You are on page 1of 7

Computational Materials Science 69 (2013) 428–434

Contents lists available at SciVerse ScienceDirect

Computational Materials Science


journal homepage: www.elsevier.com/locate/commatsci

Comparison of Green–Kubo and NEMD heat flux formulations for


thermal conductivity prediction using the Tersoff potential
Masoud H. Khadem ⇑, Aaron P. Wemhoff
Department of Mechanical Engineering, Villanova University, Villanova, PA, United States

a r t i c l e i n f o a b s t r a c t

Article history: The phonon-based thermal conductivity of graphene sheets is estimated using equilibrium and non-equi-
Received 24 August 2012 librium molecular dynamics simulations. Different possible heat flux formulations, as a result of using the
Received in revised form 10 December 2012 three-body Tersoff potential, have been examined to calculate the thermal conductivity by using equilib-
Accepted 12 December 2012
rium molecular dynamics (EMD) with Green–Kubo relations. Non-equilibrium molecular dynamics
Available online 17 January 2013
(NEMD) simulations are also performed to compare the heat flux formulations. We have found that
the selected heat flux formulation has a prominent impact on the calculated thermal conductivity. A heat
Keywords:
flux formulation choice is recommended that best maintains consistency between applied and calculated
Green-Kubo
Tersoff
heat flux values in NEMD simulations.
Molecular dynamics Ó 2012 Elsevier B.V. All rights reserved.
Heat flux formulation
NEMD

1. Introduction but the effect of the heat flux calculation method is not yet
understood.
Experimental measurements [1,2] and theoretical calculations The Tersoff potential is widely used to model the covalent inter-
[3–7] have been performed by various researchers to analyze the actions in the structure of graphene by suggesting the following
unusually high thermal conductivity of graphene. The experimental form:
measurements range from 600–5000 W/m K and are influenced by
the experimental method and the presence of a substrate [8]. In /ðr ij Þ ¼ fc ðrij Þ½/R ðrij Þ  bij /A ðr ij Þ ð1Þ
addition, the theoretical studies report a wide range of thermal con- R A
where fc ðrij Þ is the cut-off function, / ðr ij Þ and / ðr ij Þ are the repulsive
ductivity values for graphene, which is primarily due to differences and attractive parts of the potential, respectively, and bij is the three-
in modeling techniques employed by various investigators. The the- body term. More details and the coefficients for carbon–carbon
oretical studies commonly employ either atomistic simulations interactions can be found in the original papers by Tersoff [14,15].
such as molecular dynamics (MD) [9], Green’s functions [7], BTE- Recently, Lindsay and Broido [16] proposed new parameters for
based approaches [10], or theory based on the original work by graphene that model the phonon dispersion in the center of the
Klemens [11]. For example, Zhixin et al. [4] applied non-equilib- Brillouin zone more precisely, with the result of a modified thermal
rium molecular dynamics (NEMD) to predict thermal conductivity conductivity prediction. They showed that the change in Tersoff
values of 300 W/m K and 500 W/m K at 300 K for graphene in parameters results in an increase in predicted thermal conductivity
armchair and zigzag directions, respectively, but Hu et al. [3] from 1900 W/m K by the conventional Tersoff to 3500 W/m K with
obtained values of 1500 W/m K and 2000 W/m K in armchair their optimized version, yet they did not provide many details on
and zigzag directions also using NEMD at 300 K. BTE-based their MD simulation approach. This modified Tersoff potential is
approaches have also shown a large range of predictions now generally accepted by the MD community over the original
(1000–5000 W/m K) [12]. Therefore, it is concluded that the choice Tersoff when modeling thermal transport in graphene [13]. There-
of the theoretical approach impacts the predicted values of thermal fore, the optimized Tersoff potential is used in all simulations in this
conductivity, and that variations within each approach can also study.
influence the predictions. The selected potential function is known Equilibrium and non-equilibrium molecular dynamics are two
to influence the predicted thermal conductivity of graphene [13], widely used approaches to predict the thermal conductivity of a
system of atoms in a solid structure [17–19]. Schelling et al. [20]
⇑ Corresponding author. studied and compared the features of each method. They pointed
E-mail addresses: masoud.khadem@villanova.edu (M.H. Khadem), aaron. out that NEMD might contain nonlinear effects due to the applica-
wemhoff@villanova.edu (A.P. Wemhoff). tion of the required temperature gradient. They also noted that

0927-0256/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.commatsci.2012.12.016
M.H. Khadem, A.P. Wemhoff / Computational Materials Science 69 (2013) 428–434 429

while both EMD and NEMD approaches exhibit finite size effects, and 14–120 W/m K, respectively, depending on the heat flux for-
these effects are much more severe in NEMD due to the presence mulation. However, no similar study has been performed for
of interfaces at the heat source and sink. Furthermore, EMD facili- graphene. In addition, the question of which formula is the most
tates thermal conductivity prediction in all directions using one accurate for heat flux estimation for graphene using the Tersoff po-
simulation, whereas NEMD requires the use of a thermal gradient tential is still unanswered. The knowledge of which formulation
and therefore only enables the calculation of thermal conductivity provides the most reliable prediction of heat flux is key to thermal
in one direction. Therefore, EMD is particularly useful for geome- conductivity calculation of graphene using MD simulations. There-
tries where periodic boundary conditions can be applied. However, fore, this study examines the various heat flux formulations to
the basis of EMD is the fluctuation–dissipation theorem, while ascertain which appears to provide the most consistent predictions
NEMD’s basis in Fourier’s Law of conduction makes NEMD to be when compared to a direct method. This will be accomplished by
analogous to experimental measurements. Furthermore, EMD is comparing the applied heat flux using a direct method with the
often computationally more expensive and the results are more calculated heat flux using the various correlations.
sensitive to the simulation parameters [21]. NEMD simulations provide the means to apply a heat flux by
In EMD, the system is set to the desired temperature, and then a scaling the velocity of atoms in two regions of the simulation
constant energy scheme is used with the well-known Green–Kubo domain to generate a heat flow of interest. A few studies have com-
[22,23] relations to calculate the thermal conductivity tensor. MD pared the applied heat flux and the calculated heat for a system of
simulations may be applied in different statistical ensembles atoms interacting with a pairwise potential function [19,33,34].
[24], namely canonical (NVT), grandcanonical (lPT), and microca- They all found good agreement between the calculated heat flux
nonical (NVE). It is worth noting that the derivations of Green [22] in EMD and NEMD simulations. However, to the best of our knowl-
and Kubo [23] have been done in t different ensembles: the former edge, there has been no attempt to compare the heat flux values in
in microcanonical and the latter in grandcanonical. Lepri et al. [25] the presence of the Tersoff potential. The following sections use
resolved this discrepancy by noting that if the microcanonical EMD and NEMD simulations to compare the calculated thermal
ensemble is used, then the thermal conductivity might diverge conductivity values and determine how well these formulations
non-trivially unless the velocity of the center of mass of the system maintain consistency with applied heat fluxes.
is set to zero or alternatively some terms are subtracted from the
calculated heat flux vectors. However, the canonical ensemble
can also be used with the Green–Kubo formula to predict the ther- 2. Heat flux formulations
mal conductivity by applying additional thermal forces to all the
atoms [26]. Combining Eq. (2) with the statistical definition of heat flux vec-
Another issue regarding the EMD is the formulation of heat flux tor, ~
j ¼ ðjx ; jy ; jz Þ, the following equation is obtained [35]:
vectors when the atoms interact under the influence of a three-
X
body potential. Hardy [27] was the first investigator to derive the ~j ¼ 1 d r i ei
~ ð5Þ
heat flux formulation for a pairwise potential by using localization X dt i
functions for mass density, momentum density, and energy den-
sity. He managed to solve the continuity equation for energy con- where ~ ri is the position of each atom, ei is the energy, ei ¼
P
servation of the following form: 1
2
v i j2 þ i /ij Þ, of each atom in the system, and X is the system’s
ðmj~
volume. For a simulation that uses a pairwise potential, Eq. (5) takes
@E0 ð~
R; tÞ ~ 0 ð~ ~0 ~ the following form:
þ r~R  ½Q K R; tÞ þ Q V ðR; tÞ ¼ 0 ð2Þ
@t
!
X 1X ~
where ~ ~
R is the position vector, is time, E is the system energy, and Q ~j ¼ 1 v i ei þ
~ v iÞ
r ij ðF ij  ~
~ ð6Þ
~ are the kinetic and potential contributions of the total heat
and Q X i 2 i;j
flux. The superscript ‘‘0’’ designates the case where there is no bulk
motion in the system of particles. In the same study the two above In deriving Eq. (6) it was assumed that the potential is equally
terms were evaluated to be: divided between two atoms. However, Eq. (6) is invalid for a
many-body potential such as the Tersoff potential, which leads to
1 X
~ 0 ð~
Q ð~ v Þ~rij
F ij  ~ numerous proposed heat flux formulations. Guajardo-Cuellar
K R; tÞ ¼ ð3Þ
2 i;j et al. [32] have summarized the proposed heat flux formulations
used by various researchers.
X1 1

Dong et al. [28] considered the equation derived by Hardy [27]
~ 0 ð~
Q V R; tÞ ¼ v i j2 þ /ij ~
mi j~ vi ð4Þ for a pairwise potential (Eq. (6)) to neglect the effect of the third
i
2 2
atom in the triplet interaction. As result they arrived at Eq. (F1)
where mi is the mass of particle i, ~F ij is the inter-atomic force in Table 1. They have also assumed that the triplet potential is
vector between atoms i and, ~ v i is the velocity vector of atom i, equally divided between atoms i and j. This assumption has been
and /ij is the inter-atomic potential. Eqs. (3) and (4) were devel- also used by Li et. al. [29] except that they also included the effect
oped for a pairwise potential, but extension to a three-body poten- of atom k in the derivative of the potential (Eq. (F2)). In addition,
tial is not straightforward. Different heat flux formulations for Chen [30] used the localization function stated by Hardy [27],
three-body potentials have been proposed recently [18,28–30]. and after some algebraic steps they derived a theoretical formula-
For example, Chen [30] used the same approach employed by tion for the heat flux to be used with the Tersoff potential. Later
Hardy to derive the heat flux and stress formulations for Stillin- Guajardo-Cuellar et al. [32] rearranged their equation and wrote
ger–Weber [31] and Tersoff potentials. However, the results it in the form of Eq. (F4). The same authors also derived Eq. (F3),
obtained by other investigators are not identical. So, Guajardo- which is a result of the term by term differentiation of Eq. (5),
Cuellar et al. [32] compared the existing heat flux formulations and it contains the position vector of each atom with respect to
for carbon and germanium in the diamond structure by computing the origin rather than the inter-atomic distance of the interacting
the thermal conductivity using EMD simulations. Their results atoms. The above mentioned heat flux formulations are summa-
show that the values of thermal conductivity for diamond-like rized in Table 1, and further background and the assumptions
carbon and germanium vary in the range of 920–4500 W/m K made for each formulation have been discussed in detail [32].
430 M.H. Khadem, A.P. Wemhoff / Computational Materials Science 69 (2013) 428–434

Table 1
Possible formulations for the heat flux vector to be used with the Tersoff potential.

Reference Heat flux vector ð~



Dong et al. [28] (F1)

X
n X
n X
n  !
~j ¼ 1 v i ei þ
~ r ij v j 
~ ~
@/i
X i¼1 i¼1 j¼1
@r j

Li et. al. [29] (F2)


!
Xn
~j ¼ 1 ~
1
v i ei þ ~rji ðD~F j  ~
v j  D~F i  ~
1
v i Þ  ð~rjk  ~rki ÞðD~F k  ~
v kÞ
X i¼1 2 2

Guajardo-Cuellar et al. [32] (F3)


!!
X
n X
n X
n X
n X
n
~j ¼ 1 v i ei þ
~ r i ð~
~ v iÞ 
Fi  ~
1
~
ri ð~ vi  ~
F ij  ð~ v jÞ þ ~ vi  ~
F ik  ð~ vkÞ
X i¼1 i¼1
2 i¼1 j–i k–i;j

Chen [30,32] (F4)


!!
Xn
1X n X n Xn
~j ¼ 1  ~ v i ei þ rij ð
@U ij
v
~ kÞ þ ~ r ik ð
@U ij
vkÞ
~
X i¼1
2 i¼1 j–i
@~r j k–i;j
@~
rk

It is worthwhile to mention that in all the above heat flux for- conductivity calculation in both directions at once, the simulations
mulations it is assumed that there is no bulk motion in the system here are performed separately for the two directions so that we can
(i.e., the velocity of the center of mass of particles is zero). How- use rectangular graphene sheets (instead of square), which are
ever, if there is a bulk velocity in the system then the system center longer in the direction of interest for the thermal conductivity cal-
of mass velocity vector ~ v G must be subtracted from each particle’s culation. In these simulations, the graphene sheets are 10 nm long
velocity vector to obtain correct results. in the direction in which the thermal conductivity is computed and
2 nm wide. We ensured that a 2 nm width domain is large enough
3. EMD simulations to resolve the Brillion zone and avoid divergence of the thermal
conductivity prediction. Periodic boundary conditions are imposed
3.1. Method to avoid the boundary effects caused by the long mean free path
(up to 600 nm) [36] in the graphene structure. To reduce statistical
Dynamic properties such as thermal conductivity are calculated error, each 2  105 time step period is treated as an independent
in EMD based on the fluctuation dissipation and linear response simulation with new initial velocities. All of the EMD simulations
theorem. This method applies the fact that the heat flow in a sys- are performed at 300 K. The temperature is set by equilibrating un-
tem of particles in the equilibrium state fluctuates around zero. der a constant temperature (NVT) using a Berendsen thermostat
The heat flux vectors and their correlations are computed through- [37] for 2  105 time steps. The velocity of the center of mass of
out the simulations. The time needed for Heat Auto-Correlation the system is set to zero at the beginning of each simulation, which
Functions (HACF) to decay to zero is then used through the maintains a zero center of mass velocity during the simulation that
Green–Kubo relation to predict the thermal conductivity. Statisti- prevents the divergence of the heat flux auto-correlation functions
cal thermodynamics coupled with extensive algebra results in and erroneous values of thermal conductivity. After the system
the following equation for thermal conductivity: reaches the desired temperature (300 K), the thermostat is turned
Z s off and the system relaxes at a constant energy (NVE) for 105 time
1 steps to mitigate the effects of velocity scaling via the thermostat.
kx ¼ hjx ðtÞ  jx ð0Þidt ð7Þ
XkB T 2 0

where T is the equilibrium temperature, kB is the Boltzmann con-


stant, and s is the time needed for HACF decay. In order to use
Eq. (7) with EMD simulations, the following discrete form [20] is
used with the simulations:

Dt X M X
1 Nm
kx ¼ 2
jx ðm þ nÞjx ðnÞ ð8Þ
XkB T m¼1 N  m n¼1

where Dt is the simulation time step, N represents the total number


of time steps of the simulation (after equilibrium), and M is the
number of time steps for the correlation of heat flux vectors
(MDt ¼ sÞ. The outer summation in Eq. (8) represents the HACF.

3.2. Simulation details

Eqs. (F1)-(F4) are substituted into Eq. (8) to compute the ther- Fig. 1. Temperature distribution and energy variation for the first 106 time steps of
mal conductivity of graphene sheets in both zigzag and armchair an EMD simulation. The first 2  105 time steps are in the NVT ensemble, where the
directions. Although the Green–Kubo relation enables thermal temperature fluctuations are allowed by the Berendsen thermostat.
M.H. Khadem, A.P. Wemhoff / Computational Materials Science 69 (2013) 428–434 431

Fig. 1 shows the temperature distribution and energy conservation


for the first 106 time steps in the NVE portion of the simulations.
Simulations continue after equilibrium for an additional 107 time
steps in the microcanonical ensemble to sample the heat flux
vectors. Following Li et al. [29], the statistical error is reduced by
dividing the sampling stage into five individual intervals for ther-
mal conductivity prediction. The averaged thermal conductivity
in five intervals is then calculated upon the completion of the sim-
ulation. Eq. (8) is used to calculate the values of thermal conductiv-
ity upon the completion of the simulations. In our simulations,
t ¼ 0:5 fs, and X is obtained by multiplying the surface area of
the graphene sheets by the graphite interlayer nominal distance
(0.335 nm). The summation limit of HACFs (M) is chosen to be
105 time steps, which corresponds to 50 ps. This choice of M, as
will be shown, is large enough to let the HACFs to decay to zero Fig. 3. The average thermal conductivity of graphene in zigzag and armchair
directions with three heat flux formulations.
for thermal conductivity convergence. However, the value of the
upper limit can be varied for different materials and it increases
with the mean free path of the material. Furthermore, in order to
thermal conductivity achieved by Zhixin et al. [4] using NEMD sim-
obtain a good statistical average [20], the total number of time
ulations. Fig. 3 also reports armchair and zigzag thermal conductiv-
steps (N) in Eq. (8) must be much larger than M, and is therefore
ity values much lower than those obtained by Evans et al. (in the
selected to be 2  106 (1 ns) throughout this study.
range of 8000–10000 W/m K). We relate this difference to a variety
of factors including the use of a different potential model and heat
3.3. EMD results flux formulation. It can be concluded that kEq:ðF2Þ > kEq:ðF1Þ > kEq:ðF4Þ
for both zigzag and armchair directions. The same order has been
The thermal conductivity convergence of a graphene sheet in observed for carbon and germanium in the diamond structure [32].
the zigzag direction is shown in Fig. 2 by averaging four indepen- One interesting result stems from the use of Eq. (F3) for the
dent thermal conductivity curves and using Eq. (F1) for heat flux computation of the heat flux vectors. It is found that considering
calculations. The figure also shows the corresponding HACF decay this equation for heat flux calculations will result in the divergence
to indicate the thermal conductivity convergence. The figure shows of thermal conductivity and HACFs for domain sizes larger than
that although the thermal conductivity might not independently 4 nm. This has been shown in Fig. 4 where thermal conductivity
converge for each simulation, the statistical error can be reduced in the armchair direction is plotted for different sizes in the range
by averaging the thermal conductivity. The same behavior has of 2–5 nm. In addition, an oscillatory behavior in the thermal con-
been observed by Evans et al. [5]. Fig. 2 also enables the computa- ductivity starts to grow at the domain size of 4 nm and it leads to
tion of the statistical error in the predictions of thermal conductiv- the divergence of HACFs for graphene sheets larger than 8 nm. We
ity values. relate this phenomenon to the fact that Eq. (F3) is the only heat
The average thermal conductivity of graphene sheets along the flux vector formulation that is based solely on the atomic position
zigzag and armchair directions are shown in Fig. 3 using the heat ~
ri rather than the relative atomic position ~ r ij , and the use of peri-
flux formulations of Eqs. (F1), (F2), and (F4). In general, the figure odic boundary conditions results in nonrealistic atomic position
shows that while the thermal conductivity is highly influenced vectors. This phenomenon has been also pointed out by Donadio
by the selected heat flux formulation, the thermal conductivity and Galli [39] as ‘‘ill-defined’’ position vectors in periodic systems.
convergence is achieved regardless of the applied formulation. It Therefore, the use of Eq. (F3) might not be appropriate for heat flux
is also deduced that thermal conductivity values in the armchair calculations with periodic boundary conditions, and it leads to a
and zigzag directions are of the same order except for the case divergence in thermal conductivity.
where Eq. (F2) is used to compute the heat flux. Furthermore, The achieved values of thermal conductivity are tabulated in
applying Eq. (F2) provides a good agreement with EMD predictions Table 2 for both zigzag and armchair directions. We have used
of thermal conductivity values by Hu et al. [3] and Haskins et al. the exponential fits to each curve in order to calculate the
[38]. On the other hand, the results obtained by using Eqs. (F1) converged values and the associated errors (see Fig. 3). For all
and (F4) are also reasonable and in agreement with the obtained

Fig. 2. Thermal conductivity and HACF convergence for a graphene sheet in the Fig. 4. Thermal conductivity convergence in the armchair direction for different
zigzag direction when heat flux is computed by Eq. (F1). domain sizes using Eq. (F3).
432 M.H. Khadem, A.P. Wemhoff / Computational Materials Science 69 (2013) 428–434

Table 2 path of graphene compared to the simulation domain size. The


Computed thermal conductivities in W/m K of graphene using different heat flux resulting heat flux can then be written as:
formulations.
e
Direction Eq. (F1) Eq. (F2) Eq. (F3) Eq. (F4) jx ¼ ð10Þ
2Adt
Zigzag 830 ± 6% 1600 ± 8% 510 ± 20% (5 nm sheet) 250 ± 8%
Armchair 800 ± 8% 1800 ± 10% 317 ± 15% (5 nm sheet) 300 ± 10% where A is the cross sectional area of the system perpendicular to
the heat flow. The factor of 2 in the denominator of Eq. (10) stems
from the fact that the heat flows from the hot and cold regions in
two directions. The velocities of all the particles in the hot and cold
the formulations, except for the case of Eq. (F3), the corresponding
regions are scaled according to the following equation:
statistical errors are within 10%. It should be noted that the values
in the Table require a quantum correction (QC) to the classical MD v inew ¼ ~
~ v G þ að~
v iold  ~
v GÞ ð11Þ
results for comparison with experimental measurements as per-
formed by some investigators [40,41]. This is due to the high Debye where ~v G is the velocity of center of mass of the associated hot or
temperature of graphene where only at temperatures higher than cold region, and a is calculated from the following equation:
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
the Debye temperature are all the phonon modes excited. How- e
ever, this study does not impose the QC as the interest here lies a¼ 1 ð12Þ
KE
in the comparison of the existing heat flux formulations, and the P
QC affects the result of each formulation in an equivalent manner. v i j2 , is the kinetic energy of the particles in the
where KE ¼ i 12 mi j~
Furthermore, the effect of a quantum correction was recently hot or cold regions. In our simulations the bulk velocity of the sys-
shown to be insignificant on predictions of the thermal conductiv- tem of particles in hot and cold regions is controlled by applying a
ity of graphene when the system temperature is greater than 280 K Berendsen thermostat, so by setting the velocity of the center of
[42]. mass of the two regions to zero at the beginning of the simulations,
Furthermore, the effect of finite domain size must also be men- Eqs. (11) and (12) reduce to:
tioned. Although large size effects have been reported in NEMD rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
e
simulations, Schelling et al. [20] and Sellan et al. [43] have shown v inew ¼ ~
~ v iold 1 ð13Þ
KE
that these effects are lower in EMD simulations. The EMD domain
size requires that the phonon–phonon scattering processes are The simulation domain is divided into several segments and the
unalte x red by the presence of periodic boundary conditions; this temperature of each segment is recorded at each time step. The
requirement has shown to be low (e.g. 128 atoms for solid simulations are continued until a linear temperature gradient is
argon[17]). The selected 10 nm graphene sheet length used in this achieved.
study contains approximately 2000 atoms, and the comparison of
results for a 10 nm domain length with smaller domain sizes 4.2. Simulation details
shows very low size effects (within the statistical error) on the cal-
culated thermal conductivity. The simulated graphene sheets are 22 nm long and 3 nm wide.
The length of the simulation domain is divided into five regions
4. NEMD simulations for simulations calculating kx in either the zigzag or armchair direc-
tion: a hot region, a cold region, two nonlinear regions, and a middle
In this section we apply NEMD simulations that generate a heat region. Fig. 5 shows the graphene sheet geometry and locations of
flow into the system using a direct approach, and then apply the these regions. Separate simulations have been performed where
formulations of Table 1 to compare the calculated heat flux with zigzag and armchair edges appear along the long edge of the do-
the applied heat flux. main. Hot and cold regions are equally 1 nm wide, which contains
a sufficient number of atoms to calculate the average ensemble
temperature. The portion of the domain between the hot and cold
4.1. Method
regions is divided into 20 segments (each 0.5 nm wide) for statistics
NEMD simulations provide a means to calculate the thermal
conductivity in a way analogous to the experimental measure-
ments either by imposing a thermal gradient into the system of
particles or by introducing a heat flow. Both approaches enable
thermal conductivity prediction using the Fourier’s Law of
conduction:
j
kx ¼  dTx ð9Þ
dx

where x is the direction of heat flow (or thermal gradient).


The method of imposing a thermal gradient by controlling the
temperature of two regions in the system is straightforward to
implement, but it often requires long simulations to obtain a stea-
dy state linear temperature distribution. In this study, the method
proposed by Jund and Jullien [19] is used to apply a specified heat
flux by scaling the atomic velocities in two regions of the simula-
tions in such a way that at each time step a specified amount of
energy, e (set to 10% of KBT, or 2.6 meV at 300 K), is both added
to a hot region and subtracted from a cold region. Periodic bound- Fig. 5. Simulation domain of graphene sheets in the zigzag and armchair directions.
ary conditions are imposed to all boundaries to mitigate the effect The hot, cold, middle, and nonlinear regions are shown. Lengthwise dimensions are
of boundary scattering effects from the high phonon mean free shortened in this figure for clarity.
M.H. Khadem, A.P. Wemhoff / Computational Materials Science 69 (2013) 428–434 433

collection. This 10 nm-wide volume may be separated into two no


5  104 nlinear regions where the influence of the nearby hot/cold
region distorts the statistics, and a middle region where this distor-
tion is not prevalent. The statistics collected in the middle region
5  105 alone are therefore used for heat flux and thermal gradient
calculations. The nonlinearity in the profile is due to phonon scat-
tering at the boundary between the temperature-controlled regions
and the uncontrolled regions. The approximate width of the nonlin-
ear regions is 1–2 nm. The instantaneous temperature of each
segment is calculated using the following equation derived by com-
bining the definition of kinetic energy with the equipartition
theorem:
*  +
1 X 1 2
Ti ¼ v jj
mj j~ ð14Þ
3kB Ni j
2 Fig. 7. The Temperature distributions, average temperature, and the linear fit in the
middle region for the simulation of a graphene sheet in the zigzag direction.

where N i is the number of atoms in segment i. The NEMD simula-


tions start with time steps of an NVT simulation using the Berend-
method to the results obtained from simulations in the armchair
sen thermostat until the system reaches the desired temperature
and zigzag directions, the computed temperature gradients are
(300 K). The hot and cold regions are then initiated to generate a
calculated to be 10.4 K/nm and 6.2 K/nm for zigzag and armchair
heat flow using the mentioned method. The simulations continued
directions, respectively. The variation in the average temperature
for another time steps to allow for a linear temperature distribution
of the middle region for each time interval is inevitable since the
to be established. The temperature and local heat flux (based on the
system is under constant heat flux and not constant temperature.
formulations of Table 1) are recorded for the rest of the simulations
However, the slope of the independent curves through the middle
(5  105 time steps). Eq. (14) is used to calculate the thermal con-
of the domain is consistent with that for the averaged curve, show-
ductivity values at the end of each 105 time step interval to reduce
ing that the averaging windows are sufficiently large.
the statistical error by averaging over five obtained thermal conduc-
As mentioned in the previous sections, adding or removing
tivity values.
energy to create hot and cold regions will generate a heat flow in
the system. The resultant applied heat flux is calculated using Eq.
4.3. NEMD results (10). This heat flux must be the same throughout the simulation
domain. However, a nonlinear temperature gradient caused by
The temperature variation of different regions in the simulation inter-region boundary scattering affects the heat flux in the regions
of the graphene sheet in the zigzag direction is shown in Fig. 6. close to both hot and cold regions. So, only the middle region was
Although the same amount of energy is added and withdrawn from used for heat flux computation.
the hot and cold regions, the steady state temperature of the mid- Eqs. (F1)–(F4) of Table 1 are used to calculate the heat flux in
dle region is not the average temperature of hot and cold regions. this region during the steady state portion of the NEMD simula-
The number of atoms in the cold region becomes larger than that of tions. The generated heat flux and the averaged applied heat flux
the hot region due to the lower temperature, which results in a in the middle region are shown in Fig. 8. The variation in the
higher number density and larger pre-adjusted kinetic energy in applied heat flux corresponds to the temperature fluctuations in
the cold region. Therefore a larger velocity scaling parameter (see the middle region. Fig. 8 also shows that Eq. (F2) gives the closest
Eq. (12)) is associated with the hot region. However, after the estimation of the heat flux with the Tersoff potential when com-
equilibration, the average temperature of the middle region shows pared to the applied heat flux. However, the heat fluxes obtained
a constant value which demonstrates a steady state in this region. from Eqs. (F1) and (F4) are still of the same order of magnitude
Fig. 7 shows the average temperature distribution in the middle as the applied heat flux.
region for five consecutive time intervals of the simulation of a Fig. 9 shows that plotting the calculated average heat flux using
graphene sheet in the zigzag direction. The average temperature Eq. (F3) results in a much larger prediction than the applied heat
is also plotted to enable the calculation of the temperature flux. In addition, the heat flux fluctuations are substantial using this
gradient, which equals the slope of the linear fit. By applying the approach. These results are consistent with the weak convergence

Fig. 6. Temperature distribution in the hot, cold, and middle regions of the
simulation for a graphene sheet in zigzag direction. The dashed line shows a Fig. 8. The applied heat flux and the calculated heat flux and their average values in
constant time-averaged temperature in the middle region. the middle region.
434 M.H. Khadem, A.P. Wemhoff / Computational Materials Science 69 (2013) 428–434

the results of this study, EMD simulations with the Green–Kubo


relation can be applied to a system of particles interacting via
the Tersoff potential to predict the thermal conductivity with high
confidence in accuracy. The same approach used in this study can
also be implemented to investigate the performance of the heat
flux formulations with other many-body potential functions.

Acknowledgment

We acknowledge the support of National Science Foundation


(NSF) through the Grant CBET-0931507.

References

Fig. 9. The computed and averaged heat flux in the middle region using Eq. (F3). [1] A.A. Balandin, S. Ghosh, B. Wenzhong, I. Calizo, D. Teweldebrhan, M. Feng, L.
Dash line shows the average heat flux from Eq. (F3). Chun Ning, Nano Lett. 8 (2008) 902–907.
[2] S. Ghosh, I. Calizo, D. Teweldebrhan, E.P. Pokatilov, D.L. Nika, A.A. Balandin, W.
Bao, F. Miao, C.N. Lau, Appl. Phys. Lett. 92 (2008) 151911.
[3] J.N. Hu, S. Schiffli, A. Vallabhaneni, X.L. Ruan, Y.P. Chen, Appl. Phys. Lett. 97
ability for this heat flux formulation in EMD simulations (see the (2010) 133107.
explanation of Fig. 4). [4] G. Zhixin, D. Zhang, G. Xin-Gao, Appl. Phys. Lett. 95 (2009) 163103.
[5] W.J. Evans, L. Hu, P. Keblinski, Appl. Phys. Lett. 96 (2010) 203112.
The thermal conductivity values calculated based on the ap- [6] J.W. Jiang, J.S. Wang, B.W. Li, Phys. Rev. B 79 (2009) 205418.
plied heat flux and Eq. (9) are 70 W/m K and 110 W/m K for zigzag [7] Z. Huang, T.S. Fisher, J.Y. Murthy, J. Appl. Phys. 108 (2010) 094319.
and armchair directions, respectively. These values are much smal- [8] A.A. Balandin, Nat. Mater. 10 (2011) 569–581.
[9] A.J.H. McGaughey, M. Kaviany, Int. J. Heat Mass Transf. 47 (Copyright 2004,
ler than those achieved by EMD simulations. However, it is worth
IEE) (2004) 1783–1798.
noting that (1) the mechanism of heat conduction is different for [10] D.L. Nika, E.P. Pokatilov, Phys. Rev. B 79 (15) (2009).
the two methods, and (2) boundary scattering associated with [11] P.G. Klemens, D.F. Pedraza, Carbon 32 (4) (1994) 735–741.
NEMD simulations require domain sizes on the order of the mean [12] Z.W. Tan, J.-S. Wang, C.K. Gan, Nano Lett. 11 (2011) 214–219.
[13] A.P. Wemhoff, Int. J. Transp. Phenomena 13 (2012) 121–141.
free path, which is larger than the domain used in this study. As a [14] J. Tersoff, Phys. Rev. B (Condensed Matter) 37 (1988) 6991–7000.
result the values of thermal conductivity predicted by EMD and [15] T. Tersoff, Phys. Rev. Lett. 61 (25) (1988) 2879–2882.
NEMD methods for solids with large mean free path might not [16] L. Lindsay, D.A. Broido, Phys. Rev. B 82 (2010) 205441.
[17] K.V. Tretiakov, S. Scandolo, J. Chem. Phys. 120 (2004) 3765–3769.
be consistent. For example, while Zhixin Guo et al. [4] and Bi [18] C. Jianwei, C. Tahir, D. Weiqiao, W.A. Goddard, J. Chem. Phys. 113 (2000) 6888–
et al. [44] reported values in the range of 100–500 W/m K using 6900.
NEMD method with the original Tersoff potential, Evans et al. [5] [19] P. Jund, R. Jullien, Phys. Rev. B 59 (1999). 13707–13704.
[20] P.K. Schelling, S.R. Phillpot, P. Keblinski, Phys. Rev. B 65 (2002) 144306.
calculated values in the range of 4000–8000 W/m K using EMD [21] K. Esfarjani, Phys. Rev. B 84 (8) (2011).
with the same potential model. Therefore, although Sellan et al. [22] M.S. Green, J. Chem. Phys. 20 (1952) 1281–1295.
[43] and Schelling et al. [20] have found some agreement between [23] R. Kubo, J. Phys. Soc. Jpn. 12 (1957) 570–586.
[24] J.M. Haile, Molecular Dynamics Simulation: Elementary Methods, Wiley, New
EMD and NEMD in the simulations of Silicon-based materials, com- York, 1997.
paring the values of thermal conductivity obtained by EMD and [25] S. Lepri, R. Livi, A. Politi, Phys. Rep. 377 (1) (2003) 1–80.
NEMD is not beneficial and often leads to confusion. This is why [26] A. Maeda, T. Munakata, Phys. Rev. E 52 (1995) 234–239.
[27] R.J. Hardy, J. Chem. Phys. 76 (1) (1982) 622–628.
we use NEMD simulations only as a means to compare the applied
[28] D. Jianjun, O.F. Sankey, C.W. Myles, Phys. Rev. Lett. 86 (2001) 2361–2364.
heat flux using the direct method and the calculated heat flux [29] J. Li, L. Porter, S. Yip, J. Nucl. Mater. 255 (1998) 139–152.
using the various heat flux formulations. [30] Y. Chen, J. Chem. Phys. 124 (2006). 54113–54111.
[31] F.H. Stillinger, T.A. Weber, Phys. Rev. B 31 (8) (1985) 5262–5271.
[32] A. Guajardo-Cuellar, D.B. Go, M. Sen, J.Chem. Phys. 132 (2010) 104111–
5. Conclusions 104117.
[33] D.G. Cahill, R.O. Pohl, Phys. Rev. B (Condensed Matter) 35 (8) (1987) 4067–
4073.
EMD simulations were used to determine the thermal conduc- [34] R.D. Mountain, R.A. MacDonald, Phys. Rev. B (Condensed Matter) 28 (6) (1983)
tivity of graphene in both zigzag and armchair directions under 3022–3025.
different heat flux formulations for the Tersoff potential. We calcu- [35] D.A. McQuarrie, Statistical Mechanics, University Science Books, 2000.
[36] C. Berger, Z. Song, X. Li, X. Wu, N. Brown, C. Naud, D. Mayou, T. Li, J. Hass, A.N.
lated thermal conductivity values in the range of 300–2000 W/
Marchenkov, E.H. Conrad, P.N. First, W.A. de Heer, Science 312 (5777) (2006)
m K using the Green–Kubo relation for thermal conductivity pre- 1191–1196.
diction. The wide range of obtained values shows that although [37] H.J.C. Berendsen, J.P.M. Postma, W.F. van Gunsteren, A. DiNola, J.R. Haak, J.
Chem. Phys. 81 (8) (1984) 3684–3690.
the HACFs and thermal conductivity might converge, the choice
[38] J. Haskins, A. Kinaci, C. Sevik, H. Sevincli, G. Cuniberti, T. Cagin, ACS Nano 5 (5)
of heat flux formulation has a significant impact on the calculated (2011) 3779–3787.
value. NEMD simulations with a specified heat flux were used to [39] D. Donadio, G. Galli, Nano Lett. 10 (3) (2010) 847–851.
investigate different heat flux formulations with the Tersoff poten- [40] Y.H. Lee, R. Biswas, C.M. Soukoulis, C.Z. Wang, C.T. Chan, K.M. Ho, Phys. Rev. B
43 (8) (1991) 6573–6580.
tial. Our results showed that Eq. (F2) is recommended for the [41] A. Maiti, G.D. Mahan, S.T. Pantelides, Solid State Commun. 102 (7) (1997) 517–
future calculations of thermal conductivity using EMD simulations 521.
and the Tersoff potential. We also found that Eqs. (F1) and (F4) do [42] A. Cao, J. Appl. Phys. 111 (8) (2012).
[43] D.P. Sellan, E.S. Landry, J.E. Turney, A.J.H. McGaughey, C.H. Amon, Phys. Rev. B
not exhibit the same level of agreement as Eq. (F2) with the direct (Condensed Matter and Materials Physics) 81 (21) (2010) 214305 (214310
approach but still predict the heat flux in the correct range, while pp.).
using Eq. (F3) results in the divergence of HACFs and the thermal [44] B. Kedong, C. Yunfei, C. Minhua, W. Yujuan, Solid State Commun. 150 (29–30)
(2010) 1321–1324.
conductivity for graphene domain sizes larger than 5 nm. With

You might also like