ViewVC Help
View File | Revision Log | Show Annotations | View Changeset | Root Listing
root/group/interfacial/interfacial.tex
Revision: 3760
Committed: Mon Aug 1 17:57:34 2011 UTC (12 years, 11 months ago) by skuang
Content type: application/x-tex
File size: 50337 byte(s)
Log Message:
small changes to a figure and text.

File Contents

# Content
1 \documentclass[11pt]{article}
2 \usepackage{amsmath}
3 \usepackage{amssymb}
4 \usepackage{setspace}
5 \usepackage{endfloat}
6 \usepackage{caption}
7 %\usepackage{tabularx}
8 \usepackage{graphicx}
9 \usepackage{multirow}
10 %\usepackage{booktabs}
11 %\usepackage{bibentry}
12 %\usepackage{mathrsfs}
13 %\usepackage[ref]{overcite}
14 \usepackage[square, comma, sort&compress]{natbib}
15 \usepackage{url}
16 \pagestyle{plain} \pagenumbering{arabic} \oddsidemargin 0.0cm
17 \evensidemargin 0.0cm \topmargin -21pt \headsep 10pt \textheight
18 9.0in \textwidth 6.5in \brokenpenalty=10000
19
20 % double space list of tables and figures
21 \AtBeginDelayedFloats{\renewcommand{\baselinestretch}{1.66}}
22 \setlength{\abovecaptionskip}{20 pt}
23 \setlength{\belowcaptionskip}{30 pt}
24
25 %\renewcommand\citemid{\ } % no comma in optional reference note
26 \bibpunct{[}{]}{,}{n}{}{;}
27 \bibliographystyle{achemso}
28
29 \begin{document}
30
31 \title{Simulating interfacial thermal conductance at metal-solvent
32 interfaces: the role of chemical capping agents}
33
34 \author{Shenyu Kuang and J. Daniel
35 Gezelter\footnote{Corresponding author. \ Electronic mail: gezelter@nd.edu} \\
36 Department of Chemistry and Biochemistry,\\
37 University of Notre Dame\\
38 Notre Dame, Indiana 46556}
39
40 \date{\today}
41
42 \maketitle
43
44 \begin{doublespace}
45
46 \begin{abstract}
47
48 With the Non-Isotropic Velocity Scaling algorithm (NIVS) we have
49 developed, an unphysical thermal flux can be effectively set up even
50 for non-homogeneous systems like interfaces in non-equilibrium
51 molecular dynamics simulations. In this work, this algorithm is
52 applied for simulating thermal conductance at metal / organic solvent
53 interfaces with various coverages of butanethiol capping
54 agents. Different solvents and force field models were tested. Our
55 results suggest that the United-Atom models are able to provide an
56 estimate of the interfacial thermal conductivity comparable to
57 experiments in our simulations with satisfactory computational
58 efficiency. From our results, the acoustic impedance mismatch between
59 metal and liquid phase is effectively reduced by the capping
60 agents, and thus leads to interfacial thermal conductance
61 enhancement. Furthermore, this effect is closely related to the
62 capping agent coverage on the metal surfaces and the type of solvent
63 molecules, and is affected by the models used in the simulations.
64
65 \end{abstract}
66
67 \newpage
68
69 %\narrowtext
70
71 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
72 % BODY OF TEXT
73 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
74
75 \section{Introduction}
76 Due to the importance of heat flow in nanotechnology, interfacial
77 thermal conductance has been studied extensively both experimentally
78 and computationally.\cite{cahill:793} Unlike bulk materials, nanoscale
79 materials have a significant fraction of their atoms at interfaces,
80 and the chemical details of these interfaces govern the heat transfer
81 behavior. Furthermore, the interfaces are
82 heterogeneous (e.g. solid - liquid), which provides a challenge to
83 traditional methods developed for homogeneous systems.
84
85 Experimentally, various interfaces have been investigated for their
86 thermal conductance. Cahill and coworkers studied nanoscale thermal
87 transport from metal nanoparticle/fluid interfaces, to epitaxial
88 TiN/single crystal oxides interfaces, to hydrophilic and hydrophobic
89 interfaces between water and solids with different self-assembled
90 monolayers.\cite{Wilson:2002uq,PhysRevB.67.054302,doi:10.1021/jp048375k,PhysRevLett.96.186101}
91 Wang {\it et al.} studied heat transport through
92 long-chain hydrocarbon monolayers on gold substrate at individual
93 molecular level,\cite{Wang10082007} Schmidt {\it et al.} studied the
94 role of CTAB on thermal transport between gold nanorods and
95 solvent,\cite{doi:10.1021/jp8051888} and Juv\'e {\it et al.} studied
96 the cooling dynamics, which is controlled by thermal interface
97 resistence of glass-embedded metal
98 nanoparticles.\cite{PhysRevB.80.195406} Although interfaces are
99 normally considered barriers for heat transport, Alper {\it et al.}
100 suggested that specific ligands (capping agents) could completely
101 eliminate this barrier
102 ($G\rightarrow\infty$).\cite{doi:10.1021/la904855s}
103
104 Theoretical and computational models have also been used to study the
105 interfacial thermal transport in order to gain an understanding of
106 this phenomena at the molecular level. Recently, Hase and coworkers
107 employed Non-Equilibrium Molecular Dynamics (NEMD) simulations to
108 study thermal transport from hot Au(111) substrate to a self-assembled
109 monolayer of alkylthiol with relatively long chain (8-20 carbon
110 atoms).\cite{hase:2010,hase:2011} However, ensemble averaged
111 measurements for heat conductance of interfaces between the capping
112 monolayer on Au and a solvent phase have yet to be studied with their
113 approach. The comparatively low thermal flux through interfaces is
114 difficult to measure with Equilibrium
115 MD\cite{doi:10.1080/0026897031000068578} or forward NEMD simulation
116 methods. Therefore, the Reverse NEMD (RNEMD)
117 methods\cite{MullerPlathe:1997xw,kuang:164101} would have the
118 advantage of applying this difficult to measure flux (while measuring
119 the resulting gradient), given that the simulation methods being able
120 to effectively apply an unphysical flux in non-homogeneous systems.
121 Garde and coworkers\cite{garde:nl2005,garde:PhysRevLett2009} applied
122 this approach to various liquid interfaces and studied how thermal
123 conductance (or resistance) is dependent on chemistry details of
124 interfaces, e.g. hydrophobic and hydrophilic aqueous interfaces.
125
126 Recently, we have developed a Non-Isotropic Velocity Scaling (NIVS)
127 algorithm for RNEMD simulations\cite{kuang:164101}. This algorithm
128 retains the desirable features of RNEMD (conservation of linear
129 momentum and total energy, compatibility with periodic boundary
130 conditions) while establishing true thermal distributions in each of
131 the two slabs. Furthermore, it allows effective thermal exchange
132 between particles of different identities, and thus makes the study of
133 interfacial conductance much simpler.
134
135 The work presented here deals with the Au(111) surface covered to
136 varying degrees by butanethiol, a capping agent with short carbon
137 chain, and solvated with organic solvents of different molecular
138 properties. Different models were used for both the capping agent and
139 the solvent force field parameters. Using the NIVS algorithm, the
140 thermal transport across these interfaces was studied and the
141 underlying mechanism for the phenomena was investigated.
142
143 \section{Methodology}
144 \subsection{Imposd-Flux Methods in MD Simulations}
145 Steady state MD simulations have an advantage in that not many
146 trajectories are needed to study the relationship between thermal flux
147 and thermal gradients. For systems with low interfacial conductance,
148 one must have a method capable of generating or measuring relatively
149 small fluxes, compared to those required for bulk conductivity. This
150 requirement makes the calculation even more difficult for
151 slowly-converging equilibrium methods.\cite{Viscardy:2007lq} Forward
152 NEMD methods impose a gradient (and measure a flux), but at interfaces
153 it is not clear what behavior should be imposed at the boundaries
154 between materials. Imposed-flux reverse non-equilibrium
155 methods\cite{MullerPlathe:1997xw} set the flux {\it a priori} and
156 the thermal response becomes an easy-to-measure quantity. Although
157 M\"{u}ller-Plathe's original momentum swapping approach can be used
158 for exchanging energy between particles of different identity, the
159 kinetic energy transfer efficiency is affected by the mass difference
160 between the particles, which limits its application on heterogeneous
161 interfacial systems.
162
163 The non-isotropic velocity scaling (NIVS) \cite{kuang:164101} approach
164 to non-equilibrium MD simulations is able to impose a wide range of
165 kinetic energy fluxes without obvious perturbation to the velocity
166 distributions of the simulated systems. Furthermore, this approach has
167 the advantage in heterogeneous interfaces in that kinetic energy flux
168 can be applied between regions of particles of arbitary identity, and
169 the flux will not be restricted by difference in particle mass.
170
171 The NIVS algorithm scales the velocity vectors in two separate regions
172 of a simulation system with respective diagonal scaling matricies. To
173 determine these scaling factors in the matricies, a set of equations
174 including linear momentum conservation and kinetic energy conservation
175 constraints and target energy flux satisfaction is solved. With the
176 scaling operation applied to the system in a set frequency, bulk
177 temperature gradients can be easily established, and these can be used
178 for computing thermal conductivities. The NIVS algorithm conserves
179 momenta and energy and does not depend on an external thermostat.
180
181 \subsection{Defining Interfacial Thermal Conductivity ($G$)}
182
183 For an interface with relatively low interfacial conductance, and a
184 thermal flux between two distinct bulk regions, the regions on either
185 side of the interface rapidly come to a state in which the two phases
186 have relatively homogeneous (but distinct) temperatures. The
187 interfacial thermal conductivity $G$ can therefore be approximated as:
188 \begin{equation}
189 G = \frac{E_{total}}{2 t L_x L_y \left( \langle T_\mathrm{hot}\rangle -
190 \langle T_\mathrm{cold}\rangle \right)}
191 \label{lowG}
192 \end{equation}
193 where ${E_{total}}$ is the total imposed non-physical kinetic energy
194 transfer during the simulation and ${\langle T_\mathrm{hot}\rangle}$
195 and ${\langle T_\mathrm{cold}\rangle}$ are the average observed
196 temperature of the two separated phases. For an applied flux $J_z$
197 operating over a simulation time $t$ on a periodically-replicated slab
198 of dimensions $L_x \times L_y$, $E_{total} = J_z *(t)*(2 L_x L_y)$.
199
200 When the interfacial conductance is {\it not} small, there are two
201 ways to define $G$. One common way is to assume the temperature is
202 discrete on the two sides of the interface. $G$ can be calculated
203 using the applied thermal flux $J$ and the maximum temperature
204 difference measured along the thermal gradient max($\Delta T$), which
205 occurs at the Gibbs deviding surface (Figure \ref{demoPic}). This is
206 known as the Kapitza conductance, which is the inverse of the Kapitza
207 resistance.
208 \begin{equation}
209 G=\frac{J}{\Delta T}
210 \label{discreteG}
211 \end{equation}
212
213 \begin{figure}
214 \includegraphics[width=\linewidth]{method}
215 \caption{Interfacial conductance can be calculated by applying an
216 (unphysical) kinetic energy flux between two slabs, one located
217 within the metal and another on the edge of the periodic box. The
218 system responds by forming a thermal response or a gradient. In
219 bulk liquids, this gradient typically has a single slope, but in
220 interfacial systems, there are distinct thermal conductivity
221 domains. The interfacial conductance, $G$ is found by measuring the
222 temperature gap at the Gibbs dividing surface, or by using second
223 derivatives of the thermal profile.}
224 \label{demoPic}
225 \end{figure}
226
227 The other approach is to assume a continuous temperature profile along
228 the thermal gradient axis (e.g. $z$) and define $G$ at the point where
229 the magnitude of thermal conductivity ($\lambda$) change reaches its
230 maximum, given that $\lambda$ is well-defined throughout the space:
231 \begin{equation}
232 G^\prime = \Big|\frac{\partial\lambda}{\partial z}\Big|
233 = \Big|\frac{\partial}{\partial z}\left(-J_z\Big/
234 \left(\frac{\partial T}{\partial z}\right)\right)\Big|
235 = |J_z|\Big|\frac{\partial^2 T}{\partial z^2}\Big|
236 \Big/\left(\frac{\partial T}{\partial z}\right)^2
237 \label{derivativeG}
238 \end{equation}
239
240 With temperature profiles obtained from simulation, one is able to
241 approximate the first and second derivatives of $T$ with finite
242 difference methods and calculate $G^\prime$. In what follows, both
243 definitions have been used, and are compared in the results.
244
245 To investigate the interfacial conductivity at metal / solvent
246 interfaces, we have modeled a metal slab with its (111) surfaces
247 perpendicular to the $z$-axis of our simulation cells. The metal slab
248 has been prepared both with and without capping agents on the exposed
249 surface, and has been solvated with simple organic solvents, as
250 illustrated in Figure \ref{gradT}.
251
252 With the simulation cell described above, we are able to equilibrate
253 the system and impose an unphysical thermal flux between the liquid
254 and the metal phase using the NIVS algorithm. By periodically applying
255 the unphysical flux, we obtained a temperature profile and its spatial
256 derivatives. Figure \ref{gradT} shows how an applied thermal flux can
257 be used to obtain the 1st and 2nd derivatives of the temperature
258 profile.
259
260 \begin{figure}
261 \includegraphics[width=\linewidth]{gradT}
262 \caption{A sample of Au-butanethiol/hexane interfacial system and the
263 temperature profile after a kinetic energy flux is imposed to
264 it. The 1st and 2nd derivatives of the temperature profile can be
265 obtained with finite difference approximation (lower panel).}
266 \label{gradT}
267 \end{figure}
268
269 \section{Computational Details}
270 \subsection{Simulation Protocol}
271 The NIVS algorithm has been implemented in our MD simulation code,
272 OpenMD\cite{Meineke:2005gd,openmd}, and was used for our simulations.
273 Metal slabs of 6 or 11 layers of Au atoms were first equilibrated
274 under atmospheric pressure (1 atm) and 200K. After equilibration,
275 butanethiol capping agents were placed at three-fold hollow sites on
276 the Au(111) surfaces. These sites are either {\it fcc} or {\it
277 hcp} sites, although Hase {\it et al.} found that they are
278 equivalent in a heat transfer process,\cite{hase:2010} so we did not
279 distinguish between these sites in our study. The maximum butanethiol
280 capacity on Au surface is $1/3$ of the total number of surface Au
281 atoms, and the packing forms a $(\sqrt{3}\times\sqrt{3})R30^\circ$
282 structure\cite{doi:10.1021/ja00008a001,doi:10.1021/cr9801317}. A
283 series of lower coverages was also prepared by eliminating
284 butanethiols from the higher coverage surface in a regular manner. The
285 lower coverages were prepared in order to study the relation between
286 coverage and interfacial conductance.
287
288 The capping agent molecules were allowed to migrate during the
289 simulations. They distributed themselves uniformly and sampled a
290 number of three-fold sites throughout out study. Therefore, the
291 initial configuration does not noticeably affect the sampling of a
292 variety of configurations of the same coverage, and the final
293 conductance measurement would be an average effect of these
294 configurations explored in the simulations.
295
296 After the modified Au-butanethiol surface systems were equilibrated in
297 the canonical (NVT) ensemble, organic solvent molecules were packed in
298 the previously empty part of the simulation cells.\cite{packmol} Two
299 solvents were investigated, one which has little vibrational overlap
300 with the alkanethiol and which has a planar shape (toluene), and one
301 which has similar vibrational frequencies to the capping agent and
302 chain-like shape ({\it n}-hexane).
303
304 The simulation cells were not particularly extensive along the
305 $z$-axis, as a very long length scale for the thermal gradient may
306 cause excessively hot or cold temperatures in the middle of the
307 solvent region and lead to undesired phenomena such as solvent boiling
308 or freezing when a thermal flux is applied. Conversely, too few
309 solvent molecules would change the normal behavior of the liquid
310 phase. Therefore, our $N_{solvent}$ values were chosen to ensure that
311 these extreme cases did not happen to our simulations. The spacing
312 between periodic images of the gold interfaces is $45 \sim 75$\AA in
313 our simulations.
314
315 The initial configurations generated are further equilibrated with the
316 $x$ and $y$ dimensions fixed, only allowing the $z$-length scale to
317 change. This is to ensure that the equilibration of liquid phase does
318 not affect the metal's crystalline structure. Comparisons were made
319 with simulations that allowed changes of $L_x$ and $L_y$ during NPT
320 equilibration. No substantial changes in the box geometry were noticed
321 in these simulations. After ensuring the liquid phase reaches
322 equilibrium at atmospheric pressure (1 atm), further equilibration was
323 carried out under canonical (NVT) and microcanonical (NVE) ensembles.
324
325 After the systems reach equilibrium, NIVS was used to impose an
326 unphysical thermal flux between the metal and the liquid phases. Most
327 of our simulations were done under an average temperature of
328 $\sim$200K. Therefore, thermal flux usually came from the metal to the
329 liquid so that the liquid has a higher temperature and would not
330 freeze due to lowered temperatures. After this induced temperature
331 gradient had stablized, the temperature profile of the simulation cell
332 was recorded. To do this, the simulation cell is devided evenly into
333 $N$ slabs along the $z$-axis. The average temperatures of each slab
334 are recorded for 1$\sim$2 ns. When the slab width $d$ of each slab is
335 the same, the derivatives of $T$ with respect to slab number $n$ can
336 be directly used for $G^\prime$ calculations: \begin{equation}
337 G^\prime = |J_z|\Big|\frac{\partial^2 T}{\partial z^2}\Big|
338 \Big/\left(\frac{\partial T}{\partial z}\right)^2
339 = |J_z|\Big|\frac{1}{d^2}\frac{\partial^2 T}{\partial n^2}\Big|
340 \Big/\left(\frac{1}{d}\frac{\partial T}{\partial n}\right)^2
341 = |J_z|\Big|\frac{\partial^2 T}{\partial n^2}\Big|
342 \Big/\left(\frac{\partial T}{\partial n}\right)^2
343 \label{derivativeG2}
344 \end{equation}
345
346 All of the above simulation procedures use a time step of 1 fs. Each
347 equilibration stage took a minimum of 100 ps, although in some cases,
348 longer equilibration stages were utilized.
349
350 \subsection{Force Field Parameters}
351 Our simulations include a number of chemically distinct components.
352 Figure \ref{demoMol} demonstrates the sites defined for both
353 United-Atom and All-Atom models of the organic solvent and capping
354 agents in our simulations. Force field parameters are needed for
355 interactions both between the same type of particles and between
356 particles of different species.
357
358 \begin{figure}
359 \includegraphics[width=\linewidth]{structures}
360 \caption{Structures of the capping agent and solvents utilized in
361 these simulations. The chemically-distinct sites (a-e) are expanded
362 in terms of constituent atoms for both United Atom (UA) and All Atom
363 (AA) force fields. Most parameters are from
364 Refs. \protect\cite{TraPPE-UA.alkanes,TraPPE-UA.alkylbenzenes,TraPPE-UA.thiols}
365 (UA) and \protect\cite{OPLSAA} (AA). Cross-interactions with the Au
366 atoms are given in Table \ref{MnM}.}
367 \label{demoMol}
368 \end{figure}
369
370 The Au-Au interactions in metal lattice slab is described by the
371 quantum Sutton-Chen (QSC) formulation.\cite{PhysRevB.59.3527} The QSC
372 potentials include zero-point quantum corrections and are
373 reparametrized for accurate surface energies compared to the
374 Sutton-Chen potentials.\cite{Chen90}
375
376 For the two solvent molecules, {\it n}-hexane and toluene, two
377 different atomistic models were utilized. Both solvents were modeled
378 using United-Atom (UA) and All-Atom (AA) models. The TraPPE-UA
379 parameters\cite{TraPPE-UA.alkanes,TraPPE-UA.alkylbenzenes} are used
380 for our UA solvent molecules. In these models, sites are located at
381 the carbon centers for alkyl groups. Bonding interactions, including
382 bond stretches and bends and torsions, were used for intra-molecular
383 sites closer than 3 bonds. For non-bonded interactions, Lennard-Jones
384 potentials are used.
385
386 By eliminating explicit hydrogen atoms, the TraPPE-UA models are
387 simple and computationally efficient, while maintaining good accuracy.
388 However, the TraPPE-UA model for alkanes is known to predict a slighly
389 lower boiling point than experimental values. This is one of the
390 reasons we used a lower average temperature (200K) for our
391 simulations. If heat is transferred to the liquid phase during the
392 NIVS simulation, the liquid in the hot slab can actually be
393 substantially warmer than the mean temperature in the simulation. The
394 lower mean temperatures therefore prevent solvent boiling.
395
396 For UA-toluene, the non-bonded potentials between intermolecular sites
397 have a similar Lennard-Jones formulation. The toluene molecules were
398 treated as a single rigid body, so there was no need for
399 intramolecular interactions (including bonds, bends, or torsions) in
400 this solvent model.
401
402 Besides the TraPPE-UA models, AA models for both organic solvents are
403 included in our studies as well. The OPLS-AA\cite{OPLSAA} force fields
404 were used. For hexane, additional explicit hydrogen sites were
405 included. Besides bonding and non-bonded site-site interactions,
406 partial charges and the electrostatic interactions were added to each
407 CT and HC site. For toluene, a flexible model for the toluene molecule
408 was utilized which included bond, bend, torsion, and inversion
409 potentials to enforce ring planarity.
410
411 The butanethiol capping agent in our simulations, were also modeled
412 with both UA and AA model. The TraPPE-UA force field includes
413 parameters for thiol molecules\cite{TraPPE-UA.thiols} and are used for
414 UA butanethiol model in our simulations. The OPLS-AA also provides
415 parameters for alkyl thiols. However, alkyl thiols adsorbed on Au(111)
416 surfaces do not have the hydrogen atom bonded to sulfur. To derive
417 suitable parameters for butanethiol adsorbed on Au(111) surfaces, we
418 adopt the S parameters from Luedtke and Landman\cite{landman:1998} and
419 modify the parameters for the CTS atom to maintain charge neutrality
420 in the molecule. Note that the model choice (UA or AA) for the capping
421 agent can be different from the solvent. Regardless of model choice,
422 the force field parameters for interactions between capping agent and
423 solvent can be derived using Lorentz-Berthelot Mixing Rule:
424 \begin{eqnarray}
425 \sigma_{ij} & = & \frac{1}{2} \left(\sigma_{ii} + \sigma_{jj}\right) \\
426 \epsilon_{ij} & = & \sqrt{\epsilon_{ii}\epsilon_{jj}}
427 \end{eqnarray}
428
429 To describe the interactions between metal (Au) and non-metal atoms,
430 we refer to an adsorption study of alkyl thiols on gold surfaces by
431 Vlugt {\it et al.}\cite{vlugt:cpc2007154} They fitted an effective
432 Lennard-Jones form of potential parameters for the interaction between
433 Au and pseudo-atoms CH$_x$ and S based on a well-established and
434 widely-used effective potential of Hautman and Klein for the Au(111)
435 surface.\cite{hautman:4994} As our simulations require the gold slab
436 to be flexible to accommodate thermal excitation, the pair-wise form
437 of potentials they developed was used for our study.
438
439 The potentials developed from {\it ab initio} calculations by Leng
440 {\it et al.}\cite{doi:10.1021/jp034405s} are adopted for the
441 interactions between Au and aromatic C/H atoms in toluene. However,
442 the Lennard-Jones parameters between Au and other types of particles,
443 (e.g. AA alkanes) have not yet been established. For these
444 interactions, the Lorentz-Berthelot mixing rule can be used to derive
445 effective single-atom LJ parameters for the metal using the fit values
446 for toluene. These are then used to construct reasonable mixing
447 parameters for the interactions between the gold and other atoms.
448 Table \ref{MnM} summarizes the ``metal/non-metal'' parameters used in
449 our simulations.
450
451 \begin{table*}
452 \begin{minipage}{\linewidth}
453 \begin{center}
454 \caption{Non-bonded interaction parameters (including cross
455 interactions with Au atoms) for both force fields used in this
456 work.}
457 \begin{tabular}{lllllll}
458 \hline\hline
459 & Site & $\sigma_{ii}$ & $\epsilon_{ii}$ & $q_i$ &
460 $\sigma_{Au-i}$ & $\epsilon_{Au-i}$ \\
461 & & (\AA) & (kcal/mol) & ($e$) & (\AA) & (kcal/mol) \\
462 \hline
463 United Atom (UA)
464 &CH3 & 3.75 & 0.1947 & - & 3.54 & 0.2146 \\
465 &CH2 & 3.95 & 0.0914 & - & 3.54 & 0.1749 \\
466 &CHar & 3.695 & 0.1003 & - & 3.4625 & 0.1680 \\
467 &CRar & 3.88 & 0.04173 & - & 3.555 & 0.1604 \\
468 \hline
469 All Atom (AA)
470 &CT3 & 3.50 & 0.066 & -0.18 & 3.365 & 0.1373 \\
471 &CT2 & 3.50 & 0.066 & -0.12 & 3.365 & 0.1373 \\
472 &CTT & 3.50 & 0.066 & -0.065 & 3.365 & 0.1373 \\
473 &HC & 2.50 & 0.030 & 0.06 & 2.865 & 0.09256 \\
474 &CA & 3.55 & 0.070 & -0.115 & 3.173 & 0.0640 \\
475 &HA & 2.42 & 0.030 & 0.115 & 2.746 & 0.0414 \\
476 \hline
477 Both UA and AA
478 & S & 4.45 & 0.25 & - & 2.40 & 8.465 \\
479 \hline\hline
480 \end{tabular}
481 \label{MnM}
482 \end{center}
483 \end{minipage}
484 \end{table*}
485
486
487 \section{Results}
488 There are many factors contributing to the measured interfacial
489 conductance; some of these factors are physically motivated
490 (e.g. coverage of the surface by the capping agent coverage and
491 solvent identity), while some are governed by parameters of the
492 methodology (e.g. applied flux and the formulas used to obtain the
493 conductance). In this section we discuss the major physical and
494 calculational effects on the computed conductivity.
495
496 \subsection{Effects due to capping agent coverage}
497
498 A series of different initial conditions with a range of surface
499 coverages was prepared and solvated with various with both of the
500 solvent molecules. These systems were then equilibrated and their
501 interfacial thermal conductivity was measured with the NIVS
502 algorithm. Figure \ref{coverage} demonstrates the trend of conductance
503 with respect to surface coverage.
504
505 \begin{figure}
506 \includegraphics[width=\linewidth]{coverage}
507 \caption{Comparison of interfacial thermal conductivity ($G$) values
508 for the Au-butanethiol/solvent interface with various UA models and
509 different capping agent coverages at $\langle T\rangle\sim$200K.}
510 \label{coverage}
511 \end{figure}
512
513 In partially covered surfaces, the derivative definition for
514 $G^\prime$ (Eq. \ref{derivativeG}) becomes difficult to apply, as the
515 location of maximum change of $\lambda$ becomes washed out. The
516 discrete definition (Eq. \ref{discreteG}) is easier to apply, as the
517 Gibbs dividing surface is still well-defined. Therefore, $G$ (not
518 $G^\prime$) was used in this section.
519
520 From Figure \ref{coverage}, one can see the significance of the
521 presence of capping agents. When even a small fraction of the Au(111)
522 surface sites are covered with butanethiols, the conductivity exhibits
523 an enhancement by at least a factor of 3. Cappping agents are clearly
524 playing a major role in thermal transport at metal / organic solvent
525 surfaces.
526
527 We note a non-monotonic behavior in the interfacial conductance as a
528 function of surface coverage. The maximum conductance (largest $G$)
529 happens when the surfaces are about 75\% covered with butanethiol
530 caps. The reason for this behavior is not entirely clear. One
531 explanation is that incomplete butanethiol coverage allows small gaps
532 between butanethiols to form. These gaps can be filled by transient
533 solvent molecules. These solvent molecules couple very strongly with
534 the hot capping agent molecules near the surface, and can then carry
535 away (diffusively) the excess thermal energy from the surface.
536
537 There appears to be a competition between the conduction of the
538 thermal energy away from the surface by the capping agents (enhanced
539 by greater coverage) and the coupling of the capping agents with the
540 solvent (enhanced by interdigitation at lower coverages). This
541 competition would lead to the non-monotonic coverage behavior observed
542 here.
543
544 Results for rigid body toluene solvent, as well as the UA hexane, are
545 within the ranges expected from prior experimental
546 work.\cite{Wilson:2002uq,cahill:793,PhysRevB.80.195406} This suggests
547 that explicit hydrogen atoms might not be required for modeling
548 thermal transport in these systems. C-H vibrational modes do not see
549 significant excited state population at low temperatures, and are not
550 likely to carry lower frequency excitations from the solid layer into
551 the bulk liquid.
552
553 The toluene solvent does not exhibit the same behavior as hexane in
554 that $G$ remains at approximately the same magnitude when the capping
555 coverage increases from 25\% to 75\%. Toluene, as a rigid planar
556 molecule, cannot occupy the relatively small gaps between the capping
557 agents as easily as the chain-like {\it n}-hexane. The effect of
558 solvent coupling to the capping agent is therefore weaker in toluene
559 except at the very lowest coverage levels. This effect counters the
560 coverage-dependent conduction of heat away from the metal surface,
561 leading to a much flatter $G$ vs. coverage trend than is observed in
562 {\it n}-hexane.
563
564 \subsection{Effects due to Solvent \& Solvent Models}
565 In addition to UA solvent and capping agent models, AA models have
566 also been included in our simulations. In most of this work, the same
567 (UA or AA) model for solvent and capping agent was used, but it is
568 also possible to utilize different models for different components.
569 We have also included isotopic substitutions (Hydrogen to Deuterium)
570 to decrease the explicit vibrational overlap between solvent and
571 capping agent. Table \ref{modelTest} summarizes the results of these
572 studies.
573
574 \begin{table*}
575 \begin{minipage}{\linewidth}
576 \begin{center}
577
578 \caption{Computed interfacial thermal conductance ($G$ and
579 $G^\prime$) values for interfaces using various models for
580 solvent and capping agent (or without capping agent) at
581 $\langle T\rangle\sim$200K. (D stands for deuterated solvent
582 or capping agent molecules; ``Avg.'' denotes results that are
583 averages of simulations under different applied thermal flux
584 values $(J_z)$. Error estimates are indicated in
585 parentheses.)}
586
587 \begin{tabular}{llccc}
588 \hline\hline
589 Butanethiol model & Solvent & $J_z$ & $G$ & $G^\prime$ \\
590 (or bare surface) & model & (GW/m$^2$) &
591 \multicolumn{2}{c}{(MW/m$^2$/K)} \\
592 \hline
593 UA & UA hexane & Avg. & 131(9) & 87(10) \\
594 & UA hexane(D) & 1.95 & 153(5) & 136(13) \\
595 & AA hexane & Avg. & 131(6) & 122(10) \\
596 & UA toluene & 1.96 & 187(16) & 151(11) \\
597 & AA toluene & 1.89 & 200(36) & 149(53) \\
598 \hline
599 AA & UA hexane & 1.94 & 116(9) & 129(8) \\
600 & AA hexane & Avg. & 442(14) & 356(31) \\
601 & AA hexane(D) & 1.93 & 222(12) & 234(54) \\
602 & UA toluene & 1.98 & 125(25) & 97(60) \\
603 & AA toluene & 3.79 & 487(56) & 290(42) \\
604 \hline
605 AA(D) & UA hexane & 1.94 & 158(25) & 172(4) \\
606 & AA hexane & 1.92 & 243(29) & 191(11) \\
607 & AA toluene & 1.93 & 364(36) & 322(67) \\
608 \hline
609 bare & UA hexane & Avg. & 46.5(3.2) & 49.4(4.5) \\
610 & UA hexane(D) & 0.98 & 43.9(4.6) & 43.0(2.0) \\
611 & AA hexane & 0.96 & 31.0(1.4) & 29.4(1.3) \\
612 & UA toluene & 1.99 & 70.1(1.3) & 65.8(0.5) \\
613 \hline\hline
614 \end{tabular}
615 \label{modelTest}
616 \end{center}
617 \end{minipage}
618 \end{table*}
619
620 To facilitate direct comparison between force fields, systems with the
621 same capping agent and solvent were prepared with the same length
622 scales for the simulation cells.
623
624 On bare metal / solvent surfaces, different force field models for
625 hexane yield similar results for both $G$ and $G^\prime$, and these
626 two definitions agree with each other very well. This is primarily an
627 indicator of weak interactions between the metal and the solvent, and
628 is a typical case for acoustic impedance mismatch between these two
629 phases.
630
631 For the fully-covered surfaces, the choice of force field for the
632 capping agent and solvent has a large impact on the calulated values
633 of $G$ and $G^\prime$. The AA thiol to AA solvent conductivities are
634 much larger than their UA to UA counterparts, and these values exceed
635 the experimental estimates by a large measure. The AA force field
636 allows significant energy to go into C-H (or C-D) stretching modes,
637 and since these modes are high frequency, this non-quantum behavior is
638 likely responsible for the overestimate of the conductivity. Compared
639 to the AA model, the UA model yields more reasonable conductivity
640 values with much higher computational efficiency.
641
642 \subsubsection{Are electronic excitations in the metal important?}
643 Because they lack electronic excitations, the QSC and related embedded
644 atom method (EAM) models for gold are known to predict unreasonably
645 low values for bulk conductivity
646 ($\lambda$).\cite{kuang:164101,ISI:000207079300006,Clancy:1992} If the
647 conductance between the phases ($G$) is governed primarily by phonon
648 excitation (and not electronic degrees of freedom), one would expect a
649 classical model to capture most of the interfacial thermal
650 conductance. Our results for $G$ and $G^\prime$ indicate that this is
651 indeed the case, and suggest that the modeling of interfacial thermal
652 transport depends primarily on the description of the interactions
653 between the various components at the interface. When the metal is
654 chemically capped, the primary barrier to thermal conductivity appears
655 to be the interface between the capping agent and the surrounding
656 solvent, so the excitations in the metal have little impact on the
657 value of $G$.
658
659 \subsection{Effects due to methodology and simulation parameters}
660
661 We have varied the parameters of the simulations in order to
662 investigate how these factors would affect the computation of $G$. Of
663 particular interest are: 1) the length scale for the applied thermal
664 gradient (modified by increasing the amount of solvent in the system),
665 2) the sign and magnitude of the applied thermal flux, 3) the average
666 temperature of the simulation (which alters the solvent density during
667 equilibration), and 4) the definition of the interfacial conductance
668 (Eqs. (\ref{discreteG}) or (\ref{derivativeG})) used in the
669 calculation.
670
671 Systems of different lengths were prepared by altering the number of
672 solvent molecules and extending the length of the box along the $z$
673 axis to accomodate the extra solvent. Equilibration at the same
674 temperature and pressure conditions led to nearly identical surface
675 areas ($L_x$ and $L_y$) available to the metal and capping agent,
676 while the extra solvent served mainly to lengthen the axis that was
677 used to apply the thermal flux. For a given value of the applied
678 flux, the different $z$ length scale has only a weak effect on the
679 computed conductivities (Table \ref{AuThiolHexaneUA}).
680
681 \subsubsection{Effects of applied flux}
682 The NIVS algorithm allows changes in both the sign and magnitude of
683 the applied flux. It is possible to reverse the direction of heat
684 flow simply by changing the sign of the flux, and thermal gradients
685 which would be difficult to obtain experimentally ($5$ K/\AA) can be
686 easily simulated. However, the magnitude of the applied flux is not
687 arbitary if one aims to obtain a stable and reliable thermal gradient.
688 A temperature gradient can be lost in the noise if $|J_z|$ is too
689 small, and excessive $|J_z|$ values can cause phase transitions if the
690 extremes of the simulation cell become widely separated in
691 temperature. Also, if $|J_z|$ is too large for the bulk conductivity
692 of the materials, the thermal gradient will never reach a stable
693 state.
694
695 Within a reasonable range of $J_z$ values, we were able to study how
696 $G$ changes as a function of this flux. In what follows, we use
697 positive $J_z$ values to denote the case where energy is being
698 transferred by the method from the metal phase and into the liquid.
699 The resulting gradient therefore has a higher temperature in the
700 liquid phase. Negative flux values reverse this transfer, and result
701 in higher temperature metal phases. The conductance measured under
702 different applied $J_z$ values is listed in Tables
703 \ref{AuThiolHexaneUA} and \ref{AuThiolToluene}. These results do not
704 indicate that $G$ depends strongly on $J_z$ within this flux
705 range. The linear response of flux to thermal gradient simplifies our
706 investigations in that we can rely on $G$ measurement with only a
707 small number $J_z$ values.
708
709 \begin{table*}
710 \begin{minipage}{\linewidth}
711 \begin{center}
712 \caption{Computed interfacial thermal conductivity ($G$ and
713 $G^\prime$) values for the 100\% covered Au-butanethiol/hexane
714 interfaces with UA model and different hexane molecule numbers
715 at different temperatures using a range of energy
716 fluxes. Error estimates indicated in parenthesis.}
717
718 \begin{tabular}{ccccccc}
719 \hline\hline
720 $\langle T\rangle$ & $N_{hexane}$ & $\rho_{hexane}$ &
721 $J_z$ & $G$ & $G^\prime$ \\
722 (K) & & (g/cm$^3$) & (GW/m$^2$) &
723 \multicolumn{2}{c}{(MW/m$^2$/K)} \\
724 \hline
725 200 & 266 & 0.672 & -0.96 & 102(3) & 80.0(0.8) \\
726 & 200 & 0.688 & 0.96 & 125(16) & 90.2(15) \\
727 & & & 1.91 & 139(10) & 101(10) \\
728 & & & 2.83 & 141(6) & 89.9(9.8) \\
729 & 166 & 0.681 & 0.97 & 141(30) & 78(22) \\
730 & & & 1.92 & 138(4) & 98.9(9.5) \\
731 \hline
732 250 & 200 & 0.560 & 0.96 & 75(10) & 61.8(7.3) \\
733 & & & -0.95 & 49.4(0.3) & 45.7(2.1) \\
734 & 166 & 0.569 & 0.97 & 80.3(0.6) & 67(11) \\
735 & & & 1.44 & 76.2(5.0) & 64.8(3.8) \\
736 & & & -0.95 & 56.4(2.5) & 54.4(1.1) \\
737 & & & -1.85 & 47.8(1.1) & 53.5(1.5) \\
738 \hline\hline
739 \end{tabular}
740 \label{AuThiolHexaneUA}
741 \end{center}
742 \end{minipage}
743 \end{table*}
744
745 The sign of $J_z$ is a different matter, however, as this can alter
746 the temperature on the two sides of the interface. The average
747 temperature values reported are for the entire system, and not for the
748 liquid phase, so at a given $\langle T \rangle$, the system with
749 positive $J_z$ has a warmer liquid phase. This means that if the
750 liquid carries thermal energy via convective transport, {\it positive}
751 $J_z$ values will result in increased molecular motion on the liquid
752 side of the interface, and this will increase the measured
753 conductivity.
754
755 \subsubsection{Effects due to average temperature}
756
757 We also studied the effect of average system temperature on the
758 interfacial conductance. The simulations are first equilibrated in
759 the NPT ensemble at 1 atm. The TraPPE-UA model for hexane tends to
760 predict a lower boiling point (and liquid state density) than
761 experiments. This lower-density liquid phase leads to reduced contact
762 between the hexane and butanethiol, and this accounts for our
763 observation of lower conductance at higher temperatures as shown in
764 Table \ref{AuThiolHexaneUA}. In raising the average temperature from
765 200K to 250K, the density drop of $\sim$20\% in the solvent phase
766 leads to a $\sim$65\% drop in the conductance. [BUT (125-75)/125 = .4?]
767
768 Similar behavior is observed in the TraPPE-UA model for toluene,
769 although this model has better agreement with the experimental
770 densities of toluene. The expansion of the toluene liquid phase is
771 not as significant as that of the hexane (8.3\% over 100K), and this
772 limits the effect to $\sim$20\% drop in thermal conductivity (Table
773 \ref{AuThiolToluene}).
774
775 Although we have not mapped out the behavior at a large number of
776 temperatures, is clear that there will be a strong temperature
777 dependence in the interfacial conductance when the physical properties
778 of one side of the interface (notably the density) change rapidly as a
779 function of temperature.
780
781 \begin{table*}
782 \begin{minipage}{\linewidth}
783 \begin{center}
784 \caption{Computed interfacial thermal conductivity ($G$ and
785 $G^\prime$) values for a 90\% coverage Au-butanethiol/toluene
786 interface at different temperatures using a range of energy
787 fluxes. Error estimates indicated in parenthesis.}
788
789 \begin{tabular}{ccccc}
790 \hline\hline
791 $\langle T\rangle$ & $\rho_{toluene}$ & $J_z$ & $G$ & $G^\prime$ \\
792 (K) & (g/cm$^3$) & (GW/m$^2$) & \multicolumn{2}{c}{(MW/m$^2$/K)} \\
793 \hline
794 200 & 0.933 & 2.15 & 204(12) & 113(12) \\
795 & & -1.86 & 180(3) & 135(21) \\
796 & & -3.93 & 176(5) & 113(12) \\
797 \hline
798 300 & 0.855 & -1.91 & 143(5) & 125(2) \\
799 & & -4.19 & 135(9) & 113(12) \\
800 \hline\hline
801 \end{tabular}
802 \label{AuThiolToluene}
803 \end{center}
804 \end{minipage}
805 \end{table*}
806
807 Besides the lower interfacial thermal conductance, surfaces at
808 relatively high temperatures are susceptible to reconstructions,
809 particularly when butanethiols fully cover the Au(111) surface. These
810 reconstructions include surface Au atoms which migrate outward to the
811 S atom layer, and butanethiol molecules which embed into the surface
812 Au layer. The driving force for this behavior is the strong Au-S
813 interactions which are modeled here with a deep Lennard-Jones
814 potential. This phenomenon agrees with reconstructions that have beeen
815 experimentally
816 observed.\cite{doi:10.1021/j100035a033,doi:10.1021/la026493y}. Vlugt
817 {\it et al.} kept their Au(111) slab rigid so that their simulations
818 could reach 300K without surface
819 reconstructions.\cite{vlugt:cpc2007154} Since surface reconstructions
820 blur the interface, the measurement of $G$ becomes more difficult to
821 conduct at higher temperatures. For this reason, most of our
822 measurements are undertaken at $\langle T\rangle\sim$200K where
823 reconstruction is minimized.
824
825 However, when the surface is not completely covered by butanethiols,
826 the simulated system appears to be more resistent to the
827 reconstruction. Our Au / butanethiol / toluene system had the Au(111)
828 surfaces 90\% covered by butanethiols, but did not see this above
829 phenomena even at $\langle T\rangle\sim$300K. That said, we did
830 observe butanethiols migrating to neighboring three-fold sites during
831 a simulation. Since the interface persisted in these simulations,
832 were able to obtain $G$'s for these interfaces even at a relatively
833 high temperature without being affected by surface reconstructions.
834
835 \section{Discussion}
836
837 The primary result of this work is that the capping agent acts as an
838 efficient thermal coupler between solid and solvent phases. One of
839 the ways the capping agent can carry out this role is to down-shift
840 between the phonon vibrations in the solid (which carry the heat from
841 the gold) and the molecular vibrations in the liquid (which carry some
842 of the heat in the solvent).
843
844 To investigate the mechanism of interfacial thermal conductance, the
845 vibrational power spectrum was computed. Power spectra were taken for
846 individual components in different simulations. To obtain these
847 spectra, simulations were run after equilibration in the
848 microcanonical (NVE) ensemble and without a thermal
849 gradient. Snapshots of configurations were collected at a frequency
850 that is higher than that of the fastest vibrations occuring in the
851 simulations. With these configurations, the velocity auto-correlation
852 functions can be computed:
853 \begin{equation}
854 C_A (t) = \langle\vec{v}_A (t)\cdot\vec{v}_A (0)\rangle
855 \label{vCorr}
856 \end{equation}
857 The power spectrum is constructed via a Fourier transform of the
858 symmetrized velocity autocorrelation function,
859 \begin{equation}
860 \hat{f}(\omega) =
861 \int_{-\infty}^{\infty} C_A (t) e^{-2\pi it\omega}\,dt
862 \label{fourier}
863 \end{equation}
864
865 \subsection{The role of specific vibrations}
866 The vibrational spectra for gold slabs in different environments are
867 shown as in Figure \ref{specAu}. Regardless of the presence of
868 solvent, the gold surfaces which are covered by butanethiol molecules
869 exhibit an additional peak observed at a frequency of
870 $\sim$165cm$^{-1}$. We attribute this peak to the S-Au bonding
871 vibration. This vibration enables efficient thermal coupling of the
872 surface Au layer to the capping agents. Therefore, in our simulations,
873 the Au / S interfaces do not appear to be the primary barrier to
874 thermal transport when compared with the butanethiol / solvent
875 interfaces.
876
877 \begin{figure}
878 \includegraphics[width=\linewidth]{vibration}
879 \caption{Vibrational power spectra for gold in different solvent
880 environments. The presence of the butanethiol capping molecules
881 adds a vibrational peak at $\sim$165cm$^{-1}$. The butanethiol
882 spectra exhibit a corresponding peak.}
883 \label{specAu}
884 \end{figure}
885
886 Also in this figure, we show the vibrational power spectrum for the
887 bound butanethiol molecules, which also exhibits the same
888 $\sim$165cm$^{-1}$ peak.
889
890 \subsection{Overlap of power spectra}
891 A comparison of the results obtained from the two different organic
892 solvents can also provide useful information of the interfacial
893 thermal transport process. In particular, the vibrational overlap
894 between the butanethiol and the organic solvents suggests a highly
895 efficient thermal exchange between these components. Very high
896 thermal conductivity was observed when AA models were used and C-H
897 vibrations were treated classically. The presence of extra degrees of
898 freedom in the AA force field yields higher heat exchange rates
899 between the two phases and results in a much higher conductivity than
900 in the UA force field.
901
902 The similarity in the vibrational modes available to solvent and
903 capping agent can be reduced by deuterating one of the two components
904 (Fig. \ref{aahxntln}). Once either the hexanes or the butanethiols
905 are deuterated, one can observe a significantly lower $G$ and
906 $G^\prime$ values (Table \ref{modelTest}).
907
908 \begin{figure}
909 \includegraphics[width=\linewidth]{aahxntln}
910 \caption{Spectra obtained for all-atom (AA) Au / butanethiol / solvent
911 systems. When butanethiol is deuterated (lower left), its
912 vibrational overlap with hexane decreases significantly. Since
913 aromatic molecules and the butanethiol are vibrationally dissimilar,
914 the change is not as dramatic when toluene is the solvent (right).}
915 \label{aahxntln}
916 \end{figure}
917
918 For the Au / butanethiol / toluene interfaces, having the AA
919 butanethiol deuterated did not yield a significant change in the
920 measured conductance. Compared to the C-H vibrational overlap between
921 hexane and butanethiol, both of which have alkyl chains, the overlap
922 between toluene and butanethiol is not as significant and thus does
923 not contribute as much to the heat exchange process.
924
925 Previous observations of Zhang {\it et al.}\cite{hase:2010} indicate
926 that the {\it intra}molecular heat transport due to alkylthiols is
927 highly efficient. Combining our observations with those of Zhang {\it
928 et al.}, it appears that butanethiol acts as a channel to expedite
929 heat flow from the gold surface and into the alkyl chain. The
930 acoustic impedance mismatch between the metal and the liquid phase can
931 therefore be effectively reduced with the presence of suitable capping
932 agents.
933
934 Deuterated models in the UA force field did not decouple the thermal
935 transport as well as in the AA force field. The UA models, even
936 though they have eliminated the high frequency C-H vibrational
937 overlap, still have significant overlap in the lower-frequency
938 portions of the infrared spectra (Figure \ref{uahxnua}). Deuterating
939 the UA models did not decouple the low frequency region enough to
940 produce an observable difference for the results of $G$ (Table
941 \ref{modelTest}).
942
943 \begin{figure}
944 \includegraphics[width=\linewidth]{uahxnua}
945 \caption{Vibrational spectra obtained for normal (upper) and
946 deuterated (lower) hexane in Au-butanethiol/hexane
947 systems. Butanethiol spectra are shown as reference. Both hexane and
948 butanethiol were using United-Atom models.}
949 \label{uahxnua}
950 \end{figure}
951
952 \section{Conclusions}
953 The NIVS algorithm has been applied to simulations of
954 butanethiol-capped Au(111) surfaces in the presence of organic
955 solvents. This algorithm allows the application of unphysical thermal
956 flux to transfer heat between the metal and the liquid phase. With the
957 flux applied, we were able to measure the corresponding thermal
958 gradients and to obtain interfacial thermal conductivities. Under
959 steady states, 2-3 ns trajectory simulations are sufficient for
960 computation of this quantity.
961
962 Our simulations have seen significant conductance enhancement in the
963 presence of capping agent, compared with the bare gold / liquid
964 interfaces. The acoustic impedance mismatch between the metal and the
965 liquid phase is effectively eliminated by a chemically-bonded capping
966 agent. Furthermore, the coverage precentage of the capping agent plays
967 an important role in the interfacial thermal transport
968 process. Moderately low coverages allow higher contact between capping
969 agent and solvent, and thus could further enhance the heat transfer
970 process, giving a non-monotonic behavior of conductance with
971 increasing coverage.
972
973 Our results, particularly using the UA models, agree well with
974 available experimental data. The AA models tend to overestimate the
975 interfacial thermal conductance in that the classically treated C-H
976 vibrations become too easily populated. Compared to the AA models, the
977 UA models have higher computational efficiency with satisfactory
978 accuracy, and thus are preferable in modeling interfacial thermal
979 transport.
980
981 Of the two definitions for $G$, the discrete form
982 (Eq. \ref{discreteG}) was easier to use and gives out relatively
983 consistent results, while the derivative form (Eq. \ref{derivativeG})
984 is not as versatile. Although $G^\prime$ gives out comparable results
985 and follows similar trend with $G$ when measuring close to fully
986 covered or bare surfaces, the spatial resolution of $T$ profile
987 required for the use of a derivative form is limited by the number of
988 bins and the sampling required to obtain thermal gradient information.
989
990 Vlugt {\it et al.} have investigated the surface thiol structures for
991 nanocrystalline gold and pointed out that they differ from those of
992 the Au(111) surface.\cite{landman:1998,vlugt:cpc2007154} This
993 difference could also cause differences in the interfacial thermal
994 transport behavior. To investigate this problem, one would need an
995 effective method for applying thermal gradients in non-planar
996 (i.e. spherical) geometries.
997
998 \section{Acknowledgments}
999 Support for this project was provided by the National Science
1000 Foundation under grant CHE-0848243. Computational time was provided by
1001 the Center for Research Computing (CRC) at the University of Notre
1002 Dame.
1003 \newpage
1004
1005 \bibliography{interfacial}
1006
1007 \end{doublespace}
1008 \end{document}
1009