Download as pdf or txt
Download as pdf or txt
You are on page 1of 69

Review of Multi-Scale Simulation in Sub-Micron Heat

Transfer

Jayathi Y. Murthy , Sreekant V.J. Narumanchi , Jose’ A. Pascual-Gutierrez ,

Tianjiao Wang , Chunjian Ni and Sanjay R. Mathur

School of Mechanical Engineering,Purdue University

West Lafayette IN 47907



National Renewable Energy Laboratory

MC 1633, 1617 Cole Boulevard, Golden, CO 80401-3393



Fluent Inc.

10 Cavendish Court, Lebanon NH 03766

Abstract

Over the last decade, interest in the simulation of micro- and nano-scale

heat transfer has lead to the development of a variety of models and numer-

ical methods for phonon transport in semiconductors and dielectrics. These

models span direct simulation using molecular dynamics, a range of mod-

els of varying fidelity based on the Boltzmann transport equation, as well as

simpler hyperbolic extensions to the classical Fourier heat conduction equa-

tion. The paper reviews the basics of phonon transport in crystals, available

models for phonon transport, as well as numerical methods for solving the

equations resulting from these models. Recommendations are made for fu-

ture work.

1
Nomenclature

 interatomic distance

  center and neighbor discrete coefficients

volumetric specific heat

density of states for polarization 

 energy density

  angular average of energy density



phonon distribution function

equilibrium distribution function

interatomic force between atoms  and 

reciprocal lattice vector

number of atoms per unit volume

Planck constant
"!#$

%
thermal conductivity
%'&
Boltzmann constant
(
wavenumber
)
wave vector
*
domain length

2
+
atom mass
,.-  ,0/
polar and azimuthal control angles per octant

1325476 heat generation rate

8 ray direction vector

9 position vector
:
time
;
temperature
;=<
lattice temperature
;"> 7?
reference temperature
;A@
reservoir mode temperature

BCED atom displacements from equilibrium

FHG phonon group velocity

IAG phonon group velocity vector (=FJG 8 )

F sound velocity
K
crystal volume
LNM B=O
Dirac delta function
LQP
Kronecker delta

R Gruneisen constant

R  inverse relaxation time

S frequency

3
T
solid angle
U :
time step
V XW
polar and azimuthal angles

Y relaxation time

1 Introduction

During the last decade, there have been numerous advances in the synthesis and

characterization of microstructures with dimensions on the sub-micron scale, with

an attendant interest in thermal transport at these scales. In the microelectronics

arena, for example, transistor channel dimensions, currently at 180 nm, are pro-

jected to fall to 70 nm by 2008 according to the International Technology Road

Map for Semiconductors (ITRS), and many research groups have already fabri-

cated devices in the 3-40 nm scale. Because of these aggressive scaling trends

and emerging technologies such as silicon-on-insulator (SOI) and new low-k di-

electrics, thermal problems are expected to be significantly exacerbated in the


;
future. In emerging thermo-electric device designs, higher Z values are being

sought by significantly decreasing the thermal conductivity of superlattice struc-

tures by exploiting phonon confinement effects. As fabrication techniques become

4
more sophisticated, increasingly well-controlled nanostructures including superlat-

tices, quantum dots and nano-composites are being fabricated whose thermal prop-

erties must be understood. Modern micro-manufacturing techniques also require a

knowledge of sub-micron thermal transport in solids. In emerging sub-picosecond

laser processing techniques, for example, complex ablation and phase change pro-

cesses happen on scales of a few nanometers, and must be understood in order to

better control the process.

Many emerging nanostructures involve semiconductors and dielectrics in which

phonon transport is the main mechanism for heat transfer. In silicon at room tem-

perature, the mean free path of thermal phonons is in the 300 nm range, compa-

rable to the channel length of modern transistors. On the other hand, the phonon

wavelength is approximately 1-5 nm, also comparable to the length scales of many

emerging nanostructures. When the phonon mean free path is comparable to the

device scale, bulk scattering events take place less frequently, and phonons travel

more ballistically. On the other hand, as devices become smaller, surface-to-

volume ratios increase, and boundary scattering begins to play a larger role. In

heterostructures, interface reflection and transmission of phonons play a critical

role in determining both lateral and normal effective thermal conductivity [1, 2].

When phonon wavelengths become comparable to device structure scales and co-

herence effects are important, wave effects must be taken into account. These

5
physics are not captured by conventional Fourier theory for heat conduction. Al-

ternatives employing the Boltzmann transport equation (BTE) [3, 4, 5, 6] are based

on a particle view of phonons, and can admit ballistic effects. Wave physics are

admitted by solutions of either elastic continuum [7] or lattice-dynamical equa-

tions [8, 9, 10]. These latter approaches are far more computationally intensive

and efficient computational techniques are still evolving.

From a modeling and simulation view point, multi-scale effects must be ad-

dressed in two respects. The most obvious coupling between scales occurs because

engineered nanostructures are always embedded in larger systems and the interac-

tion of scales must be accurately represented. In microelectronic devices, for ex-

ample, heat generated in hot spots in the transistor channel on the 10-100 nm scale

must be dissipated in heat sinks on the scale of tens of centimeters. In ultra-fast

laser processing, the deposition of laser energy on 10-100 nm scale causes pressure

and stress waves to propagate into the bulk material, interacting on length scales

of a micron or larger [11]. Another type of coupling between scales occurs in the

development of approximate physical models. For example, phonon confinement

effects, captured straightforwardly in molecular dynamics simulations of superlat-

tices, must be represented through modified dispersion curves in simulations based

on the Boltzmann transport equation (BTE). Multi-scale simulation in both senses

is still developing and represents a fruitful area for further research.

6
In this paper, we first present a brief review of the physics of phonon transport

in Section 2, emphasizing only those points required to illuminate the discussion

in the rest of the paper. We then present existing approaches to modeling phonon

transport in Section 3 and review the numerical methods used to solve them in

Section 4. We close with a discussion of the state of modeling and simulation in

this application area, and make recommendations for future research.

2 Review of Phonon Transport

In this section we provide a brief review of the essential elements of phonon trans-

port in crystals. A more detailed treatment may be found in [12, 13, 14, 15].

In order to illustrate the idea of phonon polarizations let us consider a 1D lattice

with two different atoms arranged alternately at a distance  apart, and with masses
+\[ +^]
and . A unit cell of atoms is denoted by the index n, and the displacements

corresponding to the two types of atoms in the cell are B and D . Assuming a

linear spring with a spring constant C connecting the atoms, the corresponding

equations of motion may be written as:

]
+ [`_ B ] a M D cb D d [feg# B O
:
_]
+g] _ D ] a M B ih [ b B eg# D O
: (1)
_

7
Assuming wave solutions of the form B a B  3 B  M e  S : 7O 3B  M  ( B=O and D a

D`QB  M e  S : O7B  M  ( DjO and substituting in Eq. 1 we obtain

]
e S +k[ Bl a D` M7m b 3 B  M e  ( nOEO eg# Bl
]
e S + ] D oa B  M7m b  B  M  ( pOEO e^# D  (2)

Since the equations are homogeneous, a solution is possible only if the determinant

is zero, i.e.,

q q
q
q # e +k[ S ] e M7m b 3 B M e ( n q
q
q    OEO q s
q q a r
q e
q MEm b 3 B M ( pOEO # e +g] S ] q
q
q   q

or
] ]
+k[X+g] Sut eg# M +\[ b +v] OS b # Mm e^wQx`y ( nO azr
(3)

]
This equation can be solved for S and yields two roots, or two phonon polariza-
M( O
tions; the corresponding dependence S is called the dispersion relation. The

region in
(
space that lies between
( a|{ $C!  is called the first Brillouin zone

(
and covers all independent values of . Correspondingly, if we represent the 1-

D lattice in K-space, the vector denoting the translation between adjacent unit

cells is called the reciprocal lattice vector. The coefficient matrix corresponding to

Eq. 2 is called the dynamical matrix.

8
The group velocity is the transmission velocity of a wave packet and is given

by

FG a _ S
( (4)
_

~} X  ! (
The phase velocity F is the local velocity of the medium and given by S .
+€[‚ƒ+g]
Dispersion curves for the case are shown in Fig. 1. We see in Fig. 1

that there are two branches or polarizations corresponding to the two solutions to

Eq. 3, called the acoustic branch and the optical branch. The group velocity of the

optical branch is seen to be small, whereas, away from the Brillouin zone edge, the

acoustic branch has non-zero group velocity.

Silicon and germanium have two atoms per unit cell, but arranged in a lat-

tice with a diamond structure [13]; this structure consists of two interpenetrating

face-centered cubic lattices. In 3-D, each of the two atoms has three degrees of

freedom, leading to six equations of motion, and a „†‡„ dynamical matrix. There

are therefore six phonon polarizations, three optical and three acoustic. The op-

tical phonons consist of one longitudinal and two transverse branches, as do the

acoustic phonon branches. In the longitudinal mode, atoms vibrate in the direction

of wave propagation whereas in the transverse mode, they vibrate perpendicular

to the direction of propagation. The specific shape of the dispersion curves de-
)
pends on the direction of the wave vector . Experimental dispersion curves in

the symmetry [111] direction for germanium are shown in Fig. 2. The two trans-

9
verse polarization branches are coincident for this direction for both optical and

acoustic phonons. Curves of this type can also be found through computation by

assuming an interatomic potential, deriving a dynamical matrix similar to Eq. 2,

and finding its eigenvalues as a function of wavenumber [16].

2.1 Phonon Distribution Function

The energy of lattice vibration is quantized, and the corresponding quantum of

energy is called a phonon. The quantum of energy corresponding to a frequency

S is S . By conceiving of phonons as particles, we may speak of a distribution

;
function for phonons. The equilibrium distribution of phonons at a temperature

is given by the Bose-Einstein distribution

m
 a
(5)
B ‰ˆ‹XŠŽjŒ ‘ e m

The total lattice energy corresponding to all phonons of all polarizations may be

obtained by integrating over all phonon modes

’ a”“ M S–O  S S


|• _ (6)

— M S–O S
Here _ is the number of states with polarization  with frequency in the
˜ M S–O
range S and S b _ S . is called the density of states and can be found from

10
M( O
the known dispersion relationship S [13]. The volumetric phonon specific heat

aš™œ› and is given by


can be found as ™

]
%'& “ B 3 B  M =
B O M S–O S
a ] _
• M B M BAO e m O (7)


where B a  ŠŽ Œ  .

In speaking of phonons as a distribution of particles characterized by different

polarizations and wave vectors (and correspondingly, frequencies), we have made

a departure from a wave description. A particle view of phonons is useful when

wave effects such as phase coherence are not important; however, it is necessary

in many applications to retain important aspects of the wave description, such as

the dispersion behavior, as well as the modeling of wave interactions through scat-

tering. A significant challenge to phonon transport modeling lies in incorporating

these aspects in a particle description.

2.2 Scattering Mechanisms

Phonons traveling in a solid may be scattered by a variety of mechanisms, including

impurities, isotopes, defects, material boundaries as well as other carriers such as

electrons or other phonons. Scattering interactions may be characterized as elastic,

where the energy (and hence frequency) of the phonon remains unchanged as a

result of the scattering event, or as inelastic, where the energy and frequency are

11
changed during the event. Typically, scattering due to dislocations, impurities, and

isotopes are modeled as elastic, and only result in a change in the direction of

travel of the phonon. A more detailed discussion of this type of scattering may be

found in [15, 17]. Another important type of scattering encountered in micro- and

nano-structures is boundary scattering, whereby the phonon is reflected at material

boundaries partly specularly and partly diffusely. Boundary scattering has been

modeled as an elastic scattering event in the literature.

Two main types of inelastic scattering processes are important for thermal
, ’
transport, normal ( ) and Umklapp ( ) processes. These occur due to scatter-

ing between phonons and are called intrinsic processes. Mathematically, they are

a result of the anharmonic terms in the interatomic potential. Three-phonon pro-

cesses can involve two phonons combining to form one, or conversely, one phonon

splitting into two. Three-phonon processes are governed by energy and momentum

conservation rules, stated as:

S b SuŸž Sf (8)

) b )  ž ) ƒ\¡j¢ xpwQ£yyE£y
(9)

) b ) Ÿž )  b o¤ ¡¥¢ x¦wQ£y5yE£y


(10)

, ’ ,
Both and processes satisfy energy conservation, Eq. 8. processes satisfy

12
’
Eq. 9, which embodies the conservation of phonon momentum. processes do
,
not conserve crystal momentum directly, but must satisfy Eq. 10. processes

do not pose a resistance to heat transfer because they conserve momentum. How-

ever, they can change the frequency distribution of phonons, and can therefore

indirectly affect other scattering processes which depend on frequency, such as im-
’
purity scattering [17]. processes offer a resistance to heat transfer because they

do not conserve crystal momentum, and must be considered in models of thermal

transport. Four-phonon processes are considered important for temperatures above


’
the Debye temperature, but are not considered here. processes are strongly tem-

perature dependent [17]. This, combined with the low population of phonons at
’
low temperatures, implies that processes are not active at low temperatures; here

impurity and boundary scattering processes dominate thermal transport.

If the scattering processes are assumed to be independent of each other, the

effective relaxation time Y , i.e., time scale for scattering, may be estimated using

Mattheissen’s Rule
m m m m
a b b bz§3§3§
Y Y › Y  Y  (11)

where Y › is the time scale for Umklapp scattering, Y  is the time scale for boundary

scattering, Y is the time scale for impurity scattering and so on.

13
2.3 Phonon Confinement Effects

The dynamical matrix in Eq. 2 and the dispersion curves shown in Fig. 2 are for a

bulk crystal, i.e., a periodic crystal in which boundaries do not play a role. Phonon

transport properties such as dispersion curves, group velocity, and density of states,

are strongly modified by limiting the spatial size of the domain under considera-

tion [7, 16]. For example, dispersion curves for a free-standing silicon layer were

computed in [18] using the EDIP potential [19] with only a six atomic layers. By

applying free-standing boundary conditions in the direction normal to the layer

and considering wave vectors in the [100] direction, the dynamical matrix and its
(
eigenvalues were computed as a function of the wavenumber . Fig. 3 shows the

computed dispersion curves. We see that for six atomic layers, with two atoms

per unit cell, thirty six dispersion curves are obtained corresponding to the thirty

six degrees of freedom for this diamond lattice. The dispersion curves no longer

pass through the origin, and no longer exhibit the linear w K̃ behavior at low wave

vector. There is a cut-off frequency below which phonons do not exist. These mod-

ifications substantially change the density of states, and reduce the dimensionality

of the dependence of low-temperature specific heat on temperature. Since bound-

ary scattering dominates at low temperatures, and the boundary scattering time

scale does not depend on temperature, the dimensionality of the kT̃ dependence

drops with , in keeping with kinetic theory [13]. Since bulk thermal conductiv-

14
]
F G
ity is proportional to , flattening of dispersion curves and a reduction of group

velocity severely reduces in-plane thermal conductivity.

3 Phonon Transport Models

The fundamental challenge in devising phonon transport models is to capture the

complexities described in Section 2. We now described models for thermal trans-

port in semiconductors and dielectrics which approximate phonon behavior to var-

ious degrees.

3.1 Fourier Conduction and Variants

The simplest model for thermal transport in semiconductors and dielectrics is the

classic Fourier conduction equation

;
‰¨ § % © ;
: as© (12)
¨

This equation is a parabolic differential equation in unsteady state, and is an el-

liptic partial differential equation in steady state. It is possible to show that the

Fourier heat conduction equation approximates phonon behavior in the acousti-

cally thick limit, i.e., when there are sufficient phonon scattering events that the
M * ! FXG Y O
acoustic thickness is large; here Y is time scale on which scattering events

15
:ǻ Y . The
occur. Furthermore, temporal phenomena must occur on time scales

primary drawback of this equation is that since phonons travel with a finite group

velocity, information transfer is not instantaneous; however, Eq. 12 has infinite

wave speed. In steady state, Eq. 12 yields erroneous results for the acoustically

thin limit, where ballistic phonon transport effects are important. A discussion of

these issues has been presented in [3, 4].

A variety of hyperbolic variants of the basic Fourier equation have been sug-

gested [20] to impart finite wave speed. One popular form [21] adds a hyperbolic

time term to Eq. 12:


]
; ;
¨ ¨ § % © ;
Y :] b : a”© (13)
¨ ¨

The value of Y governs the lagging behavior, and gives the equation a finite wave

speed. However, for steady state problems, the Fourier equation is recovered, with

attendant difficulties in capturing the ballistic transport limit. It is also not straight-

forward to recover the frequency-dependent lag-time behavior resulting from a

frequency-dependent group velocity.

3.2 Models Based on Boltzmann Transport Equation (BTE)

When phase coherence effects are unimportant, a particle view of phonon trans-

port may be adopted and the semi-classical phonon Boltzmann transport equa-

tion may be used to describe the transport of a distribution function of phonons

16
 M9 ) :O
[15]:

 
¨
: b © § M IAG  O b\¬€§ ©®­H¯  a±° ¨ A : ²‚³µ´·¶J¸¹¸µº¼»¾½À¿QÁ (14)
¨ ¨

! :
Here ¬ is the acceleration _ I¦G _ and is zero in the absence of external forces. The

left hand side then represents the advective transport of the distribution function

due to the phonon group velocity vector ~I G . The right hand side represents the

rate of change of due to scattering to other wave vectors and polarizations. The

phonon distribution function is in general a function of position, wavenumber (or

frequency), direction and polarization, as well as time. This dependence makes

the equation very expensive to solve, and a number of approximations have been

developed to make it tractable. These are described in the sections that follow.

3.2.1 Relaxation Time Approximation

Because of the complexities in the scattering term, it is common to make the relax-

ation time approximation [15]

 J M S ; O e  M 9  )  : O
¨
: b © § M IAG  O a
¨ Y J?~? (15)


Here, is the equilibrium Bose-Einstein distribution function, Eq. 5, and is a
J??
function of frequency and temperature but not direction. The time scale Y is the

17
effective relaxation time associated with all scattering processes under considera-

tion and would typically be found using Mattheissen’s rule, Eq. 11. A scattering

term of this type acts to smooth the angular variation in the phonon distribution

function at any particular frequency. Indeed, in the absence of the LHS of Eq. 15,

a scattering term of the type Eq. 15 would lead to a directionally independent

at each frequency S . However, it would not permit phonons to scatter to other

frequencies.

Elastic scattering processes do not involve a change of phonon frequency and

are easily modeled using the relaxation time form of the scattering term. However,
, ’
and processes, which involve anharmonicities, are not easily modeled in this
, ’
way unless approximations are made. and processes must satisfy energy and

momentum conservation rules, Eqns. (8-10). Using a perturbed Hamiltonian, a

general expression for the three-phonon scattering term was obtained in [17] for

processes satisfying Eqns. (8-10)

 ]
° ¨ : ² ³µ´·¶J¸¹¸µº¼»µ½À¿œÁ a “ #jÆ ÇQÈ˜É )  ) µ ) `Ê Æ È $ L É S b Su e Sf`Ê
P ÂÄÃ PNÅ
N + SuS  Sf
¨
É M  b m O É·  b m Ê   e A  M   b m O Ê
(16)

where
] + ] ]
] R
Æ ÇœÈ‚É )  ) ¾ ) `Ê Æ aÌË ] ] É u S u
S Î
 f
S `
 Ê
Í  F (17)

18
   
Here, , and  are the values of the distribution function associated with
) )  )
phonon wave vectors , and  , and S , S  , and Sf are the corresponding

frequencies dictated by the dispersion curves. is the number of atoms per unit

volume, R is the Gruneisen constant. The velocity F is the sound velocity.

To cast Eq. 16 into the relaxation time form (Eq. 15), it is necessary to assume
  
the only the mode is perturbed about its equilibrium value, but that  and 

are at equilibrium. Under these conditions, it is possible to derive a relaxation-


, ’
time form for and process scattering terms, corresponding to a single mode
’
relaxation time. For processes, for example, the scattering term reduces to [17]

 ]
° ¨ : ²˜³¾´¶J¸¹¸¾º·»¾½À¿QÁ a M   e  O “ #jÆ Ç È˜É )  )   )  Ê Æ
+ È
¨ PNÂÏà PNÅ SuS  Sf
$ L É S b S  e Sf Ê L P h PN d PNÅ dlРɼ   e   J Ê
(18)

LÑM S–O L P h P  d PNÅ dlÐ


Here is the Dirac delta function and is the Kronecker delta and

is zero unless
) b )  e )  a 
. By comparing Eq. 18 with Eq. 15, a relaxation
’
time for processes may be derived as

m $ R ]
M S–O a ÒÍ  + ] “ SuS  Sf L É S b S  e Sf Ê L P h P  d PNÅ dlÐ É  J  e   J  Ê
Y F N P Â
›
(19)

It is important to understand the limitations of the relaxation time form, as

19
compared to a BTE with more complex scattering expression such as in Eq. 16.

The relaxation time form postulates that in the absence of ballistic terms (i.e., the

LHS of Eq. 15), the phonon distribution at any wave vector should relax to the

equilibrium distribution at the same frequency. This form does not allow direct

energy exchange across frequencies, whereas a scattering expression such as that

in Eq. 16 explicitly does. The relaxation time form indirectly accounts for the

effect of energy and momentum conservation rules in determining Y › , but does

not explicitly enforce the rules themselves. Anharmonic interactions appear only

indirectly through the fact that all phonons groups share the same “temperature” in

the Bose-Einstein distribution.

These features of the relaxation time form do not have deleterious consequences

for the computation of thermal conductivity, where it is possible to show that in the

acoustically thick limit, the relaxation time form of the BTE will yield a thermal

conductivity commensurate with kinetic theory [15]:

m ]
% a NP F G M ) O Y 7?? M ) O )
Í • _ (20)

–P )
Here is the specific heat for phonons of wave vector , F·G is the corresponding
7??
group velocity, and Y is the effective relaxation time. However, in applications

such as thermal transport in high-energy field effect transistors (FETs) for exam-

20
ple, a significant fraction of the energy associated with self heating originates in

electron-phonon scattering to optical phonon modes, and is then scattered from

optical to acoustic modes. Thus anharmonic scattering processes are critical for

capturing this energy cascade. A variety of models have emerged which attempt

to explicitly allow for anharmonic interactions without taking on the complexity of

Eq. 16. These are described below.

3.2.2 Gray Model

The simplest approximation to Eq. 14 is to consider phonons of all polarizations

and frequencies as the same. It is convenient to write a BTE in terms of the inte-
M9 8 :O
grated energy density ~ , which is a function of spatial position 9 , direction

8 , and time : , but not of frequency or polarization:

¨ ~  b ©   e  
: § M FHG 8 'O a b 1325476
¨ Y (21)

Here,

~ a “ °  S M S–O
• _ S ²

 m
 a $ •  _ T (22)
Ë tEÓ

21
The gray BTE employs a single phonon velocity F G in all directions, and a single

relaxation time Y . The generation term 1·25476 may be used to model energy scattered

from electrons to phonons, for example; it would be zero in most applications.

Though different choices of FG and Y are possible, it is convenient to choose FÔG

to be an average velocity of sound in the medium, and to find Y so that the bulk
[
% a È F G] Y
thermal conductivity is matched using the kinetic theory expression .

The gray model does not incorporate dispersion curves or the specific mechanisms

for phonon scattering in any detail, but is relatively inexpensive to compute. It

recovers the parabolic Fourier conduction equation in the limit of large acoustic
* ! F3G Y :0ª Y
thickness for time scales and predicts a bulk thermal conductivity
[ ]
% a È F G Y
. Majumdar and co-workers [3, 4] developed a frequency-dependant

version of the BTE in the relaxation-time approximation called the equation for

phonon radiative transfer (EPRT), accounting for frequency-dependent relaxation

times, but not for polarization or phonon conservation rules.

3.2.3 Semi-Gray Model

The next level of sophistication is to distinguish to some degree among phonon

groups by separating propagating and reservoir modes, following [22, 23]. The

reservoir mode is considered purely capacitative and has zero group velocity. It is
M :
modeled as isotropic, and only has a dependence on 9  O . The propagating mode,

22
M :
on the other hand, is responsible for phonon transport and depends on 9  8  O . The
M9 8 :O
governing equations written in terms of the energy density ~ are [23]:

[ M ;A< e ;=> J ? O e  
¨ ~ b © § M FG 8 ~ O a Et Ó
: Y
¨
;A@ M
@ ;A< e ;=> J? O e @ MÕ ; @ e ;"> 7 ? O
@ ¨ a b 1325476 (23)
: Y
¨

<
Here, represents the specific heat of all phonons grouped into the propagating
@
mode, and represents the specific heat of all phonons grouped into the reservoir
; @ ; <
mode. The reservoir mode and lattice temperature and represent the energies

contained in the reservoir mode and the total energy respectively. The term 1 25476
;> 7?
may be used to model electron-phonon scattering to optical modes. is the

reference temperature.

Energy exchange between the reservoir and propagating modes is explicitly

included through scattering terms on the RHS of Eq. 23 which are modeled on

the relaxation time approximation on the RHS of Eq. 15. In the acoustically thick
[ ]
% È
limit, the model can be shown to yield a thermal conductivity a F G Y , and

b @
its unsteady response corresponds to a total volumetric specific heat of .

The semi-gray model captures some of the complexities of the phonon dispersion

curves at a relatively low cost. Since the reservoir mode equation does not involve

direction, it is relatively inexpensive to compute, while the cost of the propagating

23
mode computation is similar to that of the gray model.

The main approximation made in the semigray model is that of a single relax-

ation time and a single phonon velocity. In [23] the propagating mode is assumed

mainly to consist of the longitudinal acoustic modes, and the reservoir mode ac-

counts for optical and transverse modes, which are assumed to have low enough

velocities so as to be considered stationary. The longitudinal mode velocity corre-

sponding to the Brillouin zone edge is chosen as F G [23]. To recover the correct
[ ]
% È
bulk thermal conductivity, it is necessary to choose Y from a F G Y . However,

a value of Y computed in this way is typically far larger than typical optical-to-

acoustic relaxation times, leading to high temperature predictions in typical FET

simulations [24]. In unsteady problems such as those involving electrostatic dis-

charge events in FETs, the choice of which phonons constitute reservoir and prop-
@
agating modes determines the values of and , and is critical for determining

the thermal response time of the device.

3.2.4 Non-Gray Model

In order to incorporate more of the physics contained in the dispersion curves, and

to resolve differences in the rates of energy exchange between different phonon

polarizations and wave vectors, a non-gray model was proposed in [6]. Here, the
M ,×Ö 'Ø e m O U 
acoustic branches are divided into _¦Ù frequency bands of extent S

24

centered about S . The number and extent of the bands are chosen to resolve the

dispersion curves adequately. A BTE is written for the energy density associated

with each band in each angular direction. Since the optical branches have low

group velocity, they are grouped into a single reservoir mode, similar to the semi-
,×Ö Ø
gray model, corresponding to band number _ ; the optical mode energy is

assumed not to have a directional dependence. Each band  is characterized by a


 
separate group velocity F (zero for the optical band), a band specific heat and
;¦
a band temperature . Isotropy is assumed, and the dispersion curves in the [100]
 
direction of silicon are used in [6] for the computation of F and . The governing

equations take the form:

  ™ m
¨ ~  b © § M F  8   O a É   e ~   Ê R Ú  s
b Û “ ° $  M ;" e ;"> 7 ? O e ~  R 
: œÜ [ à HÜ"Ý  Ë ²
¨

(24)
 d [
™ m
¨  4 a Û “ ° 4 M ; 4  e ;=> J ? O e  4 ² R 4  b 1325476
: œ Ü [ $ (25)
¨ Ë


Here, R are the inverse band-averaged relaxation times associated with the energy

transfer between bands  and  . The energy density terms are defined as:

  a  S M S–O S
•¦Þ _
 m `Œ ß
  a $ •   _ T
Ë tEÓ

25
;  ; 
Definitions of the band temperature and the interaction temperature may be

found in [6].

The non-gray model incorporates two types of scattering terms. The first term

on the RHS of Eq. 24 models elastic scattering events which do not lead to a change

in frequency. The second term on the RHS models anharmonic energy exchanges

between different bands. These interactions are governed by the inverse relaxation
 
times R which are computed using [25]. The term 1·25476 is an energy source term

and may be used to model electron-phonon scattering to the optical mode. Bulk

thermal conductivity of silicon as a function of temperature has been computed us-

ing the non-gray BTE model, and shown in Fig. 4. The match is reasonable, though

departures are seen at the maximum. The location of the maximum denotes the

point at which boundary and impurity scattering begin to dominate three-phonon

scattering processes. The computations employ a relatively simple isotope scat-

tering model which may account for the overprediction. Thermal conductivity of

undoped silicon computed using the model is shown in Fig. 5. A reasonable match

with experimental data from [26, 27, 28] is obtained.

A discussion of the properties of the non-gray model is in order. The model

is an attempt to incorporate more of the wave physics contained in the dispersion

curves into a particle view of phonon transport. It can be shown to default to

Fourier conduction for large acoustic thickness, corresponding to a bulk thermal

26
[ ] [
% aáà  È  F  Y 7??3Ã  aæà  R   , and a volumetric
conductivity of where â·ãÔää7å
ß
a à  
specific heat . In this limit, it recovers the Fourier heat flux exactly. The

model conserves energy exactly in that the scattering terms in Eqns. 24 and 25 add

up to zero when summed over all bands. It allows for more granularity in resolving

the interactions of different phonon groups through the inverse relaxation times R .

However, the energy and momentum conservation rules governing three-phonon

processes are only indirectly and approximately enforced through the relaxation

time calculation. In the absence of equilibrium, the temperatures computed by the

model must be interpreted as representative of energy. Thus the band temperature


;"
is representative of the energy contained in band  , and the overall temperature
M  $   O
is representative of the total phonon energy a à Ë .

3.3 Comparison of BTE Models

The models described above are applied to the problem of thermal prediction in a

silicon-on-insulator FET in [24]. Computations were done using a finite volume

scheme (see Section 4.1 for details). The calculation domain is shown in Fig. 6.

Heat is generated in a hot spot located in the channel region of a SOI transistor.

The transistor is fabricated on a thin silicon device layer located on top of a silicon
]
dioxide (SiO ) buried oxide layer intended to provide electrical isolation. The

lateral and bottom boundaries are held at 300 K. A heat source of 0.6 mW/ç m is

27
applied in the hot spot and the temperature field in the device is predicted using
]
Fourier computations in the SiO region and the BTE in the device layer. Other

details are available in [24]. Contours of the predicted temperature are shown in

Fig. 7 using the Fourier and BTE models. The maximum temperatures predicted

by the Fourier, gray, semi-gray and non-gray models are 320.7 K, 326.4 K, 504.9

K and 393.1K respectively. The difference in temperature prediction between the

models is attributed to the time scale for energy scattering from the optical modes,

which receive energy from electron-phonon scattering, to the acoustic modes which

transport it. In the Fourier and gray models, the propagating modes essentially see

the heat source directly, unmodulated by a scattering process. In both the semi-gray

and non-gray models, the rate of energy transfer from the optical (reservoir) mode

to the acoustic (propagating) mode is governed by a non-zero relaxation time. In

the case of the semi-gray model, Y is 72 ps. The non-gray model has a spectrum

of relaxation times for the various phonon interactions, but the shortest relaxation

times governing optical to acoustic interactions is about 21 ps. Because of the

bottleneck in energy transfer between optical and acoustic modes, the semi-gray

model predicts far higher temperatures than any of the other models. The non-gray

model also shows that the the optical modes scatter 1¹25476 preferentially to the LA

and TA modes close to the Brillouin zone edge, which then transport the energy to

the boundaries. LA modes are responsible for about 70% of the heat transfer rate

28
leaving the device layer. Details may be found in [24].

3.4 Relaxation Time Calculation

A critical component of simulations based on the BTE is the computation of relax-

ation time. For the most part, relatively simple approaches involving curve fitting

have been used. Relaxation times with adjustable parameters tuned to experimen-

tal data have been used in bulk thermal conductivity calculations in [26, 29, 30]

to model various scattering processes. Thermal conductivity calculations in [23]


[
M ; OB¥ê M; O
employed a phonon-phonon relaxation time of the form â aéè and è

and ë were found from curve fits with experimental data. For Umklapp scattering,

Majumdar [3] employed a form from [17]

m ; Vî
aíì V 3 B  ° ; ²
Y S R (26)
› ™

and found the constants ì and R by fitting to thermal conductivity data.

Klemens and co-workers [17, 25] derived relaxation rates for impurity and

three-phonon processes starting from perturbation theory. These expressions em-

ployed no arbitrary curve fits, and had only one adjustable parameter, the Gruneisen

constant. However, their most general form requires careful implementation of en-

ergy and momentum conservation rules. For low temperatures, they postulated a
,
two-step process by which low-frequency phonons undergo an process to in-

29
’
termediate frequencies, and then processes which allow them to interact with

zone-boundary phonons. These simplifications allowed the imposition of approx-

imate energy and momentum conservation rules. These intermediate-frequency


’
-process relaxation rates were used to make thermal conductivity predictions in

thin layers and FETs in [6, 24] and in quantum-well structures in [7] at room tem-

perature.

One of the problems in using these types of approximations is the difficulty of

extending them to situations in which dispersion curves are strongly modified by

confinement. For such cases, it is necessary to develop more fundamental ways of

computing relaxation times. A first attempt using Eq. 19 was recently presented

in [31]. Here, the wave-vector space was discretized into wavenumber and solid-
LNM S b S  e Sf'O L P h P  d PNÅ dlÐ
angle pixels, and the roots of the function found
)
by bisection for each discrete , thereby enforcing energy and momentum conser-

vation rigorously. Single-mode relaxation times as function of wavenumber and

polarization were computed for Umklapp processes using the dispersion curves in

the [100] direction for silicon. A good match with bulk and thin-layer thermal con-

ductivity for silicon was reported. Application to dispersion curves resulting from

confined structures has however not yet been attempted.

Once a methodology for rigorously enforcing conservation rules is available,

the possibility of discarding the relaxation time approximation may be entertained.

30
The procedures outlined in [31] may be used to directly compute the scattering term

on the RHS in Eq. 16, combined with non-gray BTE calculations. Such computa-

tions would be very intensive, and may not be warranted for simple thermal con-

ductivity computations. However, for thermal transport situations in which phonon

modes are strongly out of equilibrium, such computations could yield fundamental

insights into phonon behavior not hitherto available.

3.5 Ballistic-Diffusive Approximation to BTE

Since BTE-based models are expensive to use, a ballistic-diffusive approximation

to the gray unsteady BTE was proposed in [32] which aimed to reduce the cost

of simulation while retaining the physics of ballistic phonon transport. The ap-

proximation is identical to the modified differential approximation (MDA) in the

thermal radiation literature [33] in steady state, but incorporates unsteady transport

terms as well. The phonon energy was decomposed into two parts: the ballistic or

boundary component (b), and a component associated with the medium (m). The

former component was computed using a ray-tracing procedure for which exact

solutions are possible in simple geometries. For the medium component, a diffuse

approximation was made, resulting in a hyperbolic heat conduction-like equation,

albeit with extra fluxes due to the ballistic component. An extension to incorporate

dispersion effects was made in [34]. A good match with the exact solution for the

31
parallel plate problem [35] was found in [34] for acoustically thin problems due

to the accuracy of the ray tracing procedure, but departures from the exact solu-

tion were found near boundaries for larger acoustic thicknesses due to the failure

of the diffusion approximation in the medium component near boundaries. These

departures point to difficulties in using this class of models for thin layers and

other micro-structures where boundaries dominate, and further development and

validation is required.

3.6 Molecular Dynamics Models

Classical molecular dynamics simulations of thermal transport in bulk semiconduc-

tors as well as in superlattice structures are increasingly being published [8, 9, 36].

A system of N atoms of mass M and position 9 is considered and Newton’s second

law of motion is written for each atom:

]

+ _ 9”
]: a “ Û 
 Ã Ü"Ý  (27)
_

The inter-atomic forces are generally modeled through empirical potentials. A

general background on molecular dynamics and related computational techniques

may be found in [37].

Many published studies in the thermal transport arena have employed Lennard-

Jones (LJ) potentials [9] which are simple to implement and yield useful insights,

32
but do not necessarily yield quantitative information for many materials of engi-

neering interest. For silicon, three-body potentials such as the Stillinger-Weber [38]

or Tersoff potential [39] have been used. The Stillinger-Weber potential was used

in [8] to compute thermal conductivity of bulk silicon using a periodic domain. The

Green-Kubo formalism was employed to find thermal conductivity. Low-frequency

cut-off due to finite system size as well as artificial correlations due to periodic

boundaries were found to strongly affect bulk predictions. An extrapolation proce-

dure was used to correct for these computational artifacts.

Though most molecular dynamics simulations are still limited to tens of nanome-

ters in terms of domain size, important insights can be gained from these computa-

tions. For strong departures from equilibrium, such as those occurring in high-

energy FETs, relaxation times based on perturbation approaches such as those

in [17] are themselves questionable. Single mode relaxation times have been com-

puted from the time-decay of the total energy of atomic ensembles using molecu-

lar dynamics in [36, 40], and could perhaps be used to gain an understanding of

phonon interactions in these strongly non-equilibrium situations. Recently, trans-

port at interfaces has been studied in [41, 42] using phonon wave packets with a

well-defined frequency and polarization. These approaches are particularly useful

when accompanied by comparisons to BTE simulations and experiments.

33
4 Numerical Methods

4.1 Finite Volume Methods for BTE

We now turn to the consideration of computational techniques for solving the the

governing equations resulting from the models described in Section 3. The Fourier

conduction model and its hyperbolic variants are easily solved using a host of con-

ventional finite volume, finite element and finite difference techniques, and are not

considered further. At the other end of the computational spectrum, there are well-

established explicit methods for addressing classical molecular dynamics and the

integration of Eq. 27 [37]. We turn now to numerical techniques for the solution of

the BTE.

4.1.1 Finite Volume Scheme

The steady gray BTE in the relaxation time approximation is identical in form to

the gray radiative transfer equation (RTE) in an isotropically scattering medium.

The discrete ordinates method [43] has long been used to solve neutron transport

problems, and has been adapted to solve the RTE. A discrete ordinates approach

was used in [5] to simulate thermal fields in FETs using a gray model. In the

early 1990’s, Raithby and Chui [44] developed a finite volume method for radiative

transport which is similar in flavor to the discrete ordinates technique, but which

34
explicitly enforces radiative energy balances over control volumes. The extension

of this technique to unstructured meshes were published in [45] and applied to the

BTE in [46].

Both the discrete ordinates and finite volume techniques start with a discretiza-

tion of space, time and angle, and for non-gray BTE models, frequency or wavenum-

ber. The finite volume scheme discretizes space into arbitrary unstructured convex
$
polyhedra, and the angular space Ë into discrete non-overlapping solid angles
U T  
each centered about the discrete direction 8 . Each octant is discretized into
,.- ,0/
solid angles. The governing equation is integrated over the control vol-

ume, control angle, frequency band, and time step, yielding a balance of phonon

energy fluxes in the direction  through the control volume faces with the stor-

age and scattering terms. The LHS of Eq. 24 is the linear wave operator in the

direction 8 , and has long been studied in computational fluid dynamics. A vari-

ety of higher-order schemes [47, 48, 49] may be used to interpolate the advection

operator. Second-order implicit formulations are used for discretizing the time-

dependent and scattering terms. The resulting discretization leads to a set of alge-
 Ã
braic equations relating    , the energy density in the direction  in frequency
% r , to its spatial neighbors ~   à  :
band at the cell centroid

~   à  a”“ '     à  b Ö (28)




35
Here,   is the center coefficient, and Q  denotes the set of neighbor coefficients
 Ã r to its spatial neighbors in the same angular
connecting the value of ~   at cell

direction and frequency band. Angular neighbors, implicit in the  term in Eq. 24,
Ö
are included explicitly in , as are the anharmonic scattering terms. The basic

solution procedure solves the discrete equation set in each angular direction and

frequency band sequentially, using an algebraic multigrid scheme [50]. Picard

iteration is used to evaluate terms involving other directions and bands. Details of

the technique as well as validation against the literature may be found in [46].

4.1.2 Improvements to Accuracy

Inadequate spatial resolution in the finite volume and discrete ordinates schemes

leads to the phenomenon of false scattering whereby spatial smearing of the phonon

energy distribution is caused purely by numerical diffusion; this type of error has

been extensively investigated in the thermal radiation literature [51]. In addition

to spatial error, angular discretization leads to another type of error, the ray effect.

Since the angular space is divided into finite control angles, small scale spatial vari-

ations in phonon energy density are in effect averaged over the control angle. This

error is especially noticeable at low acoustic thicknesses in problems with varia-

tions in boundary values. It may be alleviated by adequate resolution of the angular

domain. Coelho [52] has shown how angular and spatial discretization errors tend

36
to compensate each other, so that coarse discretizations of both sometimes yield

more accurate solutions than fine discretizations of either space or angle individ-

ually. For the unsteady BTE, the ray effect results in additional errors because of

finite wave speed which do not occur in conventional thermal radiation problems.

The phonon travel time from the boundary to the interior is inadequately resolved

for coarse angular discretizations. Thus, temperature predictions, which represent

angular averages of the phonon energy density, are found to exhibit non-physical

spatial wiggles [53]. This occurs even when individual directional phonon energy

densities are accurately computed.

Fig. 8 shows unsteady temperature profiles computed using a gray BTE for the
;ðï añr
case of a 1D domain initially at temperature which has its left boundary
;‚ï a m : ï aòr
raised to at time . The acoustic thickness is zero. The curves

labeled “first order” employ an upwind difference scheme for spatial discretization

and a first-order implicit time stepping scheme. Those labeled “second order”

employ a bounded QUICK scheme with a min-mod limiter [49] and a second-

order backward Euler scheme. Two different angular discretizations, Ë Ë and

Ò Ò in the octant are considered. Though the intensities in individual directions

are shown to be monotonic [53], their summation to compute temperature is seen

to produce unphysical spatial wiggles because of phase errors. This phenomenon

persists even for relatively fine angular discretizations if the acoustic thickness is

37
small, as it frequently would be in micro- and nano-scale domains.

A remedy combining ray tracing and the finite volume scheme was suggested

in [53]. Here, the phonon energy density is decomposed into a boundary and a

medium component, similar to the ballistic-diffusive idea of Chen [32]. However,

the diffuse approximation is not made for the medium component; instead the finite

volume scheme is used to solve the modified equations. The boundary component

is solved by ray tracing accounting for time delay due to the finite phonon velocity.

Each control angle is pixelated and rays are traced through the centroid of each
# r # r
pixel. The results for a Ë Ë discretization with pixelation of each control

angle for ray tracing is shown in Fig. 9 for zero acoustic thickness. Substantial

accuracy improvements are found across all acoustic thicknesses if sufficient ray

tracing resolution is used, with only a modest penalty in computation time.

4.1.3 Improvements to Convergence Speed

The sequential solution procedure performs well for moderate acoustic thicknesses

when the primary influences determining the phonon energy density in a cell are

due to its spatial neighbors. For large acoustic thicknesses, though, the influence of

angular neighbors (through the elastic scattering terms) and other frequency bands

(through anharmonic scattering terms) becomes dominant. Because the sequential

scheme only accounts for spatial neighbors in the coefficient matrix, and relegates

38
Ö
angular and frequency influences to the term in Eq. 28, convergence speed slows

down substantially for large acoustic thicknesses. In the non-gray BTE model,

there exists a large distribution of acoustic thicknesses across frequency bands. In

thin-layer problems where the in-plane direction is long, high-frequency bands can
* ! FEG Y
have acoustic thicknesses well in excess of 100 [24]. To remedy this type

of problem, a coupled ordinate method (COMET) has been published in [54] in

the thermal radiation context, but which applies equally well to the phonon BTE.

Here, instead of solving the RTE for each direction sequentially, the discrete radia-

tion intensities at a cell in all directions are solved simultaneously, assuming spatial

neighbors to be temporarily known. This relaxation sweep is embedded in a geo-

metric multigrid scheme to accelerate convergence. The results for gray isotropic
; aór and the others at ; a m ririr (
scattering in a square cavity with one wall at

are shown in Table 1, and would apply to the gray BTE in the relaxation time ap-

proximation. Acceleration in CPU times of several orders of magnitude are found

for large acoustic thicknesses, with no penalty at low acoustic thicknesses. Though

the method in [54] is directly applicable to the gray phonon BTE, extensions to

the point-coupled procedure are necessary to account for the changed point-matrix

structure for the non-gray model or for a direct solution of Eq. 14. The primary

expense of the technique lies in the direct solution procedure at each cell, and

therefore depends on the structure of scattering terms.

39
4.2 Monte Carlo Techniques

Though the Monte Carlo simulation technique has been extensively used for ther-

mal radiation computations [33] as well as in the simulation of electron transport

in microelectronic devices [55], its use in the simulation of phonon transport is

relatively recent [56, 57]. Phonon transport in a Debye crystal was computed

in [56] assuming no dispersion and a constant (single) relaxation time. Longi-

tudinal and transverse acoustic branches and the corresponding dispersion curves

were included in [57], but not optical branches.

The starting point for the Monte Carlo technique is Eq. 14, and as such, it

should be regarded as an alternative to the finite volume or discrete ordinates

methods for solving Eq. 14. In the Monte Carlo technique, phonon bundles or

samples are drawn from the position and wave vector spaces. Each bundle repre-

sents a number of like phonons determined by a scaling factor chosen by the user.

Phonons are placed randomly in the domain, and their polarizations and frequen-

cies are determined from the Bose-Einstein distribution. Phonon wavenumbers

are then computed from dispersion curves. The domain is also divided into spa-

tial bins or cells for accounting purposes. Phonon transport is computed in two
U :
phases. In the flight or drift phase, the phonon is displaced by a distance I G .

Then, in the scattering phase, the phonon is scattered in the following way. The
, ’
relaxation time for and scattering, Y › , is computed, and correspondingly, a
Û

40
m e 3 B M e U : ! Y
probability of scattering  a  O
› . A random number between 0
Û
and 1 is drawn, and if the random number is less than  , a scattering event is as-

sumed to happen. Impurity scattering is handled in a similar manner by computing

a separate probability. Once a scattering event happens, the phonon frequency is

changed for anharmonic processes by drawing a new sample from a local Bose-

Einstein distribution based on a local pseudo-temperature computed from phonon

energies. This pseudo-equilibration step is not necessary if the scattering process

being considered is elastic.

Energy and momentum conservation rules are not directly imposed for each

scattering interaction in such a Monte Carlo technique. Energy conservation is im-

posed in an approximate way in the following sense. Before the scattering event,

a target energy is computed in each cell. After the scattering event, phonons are

added or removed from the cell to ensure that the target energy is not changed.

However, momentum conservation is not explicitly imposed, and it is unclear in

what sense Eqns. 9 and 10 are satisfied. In this sense, the Monte Carlo technique is

similar to other BTE simulations in that conservation rules appear only indirectly

through the determination of relaxation times. Anharmonic energy transfer be-

tween modes occurs in the Monte Carlo method through the pseudo-equilibration

step. Hot phonons essentially lead to a higher pseudo-temperature and appropri-

ately shift the profile of phonons being generated for the new time step.

41
The method has been successfully used to predict thermal conductivity in thin

doped and undoped silicon layers [57], but has not been used for strongly non-

equilibrium thermal computations, such as those in FETs. The advantage of the

Monte Carlo method is its flexibility and relative simplicity in programming. The

method is inherently parallel, making it suitable for large-scale simulation. Also,

it meshes easily with electron transport solvers which also employ a Monte Carlo

approach [58], making it conceivable that concurrent electro-thermal simulations

of FETs could be done within a single computational framework. However, Monte

Carlo computations also become intensive when acoustic thickness is large, as re-

sult of frequent scattering events. Furthermore, in multi-scale simulations, such

as those in FETs, Monte Carlo simulations of thermal transport in the device layer

would have to be coupled to Fourier simulations in the rest of the device. It remains

to be seen how the stochastic jitter in Monte Carlo simulations affects Fourier com-

putations, and how well overall energy is conserved.

4.3 Multi-Scale Simulation

Today classical molecular dynamics (MD) simulations are being applied to a wide

variety of problems not only in thermal transport, but also in fracture mechanics,

plasticity and other areas. However, the physical size of the sample as well as the

time for simulation is still very limited. Typically, domain sizes are restricted to

42
tens of nanometers, and time scales to only a few hundred picoseconds even on

large computers if the more complicated potentials are used. As a result, efforts

are underway to only use molecular dynamics in the regions where it is absolutely

necessary, and to couple it to continuum simulations using finite volume or finite

element schemes elsewhere. This type of approach is useful in ultra-fast laser

ablation, for example, where the ablation region undergoes complex phase change

from solid to liquid to vapor, and is all but impossible to model using continuum

techniques. However, the outer boundary conditions for the MD simulations must

come from the macro-scale material in which it is embedded; the temperature of

the macro-scale material is, in turn, coupled to the MD region. Unless the coupling

is done carefully, wave reflection at the artificial boundaries as well as a lack of

energy conservation may result.

Though there has been little published work in the thermal transport field de-

scribing coupled simulation techniques, a number of approaches have been used

in the solid mechanics literature to couple MD with finite element methods for

predicting solid deformation. The simplest method is to employ a finite element

mesh which is coincident with atomic positions at the boundaries of macro- and

microscale regions, and to use the nodal displacements from the MD computations

as boundary conditions on the FEM simulations, and vice versa. Such as approach

has been taken in [11] for laser ablation simulation. In the multi-scale method

43
of [59, 60], the finite element mesh is graded to the MD mesh in a handshake re-

gion. A handshake Hamiltonian is postulated which consists of contributions from

the MD potential function as well from the elastic solid, and the rest of the de-

velopment flows naturally follows a typical finite element formulation. This type

of method has been used to couple classical MD and finite element methods with

tight-binding MD, and has been demonstrated in fracture computations. Another

method of this type is the quasi-continuum technique [61, 62], where atomic de-

grees of freedom are removed by interpolating from a sub-set of atoms. This type

of atom removal can be made adaptive, so that in regions of large deformation, a

full MD simulation is done. The method was originally formulated for 0 K simula-

tions, but extensions have recently proposed to relax this restriction [63]. Methods

which grade the finite element mesh down to the MD mesh are problematic for two

reasons: First, they are computationally expensive, and second, the overall time

step restriction for explicit schemes is governed by the smallest mesh scale in the

system, that of the MD region.

Liu and coworkers [64] overcame this time step restriction by using a bridging

scale decomposition. Here, the displacement is decomposed into coarse and fine

scales. The coarse scale is represented everywhere in the domain, including the

MD region, and the displacement at the macro-scale finite element nodes in the

MD region is computed as a least-squares fit to the displacements from the MD

44
simulation. Similarly, the effect of the outer continuum region is to apply time-

dependent forces on the near-boundary atoms. This type of multi-scale method

can also naturally take advantage of mesh-free finite element techniques for the

continuum.

The development of multi-scale methods coupling MD and Fourier regions are

complicated by the fact the MD simulations solve for displacement, and not energy

or temperature directly; indeed temperature is related to an ensemble average of the

kinetic energies of the atoms. However, the macro-scale equations involve conser-

vation of energy, and for conduction heat transfer, are typically written in terms

of temperature. It may be possible to solve the MD problem using displacement

formulations similar to [64] and then develop a solution to the energy equation

using these displacements either to compute temperature in the MD region or to

compute an energy flux into the continuum region. For thermal transport, these

types of approaches are very much in their infancy, and represent an exciting new

area for computational research.

5 Discussion and Closure

In this paper, we have reviewed the basics of phonon transport in solids, current

modeling efforts, as well as numerical methods for solving the resulting equations.

Molecular dynamics represents the most direct way to translate classical spring-

45
mass models of crystal lattices into numerical simulations. However, for many

practical applications, such as those involving silicon for example, complex three-

body potentials make it too expensive to use more that a few tens of thousands

of atoms – far too few to obtain realistic results. Therefore care must be taken

in interpreting results, which are frequently affected by numerical artifacts. How-

ever, if treated with care, MD is most useful today in obtaining physical insights

into problems where wave and phase coherence effects are important, and where

perturbation approaches to modeling phonon scattering are likely to fail.

Models based on the Boltzmann transport equation are still evolving, and only

recently have attempts been made to include more information from dispersion

curves and to incorporate more accurate computations of relaxation times. At

present, these types of models do not account for crystal anisotropy, and still op-

erate under the relaxation time approximation. Because BTE-based models have

hitherto been relatively rudimentary, it has been difficult to make direct compar-

isons with MD simulations except in simple cases involving thermal conductivity

calculations. However, it would be useful to understand the limits of BTE-based

predictions and to determine when molecular dynamics simulations are truly nec-

essary. As BTE-based models become more complex, it is unclear whether they

will offer significant computational advantages with respect to MD simulations.

Efficient numerical methods and parallel processing are critical for the success of

46
these more complex models.

At present, multi-scale numerical methods for phonon transport are in their

infancy. Though coupled BTE/Fourier simulations are relatively simple to do, and

have been used in the simulation of FETs [24], it is as yet unclear how to couple

MD and Fourier calculations, or MD and BTE computations. These promise to be

more difficult than multi-scale computations of solid mechanics, but may finally

enable the simulation of realistic thermal transport problems in microelectronics,

laser processing, and other areas. Finally, multi-physics simulations coupling, for

example, electron and phonon transport, remain an open area for research.

Acknowledgments

Support under NSF grants CTS-0312420, CTS-0219098 and EEC-0228390, as

well as the IBM Faculty Partnership award to J. Murthy are acknowledged.

47
List of Figures

Fig. 1: Dispersion curve for 1-D crystal lattice with two atoms per unit cell.

Fig. 2: Dispersion curves at 80K for germanium in [111] direction. Taken from [13].

Fig. 3:Dispersion curves in [100] direction for 6-layer Si layer using EDIP poten-

tial( [18, 19])

Fig. 4:Bulk thermal conductivity of Si computed using non-gray model and com-

parison with experimental data from [30].

Fig. 5:Thin film thermal conductivity of Si computed using non-gray model and

comparison with experimental data from [27, 28].

Fig. 6: Geometry of SOI FET [24]

Fig. 7:Temperature field (K) in SOI FET using (a) Fourier model, (b) gray model,

(c) semi-gray model and (d) non-gray model ( [6, 24]).

Fig. 8:Temperature variation at three different time instants for zero acoustic thick-

ness using standard finite volume scheme.



Fig. 9:Temperature variation for zero acoustic thickness at =0.1, 0.5 and 1.0 with

modified finite volume scheme.

48
List of Tables

Table 1: CPU time and iterations for convergence using COMET.

49
Table 1: CPU time and iterations for convergence using COMET
* !
Acoustic thickness FœG Y Sequential Coupled
CPU s Iter CPU s Iter
m r m r
cells
0.1 1.28 18 0.53 5
1.0 2.46 35 0.57 5
10.0 20.44 297 0.93 5
100.0 239.77 3385 1.15 5
# r # r
cells
0.1 3.03 15 1.68 5
1.0 5.65 35 2.05 5
10.0 45.77 346 5.66 5
100.0 679.07 5080 10.04 5
r r cells
Ë Ë
0.1 13.46 15 6.45 5
1.0 22.46 32 10.01 5
10.0 159.71 371 35.13 5
 
100.0 2500 6000 93.81 5
r r cells
Ò Ò
0.1 58.28 14 24.31 5
1.0 100.90 29 41.44 5
10.0 659.78 371 157.13 5
 
100.0 1e5 6000 593.63 5

50
Figure 1: Dispersion curve for 1-D crystal lattice with two atoms per unit cell.
Taken from [13]

51
Figure 2: Dispersion curves at 80K for germanium in [111] direction. Taken
from [13]

52
Figure 3: Dispersion curves in [100] direction for 6-layer Si layer using EDIP
potential [18, 19]

53
Figure 4: Bulk thermal conductivity of Si computed using non-gray model and
comparison with experimental data from [30]

54
Figure 5: Thin film thermal conductivity of Si computed using non-gray model
and comparison with experimental data from [27, 28].

55
900 nm 100 nm
10 nm

Silicon heat generation 72 nm

Silicon Dioxide
243 nm

1633 nm

Figure 6: Geometry of SOI FET [24]

56
(a)

(b)

(c)

(d)

Figure 7: Temperature field in SOI FET using (a) Fourier model, (b) gray model,
(c) semi-gray model and (d) non-gray model [6, 24].

57
0.6
Exact
0.5 Second Order 8x8
First Order 8x8
Second Order 4x4
0.4


Τ 0.3
t* = 1.0
0.2

t*=0.1 t*=0.5
0.1

0
0 0.2 0.4 0.6 0.8 1
x*
Figure 8: Temperature variation at three different time instants for zero acoustic
thickness using standard finite volume scheme

58
0.6
t*= 0.1
0.5 t*= 0.5
t*= 1.0
Exact
0.4


Τ 0.3
0.2

0.1

0
0 0.2 0.4 0.6 0.8 1
x*

Figure 9: Temperature variation for zero acoustic thickness at =0.1, 0.5 and 1.0
with modified finite volume scheme

59
References

[1] G. Chen. Size and interface effects on thermal conductivity of superlattices

and periodic thin film structures. Journal of Heat Transfer, 119:220–229,

1997.

[2] G. Chen. Thermal conductivity and ballistic phonon transport in the cross-

plane direction of superlattices. Physical Review B, 57:14958–14793, 1998.

[3] A. Majumdar. Microscale heat conduction in dielectric thin films. Journal of

Heat Transfer, 115:7–16, 1993.

[4] A. Joshi and A. Majumdar. Transient ballistic and diffusive phonon heat

transport in thin films. Journal of Applied Physics, 74:31–39, 1993.

[5] P.G. Sverdrup, Y.S. Ju, and K.E. Goodson. Sub-continuum simulations of

heat conduction in silicon-on-insulator devices. Journal of Heat Transfer,

120:30–36, 1998.

[6] S.V.J. Narumanchi, J.Y. Murthy, and C.H. Amon. Computations of sub-

micron heat transport in silicon accounting for phonon dispersion. Paper

HT2003-40490, ASME Summer Heat Transfer Conference, July 2003.

60
[7] A. Balandin and K. L. Wang. Significant decrease of the lattice thermal con-

ductivity due to phonon confinement in a free-standing semiconductor quan-

tum well. Physical Review B, 58(3):1544–1549, 1998.

[8] S.G. Volz and G. Chen. Molecular-dynamics simulation of thermal conduc-

tivity of silicon crystals. Physical Review B, 61(4):2651–2656, January 2000.

[9] A.R. Abramson, C-L. Tien, and A. Majumdar. Interface and strain effects

on the thermal conductivity of heterostructures: A molecular dynamics study.

Journal of Heat Transfer, 124:963–970, 2002.

[10] A.J.H. McGaughey and M. Kaviany. Thermal conductivity decomposition

and analysis using molecular dynamics simulations, part I. Lennard-Jones ar-

gon. International Journal of Heat and Mass Transfer, 47:1783–1798, 2003.

[11] J.A. Smirnova, L.V. Zhigelei, and B.J. Garrison. A combined molecular dy-

namics and finite element method technique applied to laser induced pressure

wave propagation. Computer Physics Communications, 118:11–16, 1999.

[12] A. Majumdar. Microscale energy transport in solids. In C-L. Tien, A. Majum-

dar, and F. Gerner, editors, Microscale Energy Transfer, Series in Chemical

and Mechanical Engineering, pages 1–95. Taylor and Francis, 1997.

[13] C. Kittel. Introduction to Solid State Physics. John Wiley, 1996.

61
[14] N.W. Ashcroft and N.D. Mermin. Solid State Physics. International Thomson

Publishing, 1976.

[15] J.M. Ziman. Electrons and Phonons. Oxford University Press, London, UK,

1960.

[16] B. Yang and G. Chen. Lattice dynamics study of phonon heat conduction in

quantum wells. Physics of Lower-Dimensional Structures, 5/6:37–48, 2000.

[17] P.G. Klemens. Thermal conductivity and lattice vibrational modes. Physical

Review B, 15(12):5957–5962, 1958.

[18] Jose A. Pascual-Gutierrez and J.Y.Murthy. unpublished results, 2004.

[19] N. Bernstein, M. J. Aziz, and E. Kaxiras. Atomistic simulations of solid-

phase epitaxial growthin silicon. Phys. Rev. B., 61(10):6696–6700, 2000.

[20] D. Tzou. Macro-to-Microscale Heat Transfer: The Lagging Behavior. Taylor

and Francis, 1996.

[21] D.D. Joseph and L. Preziosi. Heat waves. Rev. Mod. Phys., 61(1):41–73,

1989.

[22] B.H. Armstrong. Two-fluid theory of thermal conductivity of dielectric crys-

tals. Physical Review B, 23:883–899, 1981.

62
[23] Y.S. Ju and K.E. Goodson. Microscale Heat Conduction in Integrated Cir-

cuits and Their Constituent Thin Films, chapter 5, pages 58–79. Kluwer

Academic Publishers, 1999.

[24] S.V.J. Narumanchi, J.Y. Murthy, and C.H. Amon. Computations of heat con-

duction in sub-micron Silicon-On-Insulator transistors accounting for phonon

dispersion and polarization. Paper HT2003-42447, ASME IMECE, Novem-

ber 2003.

[25] Y.-J. Han and P.G. Klemens. Anharmonic thermal resistivity of dielectric

crystals at low temperature. Physical Review B, 48(9):6033–6042, 1983.

[26] M.G. Holland. Analysis of lattice thermal conductivity. Physical Review, 132

(6):2461–2471, 1963.

[27] M. Asheghi, M.N. Touzelbaev, K.E. Goodson, Y.K. Leung, and S.S. Wong.

Temperature-dependent thermal conductivity of single-crystal silicon layers

in SOI substrates. Journal of Heat Transfer, 120:30–36, 1998.

[28] M. Asheghi, K. Kurabayashi, R. KAsnavi, and K.E. Goodson. Thermal con-

duction in doped single crystal silicon films. Journal of Applied Physics, 91

(8):5079–5088, 2002.

63
[29] J. Callaway. Model of lattice thermal conductivity at low temperature. Phys-

ical Review, 113:1046–1051, 1959.

[30] M.G. Holland. Phonon scattering in semiconductors from thermal conductiv-

ity studies. Physical Review, 134(2A):471–480, 1964.

[31] T. Wang and J.Y. Murthy. An improved computational procedure for phonon

relaxation times. Paper IMECD2004-61901, ASME IMECE, November

2004.

[32] G. Chen. Ballistic-diffusive heat conduction equations. Physical Review Let-

ters, 86(11):2297–2300, 2001.

[33] M. F. Modest. Radiative Heat Transfer. Series in Mechanical Engineering.

McGraw Hill, New York, NY, 1993.

[34] J.Y. Murthy and S.R. Mathur. Ballistic-diffusive approximation for phonon

transport accounting for phonon dispersion and polarization. Paper HT2003-

47491, ASME Heat Transfer Conference, July 2003.

[35] M. A. Heaslet and R. F. Warming. Radiative transport and wall temperature

slip in an absorbing planar medium. Int. J. Heat Mass Transfer, 8:979–994,

1964.

64
[36] A.J.H. McGaughey and M. Kaviany. Quantitative validation of the Boltzmann

transport equation phonon thermal conductivity model under single-mode re-

laxtion time approximation. Physical Review B, 69(9):094303(1– 12), 2004.

[37] M.P. Allen and D.J. Tildesley. Computer Simulation of Liquids. Oxford Uni-

versity Press, 1989.

[38] F.H. Stillinger and T.A. Weber. Computer simulation of local order in con-

densed phases of silicon. Physical Review B, 31(8):5262–5271, 1985.

[39] J. Tersoff. New empirical approach for the structure and energy of covalent

systems. Physical Review B, 37(2):6991–7000, 1988.

[40] A.J.C. Ladd and B. Moran. Lattice thermal conductivity: A comparison of

molecular dynmaics and anharmonic lattice dynamics. Physical Review B,

34:5058–5064, 1986.

[41] P.K. Schelling, S.R. Phillpot, and P. Klebinski. Phonon wave-packet dynam-

ics at semiconductor interfaces by molecular-dynamics simulations. Applied

Physics Letters, pages 2484–2486, 2002.

[42] P.K. Schelling and S.R. Phillpot. Multiscale simulation of phonon transport

in superlattices. Journal Applied Physics, pages 5377–5387, 2002.

65
[43] B. G. Carlson and K. D. Lathrop. Transport theory – The method of Discrete

Ordinates. In H. Greenspan, C. Kelber, and D. Okrent, editors, Computing

Methods in Reactor Physics, pages 171–266. Gordon and Breach, New York,

1968.

[44] G. D. Raithby and E. H. Chui. A finite-volume method for predicting a radiant

heat transfer in enclosures with participating media. J. Heat Transfer, 112:

415 – 423, 1990.

[45] J.Y. Murthy and S.R. Mathur. Finite volume method for radiative heat transfer

using unstructured meshes. J. Thermophys. Heat Transfer, 12(3):313–321,

Jul-Sep 1998.

[46] J.Y. Murthy and S.R. Mathur. Computation of sub-micron thermal transport

using an unstructured finite volume method. Journal of Heat Transfer, 124

(6):1176–1181, 2002.

[47] C. Hirsch. Numerical Computation of Internal and External Flows: Compu-

tational Methods for Inviscid and Viscous Flows. John Wiley & Sons, 1990.

[48] C.B. Laney. Computational Gas Dynamics. Cambridge University Press,

Cambridge, U.K., 1998.

66
[49] B.P. Leonard. A stable and accurate convective modeling procedure based

on quadratic upstream interpolation. Comput. Methods Appl. Mech. Eng., 19:

59–98, 1979.

[50] S.R. Mathur and J.Y. Murthy. A pressure based method for unstructured

meshes. Numer. Heat Transfer, 31(2):195–216, March 1997.

[51] J.C. Chai, H.S. Lee, and S.V. Patankar. Ray effects and false scattering in the

discrete ordinates method. Numerical Heat Transfer B: Fundamentals, 24:

373–389, 1993.

[52] P. Coelho. The role of ray effects and false scattering on the accuracy of

the standard and modified discrete ordinates methods. J. Quant. Spectrosc.

Radiat. Transfer, 73:231–238, 2002.

[53] J.Y. Murthy and S.R. Mathur. An improved computational procedure for sub-

micron heat conduction. J. Heat Transfer, 125(5):904–910, 2003.

[54] S.R. Mathur and J.Y. Murthy. Coupled ordinates method for multigrid accel-

eration of radiation calculations. Journal of Thermophysics and Heat Trans-

fer, 13(4):467–473, 1999.

[55] C. Jacoboni and L. Reggiani. The Monte Carlo method for the solution of

charge transport in semiconductors with applications to covalent materials.

67
Reviews of Modern Physics, 55(3):645–705, July 1983.

[56] R.B. Peterson. Direct simulation of phonon-mediated heat transfer in a Debye

crystal. Journal of Heat Transfer, 116:815–822, 1994.

[57] S. Mazumdar and A. Majumdar. Monte Carlo study of phonon transport

in solid thin films including dispersion and polarization. Journal of Heat

Transfer, 123:749–759, 2001.

[58] K. Hess. Monte Carlo Device Simulation: Full Band and Beyond. Kluwer

Academic Press, Norwell, MA, 1991.

[59] F.F. Abraham, J.Q. Broughton, N. Bernstein, and E. Kaxiras. Spanning con-

tinuum to quantum length scales in a dynamic simulation of brittle fracture.

Europhysics Letters, 44(6):783–787, 1998.

[60] J.Q. Broughton, F.F. Abraham, N. Bernstein, and E. KAxiras. Concurrent

coupling of length scales. Physical Review B, 60(4):2391–2403, 1999.

[61] E.B. Tadmor, R. Phillips, and M. Ortiz. Mixed atomistic and continuum mod-

els of deformation in solids. Langmuir, 12(19):4529–4534, 1996.

[62] V.B. Shenoy, R. Miller, E.B. Tadmor, R. Phillips, and M. Ortiz. Quasicontin-

uum models of interfacial structure and deformation. Phys. Rev. Lett., 80(4):

742–745, 1998.

68
[63] V. Shenoy, V. Shenoy, and R. Phillips. Finite temperature quasicontin-

uum methods. In V. Bulatov, T. d. l. Rubia, R. Phillips, E. Kaxiras, and

N. Ghoneim, editors, Multiscale Modeling of Materials, pages 465–471. Ma-

terials Research Society, Warrendale, PA, 1999.

[64] G.J. Wagner and W. K. Liu. Coupling of atomistic and continuum simulations

using a bridging scale decomposition. J. Computational Physics, pages 1–35,

2003. (to appear).

69

You might also like