ViewVC Help
View File | Revision Log | Show Annotations | View Changeset | Root Listing
root/group/trunk/multipole/multipole_2/multipole2.tex
Revision: 4221
Committed: Sun Sep 14 13:40:01 2014 UTC (9 years, 9 months ago) by gezelter
Content type: application/x-tex
File size: 62089 byte(s)
Log Message:
FInal Version, I think.

File Contents

# Content
1 % ****** Start of file aipsamp.tex ******
2 %
3 % This file is part of the AIP files in the AIP distribution for REVTeX 4.
4 % Version 4.1 of REVTeX, October 2009
5 %
6 % Copyright (c) 2009 American Institute of Physics.
7 %
8 % See the AIP README file for restrictions and more information.
9 %
10 % TeX'ing this file requires that you have AMS-LaTeX 2.0 installed
11 % as well as the rest of the prerequisites for REVTeX 4.1
12 %
13 % It also requires running BibTeX. The commands are as follows:
14 %
15 % 1) latex aipsamp
16 % 2) bibtex aipsamp
17 % 3) latex aipsamp
18 % 4) latex aipsamp
19 %
20 % Use this file as a source of example code for your aip document.
21 % Use the file aiptemplate.tex as a template for your document.
22 \documentclass[%
23 aip,jcp,
24 amsmath,amssymb,
25 preprint,
26 %reprint,%
27 %author-year,%
28 %author-numerical,%
29 ]{revtex4-1}
30
31 \usepackage{graphicx}% Include figure files
32 \usepackage{dcolumn}% Align table columns on decimal point
33 %\usepackage{bm}% bold math
34 %\usepackage[mathlines]{lineno}% Enable numbering of text and display math
35 %\linenumbers\relax % Commence numbering lines
36 \usepackage{amsmath}
37 \usepackage{times}
38 \usepackage{mathptmx}
39 \usepackage{tabularx}
40 \usepackage[version=3]{mhchem} % this is a great package for formatting chemical reactions
41 \usepackage{url}
42 \usepackage[english]{babel}
43
44 \newcolumntype{Y}{>{\centering\arraybackslash}X}
45
46 \begin{document}
47
48 %\preprint{AIP/123-QED}
49
50 \title{Real space electrostatics for multipoles. II. Comparisons with
51 the Ewald Sum}
52
53 \author{Madan Lamichhane}
54 \affiliation{Department of Physics, University of Notre Dame, Notre Dame, IN 46556}
55
56 \author{Kathie E. Newman}
57 \affiliation{Department of Physics, University of Notre Dame, Notre Dame, IN 46556}
58
59 \author{J. Daniel Gezelter}%
60 \email{gezelter@nd.edu.}
61 \affiliation{Department of Chemistry and Biochemistry, University of Notre Dame, Notre Dame, IN 46556
62 }
63
64 \date{\today}
65
66 \begin{abstract}
67 We report on tests of the shifted potential (SP), gradient shifted
68 force (GSF), and Taylor shifted force (TSF) real-space methods for
69 multipole interactions developed in the first paper in this series,
70 using the multipolar Ewald sum as a reference method. The tests were
71 carried out in a variety of condensed-phase environments designed to
72 test up to quadrupole-quadrupole interactions. Comparisons of the
73 energy differences between configurations, molecular forces, and
74 torques were used to analyze how well the real-space models perform
75 relative to the more computationally expensive Ewald treatment. We
76 have also investigated the energy conservation, structural, and
77 dynamical properties of the new methods in molecular dynamics
78 simulations. The SP method shows excellent agreement with
79 configurational energy differences, forces, and torques, and would
80 be suitable for use in Monte Carlo calculations. Of the two new
81 shifted-force methods, the GSF approach shows the best agreement
82 with Ewald-derived energies, forces, and torques and also exhibits
83 energy conservation properties that make it an excellent choice for
84 efficient computation of electrostatic interactions in molecular
85 dynamics simulations. Both SP and GSF are able to reproduce
86 structural and dynamical properties in the liquid models with
87 excellent fidelity.
88 \end{abstract}
89
90 %\pacs{Valid PACS appear here}% PACS, the Physics and Astronomy
91 % Classification Scheme.
92 %\keywords{Electrostatics, Multipoles, Real-space}
93
94 \maketitle
95
96 \section{\label{sec:intro}Introduction}
97 Computing the interactions between electrostatic sites is one of the
98 most expensive aspects of molecular simulations. There have been
99 significant efforts to develop practical, efficient and convergent
100 methods for handling these interactions. Ewald's method is perhaps the
101 best known and most accurate method for evaluating energies, forces,
102 and torques in explicitly-periodic simulation cells. In this approach,
103 the conditionally convergent electrostatic energy is converted into
104 two absolutely convergent contributions, one which is carried out in
105 real space with a cutoff radius, and one in reciprocal
106 space.\cite{Ewald21,deLeeuw80,Smith81,Allen87}
107
108 When carried out as originally formulated, the reciprocal-space
109 portion of the Ewald sum exhibits relatively poor computational
110 scaling, making it prohibitive for large systems. By utilizing a
111 particle mesh and three dimensional fast Fourier transforms (FFT), the
112 particle-mesh Ewald (PME), particle-particle particle-mesh Ewald
113 (P\textsuperscript{3}ME), and smooth particle mesh Ewald (SPME)
114 methods can decrease the computational cost from $O(N^2)$ down to $O(N
115 \log
116 N)$.\cite{Takada93,Gunsteren94,Gunsteren95,Darden:1993pd,Essmann:1995pb}
117
118 Because of the artificial periodicity required for the Ewald sum,
119 interfacial molecular systems such as membranes and liquid-vapor
120 interfaces require modifications to the method. Parry's extension of
121 the three dimensional Ewald sum is appropriate for slab
122 geometries.\cite{Parry:1975if} Modified Ewald methods that were
123 developed to handle two-dimensional (2-D) electrostatic
124 interactions.\cite{Parry:1975if,Parry:1976fq,Clarke77,Perram79,Rhee:1989kl}
125 These methods were originally quite computationally
126 expensive.\cite{Spohr97,Yeh99} There have been several successful
127 efforts that reduced the computational cost of 2-D lattice summations,
128 bringing them more in line with the scaling for the full 3-D
129 treatments.\cite{Yeh99,Kawata01,Arnold02,deJoannis02,Brodka04} The
130 inherent periodicity required by the Ewald method can also be
131 problematic in a number of protein/solvent and ionic solution
132 environments.\cite{Roberts94,Roberts95,Luty96,Hunenberger99a,Hunenberger99b,Weber00,Fennell:2006lq}
133
134 \subsection{Real-space methods}
135 Wolf \textit{et al.}\cite{Wolf:1999dn} proposed a real space $O(N)$
136 method for calculating electrostatic interactions between point
137 charges. They argued that the effective Coulomb interaction in most
138 condensed phase systems is effectively short
139 ranged.\cite{Wolf92,Wolf95} For an ordered lattice (e.g., when
140 computing the Madelung constant of an ionic solid), the material can
141 be considered as a set of ions interacting with neutral dipolar or
142 quadrupolar ``molecules'' giving an effective distance dependence for
143 the electrostatic interactions of $r^{-5}$ (see figure
144 \ref{fig:schematic}). If one views the \ce{NaCl} crystal as a simple
145 cubic (SC) structure with an octupolar \ce{(NaCl)4} basis, the
146 electrostatic energy per ion converges more rapidly to the Madelung
147 energy than the dipolar approximation.\cite{Wolf92} To find the
148 correct Madelung constant, Lacman suggested that the NaCl structure
149 could be constructed in a way that the finite crystal terminates with
150 complete \ce{(NaCl)4} molecules.\cite{Lacman65} The central ion sees
151 what is effectively a set of octupoles at large distances. These facts
152 suggest that the Madelung constants are relatively short ranged for
153 perfect ionic crystals.\cite{Wolf:1999dn} For this reason, careful
154 application of Wolf's method can provide accurate estimates of
155 Madelung constants using relatively short cutoff radii.
156
157 Direct truncation of interactions at a cutoff radius creates numerical
158 errors. Wolf \textit{et al.} suggest that truncation errors are due
159 to net charge remaining inside the cutoff sphere.\cite{Wolf:1999dn} To
160 neutralize this charge they proposed placing an image charge on the
161 surface of the cutoff sphere for every real charge inside the cutoff.
162 These charges are present for the evaluation of both the pair
163 interaction energy and the force, although the force expression
164 maintains a discontinuity at the cutoff sphere. In the original Wolf
165 formulation, the total energy for the charge and image were not equal
166 to the integral of the force expression, and as a result, the total
167 energy would not be conserved in molecular dynamics (MD)
168 simulations.\cite{Zahn:2002hc} Zahn \textit{et al.}, and Fennell and
169 Gezelter later proposed shifted force variants of the Wolf method with
170 commensurate force and energy expressions that do not exhibit this
171 problem.\cite{Zahn:2002hc,Fennell:2006lq} Related real-space methods
172 were also proposed by Chen \textit{et
173 al.}\cite{Chen:2004du,Chen:2006ii,Denesyuk:2008ez,Rodgers:2006nw}
174 and by Wu and Brooks.\cite{Wu:044107} Recently, Fukuda has successfully
175 used additional neutralization of higher order moments for systems of
176 point charges.\cite{Fukuda:2013sf}
177
178 \begin{figure}
179 \centering
180 \includegraphics[width=\linewidth]{schematic.eps}
181 \caption{Top: Ionic systems exhibit local clustering of dissimilar
182 charges (in the smaller grey circle), so interactions are
183 effectively charge-multipole at longer distances. With hard
184 cutoffs, motion of individual charges in and out of the cutoff
185 sphere can break the effective multipolar ordering. Bottom:
186 dipolar crystals and fluids have a similar effective
187 \textit{quadrupolar} ordering (in the smaller grey circles), and
188 orientational averaging helps to reduce the effective range of the
189 interactions in the fluid. Placement of reversed image multipoles
190 on the surface of the cutoff sphere recovers the effective
191 higher-order multipole behavior. \label{fig:schematic}}
192 \end{figure}
193
194 One can make a similar effective range argument for crystals of point
195 \textit{multipoles}. The Luttinger and Tisza treatment of energy
196 constants for dipolar lattices utilizes 24 basis vectors that contain
197 dipoles at the eight corners of a unit cube.\cite{LT} Only three of
198 these basis vectors, $X_1, Y_1, \mathrm{~and~} Z_1,$ retain net dipole
199 moments, while the rest have zero net dipole and retain contributions
200 only from higher order multipoles. The lowest-energy crystalline
201 structures are built out of basis vectors that have only residual
202 quadrupolar moments (e.g. the $Z_5$ array). In these low energy
203 structures, the effective interaction between a dipole at the center
204 of a crystal and a group of eight dipoles farther away is
205 significantly shorter ranged than the $r^{-3}$ that one would expect
206 for raw dipole-dipole interactions. Only in crystals which retain a
207 bulk dipole moment (e.g. ferroelectrics) does the analogy with the
208 ionic crystal break down -- ferroelectric dipolar crystals can exist,
209 while ionic crystals with net charge in each unit cell would be
210 unstable.
211
212 In ionic crystals, real-space truncation can break the effective
213 multipolar arrangements (see Fig. \ref{fig:schematic}), causing
214 significant swings in the electrostatic energy as individual ions move
215 back and forth across the boundary. This is why the image charges are
216 necessary for the Wolf sum to exhibit rapid convergence. Similarly,
217 the real-space truncation of point multipole interactions breaks
218 higher order multipole arrangements, and image multipoles are required
219 for real-space treatments of electrostatic energies.
220
221 The shorter effective range of electrostatic interactions is not
222 limited to perfect crystals, but can also apply in disordered fluids.
223 Even at elevated temperatures, there is local charge balance in an
224 ionic liquid, where each positive ion has surroundings dominated by
225 negative ions and vice versa. The reversed-charge images on the
226 cutoff sphere that are integral to the Wolf and damped shifted force
227 (DSF) approaches retain the effective multipolar interactions as the
228 charges traverse the cutoff boundary.
229
230 In multipolar fluids (see Fig. \ref{fig:schematic}) there is
231 significant orientational averaging that additionally reduces the
232 effect of long-range multipolar interactions. The image multipoles
233 that are introduced in the Taylor shifted force (TSF), gradient
234 shifted force (GSF), and shifted potential (SP) methods mimic this effect
235 and reduce the effective range of the multipolar interactions as
236 interacting molecules traverse each other's cutoff boundaries.
237
238 Forces and torques acting on atomic sites are fundamental in driving
239 dynamics in molecular simulations, and the DSF energy kernel provides
240 consistent energies and forces on charged atoms within the cutoff
241 sphere. Both the energy and the force go smoothly to zero as an atom
242 approaches the cutoff radius. The comparisons of the accuracy these
243 quantities between the DSF kernel and SPME was surprisingly
244 good.\cite{Fennell:2006lq} As a result, the DSF method has seen
245 increasing use in molecular systems with relatively uniform charge
246 densities.\cite{English08,Kannam:2012rr,Space12,Lawrence13,Acevedo13,Shi:2013ij,Vergne13}
247
248 \subsection{The damping function}
249 The damping function has been discussed in detail in the first paper
250 of this series.\cite{PaperI} The $1/r$ Coulombic kernel for the
251 interactions between point charges can be replaced by the
252 complementary error function $\mathrm{erfc}(\alpha r)/r$ to accelerate
253 convergence, where $\alpha$ is a damping parameter with units of
254 inverse distance. Altering the value of $\alpha$ is equivalent to
255 changing the width of Gaussian charge distributions that replace each
256 point charge, as Coulomb integrals with Gaussian charge distributions
257 produce complementary error functions when truncated at a finite
258 distance.
259
260 With moderate damping coefficients, $\alpha \sim 0.2$ \AA$^{-1}$, the DSF method
261 produced very good agreement with SPME for interaction energies,
262 forces and torques for charge-charge
263 interactions.\cite{Fennell:2006lq}
264
265 \subsection{Point multipoles in molecular modeling}
266 Coarse-graining approaches which treat entire molecular subsystems as
267 a single rigid body are now widely used. A common feature of many
268 coarse-graining approaches is simplification of the electrostatic
269 interactions between bodies so that fewer site-site interactions are
270 required to compute configurational
271 energies.\cite{Ren06,Essex10,Essex11}
272
273 Additionally, because electrons in a molecule are not localized at
274 specific points, the assignment of partial charges to atomic centers
275 is always an approximation. For increased accuracy, atomic sites can
276 also be assigned point multipoles and polarizabilities. Recently,
277 water has been modeled with point multipoles up to octupolar order
278 using the soft sticky dipole-quadrupole-octupole (SSDQO)
279 model.\cite{Ichiye10_1,Ichiye10_2,Ichiye10_3} Atom-centered point
280 multipoles up to quadrupolar order have also been coupled with point
281 polarizabilities in the high-quality AMOEBA and iAMOEBA water
282 models.\cite{Ren:2003uq,Ren:2004kx,Ponder:2010vl,Wang:2013fk} However,
283 truncating point multipoles without smoothing the forces and torques
284 can create energy conservation issues in molecular dynamics
285 simulations.
286
287 In this paper we test a set of real-space methods that were developed
288 for point multipolar interactions. These methods extend the damped
289 shifted force (DSF) and Wolf methods originally developed for
290 charge-charge interactions and generalize them for higher order
291 multipoles. The detailed mathematical development of these methods
292 has been presented in the first paper in this series, while this work
293 covers the testing of energies, forces, torques, and energy
294 conservation properties of the methods in realistic simulation
295 environments. In all cases, the methods are compared with the
296 reference method, a full multipolar Ewald treatment.\cite{Smith82,Smith98}
297
298
299 \section{\label{sec:method}Review of Methods}
300 Any real-space electrostatic method that is suitable for MD
301 simulations should have the electrostatic energy, forces and torques
302 between two sites go smoothly to zero as the distance between the
303 sites, $r_{ab}$ approaches the cutoff radius, $r_c$. Requiring
304 this continuity at the cutoff is essential for energy conservation in
305 MD simulations. The mathematical details of the shifted potential
306 (SP), gradient-shifted-force (GSF) and Taylor shifted-force (TSF)
307 methods have been discussed in detail in the previous paper in this
308 series.\cite{PaperI} Here we briefly review the new methods and
309 describe their essential features.
310
311 \subsection{Taylor-shifted force (TSF)}
312
313 The electrostatic potential energy between point multipoles can be
314 expressed as the product of two multipole operators and a Coulombic
315 kernel,
316 \begin{equation}
317 U_{ab}(r)= M_{a} M_{b} \frac{1}{r} \label{kernel}.
318 \end{equation}
319 where the multipole operator for site $a$, $M_{a}$, is
320 expressed in terms of the point charge, $C_{a}$, dipole, ${\bf D}_{a}$, and quadrupole, $\mathsf{Q}_{a}$, for object
321 $a$, etc.
322
323 The TSF potential for any multipole-multipole interaction can be
324 written
325 \begin{equation}
326 U^{\text{TSF}}= (\text{prefactor}) (\text{derivatives}) f_n(r)
327 \label{generic}
328 \end{equation}
329 where $f_n(r)$ is a shifted kernel that is appropriate for the order
330 of the interaction (see Ref. \onlinecite{PaperI}), with $n=0$ for
331 charge-charge, $n=1$ for charge-dipole, $n=2$ for charge-quadrupole
332 and dipole-dipole, $n=3$ for dipole-quadrupole, and $n=4$ for
333 quadrupole-quadrupole. To ensure smooth convergence of the energy,
334 force, and torques, a Taylor expansion with $n$ terms must be
335 performed at cutoff radius ($r_c$) to obtain $f_n(r)$.
336
337 For multipole-multipole interactions, following this procedure results
338 in separate radial functions for each of the distinct orientational
339 contributions to the potential, and ensures that the forces and
340 torques from each of these contributions will vanish at the cutoff
341 radius. For example, the direct dipole dot product
342 ($\mathbf{D}_{a}
343 \cdot \mathbf{D}_{b}$) is treated differently than the dipole-distance
344 dot products:
345 \begin{equation}
346 U_{\mathbf{D}_{a}\mathbf{D}_{b}}(r)= -\frac{1}{4\pi \epsilon_0} \left[ \left(
347 \mathbf{D}_{a} \cdot
348 \mathbf{D}_{b} \right) v_{21}(r) +
349 \left( \mathbf{D}_{a} \cdot \hat{\mathbf{r}} \right)
350 \left( \mathbf{D}_{b} \cdot \hat{\mathbf{r}} \right) v_{22}(r) \right]
351 \end{equation}
352
353 For the Taylor shifted (TSF) method with the undamped kernel,
354 $v_{21}(r) = -\frac{1}{r^3} + \frac{3 r}{r_c^4} - \frac{8}{r_c^3} +
355 \frac{6}{r r_c^2}$ and $v_{22}(r) = \frac{3}{r^3} + \frac{3 r}{r_c^4}
356 - \frac{6}{r r_c^2}$. In these functions, one can easily see the
357 connection to unmodified electrostatics as well as the smooth
358 transition to zero in both these functions as $r\rightarrow r_c$. The
359 electrostatic forces and torques acting on the central multipole due
360 to another site within the cutoff sphere are derived from
361 Eq.~\ref{generic}, accounting for the appropriate number of
362 derivatives. Complete energy, force, and torque expressions are
363 presented in the first paper in this series (Reference
364 \onlinecite{PaperI}).
365
366 \subsection{Gradient-shifted force (GSF)}
367
368 A second (and conceptually simpler) method involves shifting the
369 gradient of the raw Coulomb potential for each particular multipole
370 order. For example, the raw dipole-dipole potential energy may be
371 shifted smoothly by finding the gradient for two interacting dipoles
372 which have been projected onto the surface of the cutoff sphere
373 without changing their relative orientation,
374 \begin{equation}
375 U_{\mathbf{D}_{a}\mathbf{D}_{b}}(r) = U_{\mathbf{D}_{a}\mathbf{D}_{b}}(r) -
376 U_{\mathbf{D}_{a}\mathbf{D}_{b}}(r_c)
377 - (r_{ab}-r_c) ~~~\hat{\mathbf{r}}_{ab} \cdot
378 \nabla U_{\mathbf{D}_{a}\mathbf{D}_{b}}(r_c).
379 \end{equation}
380 Here the lab-frame orientations of the two dipoles, $\mathbf{D}_{a}$ and $\mathbf{D}_{b}$, are retained at the cutoff distance
381 (although the signs are reversed for the dipole that has been
382 projected onto the cutoff sphere). In many ways, this simpler
383 approach is closer in spirit to the original shifted force method, in
384 that it projects a neutralizing multipole (and the resulting forces
385 from this multipole) onto a cutoff sphere. The resulting functional
386 forms for the potentials, forces, and torques turn out to be quite
387 similar in form to the Taylor-shifted approach, although the radial
388 contributions are significantly less perturbed by the gradient-shifted
389 approach than they are in the Taylor-shifted method.
390
391 For the gradient shifted (GSF) method with the undamped kernel,
392 $v_{21}(r) = -\frac{1}{r^3} - \frac{3 r}{r_c^4} + \frac{4}{r_c^3}$ and
393 $v_{22}(r) = \frac{3}{r^3} + \frac{9 r}{r_c^4} - \frac{12}{r_c^3}$.
394 Again, these functions go smoothly to zero as $r\rightarrow r_c$, and
395 because the Taylor expansion retains only one term, they are
396 significantly less perturbed than the TSF functions.
397
398 In general, the gradient shifted potential between a central multipole
399 and any multipolar site inside the cutoff radius is given by,
400 \begin{equation}
401 U^{\text{GSF}} = \sum \left[ U(\mathbf{r}, \mathsf{A}, \mathsf{B}) -
402 U(r_c \hat{\mathbf{r}},\mathsf{A}, \mathsf{B}) - (r-r_c)
403 \hat{\mathbf{r}} \cdot \nabla U(r_c \hat{\mathbf{r}},\mathsf{A}, \mathsf{B}) \right]
404 \label{generic2}
405 \end{equation}
406 where the sum describes a separate force-shifting that is applied to
407 each orientational contribution to the energy. In this expression,
408 $\hat{\mathbf{r}}$ is the unit vector connecting the two multipoles
409 ($a$ and $b$) in space, and $\mathsf{A}$ and $\mathsf{B}$
410 represent the orientations the multipoles.
411
412 The third term converges more rapidly than the first two terms as a
413 function of radius, hence the contribution of the third term is very
414 small for large cutoff radii. The force and torque derived from
415 Eq. \ref{generic2} are consistent with the energy expression and
416 approach zero as $r \rightarrow r_c$. Both the GSF and TSF methods
417 can be considered generalizations of the original DSF method for
418 higher order multipole interactions. GSF and TSF are also identical up
419 to the charge-dipole interaction but generate different expressions in
420 the energy, force and torque for higher order multipole-multipole
421 interactions. Complete energy, force, and torque expressions for the
422 GSF potential are presented in the first paper in this series
423 (Reference~\onlinecite{PaperI}).
424
425
426 \subsection{Shifted potential (SP) }
427 A discontinuous truncation of the electrostatic potential at the
428 cutoff sphere introduces a severe artifact (oscillation in the
429 electrostatic energy) even for molecules with the higher-order
430 multipoles.\cite{PaperI} We have also formulated an extension of the
431 Wolf approach for point multipoles by simply projecting the image
432 multipole onto the surface of the cutoff sphere, and including the
433 interactions with the central multipole and the image. This
434 effectively shifts the total potential to zero at the cutoff radius,
435 \begin{equation}
436 U^{\text{SP}} = \sum \left[ U(\mathbf{r}, \mathsf{A}, \mathsf{B}) -
437 U(r_c \hat{\mathbf{r}},\mathsf{A}, \mathsf{B}) \right]
438 \label{eq:SP}
439 \end{equation}
440 where the sum describes separate potential shifting that is done for
441 each orientational contribution to the energy (e.g. the direct dipole
442 product contribution is shifted {\it separately} from the
443 dipole-distance terms in dipole-dipole interactions). Note that this
444 is not a simple shifting of the total potential at $r_c$. Each radial
445 contribution is shifted separately. One consequence of this is that
446 multipoles that reorient after leaving the cutoff sphere can re-enter
447 the cutoff sphere without perturbing the total energy.
448
449 For the shifted potential (SP) method with the undamped kernel,
450 $v_{21}(r) = -\frac{1}{r^3} + \frac{1}{r_c^3}$ and $v_{22}(r) =
451 \frac{3}{r^3} - \frac{3}{r_c^3}$. The potential energy between a
452 central multipole and other multipolar sites goes smoothly to zero as
453 $r \rightarrow r_c$. However, the force and torque obtained from the
454 shifted potential (SP) are discontinuous at $r_c$. MD simulations
455 will still experience energy drift while operating under the SP
456 potential, but it may be suitable for Monte Carlo approaches where the
457 configurational energy differences are the primary quantity of
458 interest.
459
460 \subsection{The Self Term}
461 In the TSF, GSF, and SP methods, a self-interaction is retained for
462 the central multipole interacting with its own image on the surface of
463 the cutoff sphere. This self interaction is nearly identical with the
464 self-terms that arise in the Ewald sum for multipoles. Complete
465 expressions for the self terms are presented in the first paper in
466 this series (Reference \onlinecite{PaperI}).
467
468
469 \section{\label{sec:methodology}Methodology}
470
471 To understand how the real-space multipole methods behave in computer
472 simulations, it is vital to test against established methods for
473 computing electrostatic interactions in periodic systems, and to
474 evaluate the size and sources of any errors that arise from the
475 real-space cutoffs. In the first paper of this series, we compared
476 the dipolar and quadrupolar energy expressions against analytic
477 expressions for ordered dipolar and quadrupolar
478 arrays.\cite{Sauer,LT,Nagai01081960,Nagai01091963} In this work, we
479 used the multipolar Ewald sum as a reference method for comparing
480 energies, forces, and torques for molecular models that mimic
481 disordered and ordered condensed-phase systems. The parameters used
482 in the test cases are given in table~\ref{tab:pars}.
483
484 \begin{table}
485 \caption{The parameters used in the systems used to evaluate the new
486 real-space methods. The most comprehensive test was a liquid
487 composed of 2000 soft DQ liquid molecules with 48 dissolved ions (24 \ce{Na+} and 24 \ce{Cl-}
488 ions). This test exercises all orders of the multipolar
489 interactions developed in the first paper.\label{tab:pars}}
490 \begin{tabularx}{\textwidth}{r|cc|YYccc|Yccc} \hline
491 & \multicolumn{2}{c|}{LJ parameters} &
492 \multicolumn{5}{c|}{Electrostatic moments} & & & & \\
493 Test system & $\sigma$& $\epsilon$ & $C$ & $D$ &
494 $Q_{xx}$ & $Q_{yy}$ & $Q_{zz}$ & mass & $I_{xx}$ & $I_{yy}$ &
495 $I_{zz}$ \\ \cline{6-8}\cline{10-12}
496 & (\AA) & (kcal/mol) & (e) & (debye) & \multicolumn{3}{c|}{(debye \AA)} & (amu) & \multicolumn{3}{c}{(amu
497 \AA\textsuperscript{2})} \\ \hline
498 Soft Dipolar fluid & 3.051 & 0.152 & & 2.35 & & & & 18.0153 & 1.77&0.6145& 1.155 \\
499 Soft Dipolar solid & 2.837 & 1.0 & & 2.35 & & & & $10^4$ & 17.6 &17.6 & 0 \\
500 Soft Quadrupolar fluid & 3.051 & 0.152 & & & -1&-1&-2.5 & 18.0153 & 1.77&0.6145&1.155 \\
501 Soft Quadrupolar solid & 2.837 & 1.0 & & & -1&-1&-2.5 & $10^4$ & 17.6&17.6&0 \\
502 Soft DQ liquid & 3.051 & 0.152 & & 2.35 & -1.35&0&-0.68 & 18.0153 & 1.77&0.6145&1.155 \\
503 \ce{Na+} & 2.579 & 0.118 & +1& & & & & 22.99 & & &\\
504 \ce{Cl-} & 4.445 & 0.1 & -1& & & & & 35.4527& & & \\ \hline
505 \end{tabularx}
506 \end{table}
507 The systems consist of pure multipolar solids (both dipole and
508 quadrupole), pure multipolar liquids (both dipole and quadrupole), a
509 fluid composed of sites containing both dipoles and quadrupoles
510 simultaneously, and a final test case that includes ions with point
511 charges in addition to the multipolar fluid. The solid-phase
512 parameters were chosen so that the systems can explore some
513 orientational freedom for the multipolar sites, while maintaining
514 relatively strict translational order. The soft DQ liquid model used
515 here based loosely on the SSDQO water
516 model,\cite{Ichiye10_1,Ichiye10_2,Ichiye10_3} but is not itself a
517 particularly accurate water model. However, the soft DQ model does
518 test dipole-dipole, dipole-quadrupole, and quadrupole-quadrupole
519 interactions at roughly the same magnitudes. The last test case, a
520 soft DQ liquid with dissolved ions, exercises \textit{all} levels of
521 the multipole-multipole interactions we have derived so far and
522 represents the most complete test of the new methods.
523
524 In the following section, we present results for the total
525 electrostatic energy, as well as the electrostatic contributions to
526 the force and torque on each molecule. These quantities have been
527 computed using the SP, TSF, and GSF methods, as well as a hard cutoff,
528 and have been compared with the values obtained from the multipolar
529 Ewald sum. In Monte Carlo (MC) simulations, the energy differences
530 between two configurations is the primary quantity that governs how
531 the simulation proceeds. These differences are the most important
532 indicators of the reliability of a method even if the absolute
533 energies are not exact. For each of the multipolar systems listed
534 above, we have compared the change in electrostatic potential energy
535 ($\Delta E$) between 250 statistically-independent configurations. In
536 molecular dynamics (MD) simulations, the forces and torques govern the
537 behavior of the simulation, so we also compute the electrostatic
538 contributions to the forces and torques.
539
540 \subsection{Implementation}
541 The real-space methods developed in the first paper in this series
542 have been implemented in our group's open source molecular simulation
543 program, OpenMD,\cite{Meineke05,openmd} which was used for all calculations in
544 this work. The complementary error function can be a relatively slow
545 function on some processors, so all of the radial functions are
546 precomputed on a fine grid and are spline-interpolated to provide
547 values when required.
548
549 Using the same simulation code, we compare to a multipolar Ewald sum
550 with a reciprocal space cutoff, $k_\mathrm{max} = 7$. Our version of
551 the Ewald sum is a re-implementation of the algorithm originally
552 proposed by Smith that does not use the particle mesh or smoothing
553 approximations.\cite{Smith82,Smith98} This implementation was tested
554 extensively against the analytic energy constants for the multipolar
555 lattices that are discussed in reference \onlinecite{PaperI}. In all
556 cases discussed below, the quantities being compared are the
557 electrostatic contributions to energies, force, and torques. All
558 other contributions to these quantities (i.e. from Lennard-Jones
559 interactions) are removed prior to the comparisons.
560
561 The convergence parameter ($\alpha$) also plays a role in the balance
562 of the real-space and reciprocal-space portions of the Ewald
563 calculation. Typical molecular mechanics packages set this to a value
564 that depends on the cutoff radius and a tolerance (typically less than
565 $1 \times 10^{-4}$ kcal/mol). Smaller tolerances are typically
566 associated with increasing accuracy at the expense of computational
567 time spent on the reciprocal-space portion of the
568 summation.\cite{Perram88,Essmann:1995pb} A default tolerance of $1 \times
569 10^{-8}$ kcal/mol was used in all Ewald calculations, resulting in
570 Ewald coefficient 0.3119 \AA$^{-1}$ for a cutoff radius of 12 \AA.
571
572 The real-space models have self-interactions that provide
573 contributions to the energies only. Although the self interaction is
574 a rapid calculation, we note that in systems with fluctuating charges
575 or point polarizabilities, the self-term is not static and must be
576 recomputed at each time step.
577
578 \subsection{Model systems}
579 To sample independent configurations of the multipolar crystals, body
580 centered cubic (bcc) crystals, which exhibit the minimum energy
581 structures for point dipoles, were generated using 3,456 molecules.
582 The multipoles were translationally locked in their respective crystal
583 sites for equilibration at a relatively low temperature (50K) so that
584 dipoles or quadrupoles could freely explore all accessible
585 orientations. The translational constraints were then removed, the
586 systems were re-equilibrated, and the crystals were simulated for an
587 additional 10 ps in the microcanonical (NVE) ensemble with an average
588 temperature of 50 K. The balance between moments of inertia and
589 particle mass were chosen to allow orientational sampling without
590 significant translational motion. Configurations were sampled at
591 equal time intervals in order to compare configurational energy
592 differences. The crystals were simulated far from the melting point
593 in order to avoid translational deformation away of the ideal lattice
594 geometry.
595
596 For dipolar, quadrupolar, and mixed-multipole \textit{liquid}
597 simulations, each system was created with 2,048 randomly-oriented
598 molecules. These were equilibrated at a temperature of 300K for 1 ns.
599 Each system was then simulated for 1 ns in the microcanonical (NVE)
600 ensemble with the Dullweber, Leimkuhler, and McLachlan (DLM)
601 symplectic splitting integrator using 1 fs
602 timesteps.\cite{Dullweber1997} We collected 250 different
603 configurations at equal time intervals. For the liquid system that
604 included ionic species, we converted 48 randomly-distributed molecules
605 into 24 \ce{Na+} and 24 \ce{Cl-} ions and re-equilibrated. After
606 equilibration, the system was run under the same conditions for 1
607 ns. A total of 250 configurations were collected. In the following
608 comparisons of energies, forces, and torques, the Lennard-Jones
609 potentials were turned off and only the purely electrostatic
610 quantities were compared with the same values obtained via the Ewald
611 sum.
612
613 \subsection{Accuracy of Energy Differences, Forces and Torques}
614 The pairwise summation techniques (outlined above) were evaluated for
615 use in MC simulations by studying the energy differences between
616 different configurations. We took the Ewald-computed energy
617 difference between two conformations to be the correct behavior. An
618 ideal performance by one of the new methods would reproduce these
619 energy differences exactly. The configurational energies being used
620 here contain only contributions from electrostatic interactions.
621 Lennard-Jones interactions were omitted from the comparison as they
622 should be identical for all methods.
623
624 Since none of the real-space methods provide exact energy differences,
625 we used least square regressions analysis for the six different
626 molecular systems to compare $\Delta E$ from Hard, SP, GSF, and TSF
627 with the multipolar Ewald reference method. A result of unity for
628 both the correlation (slope) and coefficient of determination ($R^2$)
629 for these regressions would indicate perfect agreement between the
630 real-space method and the multipolar Ewald sum.
631
632 Molecular systems were run long enough to explore independent
633 configurations and 250 configurations were recorded for comparison.
634 Each system provided 31,125 energy differences for a total of 186,750
635 data points. Similarly, the magnitudes of the forces and torques have
636 also been compared using least squares regression analysis. In the
637 forces and torques comparison, the magnitudes of the forces acting in
638 each molecule for each configuration were evaluated. For example, our
639 dipolar liquid simulation contains 2048 molecules and there are 250
640 different configurations for each system resulting in 3,072,000 data
641 points for comparison of forces and torques.
642
643 \subsection{Analysis of vector quantities}
644 Getting the magnitudes of the force and torque vectors correct is only
645 part of the issue for carrying out accurate molecular dynamics
646 simulations. Because the real space methods reweight the different
647 orientational contributions to the energies, it is also important to
648 understand how the methods impact the \textit{directionality} of the
649 force and torque vectors. Fisher developed a probability density
650 function to analyse directional data sets,
651 \begin{equation}
652 p_f(\theta) = \frac{\kappa}{2 \sinh\kappa}\sin\theta e^{\kappa \cos\theta}
653 \label{eq:pdf}
654 \end{equation}
655 where $\kappa$ measures directional dispersion of the data around the
656 mean direction.\cite{fisher53} This quantity $(\kappa)$ can be
657 estimated as a reciprocal of the circular variance.\cite{Allen91} To
658 quantify the directional error, forces obtained from the Ewald sum
659 were taken as the mean (or correct) direction and the angle between
660 the forces obtained via the Ewald sum and the real-space methods were
661 evaluated,
662 \begin{equation}
663 \cos\theta_i = \frac{\mathbf{f}_i^\mathrm{~Ewald} \cdot
664 \mathbf{f}_i^\mathrm{~GSF}}{\left|\mathbf{f}_i^\mathrm{~Ewald}\right| \left|\mathbf{f}_i^\mathrm{~GSF}\right|}
665 \end{equation}
666 The total angular displacement of the vectors was calculated as,
667 \begin{equation}
668 R = \sqrt{\left(\sum\limits_{i=1}^N \cos\theta_i\right)^2 + \left(\sum\limits_{i=1}^N \sin\theta_i\right)^2}
669 \label{eq:displacement}
670 \end{equation}
671 where $N$ is number of force vectors. The circular variance is
672 defined as
673 \begin{equation}
674 \mathrm{Var}(\theta) \approx 1/\kappa = 1 - R/N
675 \end{equation}
676 The circular variance takes on values between from 0 to 1, with 0
677 indicating a perfect directional match between the Ewald force vectors
678 and the real-space forces. Lower values of $\mathrm{Var}(\theta)$
679 correspond to higher values of $\kappa$, which indicates tighter
680 clustering of the real-space force vectors around the Ewald forces.
681
682 A similar analysis was carried out for the electrostatic contribution
683 to the molecular torques as well as forces.
684
685 \subsection{Energy conservation}
686 To test conservation the energy for the methods, the mixed molecular
687 system of 2000 soft DQ liquid molecules with 24 \ce{Na+} and 24
688 \ce{Cl-} ions was run for 1 ns in the microcanonical ensemble at an
689 average temperature of 300K. Each of the different electrostatic
690 methods (Ewald, Hard, SP, GSF, and TSF) was tested for a range of
691 different damping values. The molecular system was started with same
692 initial positions and velocities for all cutoff methods. The energy
693 drift ($\delta E_1$) and standard deviation of the energy about the
694 slope ($\delta E_0$) were evaluated from the total energy of the
695 system as a function of time. Although both measures are valuable at
696 investigating new methods for molecular dynamics, a useful interaction
697 model must allow for long simulation times with minimal energy drift.
698
699 \section{\label{sec:result}RESULTS}
700 \subsection{Configurational energy differences}
701
702 \begin{figure}
703 \centering
704 \includegraphics[width=0.6\linewidth]{energyPlot_slopeCorrelation_combined.eps}
705 \caption{Statistical analysis of the quality of configurational
706 energy differences for the real-space electrostatic methods
707 compared with the reference Ewald sum. Results with a value equal
708 to 1 (dashed line) indicate $\Delta E$ values indistinguishable
709 from those obtained using the multipolar Ewald sum. Different
710 values of the cutoff radius are indicated with different symbols
711 (9~\AA\ = circles, 12~\AA\ = squares, and 15~\AA\ = inverted
712 triangles).\label{fig:slopeCorr_energy}}
713 \end{figure}
714
715 The combined coefficient of determination and slope for all six
716 systems is shown in Figure ~\ref{fig:slopeCorr_energy}. Most of the
717 methods reproduce the Ewald configurational energy differences with
718 remarkable fidelity. Undamped hard cutoffs introduce a significant
719 amount of random scatter in the energy differences which is apparent
720 in the reduced value of $R^2$ for this method. This can be easily
721 understood as configurations which exhibit small traversals of a few
722 dipoles or quadrupoles out of the cutoff sphere will see large energy
723 jumps when hard cutoffs are used. The orientations of the multipoles
724 (particularly in the ordered crystals) mean that these energy jumps
725 can go in either direction, producing a significant amount of random
726 scatter, but no systematic error.
727
728 The TSF method produces energy differences that are highly correlated
729 with the Ewald results, but it also introduces a significant
730 systematic bias in the values of the energies, particularly for
731 smaller cutoff values. The TSF method alters the distance dependence
732 of different orientational contributions to the energy in a
733 non-uniform way, so the size of the cutoff sphere can have a large
734 effect, particularly for the crystalline systems.
735
736 Both the SP and GSF methods appear to reproduce the Ewald results with
737 excellent fidelity, particularly for moderate damping ($\alpha \approx
738 0.2$~\AA$^{-1}$) and with a commonly-used cutoff value ($r_c =
739 12$~\AA). With the exception of the undamped hard cutoff, and the TSF
740 method with short cutoffs, all of the methods would be appropriate for
741 use in Monte Carlo simulations.
742
743 \subsection{Magnitude of the force and torque vectors}
744
745 The comparisons of the magnitudes of the forces and torques for the
746 data accumulated from all six systems are shown in Figures
747 ~\ref{fig:slopeCorr_force} and \ref{fig:slopeCorr_torque}. The
748 correlation and slope for the forces agree well with the Ewald sum
749 even for the hard cutoffs.
750
751 For systems of molecules with only multipolar interactions, the pair
752 energy contributions are quite short ranged. Moreover, the force
753 decays more rapidly than the electrostatic energy, hence the hard
754 cutoff method can also produce reasonable agreement for this quantity.
755 Although the pure cutoff gives reasonably good electrostatic forces
756 for pairs of molecules included within each other's cutoff spheres,
757 the discontinuity in the force at the cutoff radius can potentially
758 cause energy conservation problems as molecules enter and leave the
759 cutoff spheres. This is discussed in detail in section
760 \ref{sec:conservation}.
761
762 The two shifted-force methods (GSF and TSF) exhibit a small amount of
763 systematic variation and scatter compared with the Ewald forces. The
764 shifted-force models intentionally perturb the forces between pairs of
765 molecules inside each other's cutoff spheres in order to correct the
766 energy conservation issues, and this perturbation is evident in the
767 statistics accumulated for the molecular forces. The GSF
768 perturbations are minimal, particularly for moderate damping and
769 commonly-used cutoff values ($r_c = 12$~\AA). The TSF method shows
770 reasonable agreement in $R^2$, but again the systematic error in the
771 forces is concerning if replication of Ewald forces is desired.
772
773 It is important to note that the forces and torques from the SP and
774 the Hard cutoffs are not identical. The SP method shifts each
775 orientational contribution separately (e.g. the dipole-dipole dot
776 product is shifted by a different function than the dipole-distance
777 products), while the hard cutoff contains no orientation-dependent
778 shifting. The forces and torques for these methods therefore diverge
779 for multipoles even though the forces for point charges are identical.
780
781 \begin{figure}
782 \centering
783 \includegraphics[width=0.6\linewidth]{forcePlot_slopeCorrelation_combined.eps}
784 \caption{Statistical analysis of the quality of the force vector
785 magnitudes for the real-space electrostatic methods compared with
786 the reference Ewald sum. Results with a value equal to 1 (dashed
787 line) indicate force magnitude values indistinguishable from those
788 obtained using the multipolar Ewald sum. Different values of the
789 cutoff radius are indicated with different symbols (9~\AA\ =
790 circles, 12~\AA\ = squares, and 15~\AA\ = inverted
791 triangles).\label{fig:slopeCorr_force}}
792 \end{figure}
793
794
795 \begin{figure}
796 \centering
797 \includegraphics[width=0.6\linewidth]{torquePlot_slopeCorrelation_combined.eps}
798 \caption{Statistical analysis of the quality of the torque vector
799 magnitudes for the real-space electrostatic methods compared with
800 the reference Ewald sum. Results with a value equal to 1 (dashed
801 line) indicate force magnitude values indistinguishable from those
802 obtained using the multipolar Ewald sum. Different values of the
803 cutoff radius are indicated with different symbols (9~\AA\ =
804 circles, 12~\AA\ = squares, and 15~\AA\ = inverted
805 triangles).\label{fig:slopeCorr_torque}}
806 \end{figure}
807
808 The torques (Fig. \ref{fig:slopeCorr_torque}) appear to be
809 significantly influenced by the choice of real-space method. The
810 torque expressions have the same distance dependence as the energies,
811 which are naturally longer-ranged expressions than the inter-site
812 forces. Torques are also quite sensitive to orientations of
813 neighboring molecules, even those that are near the cutoff distance.
814
815 The results shows that the torque from the hard cutoff method
816 reproduces the torques in quite good agreement with the Ewald sum.
817 The other real-space methods can cause some deviations, but excellent
818 agreement with the Ewald sum torques is recovered at moderate values
819 of the damping coefficient ($\alpha \approx 0.2$~\AA$^{-1}$) and cutoff
820 radius ($r_c \ge 12$~\AA). The TSF method exhibits only fair agreement
821 in the slope when compared with the Ewald torques even for larger
822 cutoff radii. It appears that the severity of the perturbations in
823 the TSF method are most in evidence for the torques.
824
825 \subsection{Directionality of the force and torque vectors}
826
827 The accurate evaluation of force and torque directions is just as
828 important for molecular dynamics simulations as the magnitudes of
829 these quantities. Force and torque vectors for all six systems were
830 analyzed using Fisher statistics, and the quality of the vector
831 directionality is shown in terms of circular variance
832 ($\mathrm{Var}(\theta)$) in
833 Fig. \ref{fig:slopeCorr_circularVariance}. The force and torque
834 vectors from the new real-space methods exhibit nearly-ideal Fisher
835 probability distributions (Eq.~\ref{eq:pdf}). Both the hard and SP
836 cutoff methods exhibit the best vectorial agreement with the Ewald
837 sum. The force and torque vectors from GSF method also show good
838 agreement with the Ewald method, which can also be systematically
839 improved by using moderate damping and a reasonable cutoff radius. For
840 $\alpha = 0.2$~\AA$^{-1}$ and $r_c = 12$~\AA, we observe
841 $\mathrm{Var}(\theta) = 0.00206$, which corresponds to a distribution
842 with 95\% of force vectors within $6.37^\circ$ of the corresponding
843 Ewald forces. The TSF method produces the poorest agreement with the
844 Ewald force directions.
845
846 Torques are again more perturbed than the forces by the new real-space
847 methods, but even here the variance is reasonably small. For the same
848 method (GSF) with the same parameters ($\alpha = 0.2$~\AA$^{-1}$, $r_c
849 = 12$~\AA), the circular variance was 0.01415, corresponds to a
850 distribution which has 95\% of torque vectors are within $16.75^\circ$
851 of the Ewald results. Again, the direction of the force and torque
852 vectors can be systematically improved by varying $\alpha$ and $r_c$.
853
854 \begin{figure}
855 \centering
856 \includegraphics[width=0.65\linewidth]{Variance_forceNtorque_modified.eps}
857 \caption{The circular variance of the direction of the force and
858 torque vectors obtained from the real-space methods around the
859 reference Ewald vectors. A variance equal to 0 (dashed line)
860 indicates direction of the force or torque vectors are
861 indistinguishable from those obtained from the Ewald sum. Here
862 different symbols represent different values of the cutoff radius
863 (9~\AA\ = circle, 12~\AA\ = square, 15~\AA\ = inverted triangle)\label{fig:slopeCorr_circularVariance}}
864 \end{figure}
865
866 \subsection{Energy conservation\label{sec:conservation}}
867
868 We have tested the conservation of energy one can expect to see with
869 the new real-space methods using the soft DQ liquid model with a small
870 fraction of solvated ions. This is a test system which exercises all
871 orders of multipole-multipole interactions derived in the first paper
872 in this series and provides the most comprehensive test of the new
873 methods. A liquid-phase system was created with 2000 liquid-phase
874 molecules and 48 dissolved ions at a density of 0.98 g cm$^{-3}$ and a
875 temperature of 300K. After equilibration in the canonical (NVT)
876 ensemble using a Nos\'e-Hoover thermostat, six
877 statistically-independent replicas of this liquid-phase system were
878 run in the microcanonical (NVE) ensemble under the Ewald, Hard, SP,
879 GSF, and TSF methods with a cutoff radius of 12~\AA. The value of the
880 damping coefficient was also varied from the undamped case ($\alpha =
881 0$) to a heavily damped case ($\alpha = 0.3$~\AA$^{-1}$) for all of
882 the real space methods. A sample was also run using the multipolar
883 Ewald sum with the same real-space cutoff.
884
885 In figure~\ref{fig:energyDrift} we show the both the linear drift in
886 energy over time, $\delta E_1$, and the standard deviation of energy
887 fluctuations around this drift $\delta E_0$. Both of the
888 shifted-force methods (GSF and TSF) provide excellent energy
889 conservation (drift less than $10^{-5}$ kcal / mol / ns / particle),
890 while the hard cutoff is essentially unusable for molecular dynamics.
891 SP provides some benefit over the hard cutoff because the energetic
892 jumps that happen as particles leave and enter the cutoff sphere are
893 somewhat reduced, but like the Wolf method for charges, the SP method
894 would not be as useful for molecular dynamics as either of the
895 shifted-force methods.
896
897 We note that for all tested values of the cutoff radius, the new
898 real-space methods can provide better energy conservation behavior
899 than the multipolar Ewald sum, even when relatively large $k$-space
900 cutoff values are utilized.
901
902 \begin{figure}
903 \centering
904 \includegraphics[width=\textwidth]{finalDrift.eps}
905 \caption{Energy conservation of the real-space methods for the soft
906 DQ liquid / ion system. $\delta \mathrm{E}_1$ is the linear drift
907 in energy over time (in kcal/mol/particle/ns) and $\delta
908 \mathrm{E}_0$ is the standard deviation of energy fluctuations
909 around this drift (in kcal/mol/particle). Points that appear in
910 the green region at the bottom exhibit better energy conservation
911 than would be obtained using common parameters for Ewald-based
912 electrostatics.\label{fig:energyDrift}}
913 \end{figure}
914
915 \subsection{Reproduction of Structural \& Dynamical Features\label{sec:structure}}
916 The most important test of the modified interaction potentials is the
917 fidelity with which they can reproduce structural features and
918 dynamical properties in a liquid. One commonly-utilized measure of
919 structural ordering is the pair distribution function, $g(r)$, which
920 measures local density deviations in relation to the bulk density. In
921 the electrostatic approaches studied here, the short-range repulsion
922 from the Lennard-Jones potential is identical for the various
923 electrostatic methods, and since short range repulsion determines much
924 of the local liquid ordering, one would not expect to see many
925 differences in $g(r)$. Indeed, the pair distributions are essentially
926 identical for all of the electrostatic methods studied (for each of
927 the different systems under investigation).
928
929 % An example of this agreement for the soft DQ liquid/ion system is
930 % shown in Fig. \ref{fig:gofr}.
931
932 % \begin{figure}
933 % \centering
934 % \includegraphics[width=\textwidth]{gofr_ssdqc.eps}
935 % \caption{The pair distribution functions, $g(r)$, for the SSDQ
936 % water/ion system obtained using the different real-space methods are
937 % essentially identical with the result from the Ewald
938 % treatment.\label{fig:gofr}}
939 % \end{figure}
940
941 There is a minor over-structuring of the first solvation shell when
942 using TSF or when overdamping with any of the real-space methods.
943 With moderate damping, GSF and SP produce pair distributions that are
944 identical (within numerical noise) to their Ewald counterparts. The
945 degree of over-structuring can be measured most easily using the
946 coordination number,
947 \begin{equation}
948 n_C = 4\pi\rho \int_{0}^{a}r^2\text{g}(r)dr,
949 \end{equation}
950 where $\rho$ is the number density of the site-site pair interactions,
951 and $a$ is the radial location of the minima following the first peak
952 in $g(r)$ ($a = 4.2$~\AA\ for the soft DQ liquid / ion system). The
953 coordination number is shown as a function of the damping coefficient
954 for all of the real space methods in Fig. \ref{fig:Props}.
955
956 A more demanding test of modified electrostatics is the average value
957 of the electrostatic energy $\langle U_\mathrm{elect} \rangle / N$
958 which is obtained by sampling the liquid-state configurations
959 experienced by a liquid evolving entirely under the influence of each
960 of the methods. In Fig. \ref{fig:Props} we demonstrate how $\langle
961 U_\mathrm{elect} \rangle / N$ varies with the damping parameter,
962 $\alpha$, for each of the methods.
963
964 As in the crystals studied in the first paper, damping is important
965 for converging the mean electrostatic energy values, particularly for
966 the two shifted force methods (GSF and TSF). A value of $\alpha
967 \approx 0.2$~\AA$^{-1}$ is sufficient to converge the SP and GSF
968 energies with a cutoff of 12 \AA, while shorter cutoffs require more
969 dramatic damping ($\alpha \approx 0.28$~\AA$^{-1}$ for $r_c = 9$~\AA).
970 Overdamping the real-space electrostatic methods occurs with $\alpha >
971 0.3$~\AA$^{-1}$, causing the estimate of the electrostatic energy to
972 drop below the Ewald results.
973
974 These ``optimal'' values of the damping coefficient for structural
975 features are similar to those observed for DSF electrostatics for
976 purely point-charge systems, and the range $\alpha= 0.175 \rightarrow
977 0.225$~\AA$^{-1}$ for $r_c = 12$~\AA\ appears to be an excellent
978 compromise for mixed charge/multipolar systems.
979
980 To test the fidelity of the electrostatic methods at reproducing
981 \textit{dynamics} in a multipolar liquid, it is also useful to look at
982 transport properties, particularly the diffusion constant,
983 \begin{equation}
984 D = \lim_{t \rightarrow \infty} \frac{1}{6 t} \langle \left|
985 \mathbf{r}(t) -\mathbf{r}(0) \right|^2 \rangle
986 \label{eq:diff}
987 \end{equation}
988 which measures long-time behavior and is sensitive to the forces on
989 the multipoles. The self-diffusion constants (D) were calculated from
990 linear fits to the long-time portion of the mean square displacement,
991 $\langle r^{2}(t) \rangle$.\cite{Allen87} In Fig. \ref{fig:Props} we
992 demonstrate how the diffusion constant depends on the choice of
993 real-space methods and the damping coefficient. Both the SP and GSF
994 methods can obtain excellent agreement with Ewald again using moderate
995 damping.
996
997 In addition to translational diffusion, orientational relaxation times
998 were calculated for comparisons with the Ewald simulations and with
999 experiments. These values were determined by calculating the
1000 orientational time correlation function,
1001 \begin{equation}
1002 C_l^\gamma(t) = \left\langle P_l\left[\hat{\mathbf{A}}_\gamma(t)
1003 \cdot\hat{\mathbf{A}}_\gamma(0)\right]\right\rangle,
1004 \label{eq:OrientCorr}
1005 \end{equation}
1006 from the same 350 ps microcanonical trajectories that were used for
1007 translational diffusion. Here, $P_l$ is the Legendre polynomial of
1008 order $l$ and $\hat{\mathbf{A}}_\gamma$ is the unit vector for body
1009 axis $\gamma$. The reference frame used for our sample dipolar
1010 systems has the $z$-axis running along the dipoles, and for the soft
1011 DQ liquid model, the $y$-axis connects the two implied hydrogen-like
1012 positions. From the orientation autocorrelation functions, we can
1013 obtain time constants for rotational relaxation either by fitting to a
1014 multi-exponential model for the orientational relaxation, or by
1015 integrating the correlation functions.
1016
1017 In a good model for water, the orientational decay times would be
1018 comparable to water orientational relaxation times from nuclear
1019 magnetic resonance (NMR). The relaxation constant obtained from
1020 $C_2^y(t)$ is normally of experimental interest because it describes
1021 the relaxation of the principle axis connecting the hydrogen
1022 atoms. Thus, $C_2^y(t)$ can be compared to the intermolecular portion
1023 of the dipole-dipole relaxation from a proton NMR signal and can
1024 provide an estimate of the NMR relaxation time constant.\cite{Impey82}
1025 In Fig. \ref{fig:Props} we compare the $\tau_2^y$ and $\tau_2^z$
1026 values for the various real-space methods over a range of different
1027 damping coefficients. The rotational relaxation for the $z$ axis
1028 primarily probes the torques on the dipoles, while the relaxation for
1029 the $y$ axis is sensitive primarily to the quadrupolar torques.
1030
1031 \begin{figure}
1032 \includegraphics[width=\textwidth]{properties.eps}
1033 \caption{Comparison of the structural and dynamic properties for the
1034 combined multipolar liquid (soft DQ liquid + ions) for all of the
1035 real-space methods with $r_c = 12$~\AA. Electrostatic energies,
1036 $\langle U_\mathrm{elect} \rangle / N$ (in kcal / mol),
1037 coordination numbers, $n_C$, diffusion constants (in $10^{-5}
1038 \mathrm{cm}^2\mathrm{s}^{-1}$), and rotational correlation times
1039 (in ps) all show excellent agreement with Ewald results for
1040 damping coefficients in the range $\alpha= 0.175 \rightarrow
1041 0.225$~\AA$^{-1}$. \label{fig:Props}}
1042 \end{figure}
1043
1044 In Fig. \ref{fig:Props} it appears that values for $D$, $\tau_2^y$,
1045 and $\tau_2^z$ using the Ewald sum are reproduced with excellent
1046 fidelity by the GSF and SP methods. All of the real space methods can
1047 be \textit{overdamped}, which reduces the effective range of multipole
1048 interactions, causing structural and dynamical changes from the
1049 correct behavior. Because overdamping weakens orientational
1050 preferences between adjacent molecules, it manifests as too-rapid
1051 orientational decay coupled with faster diffusion and
1052 over-coordination of the liquid. Underdamping is less problematic for
1053 the SP and GSF methods, as their structural and dynamical properties
1054 still reproduce the Ewald results even in the completely undamped
1055 ($\alpha = 0$) case. An optimal range for the electrostatic damping
1056 parameter appears to be $\alpha= 0.175 \rightarrow 0.225$~\AA$^{-1}$
1057 for $r_c = 12$~\AA, which similar to the optimal range found for the
1058 damped shifted force potential for point charges.\cite{Fennell:2006lq}
1059
1060 \section{CONCLUSION}
1061 In the first paper in this series, we generalized the
1062 charge-neutralized electrostatic energy originally developed by Wolf
1063 \textit{et al.}\cite{Wolf:1999dn} to multipole-multipole interactions
1064 up to quadrupolar order. The SP method is essentially a
1065 multipole-capable version of the Wolf model. The SP method for
1066 multipoles provides excellent agreement with Ewald-derived energies,
1067 forces and torques, and is suitable for Monte Carlo simulations,
1068 although the forces and torques retain discontinuities at the cutoff
1069 distance that prevents its use in molecular dynamics.
1070
1071 We also developed two natural extensions of the damped shifted-force
1072 (DSF) model originally proposed by Zahn {\it et al.} and extended by
1073 Fennell and Gezelter.\cite{Zahn:2002hc,Fennell:2006lq} The GSF and TSF
1074 approaches provide smooth truncation of energies, forces, and torques
1075 at the real-space cutoff, and both converge to DSF electrostatics for
1076 point-charge interactions. The TSF model is based on a high-order
1077 truncated Taylor expansion which can be relatively perturbative inside
1078 the cutoff sphere. The GSF model takes the gradient from an images of
1079 the interacting multipole that has been projected onto the cutoff
1080 sphere to derive shifted force and torque expressions, and is a
1081 significantly more gentle approach.
1082
1083 The GSF method produces quantitative agreement with Ewald energies,
1084 forces, and torques. It also performs well in conserving energy in MD
1085 simulations. The Taylor-shifted (TSF) model provides smooth dynamics,
1086 but these take place on a potential energy surface that is
1087 significantly perturbed from Ewald-based electrostatics. Because it
1088 performs relatively poorly compared with GSF, it may seem odd that
1089 that the TSF model was included in this work. However, the functional
1090 forms derived for the SP and GSF methods depend on the separation of
1091 orientational contributions that were made visible by the Taylor
1092 series of the electrostatic kernel at the cutoff radius. The TSF
1093 method also has the unique property that a large number of derivatives
1094 can be made to vanish at the cutoff radius. This property has proven
1095 useful in past treatments of the corrections to the Clausius-Mossotti
1096 fluctuation formula for dielectric constants.\cite{Izvekov:2008wo}
1097
1098 Reproduction of both structural and dynamical features in the liquid
1099 systems is remarkably good for both the SP and GSF models. Pair
1100 distribution functions are essentially equivalent to the same
1101 functions produced using Ewald-based electrostatics, and with moderate
1102 damping, a structural feature that directly probes the electrostatic
1103 interaction (e.g. the mean electrostatic potential energy) can also be
1104 made quantitative. Dynamical features are sensitive probes of the
1105 forces and torques produced by these methods, and even though the
1106 smooth behavior of forces is produced by perturbing the overall
1107 potential, the diffusion constants and orientational correlation times
1108 are quite close to the Ewald-based results.
1109
1110 The only cases we have found where the new GSF and SP real-space
1111 methods can be problematic are those which retain a bulk dipole moment
1112 at large distances (e.g. the $Z_1$ dipolar lattice). In ferroelectric
1113 materials, uniform weighting of the orientational contributions can be
1114 important for converging the total energy. In these cases, the
1115 damping function which causes the non-uniform weighting can be
1116 replaced by the bare electrostatic kernel, and the energies return to
1117 the expected converged values.
1118
1119 Based on the results of this work, we can conclude that the GSF method
1120 is a suitable and efficient replacement for the Ewald sum for
1121 evaluating electrostatic interactions in modern MD simulations, and
1122 the SP method would be an excellent choice for Monte Carlo
1123 simulations where smooth forces and energy conservation are not
1124 important. Both the SP and GSF methods retain excellent fidelity to
1125 the Ewald energies, forces and torques. Additionally, the energy
1126 drift and fluctuations from the GSF electrostatics are significantly
1127 better than a multipolar Ewald sum for finite-sized reciprocal spaces,
1128 and physical properties are reproduced accurately.
1129
1130 As in all purely pairwise cutoff methods, the SP, GSF and TSF methods
1131 are expected to scale approximately {\it linearly} with system size,
1132 and are easily parallelizable. This should result in substantial
1133 reductions in the computational cost of performing large simulations.
1134 With the proper use of pre-computation and spline interpolation of the
1135 radial functions, the real-space methods are essentially the same cost
1136 as a simple real-space cutoff. They require no Fourier transforms or
1137 $k$-space sums, and guarantee the smooth handling of energies, forces,
1138 and torques as multipoles cross the real-space cutoff boundary.
1139
1140 We are not suggesting that there is any flaw with the Ewald sum; in
1141 fact, it is the standard by which the SP, GSF, and TSF methods have
1142 been judged in this work. However, these results provide evidence
1143 that in the typical simulations performed today, the Ewald summation
1144 may no longer be required to obtain the level of accuracy most
1145 researchers have come to expect.
1146
1147 \begin{acknowledgments}
1148 JDG acknowledges helpful discussions with Christopher
1149 Fennell. Support for this project was provided by the National
1150 Science Foundation under grant CHE-1362211. Computational time was
1151 provided by the Center for Research Computing (CRC) at the
1152 University of Notre Dame.
1153 \end{acknowledgments}
1154
1155 %\bibliographystyle{aip}
1156 \newpage
1157 \bibliography{references}
1158 \end{document}
1159
1160 %
1161 % ****** End of file aipsamp.tex ******