Mnm m
Y
,
rn1 n
n0 m
n

46

where Mnm are multipole moments, defined as

Mnm

k
X



qj nj Yn
m j , j

j1

FIG. 12

The multipole expansion method.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

47

290

Dobrynin

where Ynm
, is the spherical harmonic polynomial [131]. If we approximate this expansion by first p 1 terms, the truncation error  is bounded
by (r/R)
p [127,128]. Thus, fixing a precision , the number of terms in the
multipole expansion is estimated to be p
log , where  r/R.
There is a duality inherent in the situation depicted in Fig. 12; if the
location of test charge Q and a system of charges qj ( j 1, . . . , k) were
interchanged, we may write another expansion that represents the potential
due to particles outside the sphere of radius r and is correct as long as j>r,
r kB TlB

1 X
n
X

m
rn Lm
n Yn
,

48

n0 m
n

where Lm
n are coefficients of the local expansion defined as
Lm
n

k
X



qj 
n
1
Yn
m j , j
j

49

j1

To illustrate the method, let us consider a square simulation box with

sides of length one centered about the origin of the coordinate system and
containing Np particles inside. Fixing the precision , we choose r/R 2,
leading to the number of terms in the multipole expansion to be p 
log2 ,
and specify that no interactions be computed for clusters of particles which
are not separated by distances r<2R. In order to impose such a condition,
we introduce a hierarchy of meshes which refine the computational box into
smaller and smaller regions (Fig. 13). Mesh level 0 is equivalent to the entire
box, while mesh level l 1 is obtained from level l by dividing each box into
four equal parts. The number of distinct boxes at mesh level l is equal to 4l.
A tree structure is imposed on this mesh hierarchy; the four boxes obtained
on level l 1 by subdivision of the box on level l are considered to be its
children. Now consider some box S at level n. It is not justifiable to apply
the multipole expansion to nearest neighbor boxes, because particles in

FIG. 13
Hierarchical subdivision of the full simulation space (2D case) into
children, grandchildren, etc.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

291

neighboring cells might be within the cutoff distance 2Rn for this level.
Nearest neighbor cells are defined as cells sharing an edge or a corner
(in 3D also a face) with a given cell. The interactions between charges in cell
S with its nearest cells are taken by summing the electrostatic interactions
directly. We apply this approximation at each step to at least next nearest
neighbors and skip squares which lie in the regions that have been treated
at previous levels. Therefore, the boxes with which the particles in box S
interact at the present level, are (1) those that are not nearest neighbors of
S and (2) those whose parents were a nearest neighbor of the parent of S
at level n
1. Figure 14 shows the squares that are in the interaction list
of cell S.
The realization of the Fast Multipole Method involves six steps.
1.

2.
3.

4.

The computational cell is subdivided into a hierarchy of meshes. (For

a three-dimensional cubic cell the root of the hierarchy is the simulation box itself, which is divided into 8 cubes. Each of these cubes is
then subdivided and so on for a prescribed number of subdivisions, l,
so that at the finest level there are 8l cubes.)
Compute the multipole moments about the center of each cell at the
finest level of hierarchy.
Upward pass. Sweep up from the smallest cells to the largest cell to
obtain the multipole moments Mnm for cells at each subdivision level.
This is done by using the translation operators that allow us to
obtain the multipole moment of the parent cell from multipole
moments of children cells (see for details [123,128,132]).
Downward pass. Sweep down from the largest (single) cell to cells
at the next hierarchy level to obtain the local expansion coefficients
Lm
n by multipole-to-local and local-to-local translations. This local

FIG. 14 Interaction list of cell S at level n. The squares at level n are separated by
thin lines, their parents at level n
1 by heavy lines. The cells in the interaction list
are labeled by I. The nearest neighbor cells are labeled by N.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

292

5.
6.

Dobrynin

expansion describes the field due to all particles in the system that are
not contained in the current cubic box, its nearest neighbors and
second nearest neighbors.
Evaluate the potential and fields for each particle using the local
expansion coefficients on the finest level.
Add the contributions from other charges in the same cell and
near cells that are not included in the multipole expansion by
direct summation.

This method was later generalized to systems with nonuniform charge

distribution by using an adaptive grid [133].
The FMM was extended by Schmidt and Lee [134,135] to systems with
periodic boundary conditions. This approach combines both FMM and
Ewald techniques. First, the FMM is initiated as in the finite case,
calculating the multipole expansion of all cells at all refinement levels. The
level zero expansion then contains the multipole expansion for all particles
in the original simulation cell. All of its periodic images have the same
multipole expansions about their centers. The FMM requires the local
expansion of the potential from periodic images except the 26 nearest
neighbors of the original simulation cell. This local expansion is obtained by
using the Ewald sum formulation for multipole-to-local translation for
image cells. After that, the algorithm continues in its downward pass as for
the finite system.
Pollock and Glosli [123] have performed a timing comparison of Ewald,
P3M, and FMM algorithms by calculating the forces and potential energies
of random periodic configurations of Np 512, 1000, 5000, 10,000, and
20,000 charges (see Fig. 15). They concluded that P3M and FMM are more
efficient than Ewald summation, but P3M is roughly four times faster than
the FMM algorithm for all Np used in this study, despite the superior
asymptotic scaling of the FMM O (Np) vs. the O (Np logNp) for P3M. (An
extrapolation based on the scaling of both methods suggests that FMM
becomes faster at some unphysically large system size Np>1060.) However,
the speed is not the only disadvantage of the FMM; it is also more diffcult to
code than P3M.

IV. POLYELECTROLYTE SOLUTIONS

A. Polyelectrolytes in Good and h Solvents
In very dilute salt-free solutions the interaction between polyelectrolyte
chains and counterions is weaker than the contribution of the configurational entropy of counterions, and counterions are thus distributed almost

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

293

FIG. 15
Timing for complete force calculation of various size systems using
standard Ewald, fast multipole method, and P3M. (From Ref. [124].)

homogeneously in the solution with only a small fraction being near the
polyelectrolyte chains. The conformation of a chain is then determined by
the chain elasticity and by the electrostatic repulsion between charged
monomers (see Section II). As a result, chains adopt elongated conformations with size L  bNuf 2 1=3 [see Eq. (10)]. With increasing polymer
concentration c, more and more counterions are drawn toward polyelectrolyte chains and, for sufficiently strong electrostatic interactions, can even
be localized on them [3,136]. This phenomenon is called counterion or
Manning condensation [3,136]. Localization of counterions leads to effective
renormalization of the total charge on a chain Qeff feffN<fN. A simple
estimate of the effective charge Qeff can be achieved by equating the
chemical potential of a counterion in a solution far away from a chain
 kBT ln (cb3feff) with that of a counterion localized within the chain
1=3
2 1=3
volume LD2e ,  kB T lnf
feff ufeff

kB Tu2=3 feff
[3,137140]. This
leads to the effective fraction of the charged monomers on the chain
!
 2=3 1=3 
cb3
exp u f
feff  f 1

50
uf 2 1=3
being a decreasing function of polymer concentration c. The parameter
u2/3f 1/3 is the Manning parameter which determines the onset of counterion
2 1=3
condensation. Thus, the chains shrink, L  bNufeff
, as polymer concentration increases. However, for small values of the Manning parameter
u2=3 f 1=3 1 the corrections to the chain size are small and can be
neglected.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

294

Dobrynin

Chains overlap when the distance between them Rcm  (N/c)1/3 becomes
of the order of their size L  N. This happens at polymer concentration
c*  N
2. In semidilute solution, c>c*, polyelectrolytes form a temporary
network with mesh size . At length scales smaller than , the intrachain
electrostatic repulsion dominates and sections of the chain with g monomer
are strongly elongated such that  g. Pure geometric arguments can be
used to obtain the concentration dependence of the correlation length , by
imposing the close packing condition for chain sections of size ,
c  g/3  
2. This leads to the correlation length   c
1/2 [26,27,29].
At length scales larger than  the electrostatic interactions are screened by
the other chains and counterions and the statistics of a chain is Gaussian
with effective bond lengths of the order of the correlation length . Thus,
the chain
size
R in the semidilute salt-free polyelectrolyte solution is
p

R   N=g  N 1=2 c
1=4 [26,29].

Polyelectrolyte solutions have a number of properties remarkably
different from solutions of uncharged polymers. In particular:
1.

2.

There is a well pronounced peak in the scattering function of a

homogeneous polyelectrolyte solution, whose position shifts with
polymer concentration [1,4,6,7,26,27,29,141145].
At low salt concentrations the main contribution to the osmotic
pressure comes from counterions [1,4,6,7,26,27,29].

Extensive molecular dynamics simulations of dilute and semidilute

polyelectrolyte solutions of chains with degree of polymerization N ranging
from 16 up to 300 were recently performed by Stevens and Kremer
[146148] and by Liao et al. [149]. In these simulations the long-range
electrostatic interactions were taken into account by the Ewald summation
method, including interactions with all periodic images of the system.
Stevens and Kremer [146148] have used a spherical approximation of
Adams and Dubey [150] for the Ewald sum while Liao et al. [149] have
applied the PME method [110]. In addition to Coulombic interactions, all
particles, including monomers and counterions, interacted via a shifted
Lennard-Jones potential with cutoff rcut 21/6
8
 

 12  6 1
<
51
ULJ r 4"LJ r
r 4 , for r  rcut
:
0,
for r > rcut
The polymers were modeled as bead-spring chains. The attractive part of the
bond potential is described by the FENE potential



1 2
r2
UFENE r
kr0 ln 1
2
52
2
r0

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

295

with the maximum extent r0 2 and the spring constant k 7"LJ/ 2. The
repulsive part of the bond potential was described by the shifted LennardJones potential. The strength of the electrostatic interactions in the simulations was controlled by the value of the Bjerrum length. In the simulations
of Stevens and Kremer [146148] most of the results were obtained for the
Bjerrum length lB equal to 0.833, which corresponds to the case of weakly
charged polyelectrolytes with the value of the coupling parameter u<1,
while in the Liao et al. [149] simulations the Bjerrum length lB was 3,
corresponding to strongly charged polyelectrolytes.
Figure 16 shows the dependence of the osmotic pressure of salt-free
polyelectrolyte solutions. Through almost the entire concentration range
considered, the osmotic pressure is proportional to the polymer concentration, supporting that it is controlled by the osmotic pressure of counterions,
both above and below the overlap concentration. There appears to be a
weak chain length dependence of the osmotic pressure for short chains.
However, this N-dependence is consistent with a 1/N correction to the
osmotic pressure due to chain translational entropy. The deviation from
linear dependence of the osmotic pressure  occurs around polymer
concentration c  0.07
3, which is above the overlap concentration for all
samples. At very high polymer concentrations, where electrostatic interactions are almost completely screened by counterions and by charges on the

FIG. 16 The osmotic pressure is plotted as a function of the monomer density for
various chain lengths.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

296

Dobrynin

FIG. 17 Plot of the correlation length  as a function of density for different chain
lengths. The slopes are
0.5 and
0.33 for the solid and dashed lines, respectively.

chains, the osmotic pressure  is proportional to c9/4, and the scaling law for
neutral polymers in a good solvent is recovered.
The correlation length  of the solution (see Fig. 17) exhibits two regimes
as observed in experiment [1,6] and predicted by the theory [26,29,143145].
Above overlap concentrations the correlation length  is chain length
independent and is inversely proportional to the square root of polymer
concentrations ( c
1=2 ). At low polymer concentrations  scales with
polymer concentration as c
1/3. The crossover between these two regimes
occurs around the overlap concentration c*.
The effect of added salt on the conformation of polyelectrolyte chains in
dilute and semidilute solutions was investigated by Stevens and Plimpton
[151]. At high salt concentrations the electrostatic interactions between
charged monomers are screened and the chain conformations are similar
to those observed for good solvent, R N 0:588 . As the salt concentration
decreases, chains become more elongated, and finally at very low salt
concentrations the chain size saturates at its salt-free value R N [146,148].

B. Polyelectrolytes in Poor Solvent

Molecular dynamics simulations of partially charged polyelectrolytes in
poor solvent conditions were performed by the Kremer group [153,154] for

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

297

FIG. 18 Dependence of the end-to-end distance for polyelectrolyte chains with

number of monomers N 94 (filled circles) and N 187 (filled squares) in poor
solvent. Insets are snapshots of chain conformations.

chains with degree of polymerization N 94 and by Liao et al. [155] for

N 94, 187. In these simulations, only every third monomer carried an
electric charge and poor solvent conditions were imposed by setting the
cutoff radius rcut for monomermonomer interactions to 2.5. Figure 18
shows the dependence of chain size on polymer concentration [153155].
At low polymer concentrations the polyelectrolyte chains form necklaces of
beads connected by strings. As the polymer concentration increases, the
fraction of the condensed counterions on the chain increases and chains
shrink by decreasing the length of the strings and the number of beads on
the chain. Eventually, at higher polymer concentrations, polymer chains
interpenetrate, leading to a concentrated polyelectrolyte solution. In this
range of polymer concentrations, the chain size is observed to increase back
towards its Gaussian value. The nonmonotonic dependence of the chain
size on polymer concentration is in qualitative agreement with theoretical
predictions [156].

C. Counterion Distribution and Condensation in

Dilute Polyelectrolyte Solutions
The effect of counterion distributions in the dilute polyelectrolyte solution
was the central subject of the molecular dynamics simulations performed by

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

298

Dobrynin

Limbach and Holm [157]. They have shown that the counterion distribution
around a quenched strongly charged polyelectrolyte chain shows an end
effect. The counterions are accumulated preferentially in the middle part of
the chain. Toward the end of the chain, the effective charge on the chain
increases, which is equivalent to a decrease in counterion concentration. The
reason for this behavior is the difference in the electrostatic potential created
by the charges of the polyelectrolyte chain. The electrostatic potential is
stronger in the middle of the chain than at its ends. Thus, an ion close to the
middle part of the chain is attracted more strongly than an ion sitting at the
chain end [158]. Besides the inhomogeneity of the local effective charge,
there is also inhomogeneity in the local chain conformations. The bonds are
stretched more in the middle of the chain. This nonuniform stretching of the
chain causes the polyelectrolyte to appear in the trumpet-like shape [158].
As one would expect, the end effects become less pronounced as salt is
added to the solution, due to the screening of the electrostatic interactions
over the Debye screening length.
For polyelectrolytes in a poor solvent, the situation is more peculiar. In
this case the end effects are smeared out over the end beads. There is a sharp
decrease in the effective charge in the region of the first strings. The
modulation in accumulation of the effective charge in the middle of the
chain repeats the necklace structure.
Condensation of counterions was studied by molecular dynamics
simulations [148,159]. These simulations have shown that in dilute
polyelectrolyte solutions at constant polymer concentrations the chain size
depends nonmonotonically on the Bjerrum length lB. First, the chain size
increases with increasing Bjerrum length lB, which is due to increasing
strength of the intrachain electrostatic repulsion between charged monomers. This will continue until the Bjerrum length lB reaches the crossover
value lB
. Above this value the counterion condensation kicks in, reducing
the effective charge on the chain, weakening intrachain electrostatic
repulsions, and chains begin to shrink [139,140,160].

D. How Good Is the DebyeHuckel Approximation?

Almost all theoretical works considering polyelectrolytes [7,26,27,29,
141145] have treated counterions and salt ions at the DebyeHuckel
(DH) level by preaveraging the electrostatic interactions between charged
monomers over the small ions degrees of freedom. In this approximation
the presence of counterions and salt ions leads to an effective screening
of the electrostatic interaction between charged monomers, which interact
via a screened electrostatic (Yukawa) potential. Of course, such an

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

299

approximation is only correct when the presence of polymer chains will only
slightly perturb the uniform distribution of the small ions in the system. This
approximation should also fail at high polymer and salt concentrations,
when excluded volume effects start to control the ionic atmosphere around
polyelectrolyte chains.
To investigate the effect of the DebyeHuckel approximation on the
solution properties, Stevens and Kremer [152] performed molecular
dynamics simulations of salt-free solutions of bead-spring polyelectrolyte
chains in which the presence of counterions was treated via a screened
Coulomb potential, and compared the results with their simulations with
explicit counterions [146,148]. To elucidate the effect of the DebyeHuckel
approximation, the dependence of the mean square end-to-end distance,
hR2e i, osmotic pressure, and chain structure factor on polymer concentration
was examined. Stevens and Kremer found that hR2e i tends to be larger at low
densities for DH simulations and is smaller at higher densities. However, the
difference in hR2e i between DH simulations and simulations with explicit
counterions is within 10%. This trend seems to be a generic feature for all N
in their simulations. The functional form and density dependence of the
chain structure factor are very close in both simulations. The most severe
DebyeHuckel approximation affects the dependence of the osmotic
pressure on polymer concentration. It appears that in the DH simulations
not only is the magnitude of the osmotic pressure incorrect, but also the
concentration dependence is wrong.
The effect of the Bjerrum length lB (strength of the Coulombic
interactions) on the applicability of the DebyeHuckel approximation was
investigated by Stevens and Plimpton [151]. They found that this
approximation works well for weak Coulomb interactions or for the
Bjerrum lengths lB smaller than  (see Fig. 19). From this figure we see that
the DebyeHuckel approximation breaks down for lB>. In this range of
parameters the chain size monotonically increases in the DH simulations,
while for simulations with full Coulomb interactions the chain shrinks as the
strength of these interactions increases. This discrepancy should not be too
surprising, because, for lB>, the Coulomb attraction of counterions to a
polyelectrolyte chain in its vicinity becomes stronger than the thermal
energy kBT, leading to effective localization (condensation) of counterions
near the polymer chain [3,136,139,140,160]. This condensation reduces the
charge on the chain forcing it to contract [139,140,160].
The DebyeHuckel approximation can be improved by considering the
full PoissonBoltzmann approach to electrostatic interactions, in which
the counterion condensation phenomena are included implicitly. Tests of
the PoissonBoltzmann approach to a cell model of rigid polyelectrolytes
[161,162] were done in references [163165] by performing molecular

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

300

Dobrynin

FIG. 19 Plot of the square-root mean square end-to-end distance as a function of

the Bjerrum length lB for DH simulation (open circles) and Coulomb simulation
(open squares). For these simulations N 32 and c 0.001
3. (From Ref. [151].)

simulations of this model. While the agreement between theory and

simulation results was excellent in the case of monovalent counterions
and weakly charged rods, it deteriorated with increase in the strength of the
electrostatic interactions and, in particular, with increase in the counterion
valence. At high concentrations of divalent counterions, computer simulations show charge oscillations which the theoretical model is unable to
reproduce.

E. Bundle Formation in Polyelectrolyte Solutions

Can like-charged polyions ever attract each other? The answer to this
question is counterintuitive and has significant implications. It is a well
established fact that many bacteriophages use multivalent cations to
package their DNA into compact and ordered forms in vivo. This
phenomena is known as DNA condensation [166,167]. Moreover, it turns
out that DNA is not the only polyelectrolyte that is able to do so. Other
stiff polyelectrolytes such as F-actin [168], tobacco mosaic virus, and
the bacteriophage fd [169] are also able to form laterally ordered
aggregates (bundles) in the presence of multivalent counterions. Such a
variety of systems strongly suggests an electrostatic origin for the bundle
formation.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

301

Recent Brownian dynamics simulations of systems of two infinitely long

similarly charged rods in the presence of their counterions [170] have
supported the electrostatic origin of aggregation by showing that the attraction between similarly charged polyions is indeed mediated by counterions.
As the strength of the electrostatic interaction increases (or temperature of
the system decreases) fluctuations in the densities of condensed counterions
along the axes of the rods become strongly anticorrelated (lockkey
structure) leading to effective attraction between rods at small separations
[171173] (see [174] for a review).
The bundle formation in systems of stiff polyelectrolytes in the presence
of divalent counterions was simulated by Stevens [175,176]. Figure 20 shows
the ordering of stiff polyelectrolytes into a network of bundles. The number
of divalent counterions condensed on the chain is extremely large and
clearly drives the aggregation. The bundling is due to attraction between two
chains caused by correlated counterion fluctuations [171,172]. The precise

FIG. 20 Bundle configuration of polyelectrolyte chains with N 64 and number of

chains in simulation cell M 16. The figure shows four periodic images in order to
exhibit the connecting chains between bundles; there are only two bundles in the
simulation cell. The counterions are dark spheres and the monomers are light
spheres. (From Ref. [175].)

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

302

Dobrynin

morphology of the bundle depends on chain stiffness, length, and

concentration. The bundling occurs for stiff long chains, and when the
chain length decreases, the bundling decreases as well. The same happens
when the chain stiffness decreases. After the bundle is assembled the chains
can only reptate within the bundle, which leads to significant slowing down
of the relaxation of the system. It is therefore possible that within achievable
simulation times, this bundle network is metastable.

V. WHAT IS NEXT?
Our understanding of polyelectrolytes in solutions has significantly
progressed during the last ten years. The properties of polyelectrolyte
chains in dilute solution seem to be understood reasonably well. The only
problem that remains to be resolved here is the electrostatic persistence
length of flexible polyelectrolytes. Computer simulations of semidilute
polyelectrolyte solutions provide reasonable qualitative agreement with
existing experimental and theoretical works. We now better understand
counterion condensation phenomena and their relation to chain attraction
and clustering.
However, a lot of interesting work remains to be done as attention turns
to the dynamics of polyelectrolyte solutions. All current simulations of
polyelectrolytes neglect the hydrodynamic interactions [177]. The inclusion
of these interactions into molecular simulations will allow us to perform
simulations of realistic dilute and semidilute polyelectrolyte solutions. It
also will help us to answer the question of how hydrodynamic interactions
are really screened in polyelectrolyte solutions. The answer to this question
is extremely important for checking the assumptions of existing theoretical
models. Hydrodynamic interactions are also long range, thus the use of
Ewald summation [178180], PME, and P3M methods can significantly
speed up the simulations. Work in this direction has just begun.
Another area which has attracted significant interest over the last few
years is polyelectrolytes at surfaces and interfaces [8,181184]. The progress
in this area has implications for different areas of science and technology
such as colloids, biophysics, biochemistry, medical science, pharmacy, food
processing, water purification, etc. Some simulation work [185191] has
already appeared, suggesting a broad spectrum of interesting physical
phenomena.
Our list of future directions cannot be complete without mentioning
polyelectrolyte gels, polyelectrolyte brushes [192195] (polyelectrolytes
grafted to a surface), and complexation of polyelectrolytes with colloids
[196200], dendrimers [201], surfactants, and proteins. We anticipate serious

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

303

simulation studies of these systems in the near future that will have
tremendous practical applications.

APPENDIX
To perform direct evaluation of the lattice sum in Eq. (16) we first introduce
the integral representation of
1
2
p
jrj




dt exp
jrj2 t2



dt exp
jrj2 t2

53

With the help of this integral representation we can split the electrostatic
energy of the unit cell U ({rj}) into two parts
 
  
 
U rj Udir rj U2 rj
where we defined the direct sum
  

Z 1
 
X
2 
Udir rj
lB X
lim p
qk qj
exp
rkj n t2 dt
M!1 
kB T

n
k, j
 


X

X
erfc rk
rj n
lB


qk qj

rk
rj n
2
n

54

k, j

and a sum
  

Z 
 
X
2 
U 2 rj
lB X
p
qk qj
dt exp
rkj n t2
kB T
 k, j
0
n

55



p R 1
In Eq. (54) erfcx 2=  x dt exp
t2 is the complementary error function. The direct sum Udir({rj}) is a rapidly convergent series and we may take
the limit M ! 1 without further ado. The second lattice sum in (55)
requires more careful attention. This sum is diffcult to evaluate as it is, so
we use the identity
 
2 
exp
rkj n t2

1
3=2 t3

2

u
d u exp
2 2irkj  u 2in  u
t
3

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

56

304

Dobrynin

to reduce it to three finite geometric series. After substitution of the integral

(56) into Eq. (55), the integration over t may be performed immediately by
using the substitution s t
1. This leads to the following expression for the
lattice sum
U2 frj g
lB X
lim
qk qj
M!1 22
kB T
k, j
lB X 2
q

p
 j j


X
d3u
exp
u2 =2 2irkj  u
exp2in  u
2
u
n

In the last equation the second term in the r.h.s. is accounting for the selfenergy of the charges with n 0 and k j. The main contribution to the
sum comes from the regions where exp(2inu) 1 for which u G
m1b1 m2b2 m3b3 is a vector of the reciprocal lattice, b,  1, 2, 3 are
primitive vectors of the reciprocal lattice, and m1, m2, m3 are integer numbers. The vectors b have the property ba . [For the cubic lattice
these vectors are b1 (/L, 0, 0), b2 (0, /L, 0), and b3 (0, 0, /L).] This
suggests dividing the integration region over u into a sum of integrals over
the volume of a reciprocal lattice unit cell (the first Brillouin zone)
VG {
/2L < </2L, for  1, 2, 3}. To do so we make the substitution u G v where v 2 VG . This leads to
!
 
X
 

  
Urec rj
lB X
2
2
lim
qk qj
exp
G = 2irkj  G IG rkj I0 rkj
M!1 2
kB T
G60
k, j
58
where we have defined
 
1
IG rkj 2



X
d 3 vgG v, rkj
exp2in  v

VG

59

and




gG v, rij v G
2 exp
v2 =2
2v  G 2irij  v
The lattice sum in the integral IG(rkj) may be easily evaluated
X
n

exp2in  v

3
M
X
Y
1 n
M

exp2in L

3
Y
sin2M 1L
1

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

sinL

60

Molecular Simulations of Charged Polymers

305

For any G 6 0 the integral IG(rij) may be written as follows

Z
3
Y
 

 1
sin2M 1L
d 3v
IG rkj gG 0, rkj 2
 VG
sinL
1
Z
3
 


 Y
1
sin2M 1L
d 3 v gG v, rkj
gG 0, rkj
2
 VG
sinL
1

61

The first integral is equal to 3/L3 and the first term is equal to /L3G2. The
second integral may be estimated by integrating by parts with the use of
sin ((2M 1)L)
((2M 1)L)
1 d cos ((2M 1)L)/d. The leading
terms are equal to zero because cos ((2M 1)/2) 0. Integration by parts
again gives terms that are O(((2M 1)L)
2) and can be neglected in the limit
M ! 1. Collecting all terms together we obtain the final expression for the
sum in the reciprocal space with G 6 0
 


Urec rj
lB X exp
G2 =2
S GS
G
3
kB T
2L G60
G2


lB X exp
2 m2 =L2 2

SmS
m
m2
2L m60

62

where we defined the charge structure factor

SG

X
j



 X
m  rj 
qj exp i2Grj
qj exp i2
L
j

63

where vector m (m1, m2, m3) with all integers m 2
1, 1 for  1, 2, 3.
Now let us evaluate I0(rkj). After some algebra it can be written as follows


I0 rkj

 

 3
Z
exp 2irkj  v
1 Y
1
sin2M 1L
3
I0 0 2
d v
2
v
 VG
sinL
1

 2 2

Z
exp
v =
1
1
2
d 3v
 VG
v2
3
 

Y
sin2M 1L
 exp 2irij  v
1
sinL
1

64

The function I0(0) is independent of rkj and its contribution to the lattice
sum cancels for electroneutral systems. The first integral can be estimated by

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

306

Dobrynin

expanding exp (2irij  v) in powers of 2irkj v. All the odd powers in this series
do not contribute to the integral due to symmetry of the VG. Thus the
expansion starts with the term
2(rkj v)2. This is the only term in which
we are interested, because all other terms in the expansion are
O(((2M 1)L)
2) and can be omitted in the limit M ! 1. The symmetry
of the integral allows us to change (rkj v)2 v2 by r2kj =3 and to perform the
integration over v obtaining 3/L3. The second integral is finite and can be
evaluated by integrating by parts leading to an asymptotic expansion in
inverse powers of (2M 1)L. This expansion starts with the term of the
order of O(((2M 1)L)
2) and once again can be neglected in the limit
M ! 1. The final result for this part of the lattice sum is

2
 

Ushape rj
2lB X
2lB X

2

3
qk qj rkj
q
r
65

j j

kB T
3L k, j
3L3  j
.

Combining all terms together, we can write the final expression for the
electrostatic energy of the cell as a sum of four terms sums in the direct and
reciprocal space, self-energy term, and shape term that depends on the net
dipole moment of the cell
 

 

X
Uel rj
erfc rkj n
lB X



qk qj
rkj n
kB T
2 n k, j


lB X exp
2 m2 =L2 2
SmS
m

m2
2L m60

2

lB X 2 2lB X

qj
q
r
66

p


j
j

3L3  j
 j

REFERENCES
1. Polyelectrolytes; Hara, M., Ed.; Marcel Dekker: New York, 1993.
2. Tanford, C. Physical Chemistry of Macromolecules; Wiley: New York, 1961.
3. Oosawa, F. Polyelectrolytes; Marcel Dekker: New York, 1971.
4. Mandel, M. In Encyclopedia of Polymer Science and Engineering, Wiley:
New York, 1988.
5. Schmitz, K.S. Macroions in Solution and Colloidal Suspension; VCH: Weinheim,
1993.
6. Forster, S.; Schmidt, M. Adv. Polym. Sci. 1995, 120, 51133.
7. Barrat, J.L.; Joanny, J.F. Adv. Chem. Phys. 1996, XCIV, 166.
8. Colloid-Polymer Interactions: From Fundamentals to Practice; Dubin, P.L.,
Farinato, R.S., Eds.; Wiley-Interscience: New York, 1999.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

9.
10.

11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.

307

Physical Chemistry of Polyelectrolytes; Radeva, T., Ed.; Marcel Dekker:

New York, 2001.
Baschanagel, J.; Binder, K.; Doruker, P.; Gusev, A.A.; Hahn, O.; Kremer, K.;
Mattice, W.L.; Muller-Plathe, F.; Murat, M.; Paul, W.; Santos, S.; Suter, U.W.;
Tries, V. Adv. Polym. Sci. 2000, 152, 41156.
Application of the Monte Carlo methods in statistical physics, Topics in
Current Physics, Binder, K., Ed.; Springer-Verlag: New York, 1987; Vol. 36.
The Monte Carlo methods in condensed matter physics. Topics in Applied
Physics, Binder, K., Ed.; Springer-Verlag: New York, 1995; Vol. 71.
Binder, K.; Heermann, D.W. An introduction. Monte Carlo Simulation in
Statistical Physics, 3rd Ed.; Springer Verlag: Berlin, 1997.
Newman, M.E.J.; Barkema, G.T. Monte Carlo Methods in Statistical Physics;
Oxford University Press: New York, 1999.
Allen, M.P.; Tildesley, D.L. Computer Simulation of Liquids; Oxford University
Press: New York, 1989.
Frenkel, D.; Smit, B. Understanding Molecular Simulations: From Algorithms to
Applications; Academic Press: San Diego, 1996.
Rapaport, D.C. The Art of Molecular Dynamics Simulations; Cambridge
University Press: New York, 1995.
Wall, F.T.; Mandel, F. J. Chem. Phys. 1975, 63, 45924595.
Mandel, F. J. Chem. Phys. 1979, 70, 39843988.
Lal, L. Mol. Phys. 1969, 17, 57.
McDonald, B.; Jan, N.; Hunter, D.L.; Steinitz, M.O. J. Phys. A 1985, 18, 2627.
Madras, N.; Socal, A.D. J. Stat. Phys. 1988, 50, 109.
Kremer, K.; Grest, G. J. Chem. Phys. 1990, 92, 50575086.
Schneider, J.; Hess, W.; Klein, R. J. Phys. A 1985, 18, 1221.
Verdier, P.H.; Stockmayer, W.H. J. Chem. Phys. 1962, 36, 227.
de Gennes, P.G.; Pincus, P.; Velasco, R.M.; Brochard, F. J. Phys. (Paris) 1976,
37, 14611473.
Khokhlov, A.R.; Khachaturian, K.A. Polymer 1982, 23, 17421750.
Rubinstein, M.; Colby, R.H.; Dobrynin, A.V. Phys. Rev. Lett. 1994, 73,
27762780.
Dobrynin, A.V.; Colby, R.H.; Rubinstein, M. Macromolecules 1995, 28,
18591871.
de Gennes, P.G. Scaling Concepts in Polymer Physics; Cornell University Press:
Ithaca, 1979.
Yu Grosberg, A.; Khokhlov, A.R. Statistical Physics of Macromolecules; AIP
Press: New York, 1994.
Brender, C. J. Chem. Phys. 1990, 92, 44684472.
Brender, C. J. Chem. Phys. 1991, 94, 32133221.
Brender, C.; Danino, M. J. Chem. Phys. 1992, 97, 21192125.
Hooper, H.H.; Blanch, H.W.; Prausnitz, J.M. Macromolecules 1990, 23,
48204829.
Christos, G.A.; Carnie, S.L. J. Chem. Phys. 1989, 91, 439453.
Higgs, P.G.; Orland, H. J. Chem. Phys. 1991, 95, 45064518.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

308
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
59.
60.
61.
62.
63.
64.
65.
66.
67.
68.
69.
70.
71.

Dobrynin
Barrat, J.L.; Boyer, D. J. Phys. II (France) 1993, 3, 343356.
Ullner, M.; Jonsson, B.; Widmark, P.O. J. Chem. Phys. 1994, 100, 33653366.
Jonsson, B.; Petersen, C.; Soderberg, B. J. Phys. Chem. 1995, 99, 12511266.
Peterson, C.; Sommelius, O.; Soderberg, B. J. Chem. Phys. 1996, 105,
52335241.
Migliorini, G.; Rostiashvili, V.G.; Vilgis, T.A. Eur. Phys. J. E. 2001, 4,
475487.
Odijk, T. J. Polym. Sci. Polym. Phys. Ed. 1977, 15, 477483.
Skolnick, J.; Fixman, M. Macromolecules 1977, 10, 944948.
Li, H.; Witten, T.A. Macromolecules 1995, 28, 59215927.
Netz, R.R.; Orland, H. Eur. Phys. J. B. 1999, 8, 8198.
Ha, B.Y.; Thirumalai, D. J. Chem. Phys. 1999, 110, 75337541.
Tricot, M. Macromolecules 1984, 17, 16981703.
Reed, C.E.; Reed, W.F. J. Chem. Phys. 1992, 97, 77667776.
Forster, S.; Schmidt, M.; Antonietti, M. J. Phys. Chem. 1992, 96,
40084014.
deNooy, A.E.J.; Besemer, A.C.; van Bekkum, H.; van Dijk, J.; Smit, J.A.M.
Macromolecules 1996, 29, 65416547.
Nishida, K.; Urakawa, H.; Kaji, K.; Gabrys, B.; Higgins, J.S. Polymer 1997,
38, 60836085.
Tanahatoe, J.J. J. Phys. Chem. B 1997, 101, 1044210445.
Beer, M.; Schmidt, M.; Muthukumar, M. Macromolecules 1997, 30,
83758385.
Barrat, J.L.; Joanny, J.F. Europhys. Lett. 1993, 24, 333338.
Bratko, D.; Dawson, K.A. J. Chem. Phys. 1993, 99, 53525364.
Ha, B.Y.; Thirumalai, D. Macromolecules 1995, 28, 577581.
Reed, C.E.; Reed, W.F. J. Chem. Phys. 1991, 94, 84798486.
Seidel, C. Ber. Bunse-Ges. Phys. Chem. 1996, 100, 757763.
Schafer, H.; Seidel, C. Macromolecules 1997, 30, 66586661.
Ullner, M.; Jonsson, B.; Peterson, C.; Sommelius, O.; Soderberg, B. J. Chem.
Phys. 1997, 107, 12791287.
Micka, U.; Kremer, K. Phys. Rev. E 1996, 54, 26532662.
Micka, U.; Kremer, K. J. Phys. Cond. Matter 1996, 8, 94639470.
Micka, U.; Kremer, K. Europhys. Lett. 1997, 38, 279284.
Cannavacciuolo, L.; Sommer, C.; Pedersen, J.S.; Schurtenberger, P. Phys. Rev.
E 2000, 62, 54095419.
Lord Rayleigh. Philos. Mag. 1882, 14, 182.
Hooper, H.H.; Beltran, S.; Sassi, A.P.; Blanch, H.W.; Prausnitz, J.M. J. Chem.
Phys. 1990, 93, 27152723.
Dobrynin, A.V.; Rubinstein, M.; Obukhov, S.P. Macromolecules 1996, 29,
29742979.
Chodanowski, P.; Stoll, S. J. Chem. Phys. 1999, 111, 60696081.
Lyulin, A.V.; Dunweg, B.; Borisov, O.V.; Darinskii, A.A. Macromolecules
1999, 32, 32643278.
Solis, F.J.; Olvera de la Cruz, M. Macromolecules 1998, 31, 55025506.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

309

72. Balazs, A.C.; Singh, C.; Zhulina, E.; Pickett, G.; Chern, S.S.; Lyatskaya, Y.
Prog. Surface Sci. 1997, 55, 181269.
73. Edwards, S.F.; King, P.R.; Pincus, P. Ferroelectrics 1980, 30, 36.
74. Higgs, P.G.; Joanny, J-F. J. Chem. Phys. 1991, 94, 15431558.
75. Dobrynin, A.V.; Rubinstein, M. J. Phys. II (France) 1995, 5, 677695.
76. Gutin, A.; Shakhnovich, E. Phys. Rev. E 1994, 50, R3322R3325.
77. Quan, C.; Kholodenko, A. J. Chem. Phys. 1988, 89, 52735281.
78. Bratko, D.; Chakraborty, A.K. J. Phys. Chem. 1996, 100, 11641173.
79. McQuarrie, D.A. Statistical Mechanics; Harper & Row: New York, 1976;
Chapter 15.
80. Kantor, Y.; Kardar, M.; Li, H. Phys. Rev. Lett. 1992, 69, 6164.
81. Kantor, Y.; Kardar, M.; Li, H. Phys. Rev. E 1994, 49, 13831392.
82. Kantor, Y.; Kardar, M. Europhys. Lett. 1994, 27, 643648.
83. Kantor, Y.; Kardar, M. Phys. Rev. E 1995, 51, 12991312.
84. Kantor, Y.; Kardar, M. Phys. Rev. E 1995, 52, 835846.
85. Yamakov, V.; Milchev, A.; Limbach, H.J.; Dunweg, B.; Everaers, R. Phys.
Rev. Lett. 2000, 85, 43054308.
86. Tanaka, M.; Yu Grosberg, A.; Pande, V.S.; Tanaka, T. Phys. Rev. E 1997, 56,
57985808.
87. Soddemann, T.; Schiessel, H.; Bluman, A. Phys. Rev. E 1998, 57, 20812090.
88. Lee, N.; Obukhov, S.P. Eur. Phys. J. B 1998, 1, 371376.
89. Victor, J.M.; Imbert, J.B. Europhys. Lett. 1993, 24, 189195.
90. Imbert, J.B.; Victor, J.M.; Tsunekawa, N.; Hiwatary, Y. Phys. Lett. 1999, 258,
9298.
91. Wittmer, J.; Johner, A.; Joanny, J-F. Europhys. Lett. 1993, 24, 263269.
92. Pande, V.S.; Yu Grosberg, A.; Joerg, C.; Kardar, M.; Tanaka, T. Phys. Rev.
Lett. 1996, 77, 35653569.
93. Brender, C.; Lax, M.; Windwer, S. J. Chem. Phys. 1981, 74, 25762581.
94. Brender, C.; Lax, M. J. Chem. Phys. 1984, 80, 886892.
95. Christos, G.A.; Carnie, S.L. Chem. Phys. Lett. 1990, 172, 249253.
96. Woodward, C.; Jonsson, B. Chem. Phys. 1991, 155, 207213.
97. Valleau, J. Chem. Phys. 1989, 129, 163.
98. Ewald, P.P. Ann. Phys. 1921, 64, 253287.
99. Evjen, H.M. Phys. Rev. 1932, 39, 675.
100. Born, M.; Huang, K. Dynamical Theory of Crystal Lattices; Oxford University
Press: London, 1954.
101. Kittel, C. Introduction to Solid State Physics, 5th Ed.; John Wiley & Sons:
New York, 1976.
102. Toukmaji, A.Y.; Board, J.A. Jr. Comp. Phys. Comm. 1996, 95, 7392.
103. Sagui, C.; Dardin, T.A. Annu. Rev. Biophys. Biomol. Struct. 1999, 28,
155179.
104. Redlack, A.; Grinday, J. J. Phys. Chem. Solids 1975, 36, 7382.
105. De Leeuw, S.W.; Perram, J.W.; Smith, E.R. Proc. R. Soc. London 1980, A373,
2756.
106. Smith, E.R. Proc. R. Soc. London 1981, A375, 475505.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

310

Dobrynin

107. Deem, M.W.; Newsam, J.M.; Sinha, S.K. J. Phys. Chem. 1990, 94, 83568359.
108. Perram, J.W.; Petersen, H.G.; De Leeuw, S.W. Molecular Physics 1988, 65,
875893.
109. Dardin, T.; York, D.; Pedersen, L. J. Chem. Phys. 1993, 98, 1008910092.
110. Essmann, U.; Perera, L.; Berkowitz, M.L.; Dardin, T.; Lee, H.; Pedersen, L.G.
J. Chem. Phys. 1995, 103, 85778593.
111. Petersen, H.G. J. Chem. Phys. 1995, 103, 36683680.
112. Saqui, C.; Dardin, T.A. In Simulation and Theory of Electrostatic Interactions
in Solutions, AIP Conference Proceedings, Prat, L.P., Hammer, G., Eds.; Vol.
492, 104113.
113. Dardin, T.A.; Toukmaji, A.; Pedersen, L.G. J. Chim. Phys. 1997, 94,
13461364.
114. Kholmurodov, K.; Smith, W.; Yasuoka, K.; Dardin, T.; Ebisazaki, T.
J. Comp. Chem. 2000, 21, 11871191.
115. Crowley, M.F.; Dardin, T.A.; Cheatham, T.E.; Deereld, D.W. J.
Supercomputing 1997, 11, 255278.
116. Press, W.H.; Teukolsky, S.A.; Vetterling, W.T.; Flannery, B.P. Numerical
Recipies in C. The Art of Scientic Computing; 2nd Ed.; Cambridge
University Press: New York, 1992.
117. Chui, C.K. An Introduction to Wavelets; Academic Press: San Diego, 1992.
118. Schoenberg, I.J. Cardinal Spline Interpolation; SIAM: Philadelphia, 1973.
119. Hockney, R.W.; Eastwood, J.W. Computer Simulations Using Particles;
McGraw-Hill: New York, 1981.
120. Luty, B.A.; Davis, M.E.; Tirony, I.G.; van Gunsteren, W.F. Mol. Simul. 1994,
14, 1120.
121. Shimada, J.; Kaneko, H.; Takada, T. J. Comp. Chem. 1993, 14, 867878.
122. Ferrell, R.; Bertschinger, E. Int. J. Mod. Phys. C 1994, 5, 933956.
123. Pollock, E.L.; Glosli, J. Comp. Phys. Comm. 1996, 95, 93110.
124. Pollock, E.L. In Simulation and Theory of Electrostatic Interactions in
Solutions, AIP Conferences Proceedings, Prat, L.P., Hammer, G., Eds.;
Vol. 492, 146158.
125. Deserno, M.; Holm, C. J. Chem. Phys. 1998, 109, 76787693.
126. Deserno, M.; Holm, C. J. Chem. Phys. 1998, 109, 76947701.
127. Greengard, L.; Rokhlin, V. J. Comp. Phys. 1987, 73, 325348.
128. Greengard, L.; Rokhlin, V. Lecture Notes in Mathematics. 1988, 1360,
121141.
129. Greengard, L.; Rokhlin, V. Chem. Scrip. 1989, 29A, 139144.
130. Greengard, L. Science 1994, 265, 909914.
131. Jackson, J.D. Classical Electrodynamics; 2nd Ed.; Wiley: New York, 1975.
132. Esselink, K. Comp. Phys. Com. 1995, 87, 375395.
133. Carrier, J.; Greengard, L.; Rokhlin, V. SIAM J. Sci. Stat. Comput. 1988,
9, 669686.
134. Schmidt, K.E.; Lee, M.A. J. Stat. Phys. 1991, 63, 12231235.
135. Schmidt, K.E.; Lee, M.A. J. Stat. Phys. 1997, 89, 411424.
136. Manning, G.S. J. Chem. Phys. 1969, 51, 924933.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Molecular Simulations of Charged Polymers

137.
138.
139.
140.
141.
142.
143.
144.
145.
146.
147.
148.
149.
150.
151.
152.
153.
154.
155.
156.
157.
158.
159.
160.
161.
162.
163.
164.
165.
166.
167.
168.
169.
170.
171.
172.
173.

311

Khokhlov, A.R. J. Phys. A 1980, 13, 979987.

Raphael, E.; Joanny, J.F. Europhys. Lett. 1990, 13, 623628.
Schiessel, H.; Pincus, P. Macromolecules 1998, 31, 79537959.
Schiessel, H. Macromolecules 1999, 32, 56735680.
Muthukumar, M. J. Chem. Phys. 1996, 105, 51835199.
Muthukumar, M. J. Chem. Phys. 1997, 107, 26192635.
Donley, J.P.; Rudnik, J.; Liu, A.J. Macromolecules 1997, 30, 11881193.
Yethiraj, A. Phys. Rev. Lett. 1998, 78, 37893792.
Yethiraj, A. J. Chem. Phys. 1998, 108, 11841192.
Stevens, M.J.; Kremer, K. Phys. Rev. Lett. 1993, 71, 22282231.
Stevens, M.J.; Kremer, K. Macromolecules 1993, 26, 47174719.
Stevens, M.J.; Kremer, K. J. Chem. Phys. 1995, 103, 16691690.
Liao, Q.; Dobrynin, A.V.; Rubinstein, M. Macromolecules 2003, 36,
33863398; 33993410.
Adams, D.; Dubey, G. J. Comp. Phys. 1987, 72, 156168.
Stevens, M.J.; Plimpton, S.J. Eur. Phys. J. B 1998, 2, 341345.
Stevens, M.J.; Kremer, K. J. Phys. II France 1996, 6, 16071613.
Micka, U.; Holm, C.; Kremer, K. Langmuir 1999, 15, 40334044.
Micka, U.; Kremer, K. Europhys. Lett. 2000, 49, 189195.
Liao, Q.; Dobrynin, A.V.; Rubinstein, M. Macromolecules 2004 (to be
published).
Dobrynin, A.V.; Rubinstein, M. Macromolecules 2001, 34, 19641972.
Limbach, H.J.; Holm, C. J. Chem. Phys. 2001, 114, 96749682.
Castelnovo, M.; Sens, P.; Joanny, J-F. Eur. Phys. J. E 2000, 1, 115125.
Winkler, R.G.; Gold, R.G.; Reineker, P. Phys. Rev. Lett. 1998, 80, 37313734.
Brilliantov, N.V.; Kuznetsov, D.V.; Klein, R. Phys. Rev. Lett. 1998, 81,
14331436.
Fuoss, R.M.; Katchalsky, A.; Lifson, S. Proc. Natl. Acad. Sci. USA. 1951, 37,
579.
Alfrey, T.; Berg, P.; Morawetz, H.J. J. Polym. Sci. 1951, 7, 543547.
Das, T.; Bratko, D.; Bhuiyan, L.B.; Outhwaite, C.W. J. Phys. Chem. 1995, 99,
410.
Deserno, M.; Holm, C.; May, S. Macromolecules 2000, 33, 199206.
Deserno, M.; Holm, C.; Kremer, K. In Physical Chemistry of Polyelectrolytes,
Radeva, T., Ed.; Marcel Dekker: New York, 2001; 59110.
Bloomeld, V.A. Biopolymers 1991, 31, 14711481.
Bloomeld, V.A. Current Opinion in Structural Biology 1996, 6, 334341.
Tang, J.X.; Janmey, P.A. J. Biol. Chem. 1996, 271, 85568563.
Tang, J.X.; Wong, S.E.; Tran, P.T.; Janmey, P.A. Ber. der. Buns. Ges. Phys.
Chem. Chem. Phys. 1996, 100, 796806.
Gronbech-Jensen, N.; Mashl, R.J.; Bruinsma, R.F.; Gelbart, W.M. Phys. Rev.
Lett. 1997, 78, 24772480.
Shklovskii, B.I. Phys. Rev. Lett. 1999, 82, 32683271.
Solis, F.J.; de la Cruz, M.O. Phys. Rev. E 1999, 60, 44964499.
Levin, Y.; Arenzon, J.J.; Stilck, J.F. Phys. Rev. Lett. 1999, 83, 2680.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

312

Dobrynin

174. Ha, B.Y.; Liu, A.J. In Physical Chemistry of Polyelectrolytes; Radeva, T., Ed.;
Marcel Dekker: New York, 2001; 163181.
175. Stevens, M.J. Phys. Rev. Lett. 1999, 82, 101103.
176. Stevens, M.J. Biophys. J. 2001, 80, 130139.
177. Doi, M.; Edwards, S.F. The Theory of Polymer Dynamics; Oxford University
Press: New York, 1986.
178. Beenakker, C.W.J. J. Chem. Phys. 1986, 85, 1581.
179. Brady, J.F.; Phillips, R.J.; Lester, J.C.; Bossis, G. J. Fluid Mech. 1988, 195,
257.
180. Hase, K.R.; Powel, R.L. J. Phys. of Fluid 2001, 13, 3244.
181. Fleer, G.J.; Cohen Stuart, M.A.; Scheutjens, J.M.H.M.; Gasgove, T.; Vincent,
B. Polymers at Interfaces; Chapman and Hall: London, 1993.
182. Bajpai, A.K. Prog. Polym. Sci. 1997, 22, 523564.
183. Kawaguchi, M.; Takahashi, A. Adv. Coll. Inter. Sci. 1992, 37, 219317.
184. Decher, G.; Eckle, M.; Schmitt, J.; Struth, B. Curr. Opin. in Coll. Inter. Sci.
1998, 3, 3240.
185. Beltran, S.; Hooper, H.H.; Blanch, H.W.; Prausnitz, J.M. Macromolecules
1991, 24, 31783184.
186. Muthukumar, M. J. Chem. Phys. 1995, 103, 47234731.
187. Mashl, R.J.; Gronbech-Jensen, N. J. Chem. Phys. 1998, 109, 46174623.
188. Kong, C.Y.; Muthukumar, M. J. Chem. Phys. 1998, 109, 15221527.
189. Mashl, R.J.; Gronbech-Jensen, N.; Fitzsimmons, M.R.; Lutt, M.; Li, D.Q.
J. Chem. Phys. 1999, 110, 22192295.
190. Yamakov, V.; Milchev, A.; Borisov, O.; Dunweg, B. J. Phys. Cond. Matt.
1999, 11, 99079923.
191. Ellis, M.; Kong, C.Y.; Muthukumar, M. J. Chem. Phys. 2000, 112, 87238729.
192. Csajka, F.S.; van der Linden, C.C.; Seidel, C. Macromol. Symp. 1999, 146,
243249.
193. Csajka, F.S.; Seidel, C. Macromolecules 2000, 33, 27282739.
194. Ennis, J.; Sjostrom, L.; Akesson, T.; Jonsson, B. Langmuir 2000, 16,
71167125.
195. Csajka, F.S.; Netz, R.R.; Seidel, C.; Joanny, J.F. Eur. Phys. J. E 2001, 4,
505513.
196. von Goeler, F. J. Chem. Phys. 1994, 100, 77967803.
197. Wallin, T.; Linse, P. J. Chem. Phys. 1998, 109, 50895100.
198. Chodanowski, P.; Stoll, S. Macromolecules 2001, 34, 23202328.
199. Mateescu, E.M.; Jeppesen, C.; Pincus, P. Eur. Phys. Lett. 1999, 46, 493498.
200. Kunze, K.K.; Netz, R.R. Phys. Rev. Lett. 2000, 85, 43984392.
201. Welch, P.; Muthukumar, M. Macromolecules 2000, 33, 61596167.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

9
Gibbs Ensemble and Histogram
Reweighting Grand Canonical
Monte Carlo Methods
ATHANASSIOS Z. PANAGIOTOPOULOS
Princeton, New Jersey, U.S.A.

I.

Princeton University,

INTRODUCTION

The phase behavior of fluids is of central importance to many technological

and scientific fields, for example in designing separations for the chemical
and pharmaceutical industries or in understanding fundamental processes in
living systems. A large body of experimental information has been gathered
over the years (e.g., see [1]), and significant efforts have been made to
understand the phenomenology of the transitions and to obtain empiricaland theoretical-based models that can be used to correlate and extend
the range of experimental data. Experimental measurements are timeconsuming and expensive. For multicomponent mixtures, measurements are
available only for a limited number of temperatures, pressures, and
compositions. Empirical models are only valid over the range of conditions
for which experimental data have been used to obtain the model parameters.
Even theoretical-based models have limited predictive abilities for conditions and systems different from the ones for which they have been tested
against using experimental data [2].
Molecular-based simulations are an increasingly important alternative
to experimental measurements and theoretical techniques for obtaining properties of fluids and materials. The focus of the present chapter
is on simulations of phase equilibrium properties of fluids. Classical
force-field-based simulations start by postulating a functional form for the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

314

Panagiotopoulos

intermolecular forces in a system. Equilibrium properties can generally be

obtained by either Monte Carlo or molecular dynamics methods. Monte
Carlo methods are based on generating configurations from the appropriate
probability distribution for a statistical mechanical ensemble, while
molecular dynamics methods generate configurations by solving Newtons
equations of motion. Calculations by simulation of simple structural and
energetic properties (such as pair correlation functions, the mean configurational energy, or pressure) are relatively straightforward, but the
prediction of the order and precise location of phase transitions is not a
simple matter. Phase transitions are collective phenomena that occur over
time and length scales that are not directly accessible by molecular dynamics
or simple constant-volume Monte Carlo simulations. Until the mid-1980s,
obtaining the phase behavior of even a simple one-component system
required a major research effort [3]. Methodological developments since
then have rendered the determination of phase equilibria by simulation
much easier than before. Most of these methodological advances have
involved development of novel Monte Carlo algorithms, which are the focus
of the present review. In addition, the sustained increases in computing
hardware capabilities have greatly expanded the range of systems that can
be studied on readily available machines. As a result, the number of
simulation studies of both model potentials and realistic systems has
dramatically increased.
A number of textbooks, research monographs, and review articles have
appeared previously in the area of the present chapter. The book by Allen
and Tildesley [4] on computer simulation methods for liquids provides an
excellent introduction to molecular dynamics and Monte Carlo methods,
but does not cover the major recent methodological advances, since it
was published in 1987. The recent book by Frenkel and Smit [5] has
comprehensive coverage of molecular simulation methods for fluids, with
particular emphasis on algorithms for phase equilibrium calculations. It
describes many of the techniques mentioned in the present chapter in
significantly more detail than is possible here. The Gibbs ensemble method
and its applications have been reviewed in [69]. Proceedings of a
workshop on simulations of phase transitions [10] and general review
articles on simulation methods and their applications (e.g., [1113]) are also
available. The present chapter follows closely an earlier review article by the
author [14].
Knowledge of the chemical potential of all components (or the free
energy) of a system as a function of temperature, density, and composition
is, of course, sufficient to determine the phase behavior. Methods to
obtain the chemical potential include thermodynamic integration, a very
general technique in which the state of interest is linked via a reversible

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Monte Carlo Methods

315

path to a state of known free energy [5], and the Widom test particle
insertion method [15]. The present chapter focuses on methods that were
specifically designed for phase equilibrium calculations. The relative
precision and accuracy of methods to obtain the chemical potential have
been examined in [16]. Applications of direct interfacial simulations, which
can be performed by either Monte Carlo or molecular dynamics
algorithms, have been reviewed by Rowlinson and Widom [17] and
Gubbins [18].
The plan of this chapter is as follows. Section II deals with the Gibbs
ensemble Monte Carlo method, which is based on simultaneous
calculations in two regions representing equilibrium phases, coupled
indirectly via particle transfers and volume changes. The method is now
commonly used for obtaining phase equilibria of fluids, because of its
simplicity and speed. A single Gibbs ensemble simulation gives a point on
the phase envelope of a multicomponent system. A number of other
methods designed for direct calculations of phase equilibria are described
in Section III. The NPT test particle method (Section III.A) is based
on chemical potential calculations. The method has roughly the same
range of applicability and limitations as the Gibbs ensemble, but requires
multiple simulations per coexistence point. GibbsDuhem integration
(Section III.B) does not require particle insertions and removals and is
applicable to transitions involving solids. It needs to start, however,
from a point on the phase envelope determined by one of the other
techniques. Pseudo-ensembles (Section III.C) provide significant flexibility
in determinations of phase equilibria under different external constraints
and can be implemented in combination with the Gibbs ensemble or
GibbsDuhem integrations. Histogram reweighting methods (Section IV)
provide the free energy and phase behavior with excellent accuracy and
can be used in the vicinity of critical points. The majority of simulation
methods for calculations of phase transitions rely on particle transfers,
which become impractical for dense systems or multisegment molecules. A
number of methods have been developed for improving the efficiency of
particle transfers and have been instrumental in enabling calculations for
realistic potential models. Configurational-bias sampling techniques that
perform smart insertions at favorable locations are described in Section
V.A. Expanded ensembles are based on gradual transfers of parts of
molecules and are described in Section V.B. The last part of this review
(Section VI) describes applications of simulations to calculations of the
phase behavior of polymeric systems. The chapter concludes with a
discussion of the relative strengths and weaknesses of the methods
discussed and provides some suggestions for possible future research
directions in the field.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

316

Panagiotopoulos

II. GIBBS ENSEMBLE MONTE CARLO

The Gibbs Ensemble Monte Carlo simulation methodology [1921] enables
direct simulations of phase equilibria in fluids. A schematic diagram of the
technique is shown in Fig. 1. Let us consider a macroscopic system with two
phases coexisting at equilibrium. Gibbs ensemble simulations are performed
in two separate microscopic regions, each within periodic boundary
conditions (denoted by the dashed lines in Fig. 1). The thermodynamic
requirements for phase coexistence are that each region should be in internal
equilibrium, and that temperature, pressure, and the chemical potentials of
all components should be the same in the two regions. System temperature
in Monte Carlo simulations is specified in advance. The remaining three
conditions are satisfied by performing three types of Monte Carlo moves
displacements of particles within each region (to satisfy internal equilibrium), fluctuations in the volume of the two regions (to satisfy equality of
pressures), and transfers of particles between regions (to satisfy equality of
chemical potentials of all components).
The acceptance criteria for the Gibbs ensemble were originally derived
from fluctuation theory [19]. An approximation was implicitly made in the
derivation that resulted in a difference in the acceptance criterion for
particle transfers proportional to 1/N relative to the exact expressions given
subsequently [20]. A full development of the statistical mechanics of the

FIG. 1 Schematic diagram of the Gibbs ensemble Monte Carlo simulation

methodology.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Monte Carlo Methods

317

ensemble was given by Smit et al. [21] and Smit and Frenkel [22], which we
follow here. A one-component system at constant temperature T, total
volume V, and total number of particles N is divided into two regions, with
volumes VI and VII V
VI, and number of particles NI and NII N
NI.
The probability density of states, }(NI,VI; N,V,T ), is proportional to
}NI , VI ; N, V, T
N!
/
expNI ln VI NII ln VII
UI NI
UII NII
NI !NII !

Smit et al. [21] used a free energy minimization procedure to show that
for a system with a first-order phase transition, the two regions in a Gibbs
ensemble simulation are expected to reach the correct equilibrium densities.
The acceptance criteria for the three types of moves can be immediately
obtained from Eq. (1). For a displacement step internal to one of the
regions, the probability of acceptance is the same as for conventional
constant-NVT simulations,
}move min1, exp
U

where U is the configurational energy change resulting from the displacement. For a volume change step during which the volume of region I is
increased by V with a corresponding decrease of the volume of region II,
"

VI V
VII
V 
NII ln
}volume min 1, exp
UI
UII NI ln
VI
VII


3
Equation (3) implies that sampling is performed uniformly in the volume
itself. The acceptance criterion for particle transfers, written here for
transfer from region II to region I is
"

}transfer

#
NII VI
min 1,
exp
UI
UII
NI 1VII

Equation (4) can be readily generalized to multicomponent systems. The

only difference is that the numbers of particles of species j in each of the
two regions, NI,j and NII,j, replace NI and NII respectively. In simulations of
multicomponent systems dilute in one component, it is possible that the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

318

Panagiotopoulos

number of particles of a species in one of the two regions becomes zero after
a successful transfer out of that region. Equation (4) in this case is taken to
imply that the probability of transfer out of an empty region is zero.
The acceptance rules to this point are for a simulation in which the total
system is at constant number of molecules, temperature, and volume. For
pure component systems, the phase rule requires that only one intensive
variable (in this case system temperature) can be independently specified
when two phases coexist. The vapor pressure is obtained from the simulation. By contrast, for multicomponent systems pressure can be specified
in advance, with the total system being considered at constant NPT. The
probability density for this case, }(NI, VI; N, P, T ) is proportional to
}NI ,VI ;N,P,T
h
i
N!
exp NI lnVI NII lnVII
UI NI
UII NII
PVI VII
/
NI !NII !
5
and the only change necessary in the algorithm is that the volume changes in
the two regions are now made independently. The acceptance criterion
for a volume change step in which the volume of region I is changed by
V, while the other region remains unchanged is then
"
#


VI V
}volume min 1, exp
UI NI ln

PV
6
VI
An interesting extension of the original methodology was proposed by
Lopes and Tildesley [23], to allow the study of more than two phases at
equilibrium. The extension is based on setting up a simulation with as many
boxes as the maximum number of phases expected to be present. Kristof and
Liszi [24,25] have proposed an implementation of the Gibbs ensemble in
which the total enthalpy, pressure, and number of particles in the total
system are kept constant. Molecular dynamics versions of the Gibbs
ensemble algorithm are also available [2628], and also in the following
Chapter 10 by M. Kotelyanskii and R. Hentschke.
The physical reason for the ability of the Gibbs ensemble to converge to a
state that contains phases at their equilibrium density in the corresponding
boxes, rather than a mixture of the two phases in each box, is the free energy
cost for creating and maintaining an interface. Essentially, in the Gibbs
ensemble, one trades off the directness of the interfacial simulation
approach with the (slight) inconvenience of setting up and maintaining
two or more boxes for the equilibrium phases. However, much smaller
system sizes can be used relative to interfacial simulations, and the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Monte Carlo Methods

319

simulations are generally stable, except at the immediate vicinity of critical

points.
Near critical points, Gibbs ensemble simulations become unstable
because the free energy penalty for creating an interface becomes small.
In a detailed study of the behavior of Gibbs ensemble simulations near
critical points, Valleau [29] concluded that it is only with extreme care . . .
that reliable information on critical parameters or the shapes of coxistence
curves may be obtained from Gibbs ensemble simulations. The cause of the
problems is that near critical points finite-size effects are present, and there
is no mechanism for controlling system size of each individual region in the
Gibbs ensemble. A better approach for dealing with systems near critical
points is provided by the histogram methods described in Section IV. The
finite-size critical behavior of the Gibbs ensemble has been examined by
Bruce [30], Mon and Binder [31], and Panagiotopoulos [32]. The standard
procedure for obtaining critical points from Gibbs ensemble simulations is
to fit subcritical coexistence data to universal scaling laws. This approach
has a weak theoretical foundation, since the universal scaling laws are only
guaranteed to be valid in the immediate vicinity of the critical point, where
simulations give the wrong (classical) behavior due to the truncation of the
correlation length at the edge of the simulation box. In many cases,
however, the resulting critical points are in reasonable agreement with more
accurate results obtained from finite-size scaling methods (Section IV.C).
In summary, the Gibbs ensemble Monte Carlo methodology provides a
direct and efficient route to the phase coexistence properties of fluids,
for calculations of moderate accuracy. The method has become a standard
tool for the simulation community, as evidenced by the large number of
applications. Histogram reweighting techniques (Section IV) have the
potential for higher accuracy, especially if equilibria at a large number of
state points are to be determined. Histogram methods are also inherently
better at determining critical points. In its original form, the Gibbs ensemble
method is not practical for multisegment or strongly interacting systems,
but development of configurational-bias sampling methods described in
Section V.A has overcome this limitation.

III. THE NPT 1 TEST PARTICLE METHOD,

GIBBSDUHEM INTEGRATION AND
PSEUDO-ENSEMBLES
A. The NPT 1Test Particle Method
The NPT test particle method [33,34] is based on calculations of the
chemical potentials for a number of state points. A phase coexistence point

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

320

Panagiotopoulos

is determined at the intersection of the vapor and liquid branches of the

chemical potential vs. pressure diagram. The Widom test particle method
[15] or any other suitable method [16] can be used to obtain the chemical
potentials. Corrections to the chemical potential of the liquid and vapor
phases can be made, using standard thermodynamic relationships, for
deviations between the pressure at which the calculations were made and the
actual coexistence pressure. Extrapolations with respect to temperature are
also possible [35].
In contrast to the Gibbs ensemble, a number of simulations are required
per coexistence point, but the number can be quite small, especially for
vaporliquid equilibrium calculations away from the critical point. For
example, for a one-component system near the triple point, the density of
the dense liquid can be obtained from a single NPT simulation at zero
pressure. The chemical potential of the liquid, in turn, determines the
density of the (near-ideal) vapor phase so that only one simulation is
required. The method has been extended to mixtures [36,37]. Significantly
lower statistical uncertainties were obtained in [37] compared to earlier
Gibbs ensemble calculations of the same Lennard-Jones binary mixtures,
but the NPT test particle method calculations were based on longer
simulations.
The NPT test particle method shares many characteristics with the
histogram reweighting methods discussed in Section IV. In particular,
histogram reweighting methods also obtain the chemical potentials and
pressures of the coexisting phase from a series of simulations. The
corrections to the chemical potentials for changes in pressure [34] and
temperature [35] are similar to the concept of reweighting of combined
histograms from grand canonical simulations to new densities and
temperatures.
Spyriouni et al. [38,39] have presented a powerful method (called
SPECS) for calculations of polymer phase behavior related to the
NPT test particle method. The method of Spyriouni et al. targets the
calculation of the phase behavior of long-chain systems for which the test
particle method for calculation of chemical potentials fails. For sufficiently
long chains, even configurational-bias sampling methods discussed in
Section V.A become impractical. For binary mixtures of a low-molecular
weight solvent (species 1) and a polymer (species 2), two parallel simulations
are performed in the ( 1,N2,P,T ) ensemble at conditions near the expected
coexistence curve. The chemical potential of component 2 is determined
through the chain increment technique [40]. Iterative calculations at
corrected values of the chemical potential of the solvent are performed until
the chemical potential of the polymer in the two phases is equal. For the
special case of a dilute solution, estimates of the chemical potentials of the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Monte Carlo Methods

321

solvent and polymer for compositions different from the original simulation
conditions can be made using standard thermodynamic relations and the
number of required iterations is significantly reduced.

B. GibbsDuhem Integration
Most methods for determination of phase equilibria by simulation rely on
particle insertions to equilibrate or determine the chemical potentials of the
components. Methods that rely on insertions experience severe difficulties
for dense or highly structured phases. If a point on the coexistence curve is
known (e.g., from Gibbs ensemble simulations), the remarkable method of
Kofke [41,42] enables the calculation of a complete phase diagram from a
series of constant-pressure simulations that do not involve any transfers of
particles. For one-component systems, the method is based on integration
of the ClausiusClapeyron equation over temperature,
 dP 
d

sat

H
V

where sat indicates that the equation holds on the saturation line, and H
is the difference in enthalpy between the two coexisting phases. The right
hand side of Eq. (7) involves only mechanical quantities that can be simply
determined in the course of a standard Monte Carlo or molecular dynamics
simulation. From the known point on the coexistence curve, a change in
temperature is chosen, and the saturation pressure at the new temperature
is predicted from Eq. (7). Two independent simulations for the corresponding phases are performed at the new temperature, with gradual changes of the
pressure as the simulations proceed to take into account the enthalpies and
densities at the new temperature as they are being calculated.
Questions related to propagation of errors and numerical stability of the
method have been addressed in [42] and [43]. Errors in initial conditions
resulting from uncertainties in the coexistence densities can propagate and
increase with distance from the starting point when the integration path is
towards the critical point [43]. Near critical points, the method suffers from
instability of a different nature. Because of the small free energy barrier
for conversion of one phase into the other, even if the coexistence pressure is
set properly, the identity of each phase is hard to maintain and large
fluctuations in density are likely. The solution to this last problem is to
borrow an idea from the Gibbs ensemble and couple the volume changes of
the two regions [42]. Extensions of the method to calculations of three-phase
coexistence lines are presented in [44] and to multicomponent systems in [43].
Unfortunately, for multicomponent systems the GibbsDuhem integration

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

322

Panagiotopoulos

method cannot avoid particle transfershowever, it avoids transfers for one

component, typically the one that is the hardest to transfer. The method and
its applications have been recently reviewed [45].
In some cases, in particular lattice and polymeric systems, volume change
moves may be hard to perform, but particle insertions and deletions may
be relatively easy, especially when using configurational-bias methods.
Escobedo and de Pablo [46,47] proposed a modification of the Gibbs
Duhem approach that is based on the expression
 d 
d

sat

u
,


where  is the density (N/V) and u the energy per particle. This method
was applied to continuous-phase polymeric systems in [46] and to lattice
models in [50].
The GibbsDuhem integration method excels in calculations of solid
fluid coexistence [48,49], for which other methods described in this chapter
are not applicable. An extension of the method that assumes that the initial
free energy difference between the two phases is known in advance, rather
than requiring it to be zero, has been proposed by Meijer and El Azhar [51].
The procedure has been used in [51] to determine the coexistence lines
of a hard-core Yukawa model for charge-stabilized colloids.

C. Pseudo-Ensembles
The GibbsDuhem integration method represents a succesful combination
of numerical methods and molecular simulations. Taking this concept
even further, Mehta and Kofke [52] proposed a pseudo-grand canonical
ensemble method in which a system maintains a constant number of
particles and temperature, but has a fluctuating volume to ensure that, at
the final density, the imposed value of the chemical potential is matched.
The formalism still requires that estimates of the chemical potential be made
during the simulation. The main advantage of the approach over more
traditional grand canonical ensemble methods is that it provides additional
flexibility with respect to the method to be used for determination of
the chemical potential. For example, the chain increment method [40]
for chain molecules, which cannot be combined with grand canonical
simulations, can be used for the chemical potential evaluations in a pseudogrand canonical simulation (as in [38]).
The same pseudo-ensemble concept has been used by Camp and Allen
[53] to obtain a pseudoGibbs method in which particle transfers are
substituted by volume fluctuations of the two phases. The volume

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Monte Carlo Methods

323

fluctuations are unrelated to the ones required for pressure equality

[Eq. (3)] but are instead designed to correct imbalances in the chemical
potentials of some of the components detected, for example, by test particle
insertions.
While the main driving force in [52] and [53] was to avoid direct particle
transfers, Escobedo and de Pablo [47] designed a pseudo-NPT method to
avoid direct volume fluctuations which may be inefficient for polymeric
systems, especially on lattices. Escobedo [54] extended the concept for
bubble-point and dew-point calculations in a pseudo-Gibbs method and
proposed extensions of the GibbsDuhem integration techniques for tracing
coexistence lines in multicomponent systems [55].

IV. HISTOGRAM REWEIGHTING GRAND

CANONICAL MONTE CARLO
Early in the history of development of simulation methods it was realized
that a single calculation can, in principle, be used to obtain information on
the properties of a system for a range of state conditions [5658]. However,
the practical application of this concept was severely limited by the
performance of computers available at the time. In more recent years,
several groups have confirmed the usefulness of this concept, first in the
context of simulations of spin systems [5961] and later for continuousspace fluids [62,63,6567]. In the following subsections, we give a
pedagogical review of histogram reweighting methods for grand canonical
Monte Carlo (GCMC) simulations as applied to one- and multicomponent
systems. In addition, the determination of critical parameters from
histogram data is briefly reviewed.

A. One-Component Systems
A GCMC simulation for a one-component system is performed as follows.
The simulation cell has a fixed volume V, and is placed under periodic
boundary conditions. The inverse temperature,  1/kBT and the chemical
potential, , are specified as input parameters to the simulation. Histogram
reweighting requires collection of data for the probability f(N,E) of
occurrence of N particles in the simulation cell with total configurational
energy in the vicinity of E. This probability distribution function follows the
relationship
f N, E

N, V, Eexp
E  N
, V, 

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

324

Panagiotopoulos

where  (N,V,E) is the microcanonical partition function (density of states)

and ( ,V,) is the grand partition function. Neither  nor are known at
this stage, but is a constant for a run at given conditions. Since the left
hand side of Eq. (9) can be easily measured in a simulation, an estimate for
 and its corresponding thermodynamic function, the entropy S(N,V,E),
can be obtained by a simple transformation of Eq. (9):
SN, V, E=kB ln N, V, E ln f N, E E
 N C

10

C is a run-specific constant. Equation (10) is meaningful only over the range

of densities and energies covered in a simulation. If two runs at different
chemical potentials and temperatures have a region of overlap in the space
of (N,E) sampled, then the entropy functions can be merged by requiring
that the functions are identical in the region of overlap. To illustrate this
concept, we make a one-dimensional projection of Eq. (9) to obtain
f N

QN, V, exp N
, V, 

11

Histograms for two runs at different chemical potentials are presented in

Fig. 2. There is a range of N over which the two runs overlap. Figure 3
shows the function lnf(N)
 N for the data of Fig. 2. From elementary
statistical mechanics, this function is related to the Helmholtz energy,
AN, V, 
ln QN, V,  ln f N
 N C

12

FIG. 2 Schematic diagram of the probability f(N) of occurrence of N particles for

two GCMC runs of a pure component system at the same volume V and temperature
T, but different chemical potentials, 1 and 2, respectively. (From [14], 2000 IOP
Publishing Ltd.)

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Monte Carlo Methods

325

FIG. 3 The function ln [f(N)]
 N for the data of Fig. 2. The figure shows the raw
curves for 1 and 2 as well as a composite curve formed by shifting the data
by the amount indicated by the arrows. (From [14], 2000 IOP Publishing Ltd.)

Figure 3 shows the raw curves for 1 and 2 as well as a composite curve
formed by shifting data for the two runs by the amount indicated by the
arrows. The combined curve provides information over the combined range
of particle numbers, N, covered by the two runs. Note that by keeping onedimensional histograms for N we are restricted to combining runs of the
same temparature, while the more general form [Eq. (10)] allows combination of runs at different temperatures.
Simulation data are subject to statistical (sampling) uncertainties, which
are particularly pronounced near the extremes of particle numbers and
energies visited during a run. When data from multiple runs are combined as
shown in Fig. 3, the question arises of how to determine the optimal amount
by which to shift the raw data in order to obtain a global free energy
function. Ferrenberg and Swendsen [68] provided a solution to this problem
by minimizing the differences between predicted and observed histograms.
In this approach, it is assumed that multiple overlapping runs, i 1, 2, . . . , R
are available for a given system. The composite probability, }(N, E; , ), of
observing N particles and energy E, if one takes into account all runs and
assumes that they have the same statistical efficiency, is
XR

fi N, E exp
E  N
}N, E; ,  XRi1
K exp
i E i i N
Ci
i1 i

13

P
where Ki is the total number of observations [Ki N,E fi(N, E )] for run i.
The constants Ci (also known as weights) are obtained by iteration

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

326

Panagiotopoulos

from the relationship

expC i

XX
E

}N, E; i , i

14

Given an initial guess for the set of weights Ci, Eqs. (13) and (14) can be
iterated until convergence. When many histograms are to be combined this
convergence of the FerrenbergSwendsen weights can take a long time.
Once this has been achieved, however, all thermodynamic quantities for the
system over the range of densities and energies covered by the histograms
can be obtained. For example, the mean configurational energy U( ,) is
hUi , 

XX
E

}N, E; ,   E

15

and the mean density ( ,)

hi , 

1 XX
}N, E; ,   N
V E N

16

The pressure of a system can be obtained from the following expression.

If the conditions for run 1 are ( 1, V, 1) and for run 2 ( 2, V, 2), then
C2
C1 ln

2 , V, 2
2 P2 V
1 P1 V
1 , V, 1

17

where P is the pressure, since ln PV. Equation (17) can be used to

obtain the absolute value of the pressure for one of the two runs, provided
that the absolute pressure can be estimated for the other run. Typically, this
is done by performing simulations for low-density states for which the
system follows the ideal-gas equation of state, PV NkBT.
Up to this point, we assumed that a system exists in a one-phase region
over the range of densities and energies sampled. If a phase transition exists,
then the system, in principle, should sample states on either side of the phase
transition, resulting in histograms with multiple peaks. This is illustrated in
Fig. 4, in which actual simulation data (from a single run) are plotted for
a simple cubic lattice homopolymer system [63] at a slightly subcritical
temperature. There are two states sampled by the run, one at low and one
at high particle numbers, corresponding to the gas and liquid states. The
conditions for phase coexistence are equality of temperature, chemical
potential, and pressurethe first two are satisfied by construction. From

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Monte Carlo Methods

327

FIG. 4 Frequency of observation of states vs. energy, E, and number of particles,

N, for a homopolymer of chain length r 8 and coordination number z 6 on a
10  10  10 simple cubic lattice. Conditions, following the notation of [63] are
T *=11.5, *
60.4. In order to reduce clutter, data are plotted only for every
third particle. (From [14], 2000 IOP Publishing Ltd.)

Eq. (17), the integral under the probability distribution function is

proportional to the pressure. In the case of two distinct phases, the integrals
should be calculated separately under the liquid and gas peaks. The
condition of equality of pressures can be satisfied by reweighting the data
until this condition is met. In Section IV.C, we discuss how near-critical
histogram data can be used to obtain precise estimates of the critical
parameters for a transition.
In the absence of phase transitions or at temperatures near a critical
point, the values of all observable quantities (such as the histograms of
energy and density) are independent of initial conditions, since free energy
barriers for transitions between states are small or nonexistent. However,
at lower temperatures, free energy barriers for nucleation of new phases
become increasingly larger. The states sampled at a given temperature and
chemical potential depend on initial conditions, a phenomenon known as
hysteresis. This is illustrated schematically in Fig. 5. For a supercritical
isotherm, T > Tc, the mean value of the density is a continuous function of
the chemical potential, and the same value is obtained for given conditions,
irrespective of the starting configuration. By contrast, for a subcritical
isotherm, when the runs are started from a low-density state, at some value
of the chemical potential, a discontinuous jump to a state of higher
density is observed. The exact location of the jump depends on the initial
state and the specific mix of Monte Carlo moves used to change the
configuration of the system. When simulations are started in a high-density

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

328

Panagiotopoulos

FIG. 5 Schematic diagram of the mean number of particles, hNi, vs. chemical
potential, , for a subcritical and a supercritical isotherm of a one-component fluid.
The curve for the supercritical isotherm has been shifted up for clarity. (From [14],
2000 IOP Publishing Ltd.)

state, the system remains on the high-density branch of the isotherm until
some value of the chemical potential that is lower than the chemical
potential of the jump from low- to high-density states.
The histogram reweighting method can be applied to systems with large
free energy barriers for transitions between states, provided that care is
taken to link all states of interest via reversible paths. One possibility is to
utilize umbrella or multicanonical sampling techniques [61,69] to artificially
enhance the frequency with which a simulation samples the intermediate
density region [62]. Multicanonical and umbrella sampling require as input
an estimate of the free energy in the intermediate density region, which has
to be obtained by trial and error. In addition, a significant fraction of
simulation time is spent sampling unphysical configurations of intermediate
density. An alternative approach is to link states by providing connections
through a supercritical path, in a process analogous to thermodynamic
integration [5]. This approach is illustrated schematically in Fig. 6. The filled
square represents the critical point for a transition, and open squares linked
by dashed lines represent tie-lines. Ellipses represent the range of particle
numbers and energies sampled by a single simulation. A near-critical
simulation samples states on both sides of the coexistence curve, while
subcritical simulations are likely to be trapped in (possibly metastable) states
on either side. However, as long as there is a continuous path linking all
states of interest, the free energies and pressures can be calculated correctly,
and an accurate phase envelope can be obtained.
An example of the application of histogram reweighting for determining
the phase behavior of a homopolymer model on the simple cubic lattice is

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Monte Carlo Methods

329

FIG. 6 Schematic diagram of the energy, E, vs. the number of particles, N, for a
one-component fluid with a phase transition. Squares linked by dashed lines are
coexisting phases joined by tie-lines and the filled square indicates the critical point
of the transition. Ellipses represent the range of particle numbers and energies
sampled during different GCMC runs. (From [14], 2000 IOP Publishing Ltd.)

illustrated in Fig. 7. The phase behavior and critical properties of the model
for a range of chain lengths have been studied in [63]. The system in this
example is for chain length r 8 and coordination number z 6. In this
example, we first performed a simulation at reduced temperature T * 11.5
and chemical potential *
60.4, for which the raw histogram data are
shown in Fig. 4. The resulting average volume fraction for the run is
indicated on Fig. 7 by the filled circle at T * 11.5. The range of volume
fractions sampled during the simulation is indicated on Fig. 7 by the arrows
originating at the run point. Because this run is near the critical point, a very
broad range of particle numbers and thus volume fractions is sampled
during this single run. The histogram from this run was then reweighted to
lower temperatures and a preliminary phase diagram was obtained. The
estimated coexistence chemical potential at T * 9 was used as input to a
new simulation, which sampled states near the saturated liquid line. The
same procedure was repeated, now with combined histograms from the first
two runs, to obtain an estimate of the coexistence chemical potential at
T * 7. A new simulation was performed to sample the properties of the
liquid at that temperature. The total time for the three runs was 10 CPU min
on a Pentium III 300 MHz processor. The final result of these three
calculations was the phase coexistence lines shown by the thick continuous
lines on Fig. 7.
Two general observations can be made in relation to this example.
First, it should be pointed out that the histogram reweighting method works
much faster on smaller system sizes. As the system size increases, relative

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

330

Panagiotopoulos

FIG. 7 Phase diagram for a homopolymer of chain length r 8 on a 10  10  10

simple cubic lattice of coordination number z 6. Filled circles give the reduced
temperature, T *, and mean volume fraction, hi, of the three runs performed.
Arrows from the run points indicate the range of densities sampled for each
simulation. The thick continuous line is the estimated phase coexistence curve.
(From [14], 2000 IOP Publishing Ltd.)

fluctutations in the number of particles and energy for a single run at

specified conditions decrease as the 1/2 power of the system volume V. This
implies that more simulations are required to obtain overlapping histograms
that cover the range of energies and densities of interest. Moreover, the
number of Monte Carlo moves required to sample properties increases
approximately linearly with system size in order to keep the number of
moves per particle constant. The computational cost of each Monte Carlo
move is proportional to system size for pairwise additive long-range
interactions and independent of system size for short-range interactions.
The net effect is that the total computational effort required to obtain a
phase diagram at a given accuracy scales as the 1.5 to 2.5 power of system
volume, respectively for short- and long-range interactions. Fortunately,
away from critical points, the effect of system size on the location of the
coexistence curves for first-order transitions is typically small. In this
example, calculations on a 153 system result in phase coexistence lines
practically indistinguishable from the ones shown in Fig. 7. The mean
absolute relative differences for the coexistence densities between the small
and large systems are 0.1% for the liquid and 1% for the (much lower
density) gas, well within the width of the coexistence lines on Fig. 7.
A second observation relates to calculations near critical points. The
coexistence lines in Fig. 7 do not extend above a temperature of T * 11.6
because above that temperature significant overlap exists between the liquid

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Monte Carlo Methods

331

and vapor peaks of the histograms. This overlap renders calculations of the
liquid and gas densities imprecise. Larger system sizes suffer less from this
effect and can be used to obtain coexistence densities near critical points. As
discussed in Section IV.C, a sequence of studies with increasing system size
is also required to obtain accurate estimates of critical points.

B. Multicomponent Systems
The histogram reweighting methodology for multicomponent systems
[6567] closely follows the one-component version described above. The
probability distribution function for observing N1 particles of component 1
and N2 particles of component 2 with configurational energy in the vicinity
of E for a GCMC simulation at imposed chemical potentials 1 and 2,
respectively, at inverse temperature  in a box of volume V is
f N1 , N2 , E

N1 , N2 , V, Eexp
E  1 N1  2 N2
1 , 2 , V, 

18

Equations (10) to (17) can be similarly extended to multicomponent systems.

The main complication in the case of multicomponent systems relative to
the one-component case is that the dimensionality of the histograms is one
plus the number of components, thus making their machine storage and
manipulation somewhat more challenging. For example, in the case of
one-component systems, it is possible to store the histograms directly as
two-dimensional arrays. The memory requirements for storing threedimensional arrays for a two-component system make it impractical to do
so. Instead, lists of observations of particle numbers and energies are
periodically stored on disk. It is important to select the frequency of
sampling of the histogram information so that only essentially independent
configurations are sampled. This implies that sampling is less frequent
at high densities for which the acceptance ratio of the insertion and
removal steps is lower. Sampling essentially independent configurations
also enforces the condition of equal statistical efficiency underlying the
FerrenbergSwendsen histogram combination equations [(13), (14)].

C. Critical Point Determination

Recent advances in the determination of critical parameters for fluids
lacking special symmetries have been based on the concept of mixed-field
finite-size scaling and have been reviewed in detail by Wilding [70]. As a
critical point is approached, the correlation length  grows without bound

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

332

Panagiotopoulos

and eventually exceeds the linear system size L of the simulation box.
Singularities and discontinuities that characterize critical behavior in the
thermodynamic limit are smeared out and shifted in finite systems. The
infinite-volume critical point of a system can, however, be extracted by
examining the size dependence of thermodynamic observables, through
finite-size scaling theory [7173]. The finite-size scaling approach proposed
by Bruce and Wilding [74,75] accounts for the lack of symmetry between
coexisting phases in most continuous-space fluids. Even though some recent
work has cast some shadow on its full reliability [76], the approach seems
to be quite robust and is easy to apply. For one-component systems, the
ordering operator, M, is proportional to a linear combination of the number
of particles N and total configurational energy U:
M / N
sU

19

where s is the field mixing parameter. For multicomponent systems, an extra

field mixing parameter appears for each added component; for example for
binary systems,
M / N1
sU
qN2

20

where q is the field mixing parameter for the number of particles of

component 2.
General finite-size scaling arguments predict that the normalized
probability distribution for the ordering operator M at criticality, }(M),
has a universal form. The order parameter distribution for the threedimensional Ising universality class is shown in Fig. 8 as a continuous line.
Also shown in Fig. 8 are data for a homopolymer of chain length r 200 on
a 50  50  50 simple cubic lattice of coordination number z 26 [63]. The
data were obtained by histogram reweighting methods, by adjusting the
chemical potential, temperature, and field mixing parameter s so as to
obtain the best possible fit to the universal distribution. The nonuniversal
constant A and the critical value of the ordering operator Mc were chosen so
that the data have zero mean and unit variance. Due to finite-size corrections to scaling, the apparent critical temperature, Tc(L), and density, c(L),
deviate from their infinite-system values, Tc(1) and c(1). They are
expected to follow the scaling relationships with respect to the simulated
system size, L:
Tc L
Tc 1 / L

1= c L
c 1 / L
1
=

21

where
, , and  are, respectively, the correction-to-scaling exponent,

the correlation length exponent, and the exponent associated with the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Monte Carlo Methods

333

FIG. 8 Ordering operator distribution for the three-dimensional Ising universality

class (continuous line) (data are courtesy of N. B. Wilding). Points are for a
homopolymer of chain length r 200 on a 50  50  50 simple cubic lattice of
coordination number z 26 [63]. The nonuniversal constant A and the critical value
of the ordering operator Mc were chosen so that the data have zero mean and unit
variance. (From [14], 2000 IOP Publishing Ltd.)

heat capacity divergence. For the three-dimensional Ising universality

class, the approximate values of these exponents are [77,80] (
, , ) 
(0.54, 0.629, 0.11). Figure 9 demonstrates these scaling relationships for
the critical temperature and density of the square-well fluid of range

3 [81].

D. Thermodynamic and Hamiltonian Scaling

Finally in this section, we would like to mention briefly two methods that
are related to histogram reweighting. Thermodynamic scaling techniques
proposed by Valleau [82] are based on calculations in the NPT, rather than
the grand canonical ( VT ) ensemble and provide information for the free
energy over a range of volumes, rather than a range of particle numbers.
Thermodynamic scaling techniques can also be designed to cover a range of
Hamiltonians (potential models) in the Gibbs [83] or grand canonical [84]
ensembles. In their Hamiltonian scaling form, the methods are particularly

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

334

Panagiotopoulos

FIG. 9 Critical temperature (a) and density (b) scaling with linear system size for
the square-fluid of range
3. Solid lines represent a least-squares fit to the points.
(From [81], 1999 American Institute of Physics.)

useful for optimizing parameters in intermolecular potential models to

reproduce experimental data such as the coexisting densities and vapor
pressures. Thermodynamic and Hamiltonian scaling methods require
estimates for the free energy of the system as a function of conditions, so
that the system can be forced to sample the range of states of interest with
roughly uniform probability, as for umbrella sampling Monte Carlo [69].

V. SMART SAMPLING FOR DIFFICULT SYSTEMS

A. Configurational-Bias Sampling
The most common bottleneck in achieving convergence in methods that rely
on particle transfers is the prohibitively low acceptance of transfer attempts.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Monte Carlo Methods

335

For dense fluid phases, especially for complex, orientation-dependent

intermolecular potentials, configurations with holes in which an extra
particle can be accommodated are highly improbable, and the converse step
of removing a particle involves a large cost in energy. Configurational-bias
sampling techniques significantly improve sampling efficiency for Gibbs or
grand canonical Monte Carlo simulations. The methods have been reviewed
in detail in [5,13], the chapter by Frenkel in [10], and in the Chapter 7 by
T.S. Jain and J. de Pablo in this book, and will only be covered briefly in
the present chapter.
Configurational-bias methods trace their ancestry to biased sampling for
lattice polymer configurations proposed by Rosenbluth and Rosenbluth
[85]. Development of configurational-bias methods for canonical and grand
canonical simulations and for continuous-space models took place in the
early 1990s [8690] and dramatically expanded the range of intermolecular
potential models that can be studied by the methods described in the
previous sections.
Configurational-bias methods are based on segment-by-segment insertions or removals of a multisegment molecule. Several trial directions are
attempted for every segment insertion, and a favorable growth direction
is preferentially selected for the segment addition. In this way, the
acceptance probability of insertions is greatly enhanced. For each segment
growth or removal step, a correction factor (often called the Rosenbluth
weight) is calculated. The product of the Rosenbluth weights of all steps is
incorporated in the overall acceptance criterion for particle insertions and
removals in order to correct for the bias introduced by the nonrandom
growth along preferential directions.

B. Expanded Ensembles
Another approach for handling multisegment molecules is based on the
concept of expanded ensembles [9194]. Expanded ensembles for chain
molecules construct a series of intermediate states for the molecule of
interest, from a noninteracting (phantom) chain to the actual chain with all
segments and interactions in place. These intermediate states can be
semipenetrable chains of the full length [91,92] or shortened versions of the
actual chain [93,94]. Estimates of the free energy of the intermediate states
are required to ensure roughly uniform sampling, as for thermodynamic
and Hamiltonian scaling methods mentioned in the previous section. The
advantage of expanded ensembles over configurational-bias methods is that
arbitrarily complex long molecules can be sampled adequately, if sufficient
computational effort is invested in constructing good approximations of the
free energies of intermediate states.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

336

Panagiotopoulos

VI. SOME APPLICATIONS TO POLYMERIC FLUIDS

The methods described in the previous sections enable fast and accurate
calculations of the phase behavior of fluids. Their availability has resulted in
a veritable explosion in the number of studies of the phase behavior of both
simple toy model and realistic potentials for fluids in the past decade.
Several reviews [69,14] have covered applications of such methods to many
different systems. Here, we focus on recent applications to polymeric fluids.
Phase equilibria in long-chain lattice polymer models were studied by
[63,78,79]. Of particular interest is a recent study by Yan and de Pablo [64]
that goes to chains of length 16,000 on the simple cubic lattice. By going
to such long chains, the authors were able to illustrate the presence of
logarithmic corrections to the exponent characterizing the decrease of
critical density with increase in chain length.
Hydrocarbon molecules are ubiquitous in industrial processes and
form the building blocks of biological systems. They are nonpolar and
consist of a small number of groups, thus making them the logical starting
point for potential model development. Siepmann, Karaborni, and Smit
[9597] used configurational-bias Gibbs ensemble simulations to obtain
an optimized potential model and the critical properties of the n-alkane
homologous series. At the time, there were conflicting experimental data
on the dependence of the critical density on chain length, which were
resolved with the help of the simulations. Spyriouni et al. [39] have studied
the phase behavior of n-hexadecane for the DoddTheodorou potential
[98] and obtained good agreement for the phase envelope but not for
the vapor pressure. Branched alkanes have been studied by [99102],
perfluorinated alkanes by [103], fluoromethanes by [104,105], and
-olephins by [106].
Three accurate united-atom potential sets for n-alkanes have appeared
recently. The TRAPPE [107] and NERD models [108] use the LennardJones (12,6) potential to describe nonbonded interactions among methyl
and methylene groups, while the model of Errington and Panagiotopoulos
[109] uses the exponential-6 functional form. All three reproduce the
experimental phase diagrams and critical points. The exponential-6 model is
slightly better with respect to representation of the vapor pressures. Figures
10 and 11 illustrate the quality of representation of experimental data for
the newer optimized models. Deviations from experimental data for the
exponential-6 united-atom model are comparable to those for a recently
developed explicit hydrogen model [110].
Alkane mixtures have been studied extensively in recent years. For
example, Chen and Siepmann investigated supercritical ethane and nheptane mixtures and obtained the free energy of transfer for n-pentane and

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Monte Carlo Methods

337

FIG. 10 Phase diagrams of selected n-alkanes. The curves from bottom to top are
for ethane, propane, butane, pentane, hexane, octane, and dodecane. Circles
represent calculations for the model of [109]. Uncertainties are smaller than the size
of the symbols. A solid line is used for experimental data and a star for the
experimental critical point. (From [109], 1999 American Chemical Society.)

n-heptane between He and n-heptane liquid phases [110]. They have also
obtained partition coefficients of alkanes between water and octanol [111].
Delhommelle et al. [112,113] studied mixtures of n-alkanes using both a
united-atom model [96] and anisotropic united-atom models originated
by Toxvaerd [114,115]. Other recent studies of alkane mixtures include
[109,116,117]. The solubility of small molecules such N2 and methane and
their mixtures in polyethylene, including effects of polymer crystallinity,
was studied in [118]. Mixtures with -olephins were studied in [38,106]. In
general, excellent agreement between experiment and simulation results is
obtained for these nonpolar mixtures, provided that the pure component
potentials have been optimized to reproduce the phase envelope and vapor

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

338

Panagiotopoulos

FIG. 11 Vapor pressures of selected n-alkanes. The curves from right to left are for
ethane, propane, butane, pentane, hexane, octane, and dodecane. Symbols are the
same as for Fig. 10. (From [109], 1999 American Chemical Society.)

pressures of the pure components. No mixture parameters are necessary for

the calculations.

VII.

CONCLUDING REMARKS

A novice researcher interested in obtaining the phase behavior of a fluid

is now faced with a bewildering choice among a number of alternative
methods and their variations. In this section, similarities and differences
among the methods, their relative performance, and their ease of
implementation will be discussed.
Simulations with an explicit interface appear, at first glance, to be
relatively simple to implement and perform. Unlike most other methods
discussed here, interfacial simulations can also be performed using
molecular dynamics codes. However, they provide an inefficient route to
the phase coexistence properties. Unless the properties of the interface itself
(or the surface tension) are of interest, other methods discussed in the
present chapter provide better alternatives.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Monte Carlo Methods

339

The majority of recent studies discussed in Section VI have been

performed using various implementations of the Gibbs ensemble, often
combined with configurational-bias methods to improve sampling for
multisegment molecules. The Gibbs method is relatively easy to implement
and provides direct information on the properties of coexisting phases from
a single simulation. One major weakness of the methodology is that it is not
applicable to solid or highly structured phases. For such systems, the only
possible choice is the GibbsDuhem integration method and its variations.
The GibbsDuhem method, however, needs to start from a point on the
coexistence curve.
The accuracy of the Gibbs ensemble method for a given amount of
computer time does not seem to match the accuracy of histogram reweighting
methods [117]. Histogram reweighting methods are also inherently better
at determining critical points via the finite-size scaling formalism. On the
negative side, the effort required to implement histogram combination and
reweighting is more than for the Gibbs ensemble. Histogram reweighting
methods are also indirect, requiring the assembly of the free energy function
of a system from a series of simulations. The efficiency of histogram methods
decreases rapidly with increase in system size. Despite these disadvantages,
they are probably the most promising for future applications.
The NPT test particle method shares with histogram reweighting
techniques the feature that it proceeds by computation of the chemical
potential for a number of state points. Histogram reweighting methods,
however, cover a range of densities and temperatures in a single simulation.
In addition, data from separate runs can be combined in a systematic way
only for histogram reweighting methods. A variation of the NPT test
particle method is the SPECS method for systems of long-chain molecules
[38,39]. The SPECS method is a good alternative to expanded-ensemble
Gibbs Monte Carlo calculations for cases for which configurational-bias
sampling methods become inefficient.
Interesting connections between many of the methods discussed in the
present chapter have been pointed out by Escobedo [54,55]. In particular,
Escobedo suggests that GibbsDuhem integration, pseudo-ensembles, and
the NPT test particle method can be considered as low-order approximations of a histogram reweighting approach.
In the area of applications, an important goal for research in coming
years will be to develop a set of pure component group-based potentials
and combining rules that can be used for general predictions of both pure
component and mixture phase behavior. Early results for realistic mixtures
[117] suggest that relatively simple intermolecular potential models can be
used to predict the phase behavior of broad classes of binary systems. For
mixtures with large differences in polar character of the components,

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

340

Panagiotopoulos

however, present models do not always result in quantitative agreement with

experiment. New models that include higher-order interactions such as
polarizability may be suitable for this purpose, a hypothesis that will need to
be tested in the future.

ACKNOWLEDGEMENTS
Research in the authors group on which this chapter is based has been
supported by the Department of Energy (Office of Basic Energy Sciences),
the National Science Foundation, the American Chemical Society
Petroleum Research Fund, and the Camille and Henry Dreyfus Foundation.

REFERENCES
1. Rowlinson, J.S.; Swinton, F. Liquids and Liquid Mixtures, 3rd Ed.; Butterworth
Scientic: London, 1982.
2. Prausnitz, J.M.; Lichtenthaler, R.N.; de Azevedo, E.G. Molecular
Thermodynamics of Fluid Phase Equilibria, 3rd Ed.; Prentice-Hall: New York,
1999.
3. Hansen, J.-P.; Verlet, L. Phys. Rev. 1969, 184, 151161.
4. Allen, M.P.; Tildesley, D.J. Computer Simulation of Liquids; Oxford University
Press: Oxford, 1987.
5. Frenkel, D.; Smit, B. Understanding Molecular Simulation; Academic Press:
London, 1996.
6. Panagiotopoulos, A.Z. Molec. Simulation 1992, 9, 123.
7. Smit. B. In Computer Simulations in Chemical Physics, Allen, M.P., Tildesley,
D.J., Eds.; NATO ASI Ser C, 1993, 397, 173209.
8. Panagiotopoulos, A.Z. In Kiran, E., Levelt Sengers, J.M.H., Eds.; NATO ASI
Ser. E, 1994, 273, 411437.
9. Panagiotopoulos, A.Z. In Observation, Prediction and Simulation of Phase
Transitions in Complex Fluids. NATO ASI Ser. C 1995, 460, 463501.
10. Baus, M.; Rull, L.F.; Ryckaert, J.-P. Eds. Observation, Prediction and
Simulation of Phase Transitions in Complex Fluids; NATO ASI Ser. C,
1995, 460.
11. Binder, K. Rep. Prog. Phys. 1997, 60, 487559.
12. Deem, M.W. AIChE Journal 1998, 44, 25692596.
13. Siepmann, J.I. Adv. Chem. Phys. 1999, 105, 443460.
14. Panagiotopoulos, A.Z. J. Phys.: Condens Matter 2000, 12, R2552.
15. Widom, B. J. Chem. Phys. 1963, 39, 28082812.
16. Kofke, D.A.; Cummings, P.T. Molec. Phys. 1997, 92, 973996.
17. Rowlinson, J.S.; Widom, B. Molecular Theory of Capillarity; Clarendon Press:
Oxford, Ch. 6, 1982.
18. Gubbins, K.E. Molec. Simulation 1989, 2, 223252.
19. Panagiotopoulos, A.Z. Molec. Phys. 1987, 61, 813826.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Monte Carlo Methods

341

20. Panagiotopoulos, A.Z.; Quirke, N.; Stapleton, M.; Tildesley, D.J. Molec. Phys.
1988, 63, 527545.
21. Smit, B.; de Smedt, P.; Frenkel, D. Molec. Phys. 1989, 68, 931950.
22. Smit, B.; Frenkel, D. Molec. Phys. 1989, 68, 951958.
23. Lopes, J.N.C.; Tildesley, D.J. Molec. Phys. 1997, 92, 187195.
24. Kristof, T.; Liszi, J. Molec. Phys. 1997, 90, 10311034.
25. Kristof, T.; Liszi, J. Molec. Phys. 1998, 94, 519525.
26. Hentschke, R.; Bast, T.; Aydt, E. et al. J. Molec. Model 1996, 2, 319326.
27. Kotelyanskii, M.J.; Hentschke, R. Molec. Simulation 1996, 17, 95112.
28. Baranyai, A.; Cummings, P.T. Molec. Simulation 1996, 17, 2125.
29. Valleau, J.P. J. Chem. Phys. 1998, 108, 29622966.
30. Bruce, A.D. Phys. Rev. E 1997, 55, 23152320.
31. Mon, K.K.; Binder, K. J. Chem. Phys. 1992, 96, 69896995.
32. Panagiotopoulos, A.Z. Int. J. Thermophys. 1989, 10, 447457.
33. Moller, D.; Fischer, J. Molec. Phys. 1990, 69, 463473; erratum in Molec. Phys.
1992, 75, 14611462.
34. Lot, A.; Vrabec, J.; Fischer, J. Molec. Phys. 1992, 76, 13191333.
35. Boda, D.; Liszi, J.; Szalai, I. Chem. Phys. Lett. 1995, 235, 140145.
36. Vrabec, J.; Fischer, J. Molec. Phys. 1995, 85, 781792.
37. Vrabec, J.; Lot, A.; Fischer, J. Fluid Phase Equil. 1995, 112, 173197.
38. Spyriouni, T.; Economou, I.G.; Theodorou, D.N. Phys. Rev. Lett. 1998, 80,
44664469.
39. Spyriouni, T.; Economou, I.G.; Theodorou, D.N. Macromolecules 1998, 31,
14301431.
40. Kumar, S.K.; Szleifer, I.; Panagiotopoulos, A.Z. Phys. Rev. Lett. 1991, 66,
29352938.
41. Kofke, D.A. Molec. Phys. 1993, 78, 13311336.
42. Kofke, D.A. J. Chem. Phys. 1993, 98, 41494162.
43. Mehta, M.; Kofke, D.A. Chem. Eng. Sci. 1994, 49, 26332645.
44. Agrawal, R.; Mehta, M.; Kofke, D.A. Int. J. Thermophys. 1994, 15,
10731083.
45. Kofke, D.A. Adv. Chem. Phys. 1999, 105, 405441.
46. Escobedo, F.A.; Pablo, J.J. J. Chem. Phys. 1997, 106, 98589868.
47. Escobedo, F.A.; Pablo, J.J. J. Chem. Phys. 1997, 106, 29112923.
48. Agrawal, R.; Kofke, D.A. Phys. Rev. Lett. 1995, 74, 122125.
49. Bolhuis, P.; Frenkel, D. J. Chem. Phys. 1997, 106, 666687.
50. Escobedo, F.A.; de Pablo, J.J. Europhys. Lett. 1997, 40, 111117.
51. Meijer, E.J.; El Azhar, F. J. Chem. Phys. 1997, 106, 46784683.
52. Mehta, M.; Kofke, D.A. Molec. Phys. 1995, 86, 139147.
53. Camp, P.J.; Allen, M.P. Molec. Phys. 1996, 88, 14591469.
54. Escobedo, F.A. J. Chem. Phys. 1998, 108, 87618772.
55. Escobedo, F.A. J. Chem. Phys. 1999, 110, 1199912010.
56. McDonald, I.R.; Singer, K. J. Chem. Phys. 1967, 47, 47664772.
57. Wood, E.E. J. Chem. Phys. 1968, 48, 415434.
58. Card, D.N.; Valleau, J.P. J. Chem. Phys. 1970, 52, 62326240.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

342
59.
60.
61.
62.
63.
64.
65.
66.
67.
68.
69.
70.
71.
72.
73.
74.
75.
76.
77.
78.
79.
80.
81.
82.
83.
84.
85.
86.
87.
88.
89.
90.
91.
92.
93.
94.

Panagiotopoulos
Bennett, C.H. J. Comp. Phys. 1976, 22, 245268.
Ferrenberg, A.M.; Swendsen, R.H. Phys. Rev. Lett. 1988, 61, 26352638.
Berg, B.A.; Neuhaus, T. Phys. Rev. Lett. 1992, 68, 912.
Wilding, N.B. Phys. Rev. E 1995, 52, 602611.
Panagiotopoulos, A.Z.; Wong, V.; Floriano, M.A. Macromolecules 1998, 31,
912918.
Yan, Q.L.; de Pablo, J.J. J. Chem. Phys. 2000, 113, 59545957.
Wilding, N.B. Phys. Rev. E 1997, 55, 66246631.
Wilding, N.B.; Schmid, F.; Nielaba, P. Phys. Rev. E 1998, 58, 22012212.
Poto, J.J.; Panagiotopoulos, A.Z. J. Chem. Phys. 1998, 109, 1091410920.
Ferrenberg, A.M.; Swendsen, R.H. Phys. Rev. Lett. 1989, 63, 11951198.
Torrie, G.M.; Valleau, J.P. J. Comp. Phys. 1977, 23, 187199.
Wilding, N.B. J. Phys.: Condens Matter 1997, 9, 585612.
Ferdinand, A.E.; Fisher, M.E. Phys. Rev. 1969, 185, 832.
Fisher, M.E.; Barber, M.N. Phys. Rev. Lett. 1972, 28, 1516.
Privman, V. Ed. Finite Size Scaling and Numerical Simulation of Statistical
Mechanical Systems, World Scientic: Singapore, 1990.
Bruce, A.D.; Wilding, N.B. Phys. Rev. Lett. 1992, 68, 193196.
Wilding, N.B.; Bruce, A.D. J. Phys.: Condens Matter 1992, 4, 30873108.
Orkoulas, G.; Fisher, M.E.; Ustun, C. J. Chem. Phys. 2000, 113, 75307545.
Sengers, J.V.; Levelt-Sengers, J.M.H. Ann. Rev. Phys. Chem. 1986, 37,
189222.
Yan, Q.; Liu, H.; Hu, Y. Macromolecules 1996, 29, 40664071.
Frauenkron, H.; Grassberger, P. J. Chem. Phys. 1997, 107, 95999608.
Liu, A.J.; Fisher, M.E. Physica A 1989, 156, 35.
Orkoulas, G.; Panagiotopoulos, A.Z. J. Chem. Phys. 1999, 110, 15811590.
Valleau, J.P. Advances in Chemical Physics 1999, 105, 369404.
Kiyohara, K.; Spyriouni, T.; Gubbins, K.E. et al. Molec. Phys. 1996, 89,
965974.
Errington, J.R.; Panagiotopoulos, A.Z. J. Chem. Phys. 1998, 109, 10931100.
Rosenbluth, M.N.; Rosenbluth, A.W. J. Chem. Phys. 1955, 23, 356359.
Frenkel, D.; Mooij, G.C.A.M.; Smit, B. J. Phys.: Condens Matter 1992, 4,
30533076.
de Pablo, J.J.; Laso, M.; Siepmann, J.I.; Suter, U.W. Molec. Phys. 1993, 80,
5563.
Siepmann, J.I.; McDonald, I.R. Molec. Phys. 1993, 79, 457473.
Mooij, G.C.A.M.; Frenkel, D.; Smit, B. J. Phys.: Condens Matter 1992, 4,
L255L259.
Laso, M.; de Pablo, J.J.; Suter, U.W. J. Chem. Phys. 1992, 97, 28172819.
Lyubartsev, A.; Martsinovski, A.; Shevkunov, S. J. Chem. Phys. 1992, 96,
17761783.
Vorontsov-Velyaminov, P.; Broukhno, A.; Kuznetsova, T. J. Phys. Chem.
1996, 100, 11531158.
Wilding, N.B.; Muller, M. J. Chem. Phys. 1994, 101, 43244330.
Escobedo, F.; de Pablo. J.J. J. Chem. Phys. 1996, 105, 43914394.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Monte Carlo Methods

343

95. Siepmann, J.I.; Karaborni, S.; Smit, B. J. Am. Chem. Soc. 1993, 115,
64546455.
96. Smit, B.; Karaborni, S.; Siepmann, J.I. J. Chem. Phys. 1995, 102, 21262140.
97. Siepmann, J.I.; Karaborni, S.; Smit, B. Nature 1993, 365, 330332.
98. Dodd, L.R.; Theodorou, D.N. Adv. Pol. Sci. 1994, 116, 249281.
99. Cui, S.T.; Cummings, P.T.; Cochran, H.D. Fluid Phase Equil. 1997, 141,
4561.
100. Siepmann, J.I.; Martin, M.G.; Mundy, C.J.; Klein, M.L. Mol. Phys. 1997, 90,
687693.
101. Martin, M.G.; Siepmann, J.I. J. Phys. Chem. B 1999, 103, 45084517.
102. Zhuravlev, N.D.; Siepmann, J.I. Fluid Phase Equil. 1997, 134, 5561.
103. Cui, S.T.; Siepmann, J.I.; Cochran, H.D.; Cummings, P.T. Fluid Phase Equil.
1998, 146, 5161.
104. Potter, S.C.; Tildesley, D.J.; Burgess, A.N. et al. Molec. Phys. 1997, 92,
825833.
105. Jedlovszky, P.; Mezei, M. J. Chem. Phys. 1999, 110, 29913002.
106. Spyriouni, T.; Economou, I.G.; Theodorou, D.N. J. Am. Chem. Soc. 1999,
121, 34073413.
107. Martin, M.G.; Siepmann, J.I. J. Phys. Chem. B 1998, 102, 25692577.
108. Nath, S.K.; Escobedo, F.A.; de Pablo, J.J. J. Chem. Phys. 1998, 108,
99059911.
109. Errington, J.R.; Panagiotopoulos, A.Z. J. Phys. Chem. B 1999, 103, 6314.
110. Chen, B.; Siepmann, J.I. J. Phys. Chem. B 1999, 103, 53705379.
111. Chen, B.; Siepmann, J.I. J. Am. Chem. Soc. 2000, 122, 64646467.
112. Delhommelle, J.; Boutin, A.; Tavitian, B.; Mackie, A.D.; Fuchs, A.H. Molec.
Phys. 1999, 96, 15171524.
113. Ungerer, P.; Beauvais, C.; Delhommelle, J.; Boutin, A.; Rousseau, B.; Fuchs,
A.H. J. Chem. Phys. 2000, 112, 54995510.
114. Padilla, P.; Toxvaerd, S. J. Chem. Phys. 1991, 95, 509519.
115. Toxvaerd, S. J. Chem. Phys. 1997, 107, 51975204.
116. Mackie, A.D.; Tavitian, B.; Boutin, A.; Fuchs, A.H. Mol. Simulation 1997,
19, 115.
117. Poto, J.J.; Errington, J.R.; Panagiotopoulos, A.Z. Molec. Phys. 1999, 97,
10731083.
118. Nath, S.K.; de Pablo, J.J. J. Phys. Chem. B 1999, 103, 35393544.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

10
Gibbs Ensemble Molecular Dynamics
MICHAEL KOTELYANSKII
New Jersey, U.S.A.
REINHARD HENTSCHKE

I.

Rudolph Technologies, Inc., Flanders,

Bergische Universitat, Wuppertal, Germany

THE METHOD

The Gibbs Ensemble was introduced by A. Panagiotopoulos [1] to

conveniently model the coexistence of two phases in equilibrium
having the same pressure and chemical potentials. It mimics the coexistence
of two phases, a liquid and its vapor in a closed vessel. In this case the total
volume occupied by the two phases and the total number of particles are
constant, but the particles are allowed to transfer between the phases.
The computer model consists of two boxes, which exchange volume and
particles in such a way that the total volume is constant, i.e., the increment
in volume of one box is accompanied by the same decrement of the other
boxs volume. Particles are transferred between the simulation boxes. It is
shown in [1] that the Monte Carlo algorithm can be constructed with the
proper acceptance criteria for volume and particle exchange moves to
simulate the two coexisting phases in equilibrium with each box containing
one of the phases. The scheme is very convenient for studying phase
diagrams, as chemical potential and pressure do not have to be specified.
The system automatically adjusts them to the coexistence values at a given
temperature by exchanging particles and volumes.
As we mentioned before in the Background chapter, any ensemble can be
implemented with both Monte Carlo and Molecular Dynamics algorithms.
Which of the two should be chosen depends on the problem at hand.
Depending on the particular application, either Molecular Dynamics or
Monte Carlo can be more complex to program, especially when complex
segmental moves are involved in MC, but MD can provide more

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

346

Kotelyanskii and Hentschke

information about system dynamics. Gibbs Ensemble Monte Carlo is

discussed in the previous Chapter 9 by A. Panagiotopoulos.
Here we describe an implementation of the Gibbs Ensemble Molecular
Dynamics (GEMD) technique, that implements the ideas underlying Gibbs
Ensemble Monte Carlo using molecular dynamics.
In molecular dynamics (MD) one numerically solves the equations of
motion for a system of N particles contained in a box of volume V having
the total potential energy U. Here we consider a molecular system consisting
of N atoms in M molecules, and we write the total potential energy U as a
sum over inter- and intramolecular interactions, i.e.,
U Uinter Uintra

The first term is the sum over atomatom pair potentials Uinter
P
 2 i,  2 j,i>j,j 1,M(r) depending on the vectors r connecting atoms
 and  in the two molecules i and j. Uintra describes the interactions within
the molecule.
In order to simulate a variable number of molecules i in each of the two
boxes we introduce an extra degree of freedom i for every molecule. i can
vary between 1 and 0, where i 1 means that molecule i is in box one,
whereas i 0 means that it is in box two. For 1>i>0 it is in a transition
state, where it is felt in both boxes. Thus, we rewrite the intermolecular
potential energy of the system as a function of the coordinates and the i as
X

Uinter

r , V1
i
j

2i, 2j, i>j, j1, M

r , V2
1
i
1
j

2i, 2j, i>j, j1, M

where V1 and V2 are the volumes of the two boxes. The two terms represent
the intermolecular potential energies of the first and the second box, respectively. Consider, for instance, two particles with i 1 and j 1, i.e., both
particles belong to box one. In this case the product (1
i)(1
j) vanishes,
and only the first term in Eq. (2) will contribute to the nonbonded potential
energy. Notice also that as soon as we apply periodic boundary conditions,
and interparticle interactions are calculated involving the particles closest
images, the distance between them, and therefore the intermolecular
potential energy, is a function of the box dimensions (or of the volume if
the shape of the box is kept fixed).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Gibbs Ensemble Molecular Dynamics

347

The transfer of particles (or molecules) is controlled by the difference

between their potential energies in the two boxes. The number of unphysical
(but necessary) transition state molecules can be made small in comparison
to the total number of molecules by introducing an additional barrier
described by the transfer potential function g(i).
Utotal Uintra Uinter

gi

A possible choice is g(i) ![tanh(ui) tanh(u(1
i))
1] if 0  i  1 and

g(i) 1 otherwise, where u is the steepness, and ! is the height of the
barrier.
The rate of particle transfer between the boxes determines the rate with
which the equilibrium between the boxes is reached. The higher the rate the
faster the two phases equilibrate. On the other hand, particles that are in the
transition state contribute nonphysically to both boxes at the same time,
and one would try to keep their number close to zero during the production
run. This tuning is achieved by properly choosing the barrier function
parameters. The higher the barrier, the harder is it for the particles to cross
from one box to another.
In the case of phase coexistence in a one-component system the
temperature, the pressure, and the chemical potential, even though the
latter two are not explicitly specified, must be equal in the two phases and
thus in the two boxes. This is achieved if every change of the volume of one
of the boxes is accompanied by an opposite but equal change of the volume
of the other box. The total volume of the two boxes is therefore constant,
while the individual volumes are variable. The GEMD equations of motion
can be written as follows:
p m r_ 
@U

p
@r
"
#
1 X p2

XkB T
_
QT  m

p_ 

pi mi _i

p_

@U
@Uinter @gi




@i
@i
@i

pV1 QP V_1

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

348

Kotelyanskii and Hentschke

@U
@V1 "
X
@


r , V1
i
j
@V1 2i, 2j, i>j, j1, M

p_ V1

r , V2
1
i
1
j

2i, 2j, i>j, j1, M

Pe1
Pe2
Here p and pi are the momenta conjugate to the Cartesian coordinates r
and transfer variable i of the particle i. The first three equations describe
the evolution of a system coupled to an external heat bath with the temperature T [2,3].  is an additional degree of freedom, describing the coupling of the system to the external thermostat, which is necessary to simulate
constant temperature. QT describes the strength of the coupling, and X
equals the number of degrees of freedom, coupled to the thermostat.
The next two equations govern the evolution of the i, and thus the
transfer of the molecules between the boxes. Note that the intramolecular
part of the potential energy Uintra is independent of the i, and only
intermolecular interactions and the additional potential g(i) appear on the
right hand side.
The last two equations are the equations of motion of the box volume V1,
where pV1 is a momentum variable conjugate to the volume of the first box
V1, and QP is a parameter governing the volume relaxation. Again only the
intermolecular interactions Uinter depend on the box size, because periodic
boundary conditions are not applied to the intramolecular interactions.
Distances describing the intramolecular interactions Uintra are smaller than
the box sizes. They are of the order of a few chemical bond lengths, while the
box sizes are usually many times larger. As the sum of the volumes of the
two boxes is constant, pV1
pV2. Note that, similar to any constantpressure MD or MC algorithm, volume changes are controlled by the
difference between the instantaneous values of the external pressures [4,5].

II. STATISTICAL MECHANICAL FOUNDATION

As we discussed in the Background chapter, to prove that integrating this set
of differential equations does indeed simulate a model system with chemical
potentials and pressures equal in both boxes, we have to show that the
trajectory averages calculated by integrating Eqs. (4) are the corresponding
thermodynamic averages.
We start with writing the Liouville equation [6] for this model, which
describes the evolution of the phase space density distribution () with

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Gibbs Ensemble Molecular Dynamics

349

time. In our case the phase space  includes Cartesian coordinates and
momenta, thermostat coupling variable , volume V1, the transfer variables
i and their conjugate momenta pi and pV1.* The stationary (timeindependent) solution of this equation, describes the distribution function
for the ensemble reproduced.

X @
X @
@
@
@

r_ 
p_ 
p_ i

_ i
@t 2i, i1, M
@r
@p
@i
@pi
i1, M
@
@
@
p_ V1
_
V_1
@V1
@pV1
@
"
#
X @_i @V_ 1 @p_ V
X
@_r @_p 
@_

1


0
@r @p
@i @V1 @pV1 @
2i, i1, M
i1, M

By direct substitution and using Eqs. (4) to evaluate partial derivatives, one
can see that the following density function does indeed satisfy the Liouville
equation with @/@t 0. The solution is given by the following expression


GEMD fr g, fp g, fi g, fpi g, V1 , pV1 , 
"
1
/ exp

Utotal fr g, fi g, V1
kB T
X

X p2
p2
p2
Q T 2

V1

2m i1, M 2mi 2Qp

2
2i, i1, M

!#
6

Here, mi is a mass assigned to the particle for motion along the virtual
transfer direction i. Provided the system is ergodic, the averages over trajectories obtained by integrating Eqs. (4) are equivalent to the averages
calculated with the distribution function GEMD (6). Below we show that,
provided the number of particles in the transition state is small, averages for
particles in each box do in fact correspond to the ensemble with constant
temperature, pressure, and chemical potential, and that in fact the pressures
and chemical potentials are equal in both boxes.
The GEMD trajectory average of a property A, which depends on
the coordinates and momenta of the atoms 1, . . . , n, constituting the

*See corresponding section in Chapter 1.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

350

Kotelyanskii and Hentschke

molecules 1, . . . , m residing in the first box, out of the total N, given by

Q0GEMD

dM drN dpN dV1 Afr1 , . . . , rn g, fp1 , . . . , pn g

!#
X
1
p2
Utotal fr g, fi g, V1
 exp

kB T
2m
2i, i1, M
Z
dM drN dpN dV1
"
!#
X
1
p2
 exp

Utotal fr g, fi g, V1
kB T
2m
2i, i1, M

hAiGEMD

Q0GEMD
"

Here we already integrated over , pi, pV1 and we have canceled the respective factors in the average and in the partition function, as indicated by the
primed Q0GEMD . By choosing a proper g(i) one can make the number of
particles in the transition state negligibly small. In this case we can assume
that i takes only two values 0 or 1, and therefore we can replace the
integration over i by the summation over all possible combinations of
the i values (0 or 1). The equation for the average becomes:
Z
M!
drn dpn dV1 Afr1 ,...,rn g,fp1 ,...,pn g
hAiGEMD 0
QGEMD m0 m!M
m!
"
!#
n
X
1
p2
Utotal fr g,V1
 exp

kB T
2m
2i,i1
"
!#
Z
M
2
X
1
p

 drN
n dpN
n exp

Utotal fr g,V2
kB T
2m
2i,im1
1

M
X

8
Q0GEMD transforms analogously. M!/(m!(M
m)!) accounts for particle distinguishability. Here we also write the potential energies in the two boxes as
Utotal({r},V1) and Utotal({r},V2), independent of the i, because we neglect
particles in the transfer state with 0<i<1, assuming i can only be equal
to 0 or 1. In this case the total potential energy is a sum of potential energies
due to the interactions inside each of the boxes, as seen from Eq. (2).
Now we are going to demonstrate, that the chemical potentials and the
pressures are indeed the same in the two simulation boxes. The second
integration over the coordinates and momenta of the N
n atoms in the
second box equals QM
m, V2 , T h3M
m M
m!, where QM
m,V2 ,T is the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Gibbs Ensemble Molecular Dynamics

351

partition function of the isochoricisothermal (M
m, V2,T) ensemble.

It can be expressed in terms of the partition function for the (M
m, P2, T )
ensemble with constant pressure and temperature for the second box using
the saddle point method (see [5,7] for details)
QM
m, V2 , T





P2 V2
C
1
QM
m, P2 , T exp
1O
M
m
kB T N
n1=2

Here C is a constant independent of m and M. Now we rewrite QM
m,P2 ,T

as [6]

2 n
exp
kB T

QM
m, P2 , T
QM, P2 , T

10

where 2 is the chemical potential in the second box. Substituting all these in
the average hA i GEMD (8), and canceling terms independent of n between
numerator and denominator, we obtain in the thermodynamic limit:

hAiGEMD


1
2 m
exp
Q00GEMD m0 m!h3M
kB T
Z
 drn dpn dV1 Afr1 , . . . , rn g, fp1 , . . . , pn g
"
!#



m
X
P2 V1
1
p2
Utotal fri g, V1
 exp

exp

kB T
kB T
2m
2i, i1
1

M
X

11
and the corresponding expression for Q00GEMD , where the double prime is a
reminder that the common factors are canceled and that we keep only the
leading contribution in the limit of a large number of particles. The above
Eq. (11) coincides with the average over the constant- ( , P, T) ensemble [8].
Notice that the average for the first box is calculated with the pressure and
chemical potential values of the second box. This proves that these values
are indeed the same in both boxes.

III. IMPLEMENTATION
Two boxes in the GEMD are implemented in the following way: Cartesian
coordinates of the particles do not change when the particle is transferred
from one box to another. Its presence in either box is determined by the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

352

Kotelyanskii and Hentschke

value of its ghost coordinate i. Thus the transfer occurs as a

particle gradually entering the box from the additional space dimension, pretty much as teleportation in a science fiction novel. The other
particles in the entrance box feel the newcomer as its i value becomes
close to 0, if it enters box one, or as i approaches 1 if the ghost appears
in box two.
Molecules are transferred as a whole, with all atoms in the molecules
having the same value of . Only nonbonded interactions depend on the box
sizes, due to the periodic boundary conditions. Thus the nonbonded
interactions depend on the fraction of the molecule present in box one
and its fraction in box two, as determined by the i.
Given a constant temperature and pressure MD code, it is not very
difficult to modify it for GEMD. Experience also shows that the
same results are obtained both with the NoseHoover constant temperature
MD algorithm, which exactly reproduces the (NVT) ensemble [2,3],
and with the approximate but numerically more stable weak coupling
approach [9].
Additional variables to store energies, pressure, and other values for two
different boxes are to be added. Additional arrays storing values of variables
i, velocities, forces, and masses are to be added.
double precision x(N),y(N),z(N),xi(N)
double precision vx(N),vy(N),vz(N),vxi(N)
double precision fx(N),fy(N),fz(N),fxi(N)
double precision mass(N),mass_xi(N)
double precision boxx_1,boxy_1,boxz_1,invbx_1,invby_1,
invbz_1
double precision boxx_2,boxy_2,boxz_2,invbx_2,invby_2,
invbz_2
double precision volume_1, volume_2
double precision virial_1, virial_2
double precision pot_energy_1, pot_energy_2,
kin_energy_1, kin_energy_2
double precision kin_energy_xi
.....................
volume_1boxx_1*boxy_1*boxz_1
volume_2boxx_2*boxy_2*boxz_2
invbx_11.d0/boxx_1
invby_11.d0/boxy_1
invbz_11.d0/boxz_1
invbx_21.d0/boxx_2
invby_21.d0/boxy_2
invbz_21.d0/boxz_2
.............................................

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Gibbs Ensemble Molecular Dynamics

353

The nonbonded potential energy, force, and virial calculations are to be

amended, incorporating Eq. (2). The code example given here is probably
far from being the fastest possible way to calculate nonbonded interactions.
Even though we assume that the potential is pairwise and spherically
symmetric, we use function calls to calculate energy, force, and virial for the
purpose of clarity. The actual implementation should be in-lined and
optimized for the interatomic potential used. See for instance [10] and other
chapters in this book. Calculation of the forces for the i coordinates should
also be added. Here, again we use a function call to add the contribution due
to the g(i), which would be more efficient to be implemented in-line.
do i1,N
1
do ji1,N
c for box 1
xi_ij xi(i)* xi(j)
dx x(i)
x(j)
dxx dx
boxx_1*anint(dx*invbx_1)
dy y(i)
y(j)
dyy dy
boxy_1*anint(dy*invby_1)
dz z(i)
z(j)
dzz dz
boxz_1*anint(dz*invbz_1)
r dxx*dxx dyy*dyy dzz*dzz
pot_ij_1 potential(r)* xi_ij
pot_energy_1 pot_energy_1 pot_ij_1
force_abs potential_deriv(r)*xi_ij
fx force_abs*dxx/r
fy force_abs*dyy/r
fz force_abs*dzz/r
fx(i) fx(i) fx
fy(i) fy(i) fy
fz(i) fz(i) fz
fx(j) fx(i)
fx
fy(j) fy(i)
fy
fz(j) fz(i)
fz
virial_1 virial_1 fx*dxxfy*dyyfz*dzz
c again for box 2
xi_ij (1.d0
xi(i)) * (1.d0
xi(j))
dxx dx
boxx_2*anint(dx*invbx_2)
dyy dy
boxy_2*anint(dy*invby_2)
dzz dz
boxz_2*anint(dz*invbz_2)
r dxx*dxx dyy*dyy dzz*dzz
pot_ij_2 potential(r)*xi_ij
pot_energy_2 pot_energy_2 pot_ij_2

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

354

Kotelyanskii and Hentschke

force_abs potential_deriv(r)*xi_ij
fx force_abs*dxx/r
fy force_abs*dyy/r
fz force_abs*dzz/r
fx(i) fx(i) fx
fy(i) fy(i) fy
fz(i) fz(i) fz
fx(j) fx(i)
fx
fy(j) fy(i)
fy
fz(j) fz(i)
fz
virial_2 virial_2 fx*dxxfy*dyyfz*dzz

c forces along the ghost direction xi

fxi(i) fxi(i)
( pot_energy_1 * xi(i)
pot_energy_2 *
(1.d0
xi(j)) )
fxi(j) fxi(j)
( pot_energy_1 * xi(j)
pot_energy_2 *
(1.d0
xi(i)) )
enddo
enddo
c additional transfer potential for fxi
do i1,N
fxi(i) fxi(i) deriv_g(xi(i))
enddo
virial_1 virial_1/3.
virial_2 virial_2/3.
.....................

The equations describing the evolution of i are to be added to the

routine, calculating next time step coordinates and velocities. The same
numerical integration algorithms used to integrate Cartesian coordinates in
principle can be used for i. However, some care is required at the interval
boundaries, where the transfer potential g(i) is discontinuous [5]. Also M
additional degrees of freedom are to be accounted for, and the corresponding kinetic energy is to be included, when the temperature is calculated.
The above example of the GEMD code is suitable for a system of atoms,
not molecules. In the molecular system, all atoms of the same molecule have
the same  value. Thus even though the forces driving molecule transfer are
calculated for each individual atom, one has to make sure that the molecules
are transferred as whole objects. This can be achieved as follows, given a
working NPT-ensemble code for the molecular system:
1.

The same changes as described above for the atomic system are to
be made.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Gibbs Ensemble Molecular Dynamics

2.
3.

355

All atoms in the same molecule are assigned the same transfer
mass mi.
After the forces fxi along the i direction are calculated for all
atoms, they are redistributed evenly between the atoms of the same
molecule, so that the total force along the i direction acting on the
molecule remains the same. Here is how it looks in a pseudo-code:
for all molecules
ftotal0
for all atoms in the molecule
ftotal ftotalfxi(atom)
end
ftotal ftotal/num_of_atoms
for all atoms in the molecule
fxi(atom)ftotal
end
end

4.

 coordinate updates for the next time step, similar to those for
Cartesian coordinates are to be added.

And of course statistic accumulation parts of the code have to be changed

accordingly. Energy, pair radial distribution function (PRDF), or diffusion
coefficient should be calculated separately for molecules in each of the
boxes.
When the phase diagrams are calculated, it is important to correctly
account for the long-range nature of the interaction potential. Even the
nonelectrostatic interactions, usually believed to be relatively short-ranged,
have to be corrected. The corrections should be applied to the pressure and
to the energy. They take into account the interactions of the atoms that are
farther apart than the cutoff distance. Usually at these distances the center
of mass pair correlation function equals one, and the corrections are easily
calculated using equations for the energy and the pressure in terms of the
interaction potential alone (see equations in Section I.C.4 in Chapter 1).
For the spherically symmetric pairwise interaction described by (r) the
long-range corrections can be calculated as:
Z1
1
E Utotal N
r4r2 dr
12
2
rcut
Z1
1
dr
4r2 dr
PV Pcut V
N
r
6
dr
rcut
Both corrections depend on the density, and they may be quite different for
the gas and liquid phases. Therefore, they strongly affect the position of the
critical point. This is illustrated in Figs. 1 and 2.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

356

Kotelyanskii and Hentschke

FIG. 1 Liquidgas coexistence densities for the Lennard-Jones (LJ) fluid.

Temperature and density are in reduced LJ units. Filled symbols: GEMD, open
symbols: GEMC [11]. Reproduced from [5]. Note that the LJ potential in this case
was cut at 2.5  and shifted to zero. Crosses represent points obtained with longer LJ
potential cutoff and applying the long-range correction. The critical point of the
normal LJ system is close to TC
1:35 and 
C 0:307.

FIG. 2 Liquidgas coexistence curve for hexane. Open circles: experimental data of
reference [12]; open triangles: Gibbs Ensemble Monte Carlo result with 13.8 A cutoff
with the long-range corrections from [13]. The large solid circle corresponds to the
experimental critical point. Open squares: GEMD using a cut and shifted LJ
potential, where the cutoff is at 10 A; solid squares: using a 15 A cutoff instead; small
solid circles at T 440K: liquid densities for a 12 A and 14 A cutoff, respectively;
pluses: result obtained for a 10 A cutoff including long-range corrections. As in
Fig. 1, the error bars indicate standard deviations. This is figure 3 taken from
reference [5].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Gibbs Ensemble Molecular Dynamics

357

FIG. 3 Mean squared displacements hr2i of the hexane carbon atoms vs. time t in
n-hexane calculated at coexistence in the liquid via GEMD (open symbols) and via
NVT-MD (solid symbols) under corresponding conditions for T 400K (circles) and
T 460K (squares). The lines (solid: GEMD; dashed: NVT-MD) are fits to the linear
range of hr2i. This is figure 6 taken from Reference [5].

IV. EXAMPLES
GEMD has been used to calculate phase diagrams for atomic as well as
molecular systems, such as Lennard-Jones spheres or n-hexane, as shown
in Figs. 1 and 2.
Being a molecular dynamics technique, GEMD provides dynamical
information such as diffusion coefficients in the coexisting phases [5]. Shown
in Fig. 3 are the mean squared displacements of the carbon atoms in
n-hexane, calculated in coexisting phases with GEMD, compared to the
results of the constant temperature molecular dynamics at the same
conditions. This shows that diffusion coefficients and other dynamical
information can be extracted from GEMD simulations, together with the
thermodynamic properties.
The Gibbs Ensemble Monte Carlo technique transfers particles between
the boxes in one stepas a whole. To be accepted, this insertion requires
a reasonable probability to find an empty space. In the dense phases this
is a problem. GEMD transfers particles continuously. Transferred atoms
gradually appear and disappear from the simulation box. This may be
a great advantage when modeling liquidliquid or liquidsolid phase
equilibria. More examples of GEMD applications and technical details can
be found in the original papers describing applications of GEMD to the
study of n-alkanes [5] and sorption in zeolites [14].

REFERENCES
1. Panagiotopoulos, A.Z. Direct Determination of Fluid Phase Equilibria by
Simulating the Gibbs Ensemble: A Review. Mol. Sim. 1992, 9, 123.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

358

Kotelyanskii and Hentschke

2. Hoover, W.G. Canonical Dynamics: Equilibrium Phase Space Distribution.

Phys. Rev. 1985, A31, 16951697.
3. Nose, S. A Molecular Dynamics Method for Simulations in the Canonical
Ensemble. Mol. Phys. 1984, 52, 255268.
4. Winkler, R.G.; Moravitz, H.; Yoon, D.Y. Novel Molecular Dynamics
Simulations at Constant Pressure. Mol. Phys. 1992, 75, 669688.
5. Kotelyanskii, M.J.; Hentschke, R. Gibbs-Ensemble Molecular Dynamics:
Liquid-Gas Equilibria for Lennard-Jones Spheres and n-Hexane. Molecular
Simulations 1996, 17, 95112.
6. Landau, L.D.; Lifshitz, E.M. Statistical Physics. Course of Theoretical Physics,
3rd Ed.; Pergamon Press: Oxford, Vol. 5, 1980.
7. Kotelyanskii, M.J.; Hentschke, R. Gibbs-Ensemble Molecular Dynamics:
Liquid-Gas Equilibrium in a Lennard-Jones System. Phys. Rev. E 1995,
51, 51165119.
8. Hill, T.L. Statistical Mechanics; Dover: New York, 1987.
9. Berendsen, H.J.C.; Postma, J.P.M.; van Gunsteren, W.F.; Di Nola, A.; Haak,
J.R. Molecular Dynamics with Coupling to an External Bath. J. Chem. Phys.
1984, 81, 36843690.
10. Allen, M.P.; Tildesley, D.J. Computer Simulation of Liquids; Clarendon Press:
Oxford, 1993.
11. Smit, B. Phase Diagrams of Lennard-Jones Fluids. J. Chem. Phys. 1992,
96, 86398640.
12. Smith, B.D.; Srivastava, R. Thermodynamic Data for Pure Compounds:
Hydrocarbons and Ketones; Elsevier: Amsterdam, 1986.
13. Siepmann, S.; Karaborni, J.I.; Smit, B. Simulating the Critical Behavior of
Complex Fluids. Nature 1993, 365, 330332.
14. Hentschke, R.; Bast, T.; Aydt, E.; Kotelyanskii, M. Gibbs-Ensemble Molecular
Dynamics: A New Method for Simulations Involving Particle Exchange.
J. Mol. Mod. 1996, 2, 319326.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

11
Modeling Polymer Crystals
GREGORY C. RUTLEDGE Massachusetts Institute of Technology,
Cambridge, Massachusetts, U.S.A.

I.

INTRODUCTION

Crystallizable polymers constitute the majority of plastics in current use.

Nevertheless, the study of polymer crystals by molecular simulations
requires concepts and techniques somewhat different from those invoked
by conventional views of polymer melts and glasses. Amorphous polymers
are liquid-like in their structure; they exhibit only short range order.
Amorphous melts are liquid-like in their dynamics as well. The methods
employed to study amorphous polymers thus rely heavily on the Monte
Carlo and molecular dynamics methods developed in the 1950s and 1960s
to study simple liquids, and on refinements which permit efficient sampling
of the vast conformation space available to a polymer coil in the amorphous state. Polymer crystals, on the other hand, exhibit long range order
similar to that found in small molecule solids; they are characterized by high
density, limited conformation space, and pronounced anisotropy. The study
of crystalline polymers benefits considerably from the solid-state physics
developed in the first half of the twentieth century. However, polymer crystals differ in several important respects from simple atomic crystals, and
the modern simulation of polymer crystals draws upon ideas from both
solid-state and fluid-state physics.

II. STRUCTURE OF POLYMER CRYSTALS

As with any material, the molecular simulation of a crystalline solid requires
the specification of the nuclear coordinates of all the atoms in the solid, or a
reasonable approximation thereof, and a force field to quantify the interaction between all the atoms in the solid. The simplification of defining a
material structure in terms of the coordinates of the nuclei of the constituent

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

360

Rutledge

atoms follows from the BornOppenheimer approximation, which separates

the full quantum mechanical problem, based on relative mass and velocity
of the subatomic particles, into an electronic part and a nuclear part. The
electronic problem consists of describing the electronic state of the solid in
the field of stationary nuclei. The nuclear problem consists of describing the
state of the nuclei in the potential field created by the rapidly equilibrated
electrons. In atomistic simulation, attention is usually on the nuclear problem, which is used to specify the structure of the material, while the electronic energy is represented approximately by an analytical force field. In
amorphous materials, it often suffices to specify structure in terms of pair
correlations over short distances. Crystalline solids, on the other hand,
possess both short range and long range order; in the ideal crystal, the
long range order is limited only by the extent of the crystal itself. Long
range order may persist in a material in one or two dimensions, as in
liquid crystals and some rotator phases, or in three dimensions (true
crystals). Just as in atomic crystals, polymer crystals exhibit a periodic
structure, the lattice, which describes how unit cells repeat in space.
The literature on point and space group symmetries in crystals and the
role of the unit cell is extensive and dates back to the early development
of X-ray diffraction methods for solids. A complete review is beyond the
scope of this chapter, but the interested reader is referred to several excellent
introductory texts [13].
The atomic structure of the crystalline solid may be described completely
by the extent or size of the crystal, the geometry of the unit cell, and the
coordinates of the atoms in one unit cell (see Fig. 1). The unit cell may be of
arbitrary shape and size; the periodicity of the lattice is defined by the 3  3
matrix h whose columns are the three vectors a, b, and c defining the edges
of the unit cell. The entire solid is composed of such unit cells, and the extent
of the crystal along the directions specified by a, b, and c is given by the
indices h, k, and l, which take integer values between
1 and 1. Within
each unit cell, the coordinates of atom j are often expressed as fractions of a,
b, and c, respectively, which are collected into the vector sj, the fractional
coordinate of atom j. The coordinate of any atom i in the crystal, which we
designate generally as qi, may be expressed as the image coordinate of the jth
atom in the Lth unit cell:
qi qj, L hsj L
where
h a, b, c and

Lt h, k, l

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Modeling Polymer Crystals

361

FIG. 1 Definition of crystallographic parameters. The points designate equivalent

locations in successive helical repeat units. a, b, and c are the unit cell vectors; qj,0 is
the coordinate of the jth atom in the L 0 unit cell;  and d are the helix parameters.

The superscript t designates the vector transpose. The perfect periodicity of

the crystal lattice exists only in an average sense. Dynamically, each atom
oscillates about its position in the lattice so that at any instant the periodicity of the crystal may be broken. However, the persistence of long range
order implies that the forces on the atoms tend to restore each atom to its
lattice site. The relevant phase space (i.e., the set of positions and momenta
accessed by the atoms) of crystals, therefore, is considerably smaller in
comparison to fluid phases. In the crystal, most or all of the atoms are
localized about well-defined points in configuration space. Long range
motion of atoms in crystals may be quite significant, but usually consists
of discrete, activated jumps between what are essentially equivalent states of
the crystal. In some cases, atomic motion is more continuous, and description of the solid as a disordered crystal may be more appropriate (see
below).
Polymer crystals exhibit several characteristics not typically encountered
elsewhere in simulations. First, the unit cells of polymers tend to consist of
tens of atoms, while the molecules themselves comprise 104 to 106 atoms.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

362

Rutledge

Rather than having a situation where a unit cell consists of one or more
molecules, as in small molecule crystals, the situation is reversed: a single
molecule participates in many unit cells. This has significant ramifications
for the correlations, both structural and dynamical, between unit cells.
Second, the periodicity of the crystal lattice usually implies a periodicity
for the conformation of the polymer chain itself. This condition places
a severe restriction on the conformation space available to the chain in
a crystalline solid. Polymer chains in crystals are more appropriately
described as helices rather than coils. A helix conformation consists
of a regular repetition of the torsion angles characteristic of a small subsection of the chain, the helix repeat unit [4,5]. Each helix has an axis and a
handedness of rotation (the direction in which the chain backbone winds
around the helix axis). The helix itself is characterized by the translational
displacement parallel to the helix axis, d, and the angle of rotation about the
helix axis, , executed by a single helix repeat unit. The coordinates of Eq. (1)
may be expressed in terms of these helical parameters as follows:
qh, Lh , n Rh, n hsh hLh Tn

where
2

cosn !h

6
Rh, n 4 sinn !h
0
Lh h0 , k0 , 0

sinn !h 0

7
05

cosn !h
0

and Tn 0, 0, nd

1
t

Here, the subscript h indexes different helices in the unit cell. In this convention, the fractional coordinates for each unique helix in the unit cell are
specified with the helix axis directed along the z-coordinate and passing
through the origin. Each helix may have a distinct setting angle, !h, and
h0 and k0 may take fractional values to indicate helices at nonprimitive lattice
positions. The lattice index l is replaced by the helix repeat index, n.
Third, nearly all known polymer crystal polymorphs exhibit chain helices
packed into the solid with axes parallel.* As a result, there exists a unique
chain continuation direction, which gives rise to a highly anisotropic

*The  phase of isotactic polypropylene presents a notable exception. In this crystal form, the
polymer chains all have the same conformation, but are arranged in layers with an angle of
about 80 between the axes of chains in successive layers [6].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Modeling Polymer Crystals

363

material on the molecular scale. In the chain direction the crystal is held
together by chemical bonds, while in the other two directions it is held
together by weaker, chain-packing forces, such as van der Waals forces or
hydrogen bonds. This combination of conformational periodicity, throughcrystal chemical bonding and directionality of strong and weak interactions
makes polymer crystals unique.
To specify a starting configuration for any polymer simulation, a reasonably good estimate of the unit cell geometry, and in many cases restrictions
on the atomic coordinates s, may be obtained by wide angle X-ray diffraction. Estimates of the crystal size can be obtained by wide or small angle
X-ray scattering, electron or optical microscopy. One of the unique features
of crystalline polymers is the chain folded lamellar morphology, common
among flexible polymers. These plate-like crystallites are the result of kinetic
factors that dominate during crystallization itself. The thicknesses of such
lamellae are also considerably less than the contour length of the chain
molecules. The ramification of this length mismatch is that the surfaces of
such lamellae involve reentry of the same molecule into the crystal, in the
form of loops or folds. The thinness of the lamellae and the molecular
connectivity of the loop or fold surfaces may have a significant effect on
the dynamical behavior of real polymer crystals. Detailed atomistic studies
of entire fold surfaces in polymer crystals are relatively few [7,8].
More recently, atomistic simulations of the crystal/amorphous interface of
a semicrystalline polymer matrix have been presented that provide insight
into the preferred topology for chain reentry along several relevant crystallographic facets [911]. The presence of disordered surfaces or interfaces
may affect the relative stability of finite-sized crystallites. Nevertheless, the
lowest energy forms of large polymer crystals are expected to be those with
fully extended chains (i.e., no fold or loop surfaces), and it is these that have
generally been the object of simulation.
Finally, a few comments about the uniqueness of polymer crystal structures and phase space localization are warranted. Almost all crystallizable
polymers exhibit polymorphism, the ability to form different crystal structures as a result of changes in thermodynamic conditions (e.g., temperature or pressure) or process history (e.g., crystallization conditions) [12].
Two or more polymorphs of a given polymer result when their crystal
structures are nearly iso-energetic, such that small changes in thermodynamic conditions or kinetic factors cause one or another, or both, to
form. Polymorphism may arise as a result of competitive conformations of
the chain, as in the case of syndiotactic polystyrene, or as a result of competitive packing modes of molecules with similar conformations, as in the
case of isotactic polypropylene. In some instances, the conformational
change may be quite subtle; isotactic polybutene, for example, exhibits

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

364

Rutledge

three crystal polymorphs in which the number of constitutional repeat

units per turn of the helix varies only from 3 to 4. Typically, the thermodynamic transition between polymorphs is discontinuous; each polymorph
represents a distinctly different local region in the total phase space of the
polymer. However, polymer crystals may also exhibit internal forms of
disorder within a particular unit cell geometry. Common types of disorder
include (i) the statistical, iso-energetic packing of helices of opposite handedness or axis direction within the unit cell, (ii) conformationally disordered chains which are not true helices, but nevertheless satisfy
approximately the lattice periodicity along the chain direction, (iii) constitutional disordered chains composed of copolymers with irregular monomer sequences, and (iv) thermal disorder due to rapid interconversion
between iso-energetic conformational or packing states. Examples of the
last include free rotation of the pendant methyl group in isotactic polypropylene or the rotation of entire chains in polyethylene (i.e., rotator
phases) at elevated temperatures. Algorithms must be evaluated carefully
when dealing with different types of disorder.

III. COMPUTATIONAL METHODS

We distinguish between two classes of calculation or simulation appropriate
for the study of polymer crystals: optimization methods and sampling
methods. Optimization methods take explicit advantage of the localization
of crystals in phase space; numerical optimization techniques are used to
identify and characterize a specific region of phase space. Molecular
mechanics and lattice dynamics fall in this group. Sampling methods generally do not take advantage of, and thus are not restricted by, assumptions
of localization. Monte Carlo and molecular dynamics fall in this group.
Each of these computational approaches is considered below. Since optimization methods, particularly lattice dynamics, are used almost exclusively
for crystalline solids, these will be presented in some detail. Sampling methods are discussed to the extent that their use for simulation of polymer
crystals requires some special considerations.

A. Optimization Methods
Consider an ideal case where the crystal exhibits none of the forms of disorder mentioned in the previous section. A given polymorph is then characterized by localization about an equilibrium point in phase space where
the potential energy surface is concave. The relevant part of the energy

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Modeling Polymer Crystals

365

surface can be approximated to arbitrary accuracy using a Taylor series

expansion in small displacements about the local minimum energy
structure, q0:


3N
3N X
3N
  
   X


@U 
1X
@2 U 
U qi U qi0
 qi
qi0 2
 qi
qi0 qj
qj0
@q
@q
@q
i 0
i j 0
i1
i1 j1

3N
3N
3N


1 X X X @3 U 

qi
qi0 qj
qj0 qk
qk0   


6 i1 j1 k1 @qi @qj @qk 0
3a
or in tensor notation:
 
1
U q U q0 qt  rU0 qt  r2 U0  q   
2
1
X
qt  ri

U0
i!
i0

3b

Here, q represents the 3N vector of coordinates for a crystal composed of N

atoms and q q
q0 is the displacement from the local minimum. There
are 3N
6 molecular degrees of freedom in s and 6 lattice degrees of freedom
in h. In classical mechanics, in the limit that temperature T approaches 0,
q also approaches 0. In this limit of small displacements, the energy and
geometry of the crystal may be approximated well by U(q0) and q0, respectively. The problem of computing the properties of a crystal phase simplifies
in this case to identification of a single point in configuration space where
the structure is in mechanical equilibrium, characterized by all the forces on
the atoms, rU0, equal to zero and all the second derivatives rrU0 positive,
indicative of a stable point on the potential energy surface. This kind of
numerical optimization problem goes by the name of molecular mechanics.
Standard techniques [13] such as steepest descent, conjugate gradient, or
modified Newton methods may be used to find the minimum in U(q) as a
function of the degrees of freedom q {h, L, s}. For large crystals, since
every unit cell in the crystal is identical, it suffices to minimize U(q) with
respect to q {h, s} with L 0 (i.e., the degrees of freedom of a single unit
cell), provided that care is taken to evaluate U(q) over a sufficiently large
number of unit cells, using techniques such as Ewald summation [14]. It is
important that the conditions of both zero gradient and positive second
derivatives be met, in order to ensure that the structure is indeed a stable
one, and not a saddle point (cf. Section IV). Taking the limit of zero

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

366

Rutledge

displacement rules out any explicit thermal (e.g., entropy) or kinetic contributions to the structure and properties of the crystal. However, this does
not imply that the calculation is necessarily representative of a crystal at 0K.
Thermal contributions may creep into the calculation implicitly, through the
parameterization of the force field or the choice of lattice parameters h.
The relatively simple molecular mechanics calculation has proven very
useful in the study of polymer crystals. It is generally the method of first
resort for any new crystal study. The primary result of such a calculation
is structural information on a very detailed level, for purposes of gaining
insight into the balance of forces or obtaining structure factors to compare
with experiments. For very large problems, the calculations may be speeded
up through the use of one or more common simplifications invoked to
reduce the length of the vector q. First, the problem of determining the
helical conformation of the chain may be decoupled from that of the intermolecular packing structure. The chain conformation having the lowest
intramolecular energy is determined first using isolated chain calculations,
appropriately constrained to ensure helicity. (This calculation can usually be
accomplished simply by setting the lattice vectors a and b to be very large.)
Subsequently, the energy of the crystal is minimized with respect to the
intermolecular degrees of freedom and with the chains themselves treated
as rigid molecules having only three translational and one rotational degree
of freedom per chain [1517]. However, such rigid chain calculations do not
reflect the effects of strong intermolecular packing interactions on conformation. Packing interactions have been shown in some cases to distort
the conformation of the chain [18,19] or even to stabilize a crystal structure containing a completely different and less stable conformation [20].
Second, the simultaneous optimization of intramolecular and intermolecular
degrees of freedom may be retained at minimal cost by taking advantage of
known symmetries within the unit cell, usually deduced from X-ray results
[21,22]. In crystallography, the smallest group of atoms that contains no
internal symmetry is called the asymmetric unit and, in many instances,
comprises a fraction of the total number of atoms in the unit cell. In this
way, the number of degrees of freedom is reduced to the coordinates of the
asymmetric unit, the lattice vectors of the unit cell, and the symmetry operations defining the coordinates of the remaining atoms within the same
unit cell.
Thermal contributions to the properties of the crystal are introduced
by taking higher order terms in the Taylor series expansion [Eq. (3)] for
the energy surface in the vicinity of the equilibrium structure. The classical
theory of lattice dynamics was originally formulated for crystals by Born
and von Karman [23,24] and later presented in the comprehensive text of
Born and Huang [25]. There are several excellent texts that discuss lattice

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Modeling Polymer Crystals

367

dynamics in great detail [2528]; here, we will only summarize the main
points.
In the harmonic approximation, all contributions to the energy beyond
second order in atomic displacements are neglected. Since the gradient rU0
evaluated at equilibrium is zero, this yields:
1
Uq Uq
Uq0 qt  rrU0  q
2
3N X
3N 2 
X
1
@ U

qi qj
2 i1 j1 @q@q

Differentiating both sides of Eq. (4) with respect to qi, one obtains an
equation for the force @U/@qi acting on coordinate qi, and hence the
equations of motion describing oscillatory displacement of the atoms
about their equilibrium positions:

mi


3N
X
d 2 qi t
@2 U 

qj t
dt2
@qi @qj 0
j1

The periodicity of the crystal permits description of the displacement of

atoms in the form of traveling waves with wavevector k. Solution for the
oscillatory displacement of coordinate i takes the form:


qi t ui exp ik  qj
!kt

where ui is the amplitude of displacement of coordinate i and !(k) is the

frequency of oscillation. There are 3N such solutions. The harmonic motion
of each atom is, in general, coupled with the motions of the other atoms in
the crystal. However, there exists a particular set of linear combinations of
these solutions, called normal modes, which are independent of each other.
Substitution of Eq. (6) into Eq. (5) yields:

3N
X


@2 U 
m i ui !
 exp
ik  qj
qi uj
@q
@q
i j 0
j1
2

This equation can now be expressed in terms of the displacements of the Nu

atoms of the unit cell, rather than the total number of atoms of the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

368

Rutledge

entire crystal. In matrix form, this becomes:

D
!2 Iu0 0

where the vector u0 consists of the 3Nu mass-weighted elements of one unit
p
cell, u0i ui mi , and D(k) is the 3Nu  3Nu mass-weighted dynamical
matrix, whose elements are defined as:




P @2 U 
exp ik  h sj
si L

L @qi @qj 0
Dij k
p
mi mj

The sum over L corresponds to a triple summation over all values of h, k,

and l in the crystal. The second derivatives of potential energy with respect
to the components of q are obtained from the converged Hessian matrix
at the potential energy minimum. The roots of Eq. (8) yield the frequencies
!i(k) for the 3Nu independent 1-dimensional oscillators, or normal modes
of vibration. Associated with each frequency is a polarization vector mi(k)
which is a linear combination of the original mass-weighted atomic displacements u0 . The polarization vectors are orthonormal, such that
mti m
j ij (here, the superscript * denotes the complex conjugate). These
are found by diagonalizing the matrix D(k):
m
1 Dm ,

10

,(k) is the diagonal matrix whose elements are the 3Nu values !i2(k). D(k) is
Hermitian, i.e., D (D*)t, so values of !2i are always real and positive for
solution about a minimum in energy. It is apparent that the set of frequencies !i(k) is a function of the wavevector, k. However, the periodicity of the
lattice reduces the set of unique wavevectors to those located in the first
Brillouin Zone (BZ), the region of reciprocal space whose boundaries lie
halfway between reciprocal lattice points, eg.,
/a<k  /a along the a*
dimension, and similarly for b* and c*; a* (b  c)/V, b* (c  a)/V and
c* (a  b)/V are the reciprocal lattice vectors, while V (a b  c) is the
unit cell volume. The k-dependence of the 3Nu values !i(k) is called a dispersion curve, of which there are 3Nu branches in the BZ. Of these 3Nu
branches, three are acoustic modes, which have the property that !i
approaches zero as the magnitude of k approaches zero; the other 3Nu
3
branches are called optic modes. Furthermore, most polymer crystal
structures have more than one chain participating in the unit cell. Each
chain has four molecular degrees of freedom (three translational, and one
.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Modeling Polymer Crystals

369

rotational about the chain axis), giving rise to 4Nchain
3 molecular optic
modes, which tend to be lower in frequency than the remaining
3Nu
4Nchain atomic optic modes. For the finite but unbounded crystal
having hmax unit cells along a, kmax unit cells along b, and lmax unit cells along
c, and periodic boundary conditions, the wavevectors along each branch are
restricted to discrete values: k 2ha
=hmax kb
=kmax lc
=lmax ,
where h takes integer values running from
hmax/2 to hmax/2, and similarly
for k and l, for a total of 3N vibration frequencies in the finite crystal. In the
limit that the crystal becomes infinitely large, the interval between allowed
wavevectors goes to zero, and both !i(k) and i(k) become continuous functions of k.
Each of the normal modes in the harmonic approximation is a traveling
wave in the crystal and represents a packet of energy; by analogy with the
wave/photon duality of electromagnetic radiation, the wave packets in crystals are called phonons. The energy of each phonon can be quantified either
classically or quantum mechanically. Since the normal modes are independent, the vibrational partition function for the crystal is simply the product
of the individual partition functions for each phonon. These are known
analytically for harmonic oscillators. The corresponding classical and
quantum mechanical partition functions for the crystal are:
classical :

quantum :

ln Qvib
cm

ln Qvib
qm

Z
dk
BZ


3Nu
X
h!
 i k
ln
kB T
i1

11a

#
3Nu
X
h!
 i k

h!i k
ln 1
exp


dk
2kB T
kB T
BZ
i1
Z

"

11b
RThe free energy is thus computed explicitly (Nk is the normalization factor
BZ dk):
 
  kB T


ln Qvib
Ah, T U q0 Avib q0 , T U q0

Nk

12

The size of crystal simulated in this fashion is determined by both the

range of interactions included in the computation of the potential U(q), and
the resolution of wavevectors in the BZ. As mentioned earlier, for periodic
systems like crystals, Ewald summation permits the efficient evaluation of
interactions for a material of infinite extent. Evaluation of thermodynamic
properties for an infinite crystal requires accurate evaluation of integrals

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

370

Rutledge

involving the dispersion relations, !i(k). Standard quadrature methods such

as GaussLegendre are used to get the highest accuracy with the fewest
number of integration points. Generally, different densities of integration
(mesh) points should be tried, to ensure convergence. Some quantities, such
as the atomic displacement tensor discussed below, may be dominated by
particular subregions of the BZ, and require special attention to precise
evaluation in these regions.
In the study of polymer crystals, lattice dynamical calculations have been
implemented in different ways. Straightforward minimization of the potential energy with respect to both fractional coordinates s and lattice parameters h simultaneously yields a single, temperature-independent structure.
The resulting normal modes are thus also temperature-independent, giving
rise to a strictly harmonic crystal [2931]. In the quasiharmonic approximation (QHA), vibrational frequencies for any lattice h are equated with those
of the harmonic crystal at the state of strain defined by h. QHA thus allows
for anharmonicities that arise due to changes in size or shape of the unit cell.
Early implementations of lattice dynamics with QHA relied on experimental
X-ray diffraction measurements for appropriate values of h at finite temperatures; the potential energy was then minimized with respect to internal
degrees of freedom s at fixed external strain, e(h, h0), referred to as the zero
static internal stress approximation (ZSISA) [32]. Reviews of these types of
calculations for numerous polymers are available [21,22]. A more predictive
and self-consistent implementation of QHA lattice dynamics, within the
ZSISA, is possible at little additional effort by minimizing the potential
energy with respect to s at fixed h; the vibrational contribution to free
energy is then computed and the total free energy A(h,T ) is minimized
with respect to the external variables in h, under the condition that potential
energy is minimal with respect to s for each choice of h. The relative contribution of Avib to A varies with temperature, resulting in a shift towards
lower frequencies, and hence larger h in most cases, at higher temperature;
the contraction observed along the chain axis in polymer crystals which
possess fully extended chains is an interesting exception. The shift in lattice
frequencies for polyethylene between 0K and 300K is shown in Fig. 2. This
shift gives rise to thermal expansion behavior. This self-consistent approach
permits estimation of material properties based on the force field alone,
without additional input from experimental data. It thus provides a powerful means for checking the validity of a force field. This self-consistent
version of lattice dynamics has also been applied to numerous polymers
[19,3336]. Finally, it is possible within the QHA to circumvent the
ZSISA, by formulating analytical derivatives of the normal mode frequencies with respect to both s and h, and then minimizing the total free
energy with respect to all degrees of freedom, both internal and external,

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Modeling Polymer Crystals

371

FIG. 2 4Nchain lattice (acoustic and external optic) normal mode branches for
polyethylene along the [100] lattice direction. Each branch is symmetric about the
 (k 0) point. 0K (dashed lines) and 300K (solid lines).

simultaneously [37]. This can be important for quantities that depend sensitively on internal strains. To date, only properties for polyethylenelike crystals described by relatively simple potentials have been obtained
by this method and evaluating the analytical derivatives at the static
geometry [38].
Although we have not discussed the specific form of the potential energy
function used to describe U(q), other than to suggest that it is an effective
potential described parametrically by the positions of atomic nuclei, it is
common to treat electrostatic forces in atomistic calculations using partial
atomic charges assigned to the various nuclei. However, this description for
electrostatics does not account for distortion of the electron distribution (the
electronic part of the BornOppenheimer separation) in the environment
of the crystal. The presence of dipoles in the anisotropic crystal can lead
to significant mutual induction, or polarization, effects. The lattice dynamical description presented up to this point can be easily extended to
include mutual induction through use of the shell model for atomic
polarizability [39]. This model introduces an additional 3N coordinates to
q, corresponding to the center of action of the electron distribution about
each atom, which may polarize the atom by situating off-center from the
nucleus itself. This results in an additional 3N normal modes, called plasmon modes, which are considerably higher in frequency than phonon
modes. The shell model has been used to study poly(vinylidene fluoride)

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

372

Rutledge

using the harmonic [40] and self-consistent quasiharmonic approximations

[35,41].
The free energy of the crystal corresponding to other ensembles may be
computed through inclusion of additional work terms in the equation for
free energy, for example:
Gr, T Uh Avib h, T
V0 rt  e

13

For mechanical work, p is the applied stress and e is the deformation strain,
defined with respect to a reference cell h0. For a cell deformed from h0 to h,
the (external) strain may be computed as:
e


1  t,
1
1
h Gh0
I
2 0

14

G is the metric tensor hth and V0 is the cell volume. A wide range of thermal
and mechanical properties may be estimated through optimization of
A(h,T ) or G(p,T ) at different temperatures and states of deformation.
Among these, the constant volume heat capacity may be computed using:

CV

kB
Nk

Z
dk
BZ

3Na
X

x2 ex ex
1
2

i1

where

h!
 i k
kB T

15

Thermal expansion coefficients may be computed as thermally induced

strains:

1 t
1 dG
1
h
a h0
2
dT 0

16

Similarly, elastic stiffness and compliance constants can be computed using

the relations
C

@r
@e

i:e:,

Cij

@i
@"j

17a

and
S C
1

@e
@r

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

17b

Modeling Polymer Crystals

373

or
C

1 @2 Ah, T
V0 @e@e

and

Cij

1 @2 Ah, T
V0 @"i @"j

17c

Here, p and e are the 6  1 vectors for stress and strain, respectively,
in Voigt notation (i.e., "i "ii for i 1,2,3; "4 2"23, "5 2"13, "6 2"12).
Lattice Gruneisen coefficients offer a measure of the thermal stress on a
crystal:
c

V
Ca
CV

18

The anisotropy typical of the polymer crystal is exemplified by the material

properties of several polymer crystals studied to date. These are shown
in Table 1.
The elastic modulus of the crystal in the chain direction is one of the most
important properties of a semicrystalline polymer and has received considerable theoretical attention. It also serves to illustrate some of the differences between various methods of calculation. Figure 3 illustrates the results
for C33 for polyethylene from several different calculations: potential energy
minimization, quasiharmonic lattice dynamics, Monte Carlo and molecular
dynamics. The open symbols in Fig. 3 are all indicative of optimizationbased calculations, while the filled symbols are derived from sampling-based
methods (cf. Section III.B). For clarity, only the results from lattice
dynamics using classical vibrations are shown; calculations on polyethylene
using quantum mechanical vibrations typically yield lower values of C33 at
low temperatures, but both classical and quantum mechanical estimates of
elastic constants converge around 300K [33,45,46]. However, other properties, notably thermal expansion coefficients and heat capacity, have been
shown to exhibit quantum mechanical effects well past 300K [33]; vibration
modes with frequencies greater than 200 cm
1, predominantly intramolecular vibrations, are highly nonclassical at room temperature. The anisotropic
nature of the crystal precludes the kind of orientational averaging, present
in amorphous simulations, which helps to mitigate the error associated with
treating these high frequency vibrations classically. Thermal vibration
accounts for about 12%, a 40 GPa drop, in the estimates of modulus at
300K. Above 300K, the lattice dynamics estimates fall off more rapidly.
Inspection of Eq. (17c) reveals that the drop in modulus may be traced to
three interrelated factors: (i) an increase in reference cell volume with temperature; (ii) a change in the potential energy surface with increasing volume

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

374

Rutledge

TABLE 1 Properties of Several Polymer Crystals Computed at 300K Using a

Self-Consistent QHA Lattice Dynamics Calculation, with Quantum Mechanical
Treatment of Vibrational Free Energy

CV (kcal mol
1 K
1)f

1(10
5 K
1)
2(10
5 K
1)
3(10
5 K
1)
1
2
3
C11 (GPa)
C22 (GPa)
C33 (GPa)
C44 (GPa)
C55 (GPa)
C66 (GPa)
C12 (GPa)
C13 (GPa)
C23 (GPa)

PEa

iPPb

PVDFc

PPTAd

PETe

0.0196
28.3
10.7

2.4
1.2
0.97

2.3
6.0
7.3
280
2.7
1.7
2.9
3.4
8.7
9.5

0.00703
8.2
8.7

0.84
0.68
0.79
0.39
9.9
10.8
55.1
4.0
4.9
1.7
3.4
7.1
4.4

0.0306
6.56
5.13

0.38
0.79
0.66

0.040
24.5
26.1
276
na
na
na
1.2
7.6
9.1

0.1145
7.9
2.9

0.57
0.76
1.07

0.24
11.9
32.2
284
14.3
0.5
6.1
11.1
10.5
14.2

0.0470
11.4
4.12

1.07
1.13
1.02

0.69
14.4
17.3
178
6.6
1.4
1.2
6.4
3.4
9.5

PE (polyethylene), orthorhombic. Source: Ref. 33.

iPP (isotactic polypropylene), -phase, monoclinic (not all properties shown).
Source: Ref. 34.
c
PVDF [poly(vinylidene uoride)], -phase, orthorhombic. Source: Ref. 41.
d
PPTA [poly( p-phenylene terephthalamide)], orthorhombic. Source: Ref. 19.
e
PET [poly(ethylene terephthalate)], triclinic (not all properties shown). Source:
Ref. 36.
f
Per mole of unit cells.
na: not available.
b

to a surface of lower curvature; (iii) the reduction in vibrational frequencies

and their contribution to elastic stiffness. Comparison between lattice
dynamics and Monte Carlo results for a comparable force field (open and
filled circles, triangles in Fig. 3) indicates reasonable agreement at low temperature, followed by a breakdown of the quasiharmonic approximation at
high temperatures. Results for different force fields, on the other hand, may
differ by as much as 50100 GPa (e.g., diamonds vs. circles or triangles in
Fig. 3).
Thermoelectrical and electromechanical properties may be computed
in similar fashion by including an electrical work term,
V0Ed, in Eq.
(13). E is the applied electric field and d is the resultant electric displacement,

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Modeling Polymer Crystals

375

FIG. 3 C33 axial modulus of polyethylene obtained using different force fields and
computational methods. All calculations are classical. Open symbols are from
optimizations (circles, triangles, and diamond are self-consistent QHA lattice
dynamics, all others are plotted at temperature of experimental lattice parameters;
potential energy minimizations are plotted at 0K). Filled symbols are Monte Carlo
(circles) or molecular dynamics (diamond, triangles). Force field of Ref. 55: open and
filled circles (from Ref. 46); force field of Ref. 30: open (from Ref. 33) and filled
(from Ref. 68) triangles; pcff force field of Biosym/MSI: open and filled diamonds
(from Ref. 66);  (from Ref. 42); (from Ref. 43); open inverted triangle (from Ref.
44); open square (from Ref. 29). Curves are drawn as guides to the eye.

related to the polarization of the crystal. For poly(vinylidene fluoride),

piezoelectric and pyroelectric properties were computed without resort to
an electric field term by computing the change in polarization with temperature and strain [40,41,47].
Wide angle X-ray diffraction provides an experimental probe of atomic
scale structure through the structure factor F(S), which is defined as:
FS



fi STi S exp
2iSt qi

19

Here, S (S0 S00)/l and |S| 2sin
/

~ S refers to the scattering vector of
X-ray analysis, with S00, and S0 being the direction of incident and scattered
radiation, respectively. Here, l is the radiation wavelength and
is the
scattering angle. fi(S), Ti(S), and qi are the atomic structure factor, the
DebyeWaller temperature factor, and the position, respectively, of atom
i. The intensity observed in an X-ray experiment is related to the power

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

376

Rutledge

spectrum of F(S) by the relation: I(S) I0F(S)F*(S). I0 is a prefactor that

accounts for such effects as absoprtion, polarization, and Lorentz scattering
[13]. Taking advantage of the separation of coordinate qi into a unit cell
translation and placement of the atom within its unit cell, as suggested by
Eq. (1), one obtains:
FS


X


exp
2iSt hL
fj STj S exp
2iSt hsj

20

For large, perfect crystals, the first summation contributes a constant prefactor Fc (hmaxkmaxlmax)2 and the scattering vector interferes constructively
(i.e., is nonzero) only for values corresponding to vertices of the reciprocal
lattice: S ha* kb* lc* h*L. The static structure factor then becomes
simply:
FS Fc



fj STj S exp
2iLt sj

21

The signature of thermal motion is observed experimentally in X-ray

diffraction as a change (usually a reduction) in F(S), and hence in the
observed intensity. The structure factor for each atom is modified by the
factor Ti(S) exp(
22StBiS), which depends on both the magnitude and
orientation of S. The crystal dynamics are embodied in the direction and
amplitude of normal mode vibrations. The amplitude of vibration depends
on the energy content of that phonon, Evib,j kBT2(@lnQj/@T). The orientation of displacement is described by the polarization vector xj for that mode.
From this, one can compute the probability density function for an atom
vibrating under the influence of a single normal mode. Since the phases of
the modes are independent, the average displacement for atom i due to the
action of all normal modes can be written as:


Bi ui uti

1
Nk Nu mi

Z
dk
BZ

3Nu
X
Evib, j k
j1

!2j k


t
i, j k i,
j k

22

A plot of the orthorhombic unit cell for polyethylene is illustrated in Fig. 4,

with the 50% probability density surfaces shown for each atom of the
unit cell. The anisotropic dynamics of the polyethylene crystal are
immediately apparent. Significantly, the largest contributions to Bi come
from the long wavelength acoustic modes, due to the inverse dependence
on !2. In experimental studies on polymers, the temperature factor is often

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Modeling Polymer Crystals

377

FIG. 4 Orthorhombic unit cell of polyethylene, with 50% probability thermal

vibration ellipsoids for each atom, computed using the force field of Ref. 55 with
8000 normal mode vibrations sampled in the Brillouin Zone. Thermal vibration
ellipsoids were plotted using Ortep-III [48].

treated as isotropic and equivalent for all atoms; in this case the corresponding B-factor is simply 82(tr(Bi)), averaged over all the atoms of the
unit cell.
The principal limitations of lattice dynamics, of course, are the shape of
the potential energy surface implied by the harmonic approximation and
the restriction to localized phase spaces. When polymorphism is an issue,
each polymorph may be examined independently using lattice dynamics.
However, crystals with internal disorder at elevated temperatures include
significant contributions from configurations that are not accurately represented by the quadratic form of the Taylor series expansion for potential
energy. Studies of atomic crystals, as well as a recent study of polyethylene
[34,49], suggest that QHA is reasonably accurate up to about 2/3 of the
melting temperature, barring the presence of solidsolid transformations at
lower temperatures. Higher order anharmonic contributions may be treated
methodically through the use of perturbation theory [5053]. However, the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

378

Rutledge

improvement usually comes at considerable computational expense. For

these situations, sampling methods such as Monte Carlo or molecular
dynamics are usually preferred.

B. Sampling Methods
Both Monte Carlo and molecular dynamics methods sample directly the
phase space of a small but representative component of the crystal, the
former by performing stochastic moves through configuration space,
the latter by following a specified trajectory according to an equation of
motion and chosen initial condition. A typical Hamiltonian for molecular
dynamics simulation is [14]:

Hp, q

3N
X
p2i
Uq
2mi
i1

23

The Hamiltonian is a constant of the phase space sampled by the simulation.

A mechanical property is one that is defined at each point in phase space,
i.e., it can be computed for each configuration of the system. The macroscopic observable Fobs is then computed as the ensemble average (Monte
Carlo) or time average (molecular dynamics) of the mechanical property Fi;
according to the ergodic hypothesis, these should be equivalent.
XN

Fobs  hFi

Fi e
Ei
lim Xi1
N
N!1
e
Ei
i1

1
lim
t!1 t

dt0 Ft0

24

Combinatorial quantities such as entropy and free energy, which depend on

the entire distribution of states sampled, require further effort to extract.
Methods based on thermodynamic relations which express these quantities
as integrals over ensemble averages of mechanical quantities, e.g., @H(l)/@l,
where l defines the state at which the Hamiltonian is evaluated, are most
often used to extract thermal properties [14]. l may be temperature, volume,
or even a change in the force field itself.
Z

  0 
@H l
dl
@l0
0

Al2
Al1
1

25

Since much of this book is dedicated to the description of powerful Monte

Carlo and molecular dynamics algorithms for the study of chain molecules,

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Modeling Polymer Crystals

379

we will not go into these methods in as much detail as with lattice dynamics.
Nevertheless, a few comments are in order.
By sampling phase space directly, the methods of Monte Carlo and
molecular dynamics can avoid the restrictive approximations of lattice
dynamics. In principle, at least, polymorphism, anharmonicity, static and
dynamic disorder may all be rigorously captured by direct simulation.
However, other limitations arise which differ from those encountered in a
lattice dynamical approach. Comparison of both approaches can provide
insight into the significance of these limitations [46].
First, both Monte Carlo and molecular dynamics simulations must
approximate the macroscopic crystal using a very small number of atoms,
typically on the order of 103104 with current computers. With periodic
boundary conditions, atoms outside of the simulation box are restricted
to remain dynamically in phase with those inside the simulation box. For
a simulation cell consisting of hmax  kmax  lmax unit cells, only those wavevectors whose magnitudes are integer multiples of 2a*/hmax, 2b*/kmax,
and 2c*/lmax are sampled in the simulation. Thus, unless hmax, kmax, and
lmax are chosen suitably large, the dynamical resolution is rather poor. This
finite size effect may affect both lattice properties, such as thermal expansion, and atomic displacements, which as noted previously are most affected
by long wavelength acoustic modes. For polyethylene, finite size effects are
noticeable in small simulations containing 2  3  6 unit cells (432 atoms) at
temperatures around 100K, but become progressively worse at higher temperatures [54,55]. Furthermore, finite size effects in polyethylene were found
to be quite anisotropic, being largest along the b direction of the lattice, but
also strongly coupled to the lattice size in the chain direction. Free or rigid
boundary conditions [56] may be used instead of periodic boundaries, but
these incur surface as well as finite size effects, the implications of which are
less thoroughly quantified.
The anisotropy of polymer crystals has other implications for simulation,
as well. Along the chain direction, intramolecular degrees of freedom may
dominate the dynamical behavior, while dynamics lateral to the chain are
predominantly intermolecular in nature. As evidenced by the large separation between internal and external mode frequencies in lattice dynamics,
these two groups are largely decoupled. In molecular dynamics, this separation of dynamical time scales manifests itself as an inefficient transport of
energy between high frequency and low frequency motions. This may be
observed as poor equilibration of kinetic energy between intramolecular and
intermolecular degrees of freedom. To overcome this problem, Ryckaert
and Klein used massive stochastic collisions [57] in their molecular dynamics
simulation [58]. The collisions serve to redistribute the kinetic energy among
all the modes of motion. In Monte Carlo simulations, this effect is observed

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

380

Rutledge

as inefficient sampling of local coordinate s and cell h degrees of freedom

simultaneously. Since the problem is one of sampling internal and external
degrees of freedom, most polymer algorithms geared towards efficient simulation of conformation space are of little use in this respect. Instead, one
introduces a combination of atomic and molecular moves. Atomic moves
consist of small displacements of individual atoms, as was done in early
simulations of atomic liquids and solids, or of small groups of bonded
atoms. This permits rapid sampling of high frequency, intramolecular
degrees of freedom. Molecular moves consist of translational displacements
of the entire chain as a rigid body, as well as rotations of chains about their
axes during a simulation. This permits rapid sampling of lower frequency
lattice degrees of freedom [54,5962]. Displacements on the order of 0.05 A
and 0.2 A are typical for atomic and molecular moves, respectively, and
510 degrees for chain rotation. For sampling over large energy barriers
in crystals with some disorder, molecular moves which are some fraction
of the lattice spacing, or which involve a symmetry operation in the crystal
have been used [61,62].
A third, less obvious limitation of sampling methods is that, due to the
heavy computational burden involved, simpler interatomic potential models
are more prevalent in Monte Carlo and molecular dynamics simulations.
For example, polarizability may be an important factor in some polymer
crystals. Nevertheless, a model such as the shell model is difficult and timeconsuming to implement in Monte Carlo or molecular dynamics simulations
and is rarely used. United atom models are quite popular in simulations of
amorphous phases due to the reduction in computational requirements for a
simulation of a given size. However, united atom models must be used with
caution in crystal phase simulations, as the neglect of structural detail in the
model may be sufficient to alter completely the symmetry of the crystal
phase itself. United atom polyethylene, for example, exhibits a hexagonal
unit cell over all temperatures, rather than the experimentally observed
orthorhombic unit cell [58,63]; such a change of structure could be reflected
in the dynamical properties as well.
The properties of polymer crystals may be computed using the same
relations as presented earlier for lattice dynamics. However, it is generally
more expedient to take advantage of the fluctuation-dissipation theorem
[64] to relate response functions to the magnitude of fluctuations at equilibrium. For example, the fluctuation formula for the constant volume
heat capacity computed in the canonical (NVT) ensemble is


E 2
h E i2
CV
kB T

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

26

Modeling Polymer Crystals

381

where it is understood that the energy E includes both kinetic and potential
contributions. A similar useful relation exists for the elastic stiffness moduli
(in the NT ensemble) [65]:
C

kB T t 
1
""
hVi

or

Cij

kB T 
1
"i "j
hVi

27

As indicated earlier, accurate calculation of the elastic moduli by this formula requires a simulation cell large enough to sample adequately the lattice
modes of vibration. For polyethylene and polypropylene, simulations consisting of 4  6  12 unit cells and 3  1  3 unit cells, respectively (corresponding to simulation cells approximately 20 A on a side in each case) have
been used with reasonable success [66]. Whereas simulations on the order of
104 to 105 Monte Carlo or molecular dynamics steps are typically required
to obtain good averages for energy, density, and many other properties,
fluctuation averages such as those used in Eqs. (27) typically require simulations one to two orders of magnitude longer in order to achieve accuracies
around 5% [67,68]. This has been attributed to slow convergence of high
frequency motions [68], as well as the difficulty in sampling external variables mentioned above. A formula that takes advantage of the correlated
fluctuations of stress and strain has been proposed to improve the convergence characteristics [67].

IV. CRYSTAL IMPERFECTIONS AND

RELATED PROCESSES
Up to this point, we have focused on the restrictive nature of the phase space
of polymer crystals. The harmonic approximation of lattice dynamics
assumes that atoms execute only small displacements from their equilibrium
positions, such that the effects of anharmonicity are low. Monte Carlo and
molecular dynamics, of course, make no such approximations. Many processes in polymer crystals involve significant rearrangements of atoms and
passage through a (series of) metastable configuration(s). These processes
include relaxations, mass transport, and plasticity, as well as lamellar thickening in flexible chain polymer crystallites. Molecular simulations can be
used to study the mechanisms by which these processes occur. The mechanism may be characterized by the (one or more) trajectories through phase
space which carry the crystal structure from a reactant state, belonging to
the region of space occupied before the process, to a product state,
belonging to the region of space occupied after the process. Short of complete melting and recrystallization of the polymer, the process must occur
through the action of a more or less localized region of disorder in the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

382

Rutledge

crystal, which constitutes a crystallographic defect. The type of simulation

used to analyze the process depends significantly on the nature of the
trajectory.
For trajectories involving configurations with energies on the order of
the thermal energy, kBT, a molecular dynamics simulation can efficiently
sample configuration space both close to an equilibrium point as well as
along the (classical) trajectories between equilibrium points. The great
power of this approach is that little or no prior knowledge of the trajectory
is required, so long as the energy requirements are low and passage along
the trajectory is sufficiently frequent to ensure observation of one or more
passages during the typical 1 ns simulation. A rule of thumb based on transition state theory (see below) is that the potential energy along the trajectory does not exceed about 10 kBT at any point.
Wunderlich and co-workers have reported extensively on the dynamics of
a united atom model of polyethylene using molecular dynamics [63]. In this
model, the crystal structure resembles the high temperature hexagonal phase
observed in polyethylene. Simply by analyzing the structure and energy of
configurations sampled in a conventional molecular dynamics simulation,
one can identify crystallographic defects, conformational relaxations, and
diffusion processes. With the united atom model, these authors observed
that gauche conformational defects formed in the otherwise all-trans chain
with a frequency on the order of 10 GHz per bond at 350K. The estimated
activation energy is on the order of 58 kBT. These gauche states were
observed to be short-lived and uncorrelated. Kink defects (i.e., having conformation gtg) were observed to diffuse rapidly and annihilate at the free
surface of the simulation. Sequential deviation in the same direction by
a string of bonds along the chain gives rise to a twist, by which the
chain may execute a 180 degree flip about its axis, with a passage time on
the order of 2 ps at 300K. Longitudinal diffusion of the chain was observed
to couple to a lattice acoustic mode. All-atom simulations revealed the onset
of translational diffusion in polyethylene above 250K, and evolution to fourfold rotational disorder above 375K, on the time scale of picoseconds [58].
For processes whose frequency of occurrence is less than 1 GHz (i.e., once
per nanosecond), the events are sufficiently rare to render molecular
dynamics of little use as a means for direct simulation, given current computer limitations. Rare events are more often modeled using some form of
transition state theory (TST) [69]. The basic assumptions of TST are that
there exists a dividing surface which separates phase space into a reactant
region and a product region, that the region of space comprising the reactant state and the dividing surface are in equilibrium, and that trajectories
passing through the dividing surface are captured in the product state. In its
conventional form, TST places this dividing surface at a first order saddle

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Modeling Polymer Crystals

383

point in the potential energy surface, defined as a stationary point with one
and only one negative eigenvalue of the dynamical matrix. The problem
then centers on the characterization of the reactant and saddle point
(i.e., transition state) configurations. For simplicity, the present discussion
adheres to this conventional interpretation.
In conventional TST, the one-way flux, or reaction rate, from reactant to
product may be written as:
kTST


z
kB T Qvib
E
exp

h Qvib, 0
kB T

28

where h is Plancks constant and  is the transmission coefficient, usually

taken to be on the order of unity. Characterizing the kinetic rates of processes dominated by rare events reduces to estimating the activation energy
E (Uz
U0) and the ratio, Qzvib /Qvib,0, of the vibrational partition functions for reactant (0) and transition (z) states. If the reactant and transition
states are reasonably well known, the methods of lattice dynamics may be
used to compute the four quantities required, Uz, Qzvib , U0, and Qvib,0. It is
worth noting that the specification of the trajectory and the associated oneway flux in TST constitutes a classical treatment of the one degree of freedom associated with motion along the trajectory; Qzvib may be evaluated
either classically or quantum mechanically using Eqs. (11), but possesses
one fewer degree of freedom than Qvib,0 due to the restriction of motion
to a specifed trajectory. At a saddle point, one imaginary frequency is
obtained in the dynamical matrix and is discarded.
The identification of the transition state configuration(s) for rare events
in polymer crystals can be a complicated task. Conformational defects of the
types discussed above in relation to molecular dynamics simulations have
received the most attention in polymer crystals, due to their association with
experimentally observed mechanical and dielectric relaxations, crystal thickening, etc. These are also illustrative of the application of TST to polymer
crystal studies.
It is generally believed that relaxations in the crystal phase occur through
the formation and propagation of local regions of disorder, called defects.
The gauche bonds, kinks, and twists observed in the molecular dynamics
simulations cited earlier are all examples of conformational defects.
Conformational defects in polymer chains are classified by the net translational and rotational lattice mismatch incurred by a defect involving a particular length of the chain (the extent of the defect). Such translational
mismatches have been called dislocations, by analogy to dislocations in
the solid-state physics literature. However, studies of relaxations in polymer
crystals have typically focused on point defects, and should not be confused

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

384

Rutledge

with the line defects of the latter; similarly, rotational mismatches are
termed disclinations. In polymers, the helical symmetry of the chain permits a combined translational/rotational mismatch that has been termed a
dispiration [70]. The identification of the structure of defects which satisfy
the constraints of a dislocation, disclination, or dispiration typically requires
considerable insight, inspiration, or brute force searching of conformation
space [7173]. For highly localized conformational defects, the process of
defect identification can be simplified by solving the conformational problem in torsion space alone, for defects up to six bonds long [74,75]. The
method involves solving a set of nonlinear equations, and also serves as
the basis of the Concerted Rotation Monte Carlo algorithm [76]. Once
the conformational problem is solved, one can construct a cell containing
a single reference chain containing the conformational defect, and thereby
compute the energy and distortion of the defect due to packing in the lattice.
Unlike perfect crystal simulations, simulations to date involving defects
have been performed almost exclusively using finite cells with rigid or free
boundary conditions so as to avoid defectdefect interactions.
Having once identified a stable defect in the crystal lattice and computed
its energy of formation, determining the kinetics of a process based on this
defect in TST requires the further identification of a saddle point along a
trajectory which displaces the defect by one lattice period; the periodicity of
the lattice permits a scenario where subsequent propagation of the defect
occurs through a sequence of similar hops in the lattice. Identification of
a suitable trajectory for defect propagation in polyethylene was performed
by introducing the defect into the lattice, as described above, and then
driving one or more coordinates (e.g., a torsion angle in the chain or setting
angle in the crystal) so that the defect is forced to move to the next unit cell
in the lattice. This type of calculation is essentially quasi-static; a constraint
on one or more rotation angles in the chain is incremented, and the potential
energy of the defect reminimized [73,77]. The process is repeated stepwise
until a potential energy barrier is crossed (i.e., the defect falls into the
product state upon minimization). However, this approach does not generally locate a first order saddle point on the potential energy surface, but
rather a maximum along the trajectory chosen by the coordinate-incrementing protocol employed. The resulting trajectories are thus sensitive to how
the defect is driven along the chain and provide upper bounds on the true
saddle point energy [77]. Alternative numerical approaches can be used to
refine the estimate of a true first order saddle point based on an adiabatic
reaction path connecting reactant to product. One particularly useful algorithm for this purpose is the conjugate peak refinement algorithm of Fischer
and Karplus [78], which requires knowledge only of the equilibrium conformation of the defect in order to identify a lowest energy path from

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Modeling Polymer Crystals

385

FIG. 5 Free energy of formation and transition state barriers (A

(Qzvib /Qvib,0)exp(
E/kBT )) to transport for two types of defects capable of producing the -relaxation in poly(vinylidene fluoride) (from Ref. 79). tgtg* ! g*tgt
(circles); g*tgt ! tgtg* (squares). Dashed curves are drawn as guides to the eye.

reactant to product. This saddle point search algorithm was used in a study
of poly(vinylidene fluoride) [79] and, more recently, polyethylene [80]. For
cases where a good initial guess of the saddle point configuration is available, more efficient gradient algorithms can be used to refine the estimate of
the saddle point configuration and energy [81]. Once the saddle point configuration is identified, Uz and Qz are readily computed using the methods
of lattice dynamics, and TST applied. Figure 5 illustrates two trajectories
computed for propagation of defects in poly(vinylidene fluoride) to account
for the -relaxation in the -crystalline form of that polymer. It is interesting
to note that, whereas heats of formation for defects may be fairly large, the
methods described above can be used to characterize the trajectory of the
defect through the lattice, regardless of whether it involves a hopping
mechanism between adjacent lattice sites, separated by large energy barriers,
or a relatively facile free-streaming motion involving only low energy barriers to propagation. Both hopping [70] and free-streaming [73] mechanisms
have also been proposed for motion of defects in polyethylene.

V. SUMMARY
There is currently a variety of methods for studying polymer crystals by
molecular simulation. As with other materials, the usual caveats regarding force field accuracy and proper execution of the simulation apply.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

386

Rutledge

The anisotropy of polymer crystals is characteristic of the orientation of

intra- and intermolecular interactions, which is also reflected in a broad
separation on the dynamical time scale. It is especially important in polymer
crystals to be cognizant of the limitations imposed by either the assumptions
on which a method is based (e.g., the quasiharmonic approximation for
lattice dynamics) or the robustness of the simulation method (e.g., ergodicity
of the simulation in Monte Carlo or molecular dynamics). In some
instances, valuable checks on these assumptions and limitations (e.g., quantum mechanical vs. classical dynamics, finite size effects and anharmonicity)
can and should be made by repeating the study using more than one
method.

REFERENCES
1. Cullity, B.D. Elements of X-Ray Diraction, 2nd Ed.; Addison-Wesley:
Reading, 1978.
2. Alexander, L.E. X-Ray Diraction Methods in Polymer Science; Wiley
Interscience: New York, 1969.
3. Guinier, A. X-Ray Diraction in Crystals, Imperfect Crystals, and Amorphous
Bodies; W.H. Freeman and Co.: San Francisco, 1963.
4. Miyazawa, T. J. Polym. Sci. 1961, 55, 215.
5. Sugeta, H.; Miyazawa, T. Biopolymers 1965, 5, 673.
6. Meille, S.V.; Bruckner, S.; Porzio, W. Macromolecules 1990, 23, 41144121.
7. Dave, R.S.; Farmer, B.L. Polymer 1988, 29, 15441554.
8. Fisher, H.P.; Eby, R.K.; Cammarata, R.C. Polymer 1994, 35, 19231930.
9. Balijepalli, S.; Rutledge, G.C. J. Chem. Phys. 1998, 109, 65236526.
10. Balijepalli, S.; Rutledge, G.C. Comput. Theor. Polym. Sci. 2000, 10 (1,2),
103113.
11. Gautam, S.; Balijepalli, S.; Rutledge, G.C. Macromolecules 2000, 33 (24),
91369145.
12. Corradini, P.; Guerra, G. Advances in Polymer Science 1991, 100, 185.
13. Press, W.H.; Teukolsky, S.A.; Vetterling, W.T.; Flannery, B.P. Numerical
Recipes in C: The Art of Scientic Computing, 2nd Ed.; Cambridge
University Press: Cambridge, 1992.
14. Allen, M.P.; Tildesley, D.J. Computer Simulation of Liquids; Clarendon Press:
Oxford, 1987.
15. McCullough, R.L. J. Macromol. Sci. Phys. 1974, 9, 97.
16. Tripathy, S.K.; Hopnger, A.J.; Taylor, P.L. J. Phys. Chem. 1981, 85,
13711380.
17. Bacon, D.J.; Geary, N.A. J. Mater. Sci. 1983, 18, 853863.
18. Rutledge, G.C.; Suter, U.W. Macromolecules 1991, 24, 19211933.
19. Lacks, D.J.; Rutledge, G.C. Macromolecules 1994, 27, 71977204.
20. Zhang, R.; Taylor, P.L. J. Chem. Phys. 1991, 94, 32073212.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Modeling Polymer Crystals

387

21. Tadokoro, H. Structure of Crystalline Polymers; Wiley and Sons: New York,
1979.
22. Tashiro, K. Prog. Polym. Sci. 1993, 18, 377435 and references therein.
23. Born, M.; von Karman, T. Phys. Zeit. 1912, 13, 297.
24. Born, M.; von Karman, T. Phys. Zeit. 1913, 14, 15.
25. Born, M.; Huang, K. Dynamical Theory of Crystal Lattices; Oxford University
Press: Oxford, 1954.
26. Willis, B.T.M.; Pryor, A.W. Thermal Vibrations in Crystallography; Cambridge
University Press: Cambridge, 1975.
27. Horton, G.K.; Maradudin, A.A. Eds. Dynamical Properties of Solids; NorthHolland: Amsterdam, 19741990; Vol. 17.
28. Dove, M.T. Introduction to Lattice Dynamics; Cambridge University Press:
Cambridge, 1993.
29. Sorensen, R.A.; Liau, W.B.; Kesner, L.; Boyd, R.H. Macromolecules 1988, 21,
200208.
30. Karasawa, N.; Dasgupta, S.; Goddard, W.A. J. Phys. Chem. 1991, 95,
22602272.
31. Boyd, R.H. In Advances in Polymer Science, Monnerie, L., Suter, U.W., Eds.;
Springer Verlag: Berlin, 1994; Vol. 116, 125.
32. Allan, N.L.; Barron, T.H.K.; Bruno, J.A.O. J. Chem. Phys. 1996, 105,
83008303.
33. Lacks, D.J.; Rutledge, G.C. J. Phys. Chem. 1994, 98, 12221231.
34. Lacks, D.J.; Rutledge, G.C. Macromolecules 1995, 28, 11151120.
35. Carbeck, J.D.; Lacks, D.J.; Rutledge, G.C. J. Chem. Phys. 1995, 103,
1034710355.
36. Rutledge, G.C. Macromolecules 1997, 30, 27852791.
37. Taylor, M.B.; Barrera, G.D.; Allan, N.L.; Barron, T.H.K. Phys. Rev. B 1997,
56 (22), 1438014390.
38. Bruno, J.A.O.; Allan, N.L.; Barron, T.H.K.; Turner, A.D. Phys. Rev. B 1998,
58 (13), 84168427.
39. Dick, B.; Overhauser, A. Phys. Rev. 1958, 112, 90.
40. Karasawa, N.; Goddard, W.A. Macromolecules 1992, 25, 72687281.
41. Carbeck, J.D.; Rutledge, G.C. Polymer 1996, 37, 50895097.
42. Anand, J.N. J. Macromol. Sci.-Phys. B1 1967, (3), 445.
43. Wobser, V.G.; Blasenbrey, S. Kolloid. Z. Z. Polym. 1970, 241. 985.
44. Tashiro, K.; Kobayashi, M.; Tadokoro, H. Macromolecules 1978, 11 (5), 914.
45. Martonak, R.; Paul, W.; Binder, K. Phys. Rev. E 1998, 57, 24252437.
46. Rutledge, G.C.; Lacks, D.J.; Martonak, R.; Binder, K. J. Chem. Phys. 1998,
108, 1027410280.
47. Tashiro, K.; Kobayashi, M.; Tadokoro, H.; Fukada, E. Macromolecules 1980,
13, 691.
48. Burnett, M.N.; Johnson, C.K. Ortep-III: Oak Ridge Thermal Ellipsoid Plot
Program for Cyrstal Stucture Illustration. Oak Ridge National Laboratory
Report ORNL-6895, 1996.
49. Lacks, D.J.; Rutledge, G.C. J. Chem. Phys. 1994, 101, 99619965.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

388
50.
51.
52.
53.
54.
55.
56.
57.
58.
59.
60.
61.
62.
63.

64.
65.
66.
67.
68.
69.
70.
71.
72.
73.
74.
75.
76.
77.
78.
79.
80.
81.

Rutledge
Shukla, R.C.; Cowley, E.R. Phys. Rev. B 1985, 31, 372378.
Shukla, R.C.; Shanes, F. Phys. Rev. B 1985, 32, 25132521.
Shukla, R.C.; Bose, S.K.; Delogu, R.F. Phys. Rev. B 1992, 45, 1281212820.
Shukla, R.C.; Lacks, D.J. J. Chem. Phys. 1997, 107, 74097417.
Martonak, R.; Paul, W.; Binder, K. Comput. Phys. Commun. 1996, 99, 28.
Martonak, R.; Paul, W.; Binder, K. J. Chem. Phys. 1997, 106, 89188930.
Liang, G.L.; Noid, D.W.; Sumpter, B.G.; Wunderlich, B. Comput. Polym. Sci.
1993, 3, 101.
Andersen, H.C. J. Chem. Phys. 1980, 72, 2384.
Ryckaert, J.-P.; Klein, M.L. J. Chem. Phys. 1986, 85, 16131620.
Yamamoto, T. J. Chem. Phys. 1985, 82, 37903794.
Yamamoto, T. J. Chem. Phys. 1988, 89, 23562365.
Foulger, S.H.; Rutledge, G.C. Macromolecules 1995, 28, 70757084.
Foulger, S.H.; Rutledge, G.C. J. Polym. Sci.-Phys. 1998, 36, 727741.
Sumpter, B.G.; Noid, D.W.; Liang, G.L.; Wunderlich, B. In Advances in
Polymer Science; Monnerie, L., Suter, U.W., Eds.; Springer Verlag: Berlin,
1994; Vol. 116, 2772 and references therein.
Van Kampen, N.G. Stochastic Processes in Physics and Chemistry; NorthHolland: Amsterdam, 1992.
Parrinello, M.; Rahman, A. J. Chem. Phys. 1982, 76 (5), 26622666.
Zehnder, M.M.; Gusev, A.A.; Suter, U.W. Rev. Inst. Francais Petrole 1996,
51 (1), 131137.
Gusev, A.; Zehnder, M.M.; Suter, U.W. Phys. Rev. B. 1996, 54, 14.
Cagin, T.; Karasawa, N.; Dasgupta, S.; Goddard, W.A. Mat. Res. Soc. Symp.
Proc. 1992, 278, 61.
Glasstone, S.; Laidler, K.; Eyring, H. The Theory of Rate Processes; McGrawHill: New York, 1941.
Reneker, D.H. J. Mazur. Polymer 1988, 29, 313.
Reneker, D.H. J. Polym. Sci. 1962, 59 (168), S39.
Boyd, R.H. J. Polym. Sci.-Phys. 1975, 13, 2345.
Manseld, M.; Boyd, R.H. J. Polym. Sci.-Phys. 1978, 16, 1227.
McMahon, P.E.; McCullough, R.L.; Schlegel, A.A. J. Appl. Phys. 1967, 38 (11),
4123.
Go, N.; Scheraga, H.A. Macromolecules 1970, 3, 178.
Dodd, L.R.; Boone, T.D.; Theodorou, D.N. Mol. Phys. 1993, 78, 961996.
Reneker, D.H.; Fanconi, B.M.; Mazur, J. J. Appl. Phys. 1977, 48 (10), 4032.
Fischer, S.; Karplus, M. Chem. Phys. Lett. 1992, 194, 252261.
Carbeck, J.D.; Rutledge, G.C. Macromolecules 1996, 29 (15), 51905199.
Mowry, S.W.; Rutledge, G.C. Macromolecules 2002, 35 (11), 45394549.
Schlegel, H.B. In Ab Initio Methods in Quantum Chemistry; Lawley, K.P, Ed.;
Wiley and Sons: New York, 1987; 249 pp.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

12
Plastic Deformation of
Bisphenol-A-Polycarbonate: Applying
an Atomistic-Continuum Model
JAE SHICK YANG and WON HO JO
Korea

Seoul National University, Seoul,

SERGE SANTOS and ULRICH W. SUTER

Hochschule (ETH), Zurich, Switzerland

I.

Eidgenossische Technische

INTRODUCTION

Atomistic modeling techniques have been employed for the investigation

of mechanical properties of amorphous polymeric solids for years [15].
Theodorou and Suter [1] successfully calculated the elastic constants of
glassy atactic polypropylene (aPP) by molecular mechanics techniques,
surmising that the entropic contributions to the elastic response are negligible
for such glasses. Mott et al. [2] strained polypropylene glasses beyond the
macroscopic elastic range and found evidence for the occurrence of diffuse
plastic unit events and estimated, by comparison with experimental results,
that the size of the plastically transforming domains in elementary sheartransformation events extends over roughly 100 A (well beyond the system
size that can conveniently be simulated). Plastic relaxation occurs through
cooperative rearrangements with rather small local transformation strains
[O(2%)]. These findings were assumed to be true for all polymeric glasses,
in contrast to the long-held belief that the elementary process of plastic
transformation is a well-localized conformational transition. Hutnik et al. [3]
used the same technique on glassy bisphenol-A-polycarbonate (BPA-PC)
and confirmed the results on aPP. All these calculations suffered from small
system size and the fact that the atomistic box shape was used as a simulation
variable, prescribed and controlled during the athermal simulations. While
this is perfectly appropriate for the elucidation of elastic properties, it is
not optimal when the natural system path is unknown, as in plastic

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

390

Yang et al.

deformation. It would be preferable to be able to simulate the behavior of

atomistic subsystems embedded in a medium (of identical properties) and free
to follow the driving forces that the gradient of (free) energy suggests. The
overall system would have to be considerably larger, however, than the ones
employed for these atomistic simulations.
The demand for larger systems in the modeling of mechanical properties
of polymeric materials requires the consideration of multiple length scales.
For very large length scales it suffices to think of the material as an elastic medium and an atomistic description is unnecessary. However, for
the plasticity to be captured at the molecular level, atomistically detailed
modeling is required and several efforts to combine the atomistic and
continuum levels for the modeling of viscoelastic properties of materials
have been published [610]. The fundamental difficulty in this endeavor lies
in the way in which the length scales are coupled. We have presented [11] an
approach at atomistic-continuum modeling, where the system consists of a
continuum matrix and an atomistic inclusion, and the inclusion-boundary
behavior and the atomistic cell behavior are connected via the common
strain transformation. This new model has proven to be consistent for
modeling the elastic deformation of a nearest-neighbor fcc crystal of argon
and glassy BPA-PC. It has the advantage of being free of artificial
constraints and assumptions beyond those intrinsic to the descriptions on
the two levels being connected.
Here, the atomistic-continuum model is applied to the study of plastic
deformation of BPA-PC. First, the elastic constants of BPA-PC are
calculated by atomistic simulation. These values are used as the elastic
constants for the matrix throughout the simulated deformation. Then, the
system, i.e., the continuum matrix with its atomistic inclusion, is deformed
stepwise up to a strain of about 0.2. The overall system is constrained
to exactly follow a predescribed deformation sequence, but the atomistic
inclusion is free to follow any strain path consistent with the misfit stresses
acting between it and the matrix. The result is a new look at the behavior of a
glassy polymeric inclusion deformed plastically in a surrounding elastic
medium.

II. MODEL
We briefly recapitulate the model, detailed previously [11], which is based on
the variable-metric total-energy approach by Gusev [12] and consists of a
system comprising an inclusion embedded in a continuous medium. The
inclusion behavior is described in atomistic detail whereas the continuum is
modeled by a displacement-based finite-element method. Since the atomistic

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Atomistic Continuum Model

391

model provides configuration-dependent material properties inaccessible to

continuum models, the inclusion in the atomistic-continuum model acts as a
magnifying glass, which allows observation of the molecular properties of
the material.

A. Continuum Model
The variable-metric total-energy approach [12] is adopted here for the
matrix. As illustrated in Fig. 1, the periodic system is described by the
scaling matrix [13,14] H [ABC], where A, B, and C are the overall systems
continuation vectors. Two kinds of nodal points are specified in the matrix:
One (xb) on the inclusion boundary and the other (xc) in the continuum
(throughout this text, vectors are written as column matrices). For convenience, the scaled coordinates [13,14] (sb and sc) are chosen as degrees of
freedom via xb Hsb and xc Hsc. These nodal points are used as vertices
of a periodic network of Delaunay tetrahedra. The Delaunay network
uniquely tessellates space without overlaps or fissures; the circumsphere of
the four nodal points of any tetrahedron does not contain another nodal
point of the system [15,16]. For each tetrahedron , the local scaling matrix
h [abc] is defined in the same manner as H for the overall system. The
local (Lagrange) strain e is assumed to be constant inside each tetrahedron
and defined as [13,14]
e

i
1 h  T 
G G
I ,
2

 
1
G h h0

FIG. 1 Sketch of atomistic-continuum model, for illustrative purposes drawn for a

two-dimensional version only. All degrees of freedom in the system are of one of four
types: H, DH, sc, and sa.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

392

Yang et al.

where the subscript 0 represents the reference state before deformation and I
is the unit matrix of order 3. The system strain esys is calculated from Eq. (1)
by substituting H and H0 for h and h0 , respectively. The strain energy in the
continuum Econt can be evaluated from the sum over all tetrahedra:
     1 X   T  
V e C e
E cont H, sb , sc
2 

where V is the volume of the reference state and C the matrix of elastic
constants of tetrahedron .

B. Atomistic Model
A spatially periodic simulation box represents the contents of the atomistic
inclusion. Again, the shape of this box is expressed by a scaling matrix
h [abc] (cf. Fig. 1). The scaled coordinates of atom sa are used as degrees of
freedom via xa hsa. The Lagrange strain tensor of the atomistic box, eatom,
can be calculated from Eq. (1) by replacing h and h0 with h and h0,
respectively. The energy in the atomistic box can be expressed as Eatom
(h,{sa}) and can be obtained via any energy model for atomistic simulation,
e.g., a force-field approach.

C. Atomistic-Continuum Model
To connect the two models with different length scales, the displacement
of the nodal points on the inclusion boundary is associated with the
b
deformation of the atomistic box by xb hh
1
0 x0 . This means that the
inclusion boundary is constrained to follow the homogeneous deformation
of the atomistic cell vectors. The degree of freedom DH is introduced to
relate h with H through
h HI DH H
1
0 h0

Thus, sb can be expressed by

b
b
sb H
1 xb H
1 hh
1
0 H0 s0 I DH s0

There are four types of degrees of freedom in the system: H and sc for the
continuum model; h and sa for the atomistic model.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Atomistic Continuum Model

393

The total potential energy in the system Esys is consequently expressed as

E sys H,DH ,fsc g,fsa g E cont H,DH ,fsc g V E atom H,DH ,fsa g,
V

V inc
V atom

where Vinc and Vatom are the volumes of the inclusion and atomistic box,
respectively (evidently, the inclusion must have the same energy density as
the atomistic box). Note that V is kept constant during deformation
because xb deforms according to h.
The length scales in the model are given by the inclusion volume fraction
fI and V, e.g., in a cubic system box of edge length L with a spherical
inclusion of diameter d, defined by an atomistic box of edge length a,
d (6V/)1/3a and L (V/fI)1/3a.
While this two-scale model allows for a natural connection between
the two scales (i.e., the inclusion-boundary behavior and the atomistic cell
behavior are connected via the common strain transformation), it also is
limited by the required uniformity of strain in the atomistic inclusion
(through the periodic continuation conditions on the atomistic cell) and in
each of the tetrahedra of the Delaunay tessellation.

III. SIMULATION METHOD

A. Model System
The system was a cube with an edge length L 100 A and one spherical
inclusion (volume fraction fI 0.18). We specified 1000 nodal points on the
inclusion boundary (xb) and 6259 in the continuum (xc). The atomistic
box consists of one BPA-PC chain with degree of polymerization X 74,
resulting in 2454 atoms. The molecular structure of BPA-PC is sketched in
Fig. 2; the dihedral angles are numbered as in previous publications [3], even
though a flexible-chain, Cartesian-coordinate approach is used here. Three
independent atomistic structures were generated and equilibrated at the
experimental density of 1.20 g cm
3 (300 K) [3], following the procedure of
[11], whereupon the energies of the atomistic structures were minimized with
respect to all degrees of freedom, including hatom, to remove residual
stress inside the atomistic boxes [17] through the Polak-Ribiere conjugate
gradient minimization method [18]. The pcff force field [1921] was used to
calculate the atomistic energy, which is composed of 14 components:
E atom Eb E
E E Eelec EvdW
Ebb0 Ebb00 Eb
E

0 Eb Eb0  E
 E

0

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

394

FIG. 2

Yang et al.

The constitutional repeat unit of BPA-PC.

where the first four terms are valence terms (bond length, bond angle,
dihedral angle, and out-of-plane angle), the next two terms nonbonded
interaction terms (electrostatic and van der Waals), and the last eight
terms coupling valence terms. For efficiency, an in-house code was constructed that yields the same energy and partial derivatives with respect to
coordinates and components of the scaling matrix as the commercial program Discover [19]. The nonbonded energy cutoff and the spline width were
8 A and 2 A, respectively. The charged-group based method [19] was
adopted for the calculation of nonbonded energies to keep the total
charge inside the cutoff to zero.

B. Elastic Deformation of the Atomistic Model

The constant strain method was used to determine the elastic constants
Catom of three independent atomistic boxes. The average values of these
elastic constants were assigned to the matrix [i.e., to all C in Eq. (2)]. A set
of small strains ("  10
4) were applied to the atomistic box in its minimumenergy configuration and the energy of the atomistic box was minimized
with respect to all degrees of freedom, except h. The elastic constants of the
atomistic box Catom were evaluated by parabolic fits to

T
1
E atom
E0atom V atom eatom Catom eatom
2

In the Voigt notation, elastic constants can be expressed as 6  6 matrices

[22]. For an isotropic material, C takes the form
1
0
2

0 0 0
B

0 0 0C
C
B
B

2
0 0 0 C
C
8
CB
B 0
0
0
0 0C
C
B
@ 0
0
0
0 0A
0
0
0
0 0

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Atomistic Continuum Model

395

where
and are the so-called Lame constants. The Youngs modulus E,
the shear modulus G, the bulk modulus B, and the Poissons ratio  are
obtained by
3
2
,

2
B
,
3

C. Plastic Deformation of the Atomistic-Continuum

Model
The deformation method developed by Mott et al. [2] was used here. The
system was subjected to deformation in small steps up to a system strain
of 0.2, the overall energy being minimized with respect to all degrees of
freedom (except H) after each step. The strain increment of the system,
iesys, was
0

B
esys B
@0
0
0

2
B
B
@0

esys

1 q

C
C  10
3 for uniaxial extension
A

10

1 q
1
0
0
C

3

2 r 0 C
A  10 for pure shear
0

11

where q and r are small values adjusted to keep the system volume constant.
Thus, for both, uniaxial extension and pure shear, the system volume Vsys
was held constant while the atomistic box was free to change its volume
Vatom [through the coupling degree of freedom DH, Eq. (3)].
The work-equivalent tensile strains in the system esys
eq and in the atomistic
are
given
by
[2]
box eatom
eq

2  sys 2 1=2
1
Tr e^
where e^ sys esys
Tresys I
3
3

1=2
2

2 
1 
Tr e^ atom
where e^ atom eatom
Tr eatom I
3
3

esys
eq

12

eatom
eq

13

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

396

Yang et al.

The internal stress tensors in the system psys and in the atomistic box patom
are calculated by [13,14]
psys

sys T
1
@E
H
,
sys
V
@H

patom

h
atom

atom T
@E
@h

14

sys
The von Mises-equivalent tensile stresses of the system eq
and of the atoatom
mistic box  eq are given by [2]

sys
eq

atom
eq

3
Trr^ sys 2
2

1=2
where

3  atom 2 1=2
Tr r^
2

1
r^ sys rsys
Trrsys I
3


1 
where r^ atom ratom
Tr ratom I
3

15

16

IV. RESULTS AND DISCUSSION

A. Elastic Deformation
In the Voigt notation, the average elastic constants for three atomistic boxes
are (accurate to approximately  0.1 GPa)
0

Catom

7:46
B 4:63
B
B 4:52
B
B 0:00
B
@ 0:11

0:16

4:63
7:08
4:38

0:02
0:15
0:01

4:52
0:00
4:38
0:02
7:03
0:06

0:06 1:47
0:10
0:01

0:03 0:03

0:11
0:15
0:10

0:01
1:51

0:06

1

0:16
0:01 C
C

0:03 C
CGPa
0:03 C
C

0:06 A
1:55

17

Comparing Eq. (17) with Eq. (8) leads us to accept the BPA-PC models as
isotropic glasses with Lame constants and moduli as listed in Table 1.
Comparison with experimental values [23] indicates that the agreement is
satisfactory.

B. Plastic Deformation
Figure 3 shows stressstrain curves from the simulation runs. Because more
than 80% of the system volume is assumed to behave perfectly elastically, the
system stress increases almost linearly with the system strain [cf. Eq. (5)].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Atomistic Continuum Model

TABLE 1

397

Comparison of Calculated Elastic Constants with Experimental Values

Property

Calculated

Experimental [23]

Lame constant
(GPa)
Lame constant (GPa)
Youngs modulus E (GPa)
Shear modulus G (GPa)
Bulk modulus B (GPa)
Poissons ratio 

4.35  0.10
1.43  0.06
3.93  0.14
1.43  0.06
5.30  0.10
0.38  0.01

4.275.55
0.81.1
2.32.5
0.81.1
5.06.1
0.420.43

FIG. 3 Stressstrain curves of the system under uniaxial extension (filled squares)
and pure shear (open circles).

However, the lines are not smooth but contain numerous small discontinuities where the system stress decreases precipitously after a small strain
increment owing to a plastic response of the inclusion; arrows mark the
particularly conspicuous drops. The two curves for uniaxial and pureshear deformations would superpose perfectly if the system behaved totally
elastically; deviations are evident after a sufficient number of plastic unit
events, which are irreversible and dissipative and probably different for

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

398

Yang et al.

FIG. 4 Stressstrain curves of the atomistic box under uniaxial extension (filled
squares) and pure shear (open circles).

different deformation modes. Figure 4, which gives the stressstrain curves

for the atomistic inclusion only, clearly shows the microstructural plastic
behavior, particularly pronounced at the loci of the arrows in Fig. 3. As has
been observed before in the simpler simulations of aPP and BPA-PC [2,3],
the atomistic stressstrain curves do not emulate the macroscopic elastic-toplastic transition, but consist of several straight lines with sudden drops in the
stress at some points. Note that the slopes of those lines between plastic
events remain nearly constant during deformation; the glassy phases
elastically strained between plastic transformations all have the same moduli.
In Fig. 5, the stress of the system is compared with that in the atomistic
box. It is apparent that this ratio assumes almost constant values between
plastic events with sudden increases at these events. Because the elastic
constants of the matrix are those calculated from the atomistic box at the
outset of the simulations, the system is homogeneous at small strains. As
the system strain increases further, the atomistic box becomes softer than the
perfectly elastic matrix (strain softening).
Interestingly (cf. Fig. 6), the work-equivalent tensile strain of the
sys
atomistic box, "atom
eq , is equal to that of the system, "eq , until far beyond
the first plastic unit events (at about 8% strain) and only starts to deviate at

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Atomistic Continuum Model

399

sys
FIG. 5 The ratio of the von Mises-equivalent tensile stress of the system, eq
, to
atom
that in the atomistic box, eq , against the system strain under uniaxial extension
(filled squares) and pure shear (open circles).

FIG. 6 The work-equivalent tensile strain of the atomistic inclusion, "atom

eq , as a
function of the system strain, "sys
eq , under uniaxial extension (filled squares) and pure
shear (open circles).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

400

Yang et al.

FIG. 7 The volume of the atomistic box, Vatom/V0atom , as a function of the atomistic
strain, "atom
eq , under uniaxial extension (filled squares) and pure shear (open circles).

about 15% strain (and thereafter becomes greater than the latter). The cause
(cf.
of this is evident when the atomistic box volume is plotted against "atom
eq
Fig. 7, note that the system volume Vsys is kept constant during deformation
under both uniaxial extension and pure shear). The volume of the atomistic
0:07 and then
box Vatom remains at its initial value V0atom until "atom
eq
gradually increases with strain while there are sharp increases at the points
where the sudden drops in the stress are observed. Hence, our simulations
indicate that the plastic unit event is a dilatational process, in which the
transforming domain assumes a reduced density while the surrounding
matrix takes up the concomitant elastic deformation. This is in apparent
contrast with the deductions by Hutnik et al. [3] who found densification
during plastic deformation for BPA-PC (experimental results show that the
plastic deformation of polymer glasses at constant pressure is a nearly
constant-volume process, with a small dilatation for vinyl polymers and a
small volume contraction for BPA-PC) [2428]. Further studies are needed
to gain a satisfactory understanding of these observations.
Because the matrix is modeled as a perfectly elastic medium, the matrix
energy is quadratic with respect to the system strain and, consequently, the
energy of the continuum portion of the model against strain must exhibit a

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Atomistic Continuum Model

401

FIG. 8 Energy change in the continuum matrix, iEcont, and in the atomistic
inclusion, ViEatom, as a function of the corresponding work-equivalent tensile
strains, under uniaxial extension (filled squares) and pure shear (open circles).

slope of 2, plotted doubly logarithmically (cf. Fig. 8). The inclusion energy
(VEatom) shows the same tendency at low strain, but deviates as plastic
events begin to soften the material.
The backbone dihedral angles in the constitutional repeat unit can be
categorized into three groups [29]: (i) 1 and 2, which determine the
conformation of the phenylene rings with respect to the isopropylidene
group; (ii) 3 and 6, which specify the conformation of the phenylene rings
with respect to the carbonate group; (iii) 4 and 5, which give the
conformation of the carbonate group. We now analyze the distribution of
the angles in these groups and their changes during deformation.
1 and 2 are strongly interdependent and prefer propeller-like
conformations. Figure 9 contains a plot of all pairs of values of 1 and 2
angles concurrently occurring at one isopropylidene moiety. The intramolecular ground states are located at (45 , 45 ) and (135 ,
45 ) [29]. Upon
deformation to "atom
0:256, these distributions do not change perceptibly
eq

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

402

Yang et al.

FIG. 9 Distribution of the dihedral angles 1 and 2. The large plot contains the
values for the pairs (1, 2) before deformation ("atom
0:0, open circles) and after
eq
deformation under pure shear ("atom

0:256,
filled
circles).
The polar plot indicates
eq
the changes in 1 and 2, where the angle on the perimeter denotes the direction of
the change and the distance from the center denotes the magnitude of the change.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Atomistic Continuum Model

403

(the pairs of angles are marked with open circles for the unstrained and with
filled circles for the deformed state). An analysis of the changes in 1 and 2
during this massive deformation experiment reveals that the conformations
change mostly from one propeller-like state to a symmetry-related one and
that the magnitude of this change often exceeds 180 ; the polar plot included
in Fig. 9 indicates the direction of the change in the (1, 2) plane on
the circumference and its magnitude as the distance from the center. The
orientation of this plot is such that the direction of change conforms to
the large plot in Fig. 9.
Angles 3 and 6 assume values of 45 at conformations of minimal
energy, while 4 and 5 are preferably at 0 (trans) and 180 (cis). Figure 10

FIG. 10 Distribution of the dihedral angles 3 (or 6) and 4 (or 5). The large plot
contains the values for the pairs [3 (or 6), 4 (or 5)] before deformation
0:0, open circles) and after deformation under pure shear ("atom
0:256,
("atom
eq
eq
filled circles). The polar plot indicates the changes in 3 (or 6) and 4 (or 5), where
the angle on the perimeter denotes the direction of the change and the distance from
the center denotes the magnitude of the change.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

404

Yang et al.

is constructed similarly to Fig. 9: The large plot exhibits pairs of dihedral

angles joined at the same oxygen atom (i.e., 3 and 4 or 6 and 5), again
with open circles for the undeformed state and with filled circles for
0:256. The polar plot shows the changes in these angle pairs upon
"atom
eq
strong deformation, again in an orientation that conforms directly to the
large plot. It is obvious that a large number of significant conformational
transitions occur during the deformation runs, but most involve state
changes of the angles adjoining the phenylene units only; changes of the
carbonate torsion angle from trans to cis (or vice versa) are rare. And again,
the torsion angle distributions do not change perceptibly upon deformation
to "atom
0:256.
eq
The absence of significant change in the distribution of the backbone
dihedral angles upon plastic deformation of glassy BPA-PC has been
experimentally observed. Utz [30] and Utz et al. [31] compressed specifically
13
C labeled samples uniaxially and in plane strain to strains of
0.68 and
found, by solid-state NMR, that the distributions of the torsion angle pairs
3 and 4 (or 6 and 5) do not change perceptibly.

V. CONCLUSIONS
The elastic constants of BPA-PC, calculated by atomistic modeling, show
relatively good agreement with experimental values. The atomisticcontinuum model is successfully applied to the large deformation behavior
of a BPA-PC system where an inclusion described in an atomistically
detailed manner deforms plastically in an elastic medium. In the atomistic
box, the stressstrain curves show typical microstructural plastic behavior
with sharp drops in the stress. Because the matrix is characterized by elastic
constants of BPA-PC calculated by atomistic modeling, the strain of the
atomistic box is equal to the system strain until a certain point. At large
deformations, the strain in the atomistic box becomes higher than the
system strain.
The volume of the atomistic box initially remains constant because the
volume of the system is kept constant during deformation under both
uniaxial extension and pure shear. At strains greater than "atom
 0:07, the
eq
volume of the atomistic box grows gradually with the strain and there are
sharp increases at several points where sudden drops in the stress in the
atomistic box are observed. No considerable change in the distribution of
the backbone dihedral angles is observed during deformation up to the
system strain 0.2.
The emerging picture of plastic unit events in the polymeric glass is that
of a sudden dilatation followed by melt-like flow.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Atomistic Continuum Model

405

REFERENCES
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.

Theodorou, D.N.; Suter, U.W. Macromolecules 1986, 19, 139.

Mott, P.H.; Argon, A.S.; Suter, A.S. Phil. Mag. A 1993, 67, 931.
Hutnik, M.; Argon, A.S.; Suter, U.W. Macromolecules 1993, 26, 1097.
Fan, C.F. Macromolecules 1995, 28, 5215.
Brown, D.; Clarke, J.H.R. Macromolecules 1991, 24, 2075.
Tadmor, E.B.; Phillips, R.; Ortiz, M. Langmuir 1996, 12, 4529.
Miller, R.; Phillips, R. Phil. Mag. A 1996, 73, 803.
Tadmor, E.B.; Ortiz, M.; Phillips, R. Phil. Mag. A 1996, 73, 1529.
Shenoy, V.B.; Miller, R.; Tadmor, E.B.; Phillips, R.; Ortiz, M. Phys. Rev. Lett.
1998, 80, 742.
Rai-Tabar, H.; Hua, L.; Cross, M. J. Phys.: Condens. Matter 1998, 10, 2375.
Santos, S.; Yang, J.S.; Gusev, A.A.; Jo, W.H.; Suter, U.W. submitted for
publication.
Gusev, A.A. J. Mech. Phys. Solids 1997, 45, 1449.
Parrinello, M.; Rahman, A. J. Appl. Phys. 1981, 52, 7182.
Parrinello, M.; Rahman, A. J. Chem. Phys. 1982, 76, 2662.
Tanemura, M.; Ogawa, T.; Ogita, N. J. Comp. Phys. 1983, 51, 191.
Mott, P.H.; Argon, A.S.; Suter, A.S. J. Comp. Phys. 1992, 101, 140.
Hutnik, M.; Argon, A.S.; Suter, U.W. Macromolecules 1991, 24, 5956.
Press, W.H.; Teukolsky, S.A.; Vetterling, W.T.; Flannery, B.P. Numerical
Recipes, 2nd Ed.; Cambridge University Press: Cambridge, 1992.
Discover 2.9.8/96.0/4.0.0 User Guide; Molecular Simulations Inc.: San Diego,
CA, 1996.
Maple, J.R.; Hwang, M.-J.; Stocksch, T.P.; Dinur, U.; Waldman, M.; Ewig,
C.S.; Hagler, A.T. J. Comp. Chem. 1994, 15, 162.
Sun, H.; Mumby, S.J.; Hagler, A.T. J. Am. Chem. Soc. 1994, 116, 2978.
Nye, J.F. Physical Properties of Crystals; Oxford University Press: Oxford,
1985.
Van Krevelen, D.W. Properties of Polymers, 3rd Ed.; Elsevier: New York, 1990.
Christ, B. In The Physics of Glassy Polymers, 2nd Ed.; Haward, R.N., Young,
R.J., Eds.; Chapman & Hall: London, 1997; Chapter 4.
Spitzig, W.A.; Richmond, O. Polym. Eng. Sci. 1979, 19, 1129.
Ruan, M.Y.; Moaddel, H.; Jamieson, A.M. Macromolecules 1992, 25, 2407.
Xie, L.; Gidley, D.W.; Hristov, H.A.; Yee, A.F. J. Polym. Sci., Polym. Phys.
Ed. 1995, 33, 77.
Hasan, O.A.; Boyce, M.C.; Li, X.S.; Berko, S. J. Polym. Sci., Polym. Phys. Ed.
1993, 31, 185.
Hutnik, M.; Gentile, F.T.; Ludovice, P.J.; Suter, U.W.; Argon, A.S.
Macromolecules 1991, 24, 5962.
Utz, M. Ph.D. thesis, Eidgenossische Technische Hochschule, Zurich, 1998.
Utz, M.; Atallah, A.S.; Robyr, P.; Widmann, A.H.; Ernst, R.R.; Suter, U.W.
Macromolecules 1999, 32, 6191.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

13
Polymer Melt Dynamics
WOLFGANG PAUL
Germany

Johannes-Gutenberg-Universitat Mainz, Mainz,

MARK D. EDIGER
Wisconsin, U.S.A.

University of WisconsinMadison, Madison,

GRANT D. SMITH

University of Utah, Salt Lake City, Utah, U.S.A.

DO Y. YOON

I.

Seoul National University, Seoul, Korea

INTRODUCTION

Polymer melts and glassy amorphous polymers constitute many of the

plastic materials we are used to in everyday life. Besides this technological importance, there is also fundamental interest in the structural
and dynamical properties of complex molecules with their connectivity
constraints. We will focus in this chapter on the use of chemically realistic
polymer models for the study of the static and dynamic behavior of polymer
melts with an excursion into the glass transition dynamics of polymer melts
as revealed by simulations of a coarse-grained bead-spring model. The basic
concepts and algorithms we will present are valid for both types of models.
With the chemically realistic models we can aim for a quantitative
comparison with experiment and a detailed mechanistic understanding of
segmental motions in real polymers. The bead-spring type models allow
for studies of universal polymer properties and in general for simulations
in a broader parameter range (lower temperatures, larger chains, longer
simulation times).

II. MODELS AND DATA STRUCTURES

The types of models we will be using have already been described in
Chapter 1. The potential energy can be written as
V V B V NB

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

408

Paul et al.

with the bonded energy terms

X
X
X
VB
Vb
V

V


given as
V b constraint

1
V b kb b
b0 2
2
1
or V
k0
cos

cos
0 2
2

or

1
V
k



0 2
2
6
X
V
An cosn

2
3
4

n0

The force constants in the bond length potential of chemically realistic

models generally are the largest of all the energy parameters and therefore
lead to high frequency oscillations. These very fast motions are often constrained in the simulation for performance purposes. For small oscillations
around the equilibrium angle the two forms for the bond angle potential V

can be transformed into each other. Typically one considers at most six
terms in the expansion of the dihedral angle potential V  (see Fig. 1) and
some of the coefficients may be zero. The parameters in the bonded interaction potentials can be mostly obtained from spectroscopic information,
with the exception of the height of the barriers in the dihedral potential. A
quantitatively accurate determination of these is only possible from quantum chemical ab initio calculations [1,2]. Structural relaxation of the chain
conformation as well as of the configuration of the melt requires transitions
between the isomeric states of the dihedral potential. The crossing of the
barriers is a thermally activated process and therefore exponentially sensitive to the height of these barriers.
For the nonbonded interactions in Eq. (1) we will consider the following
contributions. A repulsion/dispersive interaction either of the LennardJones form

12
6 !


V LJ rij 4


5
rij
rij
or of the exponential-6 form
"

6 #
r0
V E6 rij  expf
rij =g

rij

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Polymer Melt Dynamics

409

FIG. 1 Dihedral potential for polyethylene (PE). Eg is the transgauche energy

difference and Etg and Ecis are the barriers for torsional transitions.

where  and r0 are the length scales of the repulsive and dispersive parts,
respectively, and a Coulomb term
V q rij

1 qi qj
40 rij

The partial charges qi as well as repulsive interactions can be parameterized

from quantum chemical calculations. The parameters of the dispersion
forces are not true two-body interaction terms but rather are parameters
for an effective two-body potential appropriate for a condensed phase environment: they are parameterized by comparison with thermodynamic properties such as pVT data, heats of vaporization, or phase behavior. In general
one would also need to consider the polarizability of the atoms to account
for time dependent partial charges, but we can neglect these effects for the
systems we will discuss in the following.
We will consider three different classes of models: all or explicit atom
models, united atom models, and bead-spring models. All fit into the above
description and can be simulated using the same MD code such as described
for instance in [3]. The nonbonded interactions are calculated from the atom
positions and atom types of the interacting pairs. The long-range Coulombic
forces need a special treatment for an efficient simulation, and this problem is discussed in Chapter 8 by Andrey V. Dobrynin. The standard way to

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

410

Paul et al.

treat the bonded interactions is to use topology files for the 2-body
(bond lengths), 3-body (bond angles), and 4-body (dihedral angles)
interactions. These files describe the chain structure in specifying which
atom is bonded to which others, which three atoms participate in a given
bond angle degree of freedom, and which four atoms participate in a given
dihedral angle degree of freedom. The bonded interactions can then be
calculated by working through these lists. In this way a bead-spring model
and an explicit atom model just differ in the topology files specifying the
dynamic degrees of freedom, and the potential form and parameter set to
use with one specific degree of freedom.

III. STARTING STRUCTURES AND EQUILIBRATION

When the model is specified, we have to generate an equilibrated starting
configuration, from which a time series of configurations will be calculated
for later measurement of static and dynamic properties. The thermodynamic
state point for the simulation can be specified by fixing the number of
particles in the simulation box, N, and any of the pairs energy/volume
(NVE), temperature/volume (NVT), or temperature/pressure (NpT) and, of
course, the chain length, K. Experiments are done at constant pressure and
temperature. For the simulations one usually fixes the temperature by using
a thermostat like the NoseHoover [4] one and proceeds in the following
way. The simulation is performed at a preselected pressure until the time
averaged density reaches a stationary value and then one switches to a
constant density simulation at the thus determined density. For this
simulation the pressure will fluctuate around the selected equilibrium
pressure, for instance ambient pressure, and the static and dynamic
properties in the simulation will be representative of that pressure. This
argument assumes that the system size is reasonably large, so that the
fluctuations are small compared to the average value. When one uses a
chemically realistic model with a carefully validated force field, one will find
that the equilibrium density obtained in that way will only differ by about
1% from the experimental density of that polymer at the selected
temperature and pressure [2]. If the force field is not quantitatively correct,
comparison with experiment will turn out better when performed at the
same pressure and not at the same density.
To perform the equilibration discussed above, we first have to generate
a dense configuration of nonoverlapping polymers. There are two standard ways to do this. The first is to start with a big simulation volume in a
dilute situation and to stepwise decrease the volume in a first part of the
simulation until the desired pressure is approximately reached. The second

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Polymer Melt Dynamics

411

way is to estimate the target density and grow the chains into a box of the
corresponding volume without considering the nonbonded interactions.
In the first part of the simulation the repulsive part of the nonbonded
interaction is then regularized to a finite value at distance zero,
which is in turn increased in a stepwise fashion until all overlaps have
disappeared [5].
When we have generated the system at the desired temperature and
pressure/density we still have to ensure that it is equilibrated with respect
to the other observables of the system. In general, equilibration of a
system can be judged by looking at the quantity with the longest
relaxation time, which in an amorphous polymer melt is the end-to-end
distance of the chains. When we are looking at chains which are long
enough to obey random walk statistics for the end-to-end distance, this
equilibration can be quantified by looking at the Gaussianity of the endto-end vector distribution P(R) [6] or at the decay of the end-to-end vector
autocorrelation function hRt  R0i. Typically, properties which are
sensitive to shorter distance behavior equilibrate faster than those living
on larger scales. When there is no coupling between the length scales, it
may be sufficient to equilibrate the scale one wants to study. For instance,
when looking at the structural relaxation in a polymer melt as a function
of temperature, it may be sufficient to ensure equilibration on the scale of
one (or a few) statistical segments of the chain, but this can only be
decided a posteriori, if one has independent measurements to compare
with. If there is a coupling to larger scale behavior one has a quenched
structure and static as well as dynamic quantities will depend on the
quenching rate.

IV. STATIC PROPERTIES

The first static property characterizing the system is the pVT behavior,
which was already discussed in the last section. Besides this thermodynamic
information, an amorphous polymer melt is mainly characterized through
two types of structural information: the single chain structure factor and the
overall structure factor of the melt.
The single chain structure factor in Fig. 2 [7] is calculated using the united
atom positions with an unspecified value for the coherent neutron scattering
length and therefore normalized such that the q 0 value is unity. This is
valid if we are only interested in the low q behavior of the scattering of a
homopolymer (one type of scatterer), i.e., the first curved part, and the
straight line in the double logarithmic plot indicating a self-similar structure
on these length scales. This part of the structure factor can be fitted with the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

412

Paul et al.

FIG. 2 Single chain structure factor for a united atom model of polyethylene
calculated for neutron scattering [7].

Debye function, which is the analytical expression for the structure factor of
a random walk,

2 
8
SDeb qRg 4 4 expf
q2 R2g g
1 q2 R2g
q Rg
which is the best experimental way to determine the radius of gyration of the
chains. In the simulation it can be obtained in the same way and also by
direct calculation
R2g

K
1X
ri
Rcm 2
K i1

where K is the number of atoms per chain and Rcm is the center of mass of
the chain. For larger q the structure factor will be sensitive to the actual
positions of the true scattering centers in the melt, i.e., in this case the
positions of the carbons and deuteriums and their respective coherent scattering lengths. This is also true for the global structure factor of the melt,
which therefore has to be calculated as a superposition of the partial structure factors of all the atom types
Scoh q 1

M X
M
 X

 coh

q
x x coh  coh h^ q

1 1

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

10

Polymer Melt Dynamics

413

Here  and label the M different atom types, the x are their number
fractions, the coh are their coherent scattering cross sections,  is the density
of scattering centers, and the average coherent scattering cross section is
defined as
 coh

M
X

x coh

11

1

The functions h^ q F g r
1 are the Fourier transforms of the pair

correlation functions
*
+
N
N X
X


1
 jri
rj j
r
g r
Nx x i1 j1

12

For the calculation of the melt structure factor the indices i and j in the
above equation run over all atoms in the simulation volume, whereas for the
calculation of the single chain structure factor they only run over the atoms
of that chain. Equation (10) is written in a notation suggestive of neutron
scattering, when we only have nuclear scattering of the neutrons and the
scattering amplitude can be expressed as a scattering length. When we want
to compare a simulated structure with X-ray scattering experiments on that
structure, we have to take into p
account
the wavelength dependence of the
atomic form factor and replace coh by f q.
The above discussed information is, of course, contained in an explicit
atom simulation, and to a certain degree also contained quantitatively in a
united atom simulation. This information is only qualitatively contained in
a bead-spring type simulation. For the united atom model we can go beyond
calculating the scattering of the united atom centerswhich would be
similar to the bead-spring systemand reinsert the hydrogen atoms into
the simulated configurations [8]. For this, the hydrogen atoms are placed
into their T 0 mechanical equilibrium positions (C-H bond length,
C-C-H bond angle, and H-C-H bond angle are fixed to their mechanical
equilibrium values) so that all thermal fluctuations of these degrees of
freedom are neglected. Nevertheless, the distribution of scattering lengths in
the simulated volume resembles closely that of an explicit atom simulation.
Figure 3 shows the X-ray structure factor of the alkane C44H90 from a
united atom [8] and an explicit atom simulation [9]. The intensities are not
normalized as in Eq. (10). For both simulations the position and width of
the amorphous halo (first peak) correspond well with experimental results.
The amorphous halo is the most prominent feature of the structure factor

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

414

Paul et al.

FIG. 3 X-ray structure factor of C44H90 compared between a united atom and an
explicit atom simulation [8].

and representative of the amorphous state (melt as well as glass). Its

properties qualitatively agree between chemically realistic and coarsegrained models. Upon constant pressure cooling of a polymer melt,
the position of the amorphous halo moves to larger q and its width
decreases, but there is no additional structure appearing upon glassy
solidification.

V. DYNAMIC PROPERTIES
In the discussion of the dynamic behavior of polymer melts we will start
with the local reorientational motion of chain segments. Experimentally this
motion can for example be observed in Nuclear Magnetic Resonance
(NMR) experiments measuring the 13C spin lattice relaxation time. Owing
to the different chemical shift for different environments of these nuclei, this
measurement can be very specific for selected positions in the polymer chain
[810]. It turns out that this reorientation dynamics is mainly determined
by torsional transitions of dihedral angles adjacent to the relaxing CH
vector. We discussed above that these transitions are exponentially sensitive
to the barrier heights in the torsion potentials. These in turn can only be
determined with an uncertainty of about 100K even through high-level
quantum chemical calculations. Furthermore, we introduce some error in
the torsional dynamics when we use a united atom representation of the
chain. Although both errors may be small, they get exponentially amplified in the torsional transition rate and therefore a comparison with

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Polymer Melt Dynamics

415

experimental spin lattice relaxation time measurements (or their equivalent)

is a crucial test to validate the torsional force fields and to perform some
final small adjustments of barrier heights, when necessary.
For the dipolar relaxation mechanism, the spin lattice relaxation time
is sensitive to the reorientation dynamics of the CH bond vectors. Different orientations of the CH bond result in slightly different magnetic
fields at the carbon nucleus and the modulation of this field allows the
spin flips to occur. When we define e^CH as the unit vector along a CH
bond, the second Legendre polynomial of its autocorrelation function is
given by
P2 t


1
3he^CH t  e^CH 02 i
1
2

13

For an isotropic system the spin lattice relaxation time T1 and a related
quantity, the Nuclear Overhauser Enhancement (NOE), can be determined
from the power spectrum of this relaxation function.
J!

1
2

P2 tei!t dt

14

1

1
K J!H
!C 3J!C 6J!H !C
nT1
NOE 1

H
6J !H !C
J !H
!C
C J !H
!C 3J !C 6J !H !C

15

16

Here !C and !H are the Larmor frequencies of carbon and hydrogen

atoms respectively, n is the number of bound hydrogens per carbon,
K 2.29  109 s
2 for sp3 hybridization, K 2.42  109 s
2 for sp2 hybridization, and  H and  C are the gyromagnetic ratios of hydrogen and carbon
( H/ C 3.98).
The CH vector autocorrelation function can be readily measured in a
simulation and resolved for specific positions on the chain. The angle
brackets in Eq. (13) indicate an averaging procedure over chemically
identical carbons in the system and over multiple time origins along the
generated trajectory of the model system. To calculate its Fourier transform,
this relaxation function is then best fit either with some model predictions,
for instance based on local conformational dynamics plus Rouse mode
contributions [11], or with some phenomenological fit function. In our
experience, P2(t) is generally very well reproduced by a superposition of an

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

416

Paul et al.

FIG. 4 Temperature dependence of the spin lattice relaxation time for two different
alkanes as seen in experiment (performed at 75 MHz carbon resonance frequency)
and MD simulation [11].

exponential part and a stretched exponential KohlrauschWilliamsWatts

(KWW) part.
(

 )


t
t
1
A exp

P2 t A exp

KWW
exp

17

For short PE chains the relaxation behavior of the four terminal carbons at
the chain end can be resolved in the NMR experiment and the inner part of
the chain gives rise to a collective resonance. In Fig. 4 we show the behavior
of the spin lattice relaxation time for these inner carbons as a function of
temperature for two alkanes of different chain length, n-C44H90 and
n-C100H202. This seemingly local quantity depends strongly on the chain
length even for these center carbons [11]. Inspecting directly the P2 function
obtained in the simulation we can see that this is due to the fact that a few
percent of the correlations measured by this function decay not through
local motion but through a coupling to the much slower overall conformational relaxation of the chains. This effect is also nicely captured in the
simulations. For the united atom simulations of C44 we can furthermore
see two sets of data points. The one agreeing perfectly with the experimental
results stems from an improved force field, where the torsional barriers in

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Polymer Melt Dynamics

417

Fig. 1 were slightly raised to improve the agreement with the experimental
data. Other experimental observables such as dielectric relaxation [12] and
the coherent dynamic structure factor for neutron spin echo [13] have also
been calculated from simulations and used to judge the realism of the local
relaxation processes in the simulations.
Simulations of local relaxation in polymers have provided several
important qualitative insights about the mechanisms of molecular motion.
When a conformational transition occurs, nearby atoms along the chain
backbone must adjust their positions to accommodate the transition. This
cooperativity length scale is not accessible experimentally and identifies the
fundamental kinetic unit (which may be the relevant motion for dynamics
near the glass transition temperature). Simulations on polyisoprene [14],
polyethylene [8,15], and other flexible polymers indicate that 46 carbon
atoms along the chain backbone are involved in localizing a conformational
transition. Simulations have also shown significant spatial heterogeneity in
conformational transition rates [15] and the importance of motions other
than conformational transitions in local relaxation processes [16,17].
For an analysis of the overall conformational relaxation of a polymer
chain we focus here on the chain dynamics on length scales which show the
self-similar random walk behavior statically. Imagine looking at a given
polymer chain in a stroboscopic fashion and from a certain distance. The
center of mass of the chain will move (at long times it has to diffuse) and the
conformation of the chain will change from one random walk realization to
a different one. The simplest model for polymer melt dynamics which
captures these effects is the Rouse model [18], in which the polymers are
modeled by a chain of phantom particles connected by entropic springs (this
generates the Gaussian random walk conformations) moving under the
impact of stochastic forces in an unstructured background characterized by
a friction coefficient.
dri t

p
3kB T
ri1 t
2ri t ri
1 tdt 2kB T dWi t
2
bK

18

The statistical segment length or Kuhn length bK is the step length of the
random walk with the same large scale structure as the chain under study
and the segmental friction  and the strength of the stochastic force are
connected through the fluctuation-dissipation theorem
hdWi tdWj t0 i 2kB Tdtij  t
t0

19

Equation (18) can be solved analytically through a transformation to the

normal modes of a harmonic chain and many exact predictions can be

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

418

Paul et al.

derived. Since there is only a random force acting on the center of mass of a
chain moving according to the Rouse model, the center of mass motion is
purely diffusive with a diffusion coefficient D kBT/K. Of interest to us is
an analytical prediction for the single chain coherent intermediate dynamic
structure factor. This is the dynamic counterpart to the static single chain
structure factor we discussed earlier on and this quantity is measured in
Neutron Spin Echo (NSE) experiments. The Rouse model yields [18]
Sq; t



1
exp
q2 Dt
K
 2 2
K
X
q bK
2Kq2 b2K

exp

jn
mj

6
32
n;m1


K
X
p1

cos

pn
pm
2
cos
1
e
p t= R
K
K

20
)


where R K 2 b2K =32 kB T is the so-called Rouse time. For small momentum transfers q, one is only sensitive to the motion of the polymer chain as a
whole and the above equation simplifies to the scattering law for diffusive
motion S(q, t) S(q, 0) exp(
Dq2t). In this way one can determine the center
of mass self-diffusion coefficient from the NSE data and of course also from
the computer simulation. The computer simulation furthermore also has
access to this property by measuring directly the average mean square displacement of the center of mass of a chain and using Einsteins relation
hR2cm ti
t!1
6t

D lim

21

Through the diffusion coefficient the segmental friction  is determined and

from the single chain static structure factor Eq. (10) one can obtain the
radius of gyration and from that the statistical segment length bK entering
the Rouse model. Again, in the simulation this quantity can be independently determined by direct calculation of the mean squared end-to-end
distance of the chains. In this way both parameters (bK, ) entering the
analytical predictions of the Rouse model can be determined and the
model critically tested. This can be done on the coarse-grained beadspring model level as well as with a fully atomistic model and it has been
found that the Rouse model gives a good quantitative description only on its
largest intrinsic length and time scales. On small length scales (large q) the
local structure and stiffness of the chains cannot be captured by entropic
springs and on small time scales it turns out that the interactions between

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Polymer Melt Dynamics

419

the chains in the melt cannot be captured by a homogeneous background

friction  but lead to a subdiffusive motion of the centers of mass of the
chains [1922]. It also leads to a non-Gaussian distribution for the mutual
displacement |ri(t)
rj(0)| of two atoms on the same chain [22].
This, in principle, invalidates all theoretical predictions like the Rouse
result for the intermediate dynamic structure factor, because they all rely on
the dynamic Gaussian assumption
2

2
q 
hexp iq  ri t
rj 0 i exp
h ri t
rj 0 i
6




22

In the simulation one can prove that the above assumption is not fulfilled by
calculating the scattering function correctly averaging over the phase factors
[left hand side of Eq. (22)] or by employing the dynamic Gaussian assumption [right hand side of Eq. (22)]. Figure 5 shows this comparison for C100 at
509K. One can clearly see, that the dynamic Gaussian assumption leads to
an overestimation of the decay of correlations and to a wrong prediction of
the scattering.
A quantitative comparison with experiments, however, can only be performed for a chemically realistic simulation. For this the positions of the
scatterers in the time dependent generalization of Eqs. (10) and (12) run over

FIG. 5 Comparison of single chain intermediate dynamic structure factor of

polyethylene at T 509K for the correct calculation and one assuming Gaussian
distributed relative atom displacements.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

420

Paul et al.

FIG. 6 Comparison of single chain intermediate dynamic structure factor of

polyethylene at T 509K between NSE experiments and chemically realistic MD
simulations [21].

the carbon and deuterium atoms of a single chain and are evaluated with a
time displacement t. Figure 6 shows a comparison of the experimental result
for the single chain intermediate dynamic structure factor of polyethylene at
T 509K with simulations of a united atom as well as an explicit atom
model. For this comparison actually only the deuterium scattering has been
evaluated. When plotting this quantity against scaled time to account for a
20% difference in the observed center of mass self-diffusion coefficients,
the relaxation behavior of the chains on all length scales agrees perfectly
between simulation and experiment [10,21].

VI. GLASS TRANSITION

So far we have discussed static and dynamic behavior of high temperature
polymer melts of unentangled chains. Upon lowering the temperature many
polymeric systems do not form ordered structures but remain in an
amorphous state, structurally identical to the melt state but with typical
relaxation times increased by up to 14 orders of magnitude [23]. This so-called
glass transition is a gradual phenomenon with an onset that is observable
in the slightly undercooled melt around a temperature where the typical
time scales have increased by about 3 orders of magnitude from their high
temperature values. This is a range of time scales that at the moment is

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Polymer Melt Dynamics

421

accessible only for simulations of coarse-grained models when one wants

to perform equilibrium cooling. If one performs simulations at lower
temperatures, one typically is in a quench situation where the results in
general depend on the quenching rate one uses. For MD simulations this
quench rate is at least 9 orders of magnitude larger than experimental rates.
There may be properties of the supercooled melts which decouple from the
slowest degrees of freedom in the melt but it is a priori unclear which these
are and the assignment may even differ from one polymer to another. We
have already seen that even a seemingly local quantity like the CH vector
reorientation as measured by spin lattice relaxation time experiments couples
to the overall conformational relaxation of the polymer chain, which defines
the slowest mode in the system. It has been found phenomenologically in
some polymers [24], that MD simulations employing high quench rates were
actually able to identify the glass transition temperature of these polymers
through the jump in the thermal expansion coefficient rather accurately, but
this is not a general finding.
We will focus now on the temperature region where the onset of the
glassy freezing first starts to be felt. This temperature region is the regime of
applicability of mode coupling theory (MCT) [25], which predicts a slowing
down of the relaxation processes in the system because particles get trapped in cages formed by their neighbors. This caging process leads to the
development of a two-stage relaxation behavior, which is, for instance,
observable in intermediate scattering functions (experiment and simulation)
or directly in the mean square displacement of the atoms (simulation). The
theory is formulated for simple liquids like hard-sphere or soft-sphere
systems and we will therefore ask ourselves in the following why and to
what degree it can describe polymer behavior.
For this we will use a type of scattering function we have not discussed
so far, which is the incoherent intermediate scattering function. Due to its
large incoherent scattering amplitude it is mainly the self-correlation of the
hydrogen motion that is seen in incoherent neutron scattering experiments.
Sinc q; t

M
1 X
expfiq  rj t
rj 0g
M j1

23

where M is the number of hydrogens in the system and where we dropped the
scattering amplitude which only adjusts the absolute scale. For an isotropic
amorphous melt one can perform a spherical average of this equation to get
Sinc q; t

M
1 X
sinqjrj t
rj 0j
M j1 qjrj t
rj 0j

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

24

422

Paul et al.

FIG. 7 Intermediate incoherent scattering function of a bead-spring polymer of

length 10 beads, approaching the mode coupling temperature which is T 0.45 in
Lennard-Jones units [26].

We will use this equation to look at the incoherent intermediate structure

factor of a bead-spring model without any bond angle or dihedral potential
as we approach this MCT temperature region [26,27]. Here the index j runs
over all beads in the system. We will do this for the q-value of the position of
the amorphous halo in the liquid structure factor of the model. As we can
see in Fig. 7 a two-step relaxation process develops upon cooling which can
alsofor this simplified polymer modelbe reasonably well analyzed in
terms of the predictions of MCT [26]. The long time behavior of these
curves can be fitted by a KWW form and the time scale entering this function is the so-called structural or -relaxation time scale that shows the
dramatic increase by more than 14 orders of magnitude between the melt
and the glass. This time can also be defined by looking at the decay to the
30% level indicated by the horizontal line.
When we look directly at the mean square displacements of the beads in
the simulation which is shown in Fig. 8, we can see the caging by the
neighbors resulting in a plateau regime. The height of the plateau is only a
fraction of the bond length in the model (l 1), so that effectively the beads
have actually not yet noticed that they are bound to a chain and are not
free particles, which is why the theory which was developed for simple
liquids could be reasonably applied in this case [26,27]. The same behavior
can be observed in the center of mass motion of the chains. The long time

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Polymer Melt Dynamics

423

FIG. 8 Master curve for the mean square displacements of the beads and of the
centers of mass of the chains obtained by plotting the common part of the mean
square displacement curves for all temperatures as a function of time scaled by the
respective diffusion coefficient of that temperature [27].

behavior of the mean square displacements is still the same as would be

predicted by the Rouse model we discussed above. The monomer friction
coefficient, however, that sets the typical time scales for the Rouse modes,
has been strongly increased due to this caging effect that is responsible for
the glass transition.

VII.

OUTLOOK

In this section we have discussed simulations of polymer melt dynamics

using chemically realistic as well as coarse-grained polymer models. We
have seen that both have their complementary uses and that these studies
can help to test theories as well as help in the interpretation of experimental
data. With the development of ever faster computers and the invention of
highly sophisticated equilibration techniques it will become possible to
extend the range of temperatures and chain lengths that can be studied
in simulations significantly and this will further enhance the contribution
that computer simulations of polymer melts and glasses can make to
our understanding especially of the relaxation processes occurring in these
systems.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

424

Paul et al.

REFERENCES
1. Smith, G.D.; Jae, R.L. J. Phys. Chem. 1996, 100, 18718.
2. Smith, G.D.; Paul, W. J. Phys. Chem. A 1998, 102, 1200.
3. Lucretius MD code and sample data les for PE: http://www.che.utah.
edu/gdsmith/mdcode/main.html.
4. Nose, S. J. Chem. Phys. 1984, 81, 511.
5. Kremer, K.; Grest, G.S. J. Chem. Phys. 1990, 92, 5057.
6. Yoon, Do, Y. unpublished.
7. Paul, W.; Smith, G.D.; Yoon, Do.Y. Macromolecules 1997, 30, 7772.
8. Paul, W.; Yoon, Do.Y.; Smith, G.D. J. Chem. Phys. 1995, 103, 1702.
9. Smith, G.D.; Yoon, Do.Y.; Zhu, W.; Ediger, M.D. Macromolecules 1994, 27,
5563.
10. Smith, G.D.; Paul, W.; Monkenbusch, M.; Willner, L.; Richter, D.; Qiu, X.H.;
Ediger, M.D. Macromolecules 1999, 32, 8857.
11. Qiu, X.H.; Ediger, M.D. Macromolecules 2000, 33, 490.
12. Smith, G.D.; Yoon, Do.Y.; Wade, C.G.; OLeary, D.; Chen, A.; Jae, R.L.
J. Chem. Phys. 1997, 106, 3798.
13. Moe, N.E.; Ediger, M.D. Phys. Rev. E 1999, 59, 623.
14. Moe, N.E.; Ediger, M.D. Polymer 1996, 37, 1787.
15. Boyd, R.H.; Gee, R.H.; Han, J.; Hin, Y. J. Chem. Phys. 1994, 101, 788.
16. Moe, N.E.; Ediger, M.D. Macromolecules 1995, 28, 2329.
17. Moe, N.E.; Ediger, M.D. Macromolecules 1996, 29, 5484.
18. Doi, M.; Edwards, S. The Theory of Polymer Dynamics; Clarendon Press:
Oxford, 1986.
19. Paul, W.; Heermann, D.W.; Kremer, K.; Binder, K. J. Chem. Phys. 1991, 95,
7726.
20. Kopf, A.; Dunweg, B.; Paul, W. J. Chem. Phys. 1997, 107, 6945.
21. Paul, W.; Smith, G.D.; Yoon, Do.Y.; Farago, B.; Rathgeber, S.; Zirkel, A.;
Willner, L.; Richter, D. Phys. Rev. Lett. 1998, 80, 2346.
22. Smith, G.D.; Paul, W.; Monkenbusch, M.; Richter, D. J. Chem. Phys., 2001,
114, 4285.
23. Proceedings of the Third International Discussion Meeting on Relaxation in
Complex Systems, J. Non-Cryst Solids, 1998, 235237.
24. Han, J.; Ge, R.H.; Boyd, R.H. Macromolecules 1994, 27, 7781.
25. Gotze, W.; Sjogren, L. In Transport Theory and Statistical Physics; Yip, S.,
Nelson, P., Eds.; Marcel Dekker: New York, 1995; 801853.
26. Bennemann, C.; Baschnagel, J.; Paul, W. Eur. Phys. J. B 1999, 10, 323.
27. Bennemann, C.; Baschnagel, J.; Paul, W.; Binder, K. Comp. Theor. Polym. Sci.
1999, 9, 217.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

14
Sorption and Diffusion of
Small Molecules Using
Transition-State Theory
MICHAEL L. GREENFIELD
Rhode Island, U.S.A.

I.

University of Rhode Island, Kingston,

INTRODUCTION

In numerous applications, it would be useful if molecular simulation could

be used to predict how the solubility or diffusivity of a small molecule in a
polymer film varied with respect to changes in the molecular architecture
of the film, or with different film processing. For example, this ability
would enable prescreening among the many possible chain structures that
could be devised, so chemical synthesis efforts could be targeted towards
the most promising candidates. Polymer materials for sorptiondiffusion
membranes are one of many fields in which such large synthetic efforts
have been conducted [14]. However, two primary barriers prevent accurate predictions from becoming routine: (1) the absence of reliable potential energy functions for penetrantpolymer systems of interest, and (2) the
need for methods that access the long time scales inherent in smallmolecule diffusion. The development of potential energy functions is the
subject of ongoing research. Efforts focus on general force fields applicable
to wide classes of molecules [514] or to particular types of polymer chains
[1521]. (These are only some of many possible examples.) This chapter
discusses small-molecule sorption and diffusion, with a focus on simulating
penetrant diffusion through polymers over long time scales.
Different simulation methods are appropriate for different penetrant
polymer systems, depending on the dynamical state of the polymer chains.
Within a rubbery polymer or an amorphous polymer above its glass
temperature, penetrants diffuse quickly enough (D  10
6 cm2 s
1) to use
molecular dynamics simulations [22] (see also Chapter 6 of this book).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

426

Greenfield

The diffusive regime is reached after tens to hundreds of picoseconds

[2325], with longer simulation times required in order to accumulate
good statistics. Molecular dynamics has been applied to small-molecule
penetrants in numerous polymers. Early work focused on polyolefins (PE,
PP, PBD, PIB) [17,2337], and work on those systems has continued
[3845]. Other systems studied include polydimethylsiloxane [36,4648],
polystyrene [49,50], polyesters [5153], polyimides and polyamideimides [5458], polysulfones [59], poly(aromatics) [60], polybenzoxazine
[61], polyphosphazenes [62], poly(ether-ether-ketone) [63], poly(trimethylsilylpropyne) [64], and more generic representations [65,66]. Recently,
molecular dynamics simulations of charged species in polymers (for fuel
cell applications) have been conducted, [17,6776]. Molecular dynamics is
also appropriate for moderate to high penetrant concentrations in glassy
polymers, at which point the polymer chains become plasticized and the
penetrant diffusion rate increases to a liquid-like value. Equivalently,
the penetrant concentration suppresses the glass temperature to below the
temperature simulated. Nonequilibrium molecular dynamics simulations
[77,78] have been generally unsuccessful: forces beyond the linear response
regime were required to impact the diffusion rate, relatively simple
polymer models have been required, and computational times have not
provided speedups over equilibrium MD.
In a glassy polymer (amorphous and below its concentration-dependent
glass temperature), diffusion is much slower10
9 cm2s
1 is typicaland
molecular dynamics is incapable of displacing penetrant molecules
sufficiently over accessible simulation times. As an order-of-magnitude
estimate, after 10 ns an average penetrant would be displaced by
r

p 
1=2
6Dt  6  10
9 cm2 s
1  10
8 s  0:75 A

This distance is 30100 times smaller than the typical simulation cell sizes
used. Furthermore, since diffusion through a glassy polymer is thought to
occur by a hopping mechanism [79], such an average displacement more
likely would result only from statistics: one molecule with a 7.5 A jump
and nine molecules with net displacements of zero, for example. Direct
molecular dynamics simulations are thus prone to large errors in predicted
diffusivity in a glassy polymer, due to (1) mismatches between the simulation time and the expected waiting time for a jump to occur and (2) poor
sampling of the many jump rates that would be found in a macroscopic
sample.
One method for predicting diffusion rates in glassy polymers has been
proposed by Boyd and coworkers [49,52]. As in glassy polymers, a hopping
mechanism has been found in melt polymers not far above their glass

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

427

temperatures [23,33,80,81].* Molecular dynamics simulations can be

conducted at several temperatures above Tg, and an Arrhenius dependence
can be used to extrapolate (by 50100K [49,52]) into the glassy region.
However, this method requires the diffusion mechanism to remain
unchanged throughout this large extrapolation.
As an alternative, this chapter describes methods for predicting smallmolecule diffusivity that are based on transition-state theory (TST) [8284]
and kinetic Monte Carlo (KMC) [85,86]. These methods capitalize on the
proposed penetrant jump mechanism. TST was described in Chapter 1 and
is typically used to estimate the rates of chemical reactions from first
principles; here we use TST to calculate the rate of characteristic jumps for
each penetrant in a host polymer matrix. The collection of jump rates can
be combined with the penetrant jump topology and KMC to obtain the
penetrant diffusion coefficient. Other results obtainable from these
simulations are physical aspects related to the jump mechanism: the sizes
and shapes of voids accessible to penetrant molecules [87], enthalpic and
entropic contributions to the penetrant jump rate [88,89], the extent and
characteristics of chain motions that accompany each jump [90], and the
shape and structure of the jump network itself [91].
Three approaches have been applied for describing the coupling between
the jumping of a penetrant and the motions of nearby polymer chains. Each
approach leads to a different approximation for the TST rate constant. The
frozen polymer method [9294] is analogous to methods used successfully to
simulate penetrant diffusion through porous crystalline materials, such as
zeolites (see [95,96] for reviews). All polymer chains are considered fixed in
place, and transition-state theory-based rate constants are calculated from
the energy barriers found for a penetrant to pass from one local potential
energy minimum to another. The jump path is intrinsically threedimensional: (x, y, z) position of the penetrant. This method is the most
straightforward; however, in polymers it yields rate constants that lead to
diffusion coefficients much lower (by factors of 103106) than experimental
values, because neglecting chain fluctuation contributions is physically
unrealistic in a polymeric material. However, we discuss it here in order to
introduce several concepts and techniques used in other TST methods. Next,
a modification that we refer to here as the average fluctuating polymer
method was proposed [93,97]. Harmonic fluctuations of the polymer matrix
about its zero-force position were allowed, with the magnitude of the
fluctuations controlled by a parameter similar to the DebyeWaller factor

*Explicit descriptions of the hopping mechanism can also be found in most of the recent MD
papers. References [80,81] focus specically on the jump mechanism.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

428

Greenfield

in X-ray scattering. The jump path was again confined to the three degrees
of freedom of the penetrant, but now within the free energy field
characteristic of the vibrating polymer. This method has led to reasonable
agreement with experiment in particular cases [24,63,97100] and has been
implemented in a popular molecular simulation software package [101].*
However, the predictions are sensitive to the value of the DebyeWaller
parameter, and hence some judgment is required when using this technique.
In the most recent method [88,91,102], which we refer to as the explicit
polymer method, the dimensionality of the jump path was increased to
include explicitly both penetrant displacements and polymer chain motions.
Each resulting rate constant then captures the details of the particular chain
motions that accompany the jump, rather than mean-field average motions,
but at the expense of a much higher computational cost. This method has
been applied to a few penetrants in polypropylene [88,89,91], with reasonable agreement compared to experiment and correlation.
In this chapter we will begin by formulating the transition-state theory
method within the context of small-molecule diffusion in a polymer matrix,
and then show how the transition-state theory equations are applied to a
single jump in each of these three approaches. We then compare different
methods for establishing the network of possible penetrant jumps and
explain how diffusion can be simulated on that network. We close with
a summary and a discussion of the outlook for other systems. Several
appendices describe aspects of particular approaches in more detail.

II. FORMULATION OF TST METHOD

Transition-state theory is a classical methodology for treating infrequent
event processes that involve crossing a barrier significantly higher than the
thermal energy kBT [82]. The complicated dynamical paths followed by an
N-body system are reduced to a statistical description in which sampling
over local fluctuations is performed in order to answer the question, How
much time goes by (on average) before the system resides in a definably
different portion of configuration space? Such a question may be answered
phenomenologically with a set of first-order rate constants, each describing
the rate of evolution along a different available pathway [103]. Transitionstate theory provides a means to calculate such rate constants by comparing
the probability of the system lying on a dividing surface, which separates
two states, to the probability of occupying an initial state. This makes TST
*The average uctuating polymer method has been incorporated into several products of
Accelrys, Inc., such as Cerius2.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

429

a bridge between kinetics and thermodynamics, since the qualities of a rate

process are determined from equilibrium information.*
A general formulation of transition-state theory was derived in Chapter 1
and will not be repeated here. We focus instead on applying a TST approach
to diffusion in polymers. Several steps are required: stable regions in
configuration space must be located (sorption states for a penetrant
molecule), transition states and dividing surfaces between stable regions
must be identified, pathways along which the system passes between
sorption states must be followed, and the rate of each process must be
calculated. Calculating the small-molecule diffusion rate through a polymer
then requires an appropriate averaging technique for converting the jump
rate constants kTST[ ]s
1 and jump lengths [] A to a diffusion coefficient
D[ ]cm2 s
1.

A. Transition State
For transition-state theory to be applicable, the physical system must
include some high-energy region through which the system must pass in
order to convert from reactant to product. Numerous such regions
exist for a small molecule diffusing through a glassy polymer, and the socalled transition states lie within them. Mathematically, a transition state is
usually taken as a first-order saddle point on the 3N-dimensional potential
energy hypersurface, meaning that all elements of the potential energy
gradient equal zero and all eigenvalues of the Hessian matrix of second
derivatives but one are positive. In Cartesian coordinates,
@V
0 for all positions X i
@X i

The notation X i indicates the ith of the 3N Cartesian coordinates. Writing

the index as a superscript follows the conventions of Riemannian geometry
[104]. Each Eq. (1) can be written in mass-weighted Cartesian coordinates,
i
xi  m1=2
i X ; these are useful in regard to following the diffusion path.
An equivalent expression can then be formulated in generalized coordinates
(bond angles, torsion angles, etc.) by writing [88]
3N
@V X
@q j @V

@xi
@xi @q j
j1

*While the dividing surface technically describes a hypothetical equilibrium, it can be sampled
using equilibrium statistical mechanics.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

430

Greenfield

for each mass-weighted Cartesian coordinate. In matrix notation,

J T rq V 0

where J equals the Jacobian matrix of partial derivatives for transforming

from mass-weighted Cartesian coordinates to generalized coordinates,
1
0 1
@q
@q1



B @x1
@x3N C
C
B
.. C
B ..
4
J B .
C
.
C
B
@ @q3N
3N A
@q

@x1
@x3N
The scalar Jacobian factor that more typically arises in a coordinate
transformation equals det J . For a nondegenerate coordinate system,
det J 6 0; thus the potential energy gradient with respect to the generalized
coordinates equals the zero vector at the transition state.
Generalized coordinates are useful when simulating penetrantpolymer
systems because many points in configuration space are not physically
realistic: they correspond to undesired values for the bond lengths, for
example. Such generalized coordinates may be kept at constant values using
the concept of infinitely stiff degrees of freedom [105]. These degrees of
freedom are considered flexible (i.e., assuming any value according to the
Boltzmann-weighted probability density) during the formulation of the
methodology. However, they are effectively constrained to the desired
positions using springs with force constants that approach infinity.
Compared to an approach using rigid constraints throughout, this
infinitely stiff model has been shown to lead to classical statistical
mechanics results that more closely resemble the true underlying quantummechanical statistical mechanics [105]. In practice, these generalized
coordinates are kept constant throughout the penetrant jump process;
their flexibility only arises in formulating the method.
Finding a saddle point is similar to finding a local minimum or maximum:
an initial guess is refined using steps based on local function information.
However, standard methods for finding stationary points (where the
gradient equals zero) typically are designed for finding maxima or
minima, so special saddle-point-focused algorithms are required. Most
common in a chemical context are those based on work by Cerjan and Miller
[106108]. Rather than taking the step defined by a quasi-Newton approach,
these methods use a shift in eigenvalues to take a step such as [108]
h

X
i

qi  rq V

qi

i

s

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

431

that maximizes in the direction of one approximate eigenvector q1 and

minimizes along all other eigenvectors. The shift factor
s for the first
eigenvector (with a negative eigenvalue) differs from that for all other eigenvectors [107,108]. An initial estimate of the Hessian matrix (and its eigenvalues and eigenvectors) is required. Updated Hessian estimates [109] can
be used in later steps, but an exact Hessian (calculated analytically or estimated numerically) leads to the best results, especially near the transition
state. Bakers description [108] of this procedure is particularly targeted
towards implementation issues. Quick convergence to a saddle point is
aided by a good initial guess; several techniques for doing so are discussed
in Section IV. Methods for obtaining the transition state in conjunction
with the entire path [110,111] are discussed below as well.

B. Jump Pathwaythe Intrinsic Reaction

Coordinate (IRC)
The transition state defines an unstable position in configuration space. A
system displaced slightly from the transition state will generally experience a
force that directs it even further away, and the net displacement will increase
until the system thermalizes in the region surrounding a local minimum.*
Depending on the direction of the initial displacement, the system will find
itself near one of two local minima (assuming the transition state was a firstorder saddle point). These two minima are associated with this transition
state in that they share the common dividing surface that passes through it.
In applying transition-state theory, it is traditional to calculate the
Intrinsic Reaction Coordinate (IRC) [112], which is the lowest-energy path
that connects the transition state to its adjacent local minima. The term
reaction refers to the common application of TST to chemical reaction
dynamics. Here, the IRC refers to the changes in penetrant and polymer
degrees of freedom that describe the mechanism of a particular penetrant
jumping motion. This exact pathway would be followed in the limit of
infinitely small displacements with continuous application of a frictional
force equal and opposite to the forces arising from the potential energy
gradient. A true system undergoing molecular dynamics would not follow
this path exactly; rather, it would fluctuate about it, much like a single melt
polymer chain fluctuates about the chain contour in the reptation theory
of long chain diffusion [113,114].
*It is theoretically possible to displace the system such that forces redirect it to the transition
state. In practice, however, numerical imprecision in following such paths will likely lead to
forces directing the system away from the transition state.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

432

Greenfield

It is convenient to begin calculating the IRC at the transition state. Near

the transition state, fluctuations in the mass-weighted atomic positions can
be combined (by a linear transformation with a Jacobian of unity) into
fluctuations along the set of 3N eigenvectors (or normal modes) x of the
Hessian matrix evaluated at the transition state, which are found by solving
Hx
x

From the transition state, the IRC is directed along the eigenvector corresponding to the negative eigenvalue [115,116]. Motions along all other
eigenvectors locally define a multidimensional dividing surface that delineates the reactant and product regions. The IRC is orthogonal to this dividing surface. In other words, the IRC step through the transition state
(over the pass in a mountain crossing analogy) is directed along the
single direction of decreasing local curvature and orthogonal to all directions
of increasing local curvature.
Away from the transition state, the IRC is formally defined by a steepest
descent trajectory* in mass-weighted Cartesian coordinates [112],
x rx V 

and corresponds physically to a sequence of infinitesimal displacements

along the local force vector. (Higher-order terms that require the Hessian
matrix are discussed below.) The factor  (of units time squared) scales the
size of each step taken [112]. Since the velocity is set to zero in each step,
the IRC equation has also been called a path with continuous dissipation
of kinetic energy [117,118].
Extensions of the IRC to generalized coordinates have been discussed in
the literature [88,102,117]. Physically, the IRC pathway and the vibrational
frequencies and normal modes at the transition state should be independent
of the coordinate system chosen, so Eqs. (6) and (7) must be reformulated
in terms of generalized coordinates. Detailed steps are shown in Appendix A.
The resulting expression at the transition state is
Hqq q
a q

*This usage of steepest descent refers to a path constructed by taking a sequence of small
steps, each directed opposite to the local gradient. This usage diers from that in the optimization literature; there [109], the magnitude of each step is chosen such that the objective function
is minimized with respect to motion along the step vector.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

433

in which Hqq @2 V=@qi @q j . The covariant metric tensor a [119] is calculated

most easily from the Jacobian matrix for transforming from generalized
to mass-weighted Cartesian coordinates,
1
0 1
@x
@x1
B @q1    @q3N C
C
B
B .
.. C
0
B
.
9
J B .
. C
C
C
B 3N
3N
@ @x
@x A

1
@q
@q3N
J 0 is the matrix inverse of J , introduced in Eq. (4),
J J 0 J 0J I

10

Starting from Eq. (9), a is calculated by

a J 0T J 0
or
aij

X @xk @xk
k

@qi @q j

11

Equation (8) is an example of a generalized eigenvalue problem, whose

numerical solution provides the eigenvectors q and eigenvalues
of the
Hessian in the space of all generalized coordinates. The units on each side
match exactly,


X energy

kg m2
qj 2
Hqq q i
q i qj
s qi
j
a qi

X kg1=2 m2
kg m2
qj
qi qj
qi
j

and
has units of s
2, or frequency squared, as expected.
Canned routines that solve the generalized eigenvalue problem are
available in standard numerical libraries.* Because a J 0 TJ 0 , the covariant
metric tensor a must be symmetric and positive definite, which facilitates
finding the eigenvalues and eigenvectors, compared to a more general case
in which the matrix characteristics of a are unknown.
*The routines REDUC, TRED2, and TQL2 of EISPACK, available at [120], are examples of
canned routines. Matrix routines from the LAPACK [121] and LINPACK [122] packages are
also available.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

434

Greenfield

Away from the transition state, the IRC equation is written as [88,102]
a q rq V 

12

Taking a step thus requires calculating the gradient and covariant metric
tensor, and then solving a set of linear equations.

C. Narrowing the Diffusion Path to a

Localized Region
Since a penetrant jump occurs over a relatively small region within a
simulation cell, it is useful to restrict the atoms involved in the TST
approach, leaving only the penetrant and some of the polymer chain
segments. This can significantly reduce the number of variables and thus the
size of the matrix calculations (which scale as N3). Constraining the system
requires a careful manipulation of the mass-weighted Cartesian coordinates
and the generalized coordinates. The origins of the nontrivial nature lie in
the covariant metric tensor a. Since this matrix is not diagonal in general,
it couples the step along one coordinate dqi to the gradient with respect
to many q j( j 6 i). The coordinates qi and q j must correspond to part of the
same chain for aij to be nonzero.
Choosing which generalized coordinates to include is facilitated by
dividing them into three sets. One set, the bond lengths, are considered
infinitely stiff and should remain unchanged between the transition state
and the local minima. In a sufficiently large penetrantpolymer system,
another set is comprised of the many generalized coordinates defined by
atoms far from the penetrant. If a local fluctuation among polymer chains
allows a penetrant to jump, the values of far-off coordinates should not be
appreciably different at the origin state, transition state, or destination state.
A third set, those nonconstrained generalized coordinates defined by atoms
near a jumping penetrant, are flexible when describing a transition state
and IRC. Both other sets of degrees of freedom are held constrained to their
initial values (in the energy-minimized neat polymer structure) by springs
with force constants whose strengths approach infinity.
The details of narrowing the diffusion path are explained in Appendix B.
Here we quote the resulting TST equations and algorithms.
At the transition state, Eq. (3) is replaced by the similar expression
rqf V 0

13

meaning that the potential energy gradient with respect to each flexible
coordinate equals zero.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

435

The eigenvalues and eigenvectors at the transition state, with respect to

the flexible coordinates, are calculated from
H0qq qf
a0 qf

14

The step to take away from the transition state is the eigenvector qf corresponding to the negative eigenvalue of the generalized eigenvalue problem
Eq. (14). The first step along the IRC (from the transition state) is found
by the following steps:
1.
2.
3.

calculate the Hessian matrix of second derivatives with respect to the

flexible generalized coordinates, H0qq ,
calculate the covariant metric tensor in the flexible generalized
coordinates, a0,
solve for the eigenvector that corresponds to the one negative
eigenvalue.

A small step along this eigenvector qf defines the first step along the IRC.
Along the step, changes occur in only the flexible coordinates; infinitely stiff
coordinates remain unchanged.
Away from the transition state, the remaining steps along the diffusion
path are calculated from
a0 qf rqf V 

15

Tracking all but the first step along the IRC thus requires these steps:
1.
2.
3.
4.

calculate the potential energy gradient in the flexible subset of

generalized coordinates,rq fV,
calculate the covariant metric tensor a0 in the subset of flexible
coordinates,
solve a series of f linear equations for the step vector qf,
scale each term by the same small constant  .

The elements of the Jacobian matrices (and thus the elements of a0) vary
with chain conformation and must be recalculated after each step. The
IRC ends when it reaches a local minimum (the gradient rqfV approaches
zero).

D. Final State(s)
The two IRCs that begin at a transition state each terminate at a local
potential energy minimum, where the gradient equals zero. Near the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

436

Greenfield

minimum, the gradient becomes sufficiently small (a step size of order 10
10,
for example [88]) that each IRC step [Eq. (15)] is excessively short. It would
then take a large number of steps to travel the small remaining distance to
the local minimum configuration. Instead, a quasi-Newton scheme [109]
can be used at this point to reach the local minimum. Relatively few steps
should be required, and the resulting potential energy decreases should be
negligible. If large decreases in energy do occur, that indicates larger-scale
structural relaxation of the polymer matrix, and the system configuration
prior to quasi-Newton steps should be restored [i.e., the last configuration
found via Eq. (15)]. In such a case, additional regular IRC steps should be
taken despite the low gradient, and the step size should soon increase again.
This approach maintains the validity of the calculated IRC. Such behavior
could occur along a near-flat plateau on the potential energy landscape that
is surrounded by more steep regions of similar orientation. After continued
energy decreases, a small gradient should recur, and quasi-Newton steps can
be attempted again.

E. Rate Constant
The pair of rate constants for each jump path (IRC) can be calculated in
one of several ways. A formal, microscopic formulation [83] was discussed
in Chapter 1 for the rate constant,




q_ 0 qz
q0
qt
qz
z

kA!B t

q
q

1:101

The denominator equals the contribution QA of the origin sorption state A

to the entire canonical partition function Q. Evaluating QA requires some
definition of the bounds on the state volume in configuration space. The
numerator equals the average velocity along the reaction coordinate for
state points within the dividing surface region at t 0 that terminate in
the product state. Transition-state theory assumes that all state points in
this region with q_ > 0 terminate in the product state, i.e., (q(t)
qz)>0
for all t [83]. If this velocity along the IRC (in some combination of
mass-weighted Cartesian coordinates) is position-independent and separable
from all other velocities, it may be integrated over 0  q_ < 1, leading to a
prefactor kBT/h, where h is Plancks constant. The remaining numerator
terms, in combination with a definition of the dividing surface, equal the
dividing surface region contribution Qz to the canonical partition function.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

437

Equation (1.101) can thus be rewritten as

kTST


kB T Qz
Vz
V0
exp

h Q
kB T

16

or as the canonical ensemble equivalent of Eq. (1.117),

TST


kB T
Az
A0
exp

h
kB T

(from Chapter 1). The exponential in Eq. (16) arises from a difference in the
zero of energy between Qz and Q.
Several approximations to Eq. (16) can then be made. The simplest
approximation is to assume a constant prefactor and to use the potential
energy difference between the transition state and the reactant sorption
state,



Vz
V0
kTST k0 exp

kB T

17

This approach can be interpreted via Eq. (1.117) as assuming that the free
energy difference equals the potential energy difference (entropy change
equals zero) and that the temperature dependence in the exponential
dominates that in the prefactor.
A less drastic approximation is to assume that (1) potential energy
increases relative to the origin or transition state energies are adequately
described locally by a harmonic approximation, and (2) potential energies
outside these harmonic regions are sufficiently large such that the
Boltzmann factor approaches zero. With this combination of assumptions,
the origin state and dividing surface boundaries can be extended to 1, and
the partition functions can be calculated from the vibrational frequencies at
the transition state and the reactant state, with respect to the flexible degrees
of freedom. The frequencies are obtained from the eigenvalues of Eq. (14) by
i

i
2

i 2, . . . , f transition state
i 1, . . . , f reactant state

18

Note that no frequency calculation is required for the negative eigenvalue at

the transition state. Two alternate forms of the rate constant can then be

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

438

Greenfield

obtained. Within a classical statistical-mechanics model, infinitely stiff

coordinates [105] contribute a factor related to the product of their
normal mode frequencies, each of which approaches infinity [88]. The
resulting rate constant is
0Y f
TST

@Y

f
2


00
a
1
Vz
V0
A det
 00 z exp

kB T
z det a
1


1 

19

00

The matrix a
1 , defined in Appendix B, is a contravariant metric tensor

00
in the infinitely stiff coordinates. Each det (a
1 ) term can be evaluated
directly using a recursion relation [102,123] or via the Fixman relation,
Eq. (82). In a quantum mechanics-based statistical mechanics model, the
effect of infinite stiffness on each partition function is not*

lim

!1

kB T
0
hi

but instead





hi
lim 1
exp

1
!1
kB T
Consequently the contributions from infinitely stiff modes become unity and
the rate constant can be written as




h



1
exp

1
kB T
Vz
V0
kB T
exp



h Yf
kB T
hz
1


exp


2
kB T
Yf

kTST

20

The jump rate constants reported in our previous work [88] were calculated
using Eq. (20). Within the harmonic approximation, the difference in

*Despite this zero limit, the ratio of the products of these frequencies, calculated via
00z
00
(det a
1 = det a
1 0 ), is well-dened with a value near unity. See [88,123] for details.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

439

entropy between the reactant and transition states can be approximated

by [88,124]
0

1


1 00 z

S
B h  1   det a
C
ln@
Yf

 A
1
z det a
1 00 min
kB
kB T

2 
Yf

21

(classical mechanics-based) or



1
0Y
f
hz
1


exp


B 2
S
k T C


B 
C

1
lnB
@
Y
f
h A
kB
1
exp

1
kB T

22

(quantum mechanics-based). The physical interpretation is that a

higher frequency corresponds to a more narrow potential well, which
allows fewer conformations. An advantage of the harmonic approximation is its speed: only frequencies are required after the transition state
and IRC are found. One disadvantage is the harmonic approximation for
the energy, which breaks down within the origin state or dividing surface.
A second disadvantage is that it neglects any contributions along the IRC
itself.
A more exact approach is to use sampling techniques to calculate
the free energy difference between reactant and transition states, with
the rate constant calculated by Eq. (1.117). One possibility is to calculate
free energy differences between hyperplanes orthogonal to the IRC [125,126]
(see also Chapters 9 and 10 in this book). Another is to use chain-of-states
methods to sample fluctuations among different states along the IRC
simultaneously [111,118,127,128]. A difficulty with the latter for explicit
inclusion of polymer chain fluctuations is the high dimensionality of each
state point.
Finally, an improvement to the transition-state theory-based rate
constant is to use the formalism in Chapter 1 [Eqs. (1.103) and (1.104)]
to calculate the dynamical correction factor  [83,84]. Several simulations
beginning from the dividing surface and directed towards the final state are
conducted, and  equals the fraction of jumps that thermalize in the final
state. Each such simulation is short (1020 ps) because it follows the system
to a point of lower energy and higher entropy. A separation of time scales
is required so subsequent jumps out of the product state are unlikely at
short times.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

440

Greenfield

III. STARTING POINT: POLYMER MOLECULAR

STRUCTURES
For using any of the diffusion simulation methods described in this chapter,
it is necessary to begin with a set of chain configurations that represent the
polymer of interest on the molecular level. For a bulk polymer these are
likely to be at the bulk density, within a simulation box of edge length
2050 A and with periodic boundary conditions in two or three dimensions.
Larger edge lengths are more desirable in that they allow larger natural
length scales to be probed within the simulation. They are less desirable in
that they require larger simulation times, since calculation times scale with
Nm  L3m and m is typically between 1 and 2.
Constructing and equilibrating molecular-level polymer configurations
are described in detail in the literature [129135]. Such techniques are
important because much longer times than can be achieved in practice are
required in order to equilibrate the entire chain, starting from a random
configuration in a room-temperature simulation. The objective of those
methods is to ensure that representative polymer chain configurations are
used for the sorption and diffusion calculations.
For the purpose of studying diffusion, it is necessary to have a penetrant
molecule amidst the polymer chains in the molecular simulation cell. It is
possible to add the penetrant during the diffusion calculations. However, it
is also advantageous to equilibrate the polymer sample to some extent in the
presence of the penetrant molecule. This will ensure that fast ( <0.1 ns),
penetrant-induced polymer relaxations occur before beginning the diffusion
calculations. Sometimes, and especially in molecular dynamics simulations,
several penetrant molecules are added in order to improve sampling.
(Of order 10 sets of penetrant mean-squared displacements are sometimes
obtained from one simulation.) One risk of this approach is that the locally
high penetrant concentration could plasticize the polymer, artificially
increasing the penetrant diffusivity. Plasticization is desirable if this is the
concentration range of interest, but it should be avoided if the objective
is calculating diffusivity in the infinite dilution limit.
For accumulating good statistics, several different sets of polymer chain
configurations should be used. Each resulting simulation cell will likely have
a different resistance to diffusion. This is physically reasonable, because
molecular-level imaging (AFM, STM, etc.) reveals heterogeneity over nm to
m length scales [136]. The question then arises of how the results from
different calculations can be averaged. For polymer samples at overall
thermal equilibrium, the Boltzmann factor for the total energy is
an appropriate weighting factor, and lower energy structures will occur a
higher proportion of the time. For glasses, physics-based studies have shown

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

441

that an arithmetic average over different structures is most appropriate

[137]. We have applied the latter when calculating the distribution of jump
rate constants for methane in glassy atactic polypropylene [88]. In a real
glassy polymer, the overall density (and thus the volume available to a
penetrant) depends on the sample preparation, such as the quench rate
from above Tg, the strain history, and the amount of physical aging [1].
Consequently the real distribution of jump rate constants could be expected
to depend on all of these variables. Detailed accounting for these effects is
beyond the current scope of molecular simulation.

IV. FROZEN POLYMER METHOD

The most straightforward approach for simulating diffusion through a
glassy polymer is the frozen polymer method. The essential idea of the
frozen polymer method is that polymer chains remain fixed in place and
provide a static external field, through which a small-molecule penetrant
can diffuse.
This method is the least accurate for simulating small-molecule diffusion
in a polymer, because neglecting contributions from polymer chain
fluctuations is unrealistic. These limitations were learned from its initial
applications: argon in polyethylene [92]; helium, oxygen, and nitrogen
in polypropylene [93]; helium, oxygen, nitrogen, hydrogen, and argon in
polycarbonate [93,94]; helium, oxygen, nitrogen, hydrogen, and argon
in polyisobutylene [94]. Materials such as zeolites are more rigid than
polymers, making neglecting fluctuations in the surrounding material more
physically realistic. Consequently an analogous approach has been used
successfully for alkanes [85] or aromatics [138] in zeolites; see [95,96] for
reviews. Despite these limitations, the frozen polymer method provides a
useful framework for introducing steps used in more sophisticated TST
approaches. A frozen polymer has also aided with interpreting different
contributors to the diffusion mechanism, within recent molecular dynamics
simulations of helium in polypropylene [44].
Removing the polymer chains from explicit consideration directly affects
the list of flexible coordinates in the TST formulation. For a monatomic
penetrant atom or a pseudo-monatomic united atom, such as methane, the
only flexible coordinates are the mass-weighted position (x, y, z) of the
penetrant. The fixed polymer coordinates can be considered part of the force
field defining the penetrant potential energy. The covariant metric tensor a0
equals a three-by-three identity matrix.
For a linear penetrant molecule such as oxygen or nitrogen, its
position can be represented using the mass-weighted position of the first

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

442

Greenfield

atom (x1, y1, z1) and the orientation of the bond vector in spherical
coordinates,
x2 x1 cos  sin

23

y2 y1 sin  sin

24

z2 z1 cos

25

 and
are the projection of the bond on the x-y plane (0   < 2) and its
azimuthal angle relative to the z-axis (0 
 ). The mass-weighted bond
length is
m1=2 X2
X1 2 Y2
Y1 2 Z2
Z1 2 1=2

26

The Jacobian J 0 is found [using Eq. (9)] by partial differentiation of the

mass-weighted Cartesian coordinates (x1, y1, z1, x2, y2, z2) (one per row) with
respect to the generalized coordinates (x1, y1, z1, ,
, ) (one per column),
leading to
1
0
1 0 0
0
0
0
C
B0 1 0
0
0
0
C
B
C
B
0 0 1
0
0
0
C
27
J0 B
B 1 0 0
sin  sin
cos  cos
cos  sin
C
C
B
@0 1 0
cos  sin
sin  cos
sin  sin
A
0 0 1
0

sin

cos

The covariant metric tensor is found by matrix multiplication [Eq. (11)],
1
0
2
0
0

sinsin
coscos
cossin

B
0
2
0
cossin
sincos
sinsin
C
C
B
B
0
0
2
0

sin

cos
C
C
B
aB
C
0
2 sin2

0
0
C
B
sin sin
cossin

A
@ coscos
sin cos

sin

0
0
2
cossin
sin sin
cos

0
0
1
28
The bond length could also be considered infinitely stiff, in which case J 0f
would lack the last column of J 0 , and a0 (a 5  5 matrix) would lack the
last row and column of a.
For a nonlinear rigid molecule such as aromatic compounds (benzene,
toluene, xylenes), coordinates corresponding to the center of mass and the
molecular orientation can be used [138]. The required Eulerian angles , ,

denote rotations about the z axis, the x axis in this first intermediate
coordinate system, and the z axis in this second intermediate coordinate

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

system [139]. The

0
1 0 0
B
B0 1 0
B
B0 0 1
B
B
B
aB
B0 0 0
B
B
B
B
B0 0 0
@

443

resulting covariant metric tensor is*

0

0
2

0
I1 sin2
sin2

6
4 I2 sin2
cos2

C
C
C
C
0
0
C
C
C
C
I3 cos
I1
I2 sin
sin cos C
C
C
C
C
C
I3
0
A
0

3
7
5

I3 cos2

I3 cos

0 0 0 I1
I2 sin
sin cos

I1 cos2

I2 sin2

29
This metric tensor can then be used in Eqs. (14) and (15). Its block-diagonal
character separates translation and rotation steps, an advantage for deciphering diffusion mechanisms. One complication, however, is that the
Eulerian angles must be carefully monitored and redefined in case any
pass through zero [138].
In principle, the transition-state theory equations developed above [Eqs.
(13), (14), (15), (16), (1.117), (20)] can be used for each jump in the frozen
polymer method. While the individual components of the saddle-point search
have a reputation of being slow (calculating the Hessian matrix, evaluating all
eigenvalues and eigenvectors), in practice these steps are fast because the
dimensionality is small. Standard matrix routines [120122,140] are recommended for the matrix calculations. Typically following the IRC will require
significantly more computational effort than finding the transition state.
A good estimate for the transition state location will speed up the search
process and help to ensure a successful search. Geometric interpretations
of the simulation cell are often useful for this purpose. The voxel method
(useful throughout the frozen polymer method) is discussed below. Another
approach is to subdivide (or tessellate) the periodic simulation cell into
tetrahedra [87,93,141145] or cubes [146,147], with a hard-sphere representation of the atoms replacing an energy calculation. Sorption states can
be assigned to the center of tetrahedra or cubes not fully occupied by atoms
of the polymer chain, with transition states assigned to the center of surface
triangles [93]. Alternatively, tetrahedra or cubes can be lumped into sorption
states, and initial guesses for transition state locations can be taken at the
center of tetrahedra/cubes that form the intersection of sorption states
*Equation 12 in [138] [the equivalent of Eq. (29) substituted into Eq. (15)] contains a typographical error. This form of a0 is correct.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

444

Greenfield

[88,102]. The important point is using such criteria to recognize the threedimensional regions in a polymer structure that most closely resemble the
necks between (presumably) larger sorption state regions.
An alternative approach typically used within the frozen polymer method
and the average fluctuating polymer method is the so-called voxel
approach [85,94]. A finely spaced cubic grid (lattice spacing of 0.1 A, for
example) is overlaid onto the simulation cell, and the energy required to insert
a penetrant at the center of each cube is calculated. For each voxel (i.e., cube)
i, the lowest-energy neighboring cube j is identified, and the connectivity i ! j
is stored. Alternatively [85], a steepest descent calculation can be conducted
from each voxel. After looping over all voxels, the net result is a connectivity
map for the particular polymer structure chosen. Voxels with no neighbors of
lower energy correspond to local minima. Each voxel can then be labeled as
leading (or belonging) to the state defined by the local minimum in which
a connectivity path terminates. (If i ! j ! k ! , then voxels i, j, k, and
belong to state .) The dividing surface between states is defined by the
common faces of voxels in different states, and the center of the face between
the lowest-energy voxels on the dividing surface approximates the position of
the transition state. This position is not an exact solution to Eq. (13), but it is
close to one. Similarly, the connectivity path defined by adjoining, lowest-inenergy voxels approximates the IRC defined by Eqs. (14) and (15).
For a spherical penetrant, the energy at each voxel can be used to
calculate each rate constant through numerical integration of Eq. (16).
The kinetic energy contributions to each partition function are integrated
analytically in mass-weighted Cartesian coordinates,


2kB T
z
Q
ZA!B
h2



2kB T 3=2
Q
ZA
h2
z

30

31

The remaining configurational integral ZA for each initial state is a sum of

the contributions from each voxel, multiplied by the mass-weighted voxel
volume
ZA

Zj rVj m3=2

32

voxel j in
state A


Vj
exp

Vj m3=2
kB T
voxel j in
X

state A

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

33

Sorption and Diffusion by TST

445

The energy V j corresponds to a ghost penetrant molecule placed at the

center of voxel j, i.e., the voxel energy calculated previously. The configurational integral for each dividing surface is similarly evaluated, using contributions from the voxels that define the interstate boundary,



X z
X
Vj
z
Aj m
Zj Aj m
exp

34
ZA!B
kB T
voxel j on
voxel j on
A!B
boundary

A!B
boundary

The voxel energy V j is assumed constant within the volume Vj and across
the surface area Aj.
The dividing surface is a curved region, which voxels approximate as a
zig-zag of cubic surfaces. To calculate each Aj (the area of the dividing
plane that cuts through voxel j ), Gusev and Suter [97] proposed
Aj KL2v

35

where Lv is the voxel edge length and K (see Table 1) depends on the number
of neighboring voxels assigned to state B (the destination state). The fewer
neighboring voxels in the destination state, the better this approximation.
Gusev et al. [94] estimated area errors A of under 7% based on a voxel edge
of Lv 0.2 A. From Eqs. (16), (30), and (34), these errors propagate directly
into the rate constant. Smaller voxels would lead to smaller area errors,
according to
jAj  L2v jKj

36

where K is the difference between the value listed in Table 1 and the true
ratio of dividing surface size to voxel face area. Consequently decreasing
TABLE 1 Coefficients Used when Estimating the
Contribution of Each Voxel to the Dividing Surface
Area (Equation 35 from [97]). A Voxel can Have a
Maximum of 6 Nearest Neighbors
Number of neighboring
voxels in state B
1
2
3
4
5
6

K
1p
2
1.41
2.0
0
0

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

446

Greenfield

voxel size can significantly reduce dividing surface area errors. The final
voxel-based expression for the rate constant is thus



Vj
Kj L2v
exp

k
T
B
voxel j on
boundary



X
Vj
exp

L3
kB T v
voxel j
X

TST


kB T 1=2

2m

37

in state

for all voxels having the same edge length Lv.

The relationship between individual sorption states and a fully connected
jump network will be discussed in Section IX. Methods for calculating
a diffusion coefficient from the structure and rate constants are discussed
in Section X.

V. AVERAGE FLUCTUATING POLYMER METHOD

Compared to the rigid polymer method, the average fluctuating polymer
method improves the treatment of how polymer chains move during the
penetrant diffusion process. Rather than remaining fixed in place, polymer
chains execute harmonic vibrations about their equilibrium positions.
Penetrant jumps are then coupled to elastic fluctuations of the polymer
matrix and are independent of structural relaxation of the polymer chains
[24,97]. After a penetrant jump completes, chains near the final sorption
state will likely show slight elastic deviations as they swell to accommodate
the penetrant molecule. Since no chain conformation relaxations are
allowed, other polymer chains will essentially retain their initial conformation. The penetrant jump rate then depends only on the local, quasiharmonic fluctuations in the sorption state and the transition state [24,97].
1=2
The magnitude of elastic vibrations is controlled by a parameter h2 i ,
of units A. While the original formulation of this method allowed for
different parameter values for each atom type (or even every atom), traditionally a single size is used for all atoms in the polymer matrix. The fluctuations impose an additional equilibrium probability density distribution
Px
hxidx
hxi



Y

3=2
xi
hxi i2 yi
hyi i2 zi
hzi i2

2h2 i
exp

dxi
hxi i
2h2 i
i
38

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

447

where hxi is the vector of all polymer atom positions in the penetrant-free
equilibrated (i.e., preaveraged) polymer sample. The three-dimensional position of a single atom is written xi. The mean-squared fluctuation along each
average position (x, y, or z) equals h2 i; hence the parameter notation as
an RMS value. The unnormalized probability density for all penetrant
positions rp and polymer atom positions x is then written as
Prp , xdx
hxi






Y

V ip rp , xi
jxi
hxi ij2
2
3=2

2h i
exp

exp

dxi
hxi i
kB T
2h2 i
i
39
1=2

The parameter h2 i can be interpreted [24] as being proportional to T1/2,

its overall effect being an additive elastic contribution to the total energy.
The modified position distribution function is incorporated directly into
the rate constant calculation based on the voxel method, with the penetrant
positioned at the voxel center. Each single-voxel configuration integral
[Zj in Eq. (32)], referenced to the energy of the local minimum origin state,
contains an integral over polymer conformation fluctuations sampled via
1=2
the h2 i parameter,
Z
exp

Zj rp



Vrp , x
Px
hxidx
hxi
kB T

40

The Boltzmann factor depends on the positions of all atoms that interact
with the penetrant. For a pairwise additive penetrantpolymer force field,
such as the Lennard-Jones expression

Vrp

"

4ip

ip
jrp
xi j

12

6 #
ip


jrp
xi j

41

the configurational integral can be written as a product of separable terms,

Zj rp

Y
i

Y


3=2

V ip rp , xi
Zji rp
2h i
exp

kB T
i



jxi
hxi ij2
dxi
hxi i
 exp

2h2 i
2

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

42

448

Greenfield

While Eq. (42) cannot be solved analytically, it is amenable to an accurate

numerical solution, discussed in Appendix D. Applying the average fluctuating polymer method thus involves
1.
2.
3.
4.
5.

evaluating the single voxel contributions for each voxel in each

sorption state A [see Eq. (105)],
z
multiplying all contributions to evaluate Zj for each state and ZA!B
for each dividing surface [Eq. (42)],
calculating the partition functions for the state and dividing surface
[Eqs. (32) and (34)],
calculating the rate constant for each A ! B transition by Eq. (16),
using the resulting jump network to calculate the diffusivity.

The last point is discussed below in Section X.

The effect of different h2 i values on the normalized probability density
of different separations between a fixed penetrant atom and a single
fluctuating atom of the polymer is shown in Fig. 1. Only one dimension
is shown to simplify the plot, and parameters are chosen such that

FIG. 1 Probability density of different interatomic separations occurring for a

combination of the Lennard-Jones potential and the pseudo-elastic energy provided
by the DebyeWaller factor. Normalization is such that the integral over a spherical
1=2
region from 0 to 10 A equals unity. Different lines indicate different h2 i
values (in Angstroms).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

449

/kBT 0.437,  3.69 A (methanemethyl interaction in atactic polypro1=2

pylene at 233K [88]). The larger the value of h2 i , the softer the
effective potential, since smaller separations can be achieved with nonzero
probability. In other words, larger polymer chain fluctuations facilitate a
penetrant squeezing through a tight channel.
1=2
The choice of h2 i is the most difficult part of applying the fluctuating
1=2
polymer method. As seen in Fig. 1, changes in h2 i of order 0.2 A cause
qualitative differences in the distribution of atom positions. Too small a
1=2
leads to energy barriers that are too high and diffusion
value of h2 i
1=2
coefficients that are too low; the limit h2 i ! 0 corresponds to the frozen
1=2
corresponds to channel
polymer method. Too high a value of h2 i
openings being too facile, with diffusion coefficients that are artificially
1=2
large. h2 i values are also penetrant-dependent [97].
1=2
One approach that would come to mind for choosing h2 i would be to
1=2
run a molecular dynamics simulation on the pure polymer, setting h2 i
equal to an asymptotic value. However, this approach fails in an amorphous
polymer. Over long times, mean-square atom displacements increase
linearly with time due to diffusive chain motions, leading to
h2 i

1=2

p
6Dt

43

1=2

The value of h2 i at 1ns will be of order 10 times larger than the value at
1=2
10 ps. Even in the reptation regime [113,114], in which hr2i  t1/2, h2 i
would appear to increase by about 3 over this time scale, which based on
1=2
Fig. 1 would have a big effect on diffusion. Clearly the value of h2 i
depends on the time scale of interest.
1=2
Several methods have been proposed for choosing the parameter h2 i .
Its value is crucial, because the magnitude and characteristics of meansquared displacement (normal vs. anomalous diffusion) are sensitive
1=2
functions of h2 i . Regrettably, several practitioners of this method
2 1=2
values from their publications. Arizzi [93] noted that an
omit h i
analytic form is available from the X-ray scattering literature [148],
h2 i

Z
D =T

3h2 T
T
d

D

e
1 4T
42 mkB
D 2
D 0

44

where
D equals the Debye temperature. This approach yields a different

1=2
value for each atom (due to their differing masses) and h2 i
values
slightly larger than used in later simulation work. Gusev and Suter [97]
1=2
initially proposed that h2 i
should be chosen to correspond to the
elapsed time at which a penetrant is most likely to jump into another state.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

450

Greenfield

They reasoned that since the jump significantly changes the list of neighboring polymer chain atoms (and especially the distances from the atoms that
remain neighbors), it is appropriate for the elastic effects to be considered
afresh. This process [97] then entails self-consistency:
1.
2.
3.
4.
5.

1=2

guess h2 i ,
calculate the escape time distribution,
1=2
rechoose h2 i based on the most likely escape time,
recalculate the escape time distribution,
iterate to convergence.

This method implicitly requires the peak and breadth of the escape time
1=2
distribution to be equally sensitive to h2 i , since the average and breadth
of a waiting time distribution are governed by only one parameter [103].
However, the sensitivities are not necessarily the same, and Gusev et al. [24]
1=2
noted that the h2 i value so-calculated is only approximate. They next
suggested [24] conducting both MD and TST simulations, using the mod1=2
erate time behavior (t<10 ns) to provide a target for choosing h2 i :
1.
2.
3.
4.

obtain hr2(t)i from MD,

1=2
guess h2 i ,
calculate hr2(t)i from TST and kinetic Monte Carlo (a technique
described below),
1=2
modify h2 i , recalculate, and iterate to obtain better agreement.

Such a combination of MD and TST has promise for systems in which the
assumption of uncoupled elastic fluctuations is valid.

VI. EXPLICIT POLYMER METHOD

The explicit polymer method provides more molecular-level detail about
each penetrant jump, at the expense of a higher computational cost
compared to the frozen polymer and average fluctuating polymer methods.
The polymer chains participate explicitly in defining the path of a penetrant
jump. In contrast to the average fluctuating polymer method, however, the
particular conformation fluctuations that open the interstate channel are
tabulated as part of the jump path. In other words, the steps qf along the
IRC include penetrant degrees of freedom and some of the many polymer
degrees of freedom. The IRC thus traces out a multidimensional pathway
that starts in a local potential energy minimum (described as a penetrant
residing in a sorption state), passes through a multidimensional transition
state as polymer chain fluctuations open a channel and the penetrant moves
through it, and terminates in another local potential energy minimum.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

451

The general approach for applying the explicit polymer method involves
1.
2.
3.
4.

identifying multidimensional transition states,

following the IRC from each transition state to its affiliated local
minima,
calculating the rate constant for each jump,
using a kinetic Monte Carlo approach to calculate the diffusion
coefficient.

A molecular scale network spanning the original simulation cell, with its
jump paths and rate constants, can then be used when calculating the diffusivity. Alternatively, a larger network can be constructed [91] (see Section
IX) that incorporates molecular-level correlations.
The task of searching for a penetrant jump transition state was
discussed in Section II. Each multidimensional transition state is a solution
to Eq. (13), which is solved most efficiently in stages [88]. First a lowdimensional transition state is found for the penetrant by keeping the
polymer frozen (as above). Next, polymer degrees of freedom can be
introduced into the transition state search, using the transition state found
in the lower dimensionality as an initial guess. The appropriate degrees of
freedom to include are those defined by chain segments near the penetrant.
All of the selected degrees of freedom (including the penetrant) are allowed
to evolve during the transition state search in the higher dimension. The
resulting transition state has two elements of unstable equilibrium. Forces
on the penetrant are balanced such that it senses neither a force towards
the origin state nor one towards the destination state. Simultaneously,
neighboring polymer chains have withdrawn in response to penetrant
polymer forces until they are balanced by the force required to compress
the polymer matrix locally. The result is a more open channel for
penetrant diffusion.
To some extent, as the number of coordinates involved in calculating
the transition state increases, the transition state energy decreases. As an
example, Fig. 2 shows the decrease in transition state energy for a methane
penetrant in atactic polypropylene at three different transition states, as a
function of transition state dimensionality. The greater the number of
degrees of freedom, the larger the chain region able to take up this
compressive force and the smaller the transition state energy. However,
the decrease in energy differs from transition state to transition state; no
simple, quantitative correlation for the magnitude of the energy decrease
exists. The overall energy decrease relative to that in the initial state
(about an order of magnitude smaller than for the frozen polymer case)
corresponds to a rate increase by a factor of about 106. A penetrant
jump assisted by polymer motions is thus much more likely than an

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

452

Greenfield

FIG. 2 Penetrant diffusion transition state energy relative to the local minimum
energy of pure polymer, as a function of the number of coordinates considered
flexible during the transition state search. Different symbols correspond to different
transition states. The transition state energy at three flexible coordinates would be
used in the frozen polymer method.

unassisted jump, as found for the fluctuating vs. frozen polymer methods.
As seen in the figure, the transition state energy ultimately reaches a
plateau value, after which it is unaffected by including additional degrees
of freedom.
In order to reach such a plateau, the initial polymer structure (no
penetrant present) must be in a local potential energy minimum. Otherwise,
each additional polymer coordinate included will contribute a new
opportunity for chain relaxation, leading to a transition state lower in
energy than the original structure but higher than a local potential energy
minimum.
For conducting numerous penetrant diffusion simulations, it isnt
reasonable to inspect Fig. 2 by eye to choose when to stop adding
degrees of freedom; an automated procedure is required instead. One
approach [88] is to continue until all degrees of freedom defined by atoms
within the penetrant neighbor list have been added. The remaining atoms
are affected only indirectly: the penetrant interacts with atom i, which then

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

453

interacts with atom j. This approach led to about 350 degrees of freedom
(penetrant position, bond angles, torsion angles, sometimes chain start
positions and orientations) for methane in glassy atactic polypropylene.
A second approach is to identify all atoms that the penetrant is likely to
encounter during its jump between two sorption states, and then to include
all of them in the transition state search. Such a list could be created by
following the IRC in only the penetrant dimensionality, in the spirit of the
frozen polymer method. An advantage of this approach would be that all
degrees of freedom that would participate in the IRC would be identified
in advance and would move towards a zero-force conformation at the
transition state. A disadvantage is the large number of degrees of freedom
involved (of order 500 for methane in atactic polypropylene). In either case,
we found [88] it was worthwhile to increase the included number of degrees
of freedom slowly (5, 15, . . . , etc.), creating a sequence of higher dimensional
initial guesses, rather than to attempt the complete search initially. While
this would seem to require more calculation, in practice the sequential initial
guesses were good enough such that in most cases no single search required
a long time.
Next, the multidimensional IRC must be determined for this jump.
The Jacobian and the covariant metric tensor include contributions from
all f flexible coordinates, and the IRC is found by following a single step
away from the transition state [Eq. (14)] and then multiple steps along the
IRC [Eq. (15)]. Qualitative descriptions of polymer contributions to the
diffusion mechanism can be taken directly from how each flexible
coordinate changes along the IRC. For example, Fig. 3 shows different
chain conformations along one diffusion jump. Dashed and solid triangles
indicate original and instantaneous conformations for selected chain
segments. Due to coupled changes in polymer bond and torsion angles,
chains move back from the penetrant path as the jump occurs, creating a
wider channel.
Next the rate constant is calculated for each jump. In prior work [88] we
used a harmonic approximation in the flexible degrees of freedom, with the
quantum mechanics-based partition function [Eq. (20)]. Such a rate constant
calculation is fast in comparison to tracking the IRC. Using a free energy
method (Eq. (1.117) or sampling as in Ref. [111]) would yield more accurate
rate constants at a cost of a larger, more time consuming calculation.
Dynamical corrections [Eqs. (1.103) and (1.104)] would increase accuracy
but add additional computational requirements. If the software and run
time are available, these more advanced methods are recommended.
Finally, the rate constants can be used to calculate the diffusivity.
Methods based on jump network structure and kinetic Monte Carlo
simulation are discussed below.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

454

Greenfield

FIG. 3 Different chain conformations of atactic polypropylene as a methane

penetrant (sphere labeled p) jumps between two sorption states. Text labels indicate
the extent along the reaction coordinate. Differences between the positions of dashed
and solid triangles indicate the motion of chain segments to create a more open
channel for diffusion. Conformations are adapted from Movie 1 in [90].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

VII.

455

OTHER IRC METHODS

Numerical artifacts can result from curvature along the IRC. For example,
too large a step along the gradient can lead to a departure from the true
diffusion path. The gradient at the end of the step will direct the system back
towards the IRC, but due to components orthogonal to the IRC it is likely
that so-called overshoot will occur again. The net result is a sequence of
steps that repeatedly cross the IRC [126]. This leads to two disadvantages:
1.
2.

No individual calculated point is actually on the IRC itself.

Several more steps are required, leading to additional calculations.

A cartoon of this effect is shown in Fig. 4.

One proposed solution to this problem is methods that incorporate both
the local gradient and the Hessian into steps along the IRC [116,149,150].
These allow for direct incorporation of curvature within each step. Each
step requires significantly more calculation, since the Hessian is required.
However, IRC curvature is likely, and since repeated crossings can be
avoided, fewer larger steps may be necessary. Which approach is more
efficient (based on both the gradient and Hessian or on only the gradient)
depends on whether an analytic Hessian calculation is available and if so
then how long it takes.

FIG. 4 Cartoon of a zig-zag approximation of the IRC (bold dotdash line). The
thin solid path that repeatedly crosses the IRC (dotdash path) is likely to result
when no Hessian information is used, such as with Eqs. (14) and (15). Dashed and
dotted lines indicate contours of constant potential energy.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

456

Greenfield

Path-based methods [110,111,118,127,128,151153] differ from the approaches discussed above in that they attempt to find an entire reaction/
diffusion path (the IRC) between local minima, without first performing a
transition state search. Instead, the IRC is represented by a sequence of
order Nst 1050 state points in configuration space. Each state point
includes all f flexible mass-weighted Cartesian or generalized coordinates.
The different methods correspond to different algorithms, typically
[110,118,152] some sort of minimization, for solving for the resulting Nstf
coordinates. On a multidimensional potential energy surface, the solution
corresponds to a discretized approximation to a low-energy pathway
traversing as directly as possible between local minima. The f 2 analog
inspires descriptive names for these approaches: chain of states [118],
nudged elastic band [110,152], throwing ropes over mountain passes in
the dark [111]. In model problems, the local minima have not even had to
neighbor one another; the path obtained could pass through multiple
transition states and intermediate local minima. However, the high
dimensionality has prevented path-based methods from being applied to
diffusion in polymers with a flexible matrix. For example, 20 state points
along an IRC in [88] (about 350 flexible coordinates) would require a
minimization with respect to 7000 variables. Evaluating the Hessian matrix
is required in some methods.
It is also possible that via thermal motions a penetrantpolymer system
would sample different sets of pathways, each with its own transition state
structure, for a penetrant jump between two associated local minima.
The methods in [111] are particularly suited to calculating the rate in
that case.

VIII. SORPTION
Penetrant sorption has three impacts on small-molecule permeability.
A direct impact is the proportionality between permeability P and solubility
in solution-diffusion transport [1,154]
P DS

45

The sorption coefficient S equals the ratio between penetrant concentration

in the polymer phase and imposed penetrant fugacity (or pressure) [154]
S  C=f

46

The higher the sorption coefficient, the larger the number of penetrant
molecules dissolved in the polymer sample and able to diffuse across it.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

457

Another direct effect is that high penetrant concentrations can increase

polymer chain mobility (plasticization), thereby increasing the diffusion
coefficient [1]. An indirect effect concerns relative sorption among different
sorption states. From microscopic reversibility [22], the ratio of equilibrium
solubilities within different sorption states equals the ratio of jump rate
constants between them,
SA kB!A

SB kA!B

47

Consequently a jump network with a distribution of rate constants must

also have a distribution of solubility coefficients. The more broad the sorption coefficient distribution, the wider the imbalance in the forward and
reverse jump rates along a single path. Conversely, assigning the same sorption coefficient to each sorption state would require all jumps to have the
same rate constant. In addition, most sorption states will have more than
one neighboring sorption state (if each state had only one neighbor in a
random orientation, it would be geometrically unlikely to form a percolating network), and thus most solubility coefficients are governed by multiple
equivalents of Eq. (47): one per neighboring state. These latter effects of the
sorption coefficients are indirect because they are independent of the total
amount of sorption; the rate constants are unaffected if all sorption coefficients are scaled by the same value [see Eq. (47)].
Sorption predictions rely on calculating the chemical potential of the
penetrant molecule in each sorption state. Details are described in Chapters
9 and 10 of this book. In the low-concentration limit, the required solubility
coefficients for use in Eq. (47) and in diffusion calculations can be calculated
by [154,155]
 !






Npen A
V pen
1
1
ex
A
SA lim
exp

exp

=P
p!0 hVpolym i
kB T
kB T
kB T
kB T

48

The ratio hNpeniA/hVpolymi equals the number of penetrant molecules dissolved in a particular sorption state A per total volume of polymer, and P
equals the total pressure. The ensemble average brackets h  i imply averaging over all penetrant positions within a sorption state A while the polymer conformation fluctuates within the canonical ensemble according to
V polym, the total potential energy in the absence of a penetrant molecule.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

458

Greenfield

ex
A equals the excess chemical potential for a penetrant in sorption state A.
While each sorption coefficient SAP
depends inversely on the system volume,
the total sorption coefficient S ASA is an intensive quantity, since the
number of sorption states varies linearly with the total polymer volume.
Equation (48) applies to the low-concentration limit because it neglects
penetrantpenetrant interactions.
Several methods are available for calculating the excess chemical
potential of a penetrant molecule using molecular simulation. Widom
particle insertion [156,157] is the most straightforward. The potential energy
change that would occur if a penetrant were inserted at a randomly chosen
position is calculated, and the excess chemical potential is proportional to
the average of the resulting Boltzmann factor,



V pen
ex

kB T ln exp

49
kB T
This approach has been applied several times to small-molecule sorption in
polymers [46,77,158,159]. However, a difficulty arises for moderate-sized or
large penetrants: there would be insufficient space for a penetrant to be
inserted; hence the increase in potential energy is large and the predicted
sorption is low. Several alternative approaches exist, and here we list just a
few that have been applied to penetrant sorption. In mixed canonical/grand
canonical simulations, the penetrant chemical potential is specified and its
concentration fluctuates, but the amount of polymer remains constant
[155,160162]. In Gibbs ensemble Monte Carlo [163,164], molecule
exchanges between two simulation cells are used to achieve phase equilibrium. Thermodynamic integration [22,165] is based on using a coupling
parameter to relate a more easily simulated system to the system of interest,
and several recent penetrantpolymer implementations show promise
[166168]. Extended ensembles treat these coupling parameters as variables
in the simulation [168171]. For chemically similar small and large molecules, chain scission/regrowth or connectivity reassignment can be used to
estimate the chemical potential [172,173]. Combinations, such as using
an extended ensemble within a Gibbs ensemble framework [45,174], are
also useful. Other approaches [161,171,175] lead to segment-level incremental chemical potentials, which are useful for calculating relative solubility.
Finally, it was recently proposed that a polymer membrane in contact with a
penetrant phase can be simulated directly [176].
Techniques that can improve statistical sampling are often crucial in
these approaches. An excluded-volume map can improve efficiency by
focusing particle insertions into free volume voids [159,177179]. Specially
designed Monte Carlo moves allow polymer chain fluctuations within

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

459

a densely packed system [180182] (see also Chapter 7). Recent reviews
[183,184] and Chapters 9 and 10 provide in-depth descriptions of methods for calculating chemical potentials in polymer systems, and their
applications.
Another approach to calculating penetrant sorption [185] is to calculate
the penetrant energy as in the Widom insertion method, but at points along
a regularly spaced lattice, using the results to calculate the Henrys Law
constant. A related, popular technique is Gusev and Suters theory [186],
which couples particle insertion concepts and the voxel method described
above for finding transition states in three dimensions. Their approach is
particularly convenient when applying the frozen polymer method or the
fluctuating polymer method, in part because the voxel calculations are
performed anyway and in part because it has been implemented in a popular
software package [101]. It differs from the prior approach by applying a
local partition function to each sorption state. The resulting functional form
is similar to the popular dual-mode sorption model [1], with a different
Langmuir parameter bi for each state.
One aspect of a typical sorption calculation may require some
modification in order to use the results in a diffusion simulation. As
suggested by Eq. (47) and illustrated below, the sorption coefficient
distribution is required for formulating the jump network and thereby
obtaining the penetrant diffusion coefficient. Thus the sorption calculation
must be formulated such that individual results are obtained for each
sorption state, rather than a single sorption result for an entire simulation
cell. If the network generation method described below will be used (rather
than the molecular-scale jump network that emerges directly from simulated
molecular structures), it is convenient to accumulate sorption correlations
during molecular-level simulations. For example, mutual distributions of
sorption magnitude and sorption state connectivity (number of neighbors
accessible via low-energy paths from each sorption state) can be later
incorporated into a large-scale network.

IX. NETWORK STRUCTURE

For conducting a kinetic Monte Carlo simulation, the most straightforward
choice for a network of sorption states and rate constants is that of the
original molecular structure. Its key advantage is its one-to-one correspondence with the detailed polymer configuration. However, the small size of a
typical simulation box is a disadvantage. For example, in [97] it was
observed that anomalous diffusion continued until the root-mean-squared
displacement equaled the box size. From this match in length scales, it is not

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

460

Greenfield

clear if the anomalous region truly ends or is artificially terminated by

periodic boundary conditions.
The effect of periodic boundary conditions on the network structure of
an amorphous polymer is shown in Fig. 5. The nine images depict a threedimensional, 23 A edge-length simulation cell of atactic polypropylene,
with period replication shown explicitly in two dimensions. Different colors
indicate different sorption states of accessible volume [87]. Within a single
cell, no positional order is apparent. Within the set of periodically replicated
cells, however, an imposed crystallinity is apparent: there is one disordered
box per unit cell, and ones eye quickly recognizes the periodic array of

FIG. 5 Nine periodic replications of geometric sorption states in a particular

configuration of atactic polypropylene. Penetrant diffusion occurs via jumps among
these states. The states within an individual simulation cell appear disordered, but
the crystalline nature of a periodically replicated set is more visually noticeable.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

461

sorption states. A dynamical realization of this effect is shown in Figure 2 of

[99]: a hydrogen penetrant sequentially visits periodic images of the same
sorption state in a polyimide matrix.
In [91] we proposed an alternative approach for creating a penetrant jump
network. The original, detailed molecular simulation provides averages and
distributions for the number of sorption states per unit volume, the number
of connections (via low-energy diffusion paths) to neighboring sorption
states, the center-to-center distances along each connection, and the angles
between connections that share a common sorption state endpoint. These
molecular-level correlations can then be used to build up a larger network
using reverse Monte Carlo [91,187]. Sorption state positions and connectivities are first chosen at random, and Monte Carlo-type moves are used to
improve agreement between the instantaneous characteristics of the jump
network and the overall distributions obtained from detailed molecular
simulations. Such moves entail translating a sorption state position by
0.5 A, adding a connection between two previously unconnected sorption
states, or removing an existing connection. The objective function
(analogous to the potential energy) quantifies the similarity between the
instantaneous and target distributions in a least-squares sense. A schedule of
decreasing pseudo-temperatures can then be used to ease the network into
the desired configuration. In previous work [91] we generated networks with
edge lengths of 141 A, with each network connection corresponding to a
methane pathway within glassy atactic polypropylene. Larger networks are
possible as well; the limiting requirement is sufficient memory to store the
array of angles between jump paths. The computer time required to construct
a network is much smaller than that required for the molecular simulations;
typically an overnight run is sufficient.
Next it is necessary to assign relative sorption probabilities and jump rate
constants to the jump network. The distributions of solubility coefficients
[calculated by Eq. (48)] and rate constants are required. First, a cumulative
solubility coefficient distribution is calculated: PS,tot(S) equals the probability that a sorption state has a solubility coefficient of S or smaller. Next,
a uniformly distributed random number 0  <1 is selected in turn for each
sorption state. That sorption state is assigned the value SA for which
PS, tot SA 

50

If enough simulation data are available, separate distributions PS,tot(S)

could be tabulated for each amount of connectivity. For example, high
interstate connectivity can be correlated with higher relative sorption
capacity [91]. By design, the resulting network will have a distribution of
sorption coefficients that matches the desired target values.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

462

Greenfield

Rate constants must then be found that simultaneously satisfy their

desired distribution and Eq. (47). All rate constants in one direction (from
lower to higher sorption state number, for example) can start at a fixed
value, say 1 s
1, and reverse rate constants can then be calculated from
Eq. (47). Another reverse Monte Carlo simulation may then be conducted,
with the attempted move being a simultaneous increase or decrease in
the magnitude of a randomly selected pair of forward and reverse rate
constants. Again, repeating many times and using a schedule of decreasing
pseudo-temperatures can lead to reasonable agreement with the desired rate
constant distribution.
The resulting jump networks are then ready for use in a diffusion
calculation. Using different molecular-level polymer configurations to
obtain the target structure, sorption, and rate distributions increases the
likelihood that the resulting large jump network will properly reflect the
characteristics of a polymer sample.

X. KINETIC MC TO DIFFUSION COEFFICIENT

The final step in obtaining a diffusion coefficient is to simulate the dynamics
of a penetrant molecule on the network of sorption states and rate
constants. Analogous to the frozen positions of voids and channels in a
glassy polymer, the relative sorption probabilities and jump rate constants
typically remain constant throughout the diffusion simulation. For uniform
rate constants on an ordered lattice, it is possible to solve the dynamics
analytically. For the disordered network found for voids through a polymer
matrix, a numerical solution is required.
Because a penetrant is likely to explore a sorption state for a long time
before jumping, it is reasonable at low concentrations to assume that
individual penetrant jumps are uncoupled from one another and that the
sequential visiting of states is a Markov process [103]. Since each jump
occurs independently, the probability density pW(t) that a time t elapses
before the next jump occurs (the waiting time) is distributed according to a
continuous-time Poisson process [103],
pW t dt R exp
Rt dt

51

R[]s
1 equals the overall jumping rate out of all states,

R

XX
i

Ri!j t

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

52

Sorption and Diffusion by TST

463

and
Ri!j Ni tki!j

53

equals the flux from state i to state j. The average waiting time until the next
jump will occur anywhere on the network equals
Z

1
W

Rt exp
Rt dt



1
e
Rt 
1
1

Rt

te

0
0
0


R
R
R
0

tp t dt
0

54

The amount of time that elapses between penetrant jumps thus depends
explicitly on the instantaneous sorption state population and on the distribution of jump rate constants out of occupied states.
The sorption state populations evolve according to a master equation.
For a system in which each sorption state can be represented by a single
local minimum,
X
X
dpi


ki!j pi
kj!i pj
dt
j
j

55

Networks generated using methods of Section IX fall in this category.

For some systems, there is a separation of time scales and jumps can be
separated into two categories: fast jumps among different local minima
in the same sorption state and slow jumps between connected sorption
states. For this case, the master equation can be written
NB X
NA
X X
dpA
pi


ki!j
dt
p
A
B
j1 i1

!
pA

NB X
NA
X X
B

j1 i1

pj
kj!i
pB

!
pB

56

A and B represent different overall sorption states and i and j are individual
regions of states A and B, respectively. The ratio pi/pA equals the fraction of
sorption in state A that occurs within region i. After some time, the sorption
eq
state probabilities stabilize at their equilibrium values peq
i or pA that satisfy
microscopic reversibility.
The master equations can be solved numerically using kinetic Monte
Carlo simulation [85,86,155]. To begin, a number of independent,
noninteracting ghost penetrants are placed in the sorption states
P of
each network, according to the equilibrium distribution pAeq SA = SA .

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

464

Greenfield

These initial positions are stored in a vector r0. The distribution can be
sampled using the same method as described above when assigning the
sorption coefficients to each sorption state:
.

Create a cumulative distribution,

Pn

n
X

Si =

i1

Ns
X

Si

57

i1

Choose a uniform random number 0  <1 for each ghost particle

and assign it to the state n for which Pn
1  <Pn.

The number of penetrants Np should be much larger than the number of

sorption states Ns in order to sample the probabilities pAeq correctly.
The number of penetrants in a state i at a time t is denoted below by Ni(t).
To begin a kinetic Monte Carlo simulation step, the flux along each jump
and their ratio, the probabilities
qi!j t

Ri!j t
Rt

58

are calculated. Next, the expected waiting time until the next event, , is
found as follows. The instantaneous waiting time distribution [Eq. (51)]
leads to a cumulative waiting time distribution by
Zt
W
P t
pW t dt 1
exp
Rt
59
0

Choosing a probability 0  <1 then implies a particular time at which

this cumulative probability  PW( ) is reached, and solving using Eq. (59)
leads to

1
ln1

R

60

The jump that occurred in this event is chosen from the probabilities qi!j t
[from Eqs. (53) and (58) a larger rate constant implies a larger probability],
and the particular ghost penetrant leaving state i (if more than one was
present) is chosen uniformly. Selecting from the qi!j t distribution is performed in the same manner as above, using the cumulative probability
density distribution. The state label for the jumping penetrant is updated,
the occupancies Ni(t) and Nj(t) are reset to account for the jump, and the
process is then repeated many times. Each jump changes the average meansquared displacement (it can increase or decrease) and increases the elapsed
time by . In practice, only fluxes out of the sorption states i and j that

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

465

involve the most recent jump have to be recalculated according to Eq. (53);
this decreases the number of computations significantly for a large system.
The kinetic Monte Carlo simulation should begin by conducting many
steps in order to equilibrate the penetrant sorption state positions, after which
r0 is reset to zero. A production run can then begin. In the production run, the
net result of many steps (103104 per particle) is a list of penetrant sorption
state occupancies as a function of time. The mean-squared displacement can
be calculated (during or after the simulation) at each time as
N 
D
X
 E

rt
r0 2 1
xti
x0 2 yti
y0 2 zti
z0 2
N i1

61

The time-dependent mean-squared displacement will fluctuate from step to

step and will increase over the long term.
Two characteristics of the relationship between the mean-squared
displacement and time are usually of interest. Over long times, meansquared displacement will increase linearly with time, and the diffusion
coefficient D can be extracted from the results as
D

1
@hr2 i
1
lim

m
2d t!1 @t
2d

62

where m is the long-time slope on a linear plot of hr2i vs. t, and d is the
dimensionality.* Over shorter times, diffusion may be anomalous, meaning
hr2 i / t n

63

with n<1. (Note that this is opposite to so-called case II diffusion [188],
in which penetrant mean-squared displacement increases with time squared.
Here diffusion is slower than in the Fickian case.) Features of interest are
the value of the exponent n and the times and distances over which diffusion
is anomalous. Anomalous diffusion can occur due to a distribution of rate
constants [189191] or network connectivity [24,91,192]; it can also depend
on the penetrant size [42]. The essential feature is a limitation of the diffusion paths such that they cannot span the full dimensionality over the simulation time scales available.
An example of the change in mean-squared displacement with time is
shown in Fig. 6 for methane in atactic polypropylene [91]. These
calculations were performed using a network generated via the methods
described in Section IX. At times of less than 1 ns, little information is
*Even for a network whose connectivity can be described using a fractal dimension, d is the
dimensionality of the space in which the network is represented; i.e., d 2 for surface diusion
and d 3 for bulk diusion.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

466

Greenfield

FIG. 6 Mean-squared displacement for methane in glassy atactic polypropylene,

from calculations in [91]. The dotdash line indicates the box size squared. Dotted
lines indicate diffusion exponents of n 1/2 and n 1. A turnover from anomalous
to Fickian diffusion is predicted over times of 15 s.

available. These time scales accessible to MD have been coarse grained

away in calculating the rate constants. For nanosecond-scale dynamics, MD
is preferable to TST. For times up to 1 s or so, an anomalous regime is
predicted in which mean-square displacement increases with time to the 1/2
power (n  1/2). This value is consistent with Monte Carlo simulations of
anomalous diffusion [193]. After about 5 s, Fickian diffusion is predicted,
with D  10
8 cm2 s
1 at T233K. The mean-squared displacement reaches
the box size squared (shown by the dotdash line) after about 30 s, well
after the anomalous regime has ended. While data for this particular system
are not available, the prediction is consistent with extrapolated data for
an ethylene/propylene copolymer [91,194].

XI. SUMMARY AND OUTLOOK FOR OTHER SYSTEMS

Several different methods for simulating small-molecule diffusion in a glassy
polymer have been discussed. The first involved MD simulations at and
above Tg, using an Arrhenius relationship to extrapolate to temperatures

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

467

well below the glass temperature. This approach is straightforward with

typical simulation tools, but it involves a large extrapolation.
Four methods based on transition-state theory were then discussed. The
frozen polymer method is inappropriate for penetrant diffusion in glassy
polymers, because chain fluctuations are intricately coupled to the diffusion
mechanism. Path-based methods are very sound, but the resulting high
dimensionality could pose computational problems.
The average fluctuating polymer method and the explicit polymer method
both account to some degree for polymer chain contributions to each rate
constant. However, there are several differences between the two approaches.
In the average fluctuating polymer method, chains experience harmonic
position fluctuations, with nonlinear resulting energy changes used in the rate
constant calculation. In the explicit polymer method, anharmonic position
fluctuations occur, with harmonic energy fluctuations about the transition
state and local minimum configurations (as the method has been applied
previously). In the average fluctuating polymer method, all elastic position
fluctuations contribute to the rate constant. However, which fluctuation is
coupled to diffusion is not clear. In the explicit polymer method, a complete
set of orthogonal fluctuations (normal modes) about one particular configuration is included, but no other configurations are sampled (in the
harmonic approximation). However, this one configuration sampled
corresponds to the lowest-energy mechanism for the jump itself. In this
sense, the choice of configuration is analogous to the maximum term
approximation in statistical mechanics: while only one set is included, that
set is (at least one of ) the most important. The average fluctuating polymer
method depends on a time scale over which the magnitude of fluctuations
is appropriate, and results vary strongly with a single, hard-to-predict
parameter. The only time scale in applying the explicit polymer method to a
jump is the inverse of the jump rate constant itself, and it is not sensitive to
such a parameter.
The proper choice among these methods depends on the problem of
interest and the software available. While the average fluctuating polymer
method has been applied to single united-atom penetrants, how well it would
work for larger, flexible penetrants is not clear. To date, the explicit polymer
method has been applied to united atom and dimer penetrants. Since the
method already handles hundreds of polymer chain degrees of freedom,
it provides a framework for treating larger penetrants. Indeed, there are no
intrinsic scientific barriers to introducing additional penetrant coordinates,
such as bond or torsion angles. The effects of local, anharmonic changes
in polymer chain configuration should be even more important in those
cases than they are for monatomic and diatomic penetrant molecules,
which will require transition-state theory treatments beyond the harmonic

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

468

Greenfield

approximation. Instead, the barriers are methodological: energy and

coordinate derivatives must be calculated for complicated force fields and
polymer structures, and software that facilitates incorporating such
modifications needs to be developed. Reformulating in other coordinate
systems could also simplify some stages while complicating others. Working
in mass-weighted Cartesian coordinates would simplify the derivations but
increase the matrix computation requirements by about 3.5 times (3 instead
of 2 degrees of freedom per flexible atom, in computations that scale as N3).
Regardless of the method, applications of transition-state theory for
simulating penetrant diffusion are ultimately limited by the physical
diffusion mechanism. For sufficiently large penetrant molecules, the
separation of time scales that accompanies the jump mechanism could
diminish, making it more difficult to distinguish between events that do or
dont contribute to diffusive motion. How large a penetrant corresponds to
this limit is still a research question.

APPENDIX A: IRC DERIVATION IN GENERALIZED

COORDINATES
To utilize generalized coordinates, it is necessary to reformulate Eqs. (6) and
(7). First, the Hessian with respect to the generalized coordinates must be
related to the HessianP
with respect to mass-weighted Cartesian coordinates.
k
i
k
Expanding @=@xi as
k @q =@x @=@q (as done for the gradient at the
transition state) leads to
Hij

3N X
3N
3N
X
X
@qk @ql @2 V
@2 ql @V

i
j
k
l
@x @x @q @q
@xi @xj @ql
l1
k1 l1

or
0
B

1
B
@q @q2
@q3N B
B
Hij
,
,...,
@xi @xi
@xi B
B
@

@2 V
@q1 @q1
..
.

...

@2 V
@q1 @q3N
..
.

@2 V
@2 V
...
3N
1
3N
@q @q
@q @q3N
1
0
@V
B @q1 C
C

2 1
B
C
@ q
@ 2 q2
@2 q3N B
B ... C

,
,
.
.
.
,
C
B
i
j
i
j
i
j
@x @x @x @x
@x @x B
C
@ @V A
@q3N

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

10

@q1
@xj
..
.

CB
C
CB
C
CB
C
CB
C
CB
C
CB
C
A@ @q3N A
@xj

64

Sorption and Diffusion by TST

469

for each term Hij of the Hessian. Different rows and columns of H result
from taking all i 1, . . . , 3N and j 1, . . . ,3N components, respectively.
The resulting matrix equation is

2 T 
@ q
T
65
H J Hqq J
rq V
@xi @xj
Hqq is the Hessian of second derivatives with respect to generalized coordinates and @2 qT =@xi @xj , a third-rank tensor, may be thought of as a 3N  3N
matrix, in which each component is itself a vector; the ijth vector contains
the coordinate transformation second derivatives of q with respect to xi and
x j. The resulting rightmost term is a matrix for which the ijth element is the
dot product of the ijth vector with rqV. This tensor complicates the use
of an analytic Hessian matrix in generalized coordinates, since additional
derivatives are required compared to typical simulation needs.
Next, the proper equation for determining eigenvalues and eigenvectors
must be reformulated. The Hessian H can be replaced using Eq. (65), and
x (renormalized to a small size x rather than unity) is replaced using
xi

3N
X
@xi
j1

@q j

q j

or
x J 0 q

66

J 0 (introduced in the main text) is the Jacobian matrix of partial derivatives

for transforming from generalized coordinates to mass-weighted coordinates. These substitutions result in

2 T 
@ q
rq V J 0 q
J 0 q
67
J T Hqq J J 0 q
@xi @xj
This expression can be simplified after premultiplying by J 0 T and using the
identity J J 0 I , leading to

2 T 
@ q
Hqq q J 0T
rq V J 0 q
J 0T J 0 q
68
@xi @xj
The product of Jacobian matrices can then be replaced with the covariant
metric tensor a [Eq. (11)]. If we restrict attention to the transition state,
where rqV 0, we obtain Eq. (8).
Reformulating the IRC step away from the transition state [Eq. (7)] is
similar. The step x is replaced using Eq. (66) and the gradient is replaced

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

470

Greenfield

using Eq. (2), leading to

J 0 q J T rq V 

69

Premultiplying by J 0 T, replacing the J 0 TJ T product with the identity

matrix, and substituting for the covariant metric tensor using Eq. (11)
leads to Eq. (12).

APPENDIX B: IRC IN A SUBSET OF COORDINATES

The key to reducing the dimensionality of a TST problem is to reformulate
the transition state criterion and the IRC equation such that flexible
coordinates evolve as necessary, while stiff coordinates remain implicitly at
their original values. This suggests formulating incremental position changes
in terms of generalized coordinate changes q and derivatives in terms of
@=@q. The latter is particularly convenient, since a partial derivative @=@qi is
taken with all other q j(j 6 i), both flexible and stiff, held constant.
Narrowing the total step involves dividing it into flexible (qf) and stiff
(qs) components,



qf
q
qs
Changes can be made along the flexible coordinates without affecting the
stiff coordinates. The gradient vector is divided as



rqf V
rq V
rqs V
Following Weiner [119, x 6.31], we use Greek indices to denote flexible
coordinates, uppercase indices to denote stiff coordinates, and lowercase
indices to denote the entire set.
The Jacobian matrix for transforming from mass-weighted Cartesian to
generalized coordinates can be narrowed in a similar manner. A derivative
@xk =@q is taken with respect to flexible generalized coordinate q with all
other generalized coordinates held constant, so no stiff coordinates change
along the direction @x=@q . Thus the Jacobian matrix J 0 may be divided as
0

@x1
@q1
..
.

B
B
B
J B
B
@ @x3N
0

@q1

1
@x1 @x1
@x1


@q f @q f 1
@q3N C
C
.. C  0 0
..
..
..
..
Jf Js
. C
.
.
.
.
C
3N
3N
3N A
@x
@x
@x


f
f
1
@q @q
@q3N

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

70

Sorption and Diffusion by TST

471

a0
a @ -a0T

---

--

where Jf 0 and Js0 are rectangular 3N  f and 3N  (3N
f ) matrices,

respectively.
The covariant metric tensor, a, can be narrowed by subdividing it based
on the flexible and stiff coordinates [119, p. 217]
1
a0
-- A
a00

71

The f  f submatrix a0 in the upper left is the covariant metric tensor in only
the flexible coordinates. Each element
a

X @xk @xk
k

@q @q

is unaffected when q and q remain flexible, since constrained generalized

coordinates (such as qA) were already held constant implicitly when taking
the partial derivatives. The elements of a0 may also be found from the
Jacobian matrix in the reduced dimensionality as
a0 J f0T J f0

72

using the Jacobian subsets introduced above. Each element of the submatrix
a0 and its transpose a0 T involve one flexible coordinate and one stiff coordinate. All derivatives in a00 are with respect to stiff coordinates.
The contravariant metric tensor a
1 in the full space can be subdivided
in a similar manner. However, a subset of a
1 in the reduced space is not
simply a single submatrix, since the partial derivatives that define each term
a
1


X @q @q
k

@xk @xk

are taken with all other mass-weighted positions held constant, not all other
generalized coordinates. Instead, the reduced-space contravariant metric
tensor equals the matrix inverse in the reduced dimensionality,
0

a
1 a0
1
The matrix a
1 0 is sometimes denoted G [105].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

73

472

Greenfield

Several relationships among the submatrices of the covariant and

contravariant metric tensors can also be determined [105]. Since the
matrix product of the covariant and contravariant metric tensors is the
identity matrix [using Eq. (10), aa
1 J 0 TJ 0 J J T J 0 T (I ) J T I ),
direct matrix multiplication
aa

1

a0
a0T

a0
a00



a
1
0T
a
1

a
1
00
a
1

I

If
0

I 3N
f

74

(I f is an f  f identity matrix) leads to the simultaneous equations

0

a0 a
1 a0 a
1
0
1 0

aa

0T

0
1 00

aa

If

75

76
00

Rearranging Eq. (76) and postmultiplying by a
1
1 yields

0

00

1

a0
a0 a
1 a
1

and substituting this into Eq. (75) leads to


0
0
00
1
0T
a0 a
1
a
1 a
1 a
1
If
and thus
 0

0
00
1
0T
0
a
1 a
1
a
1 a
1 a
1

77

which, after converting the nomenclature, is equation 10 in [105]. Thus the

contravariant metric tensor in the flexible degrees of freedom may be
calculated directly from the contravariant metric tensor in the full dimensionality. From the similar expression

a
1
0T
a
1

a
1
00
a
1



a0
a0T

a0
a00

I

If
0

I 3N
f

78

multiplying the second row of a
1 and the first column of a leads to
0T

00

a
1 a0 a
1 a0T 0

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

79

Sorption and Diffusion by TST

473

Rearranging and premultiplying yields

00

0T

1

a0T
a
1 a
1 a0

80

which is used below. Finally, each side of Eq. (74) can be postmultiplied by

a0
a0T

0
I

Expanding the matrix products on the left hand side leads to

0

a
1

0T

00

a
1

a
1 a0 a
1 a0T

a
1 a0 a
1 a0T

00

a0
a0T

0
I


81

Substituting using the transpose of Eq. (75) and Eq. (79) leads to

a
1
00
a
1

If
a
0

a0
a0T

0
I


82

and taking the determinant of each side yields

00

det a det a
1 det a0

83

This interrelationship among the magnitude of different metric tensor determinants was first derived by Fixman [195] and is useful in discussions of the
rigid and flexible models of polymer conformation (see Appendix C).
Next, the saddle point and Hessian calculations can be narrowed to a
subset of all coordinates. At a saddle point in the full dimensionality, the
potential energy gradient equals zero with respect to all generalized
coordinates. The same criterion holds for the flexible coordinates in a
subset of the full dimensionality. For constrained degrees of freedom
A f 1, . . . ,3N, the stiff potential maintaining the constraints is of the
form
V

X1
A

2
kstiff A qA
qA
0

with each kstiffA conceptually taking on values that approach infinity

[105]. Even for a differential displacement qA, this energy contribution

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

474

Greenfield

1
A 2
2 kstiff A q

would still approach infinity, as would the force kstiff A qA .

Hence the only states of finite probability with finite forces correspond to
qA 0: stiff coordinates retain their original values at saddle points in the
reduced dimensionality. The Hessian submatrix associated with the constrained degrees of freedom also takes on a diagonal form, with eigenvalues
given by the stiff spring constants. All these eigenvalues are positive; thus it
is meaningful mathematically to discuss the transition state with respect to
a flexible subset of generalized coordinates. To summarize, a saddle point in
this system of reduced dimensionality is a point at which each element of the
potential energy gradient @V=@q equals zero for all flexible generalized
coordinates q, and each eigenvalue of the Hessian in q-space

H 

@2 V
@q @q

is positive except for the lowest, which must be negative. The form of the
eigenvalue equation to solve will be discussed next.
As in the full dimensionality, the first step of the IRC is directed along
the eigenvector corresponding to the negative eigenvalue. The eigenvalue
equation to solve in the subset of flexible coordinates may be found from the
submatrix partitions of a, H, and q. In order to distinguish between the
flexible and stiff coordinates, it is convenient to premultiply Eq. (8) by a
1,
leading to
0
@

a
1

10

00

A@

a
1

0T

a
1

a
1

H0qq

H0qq

H0T
qq

H00qq

1
A

qf
qs

qf

!
84

qs

Expanding the left hand side leads to separate equations for the flexible and
stiff coordinates,
 0

 0

0
0

1
H0qq a
1 H00qq qs
qf
a
1 H0qq a
1 H0T
qq qf a
 0T

 0T

00

1
0

1 00 00

a
H

a
H
q
a
1 H0qq a
1 H0T
f
qq
qq
qq qs
qs

85
86

As described above, qA
s ! 0 for each stiff coordinate A in the infinite
stiffness limit. Even with the force constants approaching infinity, the
right hand side of Eq. (86) approaches zero, as does the second term on
the left hand side of the same equation. Since the step qf in the flexible

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

475

coordinates is nonzero, for Eq. (86) to be satisfied the metric tensor

submatrices and the Hessian submatrices must be related such that
00

0T

1
H0qq
a
1 H0T
qq
a

87

After solving for H0T

qq in this limit and substituting the result into Eq. (85),
rearrangement leads to
h 0
00
1
0
0T
0
a
1 H0qq
a
1 a
1 a
1 H0qq qf a
1 H0qq df
qf

88

Premultiplying by a0 leads to Eq. (14).

Another alternative is to derive Eq. (14) from Eq. (8) by expanding
directly in terms of submatrices of a. Matrix multiplication leads to two sets
of equations, each of which contains qf and qs,
H0qq qf H0qq qs
a0 qf
a0 qs

89

H0T
qq qf

90

H00qq qs

0T

00

a qf
a qs

Setting qs ! 0 in the first leads directly to Eq. (14) . However, the same
limit for the second leads to the seemingly contradictory
0T
H0T
qq qf
a qf

91

an overspecified system of equations for a typical case in which there are

more stiff coordinates than flexible coordinates. However, this equation is
equivalent to Eq. (14) . The proof first requires substituting Eq. (80) into
00
Eq. (91) and premultiplying by a
1 , which leads to
00

0T

1
a
1 H0T
a0 qf
qq qf

a

92

The left hand side can then be replaced using Eq. (87), leading to
0T

0T

a
1 H0qq qf

a
1 a0 qf

93
0T

Removing the common matrix a
1 from the leftmost side of each side of
this equation yields Eq. (14).
Theodorou (D. N. Theodorou, personal communication, 1993) has
obtained the same results with a derivation that remains in the Cartesian
coordinate system. First, he partitioned the eigenvectors into flexible and

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

476

Greenfield

stiff components, finding that deviations along the stiff eigenvectors must
equal zero. Next, he invoked a harmonic approximation with respect to the
full set of mass-weighted Cartesian coordinates, ultimately obtaining the
same equations as those presented here, within a change of coordinate
system.
The IRC step in a flexible subset of coordinates away from the transition
state also follows from subdividing the step vector and the Jacobian matrix.
Begin with the IRC equation in mass-weighted Cartesian coordinates,
Eq. (7). Substitute for x using Eq. (66), but note that only a subset of J 0 is
required, since qA 0 for each stiff coordinate A. The elements of J 0 that
multiply a nonzero value of qi correspond exactly to the elements of Jf 0
defined in Eq. (70). Similarly, only the gradients with respect to flexible
generalized coordinates are nonzero
T

rqf V Jf 0 rx V

94

since the gradient with respect to stiff degrees of freedom equals zero in
either the rigid or the infinitely stiff model. Premultiplying Eq. (7) by Jf 0T ,
substituting with the flexible equivalent of Eq. (66), and using Eq. (94)
results in
T

Jf 0 Jf 0 qf rqf V 

95

From Eq. (72) above, the product Jf 0T Jf 0 equals the covariant metric tensor
in the reduced coordinate system, yielding Eq. (15).
To evaluate the rate, it is necessary to combine expressions for the
partition functions at the local minima and at the transition state. The latter
partition function is one dimension smaller than that of the former, since
integration is limited to the (3N
1)-dimensional dividing surface that
passes through the transition state. An exact differential equation for this
surface exists [112], but its evaluation is difficult; it amounts to following the
3N
1 intrinsic reaction coordinates defined by taking the first step along
an eigenvector associated with a positive eigenvalue. A common assumption
is to approximate the dividing surface near the transition state with a
hyperplane orthogonal to the eigenvector corresponding to the negative
eigenvalue; the dividing plane is orthogonal to the step taken away from the
transition state. The resulting equation can be written in a flexible subset of
generalized coordinates as
x
xz  x

qf
qzf T a0 qf

0

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

96

Sorption and Diffusion by TST

477

where x

and qf

are the eigenvectors that correspond to the negative
eigenvalue in mass-weighted Cartesian coordinates and in the flexible subset
of generalized coordinates. In terms of the normal modes that correspond
to the flexible generalized coordinates, this surface restricts integration to
modes with real frequencies. From the orthogonality of eigenvectors, all
these modes are normal to the first step along the reaction coordinate.

APPENDIX C:

CHOICE OF POLYMER MODEL

FLEXIBLE, RIGID, OR INFINITELY STIFF

An important question first recognized in the 1960s and 1970s [105,195,196]

concerns whether statistical mechanical models of polymers consider certain
degrees of freedom (traditionally bond lengths and sometimes bond angles)
to be either (1) rigidly constrained to their initial values or (2) theoretically
allowed to fluctuate, but under a harmonic potential with force constants
that approach infinity. The two approaches lead to different relative
probabilities for each polymer chain conformation and hence to different
partition functions, and Go and Scheraga [105] showed that option 2, the
flexible model in the limit of infinite stiffness, more closely resembles the
true quantum-mechanical partition function.
In Monte Carlo simulation, the choice of polymer model is governed
by the choice of attempted moves. Typically kinetic energy is integrated
analytically over all modes, and the partition function for the flexible model
in the limit of infinite stiffness results [105]. In molecular dynamics,
constraints (SHAKE, etc.) freeze kinetic energy contributions and the
partition function for the rigid model results [105,197]. To achieve sampling
from the desired partition function, it is necessary to add a pseudopotential
based on the covariant metric tensor a0 [198].
For modeling penetrant diffusion in polymers using transition-state
theory, it is important to show that the approaches for following the IRC
correspond to the flexible model in the limit of infinite stiffness, rather than
to the rigid model. At first, it would seem that reducing the dimensionality
by constraining particular generalized coordinates must correspond to the
rigid model. In the development above, the constraints on the reaction path
were incorporated directly into the basis vectors of configuration space [via
Eqs. (11), (14), (15)] without introducing a form of the potential. What
would happen if the full dimensionality were retained while force constants
for stiff degrees of freedom increased in magnitude? For an infinitely stiff
chain, any slight perturbations of the bond length would increase the
potential energy towards infinity and the probability of observing such a
state would drop exponentially to zero. However, if the basis vectors in the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

478

Greenfield
0

full dimensionality are followed (i.e., use of a
1 or a rather than a
1 or a0 in
the IRC equations), even infinitesimal steps would induce changes in the
stiff degrees of freedom, since each change depends on all elements of the
gradient,
dqA

X
X @qA @q j  @V
j

@xk @xk

@q j

97

and the coordinate transformation derivatives @qA =@xk are independent of

the potential energy and are conformation-dependent. The only way to
maintain a reasonable energy after a step is to ensure that the stiff degrees
of freedom do not change, and hence to step along the basis vectors found
using only the flexible generalized coordinates. These basis vectors are set by
0
the matrix a0 and its contravariant inverse, a
1 , as suggested by Fixman
[195]. Thus, both a rigid polymer and an infinitely stiff polymer step
along the same basis vectors, and the above development of the IRC
equation does not select one model over the other.
Why is this IRC formulation insensitive to the choice of polymer model?
In Go and Scheragas treatment [105], differences between the two models
originated when integrating the kinetic energy [197]. In the flexible model,
the independent variables were the Cartesian coordinates X and their
conjugate momenta p. The kinetic energy terms are uncoupled (a and a
1
are diagonal) in these coordinates (equation 4 in [105]), and integration is
analytic. The Jacobian determinant det J 0 that appears when converting
from the mass-weighted Cartesian representation x to the generalized
coordinates q is only a function of the bond lengths and bond angles, and
the partition function may be written as [105]
Q





Z
Vqf
2kB T 3N=2
0
det
J
exp


d
q
f
h2
kB T

98

where the Jacobian determinant is intentionally kept inside the integral.

(Only parts of it may be removed if some bond angle force constants do
not approach infinity.) The masses have been absorbed into the generalized
coordinates q.
In the rigid model, the independent variables are the generalized
coordinates q and their conjugate momenta Pr (equation 12 in [105]). The
kinetic energy
T

1 T 0
1
0
q_ a q_ Pr T a
1 Pr
2
2

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

99

Sorption and Diffusion by TST

479

couples all generalized momenta on a given chain, since the contravariant

metric tensor is nondiagonal, and integration of the kinetic energy requires
yet another change of coordinates, now from the conjugate momenta Pr to
the normal modes  of the kinetic energy. The Jacobian determinant of this
transformation is det a
1 0
1=2 , resulting in the partition function for the
rigid model,
"
#1=2






Z
Vqf
2kB T 3N=2
1
d qf
Q
exp

0
h2
kB T
det a
1

100





Z


Vqf
2kB T 3N=2
0 1=2
exp

d qf det a
Q
h2
kB T

101

or

Hence the appearance of the metric tensor inside the integral is due to the
nondiagonal nature of the kinetic energy in the constrained space.
Go and Scheraga also showed that the Jacobian determinant in the
flexible model could be expressed as the Jacobian determinant used in the
rigid model multiplied by a number of classical harmonic oscillators, each
contributing kinetic and potential energy to the Hamiltonian. This
relationship is summarized by Eq. (83). Integration over the kinetic
energy part led to cancellation of (det a0)1/2 in the configurational integral,
while integration over the potential energy led to an arbitrary constant
multiplying Q. The treatments of the two polymer models are identical in
terms of only the potential energy. Within the IRC formalism, dissipation of
kinetic energy is instantaneous, and the kinetic energy need not be treated
until rate constants (and partition functions) are evaluated. In summary, the
method of following the IRC described above only specifies the Riemannian
geometry of the reaction path; it does not choose a physical model of the
classical statistical mechanics. Such a choice is made only when calculating
the rate constants.

APPENDIX D: EVALUATING THE SINGLE VOXEL

PARTITION FUNCTION
To apply the average fluctuating polymer method, it is necessary to integrate
the Boltzmann factor over fluctuations in each polymer atom position
[Eq. (42)]. Each single voxel partition function appears to depend on the
penetrant position (i.e., the voxel location) and the locations of all

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

480

Greenfield

neighboring atoms. Fortunately, Eq. (42) can be reformulated into two

different, simpler forms, each amenable to numerical solution. Below we
detail how to calculate each individual integral in Eq. (42).
The first approach is to replace each polymer atom coordinate xi by the
sum of its original position hxii and a deviation vector,


xi hxi i 2h2 i

1=2

1
t0i
@ u0i A
v0i

102

Below we leave off each atom subscript i from the deviation vector in order
to simplify the notation. Differential changes in (xi
hxii) and (t0 , u0 , v0 )T are
related by a Jacobian

3=2
dxi
hxi idyi
hyi idzi
hzi i 2h2 i dt0 du0 dv0
The square of each penetrantpolymer atom separation can then be
written as

p 2 
p 2
D2i rp x
hxi i
2h2 it0 rp y
hyi i
2h2 iu0

p 2
rp z
hzi i
2h2 iv0
The contribution to the configurational integral can then be written as
(using a Lennard-Jones potential as an example)

3=2

Zji rp 

1
1

"

ip6
4ip ip12
exp



kB T D2 6 D2 3

1
1

!#
02

02

02

e
t e
u e
v dt0 du0 dv0

103

and its value depends on the orientation vector rp
hxi i: Three-dimensional

numerical integration could then yield the value of the integral.
Fortunately a simpler integration scheme is possible. Prior to substituting
as in Eq. (102), the coordinate system can be rotated, without any loss in
generality, such that the vector rp
hxi i lies along the z axis in the rotated
coordinate system. The penetrantpolymer atom distance simplifies to

p 2
D2i 2h2 it02 2h2 iu02 rp z
hzi i
2h2 iv0

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

481

and the contribution to the configurational integral depends only on the

original distance between the penetrant and polymer atom, jrp
hxi ij. Zji
can then be evaluated using GaussHermite integration [199], in which an
2
integrand containing e
x is approximated by
Z

f x exp
x2 dx 

1

n
X

wk f xk

104

k1

i.e., the function f evaluated at the roots xk of a GaussHermite polynomial

and scaled by weighting factors wk. For the contribution of one penetrant
polymer atom separation to the overall configurational integral,


Zji jrp
hxi ij
Z1Z

3=2

1


3=2

X
kz


V ip
exp

exp
t02 exp
u02 exp
v02 dt0 du0 dv0
kB T

1
1
2
2
0
1

BX
6X
6 4
ip
B
6
6
wkz 6
wky B
wkx exp6

@k
4k
4 kB T
y

0
B
ip12
B
 B


p 2 6
@
2h2 it0k 2 2h2 iu0k 2 rp z
hzi i
2h2 iv0k
1313
C7C7
ip6
C7C7
7C7


2 3 C
p

A5A5
2
2
0
0
0
2
2
2
2h itk 2h iuk rp z
hzi i
2h ivk

105
For an odd order n, one GaussHermite root equals zero, meaning the
energy is evaluated at the voxel center (t0 , u0 , v0 ) (0, 0, 0). Each individual
configurational integral is thus determined by the summation in Eq. (105).
The rotation of the interatomic
vector

 to along the z axis is useful because
it allows the function Zji jrp
hxij in Eq. (105) to be pretabulated in one
dimension (the separation), rather than evaluated from scratch for each
penetrantpolymer atom separation in each voxel. The pretabulated function thus depends on jrp
hxij=, h2 i1=2 =, and /kBT. The same procedure

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

482

Greenfield

can be applied for other two-body potentials by making appropriate

substitutions for V ip in Eq. (105).
The second approach is to express the deviation vector between xi and
hxii in spherical coordinates,

1=2
x
hxi i 2h2 i r sin
cos 

1=2
y
hyi i 2h2 i r sin
sin 

1=2
z
hzi i 2h2 i r cos

with Jacobian

3=2
dxi
hxi idyi
hyi idzi
hzi i 2h2 i r2 sin
drd
d
The coordinate r is allowed to vary over
1  r  1, restricting the

coordinate to 0 
 . The penetrantpolymer atom separation can then
be written
p
1=2
D2i rp x
hxi i2 rp y
hyi i2 rp z
hzi i2 2h2 ir2
2 2h2 i r
 rp x
hxi i sin
cos  rp y
hyi i sin
sin  rp z
hzi i cos

As above, the penetrantpolymer separation vector can be rotated, without
loss in generality, to lie on the z axis of the rotated coordinate system.
The components rp x
hxi i and rp y
hyi i then both equal zero, and the
separation distance is
p

2
1=2
D2i rp z
hzi i 2h2 ir2
2 2h2 i rrp z
hzi i cos

106

The  coordinate has no impact on the distance in this rotated frame,

so integration is analytic. The remaining integral for each polymer atom
contribution to the single voxel partition function is

 2
Zji jrp
hxi ij p


1
2

drr exp
r

1

"

!#
ip6
4ip ip12
dcos
exp



kB T D2i 6 D2i 3
107

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

483

A numerical scheme such as the trapezoidal rule or Simpsons rule [109] can
be used for the integral over cos
, with GaussHermite integration used
to integrate over r.
In comparing the two methods at a fixed separation of 5 A, we found
essentially identical results so long as enough terms were used. For the first
approach, 79 terms within each sum were sufficient (73 343 terms in all).
For the second approach, 79 GaussHermite terms were sufficient as well,
but in combination with of order 104 trapezoidal terms. The first method
(wholly in Cartesian coordinates) is thus recommended for pretabulating
the function Zji(Di).
Even within the approach based on Cartesian coordinates, more
GaussHermite terms are required for each coordinate at shorter separations. Figure 7 compares the modified potential energy
kBT lnZji(r) as
calculated using 316 terms. Within the repulsive regime, using too few
terms leads to oscillations about the potential of mean force calculated with
more terms. The smaller the separation included, the more terms required.

FIG. 7 Modified potential energy between a penetrant and a single polymer atom,
within the average fluctuating polymer method. Different lines correspond to the
numerical results based on different numbers of GaussHermite terms. The same
number of terms was used for each of the x, y, and z coordinates. Increasing the
number of terms decreases oscillations at small separations.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

484

Greenfield

In practice, this additional requirement is not a large burden if a

pretabulated integral Zji is constructed.

REFERENCES
1. Vieth, W.R. Diusion in and Through Polymers: Principles and Applications;
Hansen: New York, 1991.
2. Paul, D.R.; Yampolskii, Y.P. Polymeric Gas Separation Membranes; CRC
Press: Boca Raton, FL, 1994; Chapter 1.
3. Koros, W.J.; Fleming, G.K. J. Membr. Sci. 1993, 83, 180.
4. Nagai, K.; Masuda, T.; Nakagawa, T.; Freeman, B.D.; Pinnau, I. Prog. Polym.
Sci. 2001, 26, 721798.
5. Jorgensen, W.L.; Maxwell, D.S.; Tirado-Rives, J. J. Am. Chem. Soc. 1996, 118,
1122511236.
6. Rigby, D.; Sun, H.; Eichinger, B.E. Polym. Internat. 1997, 44, 311330.
7. Sun, H. J. Phys. Chem. B 1998, 102, 73387364.
8. Martin, M.G.; Siepmann, J.I. J. Phys. Chem. B 1998, 102, 25692577.
9. Martin, M.G.; Siepmann, J.I. J. Phys. Chem. B 1999, 103, 45084517.
10. Chen, B.; Siepmann, J.I. J. Phys. Chem. B 1999, 103, 53705379.
11. Wick, C.D.; Martin, M.G.; Siepmann, J.I. J. Phys. Chem. B 2000, 104,
80088016.
12. Chen, B.; Poto, J.J.; Siepmann, J.I. J. Phys. Chem. B 2001, 105, 30933104.
13. Nath, S.K.; Escobedo, F.A.; de Pablo, J.J. J. Chem. Phys. 1998, 108,
99059911.
14. Nath, S.K.; de Pablo, J.J. Mol. Phys. 2000, 98, 231238.
15. Smith, G.D.; Jae, R.L.; Yoon, D.Y. Macromolecules 1993, 26, 298304.
16. Smith, G.D.; Jae, R.L.; Yoon, D.Y. J. Phys. Chem. 1993, 97, 1275212759.
17. Muller-Plathe, F. Acta Polymerica 1994, 45, 259293.
18. Smith, G.D.; Borodin, O.; Bedrov, D. J. Phys. Chem. A 1998, 102,
1031810323.
19. Hedenqvist, M.S.; Bharadwaj, R.; Boyd, R.H.; Macromolecules 1998, 31,
15561564.
20. Holt, D.B.; Farmer, B.L. Polymer 1999, 40, 46674772.
21. Neyertz, S.; Pizzi, A.; Merlin, A.; Maigret, B.; Brown, D.; Deglise, X. J. Appl.
Polym. Sci. 2000, 78, 19391946.
22. Allen, M.P.; Tildesley, D.J. Computer Simulation of Liquids; Oxford University
Press: New York, 1987.
23. Pant, P.V.K.; Boyd, R.H. Macromolecules 1993, 26, 679686.
24. Gusev, A.A.; Muller-Plathe, F.; van Gunsteren, W.F.; Suter, U.W. Adv.
Polym. Sci. 1994, 116, 207247.
25. Gee, R.H.; Boyd, R.H. Polymer 1995, 36, 14351440.
26. Takeuchi, H. J. Chem. Phys. 1990, 93, 44904491.
27. Takeuchi, H.; Okazaki, K. J. Chem. Phys. 1990, 92, 56435652.
28. Takeuchi, H.; Roe, R.-J.; Mark, J.E. J. Chem. Phys. 1990, 93, 90429048.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
59.
60.
61.
62.
63.
64.

485

Boyd, R.H.; Pant, P.V.K. Macromolecules 1991, 24, 63256331.

Pant, P.V.K.; Boyd, R.H. Macromolecules 1992, 25, 494495.
Muller-Plathe, F. J. Chem. Phys. 1991, 94, 31923199.
Muller-Plathe, F. Chem. Phys. Lett. 1991, 177, 527535.
Muller-Plathe, F. J. Chem. Phys. 1992, 96, 32003205.
Muller-Plathe, F.; Rogers, S.C.; van Gunsteren, W.F. Macromolecules 1992,
25, 67226724.
Trohalaki, S.; Kloczkowski, A.; Mark, J.E.; Roe, R.J.; Rigby, D. Comput.
Theor. Polym. Sci. 1992, 2, 147.
Tamai, Y.; Tanaka, H.; Nakanishi, K. Macromolecules 1994, 27, 44984508.
Han, J.; Boyd, R.H. Macromolecules 1994, 27, 53655370.
Chassapis, C.S.; Petrou, J.K.; Petropoulos, J.H.; Theodorou, D.N.
Macromolecules 1996, 29, 36153624.
Takeuchi, H.; Okazaki, K. Mol. Simul. 1996, 16, 5974.
Fukuda, M.; Kuwajima, S. J. Chem. Phys. 1997, 107, 21492158.
Fukuda, M.; Kuwajima, S. J. Chem. Phys. 1998, 108, 30013009.
Cuthbert, T.R.; Wagner, N.J.; Paulaitis, M.E.; Murgia, G.; DAguanno, B.
Macromolecules 1999, 32, 50175028.
van der Vegt, N.F.A. Macromolecules 2000, 33, 31533160.
Bosho, J.H.D.; Lobo, R.F.; Wagner, N.J. Macromolecules 2001, 34,
61076116.
Nath, S.K.; Banaszak, B.J.; de Pablo, J.J. Macromolecules 2001, 34, 78417848.
Sok, R.M.; Berendsen, H.J.C.; van Gunsteren, W.F. J. Chem. Phys. 1992, 96,
46994704.
Fritz, L.; Hofmann, D. Polymer 1997, 38, 10351045.
Charati, S.G.; Stern, S.A. Macromolecules 1998, 31, 55295535.
Han, J.; Boyd, R.H. Polymer 1996, 37, 17971804.
Muller-Plathe, F. Macromolecules 1996, 29, 47824791.
Li, T.; Kildsig, D.O.; Park, K. J. Control. Release 1997, 48, 5766.
Bharadwaj, R.K.; Boyd, R.H. Polymer 1999, 40, 42294236.
Shanks, R.; Pavel, D. Mol. Simul. 2002, 28, 939969.
Smit, E.; Mulder, M.H.V.; Smolders, C.A.; Karrenbeld, H.; van Eerden, J.;
Feil, D. J. Membr. Sci. 1992, 73, 247257.
Zhang, R.; Mattice, W.L. J. Membr. Sci. 1995, 108, 1523.
Hofmann, D.; Ulbrich, J.; Fritz, L.; Paul, D. Polymer 1996, 37, 47734785.
Hofmann, D.; Fritz, L.; Paul, D. J. Membr. Sci. 1998, 144, 145159.
Hofmann, D.; Fritz, L.; Ulbrich, J.; Schepers, C.; Bohning, M. Macromol.
Theory Simul. 2000, 9, 293327.
Niemela, S.; Leppanen, J.; Sundholm, F. Polymer 1996, 37, 41554165.
Fried, J.R.; Sadat-Akhavi, M.; Mark, J.E. J. Membr. Sci. 1998, 149, 115126.
Kim, W.-K.; Mattice, W.L. Macromolecules 1998, 31, 93379344.
Fried, J.R.; Ren, P. Comput. Theor. Polym. Sci. 2000, 10, 447463.
Tocci, E.; Hofmann, D.; Paul, D.; Russo, N.; Drioli, E. Polymer 2001, 42,
521533.
Fried, J.R.; Goyal, D.K. J. Polym. Sci. Polym. Phys. 1998, 36, 519536.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

486

Greenfield

65. Sonnenburg, J.; Gao, J.; Weiner, J.H. Macromolecules 1990, 23, 46534657.
66. Rottach, D.R.; Tillman, P.A.; McCoy, J.D.; Plimpton, S.J.; Curro, J.G.
J. Chem. Phys. 1999, 111, 98229831.
67. Muller-Plathe, F. J. Chem. Phys. 1995, 103, 47454756.
68. Catlow, C.R.A.; Mills, G.E. Electrochim. Acta 1995, 40, 20572062.
69. Payne, V.A.; Lonergan, M.C.; Forsyth, M.; Ratner, M.A.; Shriver, D.F.;
de Leeuw, S.W.; Perram, J.W. Solid State Ionics 1995, 81, 171181.
70. Neyertz, S.; Brown, D. J. Chem. Phys. 1996, 104, 37973809.
71. Ennari, J.; Elomaa, M.; Sundholm, F. Polymer 1999, 40, 50355041.
72. Borodin, O.; Smith, G.D. Macromolecules 2000, 33, 22732283.
73. Snyder, J.F.; Ratner, M.A.; Shriver, D.F. J. Electrochem. Soc. 2001, 148,
A858A863.
74. Aabloo, A.; Thomas, J. Solid State Ionics 2001, 143, 8387.
75. Vishnyakov, A.; Neimark, A.V. J. Phys. Chem. B 2001, 105, 95869594.
76. Ferreira, B.A.; Muller-Plathe, F.; Bernardes, A.T.; De Almeida, W.B.
Solid State Ionics 2002, 147, 361366.
77. Muller-Plathe, F.; Rogers, S.C.; van Gunsteren, W.F. J. Chem. Phys. 1993, 98,
98959904.
78. Sunderrajan, S.; Hall, C.K.; Freeman, B.D. J. Chem. Phys. 1996, 105,
16211632.
79. Takeuchi, H. J. Chem. Phys. 1990, 93, 20622067.
80. Muller-Plathe, F.; Laaksonen, L.; van Gunsteren, W.F. J. Mol. Graphics 1993,
11, 118120.
81. Muller-Plathe, F.; Laaksonen, L.; van Gunsteren, W.F. J. Mol. Graphics 1993,
11, 125126.
82. Glasstone, S.; Laidler, K.J.; Eyring, H. The Theory of Rate Processes;
McGraw-Hill: New York, 1941.
83. Chandler, D. J. Chem. Phys. 1978, 68, 2959.
84. Voter, A.F.; Doll, J.D. J. Chem. Phys. 1985, 82, 8092.
85. June, R.L.; Bell, A.T.; Theodorou, D.N. J. Phys. Chem. 1991, 95, 88668878.
86. Fichthorn, K.A.; Weinberg, W.H. J. Chem. Phys. 1991, 95, 10901096.
87. Greeneld, M.L.; Theodorou, D.N. Macromolecules 1993, 26, 54615472.
88. Greeneld, M.L.; Theodorou, D.N. Macromolecules 1998, 31, 70687090.
89. Rallabandi, P.S.; Thompson, A.P.; Ford, D.M. Macromolecules 2000, 33,
31423152.
90. Greeneld, M.L.; Theodorou, D.N. Mol. Simul. 1997, 19, 329361.
91. Greeneld, M.L.; Theodorou, D.N. Macromolecules 2001, 34, 85418553.
92. Jagodic, F.; Borstnik, B.; Azman, A. Makromol. Chem. 1973, 173, 221231.
93. Arizzi, S. Diusion of Small Molecules in Polymeric Glasses: A Modelling
Approach. Ph.D. thesis, Massachusetts Institute of Technology, 1990.
94. Gusev, A.A.; Arizzi, S.; Suter, U.W.; Moll, D.J. J. Chem. Phys. 1993, 99,
22212227.
95. Bell, A.T.; Maginn, E.J.; Theodorou, D.N. In Handbook of Heterogeneous
Catalysis, Ertl, G., Knozinger, H., Weitkamp, J., Eds.; VCH: Weinheim, 1997.
96. Auerbach, S.M. Int. Rev. Phys. Chem. 2000, 19, 155198.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

97.
98.
99.
100.
101.
102.

103.
104.
105.
106.
107.
108.
109.

110.

111.
112.
113.
114.
115.
116.
117.
118.
119.
120.
121.
122.
123.

487

Gusev, A.A.; Suter, U.W. J. Chem. Phys. 1993, 99, 22282234.

Gusev, A.A.; Suter, U.W.; Moll, D.J. Macromolecules 1995, 28, 25822584.
Hofmann, D.; Fritz, L.; Ulbrich, J.; Paul, D. Polymer 1997, 38, 61456155.
Hofmann, D.; Fritz, L.; Ulbrich, J.; Paul, D. Comput. Theor. Polym. Sci.
2000, 10, 419436.
Products from Accelrys, Inc., http://www.accelrys.com/chemicals/polymers/
diusion.html.
Greeneld, M.L. Molecular Modeling of Dilute Penetrant Gas Diusion
in a Glassy Polymer using Multidimensional Transition-State Theory. Ph.D.
thesis, University of California, Berkeley, 1996.
van Kampen, N.G. Stochastic Processes in Physics and Chemistry, Revised
Ed.; North Holland: Amsterdam, 1992.
Synge, J.L.; Schild, A. Tensor Calculus; University of Toronto Press: Toronto,
1949.
Go , N.; Scheraga, H.A. Macromolecules 1976, 9, 535542.
Cerjan, C.J.; Miller, W.H. J. Chem. Phys. 1981, 75, 28002806.
Banerjee, A.; Adams, N.; Simons, J.; Shepard, R. J. Phys. Chem. 1985, 89,
5257.
Baker, J. J. Comput. Chem. 1986, 7, 385395.
Press, W.H.; Flannery, B.P.; Teukolsky, S.A.; Vetterling, W.T. Numerical
Recipes. The Art of Scientic Computing; Cambridge University Press:
Cambridge, 1989.
Henkelman, G.; Jonsson, G.; Jonsson, H. In Theoretical Methods in
Condensed Phase Chemistry (Progress in Theoretical Chemistry and Physics),
Schwartz, S.D., Ed.; Kluwer Academic Publishers: New York, 2000; Vol. 5,
269300.
Bolhuis, P.G.; Chandler, D.; Dellago, C.; Geissler, P.L. Ann. Rev. Phys.
Chem. 2002, 53, 291318.
Fukui, K. Acc. Chem. Res. 1981, 14, 363368.
deGennes, P.-G. Scaling Concepts in Polymer Physics; Cornell University
Press: Ithaca, 1979.
Doi, M.; Edwards, S.F. The Theory of Polymer Dynamics; Oxford University
Press: New York, 1986; Chapter 6.
Pechukas, P. J. Chem. Phys. 1976, 64, 15161521.
Page, M.; McIver, J.W., Jr. J. Chem. Phys. 1988, 88, 922935.
Banerjee, A.; Adams, N.P. Int. J. Quantum Chem. 1992, 43, 855871.
Sevick, E.M.; Bell, A.T.; Theodorou, D.N. J. Chem. Phys. 1993, 98,
31963212.
Weiner, J.H. Statistical Mechanics of Elasticity; Wiley-Interscience: New York,
1983.
http://www.netlib.org/eispack
http://www.netlib.org/lapack
http://www.netlib.org/linpack
Michael L. Greeneld, Doros N. Theodorou, J. Chem. Phys. (manuscript in
preparation).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

488

Greenfield

124. Starkweather, H.W. Polymer 1991, 32, 24432448.

125. Elber, R. J. Chem. Phys. 1990, 93, 43124321.
126. Lazaridis, T.; Tobias, D.J.; Brooks, C.L.; Paulaitis, M.E. J. Chem. Phys. 1991,
95, 76127625.
127. Pratt, L.R. J. Chem. Phys. 1986, 85, 50455048.
128. Gillilan, R.; Wilson, K.R. J. Chem. Phys. 1992, 97, 17571772.
129. Theodorou, D.N.; Suter, U.W. Macromolecules 1985, 18, 14671478.
130. Rapold, R.F.; Mattice, W.L. J. Chem. Soc. Faraday Trans. 1995, 91,
24352441.
131. Rapold, R.F.; Mattice, W.L. Macromolecules 1996, 29, 24572466.
132. Kotelyanskii, M.; Wagner, N.J.; Paulaitis, M.E. Macromolecules 1996, 29,
84978506.
133. Kotelyanskii, M. Trends Polym. Sci. 1997, 5, 192198.
134. Santos, S.; Suter, U.W.; Muller, M.; Nievergelt, J. J. Chem. Phys. 2001, 114,
97729779.
135. Muller, M.; Nievergelt, J.; Santos, S.; Suter, U.W. J. Chem. Phys. 2001, 114,
97649771.
136. Ratner, B.D.; Tsukruk, V.V. Scanning Probe Microscopy of Polymers;
American Chemical Society: Washington, DC, 1998; Symposium Series,
No. 694.
137. Fischer, K.H.; Hertz, J.A. Spin Glasses; Cambridge University Press:
Cambridge, 1991.
138. Snurr, R.Q.; Bell, A.T.; Theodorou, D.N. J. Phys. Chem. 1994, 98,
1194811961.
139. Goldstein, H. Classical Mechanics, 2nd Ed.; Addison-Wesley: Reading, MA,
1980.
140. Anderson, E.; Bai, Z.; Bischof, C.; Blackford, S.; Demmel, J.; Dongarra, J.;
Croz, J.D.; Greenbaum, A.; Hammarling, S.; McKenney, A.; Sorensen, D.
LAPACK Users Guide; Society for Industrial and Applied Mathematics:
Philadelphia, 1999; available on-line at http://www.netlib.org/lapack/lug/
141. Vorono , G. Z. Reine Angew. Math. 1908, 134, 198211.
142. Tanemura, M.; Ogawa, T.; Ogita, N. J. Comput. Phys. 1983, 51, 191207.
143. Arizzi, S.; Mott, P.H.; Suter, U.W. J. Polym. Sci. Polym. Phys. Ed. 1992, 30,
415426.
144. Gentile, F.T.; Arizzi, S.; Suter, U.W.; Ludovice, P.J. Ind. Eng. Chem. Res.
1995, 34, 41934201.
145. Putta, S.; Nemat-Nasser, S. Materials Sci. Eng. A 2001, 317, 7076.
146. Takeuchi, H.; Okazaki, K. Makromol. Chem. Macromol. Symp. 1993, 65,
8188.
147. Misra, S.; Mattice, W.L. Macromolecules 1993, 26, 72747281.
148. Cullity, B.D. Elements of X-ray Diraction, 2nd Ed.; Addison-Wesley:
Reading, MA, 1978.
149. Gonzalez, C.; Schlegel, H.B. J. Phys. Chem. 1990, 94, 55235527.
150. Fischer, S.; Karplus, M. Chem. Phys. Lett. 1992, 194, 252261.
151. Czerminski, R.; Elber, R. Int. J. Quant. Chem. 1990, 24, 167186.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Sorption and Diffusion by TST

152.
153.
154.
155.
156.
157.
158.
159.
160.

161.
162.
163.
164.
165.
166.
167.
168.
169.
170.
171.
172.
173.
174.
175.
176.
177.
178.
179.
180.
181.
182.
183.
184.

489

Mills, G.; Jonsson, H.; Schenter, G.K. Surf. Sci. 1995, 324, 305337.
Zimmer, M.F. Phys. Rev. Lett. 1995, 75, 14311434.
Petropoulos, J.H. J. Membr. Sci. 1990, 53, 229255.
Theodorou, D.N. In Diusion in Polymers; Neogi, P., Ed.; Marcel Dekker:
New York, 1996; 67142.
Widom, B. J. Chem. Phys. 1963, 39, 28082812.
Widom, B. J. Phys. Chem. 1982, 86, 869872.
Muller-Plathe, F. Macromolecules 1991, 24, 64756479.
Tamai, Y.; Tanaka, H.; Nakanishi, K. Macromolecules 1995, 28, 25442554.
Boone, T.D. Prediction of Glass-Melt Behavior and Penetrant Sorption
Thermodynamics in Vinyl Polymers via Molecular Simulations. Ph.D. thesis,
University of California, Berkeley, 1995.
Spyriouni, T.; Economou, I.G.; Theodorou, D.N. Phys. Rev. Lett. 1998, 80,
44664469.
Kofke, D.A. Adv. Chem. Phys. 1999, 105, 405441.
Panagiotopoulos, A.Z. Mol. Phys. 1987, 61, 813826.
Panagiotopoulos, A.Z. Mol. Simul. 1992, 9, 124.
Shing, K.S.; Gubbins, K.E. Mol. Phys. 1982, 46, 11091128.
Knopp, B.; Suter, U.W.; Gusev, A.A. Macromolecules 1997, 30, 61076113.
Knopp, B.; Suter, U.W. Macromolecules 1997, 30, 61146119.
van der Vegt, N.F.A.; Briels, W.J. J. Chem. Phys. 1998, 109, 75787582.
Escobedo, F.A.; de Pablo, J.J. J. Chem. Phys. 1995, 103, 27032710.
Escobedo, F.A.; de Pablo, J.J. J. Chem. Phys. 1996, 105, 43914394.
Nath, S.K.; de Pablo, J.J.; DeBellis, A.D. J. Am. Chem. Soc. 1999, 121,
42524261.
de Pablo, J.J.; Laso, M.; Suter, U.W.; Cochran, H.D. Fluid Phase Equilibria
1993, 83, 323331.
Zervopoulou, E.; Mavrantzas, V.G.; Theodorou, D.N. J. Chem. Phys. 2001,
115, 28602875.
Nath, S.K.; de Pablo, J.J. J. Phys. Chem. B 1999, 103, 35393544.
Kumar, S.K.; Szleifer, I.; Panagiotopoulos, A.Z. Phys. Rev. Lett. 1991, 66,
29352938.
Kikuchi, H.; Kuwajima, S.; Fukuda, M. J. Chem. Phys. 2001, 115,
62586265.
Deitrick, G.L.; Scriven, L.E.; Davis, H.T. J. Chem. Phys. 1989, 90, 23702385.
Stapleton, M.R.; Panagiotopoulos, A.Z. J. Chem. Phys. 1990, 92, 12851293.
Fukuda, M. J. Chem. Phys. 2000, 112, 478486.
de Pablo, J.J.; Laso, M.; Suter, U.W. J. Chem. Phys. 1992, 96, 61576162.
Siepmann, J.I.; Frenkel, D. Mol. Phys. 1992, 75, 5970.
Dodd, L.R.; Boone, T.D.; Theodorou, D.N. Mol. Phys. 1993, 78, 961996.
de Pablo, J.J.; Yan, Q.; Escobedo, F.A. Annu. Rev. Phys. Chem. 1999, 50,
377411.
Ferguson, D.M.; Siepmann, J.I.; Truhlar, D.G., Eds. Monte Carlo Methods in
Chemical Physics (Adv. Chem. Phys.); John Wiley & Sons: New York, 1999;
Vol. 105.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

490

Greenfield

185. Bezus, A.G.; Kiselev, A.V.; Lopatkin, A.A.; Du, P.Q. J. Chem. Soc. Faraday
Trans. II 1978, 74, 367379.
186. Gusev, A.A.; Suter, U.W. Phys. Rev. A 1991, 43, 64886494.
187. McGreevy, R.L.; Howe, M.A. Annu. Rev. Mater. Sci. 1992, 22, 217242.
188. Alfrey, T., Jr.; Gurnee, E.F.; Lloyd, W.G. J. Polym. Sci. C 1966, 12, 249261.
189. Havlin, S.; Ben-Avraham, D. Adv. Phys. 1987, 36, 695798.
190. Haus, J.W.; Kehr, K.W. Phys. Rep. 1987, 150, 263406.
191. Stauer, D.; Aharony, A. Introduction to Percolation Theory, 2nd Ed.; Taylor
& Francis: London, 1992.
192. Muller-Plathe, F.; Rogers, S.C.; van Gunsteren, W.F. Chem. Phys. Lett. 1992,
199, 237243.
193. Weber, H.; Paul, W. Phys. Rev. E 1996, 54, 39994007.
194. Brandrup, J.; Immergut, E.H.; Grulke, E.A. Polymer Handbook, 4th Ed.;
WileyInterscience: New York, 1999.
195. Fixman, M. Proc. Nat. Acad. Sci. USA 1974, 71, 30503053.
196. Go , N.; Scheraga, H.A. J. Chem. Phys. 1969, 51, 47514767.
197. Helfand, E. J. Chem. Phys. 1979, 71, 50005007.
198. Fixman, M. J. Chem. Phys. 1978, 69, 15271537.
199. Abramowitz, M.; Stegun, I.A. Handbook of Mathematical Functions; National
Bureau of Standards: Washington, DC, 1964; Reprint published by Dover
Publications: Mineola, New York, 1965.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

15
Coarse-Graining Techniques
K. BINDER and WOLFGANG PAUL
Mainz, Mainz, Germany

Johannes-Gutenberg-Universitat

SERGE SANTOS and ULRICH W. SUTER

Hochschule (ETH), Zurich, Switzerland

I.

Eidgenossische Technische

INTRODUCTION AND OVERVIEW

For many materials, it is difficult to span the wide range of scales of length
and time from the atomistic description to that of macroscopic phenomena.
For polymers, this multiscale problem is particularly severe, since a single
flexible macromolecule already exhibits structure on many length scales
(Fig. 1), and is a slowly relaxing object [1,2]. In addition, one often wishes to
consider systems with multiple constituents (concentrated polymer solutions
or blends, mixtures of homopolymers and block copolymers or polymers of
different architecture) which exist in several phases, and then the scales of
length and time that are of physical interest truly extend from the subatomic
scale (relevant for the electronic structure of polymers, their chemical
reactions, excited states, etc.) to the macroscopic scale (e.g., phase separated
domains of the scale of micrometers or larger, glassy relaxation on the scale
of seconds or hours, etc.) [3].
Since there is no hope of treating all these scales equally well in a single
type of computer simulation, the idea of coarse-graining has emerged
[4,5], where a mapping was attempted from an atomistically realistic
description of polymers such as polyethylene (PE) or bisphenol-Apolycarbonate (BPA-PC) towards the bond fluctuation model [6,7]; see
Fig. 2. Now coarse-grained lattice models such as the self-avoiding walk
model or the bond fluctuation model in its standard form have been used for
a long time to model polymers qualitatively, but this has the obvious
drawback that the information on the chemical structure of the considered
polymer is lost completely, and thus nothing can be said on structure
property relationships. Though these standard coarse-grained models are

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

492

Binder et al.

FIG. 1 Schematic illustration of the various length scales of a macromolecule.

Polyethylene is used as an example. From Binder [1].

FIG. 2 A chemically realistic description of a polymer chain (bisphenol-Apolycarbonate [BPA-PC] in the present example) is mapped approximately onto
the bond fluctuation model by using suitable potentials for the length of the
effective bonds and the angles between them. In this example (3:1 mapping) one
chemical repeat unit of BPA-PC containing n 12 covalent bonds along the
backbone of the chain is translated into three effective bonds. From Paul et al. [5].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Coarse-Graining Techniques

493

nevertheless useful for some purposes (see Chapter 4), the idea to be
described here is the attempt to keep some information about the chemical
structure indirectly also on the coarse-grained level. Thus, although the
lattice spacing of the model in Fig. 2 corresponds to about 2 A, and hence
clearly the detailed packing of the various atoms or atomic groups in the
polymer melt or glassy state cannot be directly represented, some
information on such properties is kept indirectly, via the potentials U(),
V(#) for the lengths, , of the effective bonds and the angles, #, between
them. Chemically different polymers are then described by different
potentials.
Obviously, constructing the effective potentials explicitly will not be very
straightforward [8,9], and also the description of physical properties by
means of such a coarse-grained lattice model cannot be expected to be
perfect. Thus, the idea has arisen to use the coarse-grained model only as an
intermediate step for a simulation that equilibrates the long wavelength
degrees of freedom of the polymer and then to carry out an inverse
mapping where in Fig. 2 the arrow is simply read backwards! However, this
step is technically very difficult, and although it has been under discussion for
the bond fluctuation model for more than a decade now, it has not yet
actually been performed successfully (only for another type of lattice model
on the face-centered cubic lattice, obtained from chains on the diamond
lattice where every second monomer is eliminated, see [10,11], has a reverse
mapping been possible). The explicit construction of potentials for coarsegrained degrees of freedom [such as U(), V(#)] and the inverse mapping has
successfully been carried out for a coarse-grained continuum model rather
than a lattice model, e.g., a bead-spring model [12,13]. The basic idea is the
same in all these methods (Fig. 3), but there is considerable freedom in
defining the details of this procedure. For example, in the case of bisphenolA-polycarbonate a 2:1 mapping to a bead-spring chain was performed
[12,13], i.e., one chemical monomer (representing 12 consecutive covalent
bonds along the chain backbone, see Fig. 2) is mapped onto two effective
monomers. Taking the geometrical centers of the carbonate group and of the
isopropylidene group as the sites of the effective monomers, it is actually
advantageous to allow for a difference between these effective monomers:
thus, the coarse-grained chain representing BPA-PC is not a bead-spring
model of a homopolymer A, A, A, A . . . , but rather a bead-spring model of
an alternating copolymer ABAB. . . . The crucial point of all these models,
however, is such that on the coarse-grained length scale the effective potentials U(), V(#) are much smaller than the atomistic potentials for the lengths
of covalent bonds or the angles between them, and torsional potentials on the
coarse-grained level can be neglected completely. As a result, on the coarsegrained level all potentials are on the same scale of energy, unlike the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

494

Binder et al.

FIG. 3 Illustration of the coarse-graining procedure, starting out from a unitedatom model for polyethylene as an atomistic description (the larger spheres show a
segment of polyethylene at a temperature T 509K taken out of a molecular
dynamics melt). One coarse-grained bond (straight line) represents the end-to-end
distance of n 5 consecutive united-atom bonds. The effective monomers (black
dots) actually are represented by the eight sites of an elementary cube of the simple
cubic lattice, if a mapping on the bond fluctuation model is used; if a mapping on a
bead-spring chain is used, the effective monomers (black dots) are represented by the
repulsive part of a suitable Lennard-Jones potential, and the effective bonds are
anharmonic springs described by the FENE potential. From Tries et al. [9].

atomistic level where the scales for the potentials of covalent bond lengths
and angles are much higher than the weak Lennard-Jones type nonbonded
forces. Hence, while on the atomistic scale a Molecular Dynamics time step
on the femtosecond scale is needed, the molecular dynamics time step for a
coarse-grained bead-spring model can be several orders of magnitude larger.
As a result, one can reach equilibrium for distinctly lower temperatures using
the coarse-grained model than with the atomistic simulation, and somewhat
larger length scales are accessible, too.
Nevertheless, one must admit that this step emphasized in Figs. 2 and 3
is only a modest first step towards the goal of bridging the gap between
the length and time scales of atomistic simulations and the scales of
macroscopic phenomena in materials science. The same caveat applies to
other interesting alternatives to this mapping procedure, which we shall not
describe in detail here, such as mapping to ellipsoidal rather than spherical
effective monomers [14,15], or generating configurations of dense polymer
melts by constructing an ensemble of curves of constant curvature, using
stream lines of a vector field that never intersect [16,2]. But how can one get
from the scales of, say, 10 nm and 100 sec, that are only accessible to the
coarse-grained models, to still much larger scales?
One line of attack to this problem is to try to use a next mapping
step from a bead-spring chain to a still much cruder model. For instance,

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Coarse-Graining Techniques

495

Murat and Kremer [17] try to map whole chains of soft ellipsoidal particles,
which then must be allowed to overlap strongly in the melt. This approach
seems particularly useful for a description of heterophase structures in
polymeric materials, such as polymer blends that are unmixed on
nanoscopic scales. It seems promising to combine this model or related
approaches [18] with efficient techniques for describing the time evolution,
making use of the fact that very soft potentials between the effective
particles allow rather large time steps, such as dissipative particle dynamics
[19]. While these approaches clearly are very promising avenues for further
research, we feel it is premature to review them here simply because this field
is at its beginning and at the same time rapidly developing.
Another line of attack is the idea of bypassing this step-to-step mapping
from one scale to the next by directly making connections between the
atomistic scale and the macroscopic continuum scale by performing
calculations on both scales simultaneously and appropriately connecting
them [20,21]. We shall briefly describe the essential ideas of this approach,
emphasizing mechanical properties, in Section III of this chapter (a first
application to plastic deformation of bisphenol-A-polycarbonate can be
found in Chapter 12 of this book). In the following section, we shall explain
the mapping alluded to in Figs. 2 and 3 in some more detail, while Section
IV summarizes a few conclusions.

II. MAPPING OF ATOMISTIC MODELS TO THE BOND

FLUCTUATION MODEL
In this section, we shall focus on polyethylene exclusively because this is a
chemically simple polymer and for this case the mapping procedure has been
investigated in greater detail than for any other polymer [9]. Using a unitedatom description for the CH2 groups, a potential of the following type
results [22,23]

N
p
1
X
j1

H lj

N
p
2
X

H

j

j1

N
p
2
X

H j

j2

HLJ ~rij

j6i

where H(j) / (j
cc) with a fixed bond length j cc 1.53 A, while the
potential for the cosine of the bond angles
j was assumed harmonic,
H

j

f

cos
j
cos
0 2 ,
2

0 1108,

f
120 kcal=mol

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

496

Binder et al.

The torsional potential was parameterized as follows,

H j

3
X

ak cosk j

k0

with constants a0 1.736, a1 4.5, a2 0.764, and a3
7 (in units

kcal/mol).
Finally, the nonbonded interaction is expressed in terms of the wellknown Lennard-Jones potential with parameters  4.5 A (position of the
) 0.09344 kcal/mol, "(CH3 CH3) 0.22644 kcal/mol,
minimum), "(CH2CH
p2
and "(CH2CH3) "CH2
CH2 "CH3
CH3 . This model is a very
good representation of reality, as the direct comparison of molecular
dynamics simulations for C100H202 at T 509K with neutron spin echo data
shows [24].
The statistical mechanical concept underlying the coarse-graining
approach is the same as for the renormalization group treatment of critical
phenomena. Suppose we have a microscopic model with degrees of freedom
x~ (e.g., the bond lengths, bond angles, etc. of a chemically realistic polymer
model). The canonical partition function of this model is given as
Z


 
exp
H x~

fx~g
where  1=kB T and the sum is to be interpreted as an integral for continuous degrees of freedom. Let us further assume that we have chosen a set of
mesoscopic degrees of freedom (e.g., the lengths and angles of the coarse~ . Then we can write
grained bonds in Fig. 3), denoted by m
Z

XX


 
exp
H x~

fm~ g fx~gm~
~.
where the sum over fx~g is restricted to a fixed set of mesoscopic variables m
~ , T by writing
Introducing a generalized free energy F m
Z




~; T
exp
F m

fm~ g
we see that this defines a mapping from the probabilities for the microscopic
configurations px~ 1=Z expf
Hx~g to those for the degrees of free~ 1=Z expf
Fm
~ , Tg.
dom pm
On the mesoscopic scale we are dealing with a free energy and not with a
Hamiltonian, so that such a mapping has to be done for each temperature

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Coarse-Graining Techniques

497

separately. The generalized free energy on the mesoscopic scale in general

~ , Tg
has to be determined numerically from the partial trace expf
Fm
P
~
expf
H
x
g.
fx~gm~
In the case of melt polymers we are in the fortunate situation, that the
large scale statistical properties of such chains are the same as for isolated
chains under Theta conditions, i.e., those of a random walk. We can therefore generate Theta chains for the atomistic model and evaluate the
probability distribution p(L,
) for adjacent coarse-grained bond lengths and
bond angles (see Fig. 4).
Using Eqs. (1)(3), short single chains are simulated by simple sampling
type Monte Carlo methods, as described by Baschnagel et al. [4,25]; see
also Chapter 4 by K. Binder et al., Polymer Models on the Lattice,
and Chapter 3 by R. Hentschke, Single Chain in Solution, for a general
explanation of the simple sampling method for polymers. However, a
problem is that a cutoff in the Lennard-Jones interaction along the chain
must be introduced in order for an isolated chain to show asymptotic
Gaussian Theta behavior. One must include the interaction of each unit with
its fourth-nearest neighbor along the chain to account for the pentane effect,
while further distant neighbor interactions are eliminated. Now n0 successive
CH2 groups are mapped onto m effective bonds. To some extent, the choice
of the pair of integers (n0 , m) is arbitrary, but it is clear that neither one of
the integers n0 /m and m should be too large, in order that the method is
practically feasible. The optimal choice of n0 and m so that the crude coarsegrained model can represent the atomistic details indirectly as faithfully as
possible, is a subtle problem, and clearly will depend on the chemical
structure of the polymer that is studied. For polyethylene, the choice m 2,
n0 10 (i.e., five CH2 groups map onto an effective bond, anticipated in
Fig. 3) was found to be the best choice, by trial and error [9]. One should
note that the freedom in this choice is constrained by the fact that a
reasonable volume fraction of occupied sites (such as  0.5) must
correspond to melt densities, and the m effective bonds with the bond
angle # 0 must reproduce the length of a piece of n0 units of an all-trans
chain, for geometric consistency.
With these conditions, the experimental density of amorphous polyethylene at room temperature leads to a correspondence between one lattice
spacing and a physical length of a 2.03 A. Figure 4 shows the probability
distributions of the coarse-grained bond length L and the bond angle #
between two such bonds, choosing n0 10 CC bonds [9]. Instead of using
these probability distributions directly (as was done in [12,13]) we choose to
characterize them by their first two moments, taking into account crosscorrelations, and record the moments hLi, hL2i, h#i, h#2 i, hL#i over a wide
range of temperatures from T 250K to T 800K. These moments are used

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

498

Binder et al.

FIG. 4 Probability distribution of (a) the coarse-grained bond length L and (b) the
bond angle
between two such coarse-grained bonds, choosing n0 10, m 2. From
Tries et al. [9].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Coarse-Graining Techniques

499

as an input in a nonlinear fitting procedure, by which the parameters of the

effective potentials U, V# for the length of the effective bonds and the
angles # between them are determined. We parameterize these potentials as
follows
U u0
0 2 1=T
h1=Tiu1
1 2
2

V # v0 cos #
c0 1=T
h1=Tiv1 cos #
c1

7
2

where h1=Ti denotes an average over the temperatures where input information from the atomistic calculation (such as shown in Fig. 4) is used [8].
From this fit the eight parameters u0, u1, 0 , 1 , v0, v1, c0, c1 are obtained,
as well as the time constant 0 that is necessary to translate the Monte Carlo
time units [ one attempted Monte Carlo step (MCS) per effective
monomer] into a physical time. As is well known, the random hopping
algorithm for the bond fluctuation model, where one selects an effective
monomer randomly and tries to displace it in a randomly chosen lattice
direction, can represent the Rouse-like or reptation-like dynamics of dense
polymer melts in a rough way [2,26]; see also Chapter 4. But the Monte
Carlo method never has an intrinsic physical time scale of its own, and just
as the density of polyethylene needs to be used to give the length unit of the
lattice model a physical meaning, one also needs physical input for the time
unit. It is desirable to map the mobility of the lattice model onto the average
jump rate of the torsional degrees of freedom of the chain, since these
motions dominate the relaxation of its overall configuration. Thus a
temperature-dependent time unit MC(T ), which one attempted MCS per
monomer corresponds to, is introduced via






1
4H
1 1 X
min 1; exp

Ator T

MC T
kB T
0 Ntor tor

Here H is the energy change in the Monte Carlo move, to be calculated

from Eqs. (7) and (8), so h. . .i is just the average acceptance factor, using the
Metropolis transition probability for Monte Carlo moves as usual [26]. The
attempt frequency of torsional jumps in the atomistic models is denoted as
1/ 0, Ntor being the number of torsional degrees of freedom in the coarseactivated jump probability. We define an
grained unit and Ator their averageP
average energy barrier hEi from tor Ator T=Ntor  exp
hEikB T, and
write
MC T ABFL 1 0 expEmin =kB T

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

10

500

Binder et al.

where ABFL1 is the acceptance probability of the Monte Carlo moves in

the athermal (T 1) bond fluctuation model, and Emin the smallest
energy barrier in the torsional potential. Equations (9) and (10) yield a
further mapping condition that the potentials U(), V(#) must fulfill, namely


min1; exp
H=kB T ABFL 1 exp
hE i
Emin =kB T

11

Including this information on the energy barriers of the torsional potential

into the mapping that fixes the parameters of the potential U(), V(#)
implies that not all the many choices of bond lengths and bond angles
occurring in the bond fluctuation model can correspond to minima of the
torsional potential. On the other hand, for temperatures in the range from
T 250K to T 800K geometrical properties such as P(L) and P(
), cf.
Fig. 4, would be dominated by the minima of the torsional potential
only. Figure 5 shows the time rescaling factor as obtained from Eq. (10),
determining the absolute scale by equating the chain self-diffusion coefficient measured in the simulation at 450K to the experimentally known
value [27]. While at T 509K MC  10
11 sec, one sees that at T 200K

FIG. 5 Time rescaling factor MC(T ) for the coarse-grained Monte Carlo model for
polyethylene plotted vs. temperature. The constant 0 in Eq. (10) was fixed by fitting
the self-diffusion constant DN for C100H202 to experiment at T 450K. From Tries
et al. [9].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Coarse-Graining Techniques

501

MC(T ) has increased to about a nanosecond. Since Monte Carlo runs of the
order of 107 MCS are easily feasible, this means that relaxation processes on
the millisecond time range can be explored. Figure 6 shows that in this way a
reasonable prediction for the temperature dependence of both static and
dynamic properties of polyethylene melts over a wide temperature range
is obtained, including the melt viscosity [27], with no further adjusted parameters. Of course, this modeling accounts neither for the breaking of
covalent bonds at high temperatures, nor for crystallization of polyethylene
(that occurs for T<Tm 414K), and a further weak point at the present
stage is that attractive intermolecular interactions have been omitted as well.
The off-lattice coarse-grained models [12,13] have not included attractive
intermolecular forces either, and thus none of these approaches yields a

FIG. 6 (a) Characteristic ratio of polyethylene plotted vs. temperature, for N 20

effective monomers (representing C100H202). Symbols denote two versions of the mapping (an exact one, needing four subsequent bonds, and an approximate two-bond
procedure). (b) Logarithm of the viscosity of C100H202 plotted vs. inverse temperature, where full dots are from a constant density simulation of the bond fluctuation
model with the potentials of Eqs. (7) and (8), using simulated self-diffusion constants
DN in the Rouse model formula  =8a3 hRg2 i=6kB T=NDN . Open triangles are
experimental data obtained by Pearson et al. [27] for C100H202 at constant (ambient)
pressure. From Tries et al. [9].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

502

FIG. 6

Binder et al.

Continued.

reasonable equation of state. It appears that using the pressure of the system
as an additional input variable into the mapping of bead-spring type offlattice models to fix the strength of the attractive part of the Lennard-Jones
interaction between nonbonded effective monomers is a promising recipe to
eliminate this weakness of current coarse-graining procedures [28].

III. ATOMISTIC-CONTINUUM MODELS: A NEW

CONCEPT FOR THE SIMULATION OF
DEFORMATION OF SOLIDS
While an explicit treatment of the microscopic, atomistic degrees of freedom
is necessary when a locally realistic approach is required, on a macroscopic
level the continuum framework is perfectly adequate when the relevant
characteristics are scalar, vector, or tensor fields (e.g., displacement fields
for mechanical properties). A combination of the two levels of description
would be useful [29,30]. Here we focus on the deformation of solids: The
elastic mechanics of homogeneous materials is well understood on the
atomistic scale and the continuum theory correctly describes it. Inelastic
deformation and inhomogeneous materials, however, require techniques
that bridge length and time scales.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Coarse-Graining Techniques

503

A number of approaches to connect multiple-scale simulation in finiteelement techniques have been published [3134]. They are able to describe macroscopically inhomogeneous strain (e.g., cracks)even dynamic
simulations have been performed [35]but invariably require artificial
constraints on the atomistic scale [36]. Recently, an approach has been
introduced that considers a system comprising an inclusion of arbitrary
shape embedded in a continuous medium [20]. The inclusion behavior
is described in an atomistically detailed manner [37], whereas the continuum
is modeled by a displacement-based Finite-Element method [38,39]. The
atomistic model provides state- and configuration-dependent material
properties, inaccessible to continuum models, and the inclusion in the
atomistic-continuum model acts as a magnifying glass into the molecular
level of the material.
The basic concept to connect both scales of simulation is illustrated in
Fig. 7. The model system is a periodic box described by a continuum,
tessellated to obtain finite elements, and containing an atomistic inclusion. Any overall strain of the atomistic box is accompanied by an identical
strain at the boundary of the inclusion. In this way, the atomistic box does
not need to be inserted in the continuum, or in any way connected (e.g., with
the nodal points describing the mesh at the boundary of the inclusion). This
coupling via the strain is the sole mechanism to transmit tension between the
continuum and the atomistic system. The shape of the periodic cells is
described by a triplet of continuation (column) vectors for each phase (see
also [21]), A, B, and C for the continuous body, with associated scaling
matrix H [ABC], and analogously a, b, and c for the atomistic inclusion,
with the scaling matrix h [abc] (see Fig. 8).
The material morphology is specified by a set of nodal points in the
continuum description. The inclusion boundary is defined by a mesh of
vertices xbi (b for boundary). The exterior of the inclusion contains the
vertices xci (c for continuum). Inside the atomistic system, the (affine)
transformations obtained by altering the scaling matrix from h0 to h can be
expressed by the overall displacement gradient tensor matrix Mh hh
1
0 .
The Lagrange strain tensor [40] of the atomistic system is then
1
" MT M
1
2

12

As is convenient and customary in infinite periodic systems, scaled coordinates are employed for the continuum part as well as for the atomistic part
(see Fig. 8). In the continuum part, sbi H
1 xbi and sci H
1 xci of the nodal
points are used as degrees of freedom (the natural state variables of the
continuum are, thus, sbi , sci , and H). Similarly, the scaled coordinates sai of

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

504

Binder et al.

FIG. 7 The key idea of the atomistic-continuum model is to connect the inclusion
boundary behavior and the atomistic cell behavior via the strain transformation they
undergo. The strain behavior of the atomistic box is identical to the strain behavior
of the inclusion boundary.

atom i are related to the Cartesian coordinates through the scaling matrix,
sai h
1 xai .
Since, as stated above, our model [20] requires that an overall strain of
the atomistic box is accompanied by an identical strain at the boundary of
the inclusion, we constrain the vertices xbi of the inclusion boundary and the
atomistic scaling matrix h to behave in concert (the vertices xci in the matrix
are not affected) according to xbi h Mhxb0i , where xb0i are the original
coordinates of the boundary vertices i. Note that the coordinates xbi h are
no longer degrees of freedom, because they depend on the scaling matrix h.
Changes in the system cell matrix H act as homogeneous deformations of
the entire continuous system [20] and the nodal points of the continuum and
the atomic positions in the inclusion should be displaced accordingly. This

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Coarse-Graining Techniques

505

FIG. 8 Sketch of the four types of degrees of freedom used in the atomisticcontinuum model: H, DH, and the scaled coordinates sc and sa of the nodes in the
continuum and the atoms in the atomistic phase. The system unit cell matrix H and
the scaled coordinates sc and sb entirely determine the finite-element configuration of
the continuum. h and sa are necessary to determine the atomistic configuration. The
nodal points on the inclusion boundary sb and the atomistic scaling matrix h are
calculated from the variables H and DH.

can be achieved by introducing a new set of degrees of freedom DH (nine

scalars), through [20]
h H1 DH H
1
0 h0

13

Both H and h undergo concerted changes (a change in H from the

reference state H0 directly affects the atomistic cell h and affinely displaces
b
the inclusion boundary vertices according to xbi HH
1
0 x0i ). The reference
state, where H H0 and h h0, corresponds to DH 0.
The energy of the overall system is the sum of contributions from the
atomistic part and the continuum part. The atomistic part contributes a
potential energy determined by the interatomic potentials (the force field)
and a kinetic energy from the momenta of the atoms. The continuum system
contributes an elastic energy as the sum of the elastic deformations of all
finite elements:
E cont H, sbi , sci

1X
V " T C  "
2 

14

where  denotes a finite element and V is the volume of the undeformed

tetrahedron  [41]. The overall system can be described by the multiscale

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

506

Binder et al.

NtT ensemble distribution function [42].



V E atom
hE atom i0 K atom E cont K cont W ext
PNtT, multi / exp

kT
15
where E atom is the potential defined by the atomistic description of the
system, hE atom i0 its value at the undeformed (reference) state, K atom the
atomistic kinetic energy, E cont the continuum elastic strain energy of
Eq. (14), K cont the associated kinetic energy (with contributions from each
finite elementhere the mass of the continuum is distributed onto the nodes
of all finite elements and the kinetic energy is estimated as if it were atomistic), Wext the external work done on the system, and the parameter
V V inclusion/Vatom, the ratio of the inclusion and the atomistic cell
volume, accounts for the difference between the inclusion volume and the
atomistic system periodic cell contained therein.
The elements of the coupling matrix DH between the periodic boxes of
overall continuum system, H, and atomistic system, h [see Eq. (13)] are
simply minimization variables for the total energy when minimum-energy
simulations are carried out, i.e., calculations at T 0 [where the two kineticenergy terms in Eq. (15) are zero]. At nonzero temperature, when the
kinetic-energy terms in Eq. (15) must be taken into account, the coupling is
driven by the balance between the internal stress of the atomistic cell and the
external stress acting on the inclusion from the resistance of deformation of
the surrounding medium [42].
To date, results have been obtained for minimum-energy type simulations
of elastic deformations of a nearest-neighbor face-centered cubic (fcc)
crystal of argon [20] with different inclusion shapes (cubic, orthorhombic,
spherical, and biaxially ellipsoidal). On bisphenol-A-polycarbonate, elastic constant calculations were also performed [20] as finite deformation
simulations to plastic unit events (see [21]). The first molecular dynamics
results on a nearest-neighbor fcc crystal of argon have also become available
[42]. The consistency of the method with thermodynamics and statistical
mechanics has been tested to a satisfactory extent [20]; e.g., the calculations with different inclusion shapes all yield identical results; the results
are independent of the method employed to calculate the elastic properties
of the system and its constituents (constant-strain and constant-stress
simulations give practically identical values).
A test-case application to the elastic constants of bisphenol-A-polycarbonate shall give an idea of the manageable complexity of this multiscale
approach. Here, a single polymer was simulated as both the atomistic and the
continuum material in order to check for possible inconsistencies. The

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Coarse-Graining Techniques

507

polycarbonate was taken to be at the experimental density of 1.2 g/cm3 (300K).

The atomistic system considered was composed of three polycarbonate
chains of twenty-five repeat units each, i.e., of 2481 atoms in an atomistic cell
of edge length 300 A, and was constructed to conform to the overall torsionangle statistics of the rotational-isomeric state model of the polymer [43] with
the PolyPack algorithm [44]. The overall atomistic-continuum model was
chosen to have an atomistic inclusion shape and the inclusion volume
fraction was taken to be 18 percent (the atomistic box accounts for about one
fifth of the system energy, since the energy is distributed approximately
uniformly). The edge length of the overall system cell was taken to be 50 A so
that V 0.83 in Eq. (15). The atomistic-continuum system comprised a total
of 26,235 scalar degrees of freedom (3  827 atoms and 6259 independent
nodal points in the continuum, which give 26,220 coordinates, and the two
scaling matrices, H with 6 and DH with 9 degrees of freedom). This is easily
within the range of modern molecular simulations.

IV. CONCLUSIONS
The recently proposed atomistic-continuum model is useful for modeling the
deformation of three-dimensional solids. In this approach, an inclusion of
arbitrary shape is surrounded by a continuum, which is modeled by a
variable-metric finite-element method. The inclusion behavior is directly
driven by that of an atomistic system modeled in detail by molecularmodeling techniques. In this manner, a solution is given to the problem of
finding a constitutive description for a material. The atomistic-continuum
model comprises four types of degrees of freedom, the appropriate combination of which permits calculation of the configuration of both the
atomistic system and the mesh elements of the continuum. The behavior of
both levels of modeling is described by their corresponding potential C- and
kinetic-energy functions which are added to give the total energy of the
system.
The major drawback of the atomistic-continuum approach lies in the
limitation to uniform strain on the level of the atomistic periodic box and,
hence, the inclusion perimeter. Crack propagation, for instance, cannot be
treated with this method.
A partial compensation for this weakness might lie in the flexibility of
the atomistic-continuum concept. The approach can be extended, in
principle, to
.

any number of atomistic inclusions in a continuum phase; also,

one can imagine continuous atomistic phases cohabiting with the
continuum;

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

508

Binder et al.

other overall properties than stress and strain; among the properties that can be simulated are those describable by scalar fields in a
continuum [39], such as dielectric constants, thermal conductivity,
electrical conductivity, solute permeability, etc.

Also, the approach provides the potential for constant-stress molecular

dynamics [40,45] with realistic retracting forces acting on the atomistic
periodic box. This would provide an advantage over the current situation,
where constant-stress molecular dynamics is weakened by the fact that the
device, which maintains constant stress, (a) acts directly and without time
delay on each atom, and (b) requires an arbitrary wall mass. The question
of efficiency for atomistic-continuum constant-stress molecular dynamics
must be answered first, of course.

ACKNOWLEDGMENTS
We thank Jorg Baschnagel and Andrei A. Gusev for many illuminating
discussions and Michael Rozmann and Volker Tries for their help and
kind support. S. S. and U. W. S. also gratefully acknowledge the financial support provided by the Swiss National Science Foundation
(Schweizerischer Nationalfonds zur Forderung der wissenschaftlichen
Forschung). K. B. and W. P. acknowledge support from the Bundesministerium fur Bildung und Forschung (BMBF, Germany) under Grants
No 03N8008C and 03N6015.

REFERENCES
1. Binder, K. Macromol. Chem.: Macromol. Symp. 1991, 50, 1.
2. Baschnagel, J.; Binder, K.; Doruker, P.; Gusev, A.A.; Hahn, O.; Kremer, K.;
Mattice, W.L.; Muller-Plathe, F.; Murat, M.; Paul, W.; Santos, S.; Suter, U.W.;
Tries, V. Adv. Polymer Sci. 2000, 152, 41.
3. Kremer, K. Multiscale problems in polymer simulations. In Multiscale
Simulations in Chemistry and Physics, Proceedings NATO ARW, Eilat/Israel,
April 58, 2000, Brandt, A., Bernholc, J., Binder, K., Eds.; IOS Press:
Amsterdam, 2001.
4. Baschnagel, J.; Binder, K.; Paul, W.; Laso, L.; Suter, U.W.; Batoulis, J.; Jilge,
W.; Burger, T. J. Chem. Phys. 1991, 95, 6014.
5. Paul, W.; Binder, K.; Kremer, K.; Heermann, D.W. Macromolecules 1991, 24,
6332.
6. Carmesin, I.; Kremer, K. Macromolecules 1988, 21, 2819.
7. Deutsch, H.-P.; Binder, K. J. Chem. Phys. 1991, 94, 2294.
8. Paul, W.; Pistoor, N. Macromolecules 1994, 27, 1249.
9. Tries, V.; Paul, W.; Baschnagel, J.; Binder, K. J. Chem. Phys.1997, 106, 738.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Coarse-Graining Techniques

509

10. Rapold, R.F.; Mattice, W.L. Macromolecules 1996, 29, 2457.

11. Doruker, P.; Mattice, W.L. Macromolecules 1997, 30, 5520.
12. Tschop, W.; Kremer, K.; Batoulis, J.; Burger, T.; Hahn, O. Acta Polymerica
1998, 49, 61.
13. Tschop, W.; Kremer, K.; Batoulis, J.; Burger, T.; Hahn, O. Acta Polymerica
1998, 49, 75.
14. Zimmer, K.M.; Heermann, D.W. J. Computer-Aided Mater. Design 1995, 2.
15. Zimmer, K.M.; Linke, A.; Heermann, D.W.; Batoulis, J.; Burger, T.
Macromol. Theory Simul. 1996, 5, 1065.
16. Santos, J. From Atomistic to Continuum Modeling of Polymers-Multiscales
Paradigms. Ph.D. thesis, Eidgenossische Technische Hochschule Zurich, Dies
No 12941, 1998.
17. Murat, M.; Kremer, K. J. Chem. Phys. 1998, 108, 4340.
18. Louis, A.A.; Bolhuis, P.G.; Hansen, J.P.; Meijer, E.J. Phys. Rev. Lett. 2000, 85,
2522.
19. Warren, P.B. Current Opinion in Colloid & Interface Science 1998, 3, 620.
20. Santos, S.; Yang, J.S.; Gusev, A.A.; Jo, W.H.; Suter, U.W. submitted for
publication.
21. Yang, J.S.; Jo, W.H.; Santos, S.; Suter, U.W. Plastic deformation of bisphenolA-polycarbonate. In Simulation Methods for Polymers, Kotelyanskii, M.,
Theodorou, D., Eds.; Marcel Dekker, 2004, Chapter 12.
22. Smith, G.D.; Yoon, D.Y. J. Chem. Phys. 1994, 100, 649.
23. Paul, W.; Yoon, D.Y.; Smith, G.D. J. Chem. Phys. 1995, 103, 1702.
24. Paul, W.; Smith, G.D.; Yoon, D.Y.; Farago, B.; Rathgeber, S.; Zirkel, A.;
Willner, L.; Richter, D. Phys. Rev. Lett. 1998, 80, 2346.
25. Baschnagel, J.; Qin, K.; Paul, W.; Binder, K. Macromolecules 1992, 25, 3117.
26. Binder, K.; Paul, W. J. Polym Sci., Part B: Polymer Phys. 1997, 35, 1.
27. Pearson, D.S.; Fetters, L.J.; Graessley, W.W.; ver Strate, G.; von Meerwall, E.
Macromolecules 1994, 27, 711.
28. Muller-Plathe, F.; Biermann, O.; Faller, R.; Meyer, H.; Reith, D.; Schmitz, H.
Autimatic multiscaling, Paper presented at the ESF-SIMU-Workshop
Multiscale Modeling of Macromolecular Systems, Mainz, Sept. 46, 2000,
unpubl.
29. Phillip, R. Journal of The Minerals Metals & Materials Society 1995, 47, 37.
30. Vellinga, W.P.; Hosson, J.T.M.D.; Vitek, V. Acta Mater. 1997, 45, 1525.
31. Ortiz, M. Computational Mechanics 1996, 18, 321.
32. Tadmor, E.B.; Phillips, R.; Ortiz, M. Langmuir 1996, 12, 4529.
33. Tadmor, E.B.; Ortiz, M.; Phillips, R. Philosophical Magazine A 1996, 73, 1529.
34. Shenoy, V.B.; Miller, R.; Tadmor, E.B.; Phillips, R.; Ortiz, M. Phys. Rev.
Letters 1998, 80, 742.
35. Rai-Tabar, H.; Hua, L.; Cross, M. J. Phys.: Condens. Matter 1998, 10, 2375.
36. Miller, R.; Phillips, R. Philosophical Magazine A 1996, 73, 803.
37. Allen, M.P.; Tildesley, D.J. Computer Simulation of Liquids; Clarendon Press:
Oxford, 1989.
38. Gusev, A.A. J. Mech. Phys. Solids 1997, 45, 1449.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

510

Binder et al.

39. Gusev, A.A. Macromolecules 2001, 34, 3081.

40. Parrinello, M.; Rahman, A. Journal of Chemical Physics. 1983, 76, 2662.
41. Thurston, R.N. In Physical Acoustics: Principles and Methods, Mason, W.P.,
Ed.; Academic Press: New York, 1964.
42. Beers, K.J.; Ozisik, R.; Suter, U.W. in preparation.
43. Hutnik, M.; Argon, A.S.; Suter, U.W. Macromolecules. 1991, 24, 5956.
44. Muller, M.; Santos, S.; Nievergelt, J.; Suter, U.W. J. Chemical Physics 2001,
114, 9764.
45. Parrinello, M.; Rahman, A. Journal of Applied Physics 1981, 52, 7182.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

16
CONNFFESSIT: Simulating
Polymer Flow
MANUEL LASO

Polytechnic University of Madrid, Madrid, Spain

TTINGER Eidgenossische Technische

HANS CHRISTIAN O
Hochschule (ETH), Zurich, Switzerland

I.

INTRODUCTION

The present chapter introduces a nontraditional approach to the numerical

calculation of complex flows of complex fluids, with especial application to
polymeric systems such as melts and solutions.
The applicability of the CONNFFESSIT (Calculation Of NonNewtonian Flows: Finite Elements and Stochastic Simulation Technique)
in its present form is limited to the solution of fluid mechanical problems of
incompressible fluids under isothermal conditions. The method is based on a
combination of traditional continuum-mechanical schemes for the integration of the mass- and momentum-conservation equations and a simulational
approach to the constitutive equation.
The macroscopic equations are solved by a time-marching scheme in
which the stress tensor is obtained from a micro simulation and treated as a
constant body force. The microscopic simulations reflect the dynamics of
the specific fluid at hand and yield the stress tensor for a given velocity field.
The required integration over deformation histories is accomplished by
integrating numerically microscopic particle trajectories for large global
ensembles simultaneously with the macroscopic equations of mass and
momentum conservation. The term trajectories in the previous sentence
refers to both real space trajectories, i.e., positions ri(t) and to configurational phase space trajectories, i.e., in the case of a dumbbell model,
connector vector Q.
Given the very large ensemble sizes required, the feasibility of such
calculations depends crucially on algorithmically optimal data structures

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

 ttinger
Laso and O

512

and computational schemes. In the present chapter a major effort has

been devoted to the detailed description of such techniques and of their
implementation on advanced computer architectures. Nontrivial twodimensional calculations for a simple noninteracting polymer dumbbell model are presented in order to demonstrate the practical use of the
method.
Although CONNFFESSIT opened up the door to so-called micromacro
rheological calculations some ten years ago (Ottinger and Laso 1992),
it is but the first member of a group of methods that have appeared along
the same line since 1992 in the context of computational non-Newtonian
fluid mechanics. Most of the newer methods rely on the concept of variance
reduction (Kloeden and Platen 1992, Ottinger 1995, Melchior and Ottinger
1995, 1996) in order to achieve a great improvement on the noise level of
micromacro calculations. Although the field of micromacro methods
(Laso 1995) is still very much in flux, some of the newer and highly original
methods (see Hulsen et al. 1997, Halin et al. 1998, Bonvin and Picasso 1998,
Bell et al. 1998) can be up to two or three orders of magnitude more efficient
for a given level of precision. And improvements are appearing all the
time. Nowadays the main applicability of CONNFFESSIT as formulated originally is in situations where fluctuations do play an important
role. For typical macroscopic problems, the newer techniques are far more
efficient and thus the methods of choice.

A. Some Definitions
Throughout the succeeding sections some terms are used frequently,
sometimes in a nonstandard way. The following definitions are included
in order to avoid confusion:
.

.
.
.
.

Domain: a region of d-dimensional space in which we seek the solution to the equations of conservation of mass and momentum coupled
to a description of the fluid (constitutive equation).
Cell: in d dimensions, the simplest d-dimensional geometric object in
which the domain is subdivided.
Element: synonymous with cell.
Finite element: a cell or element augmented with nodal points and
shape functions.
Particle: the smallest entity of the micro calculation that does not
directly interact with other similar entities and which contributes to
the stress; it can be as simple as a Hookean dumbbell, or as complex
as a whole Brownian Dynamics simulation of a colloidal dispersion
under periodic continuation or boundary conditions.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

.
.
.

513

Global ensemble: the set of all particles in the domain.

Local ensemble: the set of all particles in a cell.
Complex fluid: a (possibly non-Newtonian) fluid in which microscopic
particles are explicitly resolved.

Finally, throughout this work, and unless stated otherwise, the term
microscopic refers either i) to simplified mechanical models for polymer
molecules in which no chemical detail is explicitly present or ii) to models for
particulate fluids in which the individual particles are explicitly treated and
in which the solvent is treated stochastically in an average sense and not
resolved in individual molecules. This usage of the term microscopic is
adopted here in order to be consistent with the meaning of the term micro
as used in the literature on methods that combine continuum mechanics
with a more detailed description of the flowing material. Strictly speaking,
models falling into category i) are more microscopic than those under ii).
Atomistic and ab initio (density functional) polymer modeling reside at
increasingly more fundamental levels.

II. OVERVIEW OF RELATED FIELDS

CONNFFESSIT cuts across two different and, up to recently, largely
unrelated fields: traditional continuum-mechanical computational rheology
and stochastic dynamic methods for polymers. Brief reviews are presented
separately. A third subsection is devoted to alternative methods in
Newtonian and non-Newtonian CFD which have a direct bearing on the
subject of this work.

A. Computational Rheology
Although it is young, the field of continuum-mechanical computational
rheology has a lively and eventful history. What was initially expected to be
little more than an extension of Newtonian CFD methods to fluids with
more complex but in principle harmless constitutive equations turned out to
be a source of frustration and of unexpectedly challenging mathematical
problems in spite of the vast store of knowledge on numerical techniques
for Newtonian fluids that had been accumulating during the past
half century. The analysis and solution of what has come to be known
as the high Weissenberg-number* problem have engaged the attention of
*Rheologists resort to dimensionless groups such as the Weissenburg and Deborah numbers
and the stress ratio SR to characterize the elasticity of the ow. High WS corresponds to high
uid elasticity.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

 ttinger
Laso and O

514

a large number of both applied mathematicians and engineers and spurred a great amount of analytical and numerical diagnostic work. The
book by Crochet et al. (1984) and the review articles by Crochet (1989)
and Keunings (1989) give comprehensive if slightly dated overviews of
the field; current research in computational rheology is reported in
Moldenaers and Keunings (1992), Gallegos et al. (1994), and Ait-Kadi et
al. (1996).
For the purposes of the present work the current situation in nonNewtonian CFD can be summarized as follows:
.

Converged solutions for non-Newtonian CFD problems can be

obtained nowadays for a large variety of geometries and constitutive
equations (Debbaut and Crochet 1986, Keunings 1986a, 1986b, Van
Schaftingen and Crochet 1984, Mendelson et al. 1982, Brown et al.
1986, Marchal and Crochet 1987, Papanastasiou et al. 1987, Bush
et al. 1984, Malkus and Webster 1987), including free-surface flows
(Nickell et al. 1974, Keunings and Bousfield 1987).
Most calculations are performed in two dimensions, under isothermal
conditions and at steady state, although transient (Josse and
Finlayson 1984, Lee et al. 1984, Van Schaftingen 1986, Phan-Thien
et al. 1987, Bousfield et al. 1988) and nonisothermal (Keunings 1986a,
1986b, Sugeng et al. 1987, McClelland and Finlayson 1988,
Srinivasan and Finlayson 1988) flow calculations have been reported.
The high Weissenberg problem has been recognized to actually have
several independent causes such as irregular points in the solution
family, boundary layers and singularities, loss of evolution, and
change of type of the PDEs. Although some of the problems originally encountered still remain, a good deal of these initial difficulties
have been overcome by sophisticated numerical schemes (Keunings
1989 and references therein, Franca and Hughes 1993, Baaijens 1998).

Finally, there exists a trend away from physically simple but

numerically problematic constitutive relations towards more realistic and
algorithmically more demanding ones with better numerical behavior.
Some CEs, prominently those most appropriate for melts (DoiEdwards,
CurtissBird, reptating rope, modified reptation models) are and will
remain for a long time to come well beyond the capabilities of current
continuum-mechanical methods. Further improvements in the polymer
dynamics at the microscopic level leads to models intractable via the
standard CE approach.
The systematic use of increasingly more complex CEs calls for
an alternative way of handling this increased complexity while at the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

515

same time allowing the equations of change to be solved in complex

geometries.

B. Stochastic Dynamic Methods for Polymers

Stochastic dynamics or stochastic differential equations arise naturally in
polymer physics as a consequence of the coarse-graining required when
dealing with the many degrees of freedom and the wide spectrum of time
scales characteristic of polymers. Brownian Dynamics was first applied in
the late 1970s (Fixman 1978, McCammon 1978) and has experienced a
colossal growth since then. Although there is abundant literature on
stochastic and kinetic-theoretical topics (Van Kampen 1981, Gardiner 1983,
Gard 1988, Risken 1989, Honerkamp 1990, Kloeden and Platten 1992, Doi
and Edwards 1986, Bird et al. 1987b), a unified approach to the use of
stochastic methods in polymer dynamics has been missing until recently.
Besides filling this gap, the book by Ottinger (1995) offers a comprehensive
overview of the state of current research.
As far as the present work is concerned, the relevance of numerical
stochastic methods for polymer dynamics in micro/macro calculations
resides in their ability to yield (within error bars) exact numerical solutions
to dynamic models which are insoluble in the framework of polymer kinetic
theory. In addition, and mainly as a consequence of the correspondence
between FokkerPlanck and stochastic differential equations, complex
polymer dynamics can be mapped onto extremely efficient computational
schemes. Another reason for the efficiency of stochastic dynamic models for
polymer melts stems from the reduction of a many-chain problem to
a single-chain or two-chain representation, i.e., to linear computational
complexity in the number of particles. This circumstance permits the
treatment of global ensembles consisting of several tens of millions of
particles on current hardware, corresponding to local ensemble sizes of
O(103) particles per element.

C. Particle Methods
Besides the Finite-Difference and Finite-Element mainstream, there has
appeared a handful of alternative computational methods in Newtonian and
non-Newtonian CFD that can be collectively characterized as particulate,
in the sense that they involve discrete computational entities in addition
to or in lieu of the standard continuum-mechanical discretization devices.
In some of these methods (the first four and the last one in the following

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

 ttinger
Laso and O

516

list), the computational particles have an almost exclusively mathematical

character, whereas in the other two it is possible to attach a well-defined
physical meaning to the particles.
.

As early as 1973 Chorin (1973, 1989, 1994) introduced the twodimensional random vortex method, a particle method for the solution of the NavierStokes equations. These particles can be thought
of as carriers of vorticity. Weak solutions to the conservation equations are obtained as superpositions of point vertices, the evolution of
which is described by deterministic ODEs. A random walk technique
is used to approximate diffusion, and vorticity creation at boundaries
to represent the no-slip boundary condition. The extension to three
dimensions followed in 1982 (Beale and Majda 1982). An important
improvement in stability and smoothness was achieved by Anderson
and Greengard (1985) by removing the singularities associated with
point vertices. Anderson and Greengard (1988) and Marchioro and
Pulvirenti (1984) have written comprehensive reviews of the method.
Smoothed-Particle Hydrodynamics (SPH) was introduced by Gingold
and Monaghan (1977) and Lucy (1977). SPH is a three-dimensional
free-Lagrange algorithm that in its most basic form is truly grid-free.
SPH treats fluid elements as extended clouds of material; their centers
of mass move according to the conservation laws of fluid mechanics.
It was first applied to the solution of complex astrophysical problems,
such as colliding planets and stars. In its most widely used version,
SPH is based on integral interpolants for information at nonordered
points. The fluid is represented by spherically symmetric particles,
each having a mass that has been smoothed out in space with a
density distribution given by a suitable interpolating kernel or distribution. Ordinary differential equations for the smoothed quantities at
the location of the particles are obtained by multiplying the original
governing equations by the kernel and integrating over the domain. A
critical evaluation of the quantitative performance of SPH has been
presented in Cloutman (1991). The method seems to be most useful
for exploratory studies where modest accuracy is acceptable.
Another technique, widely used in a different context (kinetic theory,
plasma simulation) is the so-called Vortex-in-Cell Method (Hockney
and Eastwood 1981, Birdsall and Langdon 1985). In the Vortex-inCell Method a two-dimensional computational domain is divided into
cells and the vorticity is counted in each cell. The Poisson equation
relating the vorticity and the Laplacian of the stream function is
subsequently solved on the grid by an FTT algorithm, the velocity

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

517

is computed on the grid and finally interpolated at the vortex

locations. This approach bears some resemblance to SPH.
Cellular-automata and lattice-gas models are in a sense a diametrically
opposed alternative to the pure grid-free NEMD method. They rely on
the existence of a regular lattice, each site of which can have a finite
number of states usually represented by Boolean variables. Each cellular automaton evolves in time in discrete steps according to deterministic or nondeterministic rules for updating based on the states of
the neighboring automata. The average behavior of a collection of
cellular automata is found to describe the behavior of physical systems
obtained by using models with smooth continuous variables. Latticegas models with fluid dynamical features were first introduced by
Kadanoff and Swift (1968) and the first fully deterministic lattice-gas
model was introduced by Hardy et al. (1976). The main value of cellular-automata methods seems to be in the qualitative study of turbulence (Frisch et al. 1986), although they have been used to investigate
other fluid-flow phenomena with varying degrees of success (Rothman
and Keller 1988, Balasubramanian et al. 1987, Baudet et al. 1989).
A grid-free alternative to discretization methods is offered by NonEquilibrium Molecular Dynamics. The NEMD approach to fluid
dynamics was pioneered by Ashurst and Hoover in 1975 for the
investigation of momentum transport in homogeneous flows. Some
of the first calculations for nonhomogeneous flows in complex geometries were performed by Hannon, Lie, and Clementi (1986). A large
number of MD simulations has been performed since then. Most of
the very satisfactory features of such calculations arise from the fundamental description of the interactions between fluid particles, which
in this case are the atoms or molecules, and the absence of a grid. The
drawback of this approach is that it is limited to extremely small
length- and time-scales, or conversely, to extremely high rates of
deformation, O(1013) s
1 typically. The same approach has been
applied by Hoover et al. (1992) to plastic flow in nonlinear solid
mechanics.

III. TWO- AND THREE-DIMENSIONAL TECHNIQUES

A. One-Dimensional vs. Multidimensional Problems
The first application of CONNFFESSIT was in the context of computational non-Newtonian fluid mechanics in 1 1 dimensions (one spatial
one temporal) (Laso and Ottinger 1993). The practical applicability of the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

518

ttinger
Laso and O

method relies on its success in more dimensions (2 1, 3 1). The

introduction of one additional spatial dimension entails both a quantitative
and a qualitative increase in the complexity of the problem.
The obvious quantitative increase is due to the higher number of particles
and of spatial cells or elements required to describe the integration domain,
typically between one and two orders of magnitude more per additional
spatial dimension.
The qualitative increase stems from the need to perform particle tracking,
a task which has no equivalent in one dimension,* since particles never leave
the element in which they are initially located as they are carried by the
velocity field. The generic term particle tracking actually involves a number
of nontrivial individual subtasks, which must be tackled efficiently if
CONNFFESSIT is to be viable. The key to practicable micro/macro
computations is the invention of efficient schemes of reduced algorithmic
complexity.
A further increase to three spatial dimensions brings about another
increase in the size of the problem. The tracking techniques are, however,
not more complex than in the two-dimensional case. Furthermore, some of
the algorithms to be presented perform proportionally better than classical
continuum-mechanical tools the larger a problem is. For this reason, the
two-dimensional case is especially critical: it involves particle tracking in its
full complexity, yet typical 2D problem sizes are not yet in the range where
the CONNFFESSIT approach can even remotely compete with traditional
methods.
Before presenting procedural details, an overview of the general scheme
for the micro/macro approach in more than one dimension is in order.
Although it is possible to devise micro/macro methods to obtain steadystate solutions by a procedure that draws heavily from traditional tracking
schemes for integral constitutive equations (Feigl et al. 1995), the natural
implementation of the micro/macro strategy calls for a time-marching
integration of the unsteady problem. This approach allows us to obtain, in
addition to the full dynamic behavior of the system, the steady-state solution
(when one exists) by continuing integration for a sufficiently long time.
The usage of the term micro/macro is justified by the two-level
description of the fluid that lies at its core: a microscopic approach permits
evaluation of the stress tensor to be used in closing the macroscopic
conservation equations. The macro part can be thought of as corresponding
to the Eulerian description of the fluid, while the micro part has Lagrangian
character.
*Except for some multiparticle dynamic models in which particle diusion across streamlines
must be considered (Ottinger 1992).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

519

The overall CONNFFESSIT procedure for the isothermal problem is

eminently simple and is illustrated in Figs. 1 and 2 as a block diagram. This
approach allows the use of standard continuum-mechanical codes to solve
the macroscopic equations of conservation (mass and momentum).
Micro/macro interfacing takes place at the points marked A and B in Fig. 1.

FIG. 1

Basic CONNFFESSIT time-marching scheme.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

 ttinger
Laso and O

520

FIG. 2

Basic CONNFFESSIT time-marching scheme (continued).

At a given step in the time-marching scheme, the currently available velocity

field is passed to the microscopic part at A where it is used to advance both
the internal configurations of the particles and their trajectories. In B, the
stress tensor obtained by a suitable ensemble-averaging (e.g., over individual
cells) is used as a right hand side (body force) in the equation of conservation of momentum, which is solved together with the mass-conservation
equation using standard techniques (Keunings 1989, Crochet et al. 1984).

B. The Basic Data Structure

Independently of the details of the micro simulations, the information
pertaining to individual particles of the global ensemble (or at least pointers
to that information) must be kept in a suitable kind of linear array. As we will
see in Sections IV.A and IV.B.1 the organization of this indexing array into
an element-wise sorted array has far-reaching consequences on computational efficiency and, above all, on vectorization and parallelization.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

521

Several micro/macro models, CONNFFESSIT among them, share the

common feature that the particles appearing in the micro simulation are
entrained by the macroscopic flow and therefore move through a succession of elements. As a consequence, the indices of the particles residing in a
given element (the local ensemble) change with time. This circumstance
makes physically contiguous particles be noncontiguous in the linear array
containing them.
The objective of constructing an element-wise sorted array is to provide
indirect indexing capability, which can be used to bypass the noncontiguity
problem. Let the array WHERE initially contain the number of the element
in which a given particle is located. A small-scale typical arrangement
(12 particles in 3 elements) is given in the schematic below where particles
are noncontiguous:

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

522

ttinger
Laso and O

For several reasons to be seen later, it is more convenient to have

physically neighboring particles (i.e., particles in the same cell or element)
placed in neighboring positions in memory, i.e., particles 1, 3, 7, 10, and 12,
etc. While it would be possible to initialize the positions of particles so as
to ensure contiguity at the start of a calculation, the macroscopic flow
would render this initial order useless at every time step. It is however
possible to create an element-wise sorted array ORDWHERE that holds
the indices of the particles contained in each element (of which there
are NUMEL), plus two pointer arrays FIRST and LAST that mark the
start and the end of the sequence of particle indices for each element within
the sorted array. Sorting can be accomplished by any efficient algorithm,
like heapsort (Knuth 1968) or, having vectorizability in mind, by the
algorithm presented in Section IV.B.1. The following pseudo-code excerpt
shows how this can be accomplished (sort array WHERE and copy onto
ORDWHERE):
J0
DO I 1 , NUMPR - 1
IF (ORDWHERE(I) .NE. ORDWHERE(I1)) THEN
JJ1
COUNT(J) I
ENDIF
ENDDO
JJ1
COUNT(J) NUMPR
DO I 1 , NUMEL 1
LAST(I) 0
ENDDO
DO I 1, J
LAST(ORDWHERE(COUNT(I))) COUNT(I)
ENDDO

Special care of empty elements (if any) must be taken and the array
FIRST created:
DO I 2, NUMEL 1
IF (LAST(I) .EQ. 0) LAST(I) LAST(I-1)
ENDDO
COUNT(1) LAST(1)
DO I 2, NUMEL 1
COUNT(I) LAST(I) - LAST(I-1)
ENDDO
DO I 1, NUMEL 1
FIRST(I) LAST(I) 1 - COUNT(I)
ENDDO

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

523

This compact algorithm performs the required task and is fully vectorizable. The indexing array and the indirect addressing capability offered
by the information contained in the FIRST and LAST pointer arrays (above)
act de facto as an interface that makes the particles appear as if they actually
were contiguous:

The crucial point of this procedure is that it relies on efficient sorting for
restoring contiguity. As we will see in Section IV.B.1, particle index sorting
using the binary radix sort has optimal O(Npr logNpr) complexity. Besides,
indirect addressing is fully supported (and in many cases hardwareexecuted) by all vector computers. Thus, the computational overhead

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

 ttinger
Laso and O

524

associated with reordering the particles at each time step is perfectly

acceptable.

C. Point-Inclusion Algorithm
The second key operation lying at the heart of a CONNFFESSIT calculation consists in determining if a particle, the spatial coordinates of
which are known, is located within or outside an element. This seemingly
minor task acquires immense importance in more than one dimension,
since it has to be executed an inordinate number of times: it is almost
always the single most-frequently performed basic operation invoked
in CONNFFESSIT.
Generalized point-inclusion algorithms are the object of much research
in computational geometry (Mulmuley 1994, Preparata and Shamos
1990); the goal of research in this field is the creation of optimal
algorithms with respect to the number of sides of the polygon, i.e., how to
optimally determine whether or not a point is included in a polygon of a
very large number of sides. An extensive search of available literature
showed that, somewhat surprisingly, the question of how to determine
whether a point is included in a triangle or quadrilateral in the fastest
possible way is amply ignored. The only algorithm found in the literature
turned out to be a geometrically elegant one frequently used in computer
graphics and based on the evaluation of a vector product for each of
the sides but which can certainly not compete in speed with the following
procedure (Fig. 3).
During the initialization phase, before the start of the actual timemarching computation:
.
.

For each element k, determine the coordinates of some arbitrary

interior point, for example, its centroid (Xck ,Yck ).
For each side of the element, determine and store the coefficients akj
i of
the equations representing the sides fjk akj
i , x, y, where j 1, Nsides,
which are in most cases straight lines or polynomials of low degree.
Evaluate sign fjk aikj , Xcj , Xcj and store it for each of the sides of each
element.

For subsequent use, i.e., in order to determine if the lth particle at (xl, yl)
is located within the kth element:
.

k kj
Evaluate sign fjk akj
i , xl , yl ; if and only if sign fj ai , xl , yl
k kj
j
j
and sign fj ai , Xc , Xc are equal for all j, is the particle within the
element.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

525

FIG. 3 Particle-inclusion algorithm in two dimensions. The same scheme is applicable to cells with higher number of edges or in higher dimensions.

This sequence of steps corresponds to determining whether the particle in

question lies on the same side (left or right) of the sides of the element as a
known interior point (like the centroid).
Once the coefficients of the equations defining the sides are known
(the computation of which amounts to a very minor overhead during
initialization), point-inclusion can be determined with at most Nsides
function evaluations and the same number of logical comparisons. In
particular, an inclusion test for a particle in a triangular element involves at
most* three multiplications, three additions (or arithmetic comparisons),
and three Boolean comparisons.
On some vector computers like the NEC SX-3 and SX-4 that execute
several overlapping vector instructions in parallel, these operations are
performed concurrently by multiple functional arithmetic units within a
vectorized loop over element number. The algorithm just presented achieves
on such hardware for the point-inclusion test on triangular elements the
ideal maximum speed of one point-inclusion test per clock cycle.
While this algorithm can conceivably fail in the case of an isoparametric
element for which one or more of its sides has extreme curvature, in practice,
the probability of such an occurrence is vanishing.

*Since the particle fails the inclusion test as soon as a dierence in sign is detected.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

526

ttinger
Laso and O

D. Scalar Velocity-Biased Ordered Neighbor Lists

In the micro/macro method discussed in this work, the macroscopic
conservation equations are solved by discretization (typically weighted
residuals) methods. The first step in such an approach is the generation
of a mesh or grid made up of simple geometric objects like triangles
or quadrilaterals in 2D, tetrahedra or parallelepipeds in 3D. The topic of
mesh generation for FE is itself an area of intensive research (Wheatherill
et al. 1994, Knupp and Steinberg 1993, Castillo 1991). Nowadays there are
many efficient mesh generators which are able to subdivide or enmesh
the integration domain. This mesh is represented by the set of coordinates
of the nodal points (loosely speaking the vertices, or other important
points of the simple geometrical objects) and information on connectivity and on the occurrence of boundaries. This information is then
employed in the WR solution of the macroscopic equation of motion.
In micro/macro methods the grid must be additionally used as the stage
in which the individual particles evolve and through which they move.
As a consequence, in contrast to continuum-mechanical methods, in a
micro/macro calculation it is necessary to keep track of the position of
each particle, both as coordinates and as the index of the element (or
the indirect addressing pointer to the element) in which it is located
(Section IV).
Particle neighbor lists are a device widely used in MD and MC
simulations of complex fluids to speed up the calculation of interactions
(Allen and Tildesley 1987). They contain, for each particle in a simulation, a
list of those other particles that fall within a cutoff distance from the particle
under consideration (possibly under periodic continuation conditions) and
which have to be taken into account in the computation of interactions.
Since the number of neighbors depends exclusively on the value of the cutoff
and not on the total number of particles, the complexity of the calculation of
2
to O(Npr).
pairwise interactions is reduced from ONpr
The meaning of a neighbor list is different in the present context: a cell
neighbor list contains, for each element or cell in the mesh, the indices of the
elements that share at least one vertex (first neighbors) with the cell under
consideration. Second neighbors of a given cell share at least one vertex with
its first neighbors. Since the typical particle displacement during a time step
is much smaller than the size of a cell, particles that have left a cell after a
time step are to be found in most cases in a first neighbor cell or, at most, a
second neighbor. It is not necessary to check all elements in the mesh. Just
as in the case of the neighbor lists used in MD, this data structure allows
a reduction of the complexity of the localization of particles from
O(Npr)O(Nel) to O(Nel).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

527

In micro/macro methods, the grid generator is more or less a black box;

normally the generator comes from another application and in principle no
information is available about its inner workings, nor is it possible to
control in any way the numbering scheme and the resulting connectivity of
the elements. Any algorithm to be developed for particle tracking must
therefore be valid and efficient for arbitrarily generated and numbered grids.
For the sake of discussion we will assume a structured two-dimensional grid
made up of convex quadrilaterals; none of these assumptions limits the
applicability of the algorithms to be developed in this section and in Section
IV. The quadrilaterals are of different sizes and shapes in order to adapt
to the geometry and resolve the details of the domain and the velocity
and stress fields. The grid is defined as two lists of X and Y coordinates
specifying the positions of the vertex points, and cells or elements in the grid
are specified by an integer array of the form VERTEX(NUMEL,4), giving
the numbers or indices of the four vertices of each cell. Actual cell and vertex
numbers are more or less random, in the sense of being unpredictable. The
numbering is determined by the internal logic of the grid generator, with cell
and vertex numbers assigned as needed, and no guarantee that cell i has cells
i 1 and i
1 as two of its nearest neighbors.
Since the characteristic size (length) of a cell or element is very much
larger than the displacement of a typical particle in a single time step, the
particles that leave an element at a given time step are almost always to be
found either in nearest or at most in second-nearest neighbors. The task of
locating all particles after every time step is thus greatly simplified if
neighbor lists are generated. These lists of first and second neighbors can be
generated in a very compact way using the information stored in the array
VERTEX(NUMEL, 4): two cells are nearest neighbors if and only if they
have one or more vertices in common. The most straightforward algorithm
for a static mesh* compares for each cell, each of the four vertices with each
of the four vertices for all the other cells. The following code implements this
idea in a compact way: it determines the nearest neighbors for all NUMEL
cells, the number of neighbors of each cell NOFNEI1 (which in general will
vary depending on whether the cell is at a boundary or not, or for internal
cells in unstructured meshes), and creates a list of these neighbors NEI1 in a
single pass. The following implementation is such that most vector
compilers, including the finicky f77sx on the NEC-SX series, can handle
the innermost four IFs and two DO loops efficiently by unrolling, thus
yielding a count for the vectorized DO loop equal to the number of cells in

*A static mesh remains the same throughout the calculation, both the coordinates of the nodal
points and the numbering of the cells.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

528

ttinger
Laso and O

the mesh, typically an order of magnitude larger than the hardware vector
length:
DO I 1, NUMEL
J1
DO ICORNER 1, 4
DO II 1, NUMEL
IF (II .NE. I) THEN
DO IICORNER 1, 4
IF (VERTEX(II, IICORNER) .EQ. VERTEX(I, ICORNER)) THEN
REPEAT .FALSE.
DO ICHECK 1, J
IF (II .EQ. NEI1(ICHECK, I)) THEN
REPEAT .TRUE.
ENDIF
ENDDO
IF (.NOT. REPEAT) THEN
NOFNEI1(I) NOFNEI1(I) 1
NEI1(NOFNEI1(I), I) II
ENDIF
J MAX(NOFNEI1(I), 1)
ENDIF
ENDDO
ENDIF
ENDDO
ENDDO
ENDDO

Although FORTRAN 77 does not contain recursive constructs, it is

possible to emulate them via a call-stack (Kieling and Lowes 1985).
Recursive application of the above procedure would then allow us to find
second neighbors without additional coding. However, it is worth spending
some additional effort in developing a slightly more complex procedure to
determine second neighbors NEI2 and their number NOFNEI2 in a
nonrecursive way:
DO I 1, NUMEL
J1
DO INEI 1, NOFNEI1(I)
ITEST NEI1(INEI, I)
DO ICORNER 1, 4
DO II 1, NUMEL
IF ((II .NE. ITEST) .AND. (II .NE. I)) THEN
REPEAT2 .FALSE.
DO ICHECK2 1, NOFNEI1(I)
IF (II .EQ. NEI1(ICHECK2, I)) THEN

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

529

REPEAT2 .TRUE.
ENDIF
ENDDO
IF (.NOT. REPEAT2) THEN
DO IICORNER 1, 4
IF (VERTEX(II,IICORNER).EQ.
VERTEX(ITEST,ICORNER)) THEN
REPEAT .FALSE.
DO ICHECK 1, J
IF (II .EQ. NEI2(ICHECK, I)) THEN
REPEAT .TRUE.
ENDIF
ENDDO
IF (.NOT. REPEAT) THEN
NOFNEI2(I) NOFNEI2(I) 1
NEI2(NOFNEI2(I), I) II
ENDIF
J MAX(NOFNEI2(I), 1)
ENDIF
ENDDO
ENDIF
ENDIF
ENDDO
ENDDO
ENDDO
ENDDO

The slightly greater complication of this procedure stems from the need
to avoid double counting the neighbors of the first neighbors. The reward
for the more complex code is that again, in spite of the deep nesting, the DO
over the number of cells is well vectorizable and by far outperforms the
recursive construct.
The lists of first neighbors (NEI1) and second neighbors (NEI2)
generated in this way can be used as is for the task of relocating the
particles that have flowed out of a given cell. The search needs only be
performed in the cells contained in the neighbor lists.
However, if the velocity field is known, the search can be further sped
up by ordering the neighbors so that those that are downstream of the given
cell are searched first. This is the idea behind the velocity-biased neighbor
list. The natural way to implement it is to sort the list of neighbors in order
of increasing absolute value of the angle formed between the velocity vector
at some characteristic point of the central cell (e.g., the centroid) and
the vector joining the centroid of the central cell with the centroid of
a neighboring cell (angle
in Fig. 4). Thus, for deterministic particle

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

530

ttinger
Laso and O

FIG. 4 Velocity-biased neighbor lists. Numbers in circles indicate the searching

order in a velocity-biased neighbor list.

trajectories, all particles that have left a cell will necessarily be located in
downstream cells; during the search only a small fraction of the neighbor list
(seldom more than the first three elements) needs to be scanned. The scheme
just presented is equally valid for stochastic particle trajectories, although
due to their inherent variability it is no longer guaranteed that particles
leaving a cell will be located in cells downstream. However, given the
smallness of the typical displacement per step with respect to the cell size,
the velocity-biased neighbor list should be just as efficient as in the
deterministic case.
As a final remark it should be said that although the procedures for first
and second neighbor lists just presented fully exploit vectorizability, they are
still ONel2 algorithms. Their use is justified for fixed meshes because they
are straightforward, well vectorizable, and have to be invoked only once.
For moving meshes, however, the neighbor lists may have to be generated at
every integration step. In this case, ONel2 is unacceptable. As a matter of
fact, the experience gathered so far shows that, unless the neighbor list
problem is dealt with efficiently, neighbor list generation and the subsequent
particle location will become the limiting steps in large-scale calculations
(Section IV.B).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

531

IV. MOVING PARTICLES AND REMESHING

While in a one-dimensional geometry the limit to the computationally
feasible is set by the number and complexity of the molecular models or
stochastic processes to be simulated, the situation is rather different in two
or three dimensions. In these cases, the particle or computational model, be
it deterministic or stochastic, will move throughout the computational
domain following the macroscopic velocity field. Furthermore, particles
may enter or leave the domain, or both. The immediate consequence of the
macroscopic flow is the need to track the particles positions at all times
concurrently with the integration of the equations of change and of the
equations controlling the evolution of the microscopic degrees of freedom.
The need to perform particle tracking entails two attendant computational
tasks, given in Sections IV.A and IV.B.

A. Integration of Particle Trajectories

The first part of particle tracking consists in the integration of the
trajectories particles as they are carried by the macroscopic flow of the fluid.
Since the velocity field is known at all integration times, this is a timeconsuming but in principle straightforward problem. Although particle
trajectories will in most cases be deterministic, for some generalized classes
of constitutive equations (Ottinger 1992) intrinsically stochastic trajectories
need to be considered.
The integration of deterministic particle trajectories is not problematic
since it is possible to draw from the vast body of known algorithms for
ODEs. However, whereas exact deterministic trajectories must always
remain within the integration domain (except at open boundaries), the
discretized versions obtained by any numerical scheme will be subjected to a
finite discretization error. This error has two main consequences with farreaching computational effects in micro/macro methods.
The first effect is a long-time, progressive cumulation of discretization
errors (and possibly truncation errors if compressed-variable representation
is used). This cumulative effect is quite harmless for trajectories along which
the residence time tR of particles is low, in the sense of |(tR/t)u| |u|.
However, in flow fields where regions of intense recirculation exist, the
number of integration steps along a given trajectory (tR/t) is large enough
for the condition above not to be fulfilled. This situation is reflected for
example in the failure of what should be closed-loop trajectories (at steady
state) to close exactly, thus leading to inward- or outward-spiraling
trajectories. This failure is not specific to micro/macro methods but is
common to other tracking techniques, such as those used in viscoelastic flow

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

532

ttinger
Laso and O

FIG. 5 Cumulation of integration errors during particle tracking resulting in

nonclosing recirculation trajectories.

calculations for integral constitutive equations. Figure 5 shows the trajectory of a particle within the recirculation region in the 4:1 contraction
geometry obtained by tracking it in a constant steady-state velocity field for
an Oldroyd-B model.
In spite of the use of a special type of macroelement (the MDF Finite
Element of Nagtigall et al. 1979, Mercier 1979, Bernstein et al. 1985) which
allows analytical tracking, the result of the cumulation of round-off errors in
an otherwise exact technique is striking. Given the low velocities and shear
rates in this particular instance, the consequences of tracking errors are very
minor.
However, for flows with strong recirculation, the erratic behavior of the
discretized trajectories leads to a rapid depletion or enrichment of particles
in certain regions. For sufficiently strong recirculation rates and sufficiently
long integration times, the spurious migration of particles with the wrong
flow history into the wrong elements and the spurious depletion/enrichment
effect render calculations meaningless. The only remedy for this ailment is
the use of smaller time steps or higher-order numerical schemes for ODEs,
or both.
The use of higher-order schemes is often the preferred solution, in
spite of requiring several function evaluations per time step for each of

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

533

the particle trajectories being integrated. This additional effort may pay
back since:
.
.

Trajectory integration does not constitute a large fraction of the overall computational load.
Relatively inexpensive velocity-field updates can be performed independently of and between two consecutive particle configuration
updates with higher frequency than the latter if so required.

The optimal order of the tracking algorithm is thus determined by the ratio
of particle tracking to particle simulation.
Figure 6 shows a comparison of a trajectory integration similar to that
depicted in Fig. 5, but using first-order Euler (Fig. 6a), second- order, fixedstep trapezoidal rule (Fig. 6b), and 12th-order predictor-corrector Adams
Bashforth with self-adaptive step, (Fig. 6c) schemes for a fixed macroscopic
velocity field. While the increase in order does result in increased stability,
even the phenomenally accurate 12th-order scheme fails to yield closed-loop
trajectories. For simple dynamic models including a few degrees of freedom
(Oldroyd-B, CurtissBird) on average-size meshes (a few thousand cells)
with simple elements (P1-C0) and for the current hardware generation,
the best ratio of tracking accuracy to computational time is obtained
for second-order, fixed-step explicit schemes. Higher-order schemes are
advantageous only for very smooth velocity fields, where a very large
integration step can be employed.
The noisy character of micro/macro methods makes it very unlikely that
higher-order schemes will significantly increase overall performance. The
same applies to richer FEs for which velocity updates are increasingly more
expensive.
The second effect of discretization and truncation is not related to the
long-time cumulation of errors, but makes itself observable at every time
step, as particles close to impenetrable boundaries (which can only be
approached asymptotically by trajectories) actually cross them due to a
finite time integration step. In general, the existence of a nonzero outwardpointing component of the velocity of a particle leads to its leaving the
integration domain if the particle is sufficiently close to the boundary.
This effect is in general of an order of magnitude commensurate with the
time step: it acts on a fraction of all particles given by the ratio of the
perimeter @ (or bounding surface in 3D) of the domain times the average
outward-pointing velocity component times the integration step to the total
area (or volume in 3D) of the domain . This is a small ratio for any
nonpathological domain shape. In most cases, the fraction of particles
that leave the integration domain at any given time step due to the
discretization error produced by a first-order explicit Euler scheme with

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

534

ttinger
Laso and O

FIG. 6 Effect of order of tracking algorithm on closedness of trajectories: (a) firstorder Euler, (b) trapezoidal rule, (c) 12th-order adaptive AdamsBashforth PC.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

535

reasonable t lies between 10
4 and 5  10
3 and is therefore of no consequence for the numerical accuracy of the solution. Particles that are not in
the region close to the boundary do not exit the domain. Their trajectories
will drift, however, towards or away from the boundary at a rate dictated
both by the local outward- or inward-pointing velocity component and by
the discretization error.
Figure 7 shows this effect for three particles close to the reentrant corner
in the contraction geometry for P1-C0 quadrilaterals. The closer the initial
position of the particle to the separation line, the more difficult it is to turn
around the corner. Low-order algorithms lead to particle trajectories that
intersect impenetrable boundaries.
Unfortunately, even if only a small fraction of all particles leave the
domain at each time step, these events cannot be ignored. Even though they
have a negligible impact on the numerical values of the stresses computed as
element-wise averages, if a fraction of O(10
410
3) of the total number of
particles leave the domain at each time step due to discretization and
particles drift toward the boundary at a comparable rate, it does not take
more than a few thousand steps for a large fraction of particles to disappear
from the computational domain.
The most straightforward solution to this question consists in reflecting
back into the domain all lost trajectories, i.e., those that have left it at a
given time step. This reflection is acceptable and does not result in a
reduction of the order of the algorithm. However, the crucial aspect of
the problem is that it utterly destroys the vectorizable character we strive
for, since such particle-exit events occur in a geometrically localized and
structured region (the boundary region), but purely at random when considered from the viewpoint of the particle numbering scheme. Hence the
need to develop a scheme to recover and replace such lost particles.
An efficient method to handle this difficulty and retain vectorizability
must: i) make use of the numbering schemes, neighbor lists, and pointinclusion techniques explained in Section III, in order to efficiently
determine which particles have left the domain and where they are located
and ii) obtain this information within and extract it from a loop for
subsequent processing without jeopardizing vectorizability. An optimal twostep algorithm works in the following way:
.

Determine the indices of all particles that have left the element in
which they were located at the previous time step. This is performed
in the double-loop construct already explained in Section III.D using
indirect addressing to point at all particles (inner loop in the following
FORTRAN excerpt) within a given element (outer loop). If a particle
has left the element it was in, add its address to the list of particles

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

536

ttinger
Laso and O

FIG. 7 Effect of order of tracking algorithm on particle trajectories close to

impenetrable boundaries: (a) first-order Euler, (b) trapezoidal rule, (c) 12th-order
AdamsBashforth PC.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

537

that left their elements (LEFT1). A fraction of these will be in first or

second neighbors; the rest will have left the domain altogether.
COUNTLEFT1 0
DO I 1, NUMEL
DO II FIRST(I), LAST(I)
IHOLD ORDWHERE(II)
IF ( particle IHOLD is in element I) THEN
WHERE(IHOLD) I
ELSE
COUNTLEFT1 COUNTLEFT1 1
LEFT1(COUNTLEFT1) IHOLD
ENDIF
ENDDO
ENDDO

The second double loop finds which particles are in first neighbors
and spots those particles that have left the domain:
LOST 0
DO I 1, COUNTLEFT1
IHOLD LEFT1(I)
DO II 1, NOFNEI1(WHERE(IHOLD))
III NEI1(II, WHERE(IHOLD))
IF (particle IHOLD is in first neighbor element III)
THEN
WHERE(IHOLD) III
GOTO 10
ENDIF
ENDDO
LOST
LOST 1
NOTFOUND(LOST) IHOLD
10 CONTINUE
ENDDO

This code is vectorizable, determines where all particles are located,

and, with a minimum overhead, spots those particles that have left
the domain. The array NOTFOUND contains the list of indices for all
particles that have left the domain.
.

In the second step a simple pass is performed over all particles that
have left the domain in order to reflect their positions back into the
domain.
DO I 1, LOST
IHOLD NOTFOUND(I)

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

 ttinger
Laso and O

538

ELEMENT WHERE(IHOLD)
reflect position of particle back into domain
ENDDO

This would seem to complete the procedure, since all particles must
have either been located in nearest neighbors or they must have been
reflected back into the domain. We are, however, not done yet: any
reflection procedure has a loophole which is a consequence of the
finite integration time step: if a particle leaves the domain through
an element boundary that is close to an external nodal point (see Figs.
8 and 9), there will always exist the possibility that the position after
reflection is still outside of the domain or within the domain but not
in the element the particle was originally in.
These two possibilities make it necessary to perform a third and
last pass (not explicitly presented here because of its trivial vectorizability) in order to check that all reflected particle positions are
indeed within the domain and within the intended element.
The root of both of these failures is that, within the discretization
error, it is not possible to decide whether the particle moved from
ri(tj) to ri(tj1) outside the domain directly from element n or through
element n 1. The assignment of a new position to the lost particle is
therefore arbitrary, but since the number of such failures has always a
higher error order, it is irrelevant where the position of such particles should be reinitialized, apart from obviously poor choices like
ri(tj) ri(tj 1).

FIG. 8

Failed particle reflection, first case.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

FIG. 9

539

Failed particle reflection, second case.

We thus see that solving the simple problem of deterministic trajectory

integration requires some degree of sophistication if an efficient algorithm is
to result. As will often be the case, it is the exceptional (in this instance,
particles exiting a closed domain) with little physical significance that usually
makes it necessary to invest a disproportionate algorithmic effort and to
resort to rather convoluted techniques in order to reach vectorizability. At
any rate, the algorithmic complexity of trajectory integration can be kept as
O(Npr), the size of the mesh being irrelevant for integration of particle
trajectories.
The integration of stochastic particle trajectories can be performed using
the same techniques for SDEs used to advance the internal configurations of
stochastic models (Kloeden and Platten 1992). Given the intrinsic variability
of stochastic particle trajectories, the problems associated with particles
leaving the domain are more acute in this case, although they can be handled
using the same techniques. However, the imposition of reflecting boundary
conditions on stochastic trajectories is in general a nontrivial matter
(Ottinger 1995).

B. Particle Localization in the Mesh

On the one hand, the integration of particle trajectories is in principle, and
with the use of proper algorithms as well in practice, an O(Npr) algorithm,
since the information required to integrate a given trajectory is purely local
and does not depend on the size of the mesh. Even the treatment of the
anomalous cases considered in the last part of the previous section involves

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

540

ttinger
Laso and O

no more than the immediate vicinity of the element in which a particle is

located at a given time.
On the other hand, the second part of particle tracking consists in the
localization of the particle in the FE mesh (i.e., given the coordinates of the
particle and of the vertices defining all cells, find in which cell the particle
is located). After a time step the particle may still be within the same cell
or element where it was at the previous step (which will be the most
frequent case) or it may have moved to another element. In both cases it is
necessary to determine in which element the particle is located. This task
presents no great difficulties for either a fixed or a moving mesh once the
neighbor lists are available, since (Section III.D) the search is reduced to
the immediate environment of the original element. It is again an O(Npr)
procedure.
Particle localization in its simplest form calls for checking whether a
given particle lies within a given element, using some suitable point-location
criterion. This brute-force approach is O(Npr)O(Nel) and therefore unsuitable for intensive use in a time-marching scheme. As soon as particles have
been located once (for example when initializing the ensemble), the use of
a neighbor list is vastly more efficient. The use of the brute-force approach is
justified, at most, during the initialization phase. Even then, an element-byelement fill-up scheme easily bypasses the problem.
Since for fixed meshes, there is no need to update the neighbor list at
every time step, the overall effort of advancing trajectories and relocating
particles remains O(Npr) throughout the time integration. For moving grids,
however, the domain has to be reenmeshed after a certain number of
time steps [typically O(100102)], or even after each step (Zegeling 1993,
Trease et al. 1991), a task not to be overlooked: gridding a domain is
isomorphic to Voronoi-tesselating it, for which the best algorithms are
ONel5=4 (Mulmuley 1994).
The reenmeshment makes it necessary to reconstruct the neighbor lists
and to relocate all particles in the new mesh. In this case, unlike during a
time step in a static mesh, a large number of [O(Npr) and possibly all]
particles will be located in a cell with a different index, even though the cell
itself may be geometrically quite similar and close or even identical to the
one before the reenmeshment. Reenmeshment thus has a devastating effect
on the efficiency of a micro/macro method: it requires new neighbor lists
and particle relocation from scratch. Furthermore, a new neighbor list is a
priori not particularly useful.
If the neighbor list is reconstructed using the neighbor list algorithms
presented in Section III.D and brute-force particle location is used, the
combined effort is of order ONel2 ONpr ONel (the first term
corresponding to neighbor list generation, the second to particle location,

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

541

which cannot be alleviated by an element-by-element fill-up scheme as

during initialization). As a result, the generation of the neighbor lists and
relocation of particles may become the bottleneck of the whole micro/macro
simulation. Both of these obstacles can be overcome.

1. An Optimal Neighbor List Generator

The algorithms presented in Section III.D for neighbor list generation are
ONel2 . The following scheme reduces the order of complexity to an optimal
O(Nel log Nel). The key to the successful reduction of complexity is the
transformation of the double-loop construct in Section III.D into a series of
sorting operations. As in Section III.D, we assume that cells or elements in
the grid are specified by an integer array of the form VERTEX(NUMEL,4),
giving the numbers or indices of the four vertices of each cell.
First the array VERTEX(NUMEL,4) is inverted in order to obtain an
array CELL(NVERTS,4) giving the cell numbers associated with each
vertex. Further, NBITS is the number of bits needed to hold the maximum
vertex number, which is roughly equal to log2(NVERTS). For any
nonpathological two-dimensional mesh, NVERTS will be only slightly
larger than NUMEL, the difference between both stemming from vertices
belonging to elements at the domain boundary. The array inversion requires
an intermediate storage and sorting array VC(NUMEL*4). We next generate
all vertex-cell pairings. For this task, the FORTRAN 77 intrinsic routines
lshift, ior, ibits come in handy* (they are not essential for the
present algorithm, but make its implementation simpler and usually more
efficient):
INDEX 0
DO J 1, 4
DO I 1, NUMEL
INDEX INDEX 1
VC(INDEX) IOR(LSHIFT(VERTEX(I,J), NBITS), I)
ENDDO
ENDDO
*Although not ANSI-FORTRAN, the FORTRAN bitwise Boolean functions are intrinsics
from MIL-STD-1753. They are available on most FORTRAN compilers nowadays, including
those by IBM, NEC, Silicon Graphics, and Sun.
lshift(M,K) returns the value of the rst argument shifted left or right, respectively, the
number of times specied by the second (integer) argument. lshift is implemented as a logical
shift, i.e., zeros are shifted in.
ior(M,N) performs a bitwise logical OR of M and N.
ibits(M,K, LEN) extracts a subeld of LEN bits from M starting with bit position K and
extending left for LEN bits. The resulting eld is right justied and the remainig bits are set to
zero.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

 ttinger
Laso and O

542

Left shifting the vertex index by NBITS and bitwise ORing it with the cell
number, produces an array VC, the elements of which contain:
<- - - - - - - - - - - - - - one word - - - - - - - - - - - - ->
0

vertex number

cell number

vertex number

cell number

vertex number

cell number

...

...

...

vertex number

cell number

Sorting VC on the vertex number automatically sorts the corresponding

cell numbers. The most efficient general sorting algorithms are O(Nel
log Nel). Having vector computers in mind, a good way of sorting the array
VC is to use the following modification of the binary radix sort:
.

Starting with the rightmost bit of the word to be sorted, i.e., with each
element of VC, determine whether the corresponding bit position in
each word is a 0 or a 1. Compress all numbers that have 0 as the
rightmost key-bit in a second scratch array of the same length.
Following this, compress out all numbers that have 1 as the rightmost
key-bit into the remainder of the scratch array. Finally, copy the
scratch array back into the original array. Both the scratch array
and the original array now contain all the words or numbers in
the original array, sorted by their least significant bit in ascending
order.
Proceed in the same fashion with the second least significant bit of
each element in the array: all numbers with 0 as the second-rightmost
key-bit are gathered into the second array. Following this, compress
out all numbers with 1 as the second-rightmost key-bit into the
remainder of the scratch array and subsequently copy the scratch
array back into the original array.
The original array now contains all the numbers sorted on their
rightmost two key-bits; first will be all numbers with rightmost bits
00, followed by all numbers with rightmost bits 01, followed by all
numbers with rightmost bits 10, followed by all numbers with
rightmost bits 11. This process continues from right to left until the
numbers have been sorted on each of the key-bit positions.

The entire array is now sorted with the keys in ascending order. In the
following tables we see the way the binary radix sort operates on a sequence

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

543

of three-bit numbers (which may represent a three-bit sorting key in a longer

word).

Although in this example the binary radix sort operates on all bits of
a word it is, however, perfectly possible to sort words or numbers based
on only a few or just one of its bits by applying the bit sort to these bit
positions only.
The following simple implementation of the binary radix sort vectorizes
well on all vector architectures and extraordinarily well on processors where

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

544

ttinger
Laso and O

the compress operations are done by hardware vector instructions (such as

the NEC SX-3 and the deceased ETA family). A(N) is the array to be
sorted, B(N) is the scratch array, and RIGHT and LEFT are the positions
of the bit segment (key) on which sorting should take place (for example,
on a 64-bit-word machine, they could take any values that fulfill
1  LEFT  RIGHT  64). BTEST is a bit field manipulation intrinsic from
MIL-STD-1753* that tests for the state of a given bit.
DO I RIGHT, LEFT
K0
DO J 1, N
IF (BTEST(A(J), I)) THEN
ELSE
KK1
B(K) A(J)
ENDIF
ENDDO
DO J 1, N
IF (BTEST(A(J), I)) THEN
KK1
B(K) A(J)
ENDIF
ENDDO
DO J 1, N
A(J) B(J)
ENDDO
ENDDO

A machine-specific version would avoid the copying from the scratch

array to the original array for each bit position and might require only a
single BTEST function for each word at the expense of more complicated
indexing.*,y
The time to sort a list is therefore proportional to the number of entries
(four entries per cell, given the four vertices per cell) and the width of the key
in bits. If the vertices are numbered from 1 to NVERTS, where NVERTS is
not much larger than the number of cells, the number of bits of the key is
proportional to the logarithm of the number of cells. Thus the binary radix
sort operates in this case on a permutation of the integers from 1 to NUMEL

*btest(N,K) tests the Kth bit of argument N. The value of the function is .TRUE. if the bit is
1 and .FALSE. if the bit is 0.
y
Even so, while not the most ecient available, this routine runs at a good fraction of the
maximum possible speed for this algorithm on the Cray Y-MP and the NEC SX-3.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

545

with each of them appearing up to four times (for quadrilateral cells) and
becomes an O(Nel log Nel) process.
Most vertices are connected to four cells and have four entries in the
array. Those on the edges of the mesh, however, will have fewer. Vertices in
the corners of the grid (if any) may have only one. It is now necessary to
determine where in the sorted array VC each vertex begins and ends. The
following code finds each place where the vertex number changes and enters
this information into two index arrays: ILASTI, giving the last index into
the array VC for each vertex, and IGATHR, which initially is set to the first
index minus one for each vertex. IGATHR is used to extract successive cells
connected to each vertex:
ENTRY 0
DO I 1, INDEX - 1
IF (IBITS(VC(I),NBITS,NBITS) .NE. IBITS(VC(I1),NBITS,
NBITS)) THEN
ENTRY ENTRY 1
ILASTI(ENTRY) I
ENDIF
ENDDO
ENTRY ENTRY 1
ILASTI(ENTRY) INDEX
IGATHR(1) 0
DO I 2, ENTRY
IGATHR(I) ILASTI(ENTRY-1)
ENDDO

The next step consists in building up a four-column table: for each separate vertex enter one of the connected cells in a separate column. If a vertex
is shared by fewer than four cells, repeat the last cell number for that vertex
as necessary. In a last pass, any such duplicated cell numbers are set to zero:
DO I 1, 4
DO J 1, ENTRY
IF (IGATHR(J) .LT. ILASTI(J)) SIGATHR(J) IGATHR(J)1
CTABLE(J) IBITS(VC(IGATHR(J), 0, NBITS))
ENDDO
ENDDO
DO I 4, 2 -1
DO J 1, ENTRY
IF (CTABLE(J, I) .EQ. CTABLE(J, I-1)) THEN CTABLE(J, I) 0
ENDDO
ENDDO

For each cell number, the only nearest neighbor cells are those that occur
in one or more rows along with it in the array CTABLE. The following code

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

 ttinger
Laso and O

546

generates a table CELNBR that contains all cell-neighbor pairs and only cellneighbor pairs:
PAIRS 0
DO I 1, 4
DO J 1, 4
IF (I .NE. J) THEN
DO K 1, ENTRY
IF ((CTABLE(K,I) .NE. 0) .AND. (CTABLE(K,J) .NE. 0)) THEN
PAIRS PAIRS 1
CELNBR(PAIRS)
IOR (LSHIFT(CTABLE(K,I),NBITS),CTABLE(K,J))
ENDIF
ENDDO
ENDIF
ENDDO
ENDDO

The resulting array CELNBR contains:

<- - - - - - - - - - - - - - one word - - - - - - - - - - - - ->
0

cell number

neighbor

cell number

neighbor

cell number

neighbor

...

...

neighbor

cell number

neighbor

There will now be some duplicates since some cell pairs (those where two
cells share more than one vertex) will appear in CELNBR twice. Additionally,
since the cell numbers will be in no particularly useful order, sort the array
CELNBR using the cell number in each entry as the key (in the middle of the
word).
The sorted CELNBR contains at most twelve neighbor entries for each
cell.* A neighbor that shares only a corner appears once, a neighbor that
shares two corners (an edge) appears twice. Now eliminate the duplicates by
sorting on the entire cell-neighbor pair rather than on the cell number only.
But if we take into account that duplicate entries can be no more than eleven
words apart (in the event they are first and last in the neighbor list), we can

*In the case of a structured mesh, where the number of neighbors is xed.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

547

replace this sorting by the more elegant device of comparing each entry in
the array, as sorted on cell number only, with the previous eleven entries:
DO I 1, 11
DO J I1, PAIRS
IF (CELNBR(J) .EQ. CELNBR(J-I)) CELNBR(J) 0
ENDDO
ENDDO
UNIQUE 0
DO I 1, PAIRS
IF (CELNBR(I) .NE. 0) THEN
UNIQUE UNIQUE 1
CELNBR(UNIQUE) CELNBR(I)
ENDIF
ENDDO

The last step consists in finding the first and last entries for each cell
number, which can be done in exactly the same fashion as for the array VC
sorted on vertex number and will therefore not be repeated here. After that,
reformatting the table of nearest neighbors into whatever format we wish is
a trivial matter.
We thus see that while a frontal attack results in the elegant and
straightforward but ONel2 algorithm of Section III.D, recasting the
problem as a quite convoluted sequence of sorting and indirect addressing
yields an O(Nel log Nel) process and therefore optimal.* The break-even
point between the two algorithms will obviously depend on the particular
mesh being used, in particular on the ratio of boundary cells to internal cells.
The figures in Table 1 apply to the two-dimensional eccentric cylinder
meshes.
TABLE 1 Peformance Comparison for Increasingly Refined Meshes

Mesh
M0
M1
M3
M4
M5

Number
of vertices

Number
of cells

CPU-time ratio
(Cray Y-MP):
algorithm section III.D/
algorithm section IV.B.1

55
220
840
1680
16800

50
200
800
1600
16000

0.81
2.5
8.3
25
130

CPU-time ratio
for scalar processor
0.75
2.2
8.0
23
110

*The algorithm just described is a constructive proof of the equivalence of building a neighbor
list and a search.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

 ttinger
Laso and O

548

We thus see that the break-even point lies at about a few tens of cells, i.e.,
except for the coarsest grids, which in practice never occur, the algorithm
just presented is significantly more efficient. For realistic 2D computations
and, above all in three dimensions, where grid sizes of O(104105) cells are
standard, the algorithm of Section IV.B.1 is at least two orders of magnitude
faster and the only viable alternative. Finally, it is worth mentioning that the
growing CPU-time ratios apply to absolute CPU-times that in turn grow
with problem size.
Interestingly, the effort invested in developing vectorizable algorithms
pays off on scalar architectures as well. The CPU ratio between the nestedloop algorithm and the algorithm just presented is slightly more favorable
for the latter on vector machines since the nested DO loop constructs of
Section III.D contain comparatively more or deeper IFs and low-count
DOs that must be unrolled.

2. Particle Relocation
As we saw in IV.B, the problem of determining in which element a given
particle is located after the mesh has been created, renumbered, or modified
is of O(Npr)O(Nel), i.e., quadratic in problem size, instead of O(Npr) as is the
case when the mesh is static and neighbor lists can be used. The reason is
that knowledge of the neighboring cells is purposeless if the location of the
particle at the previous time step is unknown, as is the case when the domain
is reenmeshed.* A partial remedy for the O(Npr)O(Nel) complexity could
be achieved by reenmeshment procedures that keep renumbering to a
minimum. Although the amount of literature on reenmeshing procedures is
overwhelming, little attention has been given to the question of minimumrenumbering algorithms.
While the quadratic complexity of particle relocation is unavoidable the
very first time a domain is initialized,y it is possible in practice to bypass this
complexity in subsequent steps. The reconstructed neighbor list can be used
in order to avoid brute-force particle location no matter how extensive the
renumbering is. The two-step idea is straightforward and does not require
pseudo-code to illustrate it:
.

Once the updated neighbor list is available, a brute-force search

is started for inclusion in all cells of a small number of particles
(typically 20 to 100) belonging to the original cell. Since the indices

*Even if the reenmeshment would only renumber cells, the new cell number would bear no
relationship whatsoever with the number before the reenmeshment.
y
Even in this case, particle relocation can be avoided by systematic cell ll-up, resulting
in O(Npr).

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

549

of these particles are contiguous in the array ORDWHERE ( see Section

III.D) this search is very fast; furthermore, the effort spent in this
tentative search is the same, no matter what the size of the local
ensemble is. Given that most of the particles will not have left
the original cell (now renumbered beyond recognition), as soon
as the particles are tested for inclusion in the original cell (but with
the updated index), a large majority of the particles test positive. This
is all that is required to identify the new index of the original
cell.
As soon as the new cell index is known, the search is restarted, this
time for all the particles that were in the original cell but using the
reconstructed neighbor cell, thus reducing the problem to the same
one encountered for static meshes (Section III.D), i.e., to O(Npr).

The objective has been accomplished: combined use of the optimal

neighbor list generator and the relocation technique just described reduces
the ONel2 ONpr ONel complexity to O(Nel log Nel) O(Npr).
The core of the computationally crucial part of CONNFFESSIT
calculations has been described in great detail in Sections III and IV.
The remaining elements in a CONNFFESSIT calculation are relatively
straightforward. Two recent references describe in detail specific applications to the journal bearing problem (Laso et al. 1997) and to the planar
contraction flow (Laso 1998) as well as the implementation on parallel
and vector machines.

V. CONCLUSIONS AND PERSPECTIVES

The CONNFFESSIT approach is a feasible novel method for the
investigation of the fluid mechanics of complex fluids which cannot be
dealt with satisfactorily within the continuum-mechanical framework. In
order to keep the computational effort involved in the calculation of twodimensional time-dependent flows in complex geometries under control,
efficient algorithmic tools and data structures are required.
Even so, CONNFFESSIT cannot compete in speed with the most recent
micromacro methods but it is still the tool of choice when fluctuations are
important, typically in restricted geometries. Its efficient implementation,
while not complex, requires that care be exercised. The techniques presented
here are certainly not the only ones, but they are reasonably close to algorithmic optimality. Further improvements are certainly possible through
the use of increasingly complex shortcuts, although the gain will probably
be only marginal.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

 ttinger
Laso and O

550

The recent micromacro methods, CONNFFESSIT among them, can, in

spite of their youth, already be seriously considered a tool complementary to
those of classical computational rheology.

APPENDIX A: SYMBOLS
akj
i
fjk
fnl,. . .
l
n
n
Nel
Np
Npr
Nsim
Nver
Nsides
Q
ri(t)
T
t
t
u
u
Ws
Xij
Yij
Xcj
Ycj
xi
yi
zi
d
s


ith coefficient in Cartesian equations for jth side of kth

element
Cartesian equation of the jth side of the kth element
inflow factors
length along boundary between two cells (m)
number density of particles (m
3)
unitary normal vector
number of mesh cells or elements (NUMEL in FORTRAN
excerpts)
number of computational nodes
number of particles in the global ensemble (NUMPR in
FORTRAN excerpts)
number of particles per cell or element
number of vertices of the mesh (NVERTS in FORTRAN
excerpts)
number of sides of an element
connector vector (m)
position vector of particle i (m)
temperature (K)
time (s)
integration time step, or average time step (s)
velocity field (m/s)
discretization error in macroscopic velocity field per time
step (m/s)
Weissenberg number
x-coordinate of ith vertex of jth element (m)
y-coordinate of ith vertex of jth element (m)
x-coordinate of centroid of jth element (m)
y-coordinate of centroid of jth element (m)
x-component of position vector of particle i (m)
y-component of position vector of particle i (m)
z-component of position vector of particle i (m)
unit tensor
extra-stress or deviatoric part of the stress tensor (Pa)
integration domain; its area (2D) (m2) or volume (3D) (m3)

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

551

boundary of the integration domain; its perimeter (2D) (m)

or area (3D) (m2)

APPENDIX B: ABBREVIATIONS
CE
CFD
FE
MC
MD
NEMD
ODE
PC
PDE
SDE
SPH
WR

Constitutive Equation
Computational Fluid Dynamics
Finite Elements
Monte Carlo
Molecular Dynamics
Non-Equilibrium Molecular Dynamics
Ordinary Differential Equation
Predictor-Corrector
Partial Differential Equation
Stochastic Differential Equation
Smoothed-Particle Hydrodynamics
Weighted Residuals

REFERENCES
Ait-Kadi, A.; Dealy, J.M.; James, D.F.; Williams, M.C., Eds., Proceedings of the
XIIth International Congress on Rheology, Laval University, Quebec City,
1996.
Allen, M.P., Tildesley, D.J. Computer Simulation of Liquids; Clarendon Press:
Oxford, 1987.
Anderson, C.; Greengard, C. SIAM J. Math. Anal. 1985, 22, 413440.
Anderson, C.; Greengard, C., Eds., Vortex Methods; Springer: Berlin 1988.
Ashurst, W.T.; Hoover, W.G. Phys. Rev. A 1975, 11, 658667.
Baaijens, F.P.T. J. Non-Newtonian Fluid Mech. 1998, 75, 119138.
Balasubramanian, K.; Hayot, F.; Saam, W.F. Phys. Rev. A 1987, 36, 22482252.
Baudet, C.; Hulin, J.P.; Lallemand, P.; dHumieres, D. Phys. Fluids A 1989, 1,
507518.
Beale, J.T.; Majda, A. Math. Comp. 1982, 39, 127.
Bell, T.W.; Nylond, G.H.; de Pablo, J.J.; Graham, M.D. Macromolecules 1997, 30,
18061812.
Bernstein, B.; Malkus, D.S.; Olsen, E.T. Int. J. Num. Meth. Fluids 1985, 5,
4370.
Bird, R.B.; Curtiss, C.F.; Armstrong, R.C.; Hassager, O. Dynamics of Polymeric
Liquids; John Wiley: New York, 1987b; Vol. II.
Birdsall, C.K.; Langdon, A.B. Plasma Physics via Computer Simulation; McGrawHill: London, 1985.
Bonvin, J.C.; Picasso, M. J. Non-Newtonian Fluid Mech. 1999, 84, 191215.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

552

ttinger
Laso and O

Bouseld, D.W.; Keunings, R; Denn, M.M. J. Non-Newtonian Fluid Mech. 1988,

27, 205221.
Brown, R.A.; Armstrong, R.C.; Beris, A.N.; Yeh, P.W. Comp. Meth. App. Mech.
Eng. 1986, 58 (2), 201226.
Bush, M.B.; Milthorpe, J.F.; Tanner, R.I. J. Non-Newtonian Fluid Mech. 1984, 16,
3751.
Castillo, J.E., Ed., Mathematical Aspects of Numerical Grid Generation; Society for
Industrial and Applied Mathematics: Philadelphia, 1991.
Chorin, A.J. J. Fluid Mech. 1973, 57, 785796.
Chorin, A.J. Computational Fluid Mechanics: Selected Papers; Academic Press: New
York, 1989.
Chorin, A.J. Vorticity and Turbulence; Springer: Berlin, 1994.
Cloutman, L.D. An evaluation of smoothed particle hydrodynamics. In Advances in
the Free-Lagrange Method; Trease, H.E., Fritts, M.J., Crowley, W.P., Eds.;
Springer: Berlin, 1991.
Crochet, M.J. Rubber Chem. Technol. 1989, 62 (3), 426455.
Crochet, M.J.; Davies, A.R.; Walters, K. Numerical Simulation of Non- Newtonian
Flow;. Elsevier: Amsterdam, 1984.
Debbaut, B.; Crochet, M.J. J. Non-Newtonian Fluid Mech. 1986, 20, 173185.
Doi, M.; Edwards, S.F. The Theory of Polymer Dynamics; Clarendon Press: Oxford,
1986.
Feigl, K.; Laso, M.; Ottinger, H.C. Macromolecules 1995, 28 (9), 32613274.
Franca, L.P.; Hughes, T.J.R. Comp. Meth. Appl. Mech. Eng. 1993, 105, 285298.
Frisch, U.; Hasslacher, B.; Pomeau, Y. Phys. Rev. Lett. 1986, 56, 15051509.
Gallegos, C.; Guerrero, A.; Munoz, J.; Berjano, M., Eds., Proceedings IVth
European Rheology Conference, Sevilla, 1994.
Gard, T.C. Introduction to Stochastic Dierential Equations; Marcel Dekker: New
York, 1988.
Gardiner, C.W. Handbook of Stochastic Methods for Physics, Chemistry and the
Natural Sciences; Springer: Berlin, 1983.
Geurts, B.J.; Jongschaap, R.J.J. J. Rheol. 1988, 32, 353372.
Gingold, R.A.; Monaghan, J.J. Month. Not. R. Astr. Soc. 1977, 181, 375387.
Halin, P.; Lielens, G.; Keunings, R.; Legat, V. J. Non-Newtonian Fluid Mech. 1998,
79, 387403.
Hannon, L.; Lie, G.C.; Clementi, E. Phys. Lett. 1986a, 57, 695699.
Hardy, J.; de Pazzis, O.; Pomeau, Y. Phys. Rev. A 1976, 13, 19491954.
Hockney, R.W.; Eastwood, J.W.; Computer Simulation using Particles; McGrawHill: London, 1981.
Honerkamp, J. Stochastische Dynamische Systeme; VCH: Weinheim, 1990.
Hoover, W.G.; De Groot, A.J.; Hoover, C.G. Computers in Phys. 1992, 6, 155167.
Hulsen, M.A.; van Heel, A.P.G.; van den Brule, B.H.A.A. J. Non-Newtonian Fluid
Mech. 1997, 70, 79101.
Josse, S.L.; Finlayson, B.A. J. Non-Newtonian Fluid Mech. 1984, 16, 1336.
Kadano, L.P.; Swift, J. Phys. Rev. 1968, 165, 310314.
Keunings, R. J. Computat. Phys. 1986a, 62, 199220.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

553

Keunings, R. J. Non-Newtonian Fluid Mech. 1986b, 20, 209226.

Keunings, R. Simulation of viscoelastic uid ow. In Fundamentals of Computer
Modeling for Polymer Processing; Tucker, C.L., Ed.; Carl Hanser Verlag:
Munich, 1989; 402470.
Keunings, R.; Bouseld, D.W. J. Non-Newtonian Fluid Mech. 1987, 22, 219233.
Kieling, I.; Lowes, M. Programmierung mit FORTRAN 77; B.G. Teubner:
Stuttgart, 1985.
Kloeden, P.E.; Platen, E. Numerical Solution of Stochastic Dierential Equations;
Springer: Berlin, 1992.
Knupp, P.; Steinberg, S. Fundamentals of Grid Generation; CRC Press: Boca Raton,
FL, 1993.
Knuth, D.E. The Art of Computer Programming; Addison-Wesley: Reading, MA,
1968.
Laso, M. AIChEJ 1994, 40 (8), 12971311.
Laso, M. Habilitationsschrift, Eidgenossische Technische Hochschule Zurich,
1995.
Laso, M. In Dynamics of Complex Fluids; 72 Adams, M.J., Mashelkar, R.A.,
Pearson, J.R.A., Rennie, A.R., Eds.; Imperial College Press, The Royal
Society: London, 1998.
Laso, M.; Ottinger, H.C. J. Non-Newtonian Fl. Mech. 1993, 47, 120.
Laso, M.; Picasso, M.; Ottinger, H.C. AIChEJ 1997, 43, 877892.
Lucy, L.B. Astrophys. J. 1977, 82, 10131022.
Malkus, D.S.; Webster, M.F. J. Non-Newtonian Fluid Mech. 1987, 25, 221243.
Marchal, J.M.; Crochet, M.J. J. Non-Newtonian Fluid Mech. 1987, 26, 77114.
Marchioro, C.; Pulvirenti, M. Vortex Methods in Two-Dimensional Fluid Mechanics;
Springer: Berlin, 1984.
McCammon, J.A.; Northrup, S.H.; Karplus, M.; Levy, R. Biopolymers 1980, 19,
2033, 2045.
McClelland, M.A.; Finlayson, B.A. J. Non-Newtonian Fluid Mech. 1988, 27,
363374.
Melchior, M.; Ottinger, H.C. J. Chem. Phys. 1995, 103, 95069509.
Melchior, M.; Ottinger, H.C. J. Chem. Phys. 1996, 105, 33163331.
Mendelson, M.A.; Yeh, P.W.; Brown, R.A.; Armstrong, R.C. J. Non-Newtonian
Fluid Mech. 1982, 10, 3154.
Mercier, B. Int. J. Num. Meth. Eng. 1979, 14, 942945.
Moldenaers, P., Keunings, R., Eds., Proceedings of the XIth International Congress
on Rheology, Brussels, Elsevier: Amsterdam, 1992.
Mulmuley, K. Computational Geometry: An Introduction Through Randomized
Algorithms; Prentice Hall: Englewood Clis, NJ, 1994.
Ottinger, H.C. Rheol. Acta 1992, 31, 1421.
Ottinger, H.C. Stochastic Processes in Polymeric Fluids; Springer-Verlag: Berlin,
1995.
Ottinger, H.C.; Laso, M. In Theoretical and Applied Rheology, Moldenaers, P.,
Keunings, R., Eds., Proceedings XIth Int. Congress on Rheology, Brussels,
Elsevier: Amsterdam, 1992; 286288.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

554

ttinger
Laso and O

Papanastasiou, A.C.; Scriven, L.E.; Macosko, C.W. J. Non-Newtonian Fluid Mech.

1987, 22, 271288.
Preparata, F.P.; Shamos, M.I. Computational Geometry: An Introduction; Springer:
New York, 1990.
Risken, H. The Fokker-Planck Equation; Springer Series in Synergetics: Berlin, 1989.
Rothman, D.H.; Keller, J.M. J. Stat. Phys. 1988, 52, 11191128.
Srinivasan, R.; Finlayson, B.A. J. Non-Newtonian Fluid Mech. 1988, 27, 115.
Sugeng, F.; Phan-Thien, N.; Tanner, R.I. J. Rheol. 1987, 31, 3758.
Trease, H.E.; Fritts, M.F.; Crowley, W.P., Eds.; Advances in the Free-Lagrange
Method; Springer: Berlin, 1991.
Van Kampen, Stochastic Processes in Physics and Chemistry; North-Holland:
Amsterdam, 1981.
Van Schaftingen, J.J., The Linear-Stability of Numerical-Solutions of Some
Elementary Viscoelastic Flows, 1986.
Van Schaftingen, J.J.; Crochet, M.J. Int. J. Num. Meth. Fluids 1984, 4, 10651081.
Weatherill, N.P. et al. Ed., Numerical Grid Generation in Computational Fluid
Dynamics and Related Fields; Pineridge Press: Swansea, UK, 1994.
Zegeling, P.A. Moving-Grid Methods for Time-Dependent Partial Dierential
Equations, CWI Tract, Amsterdam, 1993. Also available at http://www.math.
uu.nl/people/zegeling/publist.html.

BIBLIOGRAPHY
Ackerson, B.J.; Pusey, P.N. Phys. Rev. Lett. 1988, 61, 10331036.
Aggarwal, R.; Keunings, R.; Roux, F.-X. J. Rheology 1994, 38 (2), 405419.
Akl, S.G. Parallel Sorting Algorithms; Academic Press: New York, 1985.
Allen, M.P. Simulations and phase behaviour of liquid crystals. In Observation,
Prediction and Simulation of Phase Transitions in Complex Fluids, Baus, M.,
Rull, L.F., Ryckaert, J.-P., Eds.; NATO ASI Series, 1995; 557591.
Ballal, B.Y.; Rivlin, R.S. Trans. Soc. Rheol. 1976, 20, 65101.
Bartlett, P.; van Megen, W. Physics of Hard-Sphere Colloidal Suspensions, Mehta, A.
Ed.; Springer: Berlin, 1994.
Beris, A.N.; Armstrong, R.C.; Brown, R.A. J. Non-Newtonian Fluid Mech. 1983,
13, 109148.
Beris, A.N.; Armstrong, R.C.; Brown, R.A. J. Non-Newtonian Fluid Mech. 1984,
16, 141171.
Beris, A.N.; Armstrong, R.C.; Brown, R.A. J. Non-Newtonian Fluid Mech. 1986,
19, 323347.
Beris, A.N.; Armstrong, R.C.; Brown, R.A. J. Non-Newtonian Fluid Mech. 1987,
22, 129167.
Bird, R.B.; Armstrong, R.C.; Hassager, O. Dynamics of Polymeric Liquids; John
Wiley: New York, 1987a; Vol I.
Bird, R.B.; Dotson, P.J.; Johnson, N.L. J. Non-Newtonian Fluid Mech. 1980, 7,
213235.
Brady, J.F.; Bossis, G. J. Fluid Mech. 1985, 155, 105129.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

555

Burdette, S.R. J. Non-Newtonian Fluid Mech. 1989, 32, 269294.

Chen, L.B.; Ackerson, B.J.; Zukoski, C.F. J. Rheol. 1994, 38, 193202.
Clemencon, C.; Decker, K.M.; Endo, A.; Fritscher, J.; Masuda, N.; Muller, A.;
Ruhl, R.; Sawyer, W.; de Sturler, E.; Wylie, B.J.N.; Zimmermann, F.
SPEEDUP Journal, 1994.
Cook, R.; Wilemski, G. J. Phys. Chem. 1992, 96, 40234027.
Cooke, M.H.; Bridgwater, J. Powder Bulk Solids Technol. 1979, 3, 1119.
Curtiss, C.F.; Bird, R.B. Physica. 1983, 118A, 191204.
Curtiss, C.F.; Bird, R.B.; Hassager, O. Adv. Chem. Phys. 1976, 35, 31117.
Davies, M.J.; Walters, K. In Rheology of Lubricants, Davenport, T.C., Ed.;
Halsted Press, 1973.
Doi, M.; Edwards, S.F. J. Chem. Soc., Faraday Trans. II 1978, 74, 17891832.
Doi, M.; Edwards, S.F. J. Chem. Soc., Faraday Trans. II 1979, 75, 3854.
Doi, S.; Washio. T.; Muramatsu, K.; Nkata, T. Preprints of Parallel CFD94, Kyoto
Institute of Technology, Kyoto, Japan, May 1994, pp. 31-36.
Ermak, D.L.; McCammon, J.A.; J. Chem. Phys. 1978, 69 (4), 13531360.
Erpenbeck, J.J. Phys. Rev. Lett. 1984, 52 (15), 13331335.
Evans, D.J.; Morriss, G.P. Phys. Rep. 1984, 1, 297344.
Evans, D.J.; Morriss, G.P. Phys. Rev. Lett. 1986, 56 (20), 21722175.
Fix, G.J.; Paslay, P.R. J. Appl. Mech. Trans. ASME 1967, 34, 579589.
Fixman, M. J. Chem. Phys. 1978, 69, 15271538.
Fixman, M., Macromolecules 1981, 14, 17101717.
Fletcher, C.A.J. Computational Techniques for Fluid Dynamics; Springer: Berlin,
1991.
Fuentes, Y.O.; Kim, S. AIChEJ 1992, 38, 10591078.
Goodwin, J.W. Ed., Colloidal Dispersions; Royal Society of Chemistry: London,
1982.
Hanley, H.J.M.; Morriss, G.P.; Welberry, T.R.; Evans, D.J. Physica 1988, 149A,
406414.
Homan, R.L. Trans. Soc. Rheol. 1972, 16, 155163.
Hopkins, A.J.; Woodcock, L.V. J. Chem. Soc. Faraday Trans. II 1990, 86,
21212132.
Hwa, T.; Kardar, M. Phys. Rev. Lett. 1989, 62, 18131820.
Irving, J.H.; Kirkwood, J.G. J. Chem. Phys. 1960, 18, 817829.
Joseph, D.D.; Saut, J.C. J. Non-Newtonian Fluid Mech. 1986, 20, 117141.
Kalus, J.; Lindner, P.; Homann, H.; Ibel, K.; Munch, C.; Sander, J.; Schmelzer, U.;
Selbach, J. Physica B 1991, 174, 164169.
Kamal, M.M. J. Basic Engng. ASME 1966, 88, 717724.
Keunings, R. Computers and Chemical Engng.; in press, 1995.
Keunings, R.; Halin, P. personal communication, 1995.
Kim, S.; Karniadakis, G.; Vernon, M.K. Parallel Computing in Multiphase
Flow Systems Simulations; Eds.; The Fluids Engineering Division, ASME,
1994.
Kirkwood, J.G. Macromolecules; Gordon & Breach, 1967.
Krieger, I.M. Adv. Coll. Interface Sci. 1982, 3, 111127.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

556

ttinger
Laso and O

Lakshmivarahan, S.; Dhall, S.K.; Miller, L.L. Parallel sorting algorithms. Advances
in Computers, 1984, 23, 295354.
Laun, H.M.; Bung, R.; Hess, S.; Loose, W.; Hess, O.; Hahn, K.;
Hadicke, E.; Hingmann, R.; Schmidt, F.; Lindner, P.; J. Rheol. 1992, 36 (4),
743787.
Lawler, J.V.; Muller, S.J.; Brown, R.A.; Armstrong, R.C. J. Non-Newtonian Fluid
Mech. 1986, 20, 5192.
Lee, S.J.; Denn, M.M.; Crochet, M.J.; Metzner, A.B.; Rigins, G.J. J. NonNewtonian Fluid Mech. 1984, 14, 301325.
Lees, A.W.; Edwards, S.F. J. Phys. C: Solid State Phys. 1972, 5, 19211929.
Lesieur, M. Turbulence in Fluids: Stochastic and Numerical Modelling; Kluwer:
Dordrecht, 1990.
Lighthill, J. An Informal Introduction to Theoretical Fluid Mechanics; Oxford Science
Publications: Oxford, 1986.
Mochimaru, Y., J. Non-Newtonian Fluid Mech. 1983, 12, 135152.
MPIF (Message Passing Interface Forum), Technical Report CS-94-230, University
of Tennessee, Knoxville, TN, USA, April 1994; also J. of Supercomputer
Applications 1994, 8, also available from:
http://www.mcs.anl.gov/mpi/mpi-report/mpi-report.html
Nagtigaal, J.; Parks, D.M.; Rice, J.R. Comp. Meths. Appl. Mech. Eng. 1974, 4,
153178.
Nickell, R.E.; Tanner, R.I.; Caswell, B. J. Fluid. Mech. 1981, 65, 219248.
Olivari, D. Introduction to the Modeling of Turbulence, Von Karman Institute for
Fluid Dynamics, 1993.
Ottinger, H.C. J. Chem. Phys. 1990, 92, 45404549.
Ottinger, H.C.; Laso, M. Bridging the gap between molecular models and viscoelastic ow calculations. In Lectures on Thermodynamics and Statistical
Mechanics; World Scientic, Singapore, 1994; Chapter VIII, 139153.
Parrinello, M.; Rahman, A. Phys. Rev. Lett. 1980, 45, 11961199.
Parrinello, M.; Rahman, A. J. Appl. Phys. 1981, 52, 71827190.
Parrinello, M.; Rahman, A. J. Chem. Phys. 1982, 76, 26622666.
Peterlin, A. Makr. Chem. 1961, 4446, 338346.
Peterlin, A. Kolloid-Zeitschrift 1962, 182, 110115.
Pironneau, O. Finite Element Methods for Fluids; John Wiley: New York, 1989.
Phan-Thien, N.; Tanner, R.I. J. Non-Newtonian Fluid Mech. 1981, 9, 107117.
Phan-Thien, N.; Sugeng, F.; Tanner, R.I. J. Non-Newtonian Fluid Mech. 1984, 24,
301325.
Powell, R.L. J. Stat. Phys. 1991, 62, 10731094.
Pusey, P.N. Colloidal suspensions. In Liquids, Freezing and Glass Transition;
Hansen, J.P., Levesque, D., Zinn-Justin, J., Eds.; North-Holland,
Amsterdam, 1991; 765942.
Rajagopalan, D.; Byars, J.A.; Armstrong, R.C.; Brown, R.S. J. Rheol. 1992, 36 (7),
13491375.
Reiner, M.; Hanin, M.; Harnoy, A. Israel J. Tech. 1969, 7, 273279.
Rigos, A.A.; Wilemski, G. J. Phys. Chem. 1992, 96, 39813986.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

CONNFFESSIT

557

Rivlin, R.S. J. Non-Newtonian Fluid Mech. 1979, 79, 79101.

Roberts, G.W.; Walters, K. Rheol. Acta 1992, 31, 5562.
Russel, W.B. The Dynamics of Colloidal Systems; University of Wisconsin Press,
1987.
Schoeld, A.N.; Wroth, C.P. Critical State Soil Mechanics; McGraw-Hill: London,
1968.
Sommerfeld, A. Z. Math. Physik 1904, 50, 97155.
Tanner, R.I. Aust. J. Appl. Sci. 1963, 14, 129136.
Temam, R. Navier-Stokes Equations; North-Holland: Amsterdam, 1977.
ter Horst, W.J. MSC internal report MEAH-172. Laboratory for Aero and
Hydrodynamics. Delft University of Technology, 1998.
Townsend, P. J. Non-Newtonian Fluid Mech. 1980, 6, 219243.
Ungar, L.H.; Brown, R.A. Philos. Trans. R. Soc. London 1982, A306, 347370.
Van Megen, W.; Snook, I. J. Chem. Phys. 1988, 88, 11851191.
Warner, H.R. Ind. Eng. Chem. Fundamentals 1972, 11, 379387.
Wilcox, D.C. Turbulence Modeling for CFD, DCW Industries, 1994.
Wilemski, G. J. Stat. Phys. 1991, 62 (5/6), 12391253.
Woodcock, L.V. Phys. Rev. Lett. 1985, 54 (14), 15131516.
Xue, W.; Grest, G.S. Phys. Rev. A 1989, 40 (3), 17091712.
Xue, W.; Grest, G.S. Phys. Rev. Lett. 1990, 64 (4), 419422.
Yamada, T.; Nose, S. Phys. Rev. A 1990, 42, 62826288.
Yeh, P.-W.; Kim, E.M.E.; Armstrong, R.C.; Brown, R.A. J. Non-Newtonian Fluid
Mech. 1984, 16, 173194.
Zone, O.; Vanderstraeten, O.; Keunings, R. A parallel direct solver for implicit nite
element problems based on automatic domain decomposition. In Massively
Parallel Processing: Applications and Development, Dekker, L., Smit, W.,
Zuidervaart, J.C., Eds.; Elsevier, Amsterdam, 1994; 809816.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

17
Simulation of Polymers by Dissipative
Particle Dynamics
W. K. DEN OTTER

University of Twente, Enschede, The Netherlands

J. H. R. CLARKE University of Manchester Institute of Science and

Technology, Manchester, United Kingdom

I.

INTRODUCTION

Dissipative particle dynamics (DPD) was introduced a decade ago by

Hoogerbrugge and Koelman [1,2] as a particle-based off-lattice simulation
method for the flow of complex fluids. Developed for mesoscopic length
and time scales, i.e., much larger than atomic but still much smaller than
macroscopic, DPD faithfully reproduces the relevant thermodynamics,
hydrodynamics, and the omnipresent thermal noise. It was quickly
recognized that DPD offered new and exciting opportunities for modeling
polymers, materials that are rich in physical phenomena characterized by
their mesoscopic scale. Already it has made significant impact on our
understanding of microphase separation in block copolymers and phase
behavior of polymer solutions [3].
The DPD particles loosely represent a number of co-moving fluid
molecules, much like the familiar fluid elements of classical hydrodynamics.
Because of this coarse-grained interpretation of the particle, in both the
spatial and the temporal sense, the conservative force acting between any
two neighboring particles represents the time average of the sum of all
atomic interactions between the atoms constituting the two particles,
resulting in a smooth potential. In DPD the conservative force is taken to be
of a purely repulsive nature, causing the particles to spread out evenly over
the periodic simulation box. The effects of fluctuations around the averaged
interaction, due to the subparticle dynamics, are represented by dissipative
and random forces, in analogy with the Langevin model for the Brownian
particle and the more general MoriZwanzig formalism for separating slow

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

560

den Otter and Clarke

and fast variables [4,5]. The strength of the DPD method lies in the softness
of the forces, and in the specific construction by which all forces act between
particle pairs following Newtons third law. Because of this latter condition,
momentum is conserved locally. Hydrodynamic interactions, therefore, arise
automatically during the simulation, so the system willwhen viewed on a
sufficiently large time and length scalebehave like any regular fluid [4].
As it turns out, surprisingly few particles are needed in practice to recover
hydrodynamic phenomena.
Hoogerbrugge and Koelman tested their method on a quiescent fluid, a
fluid undergoing Couette flow, and a fluid flow around an array of cylinders
[1]. The objects in the latter simulation were made by locally immobilizing
the liquid. But their pie`ce de resistance was calculating the viscosity of a
colloidal suspension (with rigid groups of particles acting as moving and
rotating colloids) as a function of the colloid volume fraction, in good
agreement with the celebrated Einstein expression at low fractions [2].
Research in this area has since continued in the group of Coveney and
Boek [6]. Using DPD, liquid mixtures have been simulated by the simple
expedient of assigning a color to every particle [79]. A rapid quench to
temperatures below the spinodal was then mimicked by instantaneously
increasing the repulsion between unlike particles to a value larger than that
between like particles. The scaling law for the growth of phase-separated
domains with time agreed well with the theoretical predictions for systems
with hydrodynamic interactions [79].
Polymers, the topic of this book, have been simulated by introducing
springs between the DPD particles, as in a coarse-grained bead-spring model
[10,11]. The results of these simulations are discussed in detail below. In
a related area, the rich phase behavior of aqueous surfactants has been
characterized; these molecules were modeled as a rigid unit of one
hydrophobic and one hydrophilic particle [12]. Even cell membranes [13]
and vesicles [14] have been simulated, using surfactants with flexible tails.
Thermotropic liquid crystalline phases have been characterized for a
system of rigid rods [15]. A list of all articles relating to DPD can be
found in [16].
In Section II we look more closely at the computational aspects of DPD,
before focusing attention on the specific application to polymer systems.
Section III describes the matching of simulation parameters to the
properties of real polymer systems, with an emphasis on the relation
between the conservative force field and the common FloryHuggins 
parameter for mixtures. The dynamics of individual polymer chains in a
solvent and in a melt are discussed in Section IV, and the ordering dynamics
of quenched block copolymer systems is described in Section V. A summary
and conclusions are given in Section VI.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dissipative Particle Dynamics

561

II. DISSIPATIVE PARTICLE DYNAMICS

The interactions between DPD particles consist of three parts, each of which
locally conserves momentum by conforming to Newtons third law. The
conservative force between any pair of neighboring particles reads
f cij a!c rij ^rij

where rij |rij| and r^ ij rij =rij are, respectively, the length and the direction
of the difference vector rij ri
rj. Here !c describes the distance dependence of the force, which is usually taken to be !c(r) (1
rij/rc), with rc the
cutoff distance of the potential. The parameter a controls the magnitude
of the force. In the case of polymers an additional conservative force f sij is
added to model the bond connecting the two particles. The force constant
and the equilibrium length of these harmonic or FENE springs are chosen
such that the average bond length coincides with the first peak in the radial
distribution function.
Between every pair of particles, whether they are bound or not, there also
acts a friction force
f dij
!d rij vij  r^ ij ^rij

with vij r_ ij , and a random force

f rij !r rij
ij t^rij

The strength and range of these two forces are related by a conventional
fluctuation-dissipation theorem, which in the current context reads as [17]
2kB T!d rij  2 !2r rij

with Boltzmanns constant kB and temperature T. The common choice for

the distance dependence is !r(r) !c(r). In Eq. (3)
ij(t) denotes a random
number of zero mean, unit variance, uncorrelated between particle pairs,
and devoid of memory, i.e., h
ij ti 0 and


h
ij t
kl t0 i ik jl il jk t
t0
5
DPD codes are usually simple modifications of molecular dynamics (MD)
programs, and since this latter correlation is inconvenient for algorithms
with a finite time step
R nh h, the actual random number to be used in the nth step
becomes ij n n
1h
ij t dt. Clearly, hij ni 0 and hij2 ni h, while

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

562

den Otter and Clarke

there still are no correlations between particle pairs, nor memory effects
between consecutive steps.
Note how the friction and random forces differ from their counterparts
in Brownian Dynamics (BD), where these forces act relative to a fixed
background. In BD, therefore, a bulk flow is not sustained, and the absence
of momentum conservation leads to a screening of the hydrodynamic
interactions. By making all forces explicitly Galilean invariant, as in Eqs. (1)
through (3), these deficiencies are avoided. The friction and random forces,
in addition to their physical interpretation, also serve as a thermostat to the
simulation; see Eq. (4). More elaborate models in which the particles are
equipped with an internal energy have been proposed [18,19], allowing for
a temperature gradient. This may prove a particularly valuable addition
to DPD, as it augments mass and momentum conservation with energy
conservation to obtain the complete set of five hydrodynamical equations
describing any real liquid [4,20]. Particles with an angular momentum
[21,22] and internal elasticity [23] have also been suggested.
The equations of motion, obtained from the above force expressions in
combination with Newtons second law,
dxi
vi
dt
dvi X c

f ij f dij f rij f sij

m
dt
j

6
7

are commonly integrated with a Verlet leapfrog algorithm [1,24]. Because of

the softness of the particle potential, the algorithm is stable up to fairly large
time steps. Unfortunately, the assumption that the friction and random
forces can be treated as constants during the time step leads to a thermostatinduced spurious increase of the temperature [25]. A number of algorithms
have been proposed to counter this effect, with mixed results [2428]. We
think that an Ermak-type approach, the de facto standard in BD, offers the
best performance [28]. But if one is willing to accept a temperature rise of,
say, two percent, it is perhaps most convenient to stick with an ordinary
leapfrog algorithm: the conversion of an existing MD code into a DPD
code is then straightforward. An interesting alternative with an Andersen
thermostat has been described by Lowe [29].

III. PARAMETERIZATION AND RELATION TO

FLORYHUGGINS THEORY
The most common system of units in DPD, as introduced by Groot and
Warren [24], is to define the mass of the particle as the unit of mass, the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dissipative Particle Dynamics

563

cutoff radius of the potential as the unit of length, and the thermal energy
kB T 13 mhv2i i as the unit of energy. The unit of time is then implicitly
defined; a unit of temperature is not needed, since the temperature only
appears as the product kBT. These units, together with a field parameter a
and a friction  of the order ten, suffice for qualitative results, like the role
played by hydrodynamics in the microphase separation of block copolymers. A further tuning is required for more quantitative results. We will
summarize here the method proposed by Groot and Warren [24], including
modifications of recent date [13].
One first chooses nmap, the number of atoms to be represented by a single
DPD particle, thereby fixing the mass of the particle. Next one chooses a
density of  particles per r3c , which combines with the weight density of the
real fluid to give a value for rc. In simulations typical values for  are 3, 4, or
higher, to ensure a sufficiently high number of nearest neighbor interactions
for a fluid-like structure.
The strength of the conservative force field can be obtained from the
requirement that the density fluctuations of the real fluid should be
reproduced. As the static structure factor for a coarse-grained system is
smaller than that of the atomic system by a factor nmap for low wave
numbers, it follows that kB TT DPD n
1
map kB TT atom . The isothermal
compressibility coefficient T of the DPD fluid is readily obtained from its
surprisingly simple equation of state [24], p  kBT 0.1a2, and this set of
equations is then solved for a.
Finally a friction coefficient is chosen, and the unit of time is set by
comparing the self-diffusion coefficient D of a particle with that of nmap
atoms. Typical values obtained this way for water with nmap 3,  3[l
3],
and  3[mlt
3/2] are a 80[ml2t
2] and the units 1[l ] 3 A, 1[t] 90 ps, and
1[m] 9  10
26 kg [13]. As the implications of nmap for the conservative
field have been realized only recently, most simulations have actually been
run using softer a values in the range of 10 to 25.
This procedure for choosing the simulation parameters is by no means
unique. One could, for instance, in the above scheme opt to define the unit
of time by matching the thermal velocity of a particle with that of nmap
atoms. The friction coefficient would then have to be tuned by matching the
diffusion coefficient, using the readily understood observation D
1 
D
1 0 g6kB T=
1 , where both D(0) and g are determined by the
conservative field. Note that there will be no physically meaningful solution
if the conservative force field itself already gives rise to a diffusion
coefficient below the experimental value. An altogether different approach,
using theoretical predictions [22,3032] for the ideal fluid (a 0) to match
dynamical properties such as the viscosity and the sound velocity, has also
been proposed [33].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

564

den Otter and Clarke

Although most simulations to date assume a linear dependence of !c and

!r on r, as introduced in the original DPD for numerical reasons, there is
little evidence to suggest that this is the required form of the potential for a
coarse-grained fluid. Kong et al. [34] tried an alternative simple potential,
but found it to have little effect on the overall results of the simulation.
In this context we mention recent attempts to place DPD on a firmer footing
through a strict definition of a particle as the matter contained in a
smoothed Voronoi cell. Dynamical equations for these cells have been
obtained in bottom-up (starting from the atomic motion [35,36]) and
top-down (starting from hydrodynamics [37]) approaches, in each case
including a fluctuation-dissipation theorem and the aforementioned energy
conservation.
An important parameter in the theory of diluted polymers, polymer
mixtures, and block copolymers is the FloryHuggins  parameter, as it
determines the propensity of the system to mix or de-mix. In DPD, the 
parameter is taken into account by systematically varying the conservative
interaction parameter between unlike particles, whilst the repulsion between
like particles remains unchanged to maintain a homogeneous compressibility. Hence, for a mixture of A and B particles one has aAA aBB and
aAB aAA a. Groot and Warren [24] argued on the basis of the DPD
equation of state that  / a. This was also found by simulating phaseseparated monomeric mixtures at various values of a, using the fraction A
in the B-rich phase and vice versa to calculate the corresponding value of .
Recent simulations using Gibbs ensemble Monte Carlo, to avoid interfacial
perturbations, have confirmed these findings [38]. For polymersolvent
mixtures an identity change Monte Carlo move was introduced,
exchanging a chain of polymer particles by an equal number of solvent
particles. This time the agreement with FloryHuggins was less satisfactory,
particularly near the critical point and in the low polymer concentration
phase. These discrepancies were attributed to the deficiencies of the mean
field model, and an eight parameter master equation was proposed to
describe all simulation data, as well as experimental data [38].
In DPD, as in the FloryHuggins model, the coexistence curve ends at an
upper critical point. Above this critical temperature the system just will not
phase separate. There exist experimental systems, however, that only phase
separate above a critical temperature. In order to simulate a lower critical
point, van Vliet et al. [39] turned the conservative force parameters into
functions of the global temperature and pressure of the system.
Pagonabarraga and Frenkel [40] extended this idea, making
the conservative
P
energy a function of the local density, Ei ex j !rij . The weight
function ! represents the spatial spread-out of the atom cloud of particle j
neighboring particle i. Any phase diagram can now be reproduced by dialing

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dissipative Particle Dynamics

565

in its excess free energy ex as a function of density and temperature. This

also opened up the possibility of studying liquidvapor interfaces with
DPD [41]. Note the similarities between this potential and the one used in
smoothed particle hydrodynamics (SPH), an alternative off-lattice particlebased method for simulating fluid flow [21,42].

IV. ROUSE AND ZIMM DYNAMICS

The first application of DPD to polymers was presented by Schlijper,
Hoogerbrugge, and Manke in 1995 [43]. They simulated a single polymer, a
chain of N particles connected by N
1 harmonic springs, in a solvent of
identical particles. The average bond length was made to coincide with the
location of the first peak in the radial distribution function of the monomer
fluid, rp  0.85 rc, by tuning the strength k and the equilibrium length leq of
the springs. There was no explicit bending potential, so the beads are best
viewed as representing fragments of a polymer at least as big as a Kuhn
segment. The radius of gyration was found to follow the scaling law
hRg i / N
1 for both bond models considered, with N ranging from 2
to 30. At   0:52, the exponent is slightly above the 0.5 characteristic of
a Gaussian chain, but well below the 0.588 of a chain in a good (e.g.,
athermal) solvent [10,11]. Given the repulsive interactions between polymer
beads, the potential energy of two coinciding beads was about five times the
thermal energy in this model, a more pronounced indication of excluded
volume effects was to be expected. The scaling law for the first relaxation
time, 1 / N, yielded   1.95 for the weak bonds (leq 0, small k) and 1.83
for the strong bonds (leq rp, large k). These results lie below the value
of 2 obtained in Rouse theory (in which hydrodynamics is completely
ignored), but above the  3 predicted by Zimm theory (which includes
hydrodynamics in a preaveraging approximation) [10,11], indicating that
hydrodynamic interactions do play a role in the DPD polymer model. On
the basis of this result, the strong spring model appeared to be better than
the weak spring model, hence it was used in successive simulations by these
authors. Later simulations showed a much better agreement with theory,
even for the weak springs, as discussed below.
The strong spring polymer model was further analyzed by looking at the
dynamic structure factors,
*
!
!
+
N
N
X
X
1
ikrj t
ikrj 0
Sk, t
e
e
8
N
j1
j1
For a 20 bead chain, the plot of ln S(k,0) vs. ln k showed a region of linear
decay with a slope corresponding to   0.61 [43]. This is a clear indication

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

566

den Otter and Clarke

of swelling due to excluded volume interactions; the overshoot of the theoretical value is typical for short chains. Scaling theory predicts that for
kRg  1 the rescaled dynamic structure factors S(k, kxt)/S(k, 0) coalesce
for x 3 (Zimm) or x  4 (Rouse) [10,11]. Although the numerical value
of x  2.7 was lower than expected, in the good company of previous polymer simulations, it evidences the impact of hydrodynamic interactions on
the motion of the polymer. A further indication hereof was found in the
distribution of relaxation times in the decay of the conformational time
correlation function,

Ct

N

1X
ri t
Rcom t  ri 0
Rcom 0
N i1

where Rcom denotes the center of mass position. In a nutshell, their approach
was to fit this function with
Ct /

i
2h exp


10

1 i
2h

by varying h and 1. The values h 0 and h 0.5 correspond to the Rouse

and to the Zimm model, respectively. In their appendix, the authors warned
that for short chains, N  10, a noticeable deviation from the long N
predictions of the Zimm and Rouse models is to be expected.
The effect of solvent quality on the conformational distribution of a
solvated polymer was investigated by Kong, Manke, Madden, and Schlijper
[34]. While all interactions between like particles were kept constant,
ass app, the interaction between polymer and solvent beads was varied,
asp ass a. For negative a the polymerpolymer interaction was more
repulsive than the polymersolvent interaction, and the polymers were
observed to swell to   0.61. The opposite case of positive a saw the
polymers collapse to   0.31 (theory: 1/3). A smooth transition between
the two extremes was observed by varying a. Defining the expansion
coefficient
Rg =R
g , with R
g the radius of gyration in the theta solvent,
a qualitative agreement was found with Florys expression,

5

3
p
N
1

A12


theory, A  4=3

3:00:05
a=ass

simulations

11

This suggests a simple linear relationship between  and a, as was also
observed by Groot and Warren [24].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dissipative Particle Dynamics

567

The effect of solvent quality on the dynamics of the polymer was studied
by calculating the conformational time correlation function as a function of
N and a [34]. While the scaling exponent  decreased from nearly 2 in the
good solvent to 0.9 in the bad solvent, the ratio / remained relatively
constant between 3 and 3.3 (theory: 3). A sheared simulation box with
LeesEdwards boundary conditions was used to calculate the impact of a
single dissolved polymer on the viscosity of the box [4446]. The largest
zero-shear viscosities were observed for the good solvent, the lowest for the
bad solvent. Upon increasing the shear rate _ , the solution was found
to shear-thin in the region surrounding _ 1  1, the effect being most
pronounced in the good solvent and nearly absent in the bad solvent. All of
these static and dynamic results, save for the   0.61 from the static
structure factor S(k,0), confirmed the rather unsatisfactory result that the
athermal solvent acted as a near-theta solvent.
The case of a single polymer in a quiescent athermal solvent was recently
revisited by Spenley [47], using chains of 2 to 100 beads in a weak spring
model. For the end-to-end vector he found the exponents  0.58  0.04
and  1.80  0.04, both in excellent agreement with the predictions for a
polymer with excluded volume and hydrodynamic interactions. It is not
clear what caused this marked difference with the above discussed results.
The verification of the scaling law for the self-diffusion coefficient, D / N ,
was more involved, as D is rather sensitive to the size of the simulation box.
After correcting for this complication, a satisfying agreement with theory
was observed.
In the melt [47] the polymers adopted a Gaussian distribution,
 0.498  0.005, as predicted by Flory. The dynamical coefficients,
 1.98  0.03 and
1.02  0.02, and the viscosity  / N, all followed
the Rouse model, showing that the interpolymer interactions were
sufficiently strong to screen the hydrodynamic interactions, despite the
softness of the potential. In fact, the potential was so soft that polymers
were seen to cross one another [47,48]; a transition for large N from Rouse
dynamics to de Gennes reptation is therefore not to be expected. In cases
where the entanglement of polymers is an issue, DPD is readily combined
with recently developed algorithms which explicitly reintroduce the
uncrossability lost in coarse-grained soft potentials [49,50].
We have performed simulations at polymer concentrations ranging from
a dilute system to a melt, for polymers of 20 to 100 beads, to study the
impact of interpolymer interactions on the dynamics [51]. The polymer
solvent repulsion was tuned to a theta point condition, rendering the
exponent  and the radius of gyration at fixed N virtually independent of
the concentration. For the lowest concentrations we found a reasonable
agreement with the Zimm theory of the theta state. When the polymer bead

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

568

den Otter and Clarke

concentration c became of the order of the overlap concentration,

c
N= 43 hRg i3 , the interpolymer contacts became numerous enough to
change the scaling laws into 1 / N  c=c
l and D / N c=c
m . This
behavior is reminiscent of semidilute polymer systems, though the chains
are arguably too short to have reached this limit of combining low
concentration with many interpolymer contacts. At   1.7, l  0.2,

0.65, and m 
0.4, the exponents are much closer to the swollen
state ( and as for the ideal system, l 14 and m
12) than the theta state
(l 1
m), though there was no other direct indication of swelling. Our
theoretical understanding of these phenomena is still incomplete.
A number of articles have discussed polymers with limited freedom of
motion. For a polymer between two parallel walls, the component of the
radius of gyration perpendicular to the wall started to collapse as the gap is
reduced to about five times the unperturbed radius of gyration, while the
relaxation time of this component simultaneously strongly increased [45,52].
A recent study described the forced flow of solvated polymers through a
square tube of comparable dimensions, showing that the polymers had a
preference for the center of the flow [53]. Polymers in a grafted polymer
brush were found to predominantly orient perpendicular to the substrate;
their ordering increased with chain length and surface coverage [54]. The
effects of (oscillatory) shear on a brush have been studied recently [55,56].
Colloidal particles present in the solvent above the brush were absorbed by
the brush when the colloidpolymer interaction was less repulsive than the
colloidsolvent interaction [48,57]. The absorption was effected by the
polymer density and the quality of the solvent with respect to the polymer.
Attaching short polymers to the surface of spherical particles strongly
increased the viscosity and shear thinning of a solution of these particles,
the more so in a good solvent [46]. In a bad solvent the particles even
aggregated [48].

V. BLOCK COPOLYMERS
In an important series of articles the Unilever group have described the
use of DPD to characterize microphase separation phenomena in linear
diblock copolymers of the general formula AnBm [24,58,59], making detailed
comparisons with both experiments and the predictions of self-consistent
field theory [60]. Block copolymers are both interesting and useful because
of their property of forming complicated aggregated structures, as the
natural tendency for segregation of the blocks composed of A and B
monomers is frustrated by the connectivity of the copolymer. The system
can therefore only reduce its free energy by coalescing the A- and B-rich

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dissipative Particle Dynamics

569

domains, forming e.g., sheets, rods, or spheres. The phase diagram is fully
determined by just two parameters, N and the fraction f of A beads.
The basic parameterization used was as discussed above, with  / a.
Groot et al. [24,58] showed that in DPD the interfacial tension between
the component homopolymers followed the same scaling law, as a function of N and , as experimental data for PS-PMMA. This is an important result, since it is the interfacial tension which drives microphase
separation. It also provides validation for the comparison of experimental
data on long chains with DPD data on short chains (but correspondingly
much larger ).
With N 41, the simulations were well outside the weak segregation
regime, and when the melts were quenched the expected order of equilibrium structures was found as the copolymers become successively more
asymmetriclamellae ( f  0.37), perforated lamellae (0.31  f  0.37), hexagonally arranged rods (0.26  f  0.31), fluid peanut-shaped micelles (at
f =0.2), and a disordered phase (at f 0.1). These stable structures emerged
via nontrivial pathways, forming a series of metastable states before
equilibrium is reached. Discrepancies between the exact positions of phase
transition points and those predicted by theory were explained in terms of
the short length of the DPD copolymer chains, N 10, which effectively
reduced the FloryHuggins parameter to N  20 [58]. The finding of
a hexagonally perforated lamellar phase is interesting since this structure
has been identified in recent experiments, while current mean field theory
predicts a gyroid phase in this region. Finite sample size effects cannot,
however, be discounted as a reason for the nonappearance of the gyroid
phase in the simulations.
In a later article [59] the importance of hydrodynamic interactions in
the formation of certain microphases was demonstrated by close comparison
of simulations using DPD and Brownian dynamics (BD). Whilst both
simulation methods describe the same conservative force field, and hence
should return the same equilibrium structure, they differ in the evolution
algorithms. As mentioned above, DPD correctly models hydrodynamic
interactions, whereas these are effectively screened in BD. Distinctly
different equilibrium structures were obtained using the two techniques
in long simulations of a quenched 2400 A3B7 polymer system, as shown in
Fig. 1. Whilst DPD ordered efficiently into the expected state of hexagonal
tubes, BD remained trapped in a structure of interconnected tubes. By
way of contrast, both DPD and BD reproduced the expected lamellar
equilibrium structure of A5B5 on similar time scales, see Fig. 2.
There is experimental evidence [61] that the formation of microphases occurs on at least three different time and length scales. In the first
stage there is phase separation on the level of the interacting beads. This is

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

570

den Otter and Clarke

FIG. 1 Comparison of the time evolution of a quenched A3B7 block copolymer as

obtained by DPD (top three snapshots) and BD (bottom). Shown are the surfaces
surrounding areas of more than 50% A, after 1500, 7500, and 15,000 units of time.
(From Ref. [59].)

R
FIG. 2 The global order parameter Sk ln Sk d 3 k as a function of time for
quenched block copolymers, as obtained in DPD and BD simulations. For A3B7
(HEX) DPD shows an activated transition from a disordered network of tubes to a
hexagonal phase of rods, whereas BD remains stuck in the former. The A5B5 (LAM)
immediately form a lamellar structure in both DPD and BD. (From Ref. [59].)

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dissipative Particle Dynamics

571

followed rapidly by a second stage in which the initially formed clusters

grow out into micelles or highly connected rod-like or planar networks,
depending on the composition. Finally there is the much slower macroscopic ordering of these percolating structures into a symmetric
superstructure, which in the case of A3B7 may occur via a nematic-like
transition.
Groot et al. [59] rationalized the different behavior of the A3B7 and A5B5
copolymer in terms of different evolution mechanisms for the final slow
ordering process. They argued that the formation of hexagonal rods
appeared to proceed by a nucleation-and-growth mechanism, starting from
the gyroid-like and therefore presumably metastable disordered network
of tubes formed in the second stage. The hydrodynamical interactions are
essential in the growth of the hexagonally ordered structure. In BD, by
contrast, hexagonal domains were formed locally, but instead of growing to
encompass the entire simulation box, they disappeared again. The lamellar
phase is formed by spinodal decomposition (i.e., this process does not
involve surmounting a free energy barrier), and thus does not rely as much
on hydrodynamics for its growth. Interestingly, DPD simulations of large
boxes with short chains (A1B3 and A2B2) grew a patchwork of ordered
structures of limited dimensions. This was also observed experimentally;
macroscopic ordering is achieved only in sheared systems.
Recently the ordering effects of steady state shear flow on microphase
separated copolymer melts has been investigated using DPD and the
LeesEdwards boundary conditions for Couette flow [62]. A 150 A3B17 melt
transformed under high shear, _ 1 5:0, from a bcc micellar structure into
cylinders aligned along the flow direction. Cylinders of A6B14 aligned with
the flow at moderate shear, _ 1 0:5, and were torn into lamellae at higher
shear rates. These lamellae, as well as those formed by the symmetric
A10B10, were oriented perpendicular to the neutral direction. All these
systems showed shear-thinning of the viscosity.

VI. CONCLUSIONS
Dissipative particle dynamics has proved to be a very versatile simulation
technique for complex flows on the mesoscopic scale, where thermodynamics, hydrodynamics, and thermal noise play important roles. The
polymer simulations reviewed here cover but a part of the phenomena that
have been, and will be, studied using DPD.
The conservative force field in DPD is extremely soft compared to those
used in atomic simulations, yet after parameterization using the compressibility or the FloryHuggins parameter it is able to capture many of the

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

572

den Otter and Clarke

essential properties of polymers at the mesoscopic level. Thermodynamic

phenomena such as the swelling of polymers in a good solvent and the phase
diagram of block copolymers are thus correctly reproduced. But DPDs
most remarkable feature is that it faithfully combines thermodynamics with
fluctuating hydrodynamics. This is achieved by expressing all forces,
including the Brownian forces that thermostat and randomize the system,
as pair interactions between the particles. Numerous simulations have
confirmed the presence and importance of hydrodynamic interactions.
In the context of polymers, we mention the observed Zimm dynamics of
solvated polymers and the insights DPD has given in regard to the pathways
followed by microphase separating block copolymer systems, including the
role played therein by hydrodynamics.

REFERENCES
1.
2.
3.
4.
5.
6.
7.
8.
9.

10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.

Hoogerbrugge, P.J.; Koelman, J.M.V.A. Europhys. Lett. 1992, 19, 155.

Koelman, J.M.V.A.; Hoogerbrugge, P.J. Europhys. Lett. 1993, 21, 363.
Warren, P.B. Curr. Opin. Colloid Interface Sci. 1998, 3, 620.
Espanol, P. Phys. Rev. E 1995, 52, 1734.
Espanol, P. Phys. Rev. E 1996, 53, 1572.
Boek, E.S.; Coveney, P.V.; Lekkerkerker, H.N.W.; vander Schoot, P. Phys.
Rev. E 1997, 55, 3124.
Coveney, P.V.; Novik, K.E. Phys. Rev. E 1996, 54, 5134.
Coveney, P.V.; Novik, K.E. Phys. Rev. E 1997, 55, 4831.
Jury, S.I.; Bladon, P.; Krishna, S.; Cates, M.E. In High Performance
Computing; Allen, R.J. et al. Ed.; Kluwer Academic: New York, NY, 1999;
185192 pp.
Doi, M.; Edwards, S.F. The Theory of Polymer Dynamics; Oxford University
Press: Oxford, UK, 1998.
Grosberg, A.Y.; Khokhlov, A.R. Statistical Physics of Macromolecules; AIP
Press: New York, NY, 1994.
Jury, S.; Bladon, P.; Cates, M.; Krishna, S.; Hagen, M.; Ruddock, N.;
Warren, P. Phys. Chem. Chem. Phys. 1999, 1, 2051.
Groot, R.D.; Rabone, K.L. Biophys. J. 2001, 81, 725.
Yamamoto, S.; Maruyama, Y.; Hyodo, S. J. Chem. Phys. 2002, 116, 5842.
Sunaidi, A.; den Otter, W.K.; Clarke, J.H.R. In preparation.
URL http://www.sfun.uned.es/mripoll/Invest/dpd.html.
Espanol, P.; Warren, P. Europhys. Lett. 1995, 30, 191.
Espanol, P. Europhys. Lett. 1997, 40, 631.
Bonet Avalos, J.; Mackie, A.D. Europhys. Lett. 1997, 40, 141.
Ripoll, M.H.E.M.; Espanol, P. J. Chem. Phys. 2001, 115, 7271.
Espanol, P. Europhys. Lett. 1997, 39, 605.
Espanol, P. Phys. Rev. E 1998, 57, 2930.
ten Bosch, B.I.M. J. Non-Newtonian Fluid Mech. 1999, 83, 231.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dissipative Particle Dynamics

573

24. Groot, R.D.; Warren, P.B. J. Chem. Phys. 1997, 107, 4423.
25. den Otter, W.K.; Clarke, J.H.R. Int. J. Mod. Phys. C 2000, 11, 1179.
26. Pagonabarraga, I.; Hagen, M.H.J.; Frenkel, D. Europhys. Lett. 1998,
42, 377.
27. Besold, G.; Vattulainen, I.; Karttunen, M.; Polson, J.M. Phys. Rev. E 2000, 62,
R7611.
28. den Otter, W.K.; Clarke, J.H.R. Europhys. Lett. 2000, 53, 426.
29. Lowe, C.P. Europhys. Lett. 1999, 47, 145.
30. Marsh, C.A.; Backx, G.; Ernst, M.H. Phys. Rev. E 1997, 56, 1676.
31. Masters, A.J.; Warren, P.B. Europhys. Lett. 1999, 48, 1.
32. Espanol, P.; Serrano, M. Phys. Rev. E 1999, 59, 6340.
33. Dzwinel, W.; Yuen, D.A. Int. J. Mod. Phys. C 2000, 11, 1.
34. Kong, Y.; Manke, C.W.; Madden, W.G.; Schlijper, A.G. J. Chem. Phys. 1997,
107, 592.
35. Flekky, E.G.; Coveney, P.V.; Fabritiis, G.D. Phys. Rev. E 2000, 62, 2140.
36. Espanol, P.; Serrano, M.; Zuniga, I. Int. J. Mod. Phys. C 1997, 8, 899.
37. Serrano, M.; Espanol, P. Phys. Rev. E 2001, 64, 046115.
38. Wijmans, C.M.; Smit, B.; Groot, R.D. J. Chem. Phys. 2001, 114, 7644.
39. van Vliet, R.E.; Hoefsloot, H.C.J.; Hamersma, P.J.; Iedema, P.D. Macromol.
Theory Simul. 2000, 9, 698.
40. Pagonabarraga, I.; Frenkel, D. J. Chem. Phys. 2001, 115, 5015.
41. Pagonabarraga, I.; Frenkel, D. Mol. Sim. 2000, 25, 167.
42. Monaghan, J.J. Annu. Rev. Astron. Astrophys. 1992, 30, 543.
43. Schlijper, A.G.; Hoogerbrugge, P.J.; Manke, C.W. J. Rheol. 1995, 39, 567.
44. Schlijper, A.G.; Manke, C.W.; Madden, W.G.; Kong, Y. Int. J. Mod. Phys. C
1997, 8, 919.
45. Kong, Y.; Manke, C.W.; Madden, W.G.; Schlijper, A.G. Tribol. Lett. 1997,
3, 133.
46. Zhang, K.; Manke, C.W. Mol. Sim. 2000, 25, 157.
47. Spenley, N. Europhys. Lett. 2000, 49, 534.
48. Gibson, J.B.; Zhang, K.; Chen, K.; Chynoweth, S.; Manke, C.W. Mol. Sim.
1999, 23, 1.
49. Terzis, A.F.; Theodorou, D.N.; Stroeks, A. Macromol. 2000, 33, 1397.
50. Padding, J.T.; Briels, W.J. J. Chem. Phys. 2001, 115, 2846.
51. den Otter, W.K.; Ali, I.; Clarke, J.H.R. In preparation.
52. Kong, Y.; Manke, C.W.; Madden, W.G.; Schlijper, A.G. Int. J. Thermophys.
1994, 15, 1093.
53. Willemsen, W.M.; Hoefsloot, H.C.J.; Iedema, P.D. J. Stat. Phys. 2002, 107,
53.
54. Malfreyt, P.; Tildesley, D.J. Langmuir 2000, 16, 4732.
55. Irfachsyad, D.T.D.; Malfreyt, P. Phys. Chem. Chem. Phys. 2002, 4, 3008.
56. Wijmans, C.M.; Smit, B. Macromolecules 2002, 35, 7138.
57. Gibson, J.B.; Chen, K.; Chynoweth, S. J. Colloid Interface Sci. 1998, 206,
464.
58. Groot, R.D.; Madden, T.J. J. Chem. Phys. 1998, 108, 8713.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

574

den Otter and Clarke

59. Groot, R.D.; Madden, T.J.; Tildesley, D.J. J. Chem. Phys. 1999, 110, 9739.
60. Matsen, M.W.; Schick, M. Curr. Opin. Colloid Interface Sci. 1996, 1, 329.
61. Balsara, N.P.; Garetz, B.A.; Newstein, M.C.; Bauer, B.J.; Prosa, T.J.
Macromolecules 1998, 31, 7668.
62. Zhang, K.; Manke, C.W. Comput. Phys. Comm. 2000, 129, 275.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

18
Dynamic Mean-Field DFT Approach
for Morphology Development
A. V. ZVELINDOVSKY, G. J. A. SEVINK, and
J. G. E. M. FRAAIJE Leiden University, Leiden, The Netherlands

I.

INTRODUCTION

For a long time, chemical engineers have analyzed macroscale properties

using a variety of continuum mechanics models. In the last decade
molecular modeling has grown to an essential part of research and
development in the chemical and pharmaceutical industries. Despite the
considerable success of both molecular and macroscale modeling, in the past
few years it has become more and more apparent that in many materials
mesoscale structures determine material properties to a very large extent.
Mesoscale structures are typically of the size of 10 to 1000 nm. The
industrial relevance of mesoscale modeling is obvious, nevertheless the
necessary general purpose computational engineering tools are absent.
We are developing a general purpose method for mesoscale softcondensed matter computer simulations, based on a functional Langevin
approach for mesoscopic phase separation dynamics of complex polymer
liquids. This project aims to consider topics of utmost importance in
chemical engineering, such as chemical reactions, convection and flow
effects, surfaces and boundaries, etc.
The morphology formation in complex liquids has been studied by many
authors using time dependent LandauGinzburg models [15]. These
models are based on traditional free energy expansion methods (Cahn
Hilliard [6], OonoPuri [7], FloryHugginsde Gennes [8]) which contain
only the basic physics of phase separation [9] and are not well suited for
specific application to the different complex industrial and biological
liquids. In contrast to these phenomenological theories we use dynamic
density functional theory [914] where we do not truncate the free energy
at a certain level, but rather retain the full polymer path integral by a

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

576

Zvelindovsky et al.

numerical procedure (see Appendix A). Recently Kawakatsu and Doi

started to use a similar approach [15,16].
Although calculation of polymer path integrals is computationally very
intensive, it allows us to describe mesoscopic dynamics of specific complex
polymer liquids [17].
Our approach uses essentially the same free energy functional as in many
self-consistent field (SCF) calculations [1821] (see also [22] for an excellent
overview). In the past decade the SCF methods have been widely used
for static calculations in various polymeric systems. The dynamic DFT
approach complements the static SCF calculations by providing a detailed
dynamical picture of system evolution, which is crucial, for example, for the
investigation of metastable intermediate mesophases and for the study of the
influence of shear and other processing conditions.
The prediction of the mesoscale morphology of complex polymer systems
is very important for the final product properties. The application area of
the proposed method includes computer simulation of such processes as
emulsion copolymerization, copolymer melts and softened polymer melts,
polymer blends, polymer surfactant stabilized emulsions, and adsorption
phenomena in aqueous surfactant systems.
In this chapter we demonstrate only several types of modulation of selfassembly in complex polymer systems. They are: shearing of concentrated
aqueous solution of amphiphilic polymer surfactant, shearing of symmetric
diblock copolymer blend, reactions in polymer mixture, surface directed
phase separation in copolymer melt.

II. DYNAMIC DENSITY FUNCTIONAL THEORY

We give a short outline of the theory of the mesoscopic dynamics
algorithms. For more details see [11]. We consider a system of n Gaussian
chains of N beads of several different species (for example, ANA BNB ,
N NA NB for a diblock copolymer, ENE PNP ENE , N NP 2NE N
NP 2NE for a symmetric triblock copolymer, etc.). The solvent can easily
be taken into account [17]. The volume of the system is V. There are
concentration fields I(r), external potentials UI(r), and intrinsic chemical
potentials I(r).
Imagine that, on a course-grained time scale, there is a certain collective
concentration field I(r) of the beads of type I (say, A or B). Given this
concentration field a free energy functional F [] can be defined as follows:
F
n ln  ln n!


XZ

UI rI rdr F nid 

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dynamic Mean-Field DFT

577

Here  is the partition functional for the ideal Gaussian chains in the
external field UI, and F nid[] is the contribution from the nonideal interactions. The free energy functional (1) is derived from an optimization criterion (see [11]) which introduces the external potential as a Lagrange
multiplier field. The external potentials and the concentration fields are
related via a density functional for ideal Gaussian chains:

I U r n

N
X

K
Is0 Trc r
Rs0

s0 1
0
Here K
Is0 is a Kronecker delta function with value 1 if bead s is of type I and
0 otherwise. The trace Trc is limited to the integration over the coordinates
of one chain

Z
Trc  N


VN

N
Y

dRs

s1

N is a normalization constant.
tion function

is the single chain configuration distribu-



1
 H G PN Us Rs
s1
e


where HG is the Gaussian chain Hamiltonian

H G

N
3 X
Rs
Rs
1 2
2
2a s2

with a the Gaussian bond length parameter. The density functional is bijective; for every set of fields {UI} there is exactly one set of fields {I}. Thus
there exists a unique inverse density functional UI[] . There is no known
closed analytical expression for the inverse density functional, but for our
purpose it is sufficient that the inverse functional can be calculated
efficiently by numerical procedures (see Appendices A and B).
We split the nonideal free energy functional formally into two parts
F nid  F c  F e 

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

578

Zvelindovsky et al.

where F e contains the excluded volume interactions, and F c the cohesive

interactions. The intrinsic chemical potentials I are defined by the
functional derivatives of the free energy:
F
F c
F e

UI r

I r
I r I r
c

UI r I r eI r

I r 

5
6

Here we have introduced the cohesive potential cI r and the excluded

volume potential eI r. For the cohesive interactions we employ a twobody mean field potential:
1X
F 
2 IJ
c

cI r 

Z Z

IJ jr
r0 jI rJ r0 dr dr0

F c X

I
J

Z
V

IJ jr
r0 jJ r0 dr0

7
8

where IJ(|r
r0 |) JI(|r
r0 |) is a cohesive interaction between beads of

type I at r and J at r0 , defined by the Gaussian kernel

3
2a2

IJ jr
r0 j  0IJ

3=2

h
i
2
exp
3=2a2 r
r0

The excluded volume interactions can be included via the correction factor
or insertion probability for each bead, c [12]:
F e 

XZ
I

I rlncr dr

10

The insertion probability is interpreted as the effective fraction of free space.

The lower the fraction of free space, the lower the insertion probability and
the higher the excess free energy. We have studied several models (van der
Waals, FloryOrwollVrij, CarnahanStarling) for the excess free energy
function. We found that the phenomenological Helfand penalty function
[last term in Eq. (13)] provides a numerically and mathematically simple
way to account for compressibility effects in the system [12]. In equilibrium
I(r) is constant; this yields the familiar self-consistent field equations for
Gaussian chains, given a proper choice for F nid. When the system is not in
equilibrium the gradient of the intrinsic chemical potential
r I acts as a

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dynamic Mean-Field DFT

579

thermodynamic force which drives collective relaxation processes. When the

Onsager coefficients are constant the stochastic diffusion equations are of
the following form
@I

r  JI
@t
JI
JI
Mr I e

11
12

where M is a mobility coefficient and e

JI is a noise field, distributed according to the fluctuation-dissipation theorem.
In Appendix B we provide some details of the numerical solution of these
equations.

III. APPLICATION
A. Pluronics in Water Mixtures
The aqueous solution of triblock copolymer Pluronic L64 EO13PO30EO13
[23] is modeled as a compressible system consisting of ideal Gaussian chain
molecules in a mean field environment. The free energy is a functional of a
set of particle concentrations {} [11,12]:
Z
n
pp ns s X


UI rI r dr
Ffg
kT ln
np !ns !
V
I
Z
1X

IJ r
r0 I rJ r0 dr dr0

2 I, J V 2
!2
Z X
H
0

I I r
I dr
2 V I

13

where np (ns) is the number of polymer (solvent) molecules,  is the intramolecular partition function, I is a component index (EO, PO, or solvent),
and V is the system volume. The external potential UI is conjugate to the
particle concentration I via the Gaussian chain density functional (2).
The average concentration is 0I and I is the particle volume. The cohesive
interactions have kernels "IJ (7) [11]. The Helfand compressibility parameter
is H [12].
The dynamics of the system is governed by the diffusion-convection
equation (11) with the periodic boundary conditions as described in [11,17].
The dynamic equations are closed by the expression for the free energy (13)
and the Gaussian chain density functional (2). Seven dimensionless

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

580

Zvelindovsky et al.

parameters enter the numerics: three exchange parameters IJ 

=220IJ
0JJ
0II , the dimensionless time  kTMh
2t (h is the mesh
size), a noise scaling parameter  ( 100 in the simulation) [13], the grid
scaling d  ah
1 1.1543 [a is the Gaussian chain bond length (4)] [14], and
the compressibility parameter 0  H 10. For simplicity we use identical
mobility coefficients and particle volumes for all components.
From comparison of Random Phase Approximation (RPA) [24] and
molecular force-field single chain structure factors (generated by Monte
Carlo), we found that the structure factor is well represented by an E3P9E3
Gaussian chain [17], which corresponds to 3 to 4 monomers per bead. The
solvent molecule is represented as a single bead.
The solventpolymer interaction parameters were calculated from vapor
pressure data of aqueous homopolymer solutions [25], using the Flory
Huggins expression [26] IJ

2{ln p/p0
ln (1

)
(1
1/N)
}, where p
is the vapor pressure and
is the polymer volume fraction. The chain length
N was determined using 13=3 (EO) or 30=9 (PO) monomers per bead. This
gives for the interaction parameters ES 1.4, PS 1.7 (here S denotes
solvent). For the EOPO interaction parameter from group contribution
methods [27] we estimated EP 3.0.
We have simulated the time evolution of the Pluronicwater mixture in a
cubic box 64  64  64. The simulation was started from the homogeneous
solution. The resulting morphologies are represented by Figs. 1 and 2.
Four phases are clearly observed in good agreement with the experiments
[23]: lamellae, bicontinuous (gyroid), hexagonal cylindrical, and micellar.
The gyroid phase appeared to be metastable, which will be shown in
Section III.C.

B. Multicolor Block Copolymers

Multicolor block copolymers are a very fascinating example which have
recently become a hot topic in polymer simulation [21]. We demonstrate
just one example obtained by our method. Figure 3 shows mesophases in a
three-color quadrublock copolymer A3B3C3A3. The side chain block A
forms lamellae and middle blocks C and D form two grids of alternating
cylinders. All blocks have equal interaction strength with each other,
AB AC BC 5.4. Our 3D results correspond well to the 2D simulation
of [21] but give a more realistic picture, being 3D results.
An even more sophisticated case is demonstrated in Fig. 4, a four-color
four arm star block copolymer. Again, all interactions are taken of the same
strength. The system develops into four interpenetrating micellar grids with
a deformed dodecahedral space arrangement.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dynamic Mean-Field DFT

581

(a)

(b)

(c)

(d)

0.10

0.25

FIG. 1 Isosurface representation of PL64 in water mixture for 70% (a), 60% (b),
55% (c), and 50% (d) polymer content. The isolevel is
EO EO 0.3.

FIG. 2 Detail of the simulation snapshots from Fig. 1. (a) LAM, (b) GYR, (c)
HEX, and (d) micellar phases are clearly visible.

C. Modulation by Shear
It is known that flow fields affect mesoscale structures in complex liquids
and polymer systems, giving rise to global orientation [28,29]. Because of its

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

582

Zvelindovsky et al.

FIG. 3 Isosurface representation of the A3B3C3A3 melt (a). The system develops
into one lamellar (b) and two cylindrical (c,d) mesostructures.

FIG. 4 Isosurface representation of the star block copolymer (A)8(B)8(C)8(D)8 melt.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dynamic Mean-Field DFT

583

industrial importance, the behavior of complex polymer liquids under shear

is an intensively studied topic, both experimentally and theoretically [2932].
Most of the work is dedicated to the stability analysis of patterns and to
phase transitions in complex polymer liquids under shear.
The time evolution of morphologies in complex liquids under shear
was also studied in computer simulations using Landau type free energies in
2D geometries such as a square cell [35,28], a rectangle [1,2], and in a
two-dimensional Couette flow [1]. Recently the shear effect in a 2D polymer system was studied using the path integral formalism for the kinetic
coefficient [15]. However, this was carried out for a model with simple
periodic boundary conditions and a conventional phenomenological free
energy.
The time evolution of the density field I(r) under simple shear
flow, vx _ y, vy vz 0, can be described by dynamic Eq. (11) with a
convective term
_I MI r  I r I
_ yrx I I
where _ is the shear rate (the time derivative of the strain g). For an
L  L  L cubic grid we use a sheared periodic boundary condition [33,34]:
x, y, z, t x iL jL, y jL, z kL, t
Figure 5 illustrates the application of our method for a 3D melt of block
copolymers A8B8 under simple steady shear flow. Applying shear speeds up
the lamellar formation in a diblock copolymer melt enormously. The
alignment qualitatively differs from the 2D caseso-called perpendicular
lamellae are formed in 3D. From experiments and stability analysis this
orientation is well known to be the most stable one (see, e.g., [31]). The
structure remains stable after switching off the shear.
Figure 6 demonstrates the formation of a perfectly aligned hexagonal
cylindrical phase in aqueous solution of 55% triblock copolymer Pluronic
L64 (cf. Figs. 1c and 2c).
An even more spectacular effect of shearing can be seen in Fig. 7. The
gyroid structure of 60% PL64 from Figs. 1b and 2b was taken as a starting
structure before applying shear. Shearing breaks the system into coexistence of two phases, lamellae and cylinders, Fig. 7 (top). Cessation of
shear does not bring the system back, but it stays as a coexistence of two
phases. This point of the phase diagram is now in perfect agreement with
the experiment [23]. The gyroid phase for this particular system was a
metastable one [32].

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

584

Zvelindovsky et al.

FIG. 5 Isosurface representation of the A8B8 melt at 75,000 and shear flow
vx _ y, vy vz 0, using the structure at 500 as a starting structure. The isolevel
is
A A 0.3. One can clearly observe the global lamellar orientation. The x- and
y-axes are indicated.

FIG. 6 Mesophases of 55% Pluronic L64 in water under shear at time 12,500,
using the structure at 2500 from Figs.1c and 2c as a starting structure before
shearing. The isosurfaces are at
P  PP 0.33. The x-axis is the velocity direction,
the y-axis is the velocity gradient direction, and the z-axis is a neutral direction.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dynamic Mean-Field DFT

585

FIG. 7 Mesophases of 60% Pluronic L64 in water at the end of shearing (top) and
after a long relaxation after cessation of shear (bottom).

D. Modulation by Reactions
In dynamic mean field density functional theory, the dynamics of the
polymer melt under simple steady shear is described by the time evolution of
density fields I. The dynamics is governed by a diffusion-convection
equation with sheared periodic boundary conditions and can readily be
extended with (second order) reaction terms in the following manner:
N
X
@I
F
Mr  I r

_ yrx I
kJK J K I
I
@t
J1, K1

14

Here M is a mobility parameter, _ is the shear rate, which is zero if no shear

is applied, I is a stochastic term which is distributed according to a fluctuation-dissipation theorem [13], and kJK is the reaction rate, which can be
either negative for reactants or positive for products. Notice that the reactive noise can be neglected here. Different order reactions or multiple reaction terms can be added without any difficulties, but as a proof of principle
we focus here on the above type of reactions. In this subsection we study the
effect of combined microphase separation, shear, and reaction to gain
insight in the mechanisms that are important in pathway-controlled morphology formation and in particular in reactive blending.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

586

Zvelindovsky et al.

We therefore study a reaction in which a homopolymer A8 couples to a

homopolymer B8 to form a diblock copolymer A8B8. The reaction is limited
to the end groups and we assume that if two blocks that can couple are close
enough, the reaction takes place. This reaction can be modeled as follows:
@hA
@t
@hB
@t
@dA
@t
@dB
@t

F

_ yrx hA
khA hB hA
hA
F
Mr  hB r

_ yrx hB
khA hB hB
hB
F
Mr  dA r

_ yrx dA khA hB dA
dA
F
Mr  dB r

_ yrx dB khA hB dB
dB
Mr  hA r

15

Here, hA (hB) is the density of A (B) beads in the homopolymer, dA (dB)
is the density of A (B) beads in the diblock copolymer, and k is the reaction
rate of the coupling reaction.
Figure 8 gives an example of formation of double layer droplets in a
sheared reactive polymer system. Initially the A/B blend was subject to shear
which resulted in the formation of elongated droplets. Then reaction on the
surface of the droplets took place after switching off the shear. That leads to
relaxation of the elongated shape of the droplets towards spherical. The
excess of polymer formed at the interface goes inside the droplets which
forms a double layer structure.

E. Modulation by Geometry Constraints

The polymer melt is modeled as a compressible system, consisting of Gaussian chain molecules in a mean field environment. The free energy functional
for copolymer melts has a form that is similar to the free energy that was
used before:
Z
Z
n X
1X
Ffg
kT ln

UI rI rdr
IJ r
r0 I rJ r0 drdr0
2
n!
2
V
I
I,J V
Z
1X
0
0

IM r
r I rM r drdr0

2 I V2
!2
Z X
H

I I r
0I dr
16
2 V I
except for an extra fourth term that contributes only in the direct vicinity of
the filler particles. This accounts for the interaction of a polymer melt with

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dynamic Mean-Field DFT

587

FIG. 8 Three-dimensional simulation of a homopolymer blend 90%/10% A8/B8.

Before the reactions were switched on, 28,000 steps of shear were performed
( _ 0:001) on the blend. The shear was stopped at 28,000 and 2500 reaction
steps were performed ( k 0.1). In the figure the isodensity surfaces at different
levels are depicted at 30,500 of hB, dA, and dB. The total volume fractions at
this time level are 80.5% A8, 0.5%B8, and 19%A8B8.

surfaces. In this equation, n is the number of polymer molecules,  is the

intramolecular partition function for ideal Gaussian chains in an external
field U, I is a component index, I are the density fields of the different bead
types I, and V is the system volume. Inside the filler particles, the densities I
of the different bead types are equal to zero. Since the density  is present in
all integrals in the definition of the free energy (16), integrals over the entire
volume V are equal to the integrals restricted to V/V0, standing for the total
volume V with exception of the volume taken by the filler particles, denoted
as V0. The filler particles considered here are constrained to the condition of
stationary position in time. The constant density field M (M represents
beads of the filler particle type) that appears in Eq. (16) is defined as
M(r) 1 for r 2 V0 and M(r) 0 for r 2 V/V0. The average concentration
is 0I and I is the particle volume. The surface interactions have kernels IM.
The Helfand compressibility parameter is H [12].
The ensemble average particle density s(r) of a certain bead s at position
r in space is
Z
R1 , . . . , RN r
Rs dR1 , . . . , dRN

s U r CMr
VN

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

17

588

Zvelindovsky et al.

where C is a normalization constant and a mask field M(r) is used that is

defined as

Mr

0
1

r 2V 0
r 2V=V 0

The time evolution of the density field I(r) can be described by a time
dependent LandauGinzburg type equation (11). The boundary conditions
that are used on the simulation box are periodic boundary conditions. For
the diffusion flux in the vicinity of the filler particles, rigid-wall boundary
conditions are used. A simple way to implement these boundary conditions
in accordance with the conservation law is to allow no flux through the filler
particle surfaces, i.e.,
r I  n 0

18

where n is the normal pointing towards the filler particle. The same boundary conditions apply to the noise I. Figure 9 demonstrates formation of
lamellar structure in an A8B8 melt in the presence of interacting walls.
In Fig. 10 one can see the same system but confined between neutral
walls. This confinement leads the system to form perpendicular to the
walls lamellae.
Thickness of the slit/film can have a drastic effect on the polymer
morphology. In Fig. 11 the same 55% PL64 in water system was confined
in between two different slits. The bulk morphology of this system is
cylindrical, Figs. 1c, 2c, and 6. Adopting conformational freedom the system
develops into cylinders, which are either parallel or perpendicular to the
wall, depending on the slit width.

IV. DISCUSSION AND CONCLUSION

In this chapter we described the theoretical basis and some applications
of a new model for computer simulation of time evolution of mesoscale
structures of complex polymer liquids. Here we describe just a few possible
technologically oriented applications of the proposed method.
One of the main questions in industrial emulsion polymerizations is to
produce a latex or polymer with any desired morphology, composition,
sequence distribution, molecular weight distribution, etc. A specific example
is given by the core-shell techniques in which one type of (co)polymer is
grown around a core of another type. In paints the core polymer may
provide gloss and mechanical stability, whereas the shell might contain a

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dynamic Mean-Field DFT

589

FIG. 9 Lamellar formation of an A8B8 copolymer melt in the presence of square

plates of one grid-cell thickness. The interaction of polymer blocks with the surface is
AM
1
1.0 and BM
11.0. (a) View of filler particle in simulation box, (b)
space filled with filler particles (the slots between filler particles are drawn as white
lines), (c) morphology of A beads (isolevel A 0.5) in one simulation box at
500, (d) the same for 2000.

rubbery polymer to provide a uniform surface coating. The final

morphology that is obtained in the production process determines the
quality of the product to a very large extent and hence prediction of the
morphology based on production process parameters is desirable. The
morphology is controlled by both thermodynamic (the microphase
separation process) and kinetic (the reaction process) principles. Several
practical applications are within direct reach. An example of a one-stage
core-shell technique that has been described in the literature concerns the
mixing of silicone oils (containing Si-H and vinyl groups) and vinyl
monomers emulsified in water. After a (cross-linking) reaction of Si-H and
Si-CH CH2 (kinetics) the hydrophilic monomers are excluded to the
surface layer (thermodynamics) and a core-shell morphology results.
Another potential application is the investigation of the stability of
polymer surfactant protected emulsions. In particular, many industrial

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

590

Zvelindovsky et al.

FIG. 10 Lamellar formation of an A8B8 copolymer melt in the presence of the same
filler particle as Fig. 9. Moreover, there is no interaction between the polymer beads
and boundaries of the filler particles. (e) View of filler particle in simulation box, (f)
morphology of A beads (isolevel A 0.5) in one simulation box at 500, (g) the
same for 4000, (h) final morphology at 10,000.

FIG. 11 Cylinder formation of 55% PL64 in water mixture in two different slits,
12h (left) and 15h (right), forming parallel (left) or perpendicular (right) cylinders.

systems contain a specified mixture of different surfactants, so as to provide

a certain stability of the interphase. Thus the application of the emulsions is
controlled by the nature of the surfactant mixture. Within the described
method it is relatively easy to study the synergetic effects of surfactant

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dynamic Mean-Field DFT

591

composition and architecture, and the way processing conditions influence

the emulsion stability and morphology.

APPENDIX A: NUMERICAL IMPLEMENTATION OF

THE PATH INTEGRAL
The method presented here is off-lattice in nature. It means that the
calculations are carried out on a (cubic) lattice, but the chain conformations
are not restricted to lattice positions. The Wageningen school [3538] has
shown that for lattice models, where the chains can only occupy lattice
positions, the number of conformations is finite and summation amounts
to an exact enumeration. In this case numerical integration is rapid and
relatively simple. The off-lattice nature of our method introduces an extra
difficulty in deriving quadrature (or numerical integration) rules. Common
rules, such as used in lattice methods, are found to give problems. First,
the conformation restriction inherent in the lattice models can lead to
unphysical singularities in the inverse structure factor for copolymer melts.
As a consequence, the mapping between particle density field and external
potential may have a null space and is no longer bijective (for more details
see [9]). Second, we found that the cubic lattice chain structure factor for
copolymer melts has strong intrinsic anisotropies, especially on bond length
scales. These anisotropies are an important source for a well known
numerical plague of lattice models, so-called lattice-artifacts.
In this appendix we describe a stencil algorithm which avoids many of the
drawbacks of quadrature rules used in classical lattice models, while the
extra computational cost is modest. The derivation consists of finding a
unique and optimal set of stencil coefficients for a convolution with a
Gaussian kernel, adapted to the special case of off-lattice density functional
calculations. Stencil coefficients are the multipliers of the function values at
corresponding grid points.
The familiar Feynman decomposition of the path integral (2) results in
the well known modified diffusion algorithm [39,40]. In our case, where
sequence space is discrete, the density functional can be expressed in terms
of Green propagators
s r / Gs rGinv
s1 r

19

The set of once integrated Greens functions Gs(r) and Ginv

s1 r are related by
the recurrence relations
Gs r Mre
Us r Gs
1 r

Us r
Ginv
Ginv
s r Mre
s1 r

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

20

592

Zvelindovsky et al.

with G0 r Ginv
N1 r 1. The linkage operator  f r is defined as a
convolution with a Gaussian kernel


 Z


3 2=3
 f r
exp
3=2a2 r
r0 2 f r0 dr0
21
2
2a
V
Note that the Green propagator method to calculate the density functional
(2) yields an exact answer.
Now, the crucial point of the algorithm is a numerical representation of
the linkage operator , given a cubic uniform grid with mesh width h, which
is both efficient and accurate. Since we work on large cubic grids and
use domain decomposition for the parallelization, FFTs and Gaussian
Quadrature rules are not considered. Efficiency is crucial, since the linkage
operation has to be repeated 2(N1) times for a single density functional
calculation. Furthermore, if the numerical representation of the linkage
operator is not accurate, its repeated application will result in error
accumulation, especially in the small q range, i.e., on chain length scales.
Large errors in the high q range are also undesirable, since they may disturb
the bijectivity of the mapping between density field and external potential.
A 27-point stencil is the most compact stencil that meets all constraints.
We reduce the number of parameters by invoking symmetry rules, similar to
the polyhedra rules used in multidimensional integration of symmetric
kernels [41]. The linkage operator on a uniform grid is expressed as:
 f r c0 f r

X ci
i

f r ri f r
ri

22

where f is an arbitrary field, ri is a lattice vector, and ci is a stencil coefficient.

It is easy to see that, due to the symmetry of the linkage operator, there is
considerable redundancy in the stencil coefficients. For a 27-point stencil
only four reduced coefficients remain: a central coefficient c0 and three
coefficients for each symmetry-reduced stencil direction; c1 in direction
(1, 0, 0), c2 in direction (1, 1, 0), and c3 in direction (1, 1, 1).
In Fourier space the linkage operator is:
Z
 f e
iqr dr q fq
V

where the Fourier multipliers  q are given by

(
q

e
a jqj =6
P3
Pd i
j1 cosq  rij
i0 ci

continuum
discrete

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dynamic Mean-Field DFT

593

Here, di is the number of independent lattice vectors in symmetry-reduced

stencil direction i (d0 1,d1 3, d2 6, and d3 4), rij denotes the jth lattice
vector in stencil direction i [e.g., r01 0, r12 (0, h, 0), r23 (0, h, h), and
r34 (h,
h,
h)].
First, we guarantee maximum isotropy on the grid by imposing that the
continuum and discrete (grid-restricted) linkage operators are identical for
q 2 f0, 0, 0, =h, 0, 0, =h, =h, 0, =h, =h, =hg
In this way, we transfer the isotropy that should have resulted from an
isotropically chosen grid (for instance, honeycomb) to the stencil. Second,
we guarantee that the long length scalings of the continuum and discrete
Fourier multipliers are identical by also equating the curvatures in the point
q(0, 0, 0). This assures the bijectivity of the density-external potential mapping. Because of symmetry, it is sufficient to equate the two curvatures in
direction (1, 0, 0). With the help of these five conditions, we determine the
four stencil coefficients ci and a/h. The unique solution of the nonlinear
system is: c0 0.171752, c1 0.137231, c2 0.0548243, c3 0.0219025, and
a/h 1.15430. As a result, the stencil depends on the ratio a/h. For other
ratio, we consider different stencil coefficients [9].

APPENDIX B: NUMERICAL SCHEME FOR SOLVING

THE DYNAMICAL EQUATIONS
In Appendix A we described the numerical implementation of the Gaussian
chain density functions that retains the bijectivity between the density fields
and the external potential fields. The intrinsic chemical potentials I F/
that act as thermodynamic driving forces in the GinzburgLandau
equations describing the dynamics, are functionals of the external potentials
and the density fields. Together, the Gaussian chain density functional and
the partial differential equations, describing the dynamics of the system,
form a closed set.
Since the external potentials are highly nonlinear functionals of the
density fields (they cannot even be inverted analytically), the partial
differential equations we have to solve numerically are in themselves highly
nonlinear. We should be very careful in choosing a method, since there are
apparent risks of introducing numerical errors (for instance by taking the
time steps too large). The CrankNicolson (CN) method, that aims to solve
differential equations by mixing implicit and explicit parts, is known to be
rather stable for this kind of problem. We use periodic boundary conditions.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

594

Zvelindovsky et al.

FIG. 12 Schematic representation of the iterative scheme. We have two nested

iterative loops: the time loop (1) for updating the density fields , and within each
time iteration an iterative loop (2) for updating the external potential fields U. We
start with an initial guess for the external potential fields. We use Eq. (17) to generate
a set of unique density fields. The cohesive chemical potential E [relation (3)] can be
calculated from the density fields by Eq. (8). The total chemical potential (4) can
now be found from Eq. 6. We update the density fields (5) [by using the old and
updated fields in Eq. (23)] and accept the density fields if the condition (26) is
satisfied. If this is not the case, the external potential fields are updated by a steepest
descent method.

The way the time integration is carried out is as follows: we start with a
homogeneous distribution of Gaussian chains I  I (where  is the mean
density of the fields) and UI 0. We solve the CN equations (to be defined
below) by an iterative process in which the external potential fields are
updated according to a steepest descent method. The updated external
potential fields are related, by the chain density functional, to unique density
fields. We calculate the intrinsic chemical potential from the set of external
potential fields and density fields and update the density fields. The density
fields are only accepted as an update for the next time step if the L2 norm of
the residual is below some predefined upper boundary. An overview of this
scheme is shown in Fig. 12.
The general expression of the CN equations for this problem is
kI
k
1
k
1
k
I
!1 Mh
2 DkI D kI 1
!1 Mh
2 Dk
1
I D I I
t
23
where k is the time index, h the mesh size, !i the CN parameters determining
the degree of explicitness versus implicitness of the method, and h
2D(fDg)
the discrete operation on the grid representing r  frg. The computational
procedure for the div-gradient operation is again important because of the
risk of introducing grid artifacts. We have developed a procedure similar to
the procedure for the numerical integration. As computational cell, again a

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dynamic Mean-Field DFT

595

27-point stencil is used. Isotropy and scaling behavior should be considered

as conditions for the stencil weights. A more detailed overview of this
scheme can be found in [13].
In dimensionless parameters
I I ( is the bead volume), I  I
( kT), and 
1Mh
2t we can rewrite the CN equations as
k
reskI
Ik
!1  D
Ik DkI

Ik
1
1
!1  D
Ik
1 Dk
1
I
I 0
24

Since we do not have an explicit expression for the external potential fields
in terms of the densities, we solve this equation for the external potential
fields. The external potential fields UI are found by an iterative scheme
UI,k 0 UI,k
1
final
UI,k p UI,k p
1 reskI, p
1 p  1

25

where p is the iteration index and  is a parameter in a bisection method that

is employed in order to find update fields Up along the direction of search
with a smaller residuals norm than the residual of the previous fields Up
1.
The problem is solved for time step k when the norm of the residual satisfies


X



reskI  < 

 I


26

In general, the CN parameter is chosen equal to 1/2. In this case

starting
P the
k
value of the residual norm for every new time iteration is
k

k,
and we
I I
P
consider as a rule of thumb an  such that   0:01k I kI k. When this
condition is satisfied, the updates of the density and external potential
fields are accepted as solutions at time step k and used as starting values
for the new iterative update scheme at time step k 1.

REFERENCES
1. Qiwei He, D.; Nauman, E.B. Chem. Engng. Sci. 1997, 52, 481496.
2. Gonnella, G.; Orlandini, E.; Yeomans, J.M. Phys. Rev. Lett. 1997, 78,
16951698.
3. Patzold, G.; Dawson, K. J. Chem. Phys. 1996, 104, 59325941.
4. Kodama, H.; Komura, S. J. Phys. II France 1997, 7 (1), 714.
5. Ohta, T.; Nozaki, H.; Doi, M. Phys. Lett. A 1990, 145 (6,7), 304308.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

596
6.
7.
8.
9.
10.
11.

12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.

28.
29.
30.
31.
32.
33.
34.
35.

Zvelindovsky et al.
Cahn, J.W.; Hilliard, J.E. J. Chem. Phys. 1958, 28, 258267.
Oono, Y.; Puri, S. Phys. Rev. Lett. 1987, 58, 836839.
de Gennes, P.G. J. Chem. Phys. 1980, 72, 47564763.
Maurits, N.M.; Fraaije, J.G.E.M. J. Chem. Phys. 1997, 106, 67306743.
Fraaije, J.G.E.M. J. Chem. Phys. 1993, 99, 92029212.
Fraaije, J.G.E.M.; van Vlimmeren, B.A.C.; Maurits, N.M.; Postma, M.; Evers,
O.A.; Homann, C.; Altevogt, P.; Goldbeck-Wood, G. J. Chem. Phys. 1997,
106 (10), 42604269.
Maurits, N.M.; van Vlimmeren, B.A.C.; Fraaije, J.G.E.M. Phys. Rev. E 1997,
56, 816825.
van Vlimmeren, B.A.C.; Fraaije, J.G.E.M. Comp. Phys. Comm. 1996, 99, 2128.
Maurits, N.M.; Altevogt, P.; Evers, O.A.; Fraaije, J.G.E.M. Comp. & Theor.
Pol. Sci. 1996, 6, 18.
Kawakatsu, T. Phys. Rev. E 1997, 56, 32403250.
Hasegawa, R.; Doi, M. Macromolecules 1997, 30, 30863089.
van Vlimmeren, B.A.C.; Maurits, N.M.; Zvelindovsky, A.V.; Sevink, G.J.A.;
Fraaije, J.G.E.M. Macromolecules 1999, 32, 646656.
Fleer, G.J., et al. Polymers at Interfaces; Chapman & Hall: London, 1993.
Doi, M. Introduction to Polymer Physics; Oxford University Press: Oxford,
1996.
Matsen, M.W.; Schick, M. Phys. Rev. Lett. 1994, 72, 26602663.
Drolet, F.; Fredrickson, G.H. Phys. Rev. Lett. 1999, 83, 43174320.
Hamley, I.W. The Physics of Block Copolymers; Oxford University Press:
Oxford, 1998.
Alexandridis, P.; Olsson, U.; Lindmann, B. Macromolecules 1995, 28,
77007710.
Leibler, L. Macromolecules 1980, 13 (6), 16021617.
Malcolm, G.N.; Rowlinson, J.S. Trans. Faraday Soc. 1957, 53, 921931.
Hill, T.L. An Introduction to Statistical Thermodynamics; Addison-Wesley:
Reading, MA, 1962.
Krevelen, D.W. Properties of Polymers, Their Correlation with Chemical
Structure; Their Numerical Estimation and Prediction from Additive Group
Contribution; Elsevier: Amsterdam, third edition, 1990.
Kodama, H.; Doi, M. Macromolecules 1996, 29 (7), 26522658.
Fredrickson, G.H. J. Rheol. 1994, 38, 10451067.
Zvelindovsky, A.V.; Sevink, G.J.A.; van Vlimmeren, B.A.C.; Maurits, N.M.;
Fraaije, J.G.E.M. Phys. Rev. E 1998, 57, R4879R4882.
Doi, M.; Chen, D. J. Chem. Phys. 1989, 90, 52715279.
Ohta, T.; Enomoto, Y.; Harden, J.L.; Doi, M. Macromolecules 1993, 26,
49284934.
Chen, Z.R.; Korneld, J.A.; Smith, S.D.; Grothaus, J.T.; Satkowski, M.M.
Science 1997, 277, 12481253.
Zvelindovsky, A.V.; Sevink, G.J.A.; Fraaije, J.G.E.M. Phys. Rev. E 2000, 62,
R3063R3066.
Scheutjens, J.M.H.M.; Fleer, G.J. J. Phys. Chem. 1979, 83, 16191635.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.

Dynamic Mean-Field DFT

597

36. Scheutjens, J.M.H.M.; Fleer, G.J. J. Phys. Chem. 1980, 84, 178190.
37. Leermakers, F.A.M.; Scheutjens, J.M.H.M. J. Chem. Phys. 1988, 89, 3264.
38. Evers, O.A.; Scheutjens, J.M.H.M.; Fleer, G.J. Macromolecules 1990, 23,
52215233.
39. de Gennes, P.-G. Scaling Concepts in Polymer Physics; Cornell University
Press: Ithaca, NY, 1979.
40. Doi, M.; Edwards, S.F. The Theory of Polymer Dynamics; Clarendon: Oxford,
1986.
41. Zwillinger, D. Handbook of Integration; Jones and Barlett Publishers: Boston,
1993.

Copyright 2004 by Marcel Dekker, Inc. All Rights Reserved.