ViewVC Help
View File | Revision Log | Show Annotations | View Changeset | Root Listing
root/group/trunk/openmdDocs/openmdDoc.tex
(Generate patch)

Comparing trunk/openmdDocs/openmdDoc.tex (file contents):
Revision 3792 by gezelter, Thu Aug 9 18:48:23 2012 UTC vs.
Revision 3793 by skuang, Mon Aug 20 21:03:35 2012 UTC

# Line 17 | Line 17
17   \textwidth 6.5in
18   \brokenpenalty=10000
19   \renewcommand{\baselinestretch}{1.2}
20 + \usepackage[square, comma, sort&compress]{natbib}
21 + \bibpunct{[}{]}{,}{n}{}{;}
22  
23 +
24   %\renewcommand\citemid{\ } % no comma in optional reference note
25   \lstset{language=C,frame=TB,basicstyle=\tiny,basicstyle=\ttfamily, %
26          xleftmargin=0.25in, xrightmargin=0.25in,captionpos=b, %
# Line 44 | Line 47
47   \newcolumntype{M}{p{1.55in}}
48  
49  
50 < \title{{\sc OpenMD}: Molecular Dynamics in the Open}
50 > \title{{\sc OpenMD-2}: Molecular Dynamics in the Open}
51  
52 < \author{Shenyu Kuang, Charles F. Vardeman II, \\
53 <  Teng Lin, Christopher J. Fennell,  Xiuquan Sun, \\
52 > \author{Shenyu Kuang, Joseph Michalka, Kelsey Stocker, James Marr, \\
53 >  Teng Lin, Charles F. Vardeman II, Christopher J. Fennell, Xiuquan Sun, \\
54    Chunlei Li, Kyle Daily, Yang Zheng, Matthew A. Meineke, and \\
55    J. Daniel Gezelter \\
56    Department of Chemistry and Biochemistry\\
# Line 496 | Line 499 | fs}^{-1}$), and body-fixed moments of inertia ($\mbox{
499   \endhead
500   \hline
501   \endfoot
502 < {\tt forceField} & string & Sets the force field. & Possible force
503 < fields are DUFF, WATER, LJ, EAM, SC, and CLAY. \\
502 > {\tt forceField} & string & Sets the base name for the force field file &
503 > OpenMD appends a {\tt .frc} to the end of this to look for a force
504 > field file.\\
505   {\tt component} & & Defines the molecular components of the system &
506   Every {\tt $<$MetaData$>$} block must have a component statement. \\
507   {\tt minimizer} & string & Chooses a minimizer & Possible minimizers
# Line 2939 | Line 2943 | Harmonic Forces are used by default
2943   \end{longtable}
2944  
2945   % \chapter{\label{section:restraints}Restraints}
2946 < % Restraints are external potentials that are added to a system to keep
2947 < % particular molecules or collections of particles close to some
2946 > % Restraints are external potentials that are added to a system to
2947 > % keep particular molecules or collections of particles close to some
2948   % reference structure.  A restraint can be a collective
2949  
2950   \chapter{\label{section:thermInt}Thermodynamic Integration}
# Line 3084 | Line 3088 | There are many ways to compute transport properties fr
3088   \chapter{\label{section:rnemd}RNEMD}
3089  
3090   There are many ways to compute transport properties from molecular
3091 < dynamic simulations.  Equilibrium Molecular Dynamics (EMD) simulations
3092 < can be used by computing relevant time correlation functions and
3093 < assuming linear response theory holds.  These approaches are generally
3094 < subject to noise and poor convergence of the relevant correlation
3095 < functions. Traditional Non-equilibrium Molecular Dynamics (NEMD)
3096 < methods impose a gradient (e.g. thermal or momentum) on a simulation.
3097 < However, the resulting flux is often difficult to
3091 > dynamics simulations.  Equilibrium Molecular Dynamics (EMD)
3092 > simulations can be used by computing relevant time correlation
3093 > functions and assuming linear response theory holds.  These approaches
3094 > are generally subject to noise and poor convergence of the relevant
3095 > correlation functions. Traditional Non-equilibrium Molecular Dynamics
3096 > (NEMD) methods impose a gradient (e.g. thermal or momentum) on a
3097 > simulation.  However, the resulting flux is often difficult to
3098   measure. Furthermore, problems arise for NEMD simulations of
3099   heterogeneous systems, such as phase-phase boundaries or interfaces,
3100   where the type of gradient to enforce at the boundary between
3101   materials is unclear.
3102  
3103 < {\it Reverse} Non-Equilibrium Molecular Dynamics (RNEMD) methods adopt a
3104 < different approach in that an unphysical {\it flux} is imposed between
3105 < different regions or ``slabs'' of the simulation box.  The response of
3106 < the system is to develop a temperature or momentum {\it gradient}
3107 < between the two regions. Since the amount of the applied flux is known
3108 < exactly, and the measurement of gradient is generally less
3109 < complicated, imposed-flux methods typically take shorter simulation
3110 < times to obtain converged results for transport properties.
3103 > {\it Reverse} Non-Equilibrium Molecular Dynamics (RNEMD) methods adopt
3104 > a different approach in that an unphysical {\it flux} is imposed
3105 > between different regions or ``slabs'' of the simulation box.  The
3106 > response of the system is to develop a temperature or momentum {\it
3107 >  gradient} between the two regions. Since the amount of the applied
3108 > flux is known exactly, and the measurement of gradient is generally
3109 > less complicated, imposed-flux methods typically take shorter
3110 > simulation times to obtain converged results for transport properties.
3111  
3112 < %RNEMD figure
3112 > \begin{figure}
3113 > \includegraphics[width=\linewidth]{rnemdDemo}
3114 > \caption{The (VSS) RNEMD approach imposes unphysical transfer of both
3115 >  linear momentum and kinetic energy between a ``hot'' slab and a
3116 >  ``cold'' slab in the simulation box.  The system responds to this
3117 >  imposed flux by generating both momentum and temperature gradients.
3118 >  The slope of the gradients can then be used to compute transport
3119 >  properties (e.g. shear viscosity and thermal conductivity).}
3120 > \label{rnemdDemo}
3121 > \end{figure}
3122 >
3123 > The original ``swapping'' approaches by M\"{u}ller-Plathe {\it et
3124 >  al.}\cite{ISI:000080382700030,MullerPlathe:1997xw} can be understood
3125 > as a sequence of imaginary elastic collisions between particles in
3126 > opposite slabs.  In each collision, the entire momentum vectors of
3127 > both particles may be exchanged to generate a thermal
3128 > flux. Alternatively, a single component of the momentum vectors may be
3129 > exchanged to generate a shear flux.  This algorithm turns out to be
3130 > quite useful in many simulations. However, the M\"{u}ller-Plathe
3131 > swapping approach perturbs the system away from ideal
3132 > Maxwell-Boltzmann distributions, and this may leads to undesirable
3133 > side-effects when the applied flux becomes large.\cite{Maginn:2010}
3134 > This limits the application of the swapping algorithm, so in OpenMD,
3135 > we implement two additional algorithms for RNEMD in addition to the
3136 > original swapping approach.
3137  
3138 + {\bf Non-Isotropic Velocity Scaling (NIVS):}\cite{kuang:164101}
3139 + Instead of having momentum exchange between {\it individual particles}
3140 + in each slab, the NIVS algorithm applies velocity scaling to all of
3141 + the selected particles in both slabs.  A combination of linear
3142 + momentum, kinetic energy, and flux constraint equations governs the
3143 + amount of velocity scaling performed at each step.  Interested readers
3144 + should consult ref. \citealp{kuang:164101} for further details on the
3145 + methodology.
3146  
3147 < RNEMD methods further its advantages by utilizing momentum- and
3148 < energy-conserving approaches to apply fluxes. The original
3149 < ``swapping'' approach by Muller-Plathe {\it et al.} %CITATIONS
3150 < can be seen as an imaginary elastic collision between selected
3151 < particles for each momentum exchange. This simple to implement
3152 < algorithm turned out to be quite useful in many simulations. However,
3117 < the approach inherently perturbs the ideal Maxwell-Boltzmann
3118 < distributions, which leads to undesirable side-effects when the
3119 < applied exchanged flux becomes quite large. %CITATION
3120 < This limits the range of flux available to the method, and also its
3121 < applications.
3147 > NIVS has been shown to be very effective at producing thermal
3148 > gradients and for computing thermal conductivities, particularly for
3149 > heterogeneous interfaces.  Although the NIVS algorithm can also be
3150 > applied to impose a directional momentum flux, thermal anisotropy was
3151 > observed in relatively high flux simulations, and the method is not
3152 > suitable for imposing a shear flux.
3153  
3154 < In OpenMD, we improve the above method by introducing two alternative
3155 < approaches:
3154 > {\bf Velocity Shearing and Scaling (VSS)}:\cite{2012MolPh.110..691K}
3155 > The third RNEMD algorithm implemented in OpenMD utilizes a series of
3156 > simultaneous velocity shearing and scaling exchanges between the two
3157 > slabs.  This method results in a set of simpler equations to satisfy
3158 > the conservation constraints while creating a desired flux between the
3159 > two slabs.
3160  
3161 < Non-Isotropic Velocity Scaling (NIVS): %CITATION
3162 < Instead of have two individual particles involved in momentum
3163 < exchange, this algorithm applies scaling to all the particles in
3164 < particular regions:
3161 > The VSS approach is versatile in that it may be used to implement both
3162 > thermal and shear transport either separately or simultaneously.
3163 > Perturbations of velocities away from the ideal Maxwell-Boltzmann
3164 > distributions are minimal, and thermal anisotropy is kept to a
3165 > minimum.  This ability to generate simultaneous thermal and shear
3166 > fluxes has been utilized to map out the shear viscosity of SPC/E water
3167 > in a wide range of temperature (90~K) just with a single simulation.
3168 > VSS-RNEMD also allows the directional momentum flux to have quite
3169 > arbitary directions, which could aid in the study of anisotropic solid
3170 > surfaces in contact with liquid environments.
3171  
3172 < %NIVS equations
3173 <
3174 < Although the above matrices can be diagonal as shown, these
3175 < coefficients cannot be always the same, in order to satisfy the linear
3176 < momentum and kinetic energy conservation constraints:
3172 > {\bf What the user needs to specify:} To carry out a RNEMD simulation,
3173 > a user must specify a number of parameters in the MetaData (.md) file.
3174 > Because the RNEMD methods have a large number of parameters, these
3175 > must be enclosed in a {\tt RNEMD\{...\}} block.  The most important
3176 > parameters to specify are the {\tt useRNEMD}, {\tt fluxType} and flux
3177 > parameters. Most other parameters (summarized in table
3178 > \ref{table:rnemd}) have reasonable default values.  {\tt fluxType}
3179 > sets up the kind of exchange that will be carried out between the two
3180 > slabs (either Kinetic Energy ({\tt KE}) or momentum ({\tt Px, Py, Pz,
3181 >  Pvector}), or some combination of these).  The flux is specified
3182 > with the use of three possible parameters: {\tt kineticFlux} for
3183 > kinetic energy exchange, as well as {\tt momentumFlux} or {\tt
3184 >  momentumFluxVector} for simulations with directional exchange.
3185  
3186 < %Conservation equations
3186 > {\bf How to process the results:} OpenMD will generate a {\tt .rnemd}
3187 > file with the same prefix as the original {\tt .md} file.  This file
3188 > contains a running average of properties of interest computed within a
3189 > set of bins that divide the simulation cell along the $z$-axis.  The
3190 > first column of the {\tt .rnemd} file is the $z$ coordinate of the
3191 > center of each bin, while following columns may contain the average
3192 > temperature, velocity, or density within each bin.  The output format
3193 > in the {\tt .rnemd} file can be altered with the {\tt outputFields},
3194 > {\tt outputBins}, and {\tt outputFileName} parameters.  A report at
3195 > the top of the {\tt .rnemd} file contains the current exchange totals
3196 > as well as the average flux applied during the simulation.  Using the
3197 > slope of the temperature or velocity gradient obtaine from the {\tt
3198 >  .rnemd} file along with the applied flux, the user can very simply
3199 > arrive at estimates of thermal conductivities ($\lambda$),
3200 > \begin{equation}
3201 > J_z = -\lambda \frac{\partial T}{\partial z},
3202 > \end{equation}
3203 > and shear viscosities ($\eta$),
3204 > \begin{equation}
3205 > j_z(p_x) = -\eta \frac{\partial \langle v_x \rangle}{\partial z}.
3206 > \end{equation}
3207 > Here, the quantities on the left hand side are the actual flux values
3208 > (in the header of the {\tt .rnemd} file), while the slopes are
3209 > obtained from linear fits to the gradients observed in the {\tt
3210 >  .rnemd} file.
3211  
3212 < And to apply a kinetic energy exchange between the two regions, the
3213 < following should be satisfied as well:
3214 <
3215 < %Flux equations
3216 <
3217 < Mathematically, any points in the 3-dimensional space of the solution
3218 < set would satisfy the equations. However, to avoid solving an
3219 < ill-conditioned high-order polynomial in actual practice, another
3220 < constraint, ${x_c=y_c}$, is applied, taking into consideration of its
3221 < physical relevance. Therefore, a quartic equation is solved in actual
3222 < practice to determine the sets of possible coefficients. To determine
3223 < which set is actually used to perform the scaling, two criteria are
3224 < mainly considered: 1. ${x,y,z\rightarrow 1}$ so that the perturbation
3225 < could be as gentle as possible. 2. ${K^x, K^y, K^z}$ have minimal
3226 < difference among each other, so that the anisotropy introduced by the
3227 < algorithm can be offset to some extend. One set of scaling
3228 < coefficients is chosen against these criteria, and the best one is
3229 < used to perform the scaling for that particular step. However, if no
3157 < solution found, the NIVS move is not performed in that step.
3158 <
3159 < Although the NIVS algorithm can also be applied to impose a
3160 < directional momentum flux, thermal anisotropy was observed in
3161 < relatively high flux simulations. %This is because...
3162 < However, the gentleness and ability to apply a wide range of kinetic
3163 < energy flux makes the method useful in thermal transport simulations,
3164 < particularly for complex and heterogeneous systems including
3165 < interfaces. %CITATION
3166 <
3167 < Velocity Shearing and Scaling (VSS): %CITATION
3168 < Learning from NIVS that imposing directional momentum flux by velocity
3169 < scaling could cause problem, we shift the approach to combine the move
3170 < of velocity shearing and scaling:
3171 <
3172 < %VSS equations
3173 <
3174 < It turned out that this approach results in a set of simpler-to-solve
3175 < equations for conservation and to satisfy momentum exchange:
3176 <
3177 < %conservation equations
3178 <
3179 < Furthermore, isotropic scaling is now possible, with the presence of
3180 < velocity shearing quantities. Only a set of simple quadratic equations
3181 < need to be solved, and the positive set of coefficients are chosen, in
3182 < order to reach minimal perturbations. Similar to the NIVS method, no
3183 < VSS is performed in a step given that no solution can be found.
3184 <
3185 < The VSS approach turned out to be versatile in both thermal and
3186 < directional momentum transport simulations. It is found that the
3187 < perturbation is minimal and undesired side-effects like thermal
3188 < anisotropy can be avoided. Another nice feature of VSS is its ability
3189 < to combine a thermal and a directional momentum flux. This feature has
3190 < been utilized to map out the shear viscosity of SPC/E water in a wide
3191 < range of temperature (90~K) just with one single simulation. Possible
3192 < applications may also include the studies of thermal-momentum coupled
3193 < transport phenomena. VSS also allows the directional momentum flux to
3194 < have quite arbitary directions, which could benefit researches of
3195 < anisotropic systems.
3212 > More complicated simulations (including interfaces) require a bit more
3213 > care.  Here the second derivative may be required to compute the
3214 > interfacial thermal conductance,
3215 > \begin{align}
3216 >  G^\prime &= \left(\nabla\lambda \cdot \mathbf{\hat{n}}\right)_{z_0} \\
3217 >  &= \frac{\partial}{\partial z}\left(-\frac{J_z}{
3218 >      \left(\frac{\partial T}{\partial z}\right)}\right)_{z_0} \\
3219 >  &= J_z\left(\frac{\partial^2 T}{\partial z^2}\right)_{z_0} \Big/
3220 >  \left(\frac{\partial T}{\partial z}\right)_{z_0}^2.
3221 >  \label{derivativeG}
3222 > \end{align}
3223 > where $z_0$ is the location of the interface between two materials and
3224 > $\mathbf{\hat{n}}$ is a unit vector normal to the interface.  We
3225 > suggest that users interested in interfacial conductance consult
3226 > reference \citealp{kuang:AuThl} for other approaches to computing $G$.
3227 > Users interested in {\it friction coefficients} at heterogeneous
3228 > interfaces may also find reference \citealp{2012MolPh.110..691K}
3229 > useful.
3230  
3231 < Table \ref{table:rnemd} summarizes the parameters used in RNEMD
3198 < simulations.
3231 > \newpage
3232  
3233   \begin{longtable}[c]{JKLM}
3234   \caption{The following keywords must be enclosed inside a {\tt RNEMD\{\}} block}
# Line 3225 | Line 3258 | default is 20 \\
3258   {\tt outputBins} & int & number of $z$-bins in the output histogram &
3259   default is 20 \\
3260   {\tt outputFields} & string & columns to print in the {\tt .rnemd}
3261 < file where each column is separated by a pipe ($\mid$) symbol. & Allowed column names are: {\sc z, temperature, velocity, density}} \\
3261 > file where each column is separated by a pipe ($\mid$) symbol. & Allowed column names are: {\sc z, temperature, velocity, density} \\
3262   \label{table:rnemd}
3263   \end{longtable}
3264  
# Line 3903 | Line 3936 | DMR-0079647.
3936   DMR-0079647.
3937  
3938  
3939 < \bibliographystyle{jcc}
3939 > \bibliographystyle{aip}
3940   \bibliography{openmdDoc}
3941  
3942   \end{document}

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines