Predicting phonon transport in semiconductornanostructures using atomistic calculations and the
Boltzmann transport equation
by
Daniel P. Sellan
A thesis submitted in conformity with the requirementsfor the degree of Doctor of Philosophy
Graduate Department of Mechanical & Industrial EngineeringUniversity of Toronto
Copyright c© 2012 by Daniel P. Sellan
Abstract
Predicting phonon transport in semiconductor nanostructures using atomistic
calculations and the Boltzmann transport equation
Daniel P. Sellan
Doctor of Philosophy
Graduate Department of Mechanical & Industrial Engineering
University of Toronto
2012
The mechanisms of thermal transport in defect-free silicon nanostructures are ex-
amined using a combination of lattice dynamics (LD) calculations and the Boltzmann
transport equation (BTE). To begin, the thermal conductivity reduction in thin films
is examined using a hierarchical method that first predicts phonon transport properties
using LD calculations, and then solves the phonon BTE using the lattice Boltzmann
method. This approach, which considers all of the phonons in the first Brillouin-zone, is
used to assess the suitability of common assumptions used to reduce the computational
effort. Specifically, we assess the validity of: (i) neglecting the contributions of optical
modes, (ii) the isotropic approximation, (iii) assuming an averaged bulk mean-free path
(i.e., the Gray approximation), and (iv) using the Matthiessen rule to combine the effect
of different scattering mechanisms. Because the frequency-dependent contributions to
thermal conductivity change as the film thickness is reduced, assumptions that are valid
for bulk are not necessarily valid for thin films.
Using knowledge gained from this study, an analytical model for the length-dependence
of thin film thermal conductivity is presented and compared to the predictions of the LD-
based calculations. The model contains no fitting parameters and only requires the bulk
lattice constant, bulk thermal conductivity, and an acoustic phonon speed as inputs. By
including the mode-dependence of the phonon lifetimes resulting from phonon-phonon
ii
and phonon-boundary scattering, the model predictions capture the approach to the bulk
thermal conductivity better than predictions made using Gray models based on a single
lifetime.
Both the model and the LD-based method are used to assess a procedure commonly
used to extract bulk thermal conductivities from length-dependent molecular dynamics
simulation data. Because the mode-dependence of thermal conductivity is not included
in the derivation of this extrapolation procedure, using it can result in significant error.
Finally, phonon transport across a silicon/vacuum-gap/silicon structure is modelled
using lattice dynamics and Landauer theory. The phonons transmit thermal energy
across the vacuum gap via atomic interactions between the leads. Because the incident
phonons do not encounter a classically impenetrable potential barrier, this mechanism is
not a tunneling phenomenon. The heat flux due to phonon transport can be 4 orders of
magnitude larger than that due to photon transport predicted from near-field radiation
theory.
iii
Acknowledgements
I would like to thank Elizabeth, my partner and best friend, for her encouragement
and support. I am blessed to have someone who shows me constant and unconditional
love and support.
I would like to thank my supervisor Cristina Amon for providing me with incredible
guidance, both academic and personal, as well as for her undivided support in aiding
my development as a young scientist. I am grateful to all the members of the Advanced
Thermo/fluid Optimization, Modelling, and Simulations (ATOMS) Laboratory at the
University of Toronto, I cannot imagine having a more encouraging group of colleagues
and friends or a more positive working environment.
I would like to thank Alan McGaughey for his tireless efforts in helping me become a
better researcher. I would also like to thank all the members of the Nanoscale Transport
Phenomena Laboratory at Carnegie Mellon University for sharing their knowledge and
helping me to expand my research capabilities.
I would like to thank Nazir Kherani, Yu Sun, and Charles Ward at the University of
Toronto for serving on my PhD thesis committee. Their guidance throughout this process
has been outstanding. I would like to thank Kenneth Goodson (Stanford University) and
Chandra Veer Singh (University of Toronto) for serving as external examiners for my PhD
defence.
I would like to thank the National Science and Engineering Research Council (NSERC)
of Canada for funding this work through their Alexander Graham Bell Canada Graduate
Scholarship program.
Lastly, and obviously not the least, I would like to thank my parents for their un-
conditional love and support as well as my twin brother Mike for his expertise in almost
everything except medicine.
iv
Contents
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Thesis Overview and Scope . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Thermal Transport Models 5
2.1 Fourier Heat Conduction Equation . . . . . . . . . . . . . . . . . . . . . 5
2.2 Boltzmann Transport Equation . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Solving the Boltzmann Transport Equation using the Lattice Boltzmann
Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.1 Lattice Boltzmann Methodology . . . . . . . . . . . . . . . . . . . 9
2.3.2 Two Dimensional Lattice Structures: D2Q7 vs. D2Q9 . . . . . . 12
3 Comparing Fourier-based and BTE-based Predictions 14
3.1 Validating the Gray LBM in the Bulk Limit . . . . . . . . . . . . . . . . 14
3.2 Modelling Sub-Continuum Thermal Transport in Thin Films: A Steady-
State Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2.1 Predicting Thermal Conductivity from LBM Results . . . . . . . 19
3.2.2 Modelling Boundary Scattering in Thermal Conductivity Predictions 20
3.2.3 Modelling Sub-Continuum Thermal Transport using Fourier-based
Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
v
3.3 Heated Surface: Transient Comparison . . . . . . . . . . . . . . . . . . . 24
3.4 Fourier-based vs. BTE Solutions: A Frequency-domain Thermoreflectance
Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.4.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4 Lattice Dynamics Calculations and the BTE 33
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2 Computational Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2.1 Overview of the Hierarchical Procedure . . . . . . . . . . . . . . . 35
4.2.2 Predicting Phonon Properties using Lattice Dynamics Calculations 36
4.2.3 Cross-plane Phonon Transport using the BTE . . . . . . . . . . . 38
4.3 Analyzing the Results and Assessing Common Assumptions . . . . . . . 41
4.3.1 Thermal Conductivity Predictions . . . . . . . . . . . . . . . . . . 41
4.3.2 Isotropic Approximation . . . . . . . . . . . . . . . . . . . . . . . 45
4.3.3 Simplified Model Based on the Gray Approximation . . . . . . . . 47
4.3.4 Boundary Scattering and the Matthiessen Rule . . . . . . . . . . 47
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5 Theoretical Relations and the BTE 50
5.1 Theoretical Relations for Phonon Properties . . . . . . . . . . . . . . . . 50
5.2 Model Derivation for Length-dependent Thermal Conductivity . . . . . . 52
5.3 Model Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.3.1 Comparison to Predictions from Lattice Dynamics Calculations . 55
5.3.2 Comparison to Gray Models . . . . . . . . . . . . . . . . . . . . . 57
5.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
6 Size Effects in MD Thermal Conductivity Predictions 59
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
6.2 Predicting Thermal Conductivity using Molecular Dynamics Simulation . 62
vi
6.2.1 Green-Kubo Method . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.2.2 Direct Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.3 Predicting Thermal Conductivity using Lattice Dynamics Calculations . 70
6.3.1 Phonon Properties . . . . . . . . . . . . . . . . . . . . . . . . . . 70
6.3.2 Assessing the Linear Extrapolation Procedure . . . . . . . . . . . 72
6.3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.4 Assessing the Linear Extrapolation Procedure using a Length-dependent
Thermal Conductivity Model . . . . . . . . . . . . . . . . . . . . . . . . 76
6.5 Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
7 Thermal Transport Across a Vacuum-Gap 81
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
7.2 Vacuum-gap Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
7.3 Phonon Transmission Coefficients . . . . . . . . . . . . . . . . . . . . . . 84
7.4 Vacuum Thermal Resistance . . . . . . . . . . . . . . . . . . . . . . . . . 89
7.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
8 Contributions and Future Research Directions 96
8.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
8.2 Suggestions for Future Study . . . . . . . . . . . . . . . . . . . . . . . . . 100
Bibliography 105
A Appendix 118
A.1 Streaming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
A.2 Initial Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
A.3 Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
A.3.1 Periodic Boundary Condition . . . . . . . . . . . . . . . . . . . . 119
A.3.2 Constant Temperature Boundary Condition . . . . . . . . . . . . 119
vii
A.3.3 Constant Heat Flux Boundary Condition . . . . . . . . . . . . . . 120
A.3.4 Adiabatic Boundary Condition . . . . . . . . . . . . . . . . . . . 122
A.3.5 Corner Nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
viii
List of Tables
3.1 Phonon properties for bulk silicon at temperature of 300 K. Reprinted
with permission from Ref. [1], Copyright 2011 Elsevier. . . . . . . . . . . 16
4.1 Thermal conductivity predictions for bulk SW silicon and thin films (cross-
plane direction) at a temperature of 300 K. . . . . . . . . . . . . . . . . . 42
6.1 Size-dependence of SW silicon thermal conductivity at a temperature of
500 K predicted using MD simulations and the Green-Kubo method. The
prediction uncertainty is the 95% confidence interval based on the results
of ten independent simulations. Reprinted with permission from Ref. [2],
Copyright 2010 American Physical Society. . . . . . . . . . . . . . . . . . 64
6.2 Bulk thermal conductivities, in W/m-K, for SW silicon and LJ argon
found using the Green-Kubo method (kGK – Sec. 6.2.1) and the direct
method (keDM – Sec. 6.2.2) in MD simulations and from lattice dynamics
calculations (kLD and keLD – Sec. 6.3). The direct method uncertainty is
estimated to be ±20% for SW Silicon and ±10% for LJ Argon based on
the prediction repeatability. The superscript e indicates that the value
was predicted using the linear extrapolation procedure. Reprinted with
permission from Ref. [2], Copyright 2010 American Physical Society. . . . 65
ix
6.3 Lattice constants, densities, and elastic constants for LJ argon and SW
silicon at a temperature of 0 K. Reprinted with permission from Ref. [2],
Copyright 2010 American Physical Society. . . . . . . . . . . . . . . . . . 78
7.1 Vacuum thermal resistance predicted by lattice dynamics calculations and
near-field radiation theory for a 1 Å-wide vacuum-gap at a temperature of
300 K. Thermal boundary resistance for a Si/Si grain boundary [Σ29(001)]
and a Si/Ge interface predicted by molecular dynamics simulation using
the Stillinger-Weber potential at a temperature of 500 K. Reprinted with
permission from Ref. [6], Copyright 2012 American Physical Society. . . . 93
x
List of Figures
2.1 D1Q3, D2Q7, and D2Q9 lattice structures and respective labelling. Reprinted
with permission from Ref. [1], Copyright 2011 Elsevier. . . . . . . . . . . 11
2.2 D2Q7 lattice implementation and boundary lattice nodes. Reprinted with
permission from Ref. [1], Copyright 2011 Elsevier. . . . . . . . . . . . . . 12
3.1 (a) Temperature and (b) heat flux profile predicted by the analytical
Fourier-based solution (solid line) and the Gray LBM using the D2Q7
(open circles) and D2Q9 (open triangles) lattices for the diffuse regime.
Reprinted with permission from Ref. [1], Copyright 2011 Elsevier. . . . . 15
3.2 Thin film model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.3 (a) Steady-state temperature profile in a silicon film with constant temper-
ature boundary conditions predicted using the Gray LBM, Kn = 0.33. (b)
Steady-state temperature profile predicted using the Fourier-based heat
equation for a silicon film with the domain extended by 2Λ. . . . . . . . 18
3.4 (a) Effective thermal conductivity normalized by the bulk value (kfilm/kbulk)
as a function of the Knudsen number (Kn) predicted by the Gray LBM
and Eq. (3.8). (b) Steady-state boundary temperature jump normalized
by the bulk value (Tjump/∆Tbulk) as a function of the Kn predicted using
the Gray LBM and Eq. (3.10). . . . . . . . . . . . . . . . . . . . . . . . . 20
xi
3.5 (a) Transient ballistic temperature profile predicted by the Gray LBM
and the hyperbolic heat equation with Kn = 0.33 and t∗ = 1. (b) Time-
dependent temperature jump on the left and right boundaries of a silicon
film predicted by the Gray LBM, Kn = 0.33. (c) Jump in the right bound-
ary temperature, Tjj, when the thermal wave hits the cold right boundary
as a function of 1/Kn. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.6 Schematic diagram of the simplified frequency-domain thermoreflectance
setup. The sample is periodic in the x and y directions and infinite in the
z direction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.7 Non-dimensional temperature, T ∗, as a function of non-dimensional sample
depth, z/Λ, for a system with a surface temperature oscillating at a non-
dimensional frequency of fτ = 0.025. Each grey line corresponds to a
given time step after the system reached steady state. . . . . . . . . . . . 29
3.8 Non-dimensional penetration depth, L/Λ, as a function of non-dimensional
frequency, fτ , for the BTE (squares) and analytical solution to the con-
duction equation (solid line). The dashed line indicates L = 2Λ. . . . . . 30
3.9 Required heat flux amplitude, h0, as a function of non-dimensional pene-
tration depth, L/Λ, for the BTE (squares) and analytical solution to the
conduction equation (solid line). . . . . . . . . . . . . . . . . . . . . . . . 31
4.1 Flow chart of the hierarchical procedure for predicting the phonon thermal
conductivity of bulk and thin films using lattice dynamics calculations
and the BTE. The theoretical/computational tools are in boxes and their
inputs and outputs are in ovals. Reprinted with permission from Ref. [3],
Copyright 2010 American Institute of Physics. . . . . . . . . . . . . . . . 34
xii
4.2 (a) Frequency dependence on wave vector magnitude, |κ|, for the entire
Brillouin zone of SW silicon at a temperature of 300 K. Dispersion curves
corresponding to the [001] direction are shown as solid lines. (b) Phonon
relaxation time dependence on frequency for SW silicon at a temperature of
300 K. Near the Brillouin zone center, the relaxation times are reasonably
represented by τ = A/ω2 (solid line) [4], where A is a constant calculated
for this data to be 2×1015 1/s. Reprinted with permission from Ref. [3],
Copyright 2010 American Institute of Physics. . . . . . . . . . . . . . . . 37
4.3 Sub-continuum temperature profile across a 17.4 nm SW silicon thin film.
Reprinted with permission from Ref. [3], Copyright 2010 American Insti-
tute of Physics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.4 Cross-plane thin film thermal conductivity normalized by the bulk value.
Phonons from the full Brillouin zone (squares) and the isotropic approx-
imation (circles) are considered. The cross-plane thermal conductivity
found using the Matthiessen rule (solid line) and a simplified model (dashed
line, see Sec. 4.3.3) are also plotted. [A] corresponds to Ref. [5]. Reprinted
with permission from Ref. [3], Copyright 2010 American Institute of Physics. 43
4.5 (a) Cross-plane thermal conductivity contribution dependence on the bulk
phonon MFP. The MFPs for each mode are sorted using a histogram with
a bin width of 2 nm. The thermal conductivity contribution is normalized
by the total value. (b) Thermal conductivity contribution dependence on
frequency for bulk and for 556 and 34.8 nm thin films. The area under
each curve is proportional to the total thermal conductivity. Reprinted
with permission from Ref. [3], Copyright 2010 American Institute of Physics. 44
5.1 In-plane and cross-plane thermal conductivity models. . . . . . . . . . . . 54
5.2 Comparison of thermal conductivity models and lattice dynamics/BTE
predictions: (a) Cross-plane, (b) In-plane. . . . . . . . . . . . . . . . . . 56
xiii
6.1 Heat current autocorrelation function (body) and its integral (insert) for
SW silicon at temperatures of 500 and 1000 K. The HCACFs are nor-
malized by their initial values. The shaded region in the insert indicates
the time range over which the HCACF integral is averaged to predict the
thermal conductivity, kGK, which is represented by a dashed line for each
temperature. Reprinted with permission from Ref. [2], Copyright 2010
American Physical Society. . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.2 Schematic diagram of the simulation cell geometry used in the direct
method simulations. Reprinted with permission from Ref. [2], Copyright
2010 American Physical Society. . . . . . . . . . . . . . . . . . . . . . . . 67
6.3 Length-dependent thermal conductivities for (a) SW silicon at temper-
atures of 500 and 1000 K and (b) LJ argon at a temperature of 40 K
predicted from the direct method and MD simulations. The dashed lines
are linear fits to the discrete data. Reprinted with permission from Ref.
[2], Copyright 2010 American Physical Society. . . . . . . . . . . . . . . . 69
6.4 Inverse of the normalized length-dependent thermal conductivities for (a)
SW silicon (T = 500 K) and (b) LJ argon (T = 40 K). The squares cor-
respond to the sample lengths used in the direct method simulations [see
Figs. 6.3(a) and 6.3(b)], but are calculated using phonon properties ob-
tained from lattice dynamics calculations. Note the difference in the scales
of the vertical axes. Reprinted with permission from Ref. [2], Copyright
2010 American Physical Society. . . . . . . . . . . . . . . . . . . . . . . . 71
6.5 Bulk thermal conductivity contribution dependence on MFP for SW silicon
(T = 500 K) and LJ argon (T = 40 K). The MFPs for each mode are sorted
using a histogram with a bin size of 2 nm for SW silicon and 0.1 nm for LJ
argon. Reprinted with permission from Ref. [2], Copyright 2010 American
Physical Society. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
xiv
6.6 Inverse of the normalized length-dependent thermal conductivity for SW
silicon at T = 500 K (solid line). The squares correspond to the data in
Fig. 6.4(a) [note that this figure has a logarithmic horizontal axis]. The
diamonds correspond to sample lengths between 4 and 8 µm. Reprinted
with permission from Ref. [2], Copyright 2010 American Physical Society. 74
6.7 Length-dependence of the thermal conductivity using Eq. (6.6) and the
bulk thermal conductivity estimated using the linear extrapolation proce-
dure [Eq. (6.8)]. Reprinted with permission from Ref. [2], Copyright 2010
American Physical Society. . . . . . . . . . . . . . . . . . . . . . . . . . . 77
7.1 Schematic diagram of the three-dimensional Si/vacuum-gap/Si structure
for LG = 1.72 Å. The shaded region between the dark silicon atoms is the
volume associated with the perfect silicon crystal. Vacuum space, shown
in white, is added to form the vacuum-gap. What we call vacuum space
is in fact a region of finite electron density. The structure is periodic
in the x and y directions and semi-infinite in the negative and positive
z directions. Reprinted with permission from Ref. [6], Copyright 2012
American Physical Society. . . . . . . . . . . . . . . . . . . . . . . . . . . 85
7.2 Frequency dependence of (a) phonon transmission coefficient [αL→R(κκκ, ν,LG)]
for 10 000 randomly sampled phonon modes in the first Brillouin zone and
(b) the fraction of transmitted phonon energy that remains in its original
mode [η(κκκ, ν,LG)] for transmitted phonon modes with αL→R(κκκ, ν,LG) >
0.01. (c) Bulk phonon density of states for the silicon leads. The phonon
density of states is calculated using a histogram with a bin width of 0.25
THz. The dashed line around 12 THz separates acoustic modes from op-
tical modes. Reprinted with permission from Ref. [6], Copyright 2012
American Physical Society. . . . . . . . . . . . . . . . . . . . . . . . . . . 87
xv
7.3 (a) Vacuum thermal resistance (RNE) as a function of vacuum-gap width
(LG) at a temperature of 300 K. (b) Total lattice energy (E) and its
derivative (dE/dLG) as a function of vacuum-gap width. The dashed
line at LG = 1.89 Å corresponds to the Stillinger-Weber silicon potential
cutoff (LcutoffG ). Reprinted with permission from Ref. [6], Copyright 2012
American Physical Society. . . . . . . . . . . . . . . . . . . . . . . . . . . 90
7.4 Frequency-dependence of vacuum thermal conductance (1/RNE) at a tem-
perature of 300 K. The mode conductances sorted using a histogram with
a bin width of 0.25 THz. The dashed line around 12 THz separates acous-
tic modes from optical modes. Reprinted with permission from Ref. [6],
Copyright 2012 American Physical Society. . . . . . . . . . . . . . . . . . 92
A.1 Unknown phonon populations for periodic boundary condition in the x
direction for D2Q7 lattice. . . . . . . . . . . . . . . . . . . . . . . . . . . 120
A.2 Unknown phonon populations for adiabatic boundary condition implemen-
tation on the top and left boundaries. . . . . . . . . . . . . . . . . . . . . 122
A.3 A node at the corner of constant temperature and adiabatic boundaries. . 124
xvi
Chapter 1
Introduction
1.1 Motivation
Many modern electronic technologies such as graphics processing units, solid-state mem-
ory, and semiconductor light-emitting diodes incorporate components with nanometer-
scale dimensions. The continued miniaturization of these components results in increased
power densities, and the effective removal of generated waste heat is critical to device
operation and reliability [7, 8, 9]. Understanding the underlying heat transport mech-
anisms in these nanostructured materials is thus of great importance. Unfortunately,
continuum-level thermal models, such as the Fourier heat equation, are invalid at such
small length scales, where boundary effects play a significant role.
By adjusting the thermal properties in continuum-level thermal models, sub-continuum
transport can be captured across sub-continuum regimes but not within them. For ex-
ample, applying the Fourier heat equation using an effective thermal conductivity can
give the correct heat flux across the film but not the correct temperature gradient within
it.
Although bulk thermal conductivity can be treated as a function of material and tem-
perature, the closely packed surfaces and interfaces found in nanostrucutres sufficiently
1
Chapter 1. Introduction 2
alter the thermal transport landscape to give thermal conductivity a geometry depen-
dence. This geometry dependence gives researchers the ability to engineer materials with
desired thermal characteristics. One example is the semiconductor superlattice (a peri-
odic nanostructure containing stacked thin films of alternating species), which has been
developed to reduce from bulk the lattice thermal conductivity while not dramatically
affecting the electrical conductivity. Such a material is beneficial in thermoelectric gen-
eration devices. Developing an understanding of the mechanisms of thermal transport in
these nanostructured materials is necessary to fully exploit their design.
Although these materials can be characterized by building the structure and mea-
suring the properties at a series of temperatures, such a process is difficult and time
consuming. For this reason, investigators have turned to numerical studies to assist in
both characterizing and designing nanostructured materials. Some examples of numerical
work include thermal conductivity prediction for silicon thin films [10, 11] and nanowires
[12], Si/Ge [13, 14, 15, 16, 17] and Si/SiGe [17, 18] superlattices, and carbon nanotubes
[19, 20].
Despite progress on the analysis of nanostructured materials, fundamental questions
remain about the nature of thermal transport at the carrier-level. For instance, how much
of the observed lattice thermal conductivity reduction in superlattices is due to carrier-
boundary scattering compared to changes in the carrier properties, such as modifications
to band structure. This limited understanding of carrier-level transport in nanostructures
is exacerbated by inadequate models of carrier transport in bulk materials, which rely
on fitting parameters and major approximations [4, 21].
1.2 Thesis Overview and Scope
The objective of this thesis is to develop, validate, and apply heat transport models that
can be used to optimize the design of nanostructured materials. To perform this re-
Chapter 1. Introduction 3
search, we use a combination of atomistic modelling techniques [e.g., molecular dynamics
(MD) simulation and lattice dynamics (LD) calculations] and formulations grounded in
statistical thermodynamics [e.g., the Boltzmann transport equation (BTE)]. This thesis
attempts to elucidate the mechanisms of transport in systems where classical and contin-
uum relations break down and the transport must be analyzed at the carrier level [i.e.,
phonon (quantized lattice vibration), electron, and photon]. My research complements
experiment and is a combination of mechanical engineering, condensed matter physics,
and computational physics. The outline of the remainder of the thesis is as follows:
In Chapter 2, details regarding thermal transport modelling techniques are intro-
duced. Continuum-level equations (e.g., Fourier heat equation) and sub-continuum equa-
tions (e.g., BTE) are defined and their limitations are discussed. Methods to solve the
BTE are contrasted, and the lattice Boltzmann method (LBM) is discussed in detail.
In Chapter 3, the steady state and transient results of Fourier-based solutions are
compared to predictions of the BTE. Effort is made to characterize the LBM solutions
to the BTE using algebraic equations. The Gray approximation is made in this chapter
for the sake of simplicity and clarity. Methods to modify the Fourier-based equations so
that they reproduce the LBM results are presented.
In Chapter 4, mode-dependent phonon properties (obtained from harmonic and an-
harmonic lattice dynamics calculations) are included in BTE-based methods to predict
the cross-plane phonon thermal conductivity of Stillinger-Weber silicon thin films as thin
as 17.4 nm. This approach, which considers all of the phonons in the first Brillouin
zone, is used to assess the suitability of common assumptions. Because the frequency-
dependent contributions to thermal conductivity change as the film thickness is reduced,
assumptions that are valid for bulk are not necessarily valid for thin films.
In Chapter 5, an analytical model for the size-dependence of thin film thermal con-
ductivity is presented and compared to the predictions made in Chapter 4 on silicon
nanostructures. The model contains no fitting parameters and only requires the bulk
Chapter 1. Introduction 4
lattice constant, bulk thermal conductivity, and an acoustic phonon speed as inputs. By
including the mode-dependence of the phonon lifetimes resulting from phonon-phonon
and phonon-boundary scattering, the model captures the approach to the bulk thermal
conductivity better than Gray models based on a single lifetime.
In Chapter 6, the models developed in Chapters 4 and 5 are used to show that be-
cause mode-dependent phonon properties are not considered in the derivation of an ex-
trapolation procedure commonly used to extract bulk thermal conductivity from length-
dependent MD data [22], its use can lead to considerable error.
In Chapter 7, phonon transport across a silicon/vacuum-gap/silicon structure is mod-
elled using lattice dynamics and Landauer theory. We show that the heat flux due to
phonon transport can be 4 orders of magnitude larger than that due to photon transport
predicted from near-field radiation theory.
In Chapter 8, the major contributions of the work presented in this thesis and sug-
gestions for future study are discussed.
Chapter 2
Thermal Transport Models
2.1 Fourier Heat Conduction Equation
Thermal conductivity, k, is in general, a second rank, symmetric tensor empirically de-
fined by the Fourier law,
q = −k∇T , (2.1)
where q is the heat flux vector and ∇T is the temperature gradient in the material.
Though formally defined by this empirical relation [Eq. (2.1)], the thermal conductivity
is related to the properties of the sub-continuum energy carriers, specifically, the electrons
and phonons in solids.
Assuming no mass transfer and constant properties, the Fourier law can be combined
with the energy equation to yield
ρCvk
∂T
∂t= ∇2T, (2.2)
where ρ is the mass density, Cv is the specific heat, and t is time. Given appropriate
initial and boundary conditions, the temperature profile within a system can be predicted
using Eq. (2.2), which is the standard basis for describing conduction heat transfer.
It is known that Eq. (2.2) predicts an infinite thermal-wave propagation-speed, which
leads to a considerable error at small time scales. A remedy for this problem (albeit an
5
Chapter 2. Thermal Transport Models 6
over simplification) is to introduce a finite thermal-wave propagation-speed, vg, to Eq.
(2.2) to account for the wave nature of the thermal transport [23]. This procedure results
in the hyperbolic heat equation (Cattaneo equation) in the following form:
1
v2g
∂2T
∂t2+ρCvk
∂T
∂t= ∇2T. (2.3)
Like the Fourier heat equation [Eq. (2.2)] from which the hyperbolic heat equation is
derived, Eq. (2.3) is unable to capture ballistic transport. In fact, Equations (2.2) and
(2.3) are only valid when:
(i) the system can be modelled as a continuum and behaves classically, and
(ii) the energy carrier transport, be it a result of electrons or phonons, is diffuse rather
than ballistic. This means that the scattering of the carriers is primarily a result
of interactions with other carriers or with lattice imperfections and impurities that
are distributed throughout the material volume
When ballistic transport is present, the study of carrier behaviour becomes crucial for
accurate predictions of thermal transport [7]. In the ballistic (sub-continuum) regime,
energy carriers travel from one side of the system to the other without scattering. In
other words, ballistic effects are present when the characteristic size of a system, L, be-
comes comparable to the phonon mean free path (MFP), which is the average distance
a phonon travels before it scatters. Ballistic carriers contribute less to thermal conduc-
tivity than they would in a bulk system because their MFPs are reduced to the system
length. Thermal conductivity is thus a length-dependent property when ballistic effects
are present. Ballistic effects manifest themselves in two different ways: (i) boundary
temperature jump and (ii) reduced heat flux (i.e., reduced thermal conductivity).
Thermal transport in insulators and semiconductors is dominated by phonons, while
in metals, electrons are the primary carriers [7]. Because insulators and semiconductors
are integral to many nano-structured devices, substantial effort has gone into developing
Chapter 2. Thermal Transport Models 7
adequate theories of phonon transport. These efforts focused primarily on employing
the Boltzmann transport equation (BTE) to describe the statistical time evolution of
individual phonon modes. Because previous works typically made major approximations
and assumptions regarding the nature of phonon dynamics to reduce the computational
effort, much of this work is qualitative or semi-empirical. In Chapters 4 and 5, the suit-
ability of some of these approximations/assumptions is investigated using a quantitative
analysis.
2.2 Boltzmann Transport Equation
Different numerical methods have been proposed in the literature for solving the BTE to
model phonon transport. Joshi and Majumdar developed an equation for phonon radia-
tive transfer that employed average phonon properties [24]. Chen proposed a Ballistic-
Diffusive equation for phonon transport, where phonons are defined as transporting in
either ballistic or diffusive regimes [25, 26]. Narumanchi et al. employed the finite vol-
ume method to solve the BTE for thermal transport modelling in electronic devices
using mode-dependent phonon properties [27, 28]. Monte Carlo simulations have also
been applied to solve the BTE and obtain phonon populations in nano-scale devices
[29, 30, 31, 32].
Another approach for solving the BTE, which is adopted in the present work, is
using the lattice Boltzmann method (LBM) [33]. The LBM is a particle-based method
that uses a finite differencing technique to solve the BTE numerically and predicts the
distribution of particles in discrete directions at discrete points in time and space. The
LBM method is discussed in detail in Sec. 2.3, but first, the BTE for phonon transport
modelling is introduced.
The time-dependent phonon BTE is given by [34]
∂f(κ, ν)
∂t+ vg(κ, ν) · ∇f(κ, ν) =
[∂f(κ, ν)
∂t
]coll
, (2.4)
Chapter 2. Thermal Transport Models 8
where f(κ, ν) is the phonon population. Here, each phonon mode is identified by its
wave vector, κ, and dispersion branch, ν. The phonon group velocity vector, vg(κ, ν), is
vg(κ, ν) =∂ω(κ, ν)
∂κ, (2.5)
where ω(κ, ν) is the phonon frequency. The left hand side of Eq. (2.4) describes the
diffusion of a system of non-interacting phonons. The term on the right hand side is the
collision operator and serves to reestablish equilibrium through phonon scattering. The
phonon modes are coupled through the collision operator.
The overriding challenge in solving the BTE is modelling the collision operator. Al-
though methods have been developed to evaluate it directly [35, 36, 37], here the re-
laxation time approximation is used to make the BTE more tractable [16, 38]. Using a
first-principles iterative approach to directly evaluate the collision operator and predict
the bulk thermal conductivity of silicon and germanium, Ward and Broido found that
the relaxation time approximation introduces 5 − 10% error (for temperatures between
100 and 800 K [37]).
Under the relaxation time approximation, phonon transport is described by a set of
mode-dependent relaxation times, τ(κ, ν), defined as the average time between scattering
events. Throughout this thesis the crystal is assumed to contain no defects, no free elec-
trons, and no internal interfaces so that τ(κ, ν) is equal to the relaxation time associated
with phonon-phonon scattering, τp−p(κ, ν). Under the relaxation time approximation the
collision term is modelled as[∂f(κ, ν)
∂t
]coll
=fBE(κ, ν)− f(κ, ν)
τp−p(κ, ν), (2.6)
where fBE(κ, ν) is the Bose-Einstein (equilibrium) distribution function,
fBE(κ, ν) =1
eχ − 1, (2.7)
and χ ≡ ~ω(κ, ν)/kBT , ~ is the Planck constant divided by 2π, kB is the Boltzmann
constant, and T is the absolute temperature. Combining Eqs. (2.4) and (2.6) yields the
Chapter 2. Thermal Transport Models 9
phonon BTE under the relaxation time approximation,
∂f(κ, ν)
∂t+ vg(κ, ν) · ∇f(κ, ν) =
fBE(κ, ν)− f(κ, ν)τp−p(κ, ν)
. (2.8)
For bulk crystals, the steady state analytical solution of Eq. (2.8) can be combined with
the Fourier law to develop an expression for the i component of thermal conductivity
[34, 39],
ki =∑ν
∑κ
cph(κ, ν)v2g,i(κ, ν)τp−p(κ, ν). (2.9)
Here, the phonon specific heat, cph(κ, ν), is [34, 39]
cph(κ, ν) =~ω(κ, ν)
V
∂fBE(κ, ν)
∂T=kBχ
2
V
eχ
[eχ − 1]2, (2.10)
where V is the volume of the computational cell used to predict the phonon properties
(see Sec. 4.2.2), and vg,i(κ, ν) is the i component of the group velocity vector.
2.3 Solving the Boltzmann Transport Equation us-
ing the Lattice Boltzmann Method
2.3.1 Lattice Boltzmann Methodology
The LBM is a discrete representation of the time-dependent BTE under the relax-
ation time approximation [i.e., Eq. (2.8)], which was first introduced as a fluid flow
modelling technique through prediction of the distribution of fictitious fluid particles
[33]. Since then, the LBM has also been employed to model transport of electrons and
phonons. Guyer [40] introduced a lattice Boltzmann computational framework for mod-
elling phonon transport using a two-dimensional hexagonal (D2Q7 ) lattice. In the present
work, the lattice labelling follows conventions introduced by Qian et al. [41]; i.e., DnQm
where n is the dimensions of space and m-1 is the number of directions on each lattice
site. Jiaung and Ho employed a square (D2Q9 ) lattice to model phonon transport [42].
Chapter 2. Thermal Transport Models 10
Ghai et al. used the LBM for coupled modelling of phonons and electrons in semicon-
ductor materials and metals [43]. The LBM for phonons was further developed in the
works of Escobar et al. [44, 45, 46] and Goicochea et al. [47, 48]. These latter works
employed the D2Q9 lattice with multiple phonon modes. Thouy et al. [49, 50] performed
single- and multi-mode phonon LBM simulations on a D2Q9 lattice; in order to have a
higher resolution for direction sampling, they further developed their model to include
24 different directions using a D2Q25 lattice. Recently, Christensen and Graham [51]
developed a coupled LBM/finite difference method to model the heat transfer in joint
micro- and macro-domains.
The LBM for phonon transport is divided into two main categories, namely: multi-
mode LBM and single-mode LBM. Multi-mode LBM considers mode-dependent phonon
properties and solves the BTE for each phonon mode. The single-mode LBM for phonon
transport considers average phonon properties and, like its name suggests, solves the
BTE to predict the phonon populations of only one representative phonon mode. In other
words, the Gray approximation is made. Multi-mode LBM can be considered as solving
a single-mode LBM (herein called Gray LBM) for multiple phonon modes and requires
modifications in numerical implementation to include energy coupling between modes.
Since the objective of this chapter is to discuss LBM methodology and implementation for
phonon transport, the Gray approximation has been adopted for the sake of simplicity
and clarity. Details regarding multi-mode LBM using phonon properties predicted by
lattice dynamics calculations are discussed in Chapter 4.
Under the Gray approximation, the phonon BTE [Eq. (2.8)] is not mode-dependent
and can be re-written as:
∂f
∂t+ vg.∇f =
fBE − fτ
. (2.11)
The first-order discretization of Eq. (2.11) in time and in the x direction results in the
discrete lattice Boltzmann equation (LBE) [33]:
fi(x+ ∆x, t+ ∆t)− fi(x, t) =∆t
τ
[fBEi (x, t)− fi(x, t)
], (2.12)
Chapter 2. Thermal Transport Models 11
2 13
1 2 3
4 5 6
1 2
3 4
5 6
7 8
Figure 2.1: D1Q3, D2Q7, and D2Q9 lattice structures and respective labelling. Reprinted
with permission from Ref. [1], Copyright 2011 Elsevier.
where subscript i represents the discretized directions on each lattice site. fi is the
corresponding direction-wise phonon population (i.e., the left and right propagation di-
rections for one-dimensional phonon transport) and ∆t is the computational time step.
The lattice spacing is related to the time step by ∆x = vg∆t.
For one-dimensional simulations, the D1Q3 lattice geometry is straightforward to
implement and is shown in Fig. 2.1. Any lattice selection in two dimensions should be
plane-filling. Two possible lattice configurations that have been extensively employed for
the fluid flow simulation using the LBM are the D2Q7 and D2Q9 lattice structures, and
are shown in Fig. 2.1.
At each time step during the LBM simulation, the phonon populations move in the
direction of the assigned discrete velocity set (along the arrows in Fig. 2.1) towards the
next lattice site (streaming step); and based on a set of collision rules, new populations
are calculated at each lattice site for the next time step (collision step). The effect
of boundary conditions and external sources are also included in the computation of
new components of phonon population on each lattice site at each time step during the
collision step.
For all the lattice structures discussed in this work, the phonon population at the
center (represented by a dot in Fig. 2.1) is reserved to keep the summation of all the
Chapter 2. Thermal Transport Models 12
123
654
123
654
123
654
123
654
123
654
constant j
constant i
∆x
∆y
x
y
Figure 2.2: D2Q7 lattice implementation and boundary lattice nodes. Reprinted with
permission from Ref. [1], Copyright 2011 Elsevier.
phonon populations (i.e., lattice energy density) on that lattice site,
fm(x, t) =m−1∑i=1
fi(x, t), (2.13)
which can then be used to calculate the temperature using Bose-Einstein statistics. De-
tails regarding LBM implementation are found in Appendix A.
2.3.2 Two Dimensional Lattice Structures: D2Q7 vs. D2Q9
All previous works on the phonon LBM modelling in two dimensions, with the exception
of Guyer [40], have preferred the D2Q9 over the D2Q7 lattice for its simple structure
and more straightforward boundary condition implementation [42, 46, 49, 52, 51]. The
inherent problem with the D2Q9 lattice, however, is that the particles on diagonal di-
rections (directions 5, 6, 7, and 8 in Fig. 2.1) travel a distance√
2 times longer than the
distance travelled by the particles on the main directions (directions 1, 2, 3, and 4 in Fig.
2.1) at each time step. In the LBM for fluid flow simulations, this numerical problem
has been accounted for by altering the direction-wise velocities of the fluid particles and
Chapter 2. Thermal Transport Models 13
introducing a weight factor for particles moving in different directions. In the Gray LBM
for phonon transport, however, only one representative mode with average phonon prop-
erties is considered. All phonons therefore have the same velocity, vg, and cannot travel
faster along the diagonals, as required for implementation of the D2Q9 lattice. We find
that the D2Q9 lattice thus introduces significant error when investigating ballistic trans-
port [1] because of the faster traveling phonons along the diagonals. This issue can be
resolved by using the D2Q7 lattice when simulating two-dimensional phonon transport
(see Fig. 2.2).
On a D2Q7 lattice, the particles move on six equi-length directions separated from
each other by an angle of π3
radians. The phonons therefore travel the same distance
during each time step (i.e., the phonons travel at the same velocity). Although using
the D2Q7 lattice resolves the issue of fictitious high speed diagonal phonons when using
the D2Q9 lattice, implementation of the D2Q7 lattice is complex. On the left and right
boundaries of the D2Q9 lattice, every other lattice site does not fall on the physical
domain boundaries (see Fig. 2.2). Boundary condition implementation at these two
boundaries should be treated with extra care (see A.2 and A.3). Geometrical constraints
of the D2Q7 lattice implies that ∆y is√
32
∆x. It also should be noted that a constant i
line is a zigzag line, as shown in Fig. 2.2. Streaming rules for the D2Q7 lattice shown in
Fig. 2.2 are presented in A.1.
Chapter 3
Comparing Fourier-based and
BTE-based Predictions
3.1 Validating the Gray LBM in the Bulk Limit
Although later chapters will show that making the Gray approximation can lead to
significant error, it is made here so that: (i) the BTE-based predictions of thermal
transport can be characterized using simple algebraic formulations (this is impossible in
cases when dispersion is included) and (ii) the transport of individual phonon modes can
be described, the results of which will later be used to aid discussions when multiple
phonon modes are considered.
To validate the Gray LBM, steady state, diffuse-regime (i.e., bulk-like) heat transfer
in a square domain with side length L = 5 µm is modelled using the Gray LBM and
the D2Q7 and D2Q9 lattices. The predicted results are then compared to an analytical
solution to the two-dimensional Fourier heat equation. The model material is silicon.
The bulk phonon properties of silicon are presented in Table 3.1.
The entire domain was initialized to a constant temperature of Tc = 299.5 K, whereas
the temperature on the top boundary (y/L = 1) was maintained at a temperature of Th =
14
Chapter 3. Comparing Fourier-based and BTE-based Predictions 15
0 0.2 0.4 0.6 0.8 1.00
0.1
0.2
0.3
0.4
0.5
0.6
0.7
x*
T*
AnalyticalD2Q7D2Q9
y/L= 0.50
y/L= 0.25
y/L= 0.75
(a) Temperature
0
1
2
3
4
5
6
q y (x
10 W
/m )2
AnalyticalD2Q7D2Q9
y/L= 0.25
y/L= 0.50
y/L= 0.75
7
(b) Heat Flux
0 0.2 0.4 0.6 0.8 1.0 x*
Figure 3.1: (a) Temperature and (b) heat flux profile predicted by the analytical Fourier-
based solution (solid line) and the Gray LBM using the D2Q7 (open circles) and D2Q9
(open triangles) lattices for the diffuse regime. Reprinted with permission from Ref. [1],
Copyright 2011 Elsevier.
300.5 K during the simulation. Domain initialization and implementation of boundary
conditions for the Gray LBM are explained in A.2 and A.3. The steady state temperature
and heat flux profiles in the y-direction (for three different y/L values: 0.25, 0.50, and
0.75), calculated using the LBM on the D2Q7 and D2Q9 lattices, are compared in Fig.
3.1 to the values determined by an analytical solution of the Fourier heat equation [53].
T ∗ and x∗ are defined as (T − Tc)/(Th − Tc) and x/L, respectively. It can be seen
that for the steady state diffuse regime, both lattice structures perform well and the
predicted results are in good agreement with the analytical Fourier-based solution. The
heat flux predicted by the D2Q9 lattice is larger than those predicted from the Fourier
heat equation because of the artificially-fast phonons travelling in the diagonal direction
on the D2Q9 lattice, and suggests that some sub-continuum effects remain even for this
relatively large system. The difference between the heat flux results of the D2Q7 lattice
and Fourier heat equation are attributed to the existence of minor sub-continuum effects.
A lattice resolution study of the steady state results for the temperature profile (a
Chapter 3. Comparing Fourier-based and BTE-based Predictions 16
Table 3.1: Phonon properties for bulk silicon at temperature of 300 K. Reprinted with
permission from Ref. [1], Copyright 2011 Elsevier.
Phonon mean free path (Λ) 40× 10−9 m
Phonon group velocity (vg) 6733 ms−1
Phonon frequency (ω) 8.1825× 1013 rads−1
Specific heat (cv) 1.66× 106 Jm−3K−1
Density (ρ) 2328 kgm−3
Bulk thermal conductivity (k) 149 WK−1m−1
scalar field) showed that at least 300 lattice sites on each side of the domain are required
to have lattice-independent steady state simulations. It should be noted that, as a
rule of thumb, the lattice resolution required to obtain lattice-independent steady state
results for the heat flux (a vector field) or transient results for either temperature or
heat flux is about 20 times higher than that required for capturing the correct steady
state temperature profile. This lattice resolution corresponds to about 50 lattice sites
per physical MFP when using the Gray LBM. A longer simulation time was required to
obtain a steady state heat flux than a steady state temperature profile. This trend has
also been reported in the molecular dynamics simulations [54].
3.2 Modelling Sub-Continuum Thermal Transport in
Thin Films: A Steady-State Comparison
Thermal transport in micro- and nano-structured materials can be significantly different
than in bulk. As dimensions are reduced, boundary effects become significant and prop-
erties predicted for the bulk phase (e.g., thermal conductivity, phonon relaxation times)
may not be suitable for modelling thermal transport. One example is the semiconductor
Chapter 3. Comparing Fourier-based and BTE-based Predictions 17
z, Cross-Plane
x, In-Plane LPhonon-Phonon
Scattering
Diffuse Phonon-Boundary
Scattering
Perio
dic
Boun
dary
Perio
dic
Boun
dary
Figure 3.2: Thin film model.
thin film, which is common in microprocessors, solid-state memory, and semiconductor
light-emitting diodes [7, 8, 55]. The reduced thermal conductivity of thin films limits
their ability to effectively remove waste heat, which is critical to device operation and
reliability [7, 8, 9].
Consider a thin film oriented such that the cross-plane direction and thermal gradient
are along the z direction, as shown in Fig. 3.2. Phonon transport within this film is first
predicted using the Gray LBM to solve the BTE. The hot and cold boundaries of the thin
film are set to temperatures of 300.5 and 299.5 K, resulting in a temperature difference
of 1 K and an average temperature of 300 K. All phonon properties correspond to a
temperature of 300 K and are presented in Table 3.1. Fixed temperature boundary
conditions are imposed using the Bose-Einstein distribution to calculate the phonon
populations at the system boundaries. This boundary condition ensures that all phonons
scatter when they interact with the system boundaries (i.e., diffuse scattering). The use
of completely diffuse boundaries is based on the idea that the reconstruction of free silicon
surfaces disrupts phonons traveling in the cross-plane direction. Similarly, for thin films
bounded by an amorphous material (such as a silicon thin film bounded by amorphous
silica layers), the transition from a crystalline to an amorphous material presents a large
disruption to the phonon propagation and will diffusely scatter the majority of incident
phonons [11].
The LBM-predicted temperature profile across a silicon thin film is presented in
Chapter 3. Comparing Fourier-based and BTE-based Predictions 18
0.8
∆T
(a)
(b)
1.0
0.6
0.80
0.2
0.4
1.00 0.2 0.4 0.6
0.8 1.00 0.2 0.4 0.6
0.8
Non
-dim
ensi
onal
1.0
0.6
0
0.2
0.4
∆T
∆T
∆T
∆T
∆T
film
film
left
left
right
right
∆L=Λ
Non-dimensional Position Along Film, z/L
Non-dimensional Position Along Film, z/L
Tem
pera
ture
, T*
Non
-dim
ensi
onal
Tem
pera
ture
, T*
0.2- 1.2
∆L=Λ
Figure 3.3: (a) Steady-state temperature profile in a silicon film with constant temper-
ature boundary conditions predicted using the Gray LBM, Kn = 0.33. (b) Steady-state
temperature profile predicted using the Fourier-based heat equation for a silicon film with
the domain extended by 2Λ.
Fig. 3.3(a). Of particular interest are the temperature discontinuities at the boundaries
(∆Tleft and ∆Tright), which are an indication of ballistic phonon transport. Such discon-
tinuities are negligible in bulk systems (i.e., ∆Tright,∆Tleft → 0 as L→∞) and are not
predicted using continuum-based analysis techniques (e.g., the Fourier heat equation).
At steady state, ∆Tleft = ∆Tright (herein referred to as Tjump).
In the ballistic (sub-continuum) regime, energy carriers travel from one side of the film
to the other without scattering when the characteristic size of a system, L, becomes com-
Chapter 3. Comparing Fourier-based and BTE-based Predictions 19
parable to the phonon MFP, Λ(κ, ν) = |vg(κ, ν)|τ(κ, ν). For convenience, the phonon
MFP is compared to the system length using the Knudsen number:
Kn(κ, ν) =|vg(κ, ν)|τ(κ, ν)
L=
Λ(κ, ν)
L. (3.1)
3.2.1 Predicting Thermal Conductivity from LBM Results
From the phonon populations predicted by the LBM, the cross-plane thermal conductiv-
ity, kz(L), can be predicted using a one-dimensional form of the Fourier Law:
kz(L) =qzL
∆Tleft + ∆Tfilm + ∆Tright, (3.2)
where qz is the z component of the heat flux vector in the thin film (qx = qy = 0),
qz =1
V
∑ν
∑κ
vg,z(κ, ν)~ω(κ, ν) [fL(κ, ν)− fR(κ, ν)] , (3.3)
where fL and fR are the populations of the phonon modes travelling in the left and right
propagation directions [56]. The temperature differences ∆Tleft, ∆Tfilm, and ∆Tright, are
calculated from the steady state temperature profile.
The thermal conductivity reduction from bulk predicted using the Gray LBM is plot-
ted versus Kn (open circles) in Fig. 3.4(a). The thermal conductivity increases mono-
tonically to the bulk value as the system length increases (Kn decreases). The length-
dependence of the steady state boundary temperature jump, Tjump, normalized by the
total temperature difference across the film ∆Tbulk = ∆Tleft + ∆Tfilm + ∆Tright, is pre-
sented in Fig. 3.4(b). At small Knudsen numbers (large length scales), the temperature
jumps at the system boundaries diminish. Reducing the domain size increases the tem-
perature jump to an asymptotic value of Tjump/∆Tbulk = 0.5 at the pure ballistic regime.
Chapter 3. Comparing Fourier-based and BTE-based Predictions 20
k
/k
bulk
Gray LBM0.6
0.4
0.8
0.2
1.0
102- 1- 0 1Kn
Tju
mp
/∆T b
ulk
10 1010
0.5
0.4
0.3
0.2
0.1
0
(a) (b)
102- 1- 0 1Kn
10 10100
film
k
bulk
film k =
11+2Kn
Gray LBM T
bulk=
Kn1+2Kn
jumpT∆
Figure 3.4: (a) Effective thermal conductivity normalized by the bulk value (kfilm/kbulk)
as a function of the Knudsen number (Kn) predicted by the Gray LBM and Eq. (3.8). (b)
Steady-state boundary temperature jump normalized by the bulk value (Tjump/∆Tbulk)
as a function of the Kn predicted using the Gray LBM and Eq. (3.10).
3.2.2 Modelling Boundary Scattering in Thermal Conductivity
Predictions
One method to model boundary scattering is to solve the time-dependent BTE numeri-
cally using the LBM and appropriate boundary conditions (as described in the previous
section). A second method is to solve the steady state BTE analytically for the bulk
thermal conductivity [resulting in Eq. (2.9)] and then correct the bulk phonon-phonon
relaxation times to account for phonon-boundary scattering. To do this, phonon scat-
tering mechanisms (i.e., phonon-phonon and phonon-boundary scattering) are combined
by assuming that they are independent using the Matthiessen rule [11, 34, 3].
The Matthiessen rule can be used to model the length dependence of τ(κ, ν, L) such
that
1
τ(κ, ν, L)=
1
τ∞(κ, ν)+
1
τb(κ, ν, L), (3.4)
where τ∞(κ, ν) and τb(κ, ν, L) are the intrinsic scattering and boundary scattering re-
laxation times. Again, the crystal is assumed to contain no defects, no free electrons,
and no internal interfaces so that τ∞(κ, ν) is equal to the relaxation time associated with
Chapter 3. Comparing Fourier-based and BTE-based Predictions 21
phonon-phonon scattering τp−p(κ, ν). The boundary scattering relaxation time is taken
to be the average time between boundary scattering events in the absence of intrinsic
scattering, i.e.,
τb(κ, ν, L) =L/2
|vg,j(κ, ν)|, (3.5)
where |vg,j(κ, ν)| is the component of the velocity vector that is perpendicular to the film
surface.
Substituting Eq. (3.4) into Eq. (2.9) and applying Eq. (3.5) leads to an expression
that describes the length dependence of thermal conductivity,
ki(L) =∑ν
∑κ
cphv2g,i(κ, ν)τ∞(κ, ν)
[1 +
2|vg,j(κ, ν)|τ∞(κ, ν)L
]−1. (3.6)
For cross-plane phonon transport i = j, for in-plane phonon transport i 6= j. As L ap-
proaches infinity, the bracketed term in Eq. (3.6) approaches unity and ki(L) approaches
the bulk value, k∞. Equation (3.6) considers mode-dependent phonon properties. To
compare Eq. (3.6) to the Gray LBM results presented in Fig. 3.4(a), the results are sim-
plified by considering only one phonon mode traveling perpendicular to the film boundary
(i.e., the cross-plane direction). Under these conditions, Eq. (3.6) becomes:
k(L) = cphv2gτ∞
[1 +
2|vg|τ∞L
]−1= k∞
1
1 + 2Kn, (3.7)
or in a more convenient form,
kfilmkbulk
=1
1 + 2Kn. (3.8)
In the LBM, phonon-phonon scattering is modelled using τp−p while phonon-boundary
scattering is modelled using appropriate boundary conditions. These scattering mecha-
nisms are thus treated independently. Agreement between results of the Gray LBM and
Eq. (3.8) is thus expected (provided that the boundary scattering is consistently incor-
porated in both methods) because the underlying treatment that scattering mechanisms
are independent is consistent. In Fig. 3.4(a), the thermal conductivity reduction pre-
dicted using Eq. (3.8) is presented (solid line) with the results obtained using the Gray
Chapter 3. Comparing Fourier-based and BTE-based Predictions 22
LBM (open circles). The results are in excellent agreement (within 1% for all film thick-
nesses). The agreement between the two methods suggests that Eq. (3.5) is appropriate
for modelling diffuse boundaries. Mode-dependent results are presented in Chapter 4.
3.2.3 Modelling Sub-Continuum Thermal Transport using Fourier-
based Methods
The conventional methods for modelling bulk thermal transport, such as the Fourier heat
equation, fail to capture sub-continuum effects. To address this issue, we have shown
that numerical methods based on the BTE can be used to accurately model phonon
transport. The complex nature and relatively high computational costs of these methods,
however, have hindered their adoption by industry. Furthermore, the wide selection of
commercially available Fourier-based packages (e.g., ANSYS), which are more efficient
at large length scales, makes these sub-continuum techniques a hard sell to industry. A
method to modify the Fourier-based heat equation so that it is capable of capturing sub-
continuum effects is therefore desirable in many industrial applications. In this section,
the Gray LBM results are used to develop a method to modify the Fourier-based heat
equation so that it can recover sub-continuum effects: (i) reduced heat flux and (ii)
boundary temperature jump.
Modifying the Fourier-based heat equation to recover the heat flux reduction in bal-
listic transport is straightforward and can be obtained by using the reduced thermal
conductivity predicted by Eq. (3.8). A method to recover the temperature jumps at the
boundaries, however, requires more effort.
To derive an analytical form for the length-dependence of Tjump shown in Fig. 3.4(b),
we begin with the following formulation that was obtained by analyzing the Gray LBM
results:
q = kfilm∆Tbulk
∆x= kbulk
∆Tfilm∆x
. (3.9)
Chapter 3. Comparing Fourier-based and BTE-based Predictions 23
Combining Eqs. (3.8) and (3.9) and using ∆Tbulk = 2Tjump + ∆Tfilm at steady state
results in a relation for the non-dimensional steady state boundary temperature jump:
Tjump∆Tbulk
=Kn
1 + 2Kn. (3.10)
In Fig. 3.4(b) the length-dependence of the boundary temperature jump predicted using
Eq. (3.10) is presented (solid line) with the results obtained using the Gray LBM (open
circles). The results are in excellent agreement (within 1% for all film thicknesses).
From the Gray LBM simulation results, we know that the steady state tempera-
ture jump at each boundary is equal. One can obtain these steady state temperature
jumps at the physical boundaries of the system using the Fourier-based heat equation
(which predicts a linear temperature profile) by extending the domain beyond the physi-
cal boundaries by a fixed value, ∆L, and implementing the physical boundary conditions
at the end of these added imaginary sections. In this case, the steady state solution is a
linear temperature profile extended across the new length of the domain, L+2∆L. Using
the proposed relation for the steady state temperature jump [Eq. (3.10)] and trigonom-
etry, we obtain ∆L = Λ. Thus, by adding extensions of ∆L = Λ to either side of the
physical domain, the Fourier-based heat equation predicts the correct temperature jumps
at the physical system boundaries. This method is more practically implemented than
using a thermal boundary resistance. This result is shown in Fig. 3.3(b) for a silicon thin
film with length such that Kn = 0.33.
3.2.4 Summary
The LBM is used to numerically solve the BTE under the relaxation time approximation.
Effort was made throughout this chapter to characterize the LBM solutions using alge-
braic equations, which are presented for the first time. These relations were then used
to develop a method to modify Fourier-based heat equations so that they reproduce the
LBM data when ballistic transport is present. The steady state heat flux reduction due
Chapter 3. Comparing Fourier-based and BTE-based Predictions 24
to ballistic transport predicted by the Gray LBM can be captured using the Fourier heat
equation and an effective thermal conductivity. To obtain the temperature jumps at the
system boundaries predicted by the Gray LBM, the system domain should be extended
by a distance equal to the phonon MFP on either side.
3.3 Heated Surface: Transient Comparison
In the previous section, steady state thermal transport is addressed. In many applica-
tions, such as intrinsic heating in graphics processing units, understanding the transient
behaviour of thermal transport is important to developing appropriate cooling strategies.
As mentioned in Sec. 2.1, the Fourier heat equation predicts that energy is conducted
away from a heated region at an infinite propagation speed. This problem can lead to
significant error when investigating systems at small time scales where the thermal en-
ergy of a hot region has not yet reached a cold region. Traditionally, the hyperbolic heat
equation is used to resolve wave effects in these systems by forcing energy to propagate at
a defined speed. To illustrate this case, the transient data corresponding to the geometry
of the steady state discussed above is presented. The Gray LBM is used to simulate
phonon transport in a silicon film with a thickness such that Kn = 0.33 and the initial
temperature is equal to the right boundary, Tc = 299.5 K. The left boundary is main-
tained at Th = 300.5 K during the simulation. This system results in a thermal wave
propagating from the hot, left boundary to the cold, right boundary when the simulation
is allowed to proceed.
The transient ballistic temperature profile predicted by the Gray LBM is presented
as a solid line in Fig. 3.5(a) for t∗ = 1, where t∗ = t/τ is the non-dimensional time. The
sharp wave front predicted by the Gray LBM is not physical and is due to the inclusion
of only one average phonon mode. If multiple phonon modes are included in the LBM
simulations, the predicted wave front would be less sharp. Figure 3.5(b) shows the time
Chapter 3. Comparing Fourier-based and BTE-based Predictions 25
0 5 10 15 200
0.1
0.2
0.3
0.4
0.5
*t
T jum
p/ ∆
T bul
k
Left BoundaryRight Boundary
T*
Modified Hyperbolic
0.4
0.2
0.2 0.40
0 0.6 0.8 1.0
0.8
1.0
0.6
*x
Gray LBM
Gray LBM
Tjj
0 2 4 6 8 10 12 141/Kn
0
0.1
0.2
0.3
0.4
0.5
T jj
(c)
(b)
(a)
T = 0.5ejj- 0.5Kn
Gray LBM
Figure 3.5: (a) Transient ballistic temperature profile predicted by the Gray LBM and the
hyperbolic heat equation with Kn = 0.33 and t∗ = 1. (b) Time-dependent temperature
jump on the left and right boundaries of a silicon film predicted by the Gray LBM, Kn
= 0.33. (c) Jump in the right boundary temperature, Tjj, when the thermal wave hits
the cold right boundary as a function of 1/Kn.
Chapter 3. Comparing Fourier-based and BTE-based Predictions 26
evolution of the temperature jump at the two boundaries. The temperature at the right
boundary remains unchanged until the heat wave travels the entire film thickness, which
is equal to 3Λ, and reaches the right boundary at t∗ = 3. Once the thermal wave hits
the cold end, the temperature jump at the right boundary follows the same trend as the
temperature jump on the left boundary. In other words, the curves are symmetric about
the steady state, asymptotic value, which is in excellent agreement with Eq. (3.10). The
jump in the right boundary temperature, Tjj, when the thermal wave hits the cold right
boundary is a function of the system length and is well described by the following relation
[see Fig. 3.5(c)]:
Tjj = 0.5e− 0.5L
Λ = 0.5e−0.5Kn . (3.11)
To model the transient ballistic thermal transport using the hyperbolic heat equation,
the time dependent temperature jumps predicted by the Gray LBM [see Fig. 3.5(b)]
were implemented numerically at the system boundaries. The temperature of the left
and right boundaries were set equal to Tleft(t∗) = Th − Tjump,left(t∗) and Tright(t∗) =
Tc + Tjump,right(t∗). The hyperbolic heat equation was then solved numerically using the
finite difference scheme. The predicted temperature profile of the modified hyperbolic
heat equation is presented as a dashed line in Fig. 3.5(a). It can be seen that the
modified hyperbolic equation can recover the temperature profile predicted from the
LBM simulation to a good extent when the transient boundary temperatures are known.
Providing a general relation for the transient boundary temperature, however, is difficult
because they depend strongly on the initial condition and system geometry, which offers
opportunity for future investigation.
Chapter 3. Comparing Fourier-based and BTE-based Predictions 27
3.4 Fourier-based vs. BTE Solutions: A Frequency-
domain Thermoreflectance Study
An increased interest in measuring the thermal transport properties of nanostructured
materials has led to the development of noncontact measurement techniques based on
photothermal phenomena. One such technique is the pump-probe-based frequency-
domain thermoreflectance (FDTR) method [57, 58]. In one form of FDTR [58], the
continuous wave (CW) pump laser is modulated and used to periodically heat the sur-
face of a sample. This periodic surface heating results in a periodic thermal wave that
propagates into the sample. The amplitude and phase of the thermal response at the
sample surface depend on the thermal properties of the sample beneath. Because the
reflectance of the sample surface is temperature dependent, the probe laser (also CW)
becomes modulated upon reflection and is used to monitor the thermal response. A
Fourier-based thermal model is then fit to phase response and/or amplitude data of the
reflected probe beam to extract the sample thermal conductivity or the thermal conduc-
tance of a buried interface.
The goal of this work is not to simulate a real FDTR experiment, but to assess the
Fourier-based model used to derive the data analysis method described in Ref. [59]. The
Fourier-based solution for a homogeneous solid is evaluated. To do so, phonon transport
is simulated using the Gray LBM and the resulting temperature profiles are compared
to those predicted by a solution to the Fourier-based conduction equation [53]. To be
consistent with the implementation of Ref. [53], thermal transport in a homogeneous,
semi-infinite solid that is heated at the surface by a periodic source is considered. In
addition to simulating bulk-like conditions, we also assess the ability for the Fourier-
based model to capture thermal transport when sub-continuum effects are present. To
do this, sinusoidal heat fluxes are chosen such that the modulation frequency of the
propagating thermal wave is comparable to the phonon relaxation time.
Chapter 3. Comparing Fourier-based and BTE-based Predictions 28
T = Constant
q =
q
�Penetration Depth,
Sam
ple
Surf
ace
Hot
Cold
Sample Thicknessx
zo
sin(ω
t)∞
Figure 3.6: Schematic diagram of the simplified frequency-domain thermoreflectance
setup. The sample is periodic in the x and y directions and infinite in the z direction.
Consider a bulk, homogeneous sample that is infinite in the x and y directions and
finite in the z direction, as shown in Fig. 3.6. Silicon is used as the model material. The
averaged, bulk phonon properties required as input to the BTE are provided in Table
3.1. The sample is initially set to a uniform temperature of Tinitial = 300 K. During the
simulation, a heat flux boundary condition [i.e., q = qo sin(ft)] is imposed at z = 0 such
that the surface-temperature oscillates at a frequency, f . The magnitude of the heat flux,
qo, is chosen such that the maximum surface temperature, Tmaxsurface, is constant and equal
to 305 K. The boundary opposite to the heated surface is maintained at a temperature of
T∞ = 300 K. To ensure that the sample is semi-infinite, the sample thicknesses is chosen
such that the propagating thermal wave decays to a temperature of 300 K well before
it reaches the boundary opposite of the heated surface. The system is allowed to evolve
in time, and the non-dimensional, steady state temperature profiles are reported. The
non-dimensional temperature, T ∗, is defined as: T ∗(z) = T (z)−T∞Tmaxsurface−T∞
.
The temperature profiles predicted by the BTE are shown in Fig. 3.7 for non-dimensional
frequency fτ = 0.025. The data was recorded after the system reached steady state. As
shown in the insert of Fig. 3.7, which is a zoom-out of the main plot, the temperature
data is symmetrical about T ∗ = 0. Each grey line corresponds to the temperature profile
Chapter 3. Comparing Fourier-based and BTE-based Predictions 29
fτ = 0.025 �/Λ = 9.2
0
0.2
0.4
0.6
0.8
1.0
Non
-dim
ensi
onal
Tem
pera
ture
, T
*
Non-dimensional Depth, z/Λ0 20 40 60 80
Analytical Fourier Solution,T = exp(-z/�)
z
1000z/Λ
−1
0
1
T*
BTE Solutions
=�
*
Figure 3.7: Non-dimensional temperature, T ∗, as a function of non-dimensional sample
depth, z/Λ, for a system with a surface temperature oscillating at a non-dimensional
frequency of fτ = 0.025. Each grey line corresponds to a given time step after the
system reached steady state.
predicted by the BTE at a given time step. The decay envelope of the temperature oscil-
lations is well described by an exponential decay (black line) predicted by an analytical
solution of the Fourier-based conduction equation [53],
T ∗ = exp (−z/L) , (3.12)
where L is the penetration depth (a fitting parameter) defined as the depth at which T ∗
= 1/e. For fτ = 0.025, the non-dimensional penetration depth is L/Λ = 9.2.
For each f investigated, an exponential curve is fit to the decay envelope of the
BTE temperature data (which are all similar with that shown in Fig. 3.7) to extract the
corresponding penetration depth. This BTE data is plotted as squares for L/Λ versus
fτ in Fig. 3.8. When fτ � 1, the oscillatory period of the propagating thermal wave
is much larger than the phonon relaxation time. In this continuum regime, the phonons
react almost instantaneously to the oscillations of the propagating thermal wave, and
the BTE-predicted penetration depths are well-described by an analytical solution of the
Chapter 3. Comparing Fourier-based and BTE-based Predictions 30
0.001 0.01 0.1 1 10 1001
10
100
Analytical Fourier Solution,
�= BTE Solution
�
� = ρc k 0.5[ ][ ]
pπfbulk
Figure 3.8: Non-dimensional penetration depth, L/Λ, as a function of non-dimensional
frequency, fτ , for the BTE (squares) and analytical solution to the conduction equation
(solid line). The dashed line indicates L = 2Λ.
Fourier-based conduction equation [53],
L =[kbulkρcpπf
]0.5, (3.13)
which is plotted as a solid line in Fig. 3.8. Here, the 1/3 coefficient typically included
in the kinetic theory expression for kbulk is not included because it is associated with
three-dimensional phonon transport and only phonons that propagate perpendicular to
the heated surface are considered in the LBM implementation [1]. Thus, kbulk = cpvgΛ.
As fτ increases and becomes comparable to unity, the BTE-predicted penetration
depths deviate from the predictions of Eq. (3.13). When fτ > 1, the oscillatory period
of the propagating thermal wave is smaller than the phonon relaxation time and the
phonons do not react to the oscillations of the thermal wave. In this sub-continuum
regime, L/Λ no longer depends on fτ . In the limit that fτ approaches infinity, L/Λ
approaches 2 (indicated by the dashed line in Fig. 3.8).
In a typical FDTR experiment, f < 20 MHz. Therefore only phonons with τ > 50 ns
correspond to fτ > 1 and can transport ballistically. Ward and Broido [37] found that in
Chapter 3. Comparing Fourier-based and BTE-based Predictions 31
0
10
20
Req
uire
d H
eat F
lux
Am
plitu
de,
h [
GW
/m K
] 2
0
Analytical Fourier Solution,
� = ρc k 0.5[ ][ ]
pπf
0h =
BTE Solution
101
2π�
Non-dimensional Penetration Depth,
bulk
ρc pkbulk
Figure 3.9: Required heat flux amplitude, h0, as a function of non-dimensional pen-
etration depth, L/Λ, for the BTE (squares) and analytical solution to the conduction
equation (solid line).
bulk silicon only a few low-frequency phonons have τ > 50 ns at a temperature of 300 K.
These findings may explain why a recent transient thermoreflectance study [60], which
was performed using silicon at T = 300 K and modulation frequencies between 3 and 12
MHz, did not observe any laser modulation frequency dependence of the measured ther-
mal conductivity. The modulation frequencies were too small to allow enough phonons
to travel ballistically to significantly affect the measured thermal conductivity.
One method to force more phonons to travel ballistically is to increase the modula-
tion frequency beyond 20 MHz. To do so, however, may be difficult due to limitations
of available optical equipment. Another method to increase fτ is to decrease tempera-
ture. Decreasing temperature increases relaxation times, thus more phonons will travel
ballistically. If enough phonons travel ballistically, the implied thermal conductivity will
depend on the modulation frequency. This trend can be seen in Fig. 3.9, where the re-
quired heat flux amplitude, h0, is plotted as a function of non-dimensional penetration
depth, L/Λ, for the BTE (squares) and an analytical solution to the conduction equa-
tion (solid line). Note that the same surface-temperature amplitude is maintained for all
Chapter 3. Comparing Fourier-based and BTE-based Predictions 32
modulation frequencies studied (the surface-temperature oscillates between a maximum
of 305 K and a minimum of 295 K for all modulation frequencies). As the modulation
frequency increases, the penetration depth decreases, and the system transitions from
a diffusive to a ballistic regime when L becomes comparable to the phonon mean free
path (i.e., L/Λ < 1). In the ballistic regime, the heat flux required to heat the BTE
system is much less than that for the Fourier-based system. This reduced heat flux is
what researchers perceive as a reduced thermal conductivity in an FDTR experiment.
3.4.1 Summary
The phonon BTE is used to show that models based on the Fourier heat conduction
equation do not to accurately describe thermal transport when sub-continuum effects are
present in an FDTR-like system. In this work, phonon transport is simplified by solving
the BTE for one phonon mode with average properties (i.e., one relaxation time, one
group velocity, and one specific heat). In bulk silicon, however, phonons with relaxation
times that span 4 orders of magnitude are present [37, 3]. Performing an FDTR measure-
ment on this system would result in some phonons transporting in a continuum regime
while others transporting in a sub-continuum regime.
Chapter 4
Lattice Dynamics Calculations and
the BTE
4.1 Introduction
In the previous chapters, the Gray approximation is made for the sake of simplicity and
clarity in the attempt to elucidate the underlying physics of phonon transport in silicon
nanostructured materials. Although the Gray approximation is used extensively in liter-
ature, it is shown in this chapter that it over simplifies the phonon picture and phonon
dispersion (i.e., multiple phonon modes) must be considered to accurately describe ther-
mal transport in systems with nanoscale features.
In this chapter, multi-mode LBM is used to predict (i) the bulk thermal conductivity
of Stillinger-Weber (SW) silicon and (ii) the cross-plane thermal conductivities of SW
films as thin as 17.4 nm, all at a temperature of 300 K. Turney et al. previously analyzed
the in-plane thermal conductivity [11]. The accuracy of the BTE-based predictions de-
pend on the accuracy of the phonon properties required as input. Given an interatomic
potential that describes the atomic interactions, mode-dependent specific heats and group
velocities can be predicted using harmonic lattice dynamics calculations [61, 62]. The
33
Chapter 4. Lattice Dynamics Calculations and the BTE 34
Phonon Properties
Force ConstantsInteratomic
Harmonic &Anharmonic
Specific Heats,Velocities, &
Relaxation Times
Bulk
Thin Film (cross-plane)
ThermalConductivity
FourierLaw
TemperatureProfile &Heat Flux
Geometry &Boundary Conditions
Steady-State BTE& Fourier Law
LBM
Lattice Dynamics
Steady-State
Figure 4.1: Flow chart of the hierarchical procedure for predicting the phonon thermal
conductivity of bulk and thin films using lattice dynamics calculations and the BTE. The
theoretical/computational tools are in boxes and their inputs and outputs are in ovals.
Reprinted with permission from Ref. [3], Copyright 2010 American Institute of Physics.
required relaxation times can be predicted from molecular dynamics (MD) simulation
[63, 64] or anharmonic lattice dynamics calculations [62, 65]. Molecular dynamics simu-
lation and anharmonic lattice dynamics calculations, the latter of which naturally include
quantum effects and is used here, are computationally expensive and can be challenging
to implement. As such, the suitability of two approximations commonly made to reduce
the computational effort are investigated. First, the isotropic approximation, where the
phonon properties of one crystalline direction are assumed to be representative of the
entire Brillouin zone. Second, the Gray approximation, where the entire Brillouin zone
is represented by a single phonon velocity and relaxation time. By investigating these
approximations, we will (i) determine when they can be used to predict the phonon prop-
erties and thermal conductivity of bulk systems without introducing significant error, and
(ii) understand how their validity changes as system lengths are reduced from bulk to
the nanometer scale.
To perform the required calculations, a hierarchical procedure (described in Sec. 4.2.1)
that uses the BTE (Secs. 2.2 and 4.2.3) and lattice dynamics calculations (Sec. 4.2.2) is
Chapter 4. Lattice Dynamics Calculations and the BTE 35
used to predict the bulk and cross-plane thin film thermal conductivities. In Sec. 4.3, we
present the thermal conductivity predictions and examine the role of optical phonons,
the isotropic and Gray approximations, and the suitability of the Matthiessen rule for
combining the effects of different scattering mechanisms. It is shown that because the
frequency-dependent contributions to thermal conductivity change as the film thickness
is reduced, approximations that are valid for bulk are not necessarily valid for thin films.
4.2 Computational Tools
4.2.1 Overview of the Hierarchical Procedure
At the atomic level, the thermal conductivity of a semiconductor is related to the trans-
port of phonons. Phonon transport is modelled using the hierarchical procedure shown
in Fig. 4.1 and described in detail throughout this chapter. Interatomic forces, which
are the derivatives of a system’s potential energy with respect to the positions of its con-
stituent atoms, are first calculated from an interatomic potential energy function. The
force constants are then used in harmonic and anharmonic lattice dynamics