A guide to the aKMC script developed in Henkelman group

Lijun Xu and Graeme Henkelman

I.

OBTAINING AKMC SCRIPTS

Journal of Chemical Physics 129, 114104 (2008) [details about algorithms used in this package]

The akmc script is part of the vtstscripts package written by Henkelman group in the University of Texas at Austin, designed for using Vienna b inia tio simulation package (VASP). The whole package is written in Perl language and is expected to be working on various platforms. In WINDOWS operation system, ActivePerl program is a possible choice to run the scripts. The whole package can be downloaded via anonymous CVS system (check the guide on http://theory.cm.utexas.edu/henkelman/code/). You HAVE TO download the whole vtstscripts package in order to run adaptive KMC simulation on VASP, simply because scripts are actually calling each other. By the way, your VASP must be compiled with VTSTTOOLS developed by Henkelman group. The standard VASP package you purchased doesnt work for this aKMC script. The webpage about AKMC is http://theory.cm.utexas.edu/vtsttools/akmc/.
II. WHAT DOES AKMC SCRIPT DO?

IV.

OPTIONAL READINGS

Graeme Henkelman and Hannes Jnsson, J. Chem. o Phys. 111, 7010 (1999) [The dimer method] Lijun Xu, Donghai Mei and Graeme Henkelman, Studying chemical reactions with DFT-based adaptive kinetic Monte Carlo simulations - methanol full decomposition pathways on Cu(100) surface, in preparation [An example of using AKMC to study chemical reactions]

V.

EXAMPLES

After reading this manual, readers are strongly recommended to run the example given on our group webpage (http://theory.cm.utexas.edu/vtsttools/akmc/) to see how the scripts and simulations work out.

VI.

GET STARTED

Given a starting system conguration, the akmc script handles the whole adaptive KMC simulation using energy and forces calculated from VASP. It rst minimizes the conguration it receives, then it will search for saddle points for the minimized conguration and quench each found saddle point into two minima directly connected with it. It can also calculates normal mode frequencies for the initial minimum and its saddles. After nishing saddle point searches, it will calculate the rate constants and prefactors with harmonic transition state theory and get a table of rates for each reaction mechanism. Using that table, it will do a KMC move to select one mechanism and update the system conguration to the end point connected with the minimum energy pathway (or mechanism). The whole process is repeated for the new system conguration. The whole operations are handled by just typing the command akmc.pl.
III. PREREQUISITE READINGS

A.

Files used by aKMC script

Create a home directory for aKMC calculations. Put the following les in the aKMC home (or designated) directory, set all the parameters and start the run by typing akmc.pl CONFIG It has all the control parameters running aKMC simulation. Details will be given in the following section. POSCAR This is the initial conguration given to the program. It will be updated with new structures as simulation goes. The le format must be consistent with VASP POSCAR requirements. KPOINTS Control the k-points setup for VASP calcualtion. POTCAR Pseudopoential for VASP calculations. INCAR sp It instructs how to run a saddle point search in VASP. It is in the same format as VASP INCAR les. INCAR min It instructs how to run a geometry minimization in VASP. It is in the same format as VASP INCAR les.

Graeme Henkelman and Hannes Jnsson, Long time o scale kinetic Monte Carlo simulations without lattice approximation and predened event table, J. Chem. Phys. 115, 9657-9666 (2001) [the original idea of akmc] Lijun Xu and Graeme Henkelman, Adaptive kinetic Monte Carlo for rst-principles accelerated dynamics,

2 INCAR dynmat It instructs how to calculate a dynamic matriix in VASP. It is in the same format as VASP INCAR les. DISPLACECAR This is the VASP DISPLACECAR le for dynmaic matrix calculation, showing how much each degree of freedom is nitely displaced. DISPLACECAR sp This le is in the same format as DISPLACECAR, but each value in the le is the Gaussian width with which a random displacement is to be created for that degree of freedom. It controls how the random saddle point searches are prepared. akmc submit.pl This is your own script that submits a job from the directory the job is supposed to run in and prints a job id as the only screen output. It has to be placed in the same folder as akmc.pl and other VTSTSCRIPTS. An example of this scripts is provided on the webpage. akmc check.pl This is your own script that takes a job id and checks the status. The only screen output should be running, or, stopped, or queue. It has to be placed in the same folder as akmc.pl and other VTSTSCRIPTS. An example of this scripts is provided on the webpage. akmc kill.pl This is your own script that kills a job. It has to be placed in the same folder as akmc.pl and other VTSTSCRIPTS. An example of this scripts is provided on the webpage. akmc sp.sub This is the submission script used by akmc submit.pl for submitting saddle point searching jobs to the queue system. akmc min.sub This is the submission script used by akmc submit.pl for minimization jobs to the queue system. akmc dynmat.sub This is the submission script used by akmc submit.pl for dynamic matrix calculation jobs to the queue system. akmc.dat This le keeps the information of each aKMC step: showing the conguration of the step, barrier, accumulated time, etc. It is also a key le for the script to gure out the stop point from the previous running of akmc.pl. st0001 After the initial minimization, this state folder is created to contain all the saddle point searching processes for the rst state. There are subfolders and les inside this directory. We take the rst one st0001 as the example to show the tree structure of the state folders. st0001/POSCAR, etc. This is the POSCAR copied from the AKMC home directory as the conguration of the state st0001. Other types of les in home directory such as POTCAR, INCAR min,INCAR sp, DISPLACECAR sp and KPOINTS will also be copied into this state folder. st0001/neighborlist.dat This neighbor list le, if created, is for random displacement. Details are in the reference. st0001/st.dat This le is the key le for a state: it records how many saddle point searches are running in the sr0001, and the evaluation of each saddle searching process: good, bad or something else. Together with jobs.dat, it does the main control of the aKMC job running. Any manual handle should target towards jobs.dat and st.dat. st0001/pr0001 This folder is the rst saddle point search for state st0001. It has subfolders and les. st0001/pr0002 This folder is the second saddle point search for state st0001. It has subfolders and les. st0001/pr0001/POSCAR POSCAR for a saddle point searching job in VASP. It is created by randomly displacing the state POSCAR. st0001/pr0001/INCAR INCAR for a saddle point seaching job in VASP. Note: all three INCAR * les are copied inside folder st0001 and pr0001. INCAR sp will be copied into INCAR automatically when the job is being submitted. akmc sp.sub will be copied into akmc.sub which is the standard submission le in the akmc submit.pl script. st0001/pr0001/POTCAR Same POTCAR le as in the home directory. st0001/pr0001/KPOINTS Same KPOINTS le as in the home directory. st0001/pr0001/inter A unnished but stopped job are cleaned up into this inter, it will be renamed to inter1, inter2, as more come in, but this inter is the most recent one. It is same as what vn.pl does in our normal VASP job cleanup.

B.

Files and folders created by aKMC script

After running aKMC, there will be new les and folders to be created automatically. In the following list, you will see them in the order of being created by the script. jobs.dat This le contains all the job information: work dirctory, job id, job type, energy, force, curavture. min This folder is created to run the very beginning minimization. The minimized structure will be used to start the search of saddle points. A minimized structure represents a state of the system.

3 st0001/pr0001/nal A converged (saddle searching) job is cleaned up into this folder. It is same as what vn.pl does in our normal VASP job cleanup. DO NOT rename this folder. Both inter and nal are required standard folder names in this package. st0001/pr0001/POSCAR sp It is the saddle point conguration le. The CONTCAR (or CENTCAR) of a converged job will be copied into this le. DO NOT rename this le. st0001/pr0001/dynmat If dynamic matrix calculation is required, this folder will be created to do the dynmat calculation. POSCAR sp will be copied into dynmat/POSCAR, INCAR dynmat wll be copied into dynmat/INCAR, etc. st0001/pr0001/dynmat/nal A converged dynmat job is cleaned up into this folder. It is same as what vn.pl does in our normal VASP job cleanup. st0001/pr0001/mins This folder contains jobs of quenching the converged saddle into two adjacent minima. st0001/pr0001/mins/min1 This folder contains the minimization of one image along the MEP. A normal VASP minimization job. st0001/pr0001/mins/min2 This folder contains the minimization of the other image along the MEP. A normal VASP minimization job. st0001/pr0001/mins/min1/nal A folder for converged minimization job. Same as before. st0001/pr0001/mins/min2/nal A folder for converged minimization job. Same as before. st0001/jobs.dat Once st0001 is done, the jobs.dat in the aKMC home directory will be copied into the state folder st0001. st0002 If st0001 is nished, KMC will make a move and update the system into a new state. Then st0002 will be created to start the saddle point searches for the new state. AaKMC continues like this until it reaches the maximum number of steps indicated in CONFIG le. StEnergyFile This is to store the energy of each state the AKMC simulation has visited. It is for the sake of avoiding reading the st.dat when the code goes back to check if the new state has appeared somewhere before. Any state in this le must be a nished state.
A. CONFIG

Here is an sample CONFIG le. MaxJobs = 12 RandSeed = 5 NumSearches = 10 AkmcSteps = 100000 SimR = 0.1 NumKT = 20 Temperature = 300.0 Edimax = 0.05 Rdimax = 0.04 SearchesAlgo = dimer BarrierMAX = 10.0 JobFile = jobs.dat StFile = st.dat AkmcFile = akmc.dat DynMatFile = freq.dat Prefactor = 1.0e12 RateTableFile = RateTableFile.dat OldRunDirFile = OldRunDirFile.dat StEnergyFile = StEnergyFile.dat GrabQuenched = 1 Population = 0.2 PrRecycleAll = 1 PrRecycle = 1 PrRecycleShift = 1 SurfaceRecShift = 1 ConvergenceAlgo = 1 equivalency = 1 DisplaceAlgo = 1 DisplaceRange = 3 NN rcut = 2.6 MaxCoordNum = 8 UseKDB = 0 File cong is in the aKMC home directory and contains the following ags. MaxJobs The maximum number of jobs can be submitted through aKMC. This number must be smaller than the maximum allowed by the queue system. For example, if the queue allows you to have at most 8 jobs running and waiting in the queue and you use three jobs to do things other than AKMC, you will have to set MaxJobs to 8-3=5. You have to tell the scripts how many jobs are allowed. There is no way for the script to gure out the number by itself. NumSearches The number of searches for convergency standard. If there are no new good saddle points found during that many straight successful searches, the searching will stop and an event table will be compiled for KMC. Default: 10. Details are in the reference. AkmcSteps The total number of KMC steps. Use 1 for mechanism study. Default: 1

VII.

FOUR KEY FILES: CONFIG, ST.DAT, JOBS.DAT AND AKMC.DAT

Those four les determine the main stream of the aKMC script.

4 SimR The distance (in ) between two images centered A around the saddle point along the minimum energy path. Those two images are expected to roll into two end points of the MEP upon minimization. Default: 0.1 Temperature The temperature in kelvin. Default: 77 NumKT The energy gap above the lowest saddle point. Saddle points above the highest point are considered unimportant to the nal rate table. The unit is kT and T is for temperature. Normally, this number is at least 10, which means 10 times thermal energy at the temperature. Default: 20 EdiMax The maximum energy dierence (in eV) between two congurations, below which geometry dierence will be calculated. Default: 0.05 RdiMax The maximum dierence (in ) between any A two corresponding atoms from two congurations, below which two geometries will be considered as identical. In aKMC, two POSACR les are frequently compared to determine if they are for the same geometry. Default: 0.05 SearchAlgo The saddle point searching mechanisms. Right now, it can only be dimer. Lanzcos may be added to the script in the future. Default: dimer. BarrierMax The maximum energy allowed before killing a search for the searching point energy being too high. Default: 10 JobFile The lename for jobs.dat in the aKMC home directory. Default: jobs.dat. StFile The le name for st.dat in the state folder (such as st0001). Default: st.dat AkmcFile The le name for akmc.dat. akmc.dat Default: GrabQuenched If it is set to be 1, the code will check if the newly converged saddle has been quenched before (not limited to this state, including all the previous states and database). Default: 0 PrRecycleAll If it is set to be 1, the code will check if the new state is one of the end points in those mechanisms have been found in previous states and database. If yes, the code copies those processes and recover those saddles for the new state. Default: 1. PrRecycle If it is set to be 1, any good saddle from the previous ( or equivalent) state will be recycled and reconverged for the new state. Default: 0. PrRecycleShift If it is set to be 1, a vector-shifting recycle will be used. Default: 0. SurfaceRecShift If it is set to be 1, the system will be a surface science problem: adsorbates on substrate. In POSCAR, adsorbates will be listed as the last group. Any good saddle from the previous state will be reconverged by adjusting the adsorbate orientation. This is not a mature technique. We do not recommend it unless you are familiar with the code itself. Default: 0. ConvergenceAlgo The key ag controlling the ending of the saddle point searches: 0 for a xed number of saddle seacrhes; 1 for a dynamic standard (coupled with NumKT ag) Equivalency If it is set to be 1, the system will test if the new sate conguration is actually equivalent to some previous states by checking the the one-to-one matching between atoms in two POSCAR les ( or the radial distribution function of adsorbates if SurfaceRecShift is set to 1.). Saddle points from any equivalent state will be reshued based on the matching and used directly. Default: 0. Screenoutput the le for keeping the screen output from running akmc.pl. Default: out.dat. DisplaceAlgo The displacing algorithm used in creating random dimer congurations: 0 for selecting one of the most under-coordinated atoms as the displacement center; 1 for one of the choosing the less (= MaxCoordNum) coordinated atoms as the displacement center; 2 for dividing the surface into islands and doing the 0 option in islands; 3 for dividing the surface atoms into islands and doing the 1 option in islands; any other values will lead to a random atom to be selected as the displacement center. Note: only those atoms (or degrees of freedom) marked as non-zero in the DISPLACECAR sp can be the candidate displacement center in any of those algorithms. Default: 1. Details are in the reference. It is important to be familiar with these algorithms. DisplaceRange The radius around the displacement center within which all of those atoms will be displaced by a random vector. Default: 3.

DynMatFile The le containing normal mode frequencies. Keep the default name unless you modify the dymmatrix.pl script. Default: freq.dat. Prefactor The dynamic matrix calculation will be disabled if this ag is turned on. The positive value will be used for every process in the coming calculations. You have to keep this ag consistent throughout your whole aKMC calculations. Default: 0 (disabled) RateTableFile The le for storing the rate table. Default: RateTableFile.dat OldRunDirFile The le listing all the database folders. Default: OldRunDirFile StEnergyFile The energy of each states conguration is listed in here. Default: StEnergyFile. Population An obsolete convergence standard. It will be only used for ConvergenceAlgo=2 which we dont recommend. Default: 0

5 NN rcut The maximum pair distance for two nearest neighbors. Default: 2.6 MaxCoordNum The maximum coordination number used in DisplaceAlgo=1. Default: 8. BoltzmanEqu The threshold for two states being in Boltzman equilibrium: if one MEP shared by two states are dominating at least that amount of percentage in each rate table. Default: 0.999999. UseKDB This is about using database to prepare saddle point searches for the new state. If it is on, random searches will be done after using the initial guesses from databases. Default: 0 KDBcuto This is a standard used in UseKDB to determine how similar two systems should be in order to exchange saddle point information with each other. Default: 0.1 Fourth line NumSearchesLeft 10 shows how many more good saddle searches without seeing a new saddle are left. The state will be taken as done if it is set to zero. Fifth line numprst 12 shows that there are 12 process folders for 12 saddle point searching jobs. Sixth line It is an optional description line showing the meaning of each column in the following parts of the le: process status quality barrier nal/(min1(ev)) nal/min2(ev) Next line: column one It is the folder name for the process: e.g., pr0001, pr0002. etc. Next line: column two It is the status for the process: done means the searching (and dynmat if needed) and quenching are all nished. Otherwise, the status would be search, or quench, or dynmat for each type of operations. If the search is killed for some reason, killed will appear in this column and a reason follows it. Next line: column three It is the quality of the saddle point: good means one and only one of the end points of the minimum energy path is same as the conguration of the state (i.e., same as the POSCAR in folder st000x); bad means neither of the end points is same as the state conguration; promising means the barrier of the saddle (relative to the initial state) is not larger than BarrierMax in CONFIG le and the saddle is being quenched to determine if it is good or not; bad can also mean other situations including negative narrier, barrier being too high, etc; repeat*pr0003 means this saddle is same as the saddle point in pr0003 and pr0003 has been nished (or at least the quenching jobs have been nished); pending means the saddle is not converged yet. Next line: column four It is the barrier of the saddle point in eV. Next line: column ve If the saddle is good, this column will record the address of the nal state of the minimum energy path (the initial state will be same as the state conguration), e.g., mins/min1 means that the nal state is in st0001/pr0001/mins/min1/POSCAR and implies that the initial state is st0001/pr0001/mins/min2/POSCAR which is same as st0001/POSCAR. If the saddle is bad, this column will be simply the energy of the state in st0001/mins/min1/ in eV. Otherwise, an na will be put in there for not available. Next line: column six If the saddle is good, this column will record energy of the state in column 5. If the saddle is bad, this column will be simply the energy of the state in st0001/mins/min2/ in eV. Otherwise, an na will be put in there for not available.

B.

ST.DAT

Here is an sample st.dat le. st0001 -376.931759 status running dynmat done NumSearchesLeft 10 numprst 12 process status quality barrier nal/(min1) nal/(min2) pr0001 done good 1.997 mins/min1 -375.549 pr0002 done good 0.595 mins/min1 -376.521 pr0003 done good 2.463 mins/min1 -375.566 pr0004 done bad 1.809 -375.988 -375.339 pr0005 killed saddleenergytoohigh search na na na pr0006 done bad 3.126 -375.409 -376.264 pr0007 killed positive curvature search na na na pr0008 done bad 0.795 -376.433 -376.931 pr0009 done repeat*pr0008 0.795 na na pr0010 done repeat*pr0002 0.595 na na pr0011 quench promising 2.531 na na pr0012 search pending na na na File st.dat is in the state folder (such as st0001) and contains the following information (Lets take st0001/st.dat as the example). First line st0001 -376.931759 for the relative address and energy of the state. Second line status running for indicating the state is not done yet. status done is for a done state. Third line dynmat running for indicating the dynamic matrix calculation for the state conguration has not nished yet. dynmat done means it is either nished or tuned o by setting the Prefactor ag in CONFIG le.

6
C. JOBS.DAT

Here is an sample jobs.dat le. location Status JobId JobType Energy Force Curv. st0008/pr0001 completed 1 dimer -374.9 0.01 -1.34 st0008/pr0001/mins/min1 queue 2 minimization - - st0008/pr0001/mins/min1 queue 3 minimization - - File jobs.dat is in the aKMC home directory and contains the following information. First line It is a mandatory description line showing the meaning of each column in the following parts of the le. joblocation JobStatus JobId JobType Energy(eV) Force(eV/A) Curvature are quite self-explanatory. column one This is the relative address of the job location. column two This is the job status: submitted, queue, running, stopped, completed or 2bsubmitted. Among them, completed means the job has really converged based on the INCAR settings. 2bsubmitted means somehow the job was not successfully submitted and the script will try next time; the reasons include the number of jobs in the queue is exceeding the MaxJobs in CONFIG or the queue system limit, or just bad les. column three This is the job type which can be dimer, minimization, or dynmat. column four to six Those are for energy, force and curvature, respectively. No values will be marked na. It is mandatory that each process in the state must have an entry in the jobs le. If the users want to add some processes in the state folder (e.g., st0001), they also need to create entries in jobs.dat for those added processes.If a trivial jobid is needed so that the script will take the new processes as nished saddle point searches, we recommend using jobid marker.

5 st0005 new st0005 -376.0 -375.7 pr0032 2.4170e-05 3.631285e-05 6 st0003 repeat st0003 -375.7 -376.0 pr0011 2.4171e-05 6.048385e-05 7 st0006 new st0006 -375.7 -376.0 pr0018 2.4171e-05 8.465485e-05 8 st0007 repeat /somewhere/st0002 -375.7 -376.0 pr0038 2.4171e-05 1.0882585e-04 9 st0008 new st0008 -375.7 -376.0 pr0029 1.6978e-07 1.0899563e-04 File akmc.dat is in the aKMC home directory and contains the following information. First line It is an mandatory description line showing the meaning of each column in the following parts of the le: akmcStep stnumber property target stenergy(eV) ftenergy(eV) totaltime(s). column one This tells the simulation is at which KMC step. column two This is the name of the state. column third This tells if the state at that time was new or not. new means that state had never appeared before (including in the possible database) while repeat means the opposite. column four This is the address of the state. It looks trivial at most time, but if the state is repeated in some database, this address will be the key to get access to the information of the repeated state (see st0007 in the sample le). column ve The chosen process from that state. column six The energy of the state, in eV. column seven The energy of the chosen state, in eV. column eight The time elapse from this state, in eV column nine The total simulation time, in seconds.

VIII. D. AKMC.DAT

WHAT HAPPENS WHEN YOU TYPE AKMC.PL

Here is an sample akmc.dat le. Step stnumber property address stenergy(eV) ftenergy(eV) ChosenPr steptime(s) totaltime(s) 1 st0001 new st0001 -374.2 -374.8 pr0002 3.7758e-07 3.7758e-07 2 st0002 new st0002 -374.8 -375.4 pr0008 1.6978e-07 5.4736e-07 3 st0003 new st0003 -375.4 -375.5 pr0012 1.0949e-07 6.5685e-07 4 st0004 new st0004 -375.5 -376.0 pr0020 1.1486e-05 1.214285e-05

First of all, you should only run akmc.pl in the aKMC home directory, i.e., where the cong le is. After you type the command of akmc.pl, the script rst reads cong le. Then it checks MaxJobs set by the cong le. It has to be a non-negative number. After that, the script decides if the dynamic matrix calculation is necessary. The next step is to gure out the status of this aKMC run. It reads the akmc.dat le to pick up its memory of the previous step and current step and gure out if the current is a new state, or if it has been started, etc. If the script nds the current step (state) is still being calculated, it will tell the code (akmc check.pl)

7 to check jobs.dat le and update the job status information. Then it will switch to the main loop by reading information from st.dat le in the current state folder and handles each process based on its status and the updated job information. The screen output will tell you the progress of the script running. it occasionally hops from one local minimum to another. That hopping could contribute most to the LTD process. Transition state theory (TST) assumes that system has to cross over some transition state (usually a narrow bottle-neck type of high energy region on the PES) which the system rarely visits, to move from one local minimum to another. TST uses a statistical estimation of how fast the transition is in terms of a rate constant at a certain temperature. For a lot of solid systems, a harmonic approximation (hTST) can be assumed for both minima and transition states. Then, TST can give a most probable transition pathway (minimum energy path or mechanism) which connects two local minima via a saddle point. The rate of the transition can be calculated from the energies and normal modes of both local minimum and saddle point. Therefore, TST provides a solution to simulate the most important part of LTD - rare events. Combined with molecular dynamics simulation of frequent evens, a multiple time scale of LTD can be achieved with good eciency.

IX.

INTRODUCTION TO ADAPTIVE KINETIC MONTE CARLO SIMULATION

The original idea about adaptive kinetic Monte Carlo (aKMC) simulation was presented in the article by Graeme Henkelman and Hannes Jnsson [G. Henkelman o and H. Jnsson, Long time scale kinetic Monte Carlo o simulations without lattice approximation and predened event table, J. Chem. Phys. 115, 9657-9666 (2001)]. A new article on the improvement of aKMC is published by Lijun Xu and Graeme Henkelman [Lijun Xu and Graeme Henkelman, Adaptive kinetic Monte Carlo for rst-principles accelerated dynamics, Journal of Chemical Physics 129, 114104 (2008)]. The following paragraphs have also been posted on the web page of Henkelman group in the University of Texas at Austin (http://theory.cm.utexas.edu/henkelman/research/ltd/).

C.

Kinetic Monte Carlo

A.

Limits of Molecular Dynamics

Long time dynamics (LTD) refers to dynamic simulation of physical and chemical processes over the time scale much longer than traditional molecular dynamics does. Molecular dynamics simulates dynamic processes by solving Newtonian equation of motion via nite dierence approximation. The time step is typically at the femtosecond level. Therefore, molecular dynamics can only simulate events eciently up to pico- or nano-seconds. However, a lot of interesting events, such as diusions and chemical reactions, can only happen at the time scale of milli-seconds, seconds or even hours or days. Those events are usually called infrequent, rare, or slow events. Obviously, it would take an enormous number of steps to simulate rare events with molecular dynamics.

Combined with hTST, kinetic Monte Carlo (KMC) simulation is very useful in long time scale dynamics. Given those rate constants calculated from transition state theory, KMC can build an event table with all the transition mechanisms (for rare events). An event can be selected from Boltzman distribution by generating a random number; fast events have larger rate constants and therefore, are more probable to be chosen. The system conguration will be updated the end point of the chosen minimum energy pathway. The whole process is repeated for the new state. By this means, the evolution of the system on the PES can be eciently simulated over a long time scale.
D. Adaptive kinetic Monte Carlo

B.

Transition State Theory

The potential energy surface (PES) is a multidimensional surface on which each energy point corresponds to a possible conguration of the system. There are regions where the system spends most of its time. Those regions are usually local minima or low energy areas corresponding to stable system congurations (e.g., a stable particle, a favorable adsorbate conformation etc). Meanwhile, there are also regions the system rarely visits. From the view of dynamics on the PES, system always vibrates around those local minima (or basin areas), while

The key step in KMC combined with hTST is to build a reliable event table for each state the system visits. It is rare that those events in the table can be known beforehand (except for some simple systems). In most cases, the event table has to be built on the y (or adaptively) for each specic state. We call this type of long time scale simulation adaptive kinetic Monte Carlo (aKMC). It is going to be our main LTD tool.

E.

Saddle point searching methods

By now, the whole LTD problem has been reduced to nding those saddle points for a given conguration. There are quite a few methods on searching for saddle points (see review). Our choice is the Dimer Method developed by Graeme Henkelman and Hannes Jnsson.

8 Its a minimum mode following method using only the rst order derivative of the potential.
F. Empirical potential and DFT

Adaptive kinetic Monte Carlo simulation based on dimer method requires an accurate evaluation of the system energy and forces on atoms. Based on how energy and forces are calculated, aKMC can be carried out through either empirical potentials or rst principle methods such as density functional theory. Our group has developed a distributed computing system (EON) for using empirical potentials in the aKMC simulation.

It takes advantage of idle computing resources on the internet. The server sends out dimer searching jobs to client computers, collects the results when the jobs are nished and does the KMC steps. If there are no reliable empirical potentials available for the system, some quantum calculations have to be used to get energy and forces for the system. We have also developed a scripting program (AKMC) to do the aKMC simulation through the VASP density functional calculations. The AKMC scripts handles every detail of the adaptive kinetic Monte Carlo simulation and frees people from tedious things such as submitting and checking jobs. A single step of aKMC simulation can be a powerful tool for mechanism studies in Surface Science and Catalysis.