You are on page 1of 122

Permeation of Polymers

A Computational Approach

FLORIAN MÜLLER-PLATHE

Habilitationsschrift
Laboratorium für Physikalische Chemie
Eidgenössische Technische Hochschule Zürich
© 1993 Florian Müller-Plathe
All rights reserved.
Table of Contents

1 Introduction 1

2 Foundations 7
2.1 Permeability 7
2.2 Diffusion coefficient 9
2.3 Ionic conductivity 14
2.4 Solubility 16

3 Machinery 19
3.1 Molecular dynamics 19
3.2 Diffusion coefficients 24
3.3 Solubility 26
3.4 Software 27

4 Motion of Gas Molecules through Amorphous Polymers 31


4.1 Hopping motion 31
4.2 Penetrants in cavities 34
4.3 The jump event 37
4.4 Anomalous diffusion 42

5 Gas Diffusion Coefficients 45


5.1 Calculation of diffusion coefficients 45
5.2 Influence of simulation parameters on calculated diffusion coefficients 47
5.3 Temperature and density dependence of diffusion coefficients 49
5.4 Where can molecular dynamics be usefully applied? 51
5.5 Can MD methods be extended to slower diffusion processes? 52

Permeation of Polymers iii


6 Gas Solubilities 59

7 Semicrystalline and Filled Polymers 67


7.1 Two-phase models 68
7.2 A Monte-Carlo strategy 70

8 Ions in Polymers 77
8.1 Ion-conducting polymers: background 77
8.2 Which phenomena can be expected in polymer electrolytes? 80
8.3 Simulations 81
8.4 Interaction sites 82
8.5 Ion association 87
8.6 Ion mobility and conductivity 89

9 Summary and Outlook 99

Appendix: Forcefield for Polyethylene Oxide 105

References 111

iv Florian Müller-Plathe
Foreword

Over time, many individuals have contributed - knowingly or not - to the success
of this work. I would like to thank all of them. Some of them appear below, in no
particular order.

Bill Smith (Daresbury) taught me most of what I know about molecular dynamics
and provided help and encouragement. With Wilfred van Gunsteren (ETH) I had
many discussions about the more complicated aspects of the chemical and the
computational side of the project. His and Herman Berendsen’s coaching during
two instructional visits to their laboratory in Groningen also helped getting me
started in molecular simulation. Some outside collaborators were involved in the
project: Dave Brown (UMIST) more on the computational side, and Steve Rog-
ers (ICI-Runcorn) on the permeation of polyisobutylene in all its aspects. Leif
Laaksonen (Helsinki) was kind enough to interface his molecular graphics soft-
ware to my programs and to even include additional features at my request.
Konrad Singer (Royal Holloway College) always was a backup for sorting out my
mistakes in statistical mechanics. Ueli Suter (ETH) and his group have often
helped with polymer structures and advice. Even though we never collaborated
officially on polymer permeation there are few people with whom I have
exchanged as many ideas about the subject as with Andrei Gusev (ETH), often
by means of heated debates. Wilfred van Gunsteren, Andrew Torda (ETH) and
Bill Smith have made valuable comments on this manuscript.

v
This work has required massive amounts of computer resources. Hence, I was
always grateful if my normal allocations were kindly augmented by somebody.
Apollo Computers (shortly before becoming part of Hewlett-Packard) donated to
Daresbury one of the then world’s-fastest workstations, on which many of the ini-
tial calculations were carried out. Later in Zürich, supercomputer time was pro-
vided generously by the ETH Computer Centre, Cray Research Inc. and the
Centro Svizzero di Calcolo Scientifico. Rudy Wilopo (Cray Research) and Gabri-
ele Jost (NEC) have volunteered valuable help with the fine tuning of the molec-
ular dynamics program.

In its initial stage the project involved a collaboration between Daresbury and ICI
plc. on macromolecular dynamics. ICI did not only provide the funding, but also
have to be credited with gently steering my activities toward synthetic (as
opposed to biological) polymers.

Zürich, October 1993 Florian Müller-Plathe

vi Florian Müller-Plathe
1 Introduction

Ich möchte aus meinen Papieren, von denen ich viele


mit Eckel ansehe, einen Auszug machen.

Ulrich Bräker
“Lebensgeschichte und Natürliche Ebentheuer des armen
Mannes im Tockenburg”
(Orell, Geßner, Füßli & Compagnie, Zürich, 1789)

The permeation of atomic and molecular species through polymers has received
attention from scientists and engineers for a long time, the first experimental
investigations dating back as far as 1831 (after ref. 1). This is hardly surprising,
since permeabilities of polymers are quite important in many areas of technol-
ogy. A few such applications are compiled in Table 1. In particular, the develop-
ment of membrane materials for new separation processes continues also to be
a growth area economically. The worldwide market for membrane technology
was estimated 1 billion US-$ already in 1988 [2]. The desire to cut energy con-
sumption and pollution is an additional driving force for the replacement of clas-
sical separation methods such as distillation by membrane processes [3].

The macroscopic description of polymer permeation was developed already in


1879 by von Wroblewski [4] and was refined in 1920 by Daynes [5] to the point
where it could be used as the theoretical framework for modern methods of mea-
suring quantitatively the transport of matter across polymer membranes (for a
review, see e.g. ref. 6). Also, phenomenological theories describing the tempera-
ture dependence of relaxation and transport processes, like the Arrhenius or Wil-
liams-Landel-Ferry [7] behaviour, are early discoveries.

Microscopic theories, that link solvation and transport of small molecules (“pene-
trants”) in polymers to atomistic models of the polymer-penetrant system, were
developed between 1950 and 1970. These theories have been extensively

1
Introduction

TABLE 1. Some aspects of polymer technology for which permeabilities are


important.

Application Type of penetrant Design goal

packaging gas, moisture high barrier


additive migration plasticisers, dyes, high barrier
etc.
gas separation gases selectivity
pervaporation vapour selectivity
(reverse) osmosis liquid selectivity
analytical chemistry ions very high selectivity
biosensors biomolecules very high selectivity
drug implants pharmaceuticals controlled speed of release
polymer electrolytes ions high ionic conductivity
monomer removal unreacted monomer low barrier

reviewed, see for example, refs. 8-11. Although they have been very successful
in rationalising, for example, the diffusivities of several penetrants in the same
polymer, most have deficiencies, in that they are either not really atomistic or that
they assume a certain molecular behaviour in an ad-hoc manner. An example of
the first is the concept of free volume [12]: Before a molecule can move, free vol-
ume has to be created in its neighbourhood. The amount of free volume avail-
able, in turn, is inferred from macroscopic quantities such as the compressibility,
and not from atomistic considerations. An example of the second is the DiBene-
detto model of diffusion which assumes that chains of an amorphous polymer
are locally aligned and that diffusants move within tubes formed by the chains
[13, 14]. It is no surprise, that none of these models could ever be developed into
a predictive tool that was sufficiently general and accurate to be used to a large
extent in the design of new polymers with specific barrier properties.

The situation has changed to some degree over the last few years. The proce-
dures for molecular simulation of liquids [15, 16] have matured over the years,
and they have been applied successfully to a special class of polymers, namely
biopolymers [17, 18]. At the same time, computers have become powerful and
affordable enough to handle the large spatial dimensions and time scales often
encountered in polymeric systems. Not surprisingly, computer simulation

2 Florian Müller-Plathe
Introduction

TABLE 2. Theoretical and computational techniques for treating transport of


small molecules in different polymer morphologies. For the methods
mentioned in this treatise, the appropriate chapter numbers are given
in parentheses.

amorphous crystalline semicrystalline


diffusion molecular dynamics transition-state Monte Carlo of
coefficient (5), transition-state theory, harmonic 2-phase model
theory (9) analysis, defects? (7)
ionic molecular dynamics ? Monte Carlo
conductivity (8) 2-phase model
(7)
solubility particle insertion (6), harmonic analysis, analytical
static approximation defects? 2-phase model
(9) (7)

became part of the polymer scientist’s tool kit [19, 20] (see, however, ref. 21 for a
very different point of view). The beginning of the decade then saw the first arti-
cles describing molecular simulation being applied to the problem of penetrants
diffusing through polymers [22-24].

It is the intention of this work to describe the rapid progress that has been made
over the past few years in this area. Trying to model, theoretically, the transport
of small molecules or ions through polymer materials one quickly realises that
characteristic time scales and length scales are very different and depend on the
polymer morphology. Table 2 lists the most important cases. It also serves as a
kind of road map to the present treatise, which deals with selected transport phe-
nomena in selected morphologies. It is commonly believed that both sorption
and transport in crystalline polymers are orders of magnitude smaller than in
amorphous polymers. This is mainly due the following reasoning: Only very few
penetrants can enter the tightly packed polymer crystallites in the first place and
those, which do, will get trapped at defects and will hardly be able to move.
Since most of the technologically interesting polymers are either amorphous or
partially crystalline, the amorphous volume fraction will very often dominate
these properties in practice.

Permeation of Polymers 3
Introduction

So far, only sorption and transport in amorphous polymers has been studied at
the atomistic level by molecular simulation. Apart from the most important contri-
bution coming from the amorphous part, there is also a practical reason: The dif-
fusion in the crystallites will be so slow that during the few nanoseconds which
can be simulated by, say, molecular dynamics methods, the diffusant will not
move at all, whence there is little use for molecular simulation. Motion of neutral
and ionic species through amorphous polymers are covered in several chapters
of the present work, both at the qualitative and quantitative level.

How to model the behaviour of small molecules in crystalline polymers, if that


should really become necessary, is presently not clear. Techniques borrowed
from solid-state science might be appropriate (e.g. static simulations, consider-
ation of crystal defects, transition state theory) which, however, still have to be
tried on polymer systems. One can also imagine that migration of defects in poly-
mer crystals will play a role.

The study of semicrystalline materials, on the other hand, is quite important


since crystallites are known to influence the barrier properties of a polymer. It is
possible if the atomistic level of description is abandoned. One chapter is
devoted to a discussion of two-phase models (an amorphous phase with known
diffusivity and solubility and an impermeable crystalline phase) and their imple-
mentations in a Monte Carlo scheme, which can extend the scope of simulation
methods also to semicrystalline and filled materials.

Although most of the examples reported in this review are from our own
research, we try to mention all work in the area which has been reported to date.
The work by others is discussed in more detail, if the results differ significantly
from our own or if they cover aspects that we have not treated, like, for example,
the temperature dependence of diffusion coefficients.

In summary, this review highlights the following topics:


• Diffusion mechanism. Molecular dynamics allows us to look microscopically
at the system as it evolves in time. The system is left to pursue its dynamics
without interference from the outside and, in particular, without any precon-
ceived ideas about the molecular motions. We can therefore expect to gain
insight about the fundamental mechanisms that underlie the diffusion of small
penetrants through polymer materials. In the last chapter, we will also see

4 Florian Müller-Plathe
Introduction

how our qualitative ideas about the diffusion mechanism have to be modified
as one moves from neutral gaseous penetrants to ions.
• Statistical mechanics of diffusion. It has been discovered that penetrant
motion in polymers can be very non-intuitive. In particular, separation of short
and long time scales can occur, which leads to phenomena such as anoma-
lous diffusion.
• Transport coefficients. As stated above, transport of matter across polymers
is inherent in many applications and the transport coefficients (permeability,
diffusion coefficient, conductivity, solubility) are important characteristics of
any polymer material. The calculation of these coefficients with an accuracy
comparable to experiment is therefore a goal for molecular dynamics and
other simulation schemes. We shall assess the results obtained thus far and
examine the computational requirements for successful calculations of trans-
port coefficients.
• Method development. A number of molecular simulation techniques have
been employed in the study of polymer-penetrant systems. Some have been
newly developed, others adapted to accommodate the special needs
encountered in polymeric systems. We shall review the ones that have been
used.
• Finally, we will try to extrapolate where we think the field is going in the near
future.

Permeation of Polymers 5
Introduction

6 Florian Müller-Plathe
2 Foundations

2.1 Permeability

The most frequently measured and reported quantity characterising the barrier
properties of a polymer film or membrane is the permeability P [25]

( quantity of permeant ) × ( film thickness )


P = (Eqn. 1)
( area ) × ( time ) × ( pressure drop across film )

This is the formulation for the permeant being a gas under two different (partial)
pressures on opposite sides of the film. Other expressions for the permeability
appropriate for the permeant being a liquid, solvent or solute can also be formu-
lated. The quantity of gaseous permeant is typically given as volume in cm3 at
standard temperature and pressure (STP: 273.15 K and 1.013×105 Pa), the film
thickness and area in cm and cm2, respectively, time in seconds and pressure in
Pa, so that we have as the unit of P (other units are being used less frequently):

cm 3 ( 273.15K ; 1.013 × 105 Pa ) × cm


unit of P: (Eqn. 2)
cm 2 × s × Pa
A direct molecular simulation of a system corresponding to the relevant part of
the experimental setup for measuring permeabilities (i.e. a membrane of a few
microns plus gas at two different pressures on either side) is impossible: The
system would be far too large for the simulation to be feasible. The process of a

7
Foundations

p1 3
2 p2

1
FIGURE 1 Steps involved in the transport of a penetrant molecule through
a polymer membrane. The penetrant here is a gas of partial
pressures p1 and p2 on the two sides of the membrane. (1)
Absorption of the gas molecule into the surface of the
membrane, (2) diffusion across the membrane, and (3)
desorption from the opposite membrane surface.

gas molecule entering the membrane, diffusing through it and desorbing on the
other side is too slow and the event too rare for the accumulation of reasonable
statistics within a feasible simulation length. Finally, simulation of the sorption/
desorption process requires some knowledge about the atomic structure of the
membrane surface. Very little is known about the detail of surfaces of amor-
phous polymers.

Fortunately, all these problems can be circumvented. It has long been recog-
nised [4] that the permeability of a polymer membrane to a gas is given by

P = DS (Eqn. 3)

where D and S are the diffusion coefficient and the solubility, respectively, of the
penetrant in the bulk polymer. The diffusion coefficient is given in cm2s-1, and the
solubility in cm3 of gas at STP per volume of polymer (cm3) and gas pressure
(Pa).

Eqn. 3 can be rationalised by considering Figure 1. The permeation process can


be broken down into three steps: absorption, diffusion and desorption. Assume
for the moment that the diffusion coefficient describes the flux of gas (in cm3 at

8 Florian Müller-Plathe
Foundations

STP) in response to a drop in the concentration (c2-c1) of the gas across the
membrane of thickness d and area A, which is Fick’s law (a more rigorous treat-
ment follows in section 2.2). I.e. the diffusion is the response of the system to the
gas concentration difference in the two opposite surface layers of the mem-
brane: V ( STP ) ⁄ At = D ( c 2 − c 1 ) ⁄ d . It is further assumed that the uptake and
release of penetrant from and into the gas phase is much faster than the diffu-
sion across the membrane, so that the latter becomes the rate-limiting step for
the permeation process. In turn, the gas concentration in the surface layer is
linked to its partial pressure in the gas phase by the solubility, i.e. c 1 = Sp 1 .
Assuming further, that the two partial pressures p1 and p2 are different but of sim-
ilar magnitude, so that the dependence of the solubility on the pressure can be
neglected, these two equations can be combined to yield Eqn. 3.

Since the conditions, under which Eqn. 3 is derived, do hold very often in prac-
tice, the permeability can be determined by calculating D and S separately. This
has important consequences for molecular simulation. Both D and S are bulk
quantities and do not depend on the membrane being a thin film. They can be
calculated for the bulk polymer under periodic boundary conditions without any
surface or interface being present, and without having to simulate explicitly the
transfer of molecules into and out of the polymer.

2.2 Diffusion coefficient


In the most general formulation, the average velocity 〈 v I 〉 of molecules (atoms,
ions) of species I is driven by a gradient in its chemical potential µI. If linear
response is assumed, i.e. if the response of the system is proportional to the per-
turbation, we have

D
〈vI 〉 = − ∇ µI (Eqn. 4)
RT
where T is the temperature in K and R is the gas constant (8.31441 J K-1 mol-1),
angle brackets imply ensemble averages. The velocity and the gradient are vec-
tors. In a general medium, the velocity does not have to be against the direction
of the gradient, but can be at some angle to it. Hence, the diffusion coefficient
takes the form of a 3×3 tensor. If, however, the medium is isotropic the diffusion
tensor is diagonal with all diagonal elements being equal; average velocity and
chemical potential gradient are colinear and the tensor D may be replaced by a

Permeation of Polymers 9
Foundations

scalar D. Liquids are isotropic media and so are amorphous polymers, at least
on a macroscopic scale. With the definition of the chemical potential Eqn. 4
becomes

D
〈vI 〉 = − [ V ∇p − S I ∇T + F I + RT∇ln a I ] (Eqn. 5)
RT I
Here, VI and SI are the specific molar volume and entropy of species I, aI is its
activity, and FI is an external force acting on molecules of species I, p is the pres-
sure. (Experimentally, this can, for example, be realised by applying an electric
potential ϕ to a system containing ions of charge zI (in elementary charges e), in
which case we have

F I = z I F∇ϕ (Eqn. 6)

with F=9.64846 ×104 C mol-1 being Faraday’s constant.) In the absence of tem-
perature and pressure gradients and external forces Eqn. 5 simplifies to

〈 v I 〉 = − D∇ln a I (Eqn. 7)

or

〈 v I 〉a I = J I = − D∇a I , (Eqn. 8)

which is Fick’s first law (JI denotes the flux of particles of species I) if activities
are equated with concentrations.

Fick’s law holds in the case of thermal and mechanical equilibrium (constant T
and p). Note however, that in technological applications the upstream and down-
stream sides of membranes can have different pressures and temperatures. In
those cases, temperature and pressure differences have to be incorporated into
the calculation of the transmembrane flux.

The external force FI in Eqn. 5 is not restricted to be an electrostatic force or, in


fact, any force that can be devised experimentally. In computer simulations, it is
possible to apply forces which are completely fictitious. Under certain conditions,
Eqn. 5 still holds. It will be seen later how this fact can be exploited in non-equi-
librium molecular dynamics simulations (Section 3.2).

In a microscopic picture, the diffusion coefficient is related to the autocorrelation


function of the particle current

10 Florian Müller-Plathe
Foundations


1
D =
3 ∫ 〈 ∑i vi ( 0 ) ∑i vi ( t ) 〉dt (Eqn. 9)
0

Here, the sums extend over all N particles i in the system, and the angle brack-
ets imply the ensemble average. If the concentration of the diffusing species is
low and the interaction between these particles is short-ranged, one can make
the approximation that the velocities of two different particles, i and j, are uncor-
related, i.e. 〈 v i ( t ) v j ( 0 ) 〉 = 0 . Hence, Eqn. 9 becomes


1
D =
3 ∫ 〈vi ( t ) vi ( 0 ) 〉dt (Eqn. 10)
0

This equation (known as the Green-Kubo relation for the diffusion coefficient)
states that the diffusion coefficient is given by the time integral of the single-par-
ticle centre-of-mass velocity autocorrelation function. Under the condition that
the long-time limit (often called the hydrodynamic limit) has been reached,
Eqn. 10 can be shown [26] to be equivalent to the Einstein equation

6Dt = 〈 R i ( t ) − R i ( 0 ) 2 〉 (Eqn. 11)

The right hand side is the so-called centre-of-mass mean-square displacement


(MSD) which is the square of the distance the particle has travelled between
time 0 and time t. The particle’s cartesian position vector is denoted by Ri. What
Eqn. 11 implies is that - in the long-time limit - the MSD is proportional to the time
elapsed. The slope of the MSD as a function of t is then given as

6D = d 〈 R i ( t ) − R i ( 0 ) 2 〉 (Eqn. 12)
dt
The three forms Eqn. 10-Eqn. 12 are completely equivalent for the calculation of
the diffusion coefficient from the microscopic motion of individual particles. The
resulting diffusion coefficients are often called tracer diffusion coefficients.

The derivation of the Einstein equation Eqn. 11 assumes that the motion of a sin-
gle particle follows a random walk, i.e. that the particle has no memory of previ-
ous steps and that therefore its motion at some time t is completely uncorrelated
with its motion at any previous time t’. If there is sufficient time between t and t’
this is the case, hence the notion of a long-time limit. At shorter time scales the

Permeation of Polymers 11
Foundations

a)

b)

FIGURE 2 Two reasons for anomalous diffusion. (a) The environment


restricts the penetrant’s motion. As a result it cannot perform a
random walk on the length scale of this drawing. (b) The
environment, here a narrow slit that has to be passed, enforces
a conformational change upon a flexible penetrant.

motion might be correlated, in which case the Einstein equation Eqn. 11 is no


longer valid, the MSD follows a different power law and no diffusion coefficient
can be defined. We can summarise the different cases:

• The mean-square displacement is quadratic in time ( 〈 R i ( t ) − R i ( 0 ) 2 〉 ∝ t 2 ).


This is equivalent to the displacement being linear in time which is the signa-
ture of a free flight. In a dense system, it occurs only at very short time scales
(< 1ps) while the particle is moving freely within its (solvent) cage until it hits
the cage wall.
• The mean-square displacement is linear in time ( 〈 R i ( t ) − R i ( 0 ) 2 〉 ∝ t ). This
is the long-time limit of Einstein diffusion which was discussed above.
• The exponent of the power law of the mean-square displacement is smaller
than 1 ( 〈 R i ( t ) − R i ( 0 ) 2 〉 ∝ t n , n < 1 ). This is the case of anomalous diffu-
sion, which may or may not occur at intermediate time scales. It is caused by
some feature of the environment which prevents the particle from performing
a random walk. For instance, there can be obstacles that restrict the space
available to the particle or that force a diffusing flexible molecule into a certain
conformation in order to pass a narrow passage (Figure 2). At long enough
length scales (and, hence, long enough time scales) the path will be again
random and a cross-over from anomalous diffusion to Einstein diffusion takes
place, at least in most natural systems. It is possible to design mathematically

12 Florian Müller-Plathe
Foundations

FIGURE 3 Power laws of the dependence of the mean-square


displacement (MSD) of a diffusing particle on time in the
different time ranges. Short times: free flight motion gives
quadratic behaviour. Intermediate times: anomalous diffusion
causes an exponent smaller than 1. Long times: Einstein
diffusion with linear time dependence. A typical MSD curve is
shown in (a). The log-log plot (b) of the same data shows the
power law more directly. The grid spacing equals 1 in both
coordinates. The slope in the log-log plot equals the exponent n.

Permeation of Polymers 13
Foundations

environments, which are self-similar on all length scales and therefore cause
anomalous diffusion to persist on all time scales. Percolation clusters are
examples of such fractal environments [27, 28].
• The exponent of the power law of mean-square displacement is larger than 1
( 〈 R i ( t ) − R i ( 0 ) 2 〉 ∝ t n , n < 1 ). This can occur if some other transport mech-
anism such as convective flux is superposed to the diffusion.

A typical mean-square displacement curve is shown in Figure 3. It shows the


free-flight (quadratic), anomalous (fractional exponent) and normal diffusive (lin-
ear) regimes. The different power laws become more manifest in the log-log
representation, where the slope equals the exponent n.

2.3 Ionic conductivity


The conductivity (we shall drop the term ionic from now on, since electronic con-
ductivity does not feature in this work) is closely related to the diffusion coeffi-
cient. In fact, Ohm’s law is an example of a linear-response relation like Eqn. 4.
In the limit of an infinitely dilute solution of ions, the contribution to the molar spe-
cific conductivity ΛI of an ionic species I is related to its diffusion coefficient by
the Nernst-Einstein relation

z 2I F 2
ΛI = DI , (Eqn. 13)
RT
see also Eqn. 6. The molar specific conductivity is the specific conductivity λ
divided by the molarity of the ion solution. The specific conductivity is typically
measured in Ω-1m-1 which is the same as siemens per meter (S/m). In concen-
trated solutions, Eqn. 13 does not apply directly anymore. In such cases, some
empirical factor f is often introduced into the Nernst-Einstein equation to make it
hold again.

z 2I F 2
ΛI = f DI , (Eqn. 14)
RT
In the microscopic picture, the frequency-dependent conductivity λ(ω) is defined
[26] by the autocorrelation function of the charge flux j ( t ) = e 2 ∑ z 2i v i ( t )
ions

14 Florian Müller-Plathe
Foundations


3Vk B Tλ ( ω ) = ∫ exp ( −iωt ) 〈j ( t ) j ( 0 ) 〉dt , (Eqn. 15)
0

which in the zero-frequency limit becomes


3Vk B Tλ = ∫ 〈j ( t ) j ( 0 ) 〉dt (Eqn. 16)
0

where V is the volume and kB Boltzmann’s constant (1.38066×10-23 JK-1). Note


the similarity to Eqn. 9. In contrast to the case of diffusion of neutral species,
here it cannot generally be assumed that different ions do not interact with each
other. Because of the long-range nature of the electrostatic interactions between
ions (proportional to r-1) compared to the dispersion interactions between two
gas molecules (leading term proportional to r-6) the concentration at which this
approximation would be safe, is much lower. Ionic conductivity is therefore a col-
lective transport phenomenon, whereas the diffusion of gas molecules at low
concentrations can be regarded as an individual molecular transport phenome-
non. The Einstein equation corresponding to Eqn. 16 is then

6tVk B Tλ = e 2 〈 ∑ ∑ z i z j [ R i ( t ) − R i ( 0 ) ] [ R j ( t ) − R j ( 0 ) ] 〉 (Eqn. 17)


i j

or

6tVk B Tλ
e2
= ∑ ∑ zi zj 〈 [ Ri ( t ) − Ri ( 0 ) ] [ Rj ( t ) − Rj ( 0 ) ] 〉
i j
= ∑ z2i 〈 [ Ri ( t ) − Ri ( 0 ) ] 2 〉 (Eqn. 18)
i
+ 2 ∑ zi zj 〈 [ Ri ( t ) − Ri ( 0 ) ] [ Rj ( t ) − Rj ( 0 ) ] 〉
j>i

The first term on the right hand side of Eqn. 18 is just a sum over individual
mean-square displacements (albeit weighted with the charges) like in the Ein-
stein equation for the diffusion coefficient Eqn. 11. Let us illustrate the connec-
tion with the diffusion coefficient by an example. Assume that there are equal
numbers N/2 of positive and negative ions and all their charges have the same
magnitude z. From Eqn. 18 we then have

Permeation of Polymers 15
Foundations

6tVk B Tλ Nz2
= ( 〈 [ Ri ( t ) − Ri ( 0 ) ] 2 〉 + 〈 [ Ri ( t ) − Ri ( 0 ) ] 2 〉 )
e2 2 + -
(Eqn. 19)
+ 2 ∑ zi zj 〈 [ Ri ( t ) − Ri ( 0 ) ] [ Rj ( t ) − Rj ( 0 ) ] 〉
j>i

The subscripts + and - denote the mean-square displacements of the positive


and negative ions, respectively. Inserting Eqn. 11 and rearranging one obtains

Ne2 z 2 1
λ= ( D + D- )
Vk B T 2 +
(Eqn. 20)
e2
3tVk B T j∑
+ zi zj 〈 [ Ri ( t ) − Ri ( 0 ) ] [ Rj ( t ) − Rj ( 0 ) ] 〉
>i

The first term on the right hand side contains only single-particle diffusion coeffi-
cients (tracer diffusion coefficients), while the sum over all the cross terms allows
for the collective behaviour of ions in an electrolyte solution. If the cross terms
can be neglected, i.e. in very dilute solutions, the Nernst-Einstein equation
Eqn. 13 is recovered

λ +λ
+ - N z2 F2 1
λ = = ( D + D- ) , (Eqn. 21)
2 N A V RT 2 +

if one recognises that (i) F = eN A , (ii) R = k B N A , (iii) N ⁄ ( NA V ) is the molarity

of the solution. Avogadro’s constant is denoted by N A = 6.02205 × 10 23 mol − 1 .

2.4 Solubility
The solubility S of a gas in a polymer describes what concentration c of gas is
obtained in the polymer in equilibrium with a certain partial pressure p of the gas
outside. The unit of gas solubility is often

cm 3 ( 273.15K ; 1.013 × 105 Pa )


unit of S: (Eqn. 22)
cm 3 × Pa
In general, the solubility itself depends upon the gas partial pressure in a compli-
cated way. For gases in polymers, the sorption equilibrium is frequently
described by the so-called dual-mode theory [6]

16 Florian Müller-Plathe
Foundations

bp
c = Hp + c ∞ (Eqn. 23)
1 + bp
Dual-mode theory assumes two distinct mechanisms of sorption reflected by the
two terms. The first term corresponds to Henry’s law with Henry’s constant H.
Unspecific sorption is assumed for this mode and the gas uptake is simply pro-
portional to the gas pressure. The second term is a Langmuir-type isotherm with
the constants b and c∞, of which the latter is the saturation concentration.
Assumption of a Langmuir isotherm implies the existence of specific sorption
sites inside the polymer. There is only a finite concentration of these sites. This
leads to a saturation behaviour for gas absorption.

The dual-mode model is valid for large varieties of gases and polymers [9]. How-
ever, other sorption models have to be used [9] as soon as the penetrant
induces significant structural or dynamical changes in the polymer such as
swelling or plasticisation. This usually happens only for larger penetrants that
are normally non-gaseous or for very high penetrant concentrations.

For the solution of permanent gases in rubbery polymers it has been shown
experimentally that one can even drop the Langmuir part of Eqn. 23. For these
systems Henry’s law often holds up to pressures of several hundred atmo-
spheres. The solubility is linked to the excess free-energy ∆G of a gas molecule
dissolved in the polymer.

∆G
S = exp ( − ) (Eqn. 24)
kB T

Henry’s constant H in its standard unit (Pa-1) can be derived from S by assuming
that the gas exhibits ideal-gas behaviour both at STP conditions and at the tem-
perature T at which one is actually doing the measurements (i.e. V/V(STP) = T/
273.15 K)

273.15K
H = S (Eqn. 25)
T

Permeation of Polymers 17
Foundations

18 Florian Müller-Plathe
3 Machinery

3.1 Molecular dynamics

The most straightforward way of studying computationally the motion of individ-


ual atoms or molecules is molecular dynamics (MD) because time is explicitly
present in its formulation. The MD method has been well described in the litera-
ture [15-20, 29] so that a short overview is sufficient here.

In its simplest form, the MD method in cartesian coordinates R works as follows.


Given a potential function U(R), which depends on the positions of all atoms,
Newton’s equation of motion for all atoms i

d2 Ri 1
= − ∇ U ( R) (Eqn. 26)
dt 2 mi i

is then solved, subject to the initial conditions R=R(0) and v=v(0) and suitable
boundary conditions (Ri, vi and mi are position and velocity and mass of atom i,
whereas R and v without subscripts denote the vectors of all atoms’ positions
and velocities). Let us look at the various points of this statement in detail.

The potential energy function U(R) describes the Born-Oppenheimer surface for
the atomic motion. It would be desirable to obtain the surface quantum-chemi-
cally, i.e. by solving the electronic Schrödinger equation of the system for every

19
Machinery

BOX 1. Force fields for molecular simulation.

A typical forcefield for a polymeric system might look like this

kd kφ
U ( R) = ∑ 2
( d − d0 ) 2 + ∑ 2 ( φ − φ0 ) 2
bonds bond angles
kθ kτ
+ ∑ 2
( θ − θ 0
) 2+
∑ 2
[ 1 − cos ( nτ − τ 0 ) ]
improper dihedrals dihedral angles

 σ ij   σ ij 
12 6
1 qi qj
+ ∑ 4ε ij  r 
ij

 r ij 
+
4πεε 0 r ij
nonbonded pairs ij

d, φ, θ and τ are actual values for bond length, bond angle, improper (har-
monic) dihedral and dihedral angles, d0, φ0, θ0 and τ0 are the equilibrium val-
ues and kd, kφ, kθ and kτ the associated force constants, n is the periodicity of
the torsional potential.
rij is the distance between the two nonbonded atoms i and j, εij and σij are
the well depth and contact radius of the Lennard-Jones potential between
these atoms, qi and qj are the charges on i and j, ε0 is the vacuum permittivity
(8.854 188 × 10-12 C2J-1m-1) and ε is the effective dielectric constant of the
medium. The Lennard-Jones parameters for pairs of unlike atoms are often
derived by mixing rules, e.g. the Lorentz-Berthelot mixing rules

ε ij = ε ii ε jj σ ij = ( σ ii + σ jj ) ⁄ 2

Nonbonded interactions are usually excluded between atoms which are


already interacting by a bond or bond angle term (first and second neigh-
bours) and they are often modified for the end atoms of a dihedral angle
(third neighbours).

Note also that different analytical forms are used for some of the terms
(angles harmonic in cosφ rather than φ, cosine expansions for dihedrals, r-9 or
exp(-Arij) instead of r-12 repulsions, etc.). Some forcefields also have cross
terms between different degrees of freedom. Additional terms may be
present for out-of-plane bendings, hydrogen bonds, etc.

20 Florian Müller-Plathe
Machinery

atom configuration. Using ab initio or semiempirical quantum-chemical methods


this has recently become possible for systems of a few atoms and simulations of
up to a several hundred time steps. Using local density methods and plane-wave
basis sets, larger systems have been treated for longer times [30]. However,
treating polymeric systems of several hundreds or thousands of atoms for sev-
eral million time steps, as is required for the problem of diffusion, is still impossi-
ble with present-day computing resources. One therefore uses analytical
representations of the potential surface, so-called forcefields. An example of the
analytical form of such a force field in given in Box 1. This one is very close to
what we use in our work.

It is also very common to use rigid bond constraints instead of flexible harmonic
bonds (as we do in our work [31, 32]). Strictly speaking, bond constraints are not
part of the forcefield, but correspond to solving the equations of motion in some
lower-dimensional space. The reasons for their use are both technical (one can
use longer time steps if the high-frequency bond vibrations are frozen out and
the difficulties with their often slow equilibration with the rest of the system are
avoided) and physical (at room temperature only the vibrational ground state will
be populated and the bond length is virtually constant).

In order to solve Newton’s equation of motion Eqn. 26 it is discretised in time.


Various integration schemes exist. In practice, it makes little difference which
one is used. We use the leap-frog scheme [15] which involves calculation of
atom velocities at half steps

∆t
v i ( t + ∆t ⁄ 2 ) = v i ( t − ∆t ⁄ 2 ) + f i ( t )
mi (Eqn. 27)
R i ( t + ∆t ) = R i ( t ) + ∆t v i ( t + ∆t ⁄ 2 )

Here, ∆t denotes the step size, and fi=-dU/dRi is the total force on atom i. New-
ton’s equation of motion conserves the total energy of the system and, therefore,
leads to a microcanonical statistical-mechanical ensemble. In practice, one is
usually more interested in ensembles in which the temperature (canonical
ensemble) or the temperature and the pressure (isothermal-isobaric ensemble)
are conserved. For this, Newton’s equation of motion has to be slightly modified
to couple the system’s temperature T (or pressure p) to a temperature (or pres-
sure) bath of temperature T0 (or pressure p0). Again, there are several such con-

Permeation of Polymers 21
Machinery

stant-temperature and constant-pressure schemes [15, 33]. We use the loose-


coupling algorithm [34] which implements a first-order feed-back

dT 1
= − ( T − T0 ) (Eqn. 28)
dt τT

dp 1
= − ( p − p0 ) (Eqn. 29)
dt τp

by uniformly scaling the atom velocities (constant T) or atom positions and box
dimensions (constant p), τT and τp being the appropriate coupling times. Note
that the loose-coupling method for constant-temperature simulation has recently
been shown to generate the correct canonical distribution [35] if the τ’s are cho-
sen large enough.

The MD simulation is started with appropriate initial conditions. That means that
for every atom its initial position and velocity has to be specified. The velocities
are randomly picked from a Maxwell-Boltzmann distribution at the desired tem-
perature. Finding starting positions is far more difficult. We are mainly concerned
with amorphous polymers. Hence, by some means amorphous structures have
to be generated that are low in energy, have the correct distribution of torsional
angles, the correct distribution of free volume and so on. It should be noted that
during the typical time of our simulations (a few nanoseconds) the polymer only
undergoes local rearrangements. The global topology of the polymer chain stays
as it started out. One cannot therefore hope for the polymer to relax to another
configuration during the simulation, and the starting structure has to be very
good. This is in contrast to faster relaxing systems such as liquids which can be
started off in almost any configuration. We use the modified rotational isomeric
state method [37]. A short summary of other methods is given in Box 2. Placing
the diffusants is much easier. We start them off in some cavity in the polymer,
which is found by randomly trying to insert them into the polymer until a place is
found where their energy is below a certain threshold. We also make sure that
any two penetrants are separated by some minimum distance as they are
inserted.

In order to avoid surface effects, it is common to employ periodic boundary con-


ditions: The primary simulation box is surrounded by periodic images of itself
and when an atom leaves the box through one side it reenters it through the

22 Florian Müller-Plathe
Machinery

BOX 2. Methods for generating amorphous polymer structures.

Little is known about the atomistic structure of an amorphous polymer. Since


the amorphous parts of polymers often have the dominant influence on a poly-
mer property, it is still desirable to model them. From the various methods that
have been used in this context we list a few here. An overview is given in ref.
36.
• Structure generation methods. The emphasis is on the generation of the
initial structure. Ideally, the generated structure needs no further refine-
ment.
- Modified rotational isomeric state (RIS) method [37]: in the classical RIS
method, a polymer chain is built in vacuum, which satisfies the known
probability of key dihedral angles in the chain, including conditional prob-
abilities. In the modified scheme, a chain is built under periodic bound-
ary conditions. The probabilities are modified to allow for the interactions
with polymer segments that have already been placed.
- Phantom chain growth [38, 39]: A polymer chain with the correct dihedral
distribution is grown in vacuum. It is then placed under periodic bound-
ary conditions and excluded-volume interactions are gradually turned
on. A variant is the pivot Monte Carlo method, in which the vacuum
chain is first refined by pivot MC moves.
- In situ polymerisation [40]: A simulation of the monomer liquid is per-
formed. At a certain point the simulation is halted and nearest-neighbour
monomers are connected, which is a travelling-salesman problem.
• Structure refinement methods. Ideally, the refinement method is so effi-
cient that the initial structure can be arbitrary.
- Standard EM/MD/MC with or without temperature and/or pressure proto-
cols.
- Continuum reptation MC [41].
- Continuum configurational bias MC [42]: Entire sections of the polymer
are deleted and regrown in MC moves.
- Standard MD with nonphysical force field: Soft-core potentials allow
polymer chains to pass through each other [43].
• Coarse-graining. The idea is to map the atomistic model of a polymer into
a more coarse-grained representation, like 1 interaction site per several
atoms or even monomers. After relaxation of the coarse-grain model, a
mapping back onto the atomistic model is done. This is especially efficient
if the coarse-grain model is a lattice model [44-46].

Permeation of Polymers 23
Machinery

opposite side. In order to keep the effect of polymer chain ends minimal we use
a single polymer chain in the simulation box. For the chain the same periodic
boundary conditions hold: if it grows out of the box on one side, it grows back
into it on the other side. The system is therefore best described as a single poly-
mer chain interacting with its own periodic images.

3.2 Diffusion coefficients

The gas diffusion coefficient D may be extracted from an MD simulation in vari-


ous ways. If the hydrodynamic limit is reached (the simulation time is long
enough), D may be calculated from either the penetrants’ instantaneous veloci-
ties or their positions by means of the Green-Kubo or Einstein relation, respec-
tively Eqns. 10-12. We generally use the route via the mean-square
displacement which is plotted against time. To the linear portion of the curve a
line is least-squares fitted from whose slope the diffusion coefficient can be cal-
culated. The ensemble average in Eqn. 12 is realised in an MD simulation as an
average over all penetrants and over all possible time origins.

There are reasons for preferring the mean-square displacement to the velocity
autocorrelation function. Firstly, the velocity autocorrelation function tends to
become noisy at larger t and it is often necessary to make assumptions about
the analytical form of its long time tail in order to calculate the integral [15],
unless excessive simulation time is spent. The Einstein relation avoids such
ambiguity. Secondly, in the leap-frog integration scheme we are using, the atom
positions are determined through a higher order than the velocities. Thirdly, the
mean-square displacement is not dominated by the short-time behaviour as is
the integral in the Green-Kubo relation (Eqn. 10), and trajectory frames have to
be saved to disk less often. Finally, the shape of the mean-square displacement
curve tells us directly if the hydrodynamic limit has been reached or if anomalous
diffusion prevails. This is very useful as we shall see below.

The methods described so far calculate D from an MD simulation which is at


equilibrium, i.e. no external forces perturb the system. Equilibrium simulations
have been used in most MD calculations of penetrant diffusion coefficients. It is,
however, also possible to obtain D from a nonequilibrium MD (NEMD) simulation
by appropriately perturbing the system and measuring its response [15, 33].
Recall that a diffusive flux is driven by a gradient in the chemical potential
(Eqn. 4), which experimentally most often corresponds to a concentration gradi-

24 Florian Müller-Plathe
Machinery

ent. With the low number of diffusing particles in a simulation it is difficult to set
up a concentration gradient and obtain the diffusion coefficient in this way. We
instead [47] used a fictitious external potential as the perturbation and realised
the gradient in the chemical potential in this way. We chose the external potential
as

N
ϕ = E ∑ ( −1 ) i Ri (Eqn. 30)
i

(E being a constant, spatially uniform fictitious field, normally chosen parallel to


one cartesian axis) which in the simulation is implemented as a constant exter-
nal force Fext on every penetrant (even number of penetrants). The potential ϕ
thus defined becomes an additional term in the Hamiltonian of the system. The
(-1)i term causes half of the penetrants to be accelerated in one direction and
half of them in the opposite direction. In this way, the total linear momentum is
conserved, i.e. one avoids acceleration of the system as a whole. Note, how-
ever, that this partitioning of the forces is by no means unique. One could
achieve the same effect by letting a force Fext act on all of the N penetrant atoms
and a force -(N/M)Fext on each of the M polymer atoms.

Applying a constant external force to the system means constantly feeding


energy into it. This energy has to be drained in some way to prevent the system
from heating up. In our calculation this is done through the loose-coupling con-
stant-temperature algorithm described before.

The diffusion coefficient is then calculated from the steady-state particle flux

kB T N
D = 〈∑ ( −1 ) i vi 〉 (Eqn. 31)
ext
NF i

The factor (-1)i causes the reverse of the velocities of particles accelerated in the
negative direction Eqn. 30 to contribute to the particle flux. We note that both
with equilibrium methods and nonequilibrium methods the whole diffusion tensor
D can be calculated. For EMD, this is done by using the appropriate mixed dis-
placement (or velocity cross correlation function), e.g.

6D xy t = 〈 [ X i ( t ) − X i ( 0 ) ] [ Y i ( t ) − Y i ( 0 ) ] 〉 (Eqn. 32)

Permeation of Polymers 25
Machinery

where Xi and Yi are the x and y component of the position of particle i. In NEMD,
one uses a perturbing force along one cartesian direction and measures the
component of the flux in another direction, e.g.

kB T N
D xy = 〈 ∑ ( − 1 ) i v y, i 〉 (Eqn. 33)
NFext
x i

The individual elements of the diffusion tensor will be important in the discussion
of finite-size effects below.

Let us conclude the description of the nonequilibrium MD method by emphasis-


ing again that the external force has no physical counterpart. The external force
is just a convenient means to set up a gradient in the chemical potential, to which
the system responds as it would to a concentration gradient, so that the calcu-
lated diffusion coefficient is the same. Only in the case of diffusing ions, the ficti-
tious force can be identified with a Coulomb force due to an applied external
electric field.

3.3 Solubility

For the rubbery polymers studied here, it suffices to calculate the solubility in the
limit of Henry’s law (see Section 2.4). The free energy for the solvation process
can then be calculated by Widom’s method of test-particle insertion [15, 19, 48,
49]: The prerequisite is an ensemble of structures of the pure host system, in this
case the polymer, which may have been generated by whatever suitable
method. We normally use MD for this purpose. Into these structures one inserts
a gas molecule at a random position (and in random orientation if the molecule
has two or more atoms) and calculates the change in potential energy ∆U asso-
ciated with the insertion. Since the interaction between guest and host involves
only pair interactions (see Box 1), a separation of this energy component can be
easily done: It is simply the sum of all pairwise interactions between all atoms of
the inserted molecule and all host atoms. The inserted molecule is never actually
left in the polymer, so that the polymer behaviour is not influenced by the pres-
ence of the gas. From many insertions into many polymer structures the follow-
ing ensemble average over Boltzmann factors is formed, from which the
Helmholtz free energy ∆A for the solvation can be obtained.

26 Florian Müller-Plathe
Machinery

∆A ∆U
− = ln 〈 exp − 〉 (Eqn. 34)
RT kB T

In a constant-pressure simulation, a correction for the fluctuating volume must


be applied and the Gibbs free energy ∆G becomes [15, 19]

∆G ∆U
− = ln 〈 Vexp − 〉 ⁄ 〈V 〉 (Eqn. 35)
RT kB T

The attractive features of the particle-insertion methods are its simple implemen-
tation, the fact that one polymer trajectory (which is the most expensive part of
the calculation) can be reused for solubilities of various penetrants and that it is
easily possible to extend the calculation systematically until convergence. Its
biggest problem is that it becomes quickly inapplicable as the penetrant size
increases. The reason is that the method relies on the pure polymer being able
to sample configurations corresponding to the full polymer-penetrant system.
This can only work if the conformations in the two systems are not too different.
In presence of large penetrants, the polymer adapts itself significantly, the con-
figurations are no longer similar and the approximation breaks down. Similarly, if
the polymer undergoes substantial rearrangement in the solution process due to
strong interactions (hydrogen bonds, complexation, large penetrant concentra-
tion) this method will not work.

3.4 Software

Back in 1988 when this work started, molecular simulation of polymers was not
quite as popular as it has recently become. Consequently, commercial providers
of chemical simulation software were only beginning to design features neces-
sary for polymer simulations into their products. There were few academic pro-
grams suitable for polymer simulation and they were often specific to a certain
class of polymers.

The lack of available software led to the development of a suite of simulation


programs, which were intended to be used for but are not restricted to polymers.
The programs have since then been extended and modified as needed, and
have evolved into a larger software package. Although much time was spent on
the development of the software, in hindsight it turned out to be a good invest-
ment for a number of reasons: It has always been possible to put extensions into

Permeation of Polymers 27
Machinery

BOX 3. Main features of the YASP molecular dynamics program (Version


3.0). For details, see ref. 51.
algorithm: leap-frog (Eqn. 27). Periodic resetting of total linear
momentum.
boundary conditions: orthorhombic periodic.
force field: functional form as described in Box 1. No generic set
of parameters.
constraints: rigid bond constraints using SHAKE [31, 32].
ensembles: microcanonical, constant T and constant p using the
loose-coupling algorithm (Eqns. 28, 29) [34], the 3
cartesian components can be coupled separately in
constant p calculations.
pressure calculation: atomic virial with long-range corrections. Contribu-
tions from internal degrees of freedom and con-
straints are calculated.
additional forces: constant external force on selected atoms (cf. Eqn.
30), harmonic position restraining for selected
atoms.
force evaluation: highly vectorised neighbour list algorithm for non-
bonded forces. Coulomb forces can be evaluated
with simple cutoff or reaction-field approximation.

the program which are not available in the standard packages. An example is the
use of external forces which would have been impossible without access to the
source code. The program could be readily ported to various hardware architec-
tures. This is quite useful in a project which depends on compute resources and
needs to exploit them as they become available. At one stage, there even was a
parallel version of the code when a significant amount of time became available
on a multi-processor Apollo DN10000 workstation [50]. The central MD code
which uses most of the computer time has also been tuned for vector supercom-
puters. Finally, since the code is non-proprietary, all the cumbersome licensing
issues can be avoided.

In the development of the code a tool kit approach was chosen, i.e. a collection
of little, highly specialised utilities in preference to one big package that does
everything. Such a system is easier to maintain and extend than one bulky pro-
gram. Particular attention, however, had to be given to the definition of standard

28 Florian Müller-Plathe
Machinery

file formats and associated libraries so that the individual units can easily com-
municate [51].

The centrepiece is a molecular dynamics code whose most important features


are summarised in Box 3. The MD program does not do much analysis by itself,
but instead writes out a coordinate and velocity trajectory at specified intervals,
which can then be processed by the analysis programs. This separation makes
it possible to run the different programs on the hardware which suits them most.
It also has the advantage that the analysis programs can be kept small and spe-
cialised and that they can even be written in languages different from the MD
code, if that is more appropriate for the problem at hand. Examples of analysis
programs are the ones which evaluate various types of mean-square displace-
ment or correlation functions, or those which perform test-particle insertions.

Besides the analysis programs, the package contains other types of tools

- The energy minimiser (molecular mechanics) implements the conjugate gradi-


ent methods of Fletcher-Reeves, Polak-Ribiere and Shanno [52].

- There is an experimental version of a bond-fluctuation Monte Carlo code [53].

- Coordinate transformers: transformation between the YASP coordinate file for-


mat and other file formats.

- A generic molecular topology generator: Generates a list of bonds, bond


angles, dihedrals, excluded neighbours and so on from either atomic cartesian
coordinates or a connectivity table by simple heuristics. These lists are generic
in that there are no assumptions about forcefield parameters or even the func-
tional form of the forcefield. Instead, interactions are labelled by symbolic
strings which may be converted to appropriate parameters by means of stan-
dard UNIX utilities (sed, awk, etc.).

- General utilities.

- Plot programs: there are a number of very primitive plot programs for use on a
Sun workstation. We generally prefer Leif Laaksonen’s SCARECROW [54] as
the graphics front-end.

Permeation of Polymers 29
Machinery

Some details of the implementation of the molecular dynamics part of the pack-
age, which goes by the name of YASP, are given in reference 51. The paralleli-
sation of YASP was described in reference 50.

30 Florian Müller-Plathe
4 Motion of Gas Molecules
through Amorphous
Polymers

4.1 Hopping motion

The simplest way of studying the diffusion of an individual penetrant is to look at


its path through space. In Figure 4 we show the typical trace of an oxygen mole-
cule diffusing through amorphous polyisobutylene. We can clearly discern two
types of motion: (i) For periods of time the penetrant stays in certain small
regions of space. During such a quasi-stationary period, it explores these
regions thoroughly, but it does not move beyond the confines of the volume it
resides in. (ii) The quasi-stationary periods are interrupted by quick leaps from
one region to another, which is close by. As a first result, we note that diffusion
proceeds by hopping of the penetrant from one binding site to the next. Similar
motion patterns have been found in all MD studies of the diffusion of small pene-
trants in amorphous polymers. It is interesting to note that a related mechanism
is responsible for the superionic conductivity of Na-β″-alumina at higher temper-
atures (~1200 K), as has recently been found by MD simulations [55]: In
between the two-dimensional alumina layers, Na+ ions diffuse by performing
hops between lattice sites. There is, however, one important difference between
inorganic crystalline superionic conductors and amorphous polymers. The dwell-
ing sites in inorganic crystals are mostly arranged regularly, following the host
lattice, whereas in a polymer these sites appear to be statistically distributed
regarding their size, their shape and their location.

31
Motion of Gas Molecules through Amorphous Polymers

FIGURE 4 Trajectory (8 ns) of the centre of mass (“trace”) of an O2


molecule in polyisobutylene at 300 K, stereo view.

So far, nothing has been said about the nature of the binding sites in polymers.
One can, however, change the representation of the penetrant trajectory to gain
further insight. In Figure 5 we show the displacement of three penetrant mole-
cules (H2, O2, and CH4 in amorphous PP [56]) from their initial positions as a
function of the simulation time. The hops can clearly be seen as steps in the
graph for all three penetrants. However, there are also differences in the motion
pattern which are mainly related to the size of the penetrants, which are ordered
H2<O2<CH4. Firstly, the larger the penetrant the more pronounced are the step-
like hops in the graph. For CH4 these are quite clear, whereas for the small H2
they are blurred to some extent. One may wonder, if penetrants smaller than H2
would have displacement curves completely without steps, which would be indic-
ative of a diffusion mechanism without discrete jump events, as is common for
liquid-like diffusion. Secondly, the behaviour during the quasi-stationary periods
is also related to the penetrant’s size. The larger the penetrant, the smaller are
its positional fluctuation at the binding site. Since these gas molecules have no

32 Florian Müller-Plathe
Motion of Gas Molecules through Amorphous Polymers

8.0

H2

6.0
|R(t) - R(0)| / nm

O2
4.0

2.0

CH4
0.0
0 500 1000 1500 2000
t / ps

FIGURE 5 Displacement of selected penetrants from their initial


positions: H2, O2 and CH4 in atactic polypropylene at 300 K
(from ref. 56). Note that the O2 and H2 curves are offset in y
by 1 and 2 nm, respectively.

specific interactions with the polymer (such as, e.g., electrostatic interactions or
hydrogen bonds), one may surmise already at this stage that the binding sites in
the polymer are merely cavities which are large enough to hold a penetrant mol-
ecule. For a given cavity, a small penetrant is able to access more of its free vol-
ume than a larger penetrant, which explains the qualitative difference in the
positional fluctuations. Thirdly, already in this qualitative picture one sees that
the mobility of a penetrant is related to its size. One may suspect the diffusion
coefficients to be ordered H2>O2>CH4, which is indeed the case. From Figure 5
it is also obvious, that the increased mobility is brought about by an increase in
the jump frequency, whereas the jump distance changes only moderately. Hence
it is the jump rate that is probably related in some way to the penetrant’s size.

Permeation of Polymers 33
Motion of Gas Molecules through Amorphous Polymers

4.2 Penetrants in cavities

Before we discuss the jump event in more detail let us look at the behaviour of
penetrant in a cavity. The larger part of the time is spent in cavities. Although
these residence periods do not contribute to diffusion, they are quite interesting
in their own right [57].

The motion of a penetrant trapped in a cavity is dominated by short time scales,


typically <2ps. After that time, most of the correlation of molecular motion is lost
due to frequent collisions with the atoms forming the cavity walls. The centre-of-
mass velocity autocorrelation function of the penetrant exhibits typical liquid-like
behaviour with a negative region due to velocity reversal when the penetrant hits
the cavity wall [57]. (In contrast to our earlier preliminary report [57], the calcula-
tions upon which the results in this work are based, use an all-atom forcefield
[47]. However, the velocity autocorrelation function is not very sensitive to the
forcefield parameters used.) For H2 (Figure 6a) this happens sooner than for O2
(Figure 6b) due to its lower mass. In order to illustrate that the translational
motion really is liquid-like, we have also calculated the velocity autocorrelation
function of pure O2 at the same temperature but four different densities (Figure
6c). At the lower densities, the typical velocity autocorrelation function of a gas
with its uniform decay can be observed. One has to go to quite high oxygen
pressures in order to see the beginning of a negative backscattering peak devel-
oping in the velocity autocorrelation function. This picture has recently been con-
firmed in ref. 58, where the authors monitor reversals in the penetrants travel
direction when it hits the cavity walls.

For molecular penetrants, one can also study their rotational behaviour inside
the cavities. We have done this for the same diatomics as above, H2 and O2.
The orientation of a diatomic molecule is described by a unit vector u along its
molecular axis. The vector is defined with respect to a laboratory coordinate sys-
tem. The autocorrelation function of this orientation vector is a measure of the
angle ∆φ(t) by which the orientation changes within a given time t.

cos ∆φ ( t ) = 〈 u ( t ) u ( 0 ) 〉 (Eqn. 36)

This quantity is shown in Figure 7. One notes, that the H2 orientational correla-
tion function has a small negative region (Figure 7a), and the O2 function has
some remnants of an oscillatory structure around 0.7 ps (Figure 7b). The typical
orientational correlation function of a molecular liquid looks very different. It has

34 Florian Müller-Plathe
Motion of Gas Molecules through Amorphous Polymers

1.0
a)

0.5

0.0
1.0
b)
〈v(t)v(0)〉

0.5

0.0

1.0

1.10 MPa c)
0.37 MPa
0.5 0.16 MPa
0.13 MPa

0.0

0.0 0.5 1.0 1.5 2.0


t / ps

FIGURE 6 Normalised centre-of-mass velocity autocorrelation functions of


penetrants (all at 300 K): (a) H2 in polyisobutylene (PIB), (b) O2 in
PIB, (c) pure O2 at various pressures for comparison.

an exponential decay, maybe after some features at very short times [15, 59],
and the decay time can be related to the rotational diffusion constant, if one
assumes Debye relaxation [26]. The orientational correlation function of diatomic
molecules in a gas, on the other hand, looks more like a damped oscillation [59].

Permeation of Polymers 35
Motion of Gas Molecules through Amorphous Polymers

1.0
a)

0.5

0.0
1.0
b)
〈u(t)u(0)〉

0.5

0.0
1.0
1.10 MPa c)
0.37 MPa
0.16 MPa
0.5 0.13 MPa

0.0

0.0 0.5 1.0 1.5 2.0


t / ps

FIGURE 7 Correlation function of the orientation vector u of the molecular


axis of diatomic penetrants at 300 K: (a) H2 in polyisobutylene
(PIB), (b) O2 in PIB, (c) pure O2 at various pressures for
comparison.

To facilitate a qualitative comparison, we have again calculated the correspond-


ing orientational correlation function for pure O2 at different pressures (Figure
7c). It shows the transition from a liquid-like relaxation at high pressures to the
development of some structure, such as a negative region, as the density
approaches that of a gas. Comparing the relaxation of the penetrants inside the
polymer to that in the pure gas/liquid one sees that small gaseous penetrants

36 Florian Müller-Plathe
Motion of Gas Molecules through Amorphous Polymers

inside the polymer show residuals of a gas-like behaviour. Reorientation of the


molecular axis does not have the signature of rotational diffusion, but rather
shows some amount of free rotation with rotational correlation times of the order
of a few tenths of a picosecond. They depend strongly on the Lennard-Jones
radii of the penetrant’s atoms, since these influence the sphericity of a diatomic,
which in turn governs the rotational motion.

In summary, we find that the short-time behaviour of small molecules inside cav-
ities is best described by considering the translational motion to be somewhat
like in a liquid and the rotational motion more like that in a gas. This apparent
contradiction can be reconciled by the concept of cavities, which have relatively
hard walls that backscatter the penetrant, but which are large enough to allow
the penetrant some degree of free rotation.

Let us conclude this section by remarking that, in principle, the rotational correla-
tion function (Eqn. 36 or higher-order Legendre polynomials in cos ∆φ ( t ) ) can
be measured by various molecular spectroscopies (IR, Raman, NMR, fluores-
cence) if suitable penetrants are chosen. Such correlation functions for small
penetrants together with atomistic modelling of the polymer-penetrant system
could yield important information about size and possibly shape distributions of
the cavities inside the polymer, as probed by small molecules. Small molecules
could be a more meaningful probe for cavities than positron-annihilation spec-
troscopy, which together with some crude assumptions is the main method used
for estimating cavity sizes in polymers [60]. However, we are not aware of any
such study.

4.3 The jump event

During a jump event, a penetrant moves from a cavity to a neighbouring cavity in


a very short time compared to the residence time. The question arises what
atomic motions are involved in such an event, and, in particular, how does the
polymer matrix participate in the event or even facilitate it.

We have repeatedly tried to relate jump events to sudden conformational


changes in the polymer. Major torsional angle changes or ring flips would be
candidates, for example. We have never been able to find any such correlation.
Rather, it appears that many degrees of freedom of the polymer are involved,
but each of them only to a small extent. The motions of individual polymer atoms

Permeation of Polymers 37
Motion of Gas Molecules through Amorphous Polymers

a) d)

b) e)

FIGURE 8
c) Schematic view of a hopping event
(“Red Sea” mechanism). Initially
(a), the penetrant is located in a cav-
ity, separated from other accessible
volume in the polymer. By means of
fluctuation, a channel opens up (b).
When it is wide enough, the pene-
trant can pass through it (c). The
channel closes behind the penetrant
(d) and the initial and final cavity
become separated again (e).

38 Florian Müller-Plathe
Motion of Gas Molecules through Amorphous Polymers

are too confusing to reveal the complex processes when atom positions are
shown by classical molecular graphics. However, when one displays the volume
accessible to the penetrant rather than individual polymer atoms the situation
becomes much clearer. A two-dimensional sketch of the situation is shown in
Figure 8. The matrix consists of polymer into which cavities are embedded.
These cavities always exist, whether there is a penetrant or not. They can fluctu-
ate in size and shape to some degree, but they do not move, at least not on typ-
ical MD time scales. Usually, they are not connected. However, by means of
fluctuations occurring naturally in the polymer, passages between the cavities
open for short times. If a penetrant is present in the vicinity and it happens to
have the right velocity component it can take advantage of such a channel and
slip into a neighbouring cavity before the channel closes again. This picture
(dubbed the Red Sea mechanism) has been found not only by ourselves [61],
but also by other groups [43, 62]. It has been found for polymers as different as
polyisobutylene, polydimethyl siloxane and polyethylene and for a variety of
small gas molecules. We therefore conclude, that it is the primary transport
mechanism for small penetrants in polymers.

However, there is a small addendum to this simple picture. We have monitored


the components of the gas molecule’s kinetic energy during the hopping events:
translational motion parallel to the molecular axis (axial), translational motion
perpendicular to the molecular axis (lateral), and rotational (or librational)
motion. Since bond constraints were used, the diatomic penetrants have no
internal degree of freedom. We defined as a hopping event any instance in
which a penetrant’s centre of mass moved more than 0.5 nm (O2) or 0.7 nm (H2)
in one picosecond. With this definition we scanned the trajectories of these mol-
ecules in PIB and extracted their kinetic energy components during a time win-
dow of several picoseconds before and after the jump. We thus obtained kinetic
energy profiles for the penetrants during the hopping events.

The kinetic energy of a part of the system as small as two atoms naturally has
large fluctuations. These fluctuations tend to obscure the energetics of single
hopping events. The result is that the energy profiles of all jumps look very differ-
ent from each other. In order to find some general features we take averages
over 340 jump events for H2 and 50 events for O2. These are displayed in Figure
9a (H2) and Figure 9b (O2). Because of the noise in the data these averages are
different from any individual jump event.

Permeation of Polymers 39
Motion of Gas Molecules through Amorphous Polymers

600
axial
lateral a)
rotation
total
500
Ekin / K

400

300

-6 -4 -2 0 2 4 6
t / ps
1100

axial b)
lateral
900 rotation
total
Ekin / K

700

500

300

-10 -8 -6 -4 -2 0 2 4 6 8 10
t / ps
FIGURE 9 Average kinetic energy profile around penetrant jumps for (a) H2
and (b) O2 in polyisobutylene at 300 K. The curves for the total,
lateral translational and axial translational are shifted upward by
100, 200, 300 K for clarity.

40 Florian Müller-Plathe
Motion of Gas Molecules through Amorphous Polymers

For both molecules, the total kinetic energy rises above the noise before, during
and after the jump. This implies that there is some energetic aspect involved in
the transition process: The penetrant has to acquire energy from somewhere in
order to pass through the channel between two cavities. This supplements the
conclusions found above. It may still be true that the rate-determining step is the
channel formation, but the hopping molecule also has to be “hot” - at least on
average - in order to accomplish the jump.

These curves also provide an estimate of the length of a jump event. The win-
dow of increased kinetic energy is roughly 2 ps wide for H2 and 4 ps for O2. From
these times, the lengths of the simulations and the number of penetrant mole-
cules we can calculate the probability that a given molecule is performing a jump
at the time of observation. The probabilities amount to some 2×10-2 and 3×10-3
for H2 and O2, respectively.

For both molecules we see that the kinetic energy hump occurs mainly in the
translational degrees of freedom. The rotation seems to be almost unaffected,
with the possible exception of the small drop immediately following the jump in
the case of O2 (Figure 9b). This falls within the fluctuations of the rotational
kinetic energy. However, we have also found it in simulations on O2 using differ-
ent oxygen potentials and different polymer potentials (see also ref. [57]). This
drop could be explained by a partial orientation of some of the penetrants when
they leave the channel.

In both cases the axial kinetic energy during the jump is slightly larger than the
lateral component, indicating a small preference of the penetrant for traversing
the channel in the direction of the molecular axis (“end-on”) over a motion per-
pendicular to the molecular axis (“edge-on”). Indeed, it is surprising that the dif-
ference between the two jump modes is not larger. There are two possible
explanations for this fact: Either the average channels are wide enough to allow
edge-on jumps as well as end-on jumps, or they are wide enough to allow a fair
amount of rotation or libration to persist during the jump so that the rapid redistri-
bution of energy between the two translational modes can continue. We also
noted in our preliminary study [57] that the axial and lateral peaks of the kinetic
energy can be made almost exactly equal, if one uses a more spherical O2
forcefield [63].

Permeation of Polymers 41
Motion of Gas Molecules through Amorphous Polymers

3.0

2.0
log10(〈|R(t) - R(0)|2〉/nm2)

〈|R(t) - R(0)|2〉 ∝ t0.97

1.0

0.0

-1.0 〈|R(t) - R(0)|2〉 ∝ t0.50

-2.0
-1.0 0.0 1.0 2.0 3.0 4.0
log10(t/ps)

FIGURE 10 Mean-square displacement of He atoms in amorphous


polyisobutylene (300 K). The log-log representation clearly
shows the crossover from anomalous diffusion (<100 ps) to
normal diffusion. From ref. 65.

These results lead to the question if there is any sign of the penetrant “forcing” its
way through the polymer. For a small molecule the only conceivable way of forc-
ing the opening of a channel would be by impacting on a cavity wall with high
velocity. Given the harmonic nature of the short-time motions and the disparity in
the masses it is very likely that the penetrant would bounce back to where it
came from rather than penetrate into and through the wall. And indeed, we have
never observed anything that one could call the forced opening of a channel,
except in non-equilibrium MD (see below). However, as discussed above, if a
channel has been formed it helps if the penetrant has a little extra energy to sur-
mount the remaining barrier.

4.4 Anomalous diffusion

In most cases, diffusion coefficients are calculated by fitting a straight line to the
mean-square displacement and determining its slope. As was said before, this is
allowed only in the long-time limit. At short times the mean-square displacement

42 Florian Müller-Plathe
Motion of Gas Molecules through Amorphous Polymers

0.8

0.6
〈|R(t) - R(0)|2〉 / nm2

0.4

0.2

0.0
0.0 2.0 4.0 6.0
t / ns

FIGURE 11 Mean-square displacement of O2 molecules in polyisobutylene


at 300 K. The open squares represent a linear fit to the curve,
the filled circle a nonlinear fit, the power n being roughly 0.9. On
this scale the power law cannot be determined unambiguously.
From ref. 65.

may obey a different power law. In most of the calculations of diffusion coeffi-
cients by MD it has been tacitly assumed that the long-time limit had been
reached in the calculation. If, on the other hand, one looks at mean-square dis-
placement curves (if they have been reported and not just the diffusion coeffi-
cients [22, 41, 43, 56, 64]) one often notes that the mean-square displacement is
a wobbly line which, however, can be interpreted as being straight if one regards
statistical uncertainty and insufficient simulation length as the source of the non-
linearity.

In a more recent study, however, we have pointed [65] out a case of anomalous
diffusion in polyisobutylene (PIB) near room temperature. For He in PIB anoma-
lous behaviour could be clearly shown, and the transition to normal diffusion at
around 0.1 ns could be captured. The log-log plot which shows this crossover, is
reproduced in Figure 10. For the much slower diffusing O2 the mean-square dis-
placement data were not accurate enough to determine unambiguously if the
curve was linear or followed a different power law (Figure 11). However, there is

Permeation of Polymers 43
Motion of Gas Molecules through Amorphous Polymers

no a priori reason to assume Einstein behaviour at short times for this molecule
or larger ones.

We believe that the anomalous behaviour is caused by a separation of time


scales consistent with the hopping pattern: Fast motion inside the cavity which
does not show random-walk like behaviour due to the restriction of available
space by the polymer matrix, and slow jumps which do form a random walk. It is
interesting to note that a similar pattern has recently been observed also for gas
molecules diffusing through zeolites [66]. The other possible cause for anoma-
lous diffusion, conformational changes enforced upon the diffusant by the envi-
ronment, can safely be ruled out here. Helium does not have other degrees of
freedom but translational ones. For O2, which is rigid in our simulations, the only
possible non-translational motion is rotation (or libration). The rotation happens
on a subpicosecond time scale (Section 4.3), which is orders of magnitude faster
than the crossover from anomalous to Einstein diffusion.

The results of this section have implications for both experiment and simulation.
The length scale at which the crossover occurs is of the order of a few nm. As
membrane thicknesses approach this order of magnitude one might ask if diffu-
sion coefficients determined for much thicker membranes will still be valid. For
calculated diffusion coefficients it means that they may have to be treated with
some caution, unless Einstein behaviour can be clearly established.

44 Florian Müller-Plathe
5 Gas Diffusion Coefficients

5.1 Calculation of diffusion coefficients

If simulation methods are to be useful in the area of penetrant diffusion in poly-


mers they have to be able to reproduce experimentally measured diffusion coef-
ficients with sufficient accuracy. Given the scatter among different experiments
we consider a method accurate enough if, for absolute diffusion coefficients, it
reproduces experiment to well within one order of magnitude. In certain cases,
for instance when only relative diffusion coefficients are important, a lower accu-
racy might suffice since errors may compensate.

While the earlier MD studies of penetrant in polymers were rather successful in


revealing qualitative aspects of the diffusion mechanism, most of them failed to
reproduce experimental diffusion coefficients. Calculated diffusion coefficients
always exceeded measured ones, usually by 1-2 orders of magnitude. Even in
view of the disagreement between different experiments this is too much for the
method to be useful as a predictive tool.

In a number of recent studies it has been shown that the earlier disagreements
were mainly due to shortcomings of the forcefields used. In particular, it has
been demonstrated that isotropic united-atom forcefields (representing whole
groups, such as methylene or methyl groups as one single “atom” with an appro-
priately adjusted radius) tend to overestimate diffusion coefficients [58, 64, 68].

45
Gas Diffusion Coefficients

TABLE 3. Diffusion coefficients calculated by molecular dynamics and from


experiment (in 10-6 cm2/s) at 300 K unless indicated otherwise.

polymer penetrant calc. expt. ref.

atactic polypropylene H2 44 5.7 56


O2 4.0 ~1.5 a) 56
CH4 0.48 ~0.6 a) 56
polydimethyl siloxane He 18 10 43
CH4 2.2 2.0 43
polyethylene CH4 0.5 0.3-0.6 64
CH4 1 58
polyisobutylene He 30 5.93 47
H2 9.2 1.52 47
O2 0.047-0.169 0.081 47, 68
CH4 0.4 (350 K) 1.7 (375 K) b) 58

a) In various natural and synthetic rubbers, which have gas diffusion coefficients very similar
to atactic polypropylene [25].
b) Reference 67.

Removing the artificial sphericity of united atoms, either by including explicit


hydrogen atoms or by using so called anisotropic united atom potentials [69] (the
interaction centre of the potential does not coincide with the position of the atom)
brings calculated diffusion coefficients close enough to experimental ones. Note,
however, that in the case of low-barrier polymers such as polydimethyl siloxane
(PDMS) or atactic polypropylene (PP), united-atom models often work. A selec-
tion of diffusion coefficients calculated by MD that show reasonable agreement
with experiment, is given in Table 3.

46 Florian Müller-Plathe
Gas Diffusion Coefficients

5.2 Influence of simulation parameters on calculated diffusion


coefficients

If a simulation fails to produce desired results, the forcefield is often the first sus-
pect. There have, therefore, been a number of studies of the effect of forcefield
and other system parameters on the calculated diffusion coefficients. Different
aspects have been evaluated:

Penetrant forcefield parameters. This nonbonded interaction is described by


the coefficients ε and σ (see Box 1 on page 20 for definitions). No penetrants
with Coulombic interactions seem to have been studied to date, with the excep-
tion of the ionic systems described in Chapter 8. The effect of changing ε is dis-
cussed below in some detail. The effect of increasing the penetrant’s size σ
should be the same as increasing the density of the polymer, or decreasing the
free volume. Experimentally, it has long been known that for different penetrants
in a given polymer their diffusion coefficients depend in some simple way on their
diameters [8]. Similar relationships among the diffusion coefficients of different
penetrant species have been found computationally (H2, O2, CH4 in PP [56]).
Sonnenburg et al. [23] have systematically varied the diameter of the penetrant
and found a relationship resembling an exponential form. We [68] have tested
two commonly used models for O2: one had a smaller σ (0.295 vs. 0.309 nm) but
a larger ε (0.51 vs. 0.36 kJ/mol) which gives rise to two counteracting effects.
Because of its smaller size the first O2 model should diffuse faster, but because
of its stronger interaction with the matrix it should diffuse slower. In spite of the
difference in ε being much larger than the difference in σ, the size argument
wins: The slimmer O2 diffuses faster. This indicates that radii have to be chosen
more carefully than interaction energies.

Polymer nonbonded forcefield parameters. The polymer-penetrant interac-


tion also depends on the nonbonded parameters of the polymer. The deficien-
cies of isotropic united-atom models have already been mentioned. These have
been explained [68] by the bad packing of united-atom chains. Even if the united
atom model provides the correct density or free volume it generates a different
distribution of free volume than either the anisotropic united-atom model or an
all-atom description. The latter pack more efficiently and leave smaller interstitial
sites which is the space where diffusion takes place. For a detailed discussion,
see ref. 68.

Permeation of Polymers 47
Gas Diffusion Coefficients

Polymer covalent forcefield parameters. Attempts have been made to corre-


late the diffusion coefficient with the mobility of the polymer atoms. Takeuchi and
Okazaki [22] studied the effect of altogether dropping the torsional potential in a
polyethylene chain, i.e. allowing unhindered rotation around backbone bonds.
This increased the diffusion coefficient of CH4 by only 60% at 300 K. Pant and
Boyd [58], however, found that raising the torsional barrier of PE by 50% brought
the diffusion of CH4 virtually to a standstill at 400 K. They explained the appar-
ently contradicting observations by a change in the diffusion mechanism. At
lower T the hopping mechanism prevails, the diffusion rate is determined by the
opening of channels. This is accomplished by cooperative small displacements
of polymer atoms, for which the torsional barrier is only one of many compo-
nents. At higher T, diffusion becomes more liquid-like, and the motion of individ-
ual polymer atoms becomes more important. This, in turn, is very much affected
by the torsional potential. Takeuchi et al. [70] also studied the effect of changing
the backbone bond angles in PE. They found a marked decrease of D as they
increased the equilibrium value from 100 to 150 degrees. This could be rational-
ised in terms of a stretching out of the chains which allows different chains to
approach each other more closely leaving less room between them for the pene-
trant.

Finite-size effects. Molecular-dynamics simulations use a only small cutout


from a macroscopic polymer, and the question always arises whether the small
part is large enough to yield reliable bulk properties in a simulation. Several
effects due to the finite system size are possible. The amorphous polymer struc-
ture that was generated initially may not be representative of the real microstruc-
ture at all. In other words, the bulk diffusion coefficient is an average over very
many little domains of the size we can simulate, and our structure may be very
far from the average if the distribution of diffusion coefficients in various domains
is wide. For one example, O2 in PIB, we have calculated diffusion coefficients in
four different polymer structures [68] and the results differed by no more than a
factor of 3, which is well within the errors from other sources [47]. We concluded
that the influence of the microstructure on the magnitude of the diffusion coeffi-
cient is not particularly large so that one can probably succeed with using only
one or two different starting structures. In cases where the scatter is larger, one
has to use more structures and average over them.

Another finite-size effect, however, is clearly visible in our simulations. When


introducing the diffusion coefficient in Section 2.2 on page 9, we mentioned that

48 Florian Müller-Plathe
Gas Diffusion Coefficients

macroscopic samples of amorphous polymers are isotropic to diffusion, which is


tantamount to the diffusion tensor being diagonal with all diagonal elements
being equal. However, in a small sample there may not be enough averaging for
isotropy to be reached. In other words, the diffusion tensor can be non-diagonal
and also the diagonal elements can differ. This is indeed the case, and an exam-
ple is given in Table 4 on page 55 (from ref. 47). One sees that for both equilib-
rium and nonequilibrium MD the diagonal elements of the diffusion tensor are
different, though similar in magnitude, and that the off-diagonal elements
(Table 4 gives only the average Dαβ for non-equilibrium methods) are finite,
although noticeably smaller than the diagonal elements. Moreover, the diffusion
tensor is not even symmetric in the nonequilibrium case. (For details, see ref.
47).

Other effects may arise because of the finite size irrespective of whether or not
the structure represents the average polymer. Such effects might, for instance,
be due to the pseudo-periodicity at a small length scale, which is enforced upon
the system by the boundary conditions, or to the fact that because of the small
system size one can only use short polymer chain lengths, which may not at all
capture the dynamics of real polymers. The latter effect is probably negligible,
since on the time scales accessible to MD the dynamics is very much independ-
ent of the chain length. Such finite-size effects can only be ruled out by trying.
Unfortunately, the possibilities for systematic studies are very limited because
computer time rapidly increases with system size. However, we have compared
simulations of the same system, for parent simulation cells of ~(2.0 nm)3 and
~(3.0 nm)3 and found the diffusion coefficients to agree within a factor of 2 [68].
We cannot rule out the presence of a finite-size effect which might show up only
at much larger length scales. However, at the scales accessible no finite-size
effect on the magnitude of the diffusion coefficient could be detected.

5.3 Temperature and density dependence of diffusion coefficients

The very first MD simulations of penetrant diffusion in polymers attempted to


study the dependence of the diffusion coefficient on system parameters such as
temperature [22, 58, 71] and density [23].

Studies of the temperature dependence of the diffusion coefficient D were aimed


at determining the overall behaviour (Arrhenius vs. Williams-Landel-Ferry WLF
equation: ln ( D ⁄ D 0 ) = c 1 [ T − T 0 ] ⁄ [ c 2 + T − T 0 ] ) and extracting the various

Permeation of Polymers 49
Gas Diffusion Coefficients

parameters of these models. Initially, for CH4 in polyethylene (PE), an Arrhenius


behaviour had been found with apparent activation energies of 6.4-10.8 kJ/mol
[22], 10.5 kJ/mol [71], 11.9 kJ/mol [41]. These calculations used isotropic united-
atom models for the polymer. However, in the temperature range of 140 to
180 ˚C the experimental activation energies are around 32-40 kJ/mol [67]. Fur-
thermore, if a wider temperature range is considered, the behaviour of CH4 and
many other gases in PE is found to obey the WLF equation rather than the
Arrhenius equation. In view of this, Pant and Boyd [58, 64] have recently studied
the temperature dependence of D for CH4 in PE and polyisobutylene (PIB) using
anisotropic united atom potentials. For PE they establish WLF behaviour in
agreement with experiment. For PIB, their Arrhenius plot shows a straight line to
a very good approximation. An experimental study [67] of the same system,
which they were apparently unaware of at the time, finds the same Arrhenius
behaviour in the temperature range of 101 to 188 ˚C, and the diffusion coeffi-
cients agree very well. Hence, also for this system, both the qualitative behav-
iour (Arrhenius) and the quantitative diffusion coefficients were correctly
“predicted”. In retrospect, the disagreement of the earlier calculated temperature
dependencies of D with experiment are attributed to the isotropic united-atom
models.

The dependence of D on the density (or equivalently the pressure) has usually
been analysed in terms of free-volume Vf. The free volume fraction is the volume
not occupied by polymer atoms divided by the total volume. There is some arbi-
trariness involved in the definition of free volume (choice of atomic radii, free vol-
ume vs. penetrant-accessible volume, etc.). The results agree, however, in that
D increases with available free volume. The relationship can often be fitted by
the simple exponential form known from free-volume theories [22, 23, 58, 70]

D = Aexp ( − B ⁄ V f ) (Eqn. 37)

where A and B are constants. Comparisons with experiment have not been done
since the polymer density range that is easily accessible in a simulation is often
difficult to achieve experimentally.

A related observation can be found in Table 3. The higher of our two values for
the diffusion coefficient for O2 in PIB [47] corresponds to a constant-volume
(canonical) simulation at the experimental density of PIB, the lower to a con-
stant-pressure simulation. We believe that this decrease is not so much brought

50 Florian Müller-Plathe
Gas Diffusion Coefficients

about by the different statistical-mechanical ensemble as such but by an


increase in density by 3% when the constant-volume constraint is relaxed.

5.4 Where can molecular dynamics be usefully applied?

It is also interesting to estimate the range where the MD method can be confi-
dently used in the calculation of diffusion coefficients (this is after assuming that
technical problems like forcefields, anomalous diffusion etc. have been sorted
out). Firstly, there are general considerations: one has to be able to simulate a
system large enough to sample the configurational statistics of the polymer suffi-
ciently. In other words, to obtain a reasonable cross section of, say, polyethylene
configurations one may need a few hundred repeat units or a few hundred to a
few thousand atoms. One might need many more repeat units if for instance the
polymer is very stiff. Equally, if the monomers are large, one will need many
more atoms. Secondly, there are factors arising from the mobility of the pene-
trant itself: At equilibrium and assuming hopping motion the diffusion coefficient
can be written as

δ2
D = (Eqn. 38)

where δ is the average jump distance and τ is the average residence time
between jumps. For sufficient statistics on the jumps one needs to sample, say,
10 jump events for a single penetrant (this is probably the bare minimum).
Assuming a mean jump distance of 0.5 nm one finds, that in a 1-ns simulation
one will encounter 10 jump events on average, if the diffusion coefficient is
roughly 4×10-6 cm2/s. Our simulation lengths are usually of the order of 10 ns
which brings the diffusion coefficients of 4×10-7 cm2/s within reach. A further
improvement is possible if one can use several penetrants at the same time and
thereby improve sampling. This is usually safe if the diffusion coefficient is not
very sensitive to the penetrant concentration as is often the case if the polymer
is rubbery and the penetrant concentration is low and the penetrants have no
long-range (electrostatic) interactions.

The foregoing argument indicates that MD can be usefully employed if the pene-
trant is a small gas molecule or if the temperature is high and the polymer has a
reasonably low barrier. At room temperature, MD is therefore useful to predict
absolute diffusion constants for gases for example in siloxanes, natural and syn-

Permeation of Polymers 51
Gas Diffusion Coefficients

thetic rubbers and polyolefines. However, the study of good barrier materials
such as polyvinylidene chloride (PVDC, Saran), that have diffusion coefficients
between 10-9 and 10-12 cm2/s, at room temperature is certainly out of the ques-
tion for several generations of supercomputers to come. Hence, the role of MD
as a tool will probably be more in areas where medium to high material through-
put is required, for instance separation membranes, rather than in the design of
new barrier materials.

5.5 Can MD methods be extended to slower diffusion processes?

Because of the long simulation times there have been a few attempts to acceler-
ate the calculations, essentially by accelerating the diffusion rate in a controlled
way [24, 47]. (General tricks for saving time in MD simulations are reviewed in
ref. 72. Many of these have already been incorporated into our MD program, see
Section 3.4.) The first attempt involved scaling down the interactions between
the penetrant and the polymer. The reasoning behind was as follows. Diffusion
can be considered an activated process which is described by an Arrhenius
equation

D = D ∞ exp ( − E A ⁄ ( k B T ) ) (Eqn. 39)

where EA is the activation energy and D∞ the would-be diffusion coefficient at


infinite T. Even if more complicated relations between diffusion coefficient and
temperature are found, for example Williams-Landel-Ferry behaviour [41], for a
small enough temperature range the Arrhenius equation should be a reasonable
approximation. It is then further assumed that the activation energy scales line-
arly with the interaction potential. For a Lennard-Jones potential (Box 1 on page
20) this is a good approximation since the magnitude of the ε parameter scales
the whole potential curve uniformly. Hence, the energy difference between a
penetrant in a cavity and a penetrant on a transition state (which is given by EA)
should scale with ε. If λ is a potential scaling factor which becomes 1 if the full
potential is used, then a simulation with 0<λ<1 can be run which should give a
diffusion coefficient D(λ) > D(λ=1). Invoking above scaling relation we have for
D(λ)

D ( λ ) = D ∞ exp ( − λE A ⁄ ( k B T ) ) (Eqn. 40)

and

52 Florian Müller-Plathe
Gas Diffusion Coefficients

-2.0

-4.0
ln D(λ)

-6.0

-8.0

-10.0
0.0 0.2 0.6 0.40.8 1.0
λ
FIGURE 12 Dependence of the diffusion coefficient of CH4 in united-atom
polyethylene on the factor λ used to scale the interaction energy
ε between the CH4 and the polymer atoms. From ref. 24.

D ( λ)
= exp [ − ( 1 − λ ) E A ⁄ ( k B T ) ] (Eqn. 41)
D ( 1)

Hence, the unscaled diffusion coefficient can be calculated from D(λ), if EA is


known. EA can be obtained by calculating D(λ) at several values of lambda and
numerically taking the derivative

d ln D ( λ ) = − E ⁄ ( k T ) (Eqn. 42)
dλ A B

Notice that scaling the potential down by λ corresponds to scaling the tempera-
ture up by 1/λ. However, this form of “temperature scaling” is selective, since it
only influences the motion of the penetrant, not the motion of the polymer matrix:
The penetrant motion is mainly influenced by polymer-penetrant interactions
whereas the polymer dynamics is dominated by polymer-polymer interactions,
because there are many more of them.

Permeation of Polymers 53
Gas Diffusion Coefficients

Is the method useful in practice? The range where it works can easily be
checked by doing the calculations for different values of λ. The equations imply a
linear relationship between lnD(λ) and λ and we simply have to decrease λ until
this linear relationship breaks down. This is illustrated in Figure 12: For methane
in polyethylene (both described as united-atom models [24]) it is found that λ can
be as small as 0.1-0.2 and the linear relation still holds approximately. If λ is
taken smaller than this, the plot of lnD(λ) vs. λ becomes highly nonlinear. It is
also not possible to reconcile this behaviour in terms of the WLF equation. This
rather dramatic change therefore has to be interpreted as a qualitative change of
the diffusion mechanism. With so small an interaction, the penetrant not only
explores existing paths through the polymer more efficiently (this is what we
want) but also finds new pathways which are probably unphysical. However, with
λ=0.1-0.2 diffusion coefficients are obtained which were about 4-5 times higher
than those with λ=1 for CH4 in polyethylene. Hence, it seems that this method
could be useful for extending the range of MD calculations moderately.

In this context it is interesting to note that a method similar in spirit has been
used to study the motion of CO in myoglobin [73]. There, CO-myoglobin interac-
tion potentials were scaled down by 1/60, and 60 mutually noninteracting CO
molecules were simulated at once. Although the study was only aimed at qualita-
tive aspects (mainly escape routes out of the binding site), our findings dis-
cussed above cast some doubt on these results.

Our other attempt to accelerate penetrant diffusion was more recent [47]. It used
the nonequilibrium molecular dynamics (NEMD) method described above (see
Section 3.2). NEMD has been very fruitfully applied in the efficient calculation of
transport coefficients, including diffusion coefficients, in liquids [15, 33]. How-
ever, for penetrants in polymers we found the method less useful. It relies on the
relationship between the magnitude of the external force and the resulting flux
being linear. If the force is chosen too small, it does not speed up the diffusion
significantly. If it is chosen too large, the linear-response relation breaks down. It
is therefore important to find out the largest force which still leaves the system in
the linear-response regime. This usually involves a fairly substantial number of
test runs. Then, while an equilibrium simulation provides the whole diffusion ten-
sor as the result of one simulation there has to be one run for each of the carte-
sian directions in the NEMD case. (Although amorphous polymers are isotropic
on a macroscopic level, a tiny sample of an amorphous polymer generally is not,

54 Florian Müller-Plathe
Gas Diffusion Coefficients

TABLE 4. Gas diffusion coefficients D (10-6 cm2/s) in polyisobutylene calculated


by equilibrium and nonequilibrium methods [47]. For comparison of
computational costs, the lengths of the simulations are also indicated.

Polymer He H2 O2
(ensemble)

Equilibrium MD
Sample 1 Dxx 34.4 2.89 0.196
(NVT) Dyy 45.6 16.8 0.255
Dzz 12.5 7.79 0.107
D 30.1 9.15 0.169
length(ns) 7.0 4.0 8.1

Sample 2 Dxx 0.0355


(NPT) Dyy 0.0759
Dzz 0.0291
D 0.0468
length(ns) 9.6

Nonequilibrium MD
Sample 1 Dxx 57.5 0.647
(NVT) Dyy 69.3 0.875
Dzz 31.7 0.187
D 52.8 0.570
Dαβ 4.27 -0.153
length(ns) 40 × 0.03 44 × 0.2
Expt. [67] D 5.93 1.52 0.081

see Section 5.2. This is in marked contrast to liquids whose fast relaxation times
ensure isotropy also on the microscopic scale.)

For the very small penetrant He we have found that NEMD saved some time
over conventional MD, whereas for the larger O2 both methods used about

Permeation of Polymers 55
Gas Diffusion Coefficients

0.010

0.008

0.006
〈vx〉 / nm ps-1

0.004

0.002

0.000

-0.002
0.0 10.0 20.0 30.0 40.0 50.0
ext -1
(F )x / kJ mol nm-1

FIGURE 13 Average x component centre-of-mass velocity of O2 molecules in


polyisobutylene at 300 K in response to an external force in x
direction [47]. The error bars indicate the scatter (standard
deviation) over several 200 ps runs at a given external force.

equal amounts of time. The reason for this is basically that the linear response
relation breaks down before the external force is large enough to induce a pene-
trant flux large enough to make the method worthwhile. The breakdown can be
understood as follows: for a small external force the penetrant will be pushed
towards the cavity walls and it is more likely that it has a velocity component in
the right direction to take advantage of a channel that opens. If the force is too
large, then the penetrant is able to break open a channel by constantly being
pressed into the cavity wall. This alters the mechanism to such an extent that the
response is no longer proportional to the force. The behaviour is displayed in
Figure 13: Up to an external force of 25 kJ mol-1nm-1 there is a linear relation
between external force and induced flux. Also, up to this force the error is rea-
sonably close to the inherent scatter of the average velocity at zero external
force. Beyond 25 kJ mol-1nm-1 the linear response breaks down, and Eqn. 31 is
no longer applicable.

56 Florian Müller-Plathe
Gas Diffusion Coefficients

-4.0
H2
-4.5
log10(D/cm2s-1)

-5.0

-5.5 O2

-6.0 CH4

-6.5

-7.0
0.20 0.25 0.30 0.35 0.40
σ/nm

FIGURE 14 Relationship between diffusion coefficient and penetrant size:


different penetrants in atactic polypropylene at 300 K [56]. The
straight line is a least-squares fit to the data points.

In addition to these problems, we found that NEMD diffusion coefficients, also


the individual components of the diffusion tensor, exceeded equilibrium MD dif-
fusion coefficients for all the cases we studied (Table 4, [47]). This is probably
due to a heating of the penetrants caused by the external force. While the ther-
mostat is quite able to keep the system as a whole at the desired bath tempera-
ture, the exchange of heat between the penetrant and its surroundings is not
sufficiently fast. As a result, the penetrants heat up selectively, whereas the rest
of the system has the correct temperature. An external force of 25 kJ mol-1nm-1
is capable of causing an increase of the average temperature from 300 K to
345 K for He atoms in PIB and to 305 K for O2 atoms, if a temperature coupling
time of 1 ps is used (Eqn. 28). We therefore conclude that the NEMD method in
its present form will probably not help to attack polymer-penetrant systems with
slower diffusion coefficients. However, the NEMD method may be more useful at
very high temperatures where the diffusion mechanism probably changes from
hopping to more liquid-like patterns [58].

Permeation of Polymers 57
Gas Diffusion Coefficients

Finally, one can use empirical relationships to extend the range of MD simula-
tions. The dependence of D on the size of the penetrant has been mentioned:
Experimentally, it is often found that the logarithm of the diffusion coefficient is
linearly related to the diameter of the penetrant [8, 9], where the “diameter” is
subject to varying definitions. Such relationships can be reproduced in MD calcu-
lations. An example is shown in Figure 14. If a calculation of small penetrants
(say He, Ne, Ar, CH4) is feasible for a given polymer and if one can assume that
there are no changes in the basic diffusion processes, one could probably
extrapolate to larger approximately spherical penetrants (Xe, SF6) with some
confidence. This assumption has, however, not been put to the test.

Another way is to study analogous systems. We have given an example [47]: In


order to understand what makes polyvinylidene chloride a better barrier than
polyvinyl chloride, we did not study this system as this would not be feasible
because of the low permeabilities. Instead we studied the structurally similar sys-
tem polyisobutylene and polypropylene (Cl exchanged for CH3) which exhibit a
similar relation regarding their barrier properties, but at a much higher overall dif-
fusion coefficient. One has, however, to be lucky to find such a suitable analo-
gous pair of polymers with higher diffusion coefficients. Hence, the study of
structurally related systems will be useful only for a few systems.

58 Florian Müller-Plathe
6 Gas Solubilities

There have been several attempts to calculate gas solubilities by the test-parti-
cle insertion method outlined in Section 3.3 on page 26. The trajectories of the
pure polymer were generated by molecular dynamics simulations. The simula-
tions cover atactic polypropylene (PP) [74], polydimethyl siloxane (PDMS) [43]
and polyisobutylene (PIB) [47] and a number of light gases: He, H2, N2, O2,
CH4. Note that all these polymers are rubbery at room temperature so that Hen-
ry’s law should be a good approximation, even up to fairly high pressures. For
example, an experimental solubility study of methane in PIB showed linear
increase of the methane concentration in the polymer with its partial pressure up
to 200 atmospheres [67].

In Table 5 we summarise the results of all gas solubility calculations by particle


insertion to date along with experimental solubilities, where known. Original data
have been converted to Henry’s constants H (in bar-1) for comparison (for the
definition, see Section 2.4 on page 16). Except for the solubilities for PDMS all
results are from our own work.

For PIB we have tried two different starting structures, and for one of them we
also compared constant-volume (at the experimental density) and constant-
pressure (atmospheric pressure) simulations. For all three guest molecules (He,
H2 and O2) we note that the specific polymer structure has a larger influence on

59
Gas Solubilities

TABLE 5. Calculated gas solubility in various polymers, given as Henry’s


constant H (in bar-1) at 300 K.

Sample Ensemble He H2 N2 O2(1)a) O2(2)b) CH4

atactic polypropylene (PP) [74]


NVT 0.12 0.78 1.5 4.0 10
expt. [25] 0.015c), 0.13c) 0.055d)
0.00014d)

polydimethyl siloxane (PDMS) [43]


NVT 0.20 11
expt. [25] 0.046 0.50

polyisobutylene (PIB) [47]


1 NVT 0.089 0.93 7.0 3.2
2 NVT 0.19 1.5 18 9.7
2 NPT 0.18 1.9 21 11
expt. [25] 0.011 0.036 0.12 0.12
a) Oxygen Lennard-Jones parameters: ε=0.5122 kJ/mol, σ=0.295 nm, bond length:
0.121 nm [75].
b) Oxygen Lennard-Jones parameters: ε=0.363 kJ/mol, σ=0.309 nm, bond length:
0.10166 nm [63].
c) Polyethylene-PP copolymer with 31 methyl groups per 100 carbon atoms.
d) PP, 50% crystalline.

the solubility than the choice of the ensemble. However, deviations between gas
solubilities in the two structures are not very large either.

For O2 in PIB, two sets of force field parameters have been tried, as mentioned
in Section 5.2. They give rise to marked differences in the solubility. One of them,
marked O(1), was optimised to give correct structures for solid O2 [75], the other
one O(2) to yield correct excess properties in liquid mixtures of O2 and noble
gases [63]. The main difference between the two parameter sets lies in the Len-
nard-Jones energy parameter (see footnotes to Table 5): O(1) has a larger ε
and, hence, a stronger interaction than O(2); O(2) has a larger σ which is, how-

60 Florian Müller-Plathe
Gas Solubilities

ever, partly made up for by a reduced bond length, so that O(2) is slightly fatter
than O(1), but not longer. The O(2) forcefield produces a slightly smaller dis-
agreement with experiment which reflects the purpose it was designed for.
Reduction of the interaction strength therefore leads to a reduction in solubility.
For small gas molecules in the absence of specific interaction sites in the poly-
mer, there is a strong connection between the Lennard-Jones ε parameter and
the calculated solubility. As we saw in Sections 5.2 and 5.5, the diffusion coeffi-
cient, on the other hand, is mainly governed by the particle’s diameter σ.

Normally, the ordering of solubilities among gases in a given polymer is repro-


duced correctly (Table 5). An apparent exception is polypropylene. Here, the
experimental values are a bit erratic since they were obtained for two different
polymers, none of which is actually atactic polypropylene, for which no measure-
ments seem to be available. The partially crystalline polypropylene must contain
a substantial fraction of isotactic material and the copolymer contains a smaller
fraction of polyethylene as well as polypropylene. However, within each polymer,
also the experimental values show a consistent ordering. The ability to repro-
duce the correct ordering among penetrants is some indication that the method
might prove useful to predict sorption characteristics at the level of semiquantita-
tive comparison. As was said before (Section 3.3 on page 26), the computation-
ally intensive part of the particle-insertion method is the calculation of the pure-
polymer trajectory. Once this exists, it is very cheap to calculate solubilities for a
large variety of small penetrant molecules, which can be used for comparison.

In absolute numbers, all calculated solubilities exceed the corresponding experi-


mental values by 1 to 2 orders of magnitude. At first sight, this disagreement
looks awful. Translating Henry’s constants into excess free energies, however,
one finds that our calculated chemical potentials exceed measured values by 5-
8 kJ mol-1. This kind of error is not uncommon in calculations of free energies by
molecular simulation. One also has to remember that the solubilities are related
to absolute free energies, rather than relative free energies and are more difficult
to determine by simulation. Relative free energies, on the other hand, are suffi-
cient in many other application areas. Researchers are, for instance, rarely inter-
ested in the absolute value of the binding constant in a drug-receptor interaction,
since it is often enough to know if drug A binds better than drug B and by how
much.

Permeation of Polymers 61
Gas Solubilities

10.0

a)
He
5.0
∆A / kJ mol-1

H2
0.0
N2 O2

-5.0
CH4
-10.0
0.0 0.5 1.0 1.5
6.0

4.0 b)
He
2.0
∆A / kJ mol-1

0.0
H2
-2.0

-4.0 O2(1)
O2(2)
-6.0

-8.0
-10.0
0.0 0.2 0.4 0.6 0.8 1.0
Σε / kJ mol-1

FIGURE 15 Free energy of absorption depending on interaction with the


polymer matrix. As a measure of the interaction strength the sum
of all Lennard-Jones ε’s of the gas molecule is taken. (a) Gases in
atactic polypropylene [74], (b) gases in polyisobutylene [47]. For
the two different oxygen models, see footnotes of Table 5.

62 Florian Müller-Plathe
Gas Solubilities

Relative free energies correspond to ratios of Henry’s constants, and if we look


at those ratios we find that they typically agree with experimental ratios to with a
factor of 2-5. While closer agreement with experiment is certainly desirable, this
is already much better than the disagreement for the absolute Henry’s con-
stants. Further tests are relationships between the free energy of absorption and
the strength of the interaction between the gas molecule and the host polymer.
Assuming that mixing rules of the type given in Box 1 on page 20 are applicable
for obtaining Lennard-Jones coefficients for interactions between polymer atoms
and penetrant atoms, the ε of the gas should be a measure of this interaction
strength. It has long been known from experiments that for a given polymer
there is a linear relationship between the free energy of absorption and the ε of
the gas, or alternatively its boiling temperature or critical temperature [8]. We
show such plots in Figure 15 for both atactic polypropylene (a) and polyisobuty-
lene (b). The solid lines are least-squares fits to the data points. The relations
are linear to a good approximation. This hints at the possibility of extrapolating
solubilities of larger penetrants, which are intractable by MD simulation, from a
series of calculations for small molecules. The idea behind this is very similar to
the extrapolation for diffusion coefficients introduced in Section 5.5. The only dif-
ference is that for the diffusion coefficient a scaling with the penetrant’s σ is
assumed, whereas for the solubility a scaling with its ε is more appropriate.

One may speculate what the reason could be for the disagreement between cal-
culated solubilities and experiment. We can rule out the particle-insertion
method as such. Firstly, it is possible to keep inserting until the free energy is
converged. In some cases we have converged the free energy for every polymer
configuration and we know that our problems are not due to poor sampling dur-
ing particle insertion. Secondly, the condition under which the method is bound
to fail is the application to excessively large penetrants. As pointed out in Sec-
tion 3.3, this is because the polymer - left to itself - will not create holes big
enough to hold large penetrants. This should result in solubilities which are too
small compared to experiment. However, our solubilities all are far too large.
Moreover, when it was tried to extend the method to larger and more flexible
molecules [76] it was found that the method would work for molecules as big as
propane.

The second possible source of discrepancy with experiment is the forcefield


used to describe the polymer-penetrant interaction. Although we find the strong
correlation between solubility and potential energy function mentioned above,

Permeation of Polymers 63
Gas Solubilities

we do not believe that this is to blame either. For instance, in PIB we find that the
calculated solubility for He roughly equals the experimental solubility of O2. This
would mean that in order to calculate a correct O2 solubility one would have to
assign forcefield parameters corresponding to He, or possibly mutilate the O2
model in some other totally unphysical way. We also do not believe that a
change in the functional form of the nonbonded interactions, like an r-9 or expo-
nential repulsion, would be sufficient to lower the solubilities substantially.

Examining our simulations in more detail we found that the bulk of the solubility
is contributed by one big hole in the polymer. In one PIB simulation, for example,
this cavity is large enough to hold a whole propane molecule and it is compen-
sated by a higher density elsewhere. One can imagine that real amorphous poly-
mers show a density that is more uniform at the microscopic level in which case
there would probably not be such large holes and the solubilities would be lower.
Or at least, in a polymer sample as small as our primary simulation cell, the prob-
ability of finding so large a cavity should be essentially zero. We think that the
presence of large cavities in the structures, which are probably not present in
reality, is most likely the cause of the disagreement with experiment. The reason
for the inhomogeneous density responsible for the occurrence of holes could be
either the forcefield parameters describing interactions within the polymer or the
starting structures. It has been brought to our attention that within the framework
of a different atomistic computational technique reasonable agreement with
experimental solubilities could be obtained by using one of the commercial force-
fields which has much larger atom radii [77]. Larger radii of polymer atoms per
se already reduce the size of cavities. Cavity sizes are reduced even further
because polymer chains try to avoid each other more and expand into pockets of
free volume. However, the price to pay was that this particular forcefields leads
to an expansion of the polymer so that its calculated density became some 20%
lower than its experimental density. It therefore seems that increasing atomic
radii and clamping the system at experimental density is more of a cure of the
symptoms rather than of the underlying problem.

We rather think the way forward is to generate more realistic polymer starting
structures. The modified rotational isomeric state method used in our work (and
the soft-core molecular dynamics used in the work on PDMS [43]) apparently
provide starting structures which are good enough to calculate diffusion coeffi-
cients. However, the quality of the initial structures might not be good enough for
the calculation of solubilities. Since, even over the relatively long simulation

64 Florian Müller-Plathe
Gas Solubilities

times of around 10 ns not much structural relaxation is going on in the polymer,


deficiencies of the starting structures will persist during the whole simulation.
While it is not guaranteed that other methods (see Box 2 on page 23) will gener-
ate structures better suited for solubility calculations, it would certainly be a good
idea to try out several of them in order to put this hypothesis to the test. Such
empirical test of starting structures, which involves calculating several properties
of interest from them and comparing to experiment, is probably the only way to
assess their quality. It appears to be very difficult to come up with hard structural
quantities against which they could be judged [36].

Permeation of Polymers 65
Gas Solubilities

66 Florian Müller-Plathe
7 Semicrystalline and
Filled Polymers

So far, we have only worried about solvation and transport in amorphous poly-
mers. The reasons for this were given in the introduction: sorption and transport
mainly happen in the amorphous phase of a polymer so that the contribution of
the amorphous volume fraction will dominate these quantities. The reality of the
polymer world, on the other hand, is that many technical polymers are semicrys-
talline and that partial crystallinity is responsible for small but important effects
on these properties, even if the amorphous region contributes the lion´s share. In
addition, technical plastics often contain organic or inorganic fillers (silica, alu-
mina, mica, talc, sand, cellulose, sawdust, other polymer, ...) which are particles
whose diameter typically is a few microns, dispersed in the polymer. They will
also have an effect on the permeability. Hence, it would be nice if the range of
molecular dynamics techniques developed in the preceding chapters could be
extended to semicrystalline or filled polymers.

Straightforward atomistic MD simulations of a semicrystalline material are out of


the question altogether, since crystallite dimensions range from, say 10 nm to
well above microns. To make things worse, crystallites often aggregate into
larger domains of parallel packing or into spherulites, which can have macro-
scopic dimensions. In contrast, typical MD simulations use parent cells with a
length of 2-5 nm. One, therefore, has to adopt a less microscopic viewpoint and
develop an analytical or computational model which accounts for the modifica-

67
Semicrystalline and Filled Polymers

tion of the permeability of the amorphous material by the presence of crystallites


or filler particles.

7.1 Two-phase models

The extent to which semicrystallinity influences permeation properties depends


on many details of the polymer-penetrant system under study [10]. Very complex
phenomena can arise, which are not easily handled by simple models. We can
only develop approximate models for dealing with semicrystallinity if we make a
number of simplifying assumptions:
• The penetrant is only soluble in the amorphous regions of the polymer. This
assumption is reasonable if chain packing in the crystallites is dense enough
so as not to allow penetrant molecules to enter these regions. While this is
never rigorously true, penetrant solubilities in crystalline regions are often
several orders of magnitude smaller than in amorphous regions. An example
for a violation of this assumption are large penetrants that are commensurate
with the polymer crystal structure. One can, for example, imagine that a long
alkane chain happily integrates into a crystal of polyethylene.
• A less stringent assumption than the above is that, while the penetrant may
have a low but not negligible solubility in the crystallites, its mobility in the
crystalline phase is significantly lower than in the amorphous phase. This
assumption leads to a concept of penetrant immobilisation: The crystallites
might contain an appreciable amount of penetrant but this amount is trapped
and will play no significant role in the transport. In a crystal of tightly-packed
chains diffusion will involve crystal defects. If the defects are sufficiently sepa-
rated penetrant hopping between them will be very rare, and so will be any
transport which involves defect migration. Therefore, it is not unreasonable to
assume that penetrants have a lower mobility in the crystalline than in the
amorphous phase.
• The penetrant does not alter the properties of the polymer, neither in the crys-
talline nor in the amorphous phase. In particular, the penetrant at the concen-
tration under consideration must not swell the polymer. Swollen polymers
often offer lower resistance to a diffusant than unswollen ones. The amor-
phous and crystalline volume fractions are swollen to very different extents by
a given penetrant concentration. Finally, crystallites act as effective cross-
links also for the polymer chains in the amorphous phase thus inhibiting

68 Florian Müller-Plathe
Semicrystalline and Filled Polymers

swelling to some degree, so that the swelling of the amorphous part of a


semicrystalline polymer may depend strongly on the degree of crystallinity. All
these effects make the theoretical treatment of diffusion in semicrystalline
polymers very complicated if swelling occurs. To a lesser extent, transport
properties can be modified by plasticisation of the polymer (softening by intro-
duction of low-molecular-weight compounds). Hence, if the penetrant is a
good plasticiser and its concentration is high enough, similar complications
might be expected. However, for small penetrants at sufficiently low concen-
trations one can expect that both effects will be negligible.
• The transport properties of the amorphous phase are unaffected by the pres-
ence of crystallites. There have been speculations as to how the crystallites
might influence the structure and the mobility of polymer chains in the amor-
phous phase [78]. The “cross-linking” of polymer chains by crystallites has
already been mentioned. Fixing the chain in a crystallite may reduce the
mobility of the part that extends into the amorphous phase. If the chain mobil-
ity is crucial for the penetrant transport, i.e. for large penetrants, it will have an
effect. For the small penetrants studied here it is probably less of a problem,
since penetrant hopping is facilitated by local fluctuations of the polymer
atoms (section 4.3), which are largely unaffected by the polymer being tied up
in a crystallite a few monomers along the chain. Historically, however, a
“chain immobilisation factor” has been introduced also into the treatment of
small-molecule permeation [79, 80]. This empirical parameter was introduced
to explain the disagreement between the predictions of geometrical models
and experiment.
There have been speculations about an additional effect in the interface layer
between amorphous and crystalline phase. One can imagine that, close to
the interface, the crystallites perturb the amorphous phase sufficiently so that
the polymer cannot reach its amorphous density. In this case there could be
an increased diffusivity along the interface. As far as experimental results go
it has been found that a three-phase model (the interfacial region being the
third phase) gives a better fit to experimental permeabilities in a lamellar poly-
styrene-polybutadiene system than a simple two phase model [81]. In the
area of inorganic superionic conductors it has been found that addition of an
insulating material (Al2O3) to a conducting phase (LiI) significantly increases
the conductivity, which can only be explained by a highly enhanced conduc-
tivity along the interfaces [82]. While it would be easy to build special inter-
face properties into coarse-grained models of penetrant diffusion, it is difficult

Permeation of Polymers 69
Semicrystalline and Filled Polymers

to predict a priori the effective diffusion coefficients in the interface layers,


since their atomistic structure is not known. To our knowledge, atomistic com-
puter modelling of interfaces between amorphous and crystalline regions in
polymers has not yet been done.

The assumptions above lead to a model of two independent phases, one amor-
phous and one crystalline, and to simple expressions relating the solubility in the
semicrystalline material to the known solubilities in the two phases. The semi-
crystalline solubility Ssemi becomes an average of the amorphous and crystalline
solubilities, Samorph and Scryst appropriately weighted with the respective volume
fractions. In the limit of the crystalline solubility being negligible Ssemi becomes

S semi = v amorph S amorph , (Eqn. 43)

where vamorph is the amorphous volume fraction, i.e. the volume occupied by
amorphous polymer divided by the total volume.

Within the two-phase model a number of analytical expressions have been


developed which relate the diffusion coefficient of the semicrystalline material
Dsemi to the known diffusion coefficient in the amorphous material Damorph. All of
them have to make further simplifying assumptions about the geometry and
arrangement of the crystallites: They have to have regular shapes (spheres,
ellipsoids, rectangles, etc.), they often have to be all of the same size and shape,
and they have to be either arranged in a regular lattice or their concentration has
to be very low. An overview of the analytical models can be found in ref. 83. To
summarise, it can be said that the existing analytical models are only capable of
handling a number of cases which are restricted in size, shape, arrangement or
concentration of the obstacles, and that they have not been generalised to
account for the large variety of semicrystallinity found in real polymers.

7.2 A Monte-Carlo strategy

We have developed a simple lattice Monte Carlo scheme to demonstrate that a


general computational procedure can be found. In our pilot study [84] the model
was restricted to two dimensions, not so much because of the computational
cost (these calculations are orders of magnitude cheaper than MD simulations)
but because the systems and results can be visualised more easily. The system
is described as a 1024×1024 square lattice. A single random walker moves by

70 Florian Müller-Plathe
Semicrystalline and Filled Polymers

hopping between adjacent lattice positions. The step direction of the random
walker is chosen from the four possibilities with equal probability. In the absence
of any obstacles, the walker performs normal diffusion on the lattice and the dif-
fusion coefficient is given analytically by

δ2
D0 = (Eqn. 44)

with δ being the lattice spacing, which we also take as the walker’s diameter, and
τ is the time for one step. We adopt δ and τ as units of length and time and set
them both to 1. The factor 4 in Eqn. 44 rather than 6 (see Eqn. 38) arises
because the diffusion is two-dimensional rather than three-dimensional.

Rectangular obstacles are placed onto the square lattice. These are simply
taken as inaccessible space: If the random walker tries to move onto a lattice
site which is marked as obstacle the move is rejected and a new move is tried.
The time increases by τ irrespective of whether or not the trial move is success-
ful. The connection with semicrystalline or filled polymers is made by regarding
the available space as the amorphous polymer (with a diffusion coefficient
Damorph=D0) and the obstacles as crystallites or filler particles embedded into the
amorphous background. The model provides only a geometric impedance to the
random walker, none of the other effects listed above is being taken into
account. It would be easy to incorporate effects like interface-enhanced diffusion
or chain-immobilisation by assigning position-dependent diffusion coefficients to
areas near the obstacles. However, the difficulty is obtaining these parameters in
the first place.

One can get an idea of the absolute time scale and length scale by going back to
the jump model of diffusion and recognising that δ is the mean jump length and τ
is the average time between jumps. For a gas molecule, δ might be of the order
of 0.5-1 nm, which yields a system size of approximately 0.5-1 µm. The waiting
time between jumps is at least a few tens of ps. (It is not bounded from above.)
Trajectories of 107 steps can easily be calculated on a small workstation so that
effective times of milliseconds can be reached. If, on the other hand, one is
merely interested in how a known D0 (for example, from MD simulations) is mod-
ified by a certain arrangement of crystallites it is sufficient to calculate a relative
diffusion coefficient D/D0. One then does not have to worry about absolute times
and lengths, and results become generic in that they only depend on the size,

Permeation of Polymers 71
Semicrystalline and Filled Polymers

FIGURE 16 Two-dimensional periodic 2-phase model of a semicrystalline


polymer. The black obstacles represent crystallites and they are
taken to be impermeable to penetrants. The obstacles have
been placed so as to guarantee openings of at least 5 penetrant
diameters between them. To the left regular square obstacles
(50×50 penetrant diameters) at 44% coverage are shown. The
obstacles on the right have the same area but a larger
anisometry (10×250) and lower coverage (32%). A tendency to
form domains of parallel crystallites, that can be seen here, is a
consequence of the algorithm for placing the obstacles.

shape and arrangement of the obstacles and not any longer on the nature of the
polymer or penetrant. The diffusion coefficient in the presence of obstacles is
calculated by the two-dimensional Einstein equation, cf. Eqn. 11

4Dt = 〈 R ( t ) − R ( 0 ) 2 〉 (Eqn. 45)

In order to use the model on a realistic polymer one would need to input a typical
crystallite distribution. This is difficult to predict a priori but can be obtained from
electron micrographs of the polymer. For our purposes, we use an algorithm
which distributes obstacles of specified size and aspect ratio randomly in the
system, either to a specified concentration or to the maximum possible concen-
tration. We impose the additional constraint that between any two obstacles
there is a gap of at least 5δ. This is to avoid obstacles touching each other which
ultimately could lead to the percolation of obstacles with the trivial result D=0.

72 Florian Müller-Plathe
Semicrystalline and Filled Polymers

Two examples are shown in Figure 16. It is generally the case that the algorithm
reaches higher densities for regular obstacles than for very anisometric ones.
The properties of this algorithm were later investigated more thoroughly [85].

With the simple model a number of interesting results have already been
obtained [84]:
• The effectiveness of crystallites in slowing the diffusion depends on their
area: At a given crystalline volume fraction vcryst, the larger the crystallites,
the less effective they are as a barrier. Also, an experimental relation [79] has
been confirmed

D 0 ⁄ D = ( v amorph ) − n (Eqn. 46)


in which the exponent n appeared to be specific to the polymer-penetrant
system under study. The authors found two different exponents for linear and
branched polyethylene which they ascribed to different anisotropies of the
crystallites. We obtained the same result (different n) for obstacles which
were all square but had different sizes (see Figure 17). Also, the size depen-
dence provides a simple explanation for another experimental finding which
is discussed in detail in section 4.3 of ref. 11: While in general it is true that
with an increasing degree of crystallinity penetrant diffusion coefficients
become smaller, the contrary has been found for gases in polypropylene:
when a quenched polymer sample was subsequently annealed the diffusion
coefficients increased while the crystallinity increased as well. This was
explained by the formation of amorphous pores percolating through the crys-
tallite while the crystallite is thickening. Our model, on the other hand, sug-
gests that large crystal defects do not have to be involved. If the increase of
crystallinity happens together with the growth of large crystals at the expense
of smaller ones, then the effect due to increasing crystal size might override
the effect due to increasing crystalline volume fraction.
• The barrier properties of crystallites depend on their shape. The more aniso-
metric the crystallites are, the more effective they act as impediments to diffu-
sion. We compared obstacles of the same area but of different aspect ratios,
ranging from 20×20 to 400×1. As an example, to reach a D/D0 of 0.64 one
needs a vcryst of roughly 0.4 of 20×20 obstacles, but only 0.02 of 400×1
obstacles. At the same time, an old analytical relation between the diffusion
coefficient, vcryst and an anisometry parameter x due to Fricke [86] could be

Permeation of Polymers 73
Semicrystalline and Filled Polymers

0.5

0.4
100 x 100 20 x 20
n = 0.62 n = 0.94
0.3
ln(D0/D)

0.2
300 x 300
n = 0.29
0.1

0.0
-1.0 -0.8 -0.6 -0.4 -0.2 0.0
ln vamorph

FIGURE 17 Dependence of geometric impedance upon amorphous volume


fraction for square obstacles of different sizes. Also shown are the
exponents n (Eqn. 45). From ref. 84.

shown to hold after being modified to account for the percolation of crystal-
lites.

D − D∞ x ( 1 − v cryst ⁄ v ∞ )
= , (Eqn. 47)
D0 − D∞ x + v cryst ⁄ v ∞
where v∞ is the highest crystalline volume fraction that can be reached before
percolation starts, and D∞ is the diffusion coefficient at this crystalline volume
fraction. The anisometry parameter x found by fitting the data to the form of
(Eqn. 47) could be shown to correlate with the obstacle’s aspect ratio. A plot
of our data and a fit to Fricke’s law are shown in Figure 18.
• If largely anisometric crystallites form oriented domains (cf. Figure 16), then
in these domains diffusion parallel to the crystallites is preferred over diffusion
perpendicular to the crystallites. This fact has been frequently observed
experimentally, for instance in block copolymers which assume a layered

74 Florian Müller-Plathe
Semicrystalline and Filled Polymers

1.0

0.8
(D - D∞) / (D0 - D∞)

0.6

0.4

0.2

0.0
0.0 0.2 0.4 0.6 0.8 1.0
vcryst / v∞

FIGURE 18 Dependence of the penetrant diffusion coefficient on the


crystalline volume fraction and the obstacles’ aspect ratio. All
obstacles have the same volume of 400 δ2, but different shapes.
Along the arrow: 20×20, 8×50, 2×200, 1×400. Together with the
raw data (■ and error bars) least-square fits to Fricke’s law (Eqn.
47) are shown. From the fits we obtain the anisometry
parameters x: 9.8, 0.79, 0.17, 0.14. From ref. 84.

morphology and where the two polymers have different permeabilities (e.g.
refs. 81, 87, 88) or in polymers with platelet-shaped fillers [10].

We can summarise that our simple two-phase model can explain some of the
observed features arising in heterogenous media such as semicrystalline or
filled polymers. Many effects can only be explained by allowing for variation in
sizes, shapes, arrangements and orientations of the crystallites, which is quite
practical in a computational Monte Carlo model, but not in analytical models.
What has not yet been explained by any two-phase model is that the depen-
dence of the permeability on the crystalline volume fraction is sometimes even
stronger than observed here, as has been rightly pointed out in ref. 90. We

Permeation of Polymers 75
Semicrystalline and Filled Polymers

expect, however, that the two-phase model can be taken further by allowing for
partial percolation of crystallites, which was explicitly excluded from our simula-
tion, and that more drastic effects of crystallinity on diffusion coefficients will then
become prominent. Another possible extension has already been mentioned:
interface effects and modifications of the transport properties of the amorphous
phase close to the crystallites could be added which would, in effect, abandon a
two-phase picture.

76 Florian Müller-Plathe
8 Ions in Polymers

This chapter describes our own recent work on ionically conducting polymers. In
contrast to our work on gas diffusion, this material describes, in parts, work still in
progress and has not been published previously. This chapter, therefore, con-
tains some introductory remarks as well as details on the computational models
used. These have not been given for the gas diffusion work, since they can be
found in the original literature.

8.1 Ion-conducting polymers: background


The polymeric systems that show transport of ions come in many different
classes. Polymer gels which contain a high fraction of water or another polar sol-
vent (“hydrogels”) conduct ions essentially through the liquid phase. Such gels
have been known for many decades. One example of their technical application
are ion-exchange resins, which can also be cast into thin films (this was already
found out in the 1940’s). Such membranes can be made relatively specific to
selected cations or anions, and they form, for example, the basic ingredient of
electrodialysis, which is the most energy-efficient process for water desalination.
By doping the polymer with an ion-specific complexant the selectivity of wet poly-
mer membranes to certain ionic species can be increased to such an extent that
the membranes can be used as separating device in analytical chemistry.

77
Ions in Polymers

The most recent addition to the family of ion-conducting polymers are the dry
conductors. These are polymers capable of dissolving salts without the addition
of any solvent. Their existence has only been discovered some 20 years ago.
(For reviews on the field, see e.g. refs. 91-93.) Suitable salts are made up of
mono- or divalent cations (most commonly Li+ and Na+) and anions that are soft
bases (e.g. ClO4-, CF3CO2-, I-, SCN-). Important criteria for a salt to dissolve in a
polymer are its lattice energy, which must not be too large, and the requirement
that a hard cation is combined with a soft anion.

From their discovery these electrolytes were rapidly developed into components
of consumer products. The main commercial driving force behind the develop-
ment is the design of new batteries, rechargeable as well as non-rechargeable.
These are mainly based on Li anodes and cathodes made of some inorganic
crystal that can intercalate Li+ ions, such as TiS2 or VOx, separated by a Li+-con-
ducting electrolyte. Polymers have become the choice as electrolytes for these
batteries for a variety of technical reasons, an important one being that they can
be cast into very thin films with a high surface-to-thickness ratio. This geometry
offers a low resistance which more than makes up for the intrinsic conductivity of
the polymer which might be lower than that of other ionic conductors.

The prototype of an ion-dissolving polymer is polyethylene oxide (PEO, (-CH2-


CH2-O-)n, [94]). This is not surprising since its linear oligomers, the polyethylene
glycols (PEGs), can complex cations and its cyclic oligomers, the crown ethers,
even more so. In particular, the cyclic hexamer, 18-crown-6, is used frequently
because of its metal-binding capabilities. Cation-binding by 18-crown-6 and sim-
ilar multi-dentate ligands has also been studied by molecular simulation [95-97].
Following the discovery of salt-complexing by PEO, a number of other polymers
were found to have the same property, among them polypropylene oxide (PPO)
and polyethylene imine (PEI) and several modifications (e.g. copolymers,
blends) of PEO. Still, PEO seems to be the most popular of these materials.

After the discovery that certain salts dissolve in certain polymers it did not take
long to find ionic conductivity in these systems. Measured conductivities slightly
above room temperature can reach 10-4 S/cm. (For comparison, a 0.1 m aque-
ous solution of NaCl has a conductivity of around 10-2 S/cm [98].)

The melting point of crystalline PEO is around +65˚C, the glass transition tem-
perature of amorphous PEO around -60˚C [94]. Therefore, PEO has a high

78 Florian Müller-Plathe
Ions in Polymers

degree of crystallinity, roughly 60% at room temperature. The crystallites, which


can also cluster into spherulites, are made up of PEO helices. Helices are stable
because PEO prefers a gauche conformation around the C-C bonds, the C-O
bonds being preferentially trans. Partial crystallinity is also found for many poly-
mer-salt complexes. Initially, it had been presumed that the crystalline domains
were responsible for ion transport, with ion motion inside the helices suggested
as the primary mechanism. It was, however, established by NMR spectroscopy
(spin-lattice relaxation) [99] that diffusion occurs mainly through the amorphous
phase, which is not surprising given the analogy with gas diffusion. Since the
amorphous phase is the predominant contributor to the conductivity, decreasing
the degree of crystallinity of PEO-salt systems is one design goal and many suc-
cessful attempts have been reported. These include [100] cross-linking, block
copolymers with PPO with short block lengths, or the introduction of small links
(urethane, carbonate, siloxane, phosphate and particularly oxymethylene) that
act as helix breakers. An alternative is the use of PPO which normally is atactic
and therefore amorphous, but which is not quite as good a host system for salts
in the first place.

The conductivity is not only sensitive to the degree of crystallinity, temperature


and the like but also very much to the salt concentration. The concentration
dependence usually shows a maximum due to competing effects (for an intro-
duction, see ref. 101). As the concentration increases, the conductivity
increases due to the availability of more charge carriers. At the same time, how-
ever, formation of ion pairs and higher clusters sets in [102], which leads to a
decrease in the conductivity, as in every ionic solution at higher concentrations.
In addition, there is an effect typical of polymeric hosts which is not present in
low-molecular-weight solvents: At higher salt concentration, the polymer
becomes stiffer, as can for instance be seen from the rise of the glass-transition
temperature [103], due to a effective cross-linking by the salt. Lower chain
mobility accounts for an additional lowering of the conductivity. The occurrence
of maxima in the conductivity (often around 8-16 polymer oxygens per cation)
has, for a long time, put a ceiling on the conductivities achievable with polymer
electrolytes.

Very recently, Angell et al. [104] have approached the problem of obtaining high
conductivities from an entirely different angle. They realised that, as the salt con-
centration is increased, the glass transition temperature would eventually pass
through a maximum and, hence, they started from a multi-component eutectic

Permeation of Polymers 79
Ions in Polymers

mixture of salts with high ionic conductivity and reinforced this with some 10% of
polymer to make the system rubbery. Such a system bears more similarity to a
hydrogel than to the traditional dry ionic conductors, in the sense that it is a poly-
mer swollen by a solution, the difference being that the “solvent” is a molten salt
rather than a molecular liquid. Rather high conductivities could be realised for
these systems.

8.2 Which phenomena can be expected in polymer electrolytes?

Let us now turn to the difference in the behaviour of small neutral molecules and
small ions inside rubbery polymers. As we have seen, neutral molecules are
absorbed into pre-existing small cavities in the polymer, and they diffuse by hop-
ping in a network of cavities. The cavity, the free volume big enough to hold the
penetrant, is the basic interaction site. It spends most of its time where it can do
this most comfortably, and it takes advantage of channel formation in the poly-
mer to accomplish diffusion. Most of its behaviour - with a few notable excep-
tions (Section 4.3 on page 37) - is governed by the availability of free volume.

Polar penetrants inside polar polymers, and, as an extreme case, ionic species
have electrostatic and polarisation interactions in addition to the excluded-vol-
ume and van-der-Waals interactions also found in nonpolar systems. One can
expect a wealth of new phenomena as a direct consequence of these strong and
very specific interactions. Candidates for such effects are among others:
• Formation of specific sorption sites. Because of the strong interaction (com-
plexation), an ion might be able to shape the interaction site to its require-
ments. This is in contrast to a small gas molecule which just takes advantage
of the host system as it finds it.
• A diffusion mechanism different from simple hopping since intermediate com-
plexation might lower transition states.
• Pair formation. Ions might attract each other to form pairs already at relatively
low concentrations. This will complicate the diffusion process and also influ-
ence the conductivity since a diffusing ion pair does not contribute to the
charge flux.
• Long-range correlation. The interaction between ions of like or unlike signed
charge has a longer range than nonpolar Lennard-Jones interactions. Hence,
correlated motion becomes more likely at low concentrations.

80 Florian Müller-Plathe
Ions in Polymers

• Change of the polymer properties. It is conceivable that ions complexed by


several polymer chains can mediate an additional attraction between the
chains. An example is the compatibilising effect of LiClO4 on otherwise
immiscible blends of PEO and polyvinyl pyridine [105]. Therefore, ions in
large concentrations might change the transport and other properties of the
host polymer to a much greater extent than gas molecules.

Some of these phenomena lend themselves to study at the atomistic level by


suitable spectroscopic methods or by simulation.

8.3 Simulations
We have undertaken a series of simulations of lithium iodide LiI in polyethylene
oxide. We must emphasise that these simulations are still exploratory. At
present, we are not yet aiming at experimental accuracy, but our purposes are to
determine, firstly, if and how far molecular simulation is useful in the study of
ionic conduction in polymers and, secondly, if we can reproduce some of the
more qualitative features encountered in these systems.

Our model of the system is therefore only a rough approximation. The forcefield
parameters used are off-the-shelf without special tuning to the problem of ions in
polymers. An exception is the treatment of the gauche-effect in PEO, which has
to be reproduced correctly, since otherwise essential properties of the polymer
would change dramatically. How that effect was modelled is described in Appen-
dix 1 (see also ref. 106) together with the final PEO forcefield. For Li+ and I- we
use nominal charges (+1 and -1), the ε values of the isoelectronic rare gas
atoms He and Xe (0.084807 and 1.8308 kJ/mol), the σ were adjusted to produce
a minimum in the Li+-I- potential curve at 0.296 nm which falls between the sum
of the ionic radii (0.288 nm) and the distance in the LiI crystal (0.3025 nm). This
gave σ values of 0.208 and 0.572 nm. Later we found parameters for aqueous
solutions of LiI [111], which are similar (ε: 0.149 and 0.408 kJ/mol, σ: 0.237 and
0.540 nm). These have probably been optimised with more care. However, we
stuck to our original parameters to facilitate comparison with calculations that
were already done. Nonbonded interactions are truncated at 1.0 nm and a
shifted potential [15, 51] is used.

In order to study motions of the ions inside the polymer we have carried out
some initial test simulations with this force field. We found that with the full force

Permeation of Polymers 81
Ions in Polymers

field no ion diffusion takes place on a nanosecond time scale. The ions are
bound too strongly to the polymer. We have therefore reduced all electrostatic
interactions by using an effective dielectric constant ε of 3 (see Box 1 on page
20). Incidentally, this value has also been used in the early RIS calculations on
PEO [112]. This change produced the desired effect to make the ions more
mobile. However, a few caveats are in order, particularly in view of what has
been said about the use of reduced force fields (Section 5.5). We can definitely
not expect quantitative agreement between calculated and experimental trans-
port properties. Also, the density in constant-NPT calculations will probably be
too low. However, the change in the force field is not too large, so that we expect
to obtain insight into, e.g., the diffusion mechanism at a qualitative level.

We ran a series of NPT calculations at two temperatures (400 and 450 K) of LiI in
a PEO system of 400 monomers with varying concentrations of LiI: 100, 50, 25,
16 ether oxygens per LiI (this measure of ion concentration (EO)xLiI is rather the
inverse of the concentration, but is the quantity most used in the literature).

8.4 Interaction sites


Relatively little is known about the immediate environment of the dissolved ions.
Limited information is available on crystalline polymer salt complexes from ori-
ented fiber X-ray diffraction [92]. In a polycrystalline sample of (EO)3NaI the
sodium has been found to be surrounded on average by 3 oxygens and 2
iodides. EXAFS studies on amorphous polymer-salt systems appear to be less
conclusive [92].

We have looked at the immediate environments of positive and negative ions in


PEO. To this end we have extracted the positions and types of all atoms within a
sphere of a certain radius around the ion of interest. For Li+ ions, we set the
radius to 0.4 nm, for I- to 0.45 or 0.5 nm, because of its larger atomic radius.

The Li+ ions interact most strongly with the negative partial charge on the poly-
mer oxygens and with the I- ions. Hence, it is no surprise that those species are
predominantly found nearest to the lithium ions. For instance, an instantaneous
snapshot of Li+ environments in a (EO)25LiI system at 400 K reveals that of 16
Li+, 10 are complexed by four or more oxygens, and 13 by one or two iodide
ions. To get some idea of the evolution of the Li+ environment, we have recorded
the changes over a time span of 640 ps. We found moderate changes in the

82 Florian Müller-Plathe
Ions in Polymers

iodide shell: half of the Li+ stayed with the same iodide ions, the other gained or
lost one iodide. For the contact with the polymer oxygens we find that there are
many instances of acquiring or loosing one, two or three ligands, and much
fewer where changes by four or more oxygens take place. This is easily under-
stood, since a small change of oxygen ligands means only a local reconfigura-
tion of the binding pocket, whereas a big change in number involves the
migration of the ion into a different location.

At is also interesting to see the relation between the oxygens that coordinate to
the lithium ions. In most cases, the oxygens belong to one or two segments of
PEO (as a segment we define any contiguous stretch along the polymer chain,
this is not to be confused with the usage of the term segment for monomer): of
the 16 lithiums, 11 are complexed by oxygens from one segment, 3 are com-
plexed by oxygens from two segments, none have contacts to oxygens of three
or more segments. That means that PEO prefers to act as a multidentate ligand
for Li+, a property carried over from its linear and cyclic oligomers. In the evolu-
tion of the systems, we find the trend towards fewer and longer segments
around a given Li+. For example, a lithium that may have been complexed by
two segments of three oxygens initially, may end up being complexed by one
segment of 5 oxygens and an iodide.

Typical lithium binding situations are shown in Figures 19 and 20. In Figure 19,
the lithium is held in an almost planar PEO loop coordinated to 5 oxygens. A
sixth coordination site is saturated by an iodide. The motif of a quasi-planar PEO
arrangement is also found in Figure 20. This time 4 oxygens are coordinated to
the lithium ion and it is located out of the plane of the oxygens. The saturation of
the remaining positions is here done by two oxygens of a different segment. Inci-
dentally, both figures show the same Li+ ion that has migrated to a different site
between the two snapshots which are 600 ps apart.

While Li+ sorption (binding) is characterised by complexation by the ether oxy-


gens, the I- sites look much more like the sorption sites for gas molecules
described before. The I- interact mostly with the hydrogen atoms of PEO and
with the Li+ ions. The hydrogens carry only a small positive partial charge (see
Table A.1 in Appendix 1) so that electrostatic interactions play only a minor role
in holding the anion in place. The C atoms of the polymer have a higher partial
charge than the hydrogens, but are shielded by the hydrogens so the iodides
cannot get close. In essence, the iodides are therefore occupying cavities in the

Permeation of Polymers 83
Ions in Polymers

Y X Z X

Y Y

Z X Z X

FIGURE 19 Local environment of a Li+ ion in PEO, different views. The Li+ is
complexed by the 5 ether oxygens of a single chain segment of
polyethylene oxide. The 6th coordination site is taken up by an I-.
The system’s temperature is 400 K, its stoichiometry (EO)25LiI.
Legend: Li+ (black), I- (light grey), C (white), O (dark grey),
hydrogens are omitted for clarity.
polymer where they are trapped because of their size rather than any strong and
specific interaction with the matrix. Consequently, the iodides are not able to
wrap the polymer around themselves, as the lithium ions do. This manifests itself

84 Florian Müller-Plathe
Ions in Polymers

Y X Z X

Y Y

Z X Z X

FIGURE 20 Local environment of a Li+ ion in PEO, different views. The Li+ is
complexed by 4 ether oxygens of one chain segment of
polyethylene oxide and by 2 oxygen of another segment. The
system’s temperature is 400 K, its stoichiometry (EO)25LiI.
Legend: Li+ (black), C (white), O (dark grey), hydrogens are
omitted for clarity

in the large number of polymer segments in contact with any one iodide. In a
0.45 nm radius around an iodide one finds anything between 5 and 15 hydro-
gens which belong to between 3 and 8 different polymer segments (this is for the
same configuration the Li+ examples were drawn from). In contrast, the I- do

Permeation of Polymers 85
Ions in Polymers

maintain strong interactions with the cations in the system. We do find evidence
for this in the fact that every I- has at least 1 Li+ close by, many have two. Over
the time of 640 ps some of the cations are exchanged, but there is always at
least one within 0.45 nm. This is different from the cation environment. The Li+
can live without an iodide near by, but not vice versa. The reason for this is that
the ether oxygens can effectively compete with I- for the lithium coordination
sites, especially if they come as multidentate ligands as we saw, but the polymer
hydrogens are no match for the Li+ in competing for the complexation of I-. As a
consequence of some Li+ being far away from any iodide, other Li+ are bound to
two or more iodides.

From the consideration of the totally different environments found for cations and
anions it appears that the driving force for solvation is the (partial) complexation
of Li+ by the polymer. This releases enough energy to offset the lattice energy of
LiI that has to be overcome in the dissolution process. The polymer is a much
better “solvent” for the cations than for the anions. The iodide has few favourable
interactions with the polymer and is merely being dragged along into the polymer
by the cation to maintain electroneutrality. Because of its poor interaction with
the polymer it tries to associate with any cation which is at hand. There also
exists infrared-spectroscopic evidence of ion association in polymer electrolytes
[92]: some changes in vibrational frequencies of polyatomic anions upon dissolu-
tion in the polymer can be explained in terms of formation of ion pairs or larger
aggregates. There is also evidence that anions are actively involved in transport-
ing charge through the polymer: different Lithium salts have different conductivi-
ties [113], and conduction has different activation energies. This effect has been
put down to the anion influencing the cation mobility and to the presence of
residual water which, however, could not be traced spectroscopically. Our simu-
lations hint at a different explanation: the anions can be mobile as well. The cat-
ions are much smaller than the anions and should be able to travel through the
polymer more easily (see the discussion about the effect of penetrant size on
transport properties in Section 5.5). However, the anions are bound to the poly-
mer much more weakly so their migration does not involve any activation pro-
cess other than the formation of a large enough channel. They need not unbind
from the polymer, go through a state without many strong interactions and rebind
at some other site, like the cations. Hence, it is conceivable that their mobility is
not all that small. If, on the other hand, the anions contribute to the current in the
polymer, a change of anion will obviously change the conductivity.

86 Florian Müller-Plathe
Ions in Polymers

0.8

0.6 (EO)100LiI
(EO)50LiI
(EO)25LiI
(EO)16LiI
0.4

0.2

0.0
0 1 2 3 4 5 6 7 8 9 10
cluster size (ions)

FIGURE 21 Distribution of ion clusters inside PEO at 400 K for different ion
concentrations. The distribution is normalised so that its integral
is unity. An ion is considered member of a cluster if its distance to
at least one other cluster member is less than 0.5 nm.

8.5 Ion association


Because the attraction between cation and anion is the most favourable individ-
ual interaction in the system, one can expect association to occur. Ion pairing
was already mentioned in the last section. In principle, also larger clusters of
ions could form. Understanding clustering is quite important for a variety of rea-
sons. The mobility of a cluster of ions can be radically different from the mobility
of isolated ions. Ion aggregation lowers the number of charge carriers: Forma-
tion of electroneutral clusters (pairs, quadruples, etc.) effectively annihilates the
conductivity of the ions involved, formation of charged clusters still retains some
conductivity, but at a reduced level. Finally, the attraction between ions can ulti-
mately lead to phase separation into salt crystals on one hand and a salt-poly-
mer phase on the other. (An example of a phase diagram is given in ref. 92.)

Permeation of Polymers 87
Ions in Polymers

0.5

Li+ - I-
0.4 Li+ - Li+
〈δRi δRj〉 / 〈δRi〉 〈δRj〉

I- - I-

0.3

0.2

0.1

0.0
0.0 0.5 1.0 1.5
〈|Ri-Rj|〉 / nm
FIGURE 22 Positional cross correlation between ions in (EO)16 LiI at 400 K.
Averaging of the cross correlations has been done for all pairs of
ions over time intervals of 40 ps, with δRi=Ri-〈Ri〉. These averages
are plotted against the average distance between the two ions
during the same time interval. Data points displayed are averages
over all calculated data points within intervals of 0.05 nm.

We have analysed the distribution of ion cluster sizes in the PEO-LiI systems.
We define a cluster as follows: An ion is considered to be member of a cluster if
its distance from at least one other cluster member is less than or equal to a cer-
tain cutoff distance. We have used a cutoff distance of 0.5 nm, thereby allowing
for very loose contacts within a cluster. For comparison, the Li-I distance in crys-
talline LiI is more like 0.3 nm.

The distributions are shown in Figure 21 for several stoichiometries. Sizes of the
statistical samples are restricted; there is only one polymer starting structure,
and at low salt concentrations there is not much averaging over salt configura-
tions either. The distributions should therefore not be taken as absolute num-
bers. Nonetheless, they are indicative of the trends. At the lowest salt
concentration (or highest concentration of oxyethylene units) (EO)100LiI the most

88 Florian Müller-Plathe
Ions in Polymers

common clusters are ion pairs with a large fraction of isolated ions being present
as well. At this concentration, there is only a tiny transient population of triples,
and no larger clusters at all. Going to higher salt concentrations, the number of
isolated ions and ion pairs drops. Still, even at the highest concentration, ion
pairs remain the most important cluster. However, larger cluster sizes continue
to become more important with increasing ion concentration. At the highest ion
concentration studied here, (EO)16LiI, clusters of size 3-7 are present in appre-
ciable numbers.

The trend towards larger cluster sizes with increasing salt concentration is, of
course, hardly surprising. On the other hand, clusters remain small, and on the
time scale accessible to simulation (between one and several nanoseconds in
these simulations) no phase separation appears to take place for any of the salt
concentrations studied so far.

A measure of the dynamic correlation between ions i and j is the positional cross
correlation 〈 ( R i − 〈 R i 〉) ( R j − 〈 R j 〉) 〉 ⁄ [ 〈 R i − 〈 R i 〉 〉 〈 R j − 〈 R j 〉 〉] . We have calcu-
lated this quantity for all ion pairs and plotted it against the average distance of
the ions (Figure 22). One observes that positional fluctuations of ions are corre-
lated up to distances larger than the potential cutoff (1 nm). The correlation for
ions of opposite charge is larger than for ions of like charge, since the latter can-
not approach each other as closely. The correlation among the I- seems to be
slightly larger than the correlation among the Li+. A possible explanation is the
complexation of the Li+ which leads to a better screening.

We conclude that there is evidence for both static and dynamic correlation
among the ions. Therefore, the simple concept of independent motion of ions
should not be applied to these ionic systems.

8.6 Ion mobility and conductivity

The motion of ions through a polymer follows the principles found for neutral
small penetrants (Chapter 4) only to some extent. The presence of strong and
specific interactions changes several aspects of the motion pattern.

From the foregoing sections one would expect major effects only for the cations
but not for the anions, since the anion-polymer interaction is weaker and less
specific than the cation-polymer interaction. This is indeed the case. Examining

Permeation of Polymers 89
Ions in Polymers

O96
-
O124
O1342
-
O1363
O1769

-
O1818
O2357
I-2809
I-2827
0.0 500.0 1000.0 1500.0
t / ps
FIGURE 23 Time evolution of the local environment of an individual Li+ ion
(atom index 2808) in a (EO)25LiI system at 400 K. If an oxygen or
iodine is within 0.4 nm of the Li+ at a given time this is indicated by
a filled square. Carbon and hydrogens are not shown since the
only ones within this distance are connected with one of the
oxygens. Oxygens belonging to contiguous polymer segments are
grouped together. Subscripts indicate atom sequence numbers.

the motion of individual I- ions we find a behaviour very similar to the hopping
motion of gas molecules. The main difference worth noting is the larger correla-
tion with the motion of other ions. While gas molecules become correlated only
at very high concentrations, ions show some correlation with one another. For I-
correlated motion mainly involves the Li+ ions close by. As described in Sections
8.4 and 8.5, the iodides usually associate with at least one lithium ion. Correla-
tion involves two patterns: the ions can travel together as a ion pair, or by the
motion of one ion, the other one is released and rearranges itself into a new
environment.

The motion of Li+ shows much more interesting patterns. An example is given in
Figure 23. Three polymer segments (A: O96-O124, B: O1342-O1363, C: O1769-
O1818) and two iodines (I2809 and I2827) are involved in forming the local envi-
ronment of this Li+ ion at various stages, apart from a brief intrusion by O2357.

90 Florian Müller-Plathe
Ions in Polymers

O68

O1370- O1426

I-2811
I-2821
I-2823
I-2825
0.0 200.0 400.0 600.0 800.0
t / ps
FIGURE 24 Time evolution of the local environment of an individual Li+ ion
(atom index 2804) in a (EO)25LiI system at 450 K. For details of
the representation, see legend to Figure 23.

Initially, the lithium is complexed by segment A and I2827. At ~100 ps it migrates


to a site mainly formed by segments B and C. Until ~400 ps there is some com-
petition between these two segments. It is also in this part where O2357
becomes a ligand for a short spell, probably stabilising a transition state. At ~400
ps the accompanying iodine is lost and segment C becomes the main ligand with
five of its oxygens connecting to the lithium. Between 500 and 950 ps, it is grad-
ually being replaced by segment B again. From ~700 ps, the other iodine (I2809)
is establishing itself as a permanent complexant.

An example of another motion pattern frequently observed for Li+ is shown in


Figure 24. Here, the complexation essentially involves only one polymer seg-
ment, but several different iodides. The polymer segment contains 9 monomers,
too many to fit around one lithium ion. Instead, different pieces of the segment
become involved with the lithium at different times. These pieces are always
contiguous. The resulting motion is sketched in Figure 25. One could describe
the rearrangement either as the polymer slithering around the lithium or the lith-
ium shifting along the polymer chain. From closer inspection of the event one

Permeation of Polymers 91
Ions in Polymers

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

7
8
8
9
9

FIGURE 25 Cartoon of a Li+ ion (black) moving along a polymer segment. The
polymer oxygens are shown as white circles with monomer
numbers, the rest of the polymer is not shown. On the left, the Li+
is complexed by monomers 2-6. On the right, it has been shifted by
one monomer unit and is now complexed by monomers 3-7. Both
the ion and the polymer are spatially rearranged. Hence, the
motion is neither a “rolling” of the ion along the polymer chain nor a
“slithering” of the polymer around the ion, but a combination of
both.

finds that both partners move substantially. Hence, events like this one should be
regarded as combinations of both motions. Also in this trajectory we observe a
possible transition state at ~480 ps where the Li+ is briefly complexed by O68 and
I-2825.

From the trajectories of the ions one can calculate their mean-square displace-
ments (MSD). We calculate the MSD for both ionic species as well as the collec-
tive MSD, which is needed to describe ionic conductivity (Eqn. 17 on page 15).
As an example of a calculated MSD we show the (EO)25LiI system at 450 K
(Figure 26). It contains a few features worth pointing out. Firstly, the collective
MSD is lower than the tracer MSD’s characteristic of independent motion. This is
computational evidence for the assumptions underlying the Nernst-Einstein
equation being invalid for this system. In every case we find the collective MSD
to fall significantly below the tracer MSD’s, the separation being up to one order

92 Florian Müller-Plathe
Ions in Polymers

1.0
D(coll.) = 8.4 10-7 cm2/s
0.8 D(Li++I-) = 16 10-7 cm2/s
D(Li+) = 13 10-7 cm2/s
D(I-) = 19 10-7 cm2/s
MSD / nm2

0.6

0.4

0.2

0.0
0.0 200.0400.0 600.0 800.0
t / ps
FIGURE 26 Mean-square displacements for LiI in polyethylene oxide,
(EO)25LiI, 450 K. From top to bottom: I-, Li++I-, Li+, collective.
Diffusion coefficients calculated from the slope of the appropriate
MSD are indicated.

of magnitude in some of the other calculations. Secondly, in this system the


MSD of I- is above that of Li+, meaning that of the two species, I- diffuses faster.
This feature is not universal. On most occasions, the cation is the faster diffusing
species, the anion only winning sometimes. However, the difference between
the two is mostly insignificant. This supports our earlier conclusion that in the
PEO-LiI system both ions are mobile and contribute to the electric current.
Thirdly, the collective MSD is a property of the system as a whole, not of individ-
ual ions. Consequently, for the tracer MSD’s one can average over all ions to
improve statistics. For the collective MSD there is no such possibility, since there
is only one collective MSD. The difference of the two MSD’s in sample size man-
ifests itself in the statistical noise. There are 16 ions of either sort in the system.
Hence, the statistical uncertainty of the collective MSD should be approximately
16 = 4 times greater than that of either tracer MSD. The MSD’s in Figure 26
are roughly in line with this expectation. The two facts that correlations between
ions are not negligible, so that one really has to use the collective MSD for the
calculation of conductivities, and that its statistics are poorer than for the tracer

Permeation of Polymers 93
Ions in Polymers

TABLE 6. Diffusion coefficients D (10-7 cm2/s) and specific ionic conductivities


λ (S/m) in LiI-polyethylene oxide systems. The PEO chain has 400
monomers, the number of LiI in the simulation is 4, 8, 16 and 25.

Composition T/K V/nm3 D(Li+) D(I-) D(Li++I-) D(coll.)


λ(Li+) λ(I-) λ(Li++I-) λ(coll.)

(EO)100LiI 400 27.6 5.4 1.9 3.7 0.11


0.018 0.0064 0.025 0.00074
450 30.0 28 18 22 3.1
0.077 0.050 0.12 0.017
(EO)50LiI 400 28.4 5.4 3.1 4.3 1.3
0.035 0.020 0.056 0.017
450 28.5 6.3 3.5 4.9 2.8
0.037 0.020 0.057 0.032
(EO)25LiI 400 29.8 3.9 1.4 2.8 2.0
0.049 0.017 0.070 0.050
450 32.2 13 19 16 8.4
0.13 0.20 0.33 0.17
(EO)16LiI 400 31.4 2.5 2.2 2.4 2.5
0.043 0.038 0.083 0.086
450 33.8 25 23 24 4.7
0.38 0.35 0.73 0.14

MSD’s, has the unfortunate consequence that simulations will have to be longer
still for the calculation of ionic conductivities than for the calculation of gas per-
meabilities.

The calculated diffusion coefficients and conductivities are listed in Table 6. One
should note, that for some of these values the underlying mean-square displace-
ments are not very well converged. In general, the MSD’s are too noisy to estab-
lish unambiguously if Einstein behaviour has been reached or not. However, in
view of the errors introduced by the reduced force field (Section 8.3) these trans-
port coefficient should still be adequate for qualitative comparisons among ion
mobilities in the systems studied. The data are shown in Figure 27. Conductivi-

94 Florian Müller-Plathe
Ions in Polymers

0.8

400 K, true
400 K, Nernst-Einstein
0.6 450 K, true
450 K, Nernst-Einstein
λ / S m-1

0.4

0.2

0.0
0.00 0.02 0.04 0.06 0.08
(LiI) / (EO)
FIGURE 27 Specific ionic conductivity of the LiI-polyethylene oxide system as
a function of salt concentration and temperature. Filled symbols
indicate conductivities calculated from collective motion (Eqn. 18),
open symbols indicate conductivities calculated assuming
independent ions (Nernst-Einstein equation Eqn. 21).

ties generally increase with temperature and salt concentration. The Nernst-Ein-
stein approximation produces conductivities which disagree with the
conductivities calculated accounting for ion-ion correlations. Mostly, the Nernst-
Einstein conductivities are too high. For high salt concentrations at 450 K this
even leads to a qualitative difference: The Nernst-Einstein conductivity contin-
ues to rise with salt concentration while the true conductivity goes through a
maximum, or at least levels off. This levelling-off is more in line with experimen-
tal observations. It also occurs in the range of salt concentrations where it is
found experimentally: Li+/EO=0.02-0.1 depending mainly on temperature and
counter ion, see, for example refs. 91, 101, 103, 105, 107, 108. The split
between the conductivities calculated assuming correlated and assuming inde-
pendent motion of the ions corresponds to the difference between measured
electrical conductivities and those inferred from measured radio-tracer diffusion
coefficients. Such measurements are available for tetraglyme-NaSCN and PEO-

Permeation of Polymers 95
Ions in Polymers

0.4

400 K, Li+
400 K, I-
0.3 450 K, Li+
450 K, I-
λ / S m-1

0.2

0.1

0.0
0.00 0.02 0.04 0.06 0.08
(LiI) / (EO)

FIGURE 28 Partial specific conductivities due to Li+ and I- ions in polyethylene


oxide, and their dependence on composition and temperature.
These are calculated from the tracer diffusion coefficients of both
species.

NaSCN [109, 110]. They yield factors between 1 and 3 (with large error bars)
between the two quantities.

In order to estimate the relative importance of the two ionic species we have also
calculated their contributions to the conductivity (Table 6 and Figure 28). This
was done in the Nernst-Einstein approximation. The results show, that while at
the lower temperature the cation might contribute more than the anion, at the
higher temperature both ions are equally mobile and contribute similarly to the
diffusion.

In both Figure 27 and Figure 28 one observes a glitch in the conductivity for the
composition (EO)50LiI (LiI/EO=0.02) at 450 K. The reason for this is probably
that at this composition the usual expansion of the simulation box as the temper-
ature rises from 400 to 450 K does not take place. This is depicted in Figure 29,
where we show the composition dependence of the volume or, rather, the den-

96 Florian Müller-Plathe
Ions in Polymers

PEO mass density / g cm-3 1.1

1.0

0.9 400 K
450 K

0.8
0.00 0.02 0.04 0.06 0.08
(LiI) / (EO)

FIGURE 29 Density of the PEO-LiI systems at various compositions. The


hypothetical density of polyethylene oxide is given, i.e. mPEO/V,
which ignores the mass of LiI. While this contribution is not small,
the “pure-PEO” density reflects better the expansion with higher
LiI content.

sity, in order to use a more familiar description. We do not show the true mass
density of the system (which would include LiI) but just the polymer part of it,
which is inversely proportional to the volume changes. The reason for the failure
to expand upon heating is not completely clear. We believe that the system is
trapped at a metastable state of higher density and that, in spite of a simulation
time of around 2 ns, some yet-to-be-resolved entanglement prevents it from
reaching its equilibrium density. Apart from that, the density decreases uniformly
and smoothly with the salt content.

It is difficult to compare our calculated results exactly with experiment, because


there are no conductivity measurements on LiI in amorphous polyethylene
oxide. Pure PEO tends to be semicrystalline, and amorphous materials contain
other polymer or additives as well as PEO. It is, however, clear that despite
these details, conductivities should be more of the order of 10-5 to 10-4 S/m,
rather than the 10-2 to 10-1 S/m calculated in the present work. We believe that

Permeation of Polymers 97
Ions in Polymers

this is a direct consequence of the reduced force field and it could be easily rem-
edied if this would not incur prohibitive computational costs.

98 Florian Müller-Plathe
9 Summary and Outlook

Have I given you a clue?

Baloo
“The Jungle Book”
(W. Disney, 1965)

This treatise has outlined the application of computer modelling at the atomistic
level to systems of small penetrants in amorphous polymers. One may ask what
has been learned after all the effort and in how far are the results can be gener-
alised. In this chapter we give a summary of the most important findings.

Gas diffusion in polymers. The transport of small molecules around room tem-
perature was shown to happen by hopping from cavity to cavity. The cavities
always exists in the polymer, if there are penetrants or not. They are connected
by a network of transient channels, which open and close due to normal fluctua-
tions in the polymer. The channels are randomly directed and do not necessarily
follow the polymer chains. The rate determining step appears to be the opening
of a channel. However, there remains an energetic barrier to be overcome, so it
helps if the hopping penetrants are translationally hot. There is some evidence,
that at high temperatures the hopping mechanism becomes less important and a
more fluid-like transport mechanism takes over. It is unclear, how far the hopping
mechanism holds as the penetrant size increases. Unfortunately, direct molecu-
lar dynamics simulation of the diffusion of larger penetrants, which could clarify
this point, will be unfeasible for some time to come. It would, however, be unrea-
sonable to assume a priori that the motion of large flexible penetrants, that
involves major rearrangement of the polymer, follows exactly the rules as the
small penetrants investigated to date.

99
Summary and Outlook

Another important qualitative result, surprising at the time but not in hindsight, is
the discovery of anomalous diffusion in polymers. As outlined in Section 4.4, it
has important implications both for simulation and for experiment.

Ion diffusion in polymers. Ions diffuse in a similar way as gas molecules, in


that they hop between certain, favourable dwelling sites. Anions, which are only
weakly bound in a polymer like polyethylene oxide, diffuse almost completely like
gas molecules. They are captured in cavities of free space, and they move as
channels between those cavities become available. For cations, however, there
is less similarity with gas diffusion. Because of the strong complexation by poly-
ethylene oxide they induce major changes in the polymer and they customise
the binding sites to their needs, at least to some extent. A frequently observed
pattern is the polymer binding the metal ion as a multidentate ligand, with the
number of consecutive ether oxygens involved in the complexation ranging
between 3 and 6. Diffusional hops involve major rearrangements in the polymer.
They are facilitated by intermediate complexation of the cation (often by polymer
or counterions not part of either binding site), which presumably lowers the tran-
sition barrier. As a slightly different mechanism we also observed cations moving
along oxygens in a polymer chain. These results were obtained for a polyethyl-
ene oxide host. One would expect that polymers with a significantly different
complexation behaviour towards cations and anions would produce a different
qualitative motion pattern.

There is a considerable degree of ion-ion correlation. This is clearly evidenced


by clustering of ions into long-lived ion pairs or larger aggregates and by the
long-range positional correlations between them. It also shows up in the conduc-
tivities calculated from correlated mean-square displacements and from individ-
ual mean-square displacements, assuming the Nernst-Einstein equation to be
valid. The two quantities are sometimes radically different, which supports the
notion of correlated motion.

Transport coefficients. The range where MD simulations can be applied in the


quantitative calculation of gas diffusion coefficients has been described in detail
in Section 5.4 on page 51, and the limitations of the method have been dis-
cussed there. It is, however, safe to state that within these limits the method has
proven useful. It yields diffusion coefficients with an accuracy comparable to
experiment, it can distinguish between two penetrants in the same polymer and
also between two different polymers hosting the same penetrant. It is, therefore,

100 Florian Müller-Plathe


Summary and Outlook

a potentially useful tool in the design of new polymers with specific barrier prop-
erties to certain classes of penetrants. Since, however, the accurate calculation
of a diffusion coefficient is still quite costly, in the near future the MD method -
even with some of the tricks to push back its limits (Section 5.5 on page 52) - will
not become the routine tool to be applied to every polymer permeation problem.

The calculation of reliable gas solubilities appears to be much harder than the
calculation of diffusion coefficients. This quantity seems to be more sensitive not
only to the forcefield but to the polymer starting structures as well. At present, it
is possible to reproduce trends among the solubilities of several penetrants.
Absolute solubilities seem to come out universally too high. We expect, how-
ever, that the quality of solubility calculations will increase substantially with
every advance in the generation of amorphous polymer starting structures.

Because the ionic conductivity is a collective phenomenon which cannot be


described in terms of the mobility of individual ions (at least not at concentra-
tions of practical interest) its determination from simulation is more expensive
than the calculation of diffusion coefficients, while it is conceptually straightfor-
ward. It is just that longer sampling is needed to come up with statistically mean-
ingful averages. However, the factor between the two required sampling lengths
is not too large, since the number of penetrants or ions in a simulation is nor-
mally not very large either, so that averaging is already poor for tracer diffusion
coefficients. We predict that in the near future one will see many more success-
ful MD calculations of ionic conductivities in polymers, since the technique will
directly benefit from increases in computer speed which are more or less immi-
nent, both through advances in processor performance and in parallel process-
ing [114].

The calculation of transport coefficients in amorphous polymers by atomistic


simulation can be combined with two-phase models to estimate the effects of
semicrystallinity.

The future. This is always hard to extrapolate. However, there are already a few
trends emerging.

Recently, a very elegant and powerful transition state technique has been devel-
oped by Gusev and Suter [115-119], which allows the calculation of both gas sol-
ubilities [115] and diffusion coefficients [116-119] from the atomistic structure of a

Permeation of Polymers 101


Summary and Outlook

polymer. The method works as follows. In a first step an amorphous polymer


structure is generated and minimised. The configurational energy of a would-be
penetrant molecule is then calculated on a dense cartesian grid of meshpoints
which encompasses the whole structure. In the later version of their method a
vibrational smearing of the positions of the host atoms is incorporated, which
assumes independent isotropic harmonic motion of the polymer atoms. The sol-
ubility then follows from the configurational integral (similar to Eqn. 34), which is
approximated as a sum over the meshpoints. For the diffusion coefficient a fur-
ther step is necessary. One finds all the energy minima among the meshpoints
and the crest surfaces that separate them. Transition state theory then gives an
expression for the probability of a penetrant crossing from one minimum to an
adjacent one. After this analysis one ends up with a network of minima with tran-
sition probabilities between them. One finally has to perform a Monte-Carlo sim-
ulation of a single particle in the network. The mean-square displacement of the
penetrant yields the diffusion coefficient. Work is in progress elsewhere to
extend the transition-state theory to include explicitly a limited number of poly-
mer degrees of freedom [120].

It is clear that this technique allows it to extend the calculation to timescales far
beyond those accessible to MD. (Gusev and Suter have almost reached millisec-
onds). At the same time, the treatment of many more different starting structures
is affordable. For gases in rubbery and glassy polymers, very good agreement
with experiment has been obtained. One of the assumptions implicit in the theory
is that the polymer structure is essentially rigid, with the atoms merely oscillating
around their equilibrium positions. This assumption is very much what resulted
from the molecular dynamics simulations: the opening of channels is due to ther-
mal fluctuations of the polymer. In cases, where both methods were applied to
the same system, they gave the same results within the respective error margins
[121].

With the advent of the much faster transition-state technique, is the MD method
already out-dated? We believe that this is not quite the case. While the transi-
tion-state method is bound to become the method of choice for routine calcula-
tions of gas transport coefficients in polymers, MD simulations make fewer
presuppositions and will therefore continue to play a role in cases where the
motion patterns are not already clear but still have to be elucidated.

102 Florian Müller-Plathe


Summary and Outlook

An example of future use of the MD techniques is already given in this work.


Due to their strong interaction with the polymer, ions have the power to deform
the polymer host substantially. Hence, it would be inappropriate to assume a
static polymer structure with thermal smearing of the atom positions. A similar
argument can be made for other cases of strong polymer-penetrant interactions
like hydrogen bonds and for large penetrants.

Another area where deformations of the polymer by the penetrant become dom-
inant, is that of polymer swelling. At higher penetrant concentrations, or espe-
cially if the bulk penetrant is a solvent for the polymer, it will cause swelling,
gelation and ultimately dissolution of the polymer. Particularly the range of inter-
mediate solvent uptake (say 5-50%) is not covered by any analytical theory or
simple model, whence there is considerable scope for molecular simulation.

Initial work has been undertaken in this laboratory [122] with the aim of general-
ising techniques of grand-canonical molecular dynamics to characterise the sol-
vent uptake by a polymer, given that the solvent outside the polymer has a
certain chemical potential.

Permeation of Polymers 103


Summary and Outlook

104 Florian Müller-Plathe


APPENDIX 1 Forcefield for
Polyethylene Oxide

While our work on gas diffusion was performed completely without the use of
Coulomb potentials (all our host polymers were polyolefines so that the differ-
ence in partial charge between C and H could safely be ignored, and the guest
molecules all had neither a charge nor a dipole moment), the interaction
between ions and a polar polymer cannot be modelled without them. We there-
fore need a proper set of atomic charges. A second complication arises in the
description of polyethylene oxide: The dihedral angles OCCO prefer a gauche
over a trans conformation, which is somewhat counterintuitive. This preference
gives rise to the formation of helices in crystalline PEO. Also in the amorphous
state, gauche is preferred over trans by some 1-2 kJ/mol as was shown by rota-
tional isomeric state considerations [112, 123-125]. This “gauche-effect” has to
be accounted for by the forcefield.

The forcefields for alkyl ethers available either did not reproduce the gauche
effect or they used united-atom models for the methylene groups. From our
experiences in gas permeation [68] we believe that isotropic united-atom force-
fields should be shunned in calculations of transport properties (see also Sec-
tions 5.1 and 5.2 on page 45ff). We therefore had to develop our own parameter
set. We proceeded by performing ab initio quantum-chemical calculations [106]
at the level of second-order many-body perturbation theory using a polarised
split-valence basis set (6-31G**, in jargon) on the smallest analogue of PEO,
which comprises the 1,2-ethylene diether motif, namely dimethoxy ethane

105
(DME, CH3O-CH2-CH2-OCH3). Experimental evidence suggests that, while the
preferred conformation of DME at 0 K in vacuum is all-trans [126], in any bulk
medium (crystal, fluid, solution) at finite temperature it adopts a gauche confor-
mation around the central C-C bond [127, 128]. Because the gauche effect is
present also in this molecule that has no possibility of forming helices, it can be
concluded that the gauche effect is a cause and not a consequence of helix for-
mation in PEO.

In agreement with all previous quantum-chemical calculations on this molecule


we found that, in vacuum, the all-trans conformation was more stable than the
trans-gauche-trans conformation by some 2.3 kJ/mol. (We have learned that a
more recent calculation [129], which used a much larger basis set including dif-
fuse and d polarisation functions, arrived at a trans-gauche separation of 0.9 kJ/
mol.) We concluded that the gauche effect is not intrinsic to DME, but must be
due to its environment. Having ruled out helicity as a cause, still a few environ-
mental effects remained as possible culprits, such as entropy differences, pack-
ing, favourable contacts with partial charges in other molecules and so forth. We
decided to try the effect of a dielectric continuum, since one of the properties that
most distinguishes the two conformers is the dipole moment. The trans con-
former has, by virtue of its symmetry (C2h), no dipole moment, while for the
trans-gauche-trans conformer (C2) we calculated a gas-phase dipole moment of
1.44 debye.

We included the effect of a polarisable dielectric with dielectric constant ε by


means of the simplest reaction field method. A dipole µ in a spherical cavity of
radius a polarises the continuum surrounding the cavity, thereby inducing an
electric field, the so called reaction field ERF
2 ( ε − 1)
E RF = µ (Eqn. A.1)
4πε 0 ( 2ε + 1 ) a 3

The reaction field, in turn, interacts with the dipole, leading to a stabilisation by
− ( E RF ⋅ µ ) . Since the dipolar molecule is also polarisable, the reaction field will
induce an increase in the molecule’s dipole moment, which then will lead to a
larger reaction field and so on. The method can easily be incorporated into a
quantum-chemical method if this contains two features: Calculation of the dipole
moment and treatment of a molecule in the presence of an external electric field.
One then starts with a calculation of the molecule’s gas-phase dipole moment,

106 Florian Müller-Plathe


40.0

30.0
U-Utrans / kJ mol-1

20.0

10.0

0.0

-10.0
0.0 60.0 120.0 180.0
dihedral angle (OCCO) / degrees

FIGURE A.1 Potential-energy (normalised to Utrans=0) of dimethoxy ethane


depending on the central dihedral. All other degrees of freedom
are optimised. The vacuum curve (ε=1) is marked by filled circles
(•), the solution curve (ε=80, a=0.4 nm) by filled squares (■).

calculates the reaction field (Eqn. A.1), calculates the dipole moment in the
presence of the field, and iterates this scheme to self consistency which usually
is reached after a few steps.

We have calculated the potential energy of DME for fixed values of the central
dihedral, all other degrees of freedom being optimised. The resulting for ε=1
(vacuum) and ε=80 and a=0.4 nm (aqueous solution) curves are displayed in
Figure A.1. It is clear that the potential energy gap between trans and gauche is
reduced to zero when the polarisable continuum is introduced. Although we still
have not reached complete agreement with experiment (the gauche conformer
should be below the trans conformer), we have convinced ourselves that the
dominant contribution to the gauche effect comes from the presence of a polar-
isable continuum. Effects not considered in the very simple reaction field model

Permeation of Polymers 107


(basis set error, non-spherical cavity, higher multipoles, explicit solvent interac-
tions, dynamic phenomena) might easily account for the remaining disagree-
ment. Combining the smaller vacuum trans-gauche separation of 0.9 kJ/mol of
the more recent calculations [129] with our reaction-field correction (-2.3 kJ/mol)
would yield a trans-gauche separation of -1.4 kJ/mol, which is very much in line
with experimental observations.

From the wave function we also calculated the partial atomic charges by an elec-
trostatic potential fitting procedure, the so-called Merz-Singh-Kollman method
(see ref. 130). With these charges, and bonded and nonbonded parameters
used before in our polymer work we constructed an all-atom forcefield for DME
which has the form given in Box 1 on page 20. This raw forcefield yielded trans
as the most stable conformation of the C-C dihedral. The potential for this dihe-
dral was then modified in the following way:
12 6
k3  σ O-O   σ O-O 
U ( τ) = [ 1 − cos ( 3τ − π ) ] + 4ε O-O −
2  r O-O   r O-O 
(Eqn. A.2)
k2 1 q 2O
+ [ 1 − cos ( 2τ − π ⁄ 2 ) ] +
2 4πεε 0 r O-O

The physical motivation for this functional form is the following. The terms on the
first line contain the standard description of a dihedral potential with three min-
ima (trans, gauche+ and gauche-). The threefold cosine term provides the trans-
gauche barrier and parts of the cis barrier, whereas the 1-4 Lennard-Jones inter-
action between the oxygens contributes most of the cis barrier and also causes
the energy difference between the trans and gauche minima. The additional two-
fold cosine potential with minima at ±90° and a maximum at the trans conforma-
tion (180°) adjusts the gauche-trans energy difference. The oxygen charges
were modified for this specific 1-4 interaction (the YASP program allows one to
modify every individual nonbonded interaction, if desired, without a great perfor-
mance penalty [51]) to account for the dependence of the gauche-trans equilib-
rium on the dielectric constant. The parameters k2, εO-O, σO-O, and qO were then
changed until our DME model, after energy minimisation, adopted a trans state
at ε=1, which was 2.3 kJ/mol more stable than gauche, and a gauche state at
ε=80, which was more stable than trans by 2 kJ/mol. This forcefield was then
used in our PEO simulations. It is listed in Table A.1.

108 Florian Müller-Plathe


TABLE A.1 Forcefield for polyethylene oxide. For parameter definitions, see
Box 1 on page 20.

atoms m/amu ε/kJ mol-1 σ/nm q/e


H 1.00782 0.301 0.232 0.0355 a)
C 12.0 0.444 0.321 0.103
O 15.9949 0.565 0.303 -0.348
(1-2 and 1-3 interactions excluded, 1-4 interactions unmodified except O-O,
see below)

bonds d0/nm kd
C-H 0.1096 rigid
C-C 0.1510 rigid
C-O 0.1416 rigid

bond angles φ0/degrees kφ/kJ mol-1rad-2


HCH 109.45 306.4
HCO 109.45 350.0
HCC 109.45 366.9
CCO 107.7 460.4
COC 109.45 334.8

dihedral angles τ0/degrees n kτ/kJ mol-1


HCOC(end groups) 180 3 11.76
COCC 180 3 11.76
OCCO 180 3 12.4
90 2 4.0
modified nonbonded parameters for O-O 1-4 interactions
εO-O=0.2 kJ mol-1, σO-O=0.32 nm, qO=-0.25 e

a) On the terminal methyl groups H charges were set to 0.0237 e in order to maintain
overall electroneutrality.

Permeation of Polymers 109


110 Florian Müller-Plathe
References

1. J. D. Ferry, Macromolecules 24, 5237 (1991).


2. H. Strathmann, SWISS CHEM 10(3), 8 (1988).
3. J. A. Howell (Ed.), The Membrane Alternative: Energy Implications for
Industry, Report No. 21 of The Watt Committee on Energy (Elsevier, Lon-
don, 1990).
4. S. von Wroblewski, Ann. Phys. (Leipzig) 8, 29 (1879).
5. H. A. Daynes, Proc. R. Soc. London Ser. A 97, 286 (1920).
6. V. T. Stannett, W. J. Koros, D. R. Paul, H. K. Lonsdale and R. W. Baker,
Adv. Polym. Sci. 32, 69 (1979).
7. M. L. Williams, R. F. Landel and J. D. Ferry, J. Am. Chem. Soc. 77, 3701
(1955).
8. J. Crank and G. S. Park (Eds.), Diffusion in Polymers (Academic, London,
1968).
9. J. Comyn (Ed.), Polymer Permeability (Elsevier, London, 1985).
10. W. J. Koros (Ed.), Barrier Polymers and Structures, ACS Symposium
Series 423 (American Chemical Society, Washington D.C., 1990).
11. W. R. Vieth, Diffusion in and through Polymers (Hanser, München, 1991).
12. M. H. Cohen and D. Turnbull, J. Chem. Phys. 31, 1164 (1959).
13. A. T. DiBenedetto, J. Polym. Sci. A 1, 3477 (1963).
14. R. J. Pace and A. Datyner, J. Polym. Sci. A: Polym. Phys. Ed. 17, 437
(1979).
15. M. P. Allen and D. J. Tildesley, Computer Simulation of Liquids (Oxford
University Publishers, Oxford, 1987).

111
References

16. G. Ciccotti, D. Frenkel and I. R. MacDonald (Eds.), Simulation of Liquids


and Solids (North-Holland, Amsterdam, 1987).
17. J. A. McCammon and S. C. Harvey, Dynamics of Proteins and Nucleic
Acids (Cambridge University Press, Cambridge, 1987).
18. W. F. van Gunsteren and P. K. Weiner (Eds.), Computer Simulation of Bio-
molecular Systems (ESCOM, Leiden, 1989).
19. C. R. A. Catlow, S. C. Parker and M. P. Allen (Eds.), Computer Simulation
of Fluids Polymers and Solids (Kluwer, Dordrecht, 1990).
20. R. J. Roe (Ed.), Computer Simulation of Polymers (Prentice Hall, Engle-
wood Cliffs,1991).
21. G. Attard, Trends in Polym. Sci. 1, 127 (1993).
22. H. Takeuchi and K. Okazaki, J. Chem. Phys. 92, 5643 (1990).
23. J. Sonnenburg, J. Gao and J. H. Weiner, Macromolecules 23, 4653
(1990).
24. F. Müller-Plathe, J. Chem. Phys. 94, 3192 (1991).
25. S. Pauly in: J. Brandrup and E. H. Immergut (Eds.), Polymer Handbook
3rd Edition (Wiley, New York, 1989).
26. D. A. McQuarrie, Statistical Mechanics (Harper and Row, New York,
1976).
27. J. W. Haus and K. W. Kehr, Phys. Rept. 150, 265 (1987).
28. S. Havlin and D. Ben-Avraham, Adv. Phys. 36, 695 (1987).
29. W. F. van Gunsteren and H. J. C. Berendsen, Angew. Chem. Int. Ed. Engl.
29, 992 (1990).
30. R. Car and M. Parrinello, Phys. Rev. Lett. 55, 2471 (1985).
31. J.-P. Ryckaert, G. Ciccotti and H. J. C. Berendsen, J. Comput. Phys. 23,
327 (1977).
32. F. Müller-Plathe and D. Brown, Comput. Phys. Commun. 64, 7 (1991).
33. D. J. Evans and G. P. Morriss, Statistical Mechanics of Nonequilibrium
Liquids (Academic, London, 1990).
34. H. J. C. Berendsen, J. P. M. Postma, W. F. van Gunsteren, A. DiNola and
J. R. Haak, J. Chem. Phys. 81, 3684 (1984).
35. A. P. Heiner, Ph.D. thesis (University of Groningen, Groningen, 1992).
36. F. Müller-Plathe, Generating Starting Structures for Polymer Simulations,
Report on a CECAM workshop held April 13-16, 1993, in Orsay, France
(unpublished).
37. D. N. Theodorou and U. W. Suter, Macromolecules 18, 1467 (1985).
38. D. Brown and J. H. R. Clarke, J. Chem. Phys. 84, 2858 (1986).
39. J. I. McKechnie, D. Brown and J. H. R. Clarke, Macromolecules 25, 1562
(1992).
40. R. Khare, M. E. Paulaitis and S. R. Lustig, Macromolecules (in press).

112 Florian Müller-Plathe


References

41. R. H. Boyd and P. V. K. Pant, Macromolecules 24, 6325 (1991).


42. J. J. de Pablo, M. Laso and U. W. Suter, J. Chem. Phys. 96, 6157 (1992).
43. R. M. Sok, H. J. C. Berendsen and W. F. van Gunsteren, J. Chem. Phys.
96, 4699 (1992).
44. W. Paul and K. Binder, Polymer Preprints (Polymer Chemistry Division of
the American Chemical Society) 33(1), 535 (1992).
45. W. Paul, K. Binder, K. Kremer and D. W. Heermann, Macromolecules 24,
6332 (1991).
46. J. Baschnagel, K. Binder, W. Paul, M. Laso, U. W. Suter, I. Batoulis, W.
Jilge and T. Bürger, J. Chem. Phys. 95, 6014 (1991).
47. F. Müller-Plathe, S. C. Rogers and W. F. van Gunsteren, J. Chem. Phys.
98, 9895 (1993).
48. B. Widom, J. Chem. Phys. 39, 2808 (1963).
49. J. P. Hansen and I. R. McDonald, Theory of Simple Liquids 2nd Ed. (Aca-
demic Press, London, 1986).
50. F. Müller-Plathe, Comput. Phys. Commun. 61, 285 (1990).
51. F. Müller-Plathe, Comput. Phys. Commun. (in press).
52. S. J. Watowich, E. S. Meyer, R. Hagstrom and R. Josephs, J. Comp.
Chem. 9, 650 (1988).
53. I. Carmesin and K. Kremer, Macromolecules 21, 2819 (1988).
54. L. Laaksonen, J. Mol. Graphics 10, 33 (1992).
55. W. Smith and M. J. Gillan, J. Phys. Condens. Matter 4, 3215 (1992).
56. F. Müller-Plathe, J. Chem. Phys. 96, 3200 (1992).
57. F. Müller-Plathe and W. F. van Gunsteren, Polymer Preprints (Polymer
Chemistry Division of the American Chemical Society) 33(1), 633 (1992).
58. P. V. K. Pant and R. H. Boyd, Macromolecules 26, 679 (1993).
59. D. Chandler, Introduction to Modern Statistical Mechanics (Oxford Uni-
versity Press, New York, 1987).
60. J.-E. Kluin, Z. Yu, S. Vleeshouwers, J. D. McGervey, A. M. Jamieson, R.
Simha and K. Sommer, Macromolecules 26, 1853 (1993).
61. F. Müller-Plathe, L. Laaksonen and W. F. van Gunsteren, J. Mol. Graphics
11, 118 (1993).
62. H. Takeuchi, J. Chem. Phys. 93, 2062 (1990).
63. J. Fischer and S. Lago, J. Chem. Phys. 78, 5750 (1983).
64. P. V. K. Pant and R. H. Boyd, Macromolecules 25, 494 (1992).
65. F. Müller-Plathe, S. C. Rogers and W. F. van Gunsteren, Chem. Phys.
Lett. 199, 237 (1992).
66. S. El Amrani and M. Kolb, J. Chem. Phys. 98, 1509 (1993).
67. J. L. Lundberg, E. J. Mooney and C. E. Rogers, J. Polym. Sci. A 7, 947
(1969).

Permeation of Polymers 113


References

68. F. Müller-Plathe, S. C. Rogers and W. F. van Gunsteren, Macromolecules


25, 6722 (1992).
69. S. Toxvaerd, J. Chem. Phys. 93, 4290 (1990).
70. H. Takeuchi, R.-J. Roe and J. E. Mark, J. Chem. Phys. 93, 9042 (1990).
71. S. Trohalaki, D. Rigby, A. Kloczkowski, J. E. Mark and R. J. Roe, Polymer
Preprints (Polymer Chemistry Division of the American Chemical Society)
30(2), 23 (1989).
72. W. F. van Gunsteren, in: R. Lavery, J.-L. Rivail and J. Smith (Eds.),
Advances in Biomolecular Simulations, AIP Conference Proceedings Vol.
239 (American Institute of Physics, New York, 1991).
73. R. Elber and M. Karplus, J. Am. Chem. Soc. 112, 9161 (1990).
74. F. Müller-Plathe, Macromolecules 24, 6475 (1991).
75. C. A. English and J. A. Venables, Proc. R. Soc. London Ser. A 340, 57
(1974).
76. T. Beutler, F. Müller-Plathe and W. F. van Gunsteren (unpublished).
77. A. A. Gusev, private communication.
78. D. Jeschke and H. A. Stuart, Z. Naturforsch. A 16, 37 (1961).
79. A. S. Michaels and R. B. Parker Jr., J. Polym. Sci. 41, 53 (1959).
80. A. S. Michaels and H. J. Bixler., J. Polym. Sci. 50, 413 (1961).
81. J. Csernica, R. F. Baddour and R. E. Cohen, Macromolecules 20, 2468
(1987).
82. A. Bunde, W. Dieterich and E. Roman, Solid State Ionics 18&19, 147
(1986).
83. R. M. Barrer, in: J. Crank and G. S. Park (Eds.), Diffusion in Polymers
(Academic, London, 1968).
84. F. Müller-Plathe, Chem. Phys. Lett. 177, 527 (1991).
85. R. D. Vigil and R. M. Ziff, J. Chem. Phys. 93, 8270 (1990).
86. H. Fricke, Phys. Rev. 24, 575 (1924).
87. J. Csernica, R. F. Baddour and R. E. Cohen, Macromolecules 22, 1493
(1989).
88. J. Csernica, R. F. Baddour and R. E. Cohen, Macromolecules 23, 1429
(1990).
89. T. C. Bissot, in: W. J. Koros (Ed.), Barrier Polymers and Structures, ACS
Symposium Series 423 (American Chemical Society, Washington D.C.,
1990).
90. P. Y. Furlan, Macromolecules 25, 6516 (1992).
91. M. B. Armand, in: J. R. MacCallum and C. A. Vincent (Eds.), Polymer
Electrolyte Reviews I (Elsevier, London, 1987).
92. M. A. Ratner and D. F. Shriver, Chem. Rev. 88, 109 (1988).
93. M. Gauthier, A. Bélanger, B. Kapfer, G. Vassort and M. B. Armand, in: J.

114 Florian Müller-Plathe


References

R. MacCallum and C. A. Vincent (Eds.), Polymer Electrolyte Reviews II


(Elsevier, London, 1989).
94. F. E. Bailey and J. V. Koleske, Poly (ethylene oxide) (Academic Press,
New York, 1976).
95. G. Wipff, P. Weiner and P. A. Kollman, J. Am. Chem. Soc. 104, 3249
(1982).
96. B. Owenson, R. D. MacElroy and A. Pohorille, J. Am. Chem. Soc. 110,
6992 (1988).
97. Y. Sun and P. A. Kollman, J. Chem. Phys. 97, 5108 (1992).
98. R. C. Weast and M. J. Astle (Eds.), CRC Handbook of Chemistry and
Physics 62nd Edition (CRC Press, Boca Raton, 1981).
99. C. Berthier, W. Gorecki, M. Minier, M. B. Armand, J. M. Chabagno and P.
Rigaud, Solid State Ionics 11, 91 (1983).
100. C. Booth, C. V. Nicholas and D. J. Wilson, in: J. R. MacCallum and C. A.
Vincent (Eds.), Polymer Electrolyte Reviews II (Elsevier, London, 1989).
101. M. Watanabe and N. Ogata, Brit. Polym. J. 20, 181 (1988).
102. J. R. MacCallum and C. A. Vincent, in: J. R. MacCallum and C. A. Vincent
(Eds.), Polymer Electrolyte Reviews I (Elsevier, London, 1987).
103. M. Watanabe, S. Nagano, K. Sanui and N. Ogata, Solid State Ionics
18&19, 338 (1986).
104. C. A. Angell, C. Liu and E. Sanchez, Nature 362, 137 (1993).
105. J. Li and I. M. Khan, Macromolecules 26, 4544 (1993).
106. F. Müller-Plathe and W. F. van Gunsteren, Macromolecules (submitted).
107. P. Ferloni, G. Chiodelli, A. Magistris and M. Sanesi, Solid State Ionics
18&19, 265 (1986).
108. J. P. Lemmon and M. M. Lerner, Macromolecules 25, 2907 (1992).
109. C. Bridges, A. V. Chadwick and M. R. Worboys, Brit. Polym. J. 20, 207
(1988).
110. A. A. Al-Mudaris and A. V. Chadwick, Brit. Polym. J. 20, 213 (1988).
111. Gy. Szasz, K. Heinzinger and W.O. Riede, Z. Naturforsch 36a, 1067
(1981).
112. J. E. Mark and P. J. Flory, J. Am. Chem. Soc. 87, 1415 (1965).
113. E. A. Rietman, M. L. Kaplan and R. J. Cava, Solid State Ionics 17, 67
(1985).
114. W. Scott, A. Gunzinger, B. Bäumle, P. Kohler, U. Müller, H.-R. vonder
Mühll, A. Eichenberger, N. Ironmonger, W. Guggenbühl, F. Müller-Plathe
and W. F. van Gunsteren, Comput. Phys. Commun. 75, 65 (1993).
115. A. A. Gusev and U. W. Suter, Phys. Rev. A 43, 6488 (1991).
116. A. A. Gusev and U. W. Suter, Polymer Preprints (Polymer Chemistry Divi-
sion of the American Chemical Society) 33(1), 631 (1992).

Permeation of Polymers 115


References

117. A. A. Gusev, S. Arizzi, U. W. Suter and D. J. Moll, J. Chem. Phys. 99,


2221 (1993).
118. A. A. Gusev and U. W. Suter, J. Chem. Phys. 99, 2228 (1993).
119. A. A. Gusev and U. W. Suter, Computer-Aided Mat. Des. (in print).
120. M. L. Greenfield and D. N. Theodorou, Macromolecules 26, 5461 (1993).
121. A. A. Gusev, F. Müller-Plathe, U. W. Suter and W. F. van Gunsteren, Adv.
Polym. Sci. (in preparation).
122. E. U. Wallenborn, Diplomarbeit (Laboratory of Physical Chemistry ETH,
Zürich, 1993).
123. J. E. Mark and P. J. Flory, J. Am. Chem. Soc. 88, 3702 (1966).
124. A. Abe and J. E. Mark, J. Am. Chem. Soc. 98, 6468 (1976).
125. A. Abe, K. Tasaki and J. E. Mark, Polym. J. 17, 883 (1985).
126. H. Yoshida, I. Kaneko, H. Matsuura, Y. Ogawa and M. Tasumi, Chem.
Phys. Lett. 196, 601 (1992).
127. K. Tasaki and A. Abe, Polym. J. 17, 641 (1985).
128. H. Matsuura, K. Fukuhara and H. Tamaoki, J. Mol. Struct. 156, 293
(1987).
129. G. Smith, R. Jaffe and D. Yoon, J. Phys. Chem. (in press).
130. M. J. Frisch, G. W. Trucks, M. Head-Gordon, P. M. W. Gill, M. W. Wong, J.
B. Foresman, B. G. Johnson, H. B. Schlegel, M. A. Robb, E. S. Replogle,
R. Gomperts, J. L. Andres, K. Raghavachari, J. S. Binkley, C. Gonzalez,
R. L. Martin, D. J. Fox, D. J. Defrees, J. Baker, J. J. P. Stewart, J. A.
Pople, Gaussian 92 Revision C (Gaussian Inc., Pittsburgh PA, 1992).

116 Florian Müller-Plathe

You might also like