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 effi