THE JOURNAL OF CHEMICAL PHYSICS 124, 144909 共2006兲

Analysis of uncertainties in polymer viscoelastic properties obtained

from equilibrium computer simulations
Suchira Sen and Sanat K. Kumar
Isermann Department of Chemical and Biological Engineering, Rensselaer Polytechnic Institute, Troy,
New York 12180
Pawel Keblinskia兲
Department of Materials Science and Engineering, Rensselaer Polytechnic Institute, Troy, New York 12180
共Received 15 December 2005; accepted 21 February 2006; published online 13 April 2006兲

We critically evaluate the uncertainties in the stress autocorrelation function obtained from
equilibrium molecular dynamics simulation of model polymer melts. This quantity is central to
evaluating transport properties, e.g., the complex modulus and the viscosity. In contrast to the
intuitive expectation that simulations have to be run five to six orders of magnitude longer than the
chain relaxation time to reduce uncertainties to acceptable levels, our analysis shows that the
majority of the uncertainty is associated with rapidly oscillating bonded interactions. These
fluctuations occur on time scales which are approximately 104 times shorter than the relaxation time
of a chain of length 80. Consequently, the effects of these oscillations on the stress autocorrelation
function can be dramatically reduced by 共i兲 conducting long simulations 共typically 106 times longer
than the bond relaxation times or only 102 chain relaxation times兲 and 共ii兲 by performing running
averages with time windows whose time scales are much longer than these oscillations. Conducting
such long simulations also allows for the accurate determination of the melt viscosity and the
low-frequency complex modulus, but performing running averages do not impact these quantities
since they are time integrals of the stress autocorrelation function. © 2006 American Institute of
Physics. 关DOI: 10.1063/1.2186637兴

INTRODUCTION tions and Green-Kubo relationships is very appealing since it

The dynamics and viscoelastic properties of polymer
melts are now generally well understood from a theoretical
viewpoint.1–5 However, conducting computer simulations to
verify these predictions suffer from the difficulty that the
relaxation times of long entangled chains ␶R are expected to
increase as N3, where N is the chain length of the polymers.
A popular means to circumvent this difficulty is to use non-
equilibrium molecular dynamics 共MD兲 simulations, where a
uniform shear is imposed on the simulation cell. The analysis
of the stress versus shear rate relationship then allows one to
determine the shear dependent viscosity, from which the ex-
trapolated zero shear rate limit can be obtained.6 More re-
cently, several groups have used equilibrium MD simulations
to determine the stress autocorrelation function, 具␴共t*兲␴共0兲典
共see nomenclature for definition of t*兲. This quantity allows
for the evaluation of the viscosity, and the frequency depen-
dent storage and loss moduli, through the use of the Green-
Kubo relationships.7–9 For example, the viscosity ␩ is calcu-
lated as follows:
␩=
V
k BT
冕
0
⬁
具␴共t兲␴共0兲典dt,
where V is the system volume, T is the temperature, and kB is
the Boltzmann constant. The use of equilibrium MD simula-
a兲
Author to whom correspondence should be addressed. Electronic mail:
keblip@rpi.edu
allows for the evaluation of numerous quantities of interest
from a single simulation.
Recently, Vladkov and Barrat10 examined several equi-
librium and nonequilibrium methods of obtaining viscoelas-
tic properties of polymer melts. Based on their analysis,
these workers suggested that equilibrium simulations 共in
conjunction with the Green-Kubo formalism兲 give poor esti-
mates for viscoelastic properties for typical simulation
lengths 共i.e., 102␶R兲. These workers10 assumed that the time
dependent stress ␴共t兲 obeys Gaussian statistics. Then, the
uncertainty in the time averaged correlation function,
具␴共t*兲␴共0兲典, is estimated to be
␦共具␴共t*兲␴共0兲典兲 = 冑 2␶c
*
trun
具␴2典, 共2兲
where ␶c, the characteristic correlation time of the data, is
assumed to be ␶R, the chain relaxation time. For a typical
polymer melt comprised of chains of length 80, Fig. 1 shows
that 具␴共0兲␴共0兲典 ⬃ 102, and that the stress autocorrelation
function at t* = ␶R is 具␴共␶R兲␴共0兲典 ⬃ 10−2. Consequently, Vlad-
共1兲
kov and Barrat estimate that run times have to be at least
106␶R to obtain stress autocorrelation functions with manage-
able uncertainties. Given that the time step in a typical MD
simulation is t* = 0.01, and the chain relaxation time is
⬃1500, this implies a simulation in excess of 1011 MD time
steps. Since a very efficient MD simulation for this system
size can perform 107 time steps in a day, such simulation
times are out of range of any current calculation.

0021-9606/2006/124共14兲/144909/6/$23.00 124, 144909-1 © 2006 American Institute of Physics

Downloaded 03 May 2006 to 165.123.34.86. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp
144909-2 Sen, Kumar, and Keblinski
FIG. 1. 共a兲 Normalized stress autocorrelation function for N = 80 polymer
melt for short time scales where it exhibits oscillatory decay. 共b兲 The log-log
plot of the stress autocorrelation function for intermediate and long time
scales 共points兲 and the running average 共solid curve兲. 共c兲 The running aver-
age of the stress autocorrelation function on a log-normal scale with an
exponential fit to the terminal relaxation process.
In contrast to these pessimistic conclusions, several
authors7,8 have used much shorter equilibrium MD simula-
tions to calculate the stress autocorrelation function. The vis-
cosity obtained following Eq. 共1兲 compares well with the
values obtained by extrapolating nonequilibrium MD simu-
lations to the limit of zero strain rate. To resolve these dis-
crepancies and to critically assess the usefulness of equilib-
rium MD simulations in this context, here we present an
analysis of the uncertainties characterizing the stress autocor-
relation function and demonstrate that they are orders of
magnitude smaller than that estimated by Vladkov and Bar-
rat. We show that, surprisingly, the largest component of the
error is not due to the stress relaxation associated with the
longest relaxation time in the system, but rather due to high-
frequency bond vibrations characterized by very short relax-
ation time. Consequently, the stress autocorrelation function,
the viscosity, and the complex modulus are characterized by
much smaller uncertainties than those anticipated following
Eq. 共2兲 with ␶c = ␶R. Our findings are consistent with the good
agreement of our ␩ values with those obtained from nonequi-
librium MD simulations. Perhaps even more importantly, we
Downloaded 03 May 2006 to 165.123.34.86. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp
J. Chem. Phys. 124, 144909 共2006兲
conclude that equilibrium MD simulations with run times as
short as 100␶R can provide accurate estimates of the linear
rheological properties of polymer melts and solutions.
SIMULATION MODEL AND METHODS
Our simulations have been described in detail in a prior
publication.8 In brief, we employ the Kremer-Grest model of
a polymer chain. The interactions between nonbonded mono-
mers are described by a shifted, purely repulsive Lennard-
Jones 共LJ兲 potential: U共r兲 = 4␧关共␴ / r兲12 − 共␴ / r兲6兴 + ␧ for r
⬍ 21/6␴, and U共r兲 = 0 for r ⬎ 21/6␴. In addition to this LJ
potential, adjacent bonded monomers interact via a stiff fi-
nitely extensible nonlinear elastic 共FENE兲 potential: VFENE
= −k共R20 / 2兲ln共1 − 共r / R0兲2兲, with k = 30kBT / R20 and R0 = 1.5␴.
The combination of these two potentials constrians the dis-
tance between bonded monomers to about 1␴. In the simu-
lations reported here we use a total of 2400 monomers em-
bedded in a periodic simulation box which is 14.1␴ on a
side: this corresponds to a reduced segment density of ␳*
= 0.85. All quantities are given in reduced units defined at the
end of the paper. We performed constant volume simulations
of monodisperse polymer melts of varying chain lengths.
However, for the purpose of the uncertainty analysis we fo-
cus on melts composed of N = 80 chains. After structure
preparation and equilibration at a reduced temperature, T*
= 1, we switched to the NVE ensemble and collected data
over 2 ⫻ 108 MD steps. The NVE MD simulations use a fifth
order Gear algorithm with a time step ␦t* = 0.001 ensuring
energy conservation to within 0.5% over the entire simula-
tion run. The total simulation time was 200 000, while the
chain relaxation time was ⬃1500.
RESULTS AND ANALYSIS OF UNCERTAINTY
The stress autocorrelation function plotted in Fig. 1共a兲
displays oscillatory behavior at short times, t* ⬍ 1, with an
oscillatory period of about t* = 0.1. This is attributed to the
rapid vibrations of the stiff bonds connecting the bonded
monomers. As we will demonstrate, these vibrations are the
largest source of uncertainty in the calculated transport prop-
erties. In Fig. 1共b兲 we show the stress autocorrelation func-
tion for intermediate and long times. Up to t* = 10 the scatter
in the data is small. For t* ⬎ 10 the scatter of the data be-
comes significant and for t* ⬎ 100 the data scatter is one
order of magnitude or more. Such a large scatter appears to
suggest that the stress autocorrelation function cannot be cal-
culated with reasonable accuracy beyond t* = 100. If we use
Eq. 共2兲 with ␶c = ␶R ⬃ 1500 and trun ⬃ 200 000, we predict an
error of ␦共具␴共t*兲␴共0兲典兲 ⬃ 10, independent of t*. This result
can be visually seen to be incorrect for all t* values, since the
largest uncertainties appear to only be of O共10−1兲. Thus, on
this basis alone, we conclude that the errors obtained are at
least two orders of magnitude smaller than that expected by
Eq. 共2兲 with ␶c = ␶R. Alternatively, this result suggests that
the correlation time is actually ␶c ⬃ 10−4␶R ⬃ 0.1, i.e., that
the time constant of the process which dominates the uncer-
tainty is small, and comparable to the bond oscillation time
关Fig. 1共a兲兴.
144909-3 Analysis of uncertainties in polymer viscoelastic properties J. Chem. Phys. 124, 144909 共2006兲

FIG. 2. Standard deviation of the stress autocorrelation values ␴共␶R兲␴共0兲, as

a function of the separation between the data points used in calculations.

With the goal of reducing these large uncertainties at

long times, we performed running averages of the stress au-
tocorrelation function. For each t* we averaged from 0.9t* to
1.1t*. This averaging procedure preserves short time oscilla-
tory features of the stress autocorrelation function essentially
unchanged, but, as demonstrated in Fig. 1共b兲, it significantly
reduces the noise at large times, apparently allowing us to
evaluate the stress autocorrelation function up to t* ⬃1000,
where the value of the stress autocorrelation function is only
O共10−2兲. This observation is further supported by Fig. 1共c兲
where the relaxation process is fitted to an exponential, re-
sulting in a ␶R ⬇ 1500.
Analysis of uncertainty
To lend credence to the suggestion that the largest con-
tribution to the uncertainty is the bond fluctuation we per-
formed the following analysis. First, we calculated
具␴共t* + ␶R兲␴共t*兲典 with 130 time origins, t*, each separated by
␶R. 共This corresponds to the whole NVE simulation run.兲 The
standard deviation of ␴共t* + ␶R兲␴共t*兲 relative to its average,
具␴共t* + ␶R兲␴共t*兲典, is very large, ⬃40 共Fig. 2兲, and comparable
to 具␴2典 ⬃ 70. This result is consistent with Eq. 共2兲 which
yields an uncertainty ␦共具␴共␶R兲␴共0兲典兲 = 冑共2 / 130兲具␴2典. Next,
we calculated 130 stress correlation values, where each cor-
relation value is itself a block average,
i=k
1 shown in Figs. 1共b兲 and 1共c兲 have much smaller uncertain-
␴共␶R兲␴共0兲kn = 兺 ␴共␶R关n + i/k兴兲␴共␶R关n − 1 + i/k兴兲,
k i=1
where n labels the 130 time windows and k is the number of
data points used in calculating the block average stress auto-
correlation function in each window. The value of k deter-
mines the separation 共=␶R / k兲 between data points used in
these calculations. We selected k = 1, 15, 150, 1500, 15 000,
and 150 000 for our analysis, with k = 150 000 corresponding
to a time separation of t* = 0.01, which is comparable to the
size of a typical MD time step. The standard deviation of
these 130 ␴共␶R兲␴共0兲kn values as a function of the data sepa-
ration 共␶R / k兲 is plotted in Fig. 2. The standard deviation de-
creases monotonically with an increasing number of data
points used in the calculation of ␴共␶R兲␴共0兲kn. Since the stan-
dard deviation is proportional to the inverse of the square
root of number of points used in the calculation of
␴共␶R兲␴共0兲kn for data point separations t* ⬎ ⬃ 0.1, we suggest
that the uncertainties in the ␴共␶R + t*兲␴共t*兲 values are statis-
FIG. 3. “Error” defined as the difference between the stress autocorrelation
function and its running average as a function of time for two time ranges.
tically uncorrelated in this regime of t*. In particular, using a
data separation of t* = 0.1 leads to the reduction of the stan-
dard deviation by two orders of magnitude, i.e., from 40 to
0.4. This crucial insight reveals that the uncertainty is related
to a process with a time scale corresponding to t* ⬃ 0.1,
which is four orders of magnitude smaller than ␶R. This pro-
vides a consistent picture with our earlier estimates.
The smallest uncertainty of our calculations is for the
data separation of t* = 0.01, and is estimated by the standard
deviation of a single “measurement” divided by the square
root of the number of measurements, i.e., 0.24/ 冑130⬇ 0.02.
While this uncertainty is quite small, Fig. 1共c兲 suggests that
the value of the stress autocorrelation function at t* = ␶R is
also about 0.02, implying a relative uncertainty of 100%.
While this conclusion is reasonable in light of the data points
in Fig. 1共b兲, it is also apparent that the running averages
共3兲 ties.
To understand why the running average has much
smaller uncertainties we analyzed the uncertainty as a func-
tion of time. We defined an “error” function as the difference
between the stress autocorrelation function and its running
average, and plotted this quantity as a function of time. Fig-
ure 3 shows the time variation of this difference at short
times, t* ⬃ 5, and for t* ⬃ ␶R. In both time ranges the error is
dominated by short time scale oscillations, rather than by
white noise. This is the likely reason that running averages
provide such an improvement in the data quality. They sim-
ply average out short time oscillations, while preserving the
nonoscillatory characteristics of 具␴共t* + ␶R兲␴共t*兲典. We also
note that the average error according to Fig. 3 is about 0.02
independent of the time window considered. This is again
consistent with the fact that most of the error is due to high-
frequency bond oscillations.

Downloaded 03 May 2006 to 165.123.34.86. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp
144909-4 Sen, Kumar, and Keblinski
FIG. 4. Short time scale behavior of the error autocorrelation function ex-
hibiting oscillatory decay similarly as the stress autocorrelation function
shown in Fig. 1共a兲.
DISCUSSION
Our analysis of the uncertainty in the stress autocorrela-
tion function calculated for the bead-spring polymer melt
demonstrates that it is dominated by the stress associated
with high-frequency bond oscillations. To further understand
these issues we show the uncertainty autocorrelation function
in Fig. 4. This function is characterized by short time decay-
ing oscillations, with the oscillation period being the same as
the one characterizing the stress autocorrelation function 关see
Fig. 1共a兲兴. The relaxation time constant estimated from the
data in Fig. 4 is about 0.15t*. If we take this short time,
rather than the polymer relaxation time, as ␶c, our estimate
for the uncertainty following Eq. 共2兲 becomes
␦共具␴共t*兲␴共0兲典兲 = 冑 2 ⫻ 0.15 2
200 000
具␴ 典 ⬇ 0.08,
which is in the same order of magnitude as our results 共see
Fig. 3兲. As we already discussed, this uncertainty is further
reduced by calculating running averages.
An important question is why this oscillatory component
dominates over the uncertainties related to events occurring
at the chain relaxation time. To understand the relative con-
tributions of these two quantities to the overall uncertainty
we compare 冑␶bond具␴共0兲2典 and 冑␶R具␴共␶R兲␴共0兲典. 关Since the
stress autocorrelation function varies as 共t*兲−0.5 in the Rouse
regime 关Fig. 1共c兲兴 this product is expected to be constant for
any Rouse mode of the chain.兴 Clearly, these terms will sig-
nify the relative importance of the bond fluctuations and
chain relaxation in the estimation of uncertainties following
Eq. 共2兲. Here ␶bond ⬃ 0.15 corresponds to the bond relaxation
time, and all other values are obtained from Fig. 1. The nu-
merical value for the bond contribution is 冑␶bond具␴共0兲2典
⬃ 20, while that for the chain contribution is
冑␶R具␴共␶R兲␴共0兲典 ⬃ 0.4. Thus, we end up with the relatively
surprising conclusion that the uncertainties arising from bond
fluctuations dominate the contribution from chain relaxation:
the bond fluctuations then determine the ultimate uncertainty
in the transport properties determined from an equilibrium
simulation. An alternative means of understanding the suc-
cess of equilibrium calculations in estimating the stress au-
tocorrelation function is that Vladkov and Barrat incorrectly
suggest that the run time independent part of Eq. 共2兲 is pro-
portional to 冑␶R具␴共0兲2典: rather they should have employed
冑␶bond具␴共0兲2典. In the absence of bond fluctuations, e.g., if one
Downloaded 03 May 2006 to 165.123.34.86. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp
J. Chem. Phys. 124, 144909 共2006兲
FIG. 5. Viscosity integrals of the stress autocorrelation function showing
that both “raw” data and running averages of the stress autocorrelation func-
tion 关see Fig. 1共b兲兴 yield the same results upon integration.
used fixed length bonds, the relevant quantity on the right
side of Eq. 共2兲 would be much smaller, 冑␶R具␴共␶R兲␴共0兲典. The
uncertainty in the stress autocorrelation function would thus
be even smaller in these cases, emphasizing again that equi-
librium MD simulations can yield reliable estimates of trans-
port properties in these cases.
Finally, we note another important point that integrated
quantities such as the viscosity and the complex modulus
共for frequencies much smaller than the bond fluctuation兲 are
not affected by running average procedures. For example, the
viscosity is obtained via Eq. 共1兲. To illustrate the uncertainty
propagation in this quantity we plot the viscosity integrals
共Fig. 5兲,
共4兲 ␩共t兲 =
V
k BT
冕
0
t
具␴共t⬘兲␴共0兲典dt⬘ , 共5兲
as a function of time. The integrals are performed on the
“raw data” stress autocorrelation function, as well as on the
stress autocorrelation function smoothened by the running
average procedure. Figure 5 emphasizes that both integrals
are indistinguishable from each other. This is hardly surpris-
ing since the relatively rapid bond oscillations are simply
averaged out in this integration procedure. This demonstrates
that the running average procedure is just a convenient way
of smoothing the stress autocorrelation data, but that it plays
a minimal role in determining time integrated quantities.
The uncertainty of the viscosity integral can be analyzed
quantitatively using the following procedure. First, we for-
mally decompose the stress autocorrelation function obtained
in the simulations into an “exact” value and an “error.”
␩共t兲 = ␩共t兲 + ⌬␩共t兲 =
V
k BT
冕
0
t
具␴共t⬘兲␴共0兲典exactdt⬘
+
V
k BT
冕 t
0
具␴共t⬘兲␴共0兲典errordt⬘ . 共6兲
The second term in Eq. 共6兲 can be considered as an integral
over a random walk. By analogy to molecular diffusion, the
diffusion constant of the random walker can be evaluated as
D = 兰⬁0 具V共t兲V共0兲典dt, where V共t兲 is the “velocity” of the stress
autocorrelation function error defined as V共t兲 = 共V / kBT兲
⫻关d共具␴共t兲␴共0兲典error兲 / dt兴. Figures 6共a兲 and 6共b兲 show running
integrals corresponding to the D values. We calculated two
144909-5 Analysis of uncertainties in polymer viscoelastic properties
FIG. 6. 共a兲 Short and 共b兲 long time behaviors of the running integrals of the
stress error velocity autocorrelations function for the two data sets discussed
in the text. Both integrals converge to the same value at early times t* ⬍ 1
emphasizing that the most important component of the error is associated
with molecular oscillations.
integrals. The first one uses 具V共t兲V共0兲典 obtained from the
error data up to t* = 1500, where the error is defined as the
difference between the stress autocorrelation function and its
running average. The second one uses the error data from
t* = 4000 to t* = 10 000, where the error is defined simply as
the stress autocorrelation function itself. As shown in Fig.
6共a兲, both integrals are essentially the same and converge to
a constant value at very early times t* ⬍ 1 关see also Fig.
6共b兲兴. This demonstrates that the error is correlated only over
a time scale associated with molecular motion, and the errors
associated with polymeric time scales are negligible in com-
parison. The diffusion constant of the error, D = 3 ⫻ 10−3 关see
Fig. 6共b兲兴 allows us to evaluate the error term of the viscosity
integral as 冑具⌬␩共t兲2典 = 冑2Dt. Using the data plotted in Fig. 5
and taking t* = 3000, for which ␩共t兲 is essentially converged,
one obtains
冑具⌬␩共3000兲2典 冑2 ⫻ 共3 ⫻ 10−3兲 ⫻ 3000
⬇ ⬇ 10 % . 共7兲
␩共3000兲 50
The above results demonstrate that it is possible to obtain the
viscosity of the polymeric system characterized by a relax-
ation time ␶R ⬃ 1500, provided that the simulations run is 100
times the relaxation time. We also note that integrating the
stress autocorrelation function beyond t* = 2␶R, only increases
the error since the exact part of the integral is essentially
converged and the noise is just increasing. For example, at
t* = 30 000 the error increases to over 30%.
An alternative way of estimating the error in the viscos-
ity is to evaluate the integrals of the three independent off-
diagonal components of the stress autocorrelation function.
Such integrals, along with their average, are shown in Fig. 7.
At t* = 3000, the three integrals are within ⬃20% from the
average value, yielding an estimate for the error of
Downloaded 03 May 2006 to 165.123.34.86. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp
J. Chem. Phys. 124, 144909 共2006兲
FIG. 7. Viscosity integrals of stress-stress autocorrelation functions of the
three independent stress components. The solid line represents the average
viscosity integral.
20% / 冑3 ⬇ 12%, which is almost the same as the error ob-
tained from Eq. 共7兲. The data in Fig. 7 show that the integral
of the average stress correlation function is smoother and has
a better defined convergent value than the integrals of the
independent stress components. While we believe that this
result is a simple consequence of error reduction on averag-
ing, we cannot exclude the possibility of some systematic
error cancellation, that is outside our understanding. We em-
phasize again that the integration of the stress autocorrelation
function beyond t* = 2␶R increases the error, which is even
more pronounced for the integrals of the autocorrelation
functions of the independent stress components. We want to
point out that the data presented in Fig. 7 are good agreement
with those presented in Ref. 10, including the uncertainty.
CONCLUSIONS
We have conclusively demonstrated that the uncertain-
ties associated with the stress autocorrelation function ob-
tained from equilibrium molecular dynamics can be made
arbitrarily small with increasing simulation time. More im-
portantly, quantities such as the viscosity, the modulus, and
the stress correlation functions have manageable errors even
for simulation times as small as O共100␶R兲. These results are
a direct manifestation of the fact that the dominant error in
these calculations can be attributed to the rapidly varying
bond oscillation, which has a time constant which is O共104兲
smaller than ␶R for chain sizes N = 80. The construction of
models with fixed bond lengths, which presumably will re-
move this contribution, will provide even smaller uncertain-
ties in calculated transport coefficients. We thus strongly ar-
gue that equilibrium molecular dynamics simulations are an
excellent vehicle to quantitatively enumerate the linear rheo-
logical properties of polymeric systems.
ACKNOWLEDGMENTS
Financial support for this work was provided by the Na-
tional Science Foundation through a GOALI Grant No. CMS
0310596. The authors thank Professor Venkat Ganesan 共Uni-
versity of Texas兲 for bringing the work of Vladkov and Bar-
rat to our attention. Additionally, we thank Dr. Valdkov and
144909-6 Sen, Kumar, and Keblinski J. Chem. Phys. 124, 144909 共2006兲

1
Dr. Barrat for candid discussions of their paper and its rela- P. G. deGennes, Scaling Concepts in Polymer Physics 共Cornell Univer-
tionship to this work. sity Press, Ithaca, NY, 1979兲.
2
M. Doi and S. F. Edwards, The Theory of Polymer Dynamics 共Clarendon,
Oxford, 1986兲.
NOMENCLATURE 3
T. Kreer, J. Baschnagel, M. Muller, and K. Binder, Macromolecules 34,
Length parameter ⫽ ␴ 1105 共2001兲.
4
Well depth in LJ W. Paul, K. Binder, D. W. Heermann, and K. Kremer, J. Chem. Phys. 95,
interaction ⫽ ␧ 7726 共1991兲.
t* = t / ␶; ␶ = ␴冑m / ␧
5
K. Kremer and G. Grest, J. Chem. Phys. 92, 5057 共1990兲.
Reduced time ⫽ 6
M. Kroger, W. Loose, and S. Hess, J. Rheol. 37, 1057 共1993兲.
Reduced 7
Z. Xu, R. Khare, J. J. de Pablo, and S. Kim, J. Chem. Phys. 106, 8285
temperature ⫽ T * = k BT / ␧ 共1997兲.
␳* = 0.85= ␳␴3
8
Reduced density ⫽ S. Sen, S. K. Kumar, and P. Keblinski, Macromolecules 38, 650 共2005兲.
9
D. Bedrov, G. D. Smith, and J. F. Douglas, Europhys. Lett. 59, 384
Reduced
共2002兲.
viscosity ⫽ ␩* = ␩ / kBT␴−3␶ 10
M. Vladkov and J.-L. Barrat, Macromolecular Theory and Simulation 共in
Reduced stress press兲, also available on the web at http://arxiv.org/abs/cond-mat/
correlation ⫽ 具␴*xy共t兲␴*xy共t兲典 = 具␴xy共t兲␴xy共t兲典 / kBT␴−3 0507229.

Downloaded 03 May 2006 to 165.123.34.86. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp