=Paper=
{{Paper
|id=Vol-1787/567-572-paper-99
|storemode=property
|title=Parametrization of the Reactive MD Force Field for Zn-O-H Systems
|pdfUrl=https://ceur-ws.org/Vol-1787/567-572-paper-99.pdf
|volume=Vol-1787
|authors=Konstantin Shefov,Margarita Stepanova,Sofia Kotriakhova
}}
==Parametrization of the Reactive MD Force Field for Zn-O-H Systems==
https://ceur-ws.org/Vol-1787/567-572-paper-99.pdf
Parametrization of the Reactive MD Force Field
for Zn-O-H Systems
K. S. Shefova, M. M. Stepanovab, S. N. Kotriakhovac
St. Petersburg State Institute University, 7-9 University emb, St. Petersburg, 199034, Russia
E-mail: a k.s.shefov@gmail.com, b mstep@mms.nw.ru, c dangemy@gmail.com
We describe a procedure of optimizing the molecular dynamic force field for Zn-O-H chemical systems by
means of a new parallel algorithm of a multifactorial search for the global minimum. This algorithm allows one
to obtain numerous parameters of the ReaxFF classical force field based on quantum chemical computations of
various characteristics of simple compounds. The force field may be then used for simulating of large-scale
chemical systems consisting of the same elements by means of classical molecular dynamics. Our current im-
plementation of the algorithm is done in C++ using MPI. We compare characteristics of simple compounds, ob-
tained by 1) quantum chemical techniques, 2) molecular-dynamic methods using reference parameters of the
force field, and 3) MD methods using optimized parameters of the force field. With the optimized parameter set
we perform MD simulations (using LAMMPS package) of crystals of zinc and zinc oxide of various modifica-
tions at the room temperature. Finally, we compare results of the parameter optimization procedure by means of
the algorithm described above and results of a parallel implementation of an evolutionary approach to minimum
search using dynamic models of Zn and ZnO crystals. Also we discuss advantages and disadvantages of the both
methods and their efficiency for extremal problems. All computations are performed with machines of the dis-
tributed scientific complex of the Faculty of Physics of St Petersburg State University.
Keywords: numerical simulation, chemically reactive systems, reactive force field, molecular dynamics,
parameter optimization, parallel algorithm, absolute extremum search
© 2016 Konstantin S. Shefov, Margarita M. Stepanova, Sofia N. Kotriakhova
567
Introduction
Reactive molecular-dynamic (MD) force field ReaxFF is a function of a large number of parame-
ters. Searching optimal values of these parameters is a complicated problem. To solve it one may use
different approaches, each of them having its own advantages and disadvantages. One of the ap-
proaches is the multifactorial global search algorithm (MGSA). The aim of this work is to estimate
efficiency of MGSA technique being applied to parameter search. We compare quality of parameter
sets for Zn-O-H systems obtained by different methods: MGSA, evolutionary algorithm and one-
parameter search technique [Raymand, Duin, …, 2010]. Quality of a parameter set is estimated 1) by
comparing of physical characteristics obtained by MD-methods and quantum chemistry methods, 2)
by stability of crystal lattices of Zn and ZnO during their simulation under room temperature, and 3)
by simulation of ZnO crystal growth. For MD-simulations we use package LAMMPS.
Global Search Algorithm
The global search algorithm (GSA)
[Strongin, 1978] allows one to obtain an absolute
minimum of a target function on a segment. It is
based on probability approach. Assume we already
have a set of known target function values in sev-
eral points on a segment. We use these known
values to find an interval between neighboring
points on the segment, where absolute minimum
location is the most probable. In this interval we
take a point corresponding to a mathematical ex-
pectation of position of the minimum and compute
target function value in it. The point is added to
the list of known values and we pass to the next
iteration. The algorithm terminates, when distance
between points of the segment of two consequent
iterations becomes less than a given criteria. In
multidimensional variant of the algorithm a func-
tion of many variables is mapped on a function of
a single variable with a help of Peano scans. Mul-
tidimensional definition domain (hypercube) is
mapped on a segment of the real axis. To make the
global search algorithm parallel we use a tech-
nique of rotating scans [Strongin, Gergel, …,
2009]. Each parallel process operates with its own
scan rotated by angles ±π/2 with respect to the
basic scan in particular pair of dimensions. In total
one can perform N∙(N - 1) of such rotations for N-
Fig. 1. MGSA parallel implementation flowchart
dimensional domain of definition. Thus, in total
the program uses N∙(N - 1) + 1 processes, each of
them performs GSA and at each iteration sends its
result to all other processes. Parallel algorithm
speeds up convergence and compensates points proximity information loss caused by use of a scan.
As a target function we use function (1).
568
= − + − (1)
Here are potential energies of training set models (simple chemical compounds, which opti-
mization is based on), are components of forces acting on atom α of model k, L is number of
models in the set, are numbers of atoms in model k, and are weight factors. Indices QC and Re-
axFF mean that corresponding characteristics are obtained either by quantum chemistry methods of by
molecular dynamics, respectively. and depend on parameters , , … , of the
force field. In our case we use multifactorial algorithm, in which in addition to the basic target func-
tion (1) we also use several functions-limitations on particular groups of addends in (1). This allows us
to set all the weights equal to 1 and not to solve a complicated problem of weights choice.
MGSA parallel implementation flowchart is in Fig. 1. Here points of a function definition domain
(a grid on a hypercube) are named , and values of a target function (or functions-limitations) in them
are named . Values of index (number of satisfied limitations in ) are named . Points of a grid on
the segment [0; 1], on which the hypercube is mapped, are named . Double frames in the flowchart
show that on a particular step data exchange between parallel processes takes place. Number of paral-
lel processes is named P. For more details of the algorithm see paper [Stepanova, Shefov, …, 2016].
Quantum chemistry computations
Basic compounds that are included in the training set are presented in Table 1. QC computations
are performed with the same methods that are used in the work [Raymand, Duin, …, 2010] that is used
for comparison.
Table 1. Compounds computed by quantum chemistry methods
Compound Modifications Comp. method
ZnO lattices P63mc, F-43m, CRYSTAL09, DFT/B3LYP
Fm-3m, Pm-3m Gauss basis, k-points 10x10x10
Zn lattices hcp, fcc, ABINIT, DFT/PAW/PBE
bcc, sc planewave basis
k-points 8x8x8, Ecutoff=30 Ha
ZnO thin films (10 layers) (1 0 -1 0) P63mc, (1 1 -2 0) P63mc, CRYSTAL09, DFT/B3LYP
(1 1 0) Fm-3m, (1 0 0) Fm-3m Gauss basis, k-points 10x10
Molecule ZnOH2 Different bond lengths Zn-OH, GAUSSIAN09 DFT/B3LYP
different angles OH-Zn-OH and Zn-O- basis 6-311+G
H
Molecule OHZnOZnOH Different angles OHZn-O-ZnOH GAUSSIAN09, DFT/B3LYP
basis 6-311+G
8 ZnO cells in vacuum P63mc, F-43m, Fm-3m GAUSSIAN09, DFT/B3LYP
Gauss basis
In Table 2 and Table 3 there are shown results of computations of ZnO and Zn crystal lattice con-
stants and bulk moduli (BM) compared to computations from the work [Raymand, Duin, …, 2010]
and physical experiment data. Sign "+" marks those crystals that are observed in nature.
Parameters optimization
Parameters of the ReaxFF force filed are optimized with the multifactorial global search algo-
rithm [Stepanova, Shefov, …, 2016]. We used function (1) as the target one. The optimization is done
under condition that a sum of every group of addends in (1) that corresponds to each particular mole-
cule or crystal does not exceed 1/8 of the estimate of the mean value of this sum. The mean value is
569
estimated by computing every sum value in 65000 points uniformly distributed over the search do-
main.
Table 2. Lattice constants and bulk moduli (BM) of zinc oxide crystals
Cryst. Charact. QC comp. QC [Raymand, ReaxFF opt. ReaxFF [Raymand, Exper.
Duin, …, 2010] Duin, …, 2010]
P63mc+ a,A 3.28 3.28 3.27 3.29 3.25
P63mc+ c, A 5.28 5.28 5.24 5.28 5.21
P63mc+ BM, GPa 143 136 138 144 141, 143
F-43m a, A 4.62 4.60 4.62 4.62 N/A
F-43m BM, GPa 144 162 133 130 N/A
Fm-3m+ a, A 4.34 4.30 4.35 4.29 4.27
Fm-3m+ BM, GPa 183 202 242 283 203, 228
Pm-3m a, A 2.68 2.68 2.77 2.61 N/A
Pm-3m BM, GPa 178 183 211 407 N/A
Table 3. Lattice constants and bulk moduli (BM) of zinc metal crystals
Cryst. Charact. QC comp. QC [Raymand, ReaxFF opt. ReaxFF [Raymand, Duin, Exper.
Duin, …, 2010] …, 2010]
hcp+ a, A 2.65 2.63 2.74 2.73 2.67
hcp+ c, A 4.98 5.06 4.47 4.46 4.95
hcp+ BM, GPa 74.1 66.7 81.3 87.7 64.5 - 75.1
fcc a, A 3.90 3.86 3.86 3.86 N/A
fcc BM, GPa 70.7 81.8 101.2 112.4 N/A
bcc a, A 3.12 3.06 3.08 3.06 N/A
bcc BM, GPa 67.8 84.6 73.8 73.5 N/A
sc a, A 2.64 2.71 2.76 2.71 N/A
sc BM, GPa 47.4 64.2 34.5 30.2 N/A
The full set of parameters has been preliminarily sorted by descending of their influence on the
target function [Stepanova, Shefov, …, 2016]. From that list first 24 parameters are optimized. They
are chosen by the largest correlation with the target function addends. The search for minimum is per-
formed on a grid of 4097 points per each parameter. Four parameters are optimized simultaneously.
The algorithm has performed 6 passes one time for each 4 parameters. Target function value was re-
duced 1.5 times during six passes. Each run takes from 36 to 48 hours to finish. Target function value
is computed in 1 million grid points in average during each run.
Fig. 2 shows relative values of all 24 parameters that take part in the optimization. The values are
measured in per cent of control values [Raymand, Duin, …, 2010] ("Control") taken for 100 %.
"MGSA" - parameters optimized by multifactorial algorithm, "EA" - ones optimized by evolutionary
algorithm. One can see that the values of parameters differ from method to method, and in some cases
they differ relatively a lot. Each method has its own target function, and it results in its own minimum.
Below we check how well each of these optimized parameter sets can simulate a real physical system.
Dynamic simulation of crystals
For three sets of parameters (Control, MGSA, EA) we performed simulation of crystals (600 at-
oms) of zinc and zinc oxide in vacuum at temperature of 300 K. The simulations were performed for
all crystal lattices that appear in Tab. 2 and Tab. 3. Those modifications of crystals that occur in nature
are adequately simulated with both the parameter set obtained by MGSA and the set obtained by the
evolutionary algorithm. This means their lattices retain their density and structure. The rest of crystal
modifications either partly deform or completely lose their form. In Tab. 4 as an example we show
570
snapshots of ZnO (P63mc) crystal in the end of simulation for three different parameter sets. One can
see stable enough crystal structure in all the cases.
Fig. 2. Comparison of ReaxFF parameters, optimized by different techniques
Table 4. Simulation of ZnO (P63mc) crystal at 300 K
Crystal Control params MGSA params EA params
ZnO (P63mc)
Using parameters obtained by MGSA we simulate growth of ZnO crystal layer on a (0 0 1) slab
of wurtzite ZnO. In Tab. 5 there are illustrated system states with the simulation time of 10, 100 and
300 ps. When the time is 300 ps we observe formation of the first layer of crystal and partly of the se-
cond layer. Quality of the particular simulation is not worse than the one in paper [Raymand, Duin, …,
2010].
Table 5. Simulation of ZnO (P63mc) crystal growth
t=10 ps t=100 ps t=300ps
Conclusion
Both MGSA and the evolutionary algorithm allow one to obtain parameters that give adequate
results when simulating particular Zn-O-H systems. Absolute values of parameters differ which is
caused by use of different target functions. MGSA converges relatively slow but achieves absolute
minimum with a grid precision. Evolutionary approach allow one to quickly localize those domains
where minima take place but searching of precise minimum location may turn out to be a very compli-
571
cated task depending on the form of multidimensional surface. Experience shows that in the given par-
ticular problem an absolute minimum of a target function does not always give the best solution since
the training set is inevitably limited. For this problem the most effective step would be introduction of
additional search criteria. This is implemented for MGSA [Stepanova, Shefov, …, 2016] but is also
reasonable for the evolutionary algorithm. Both approaches discussed here are applicable for ReaxFF
parameters search for various chemical systems.
References
Raymand D., van Duin A. C. T, Spangberg D., Goddard W. A. III, Hermansson K. Water adsorption
on stepped ZnO surfaces from MD simulation // Surface Science. — 2010. — Vol. 604. —
P. 741–752.
Стронгин Р.Г. Численные методы в многоэкстремальных задачах. — М.: «Наука». 1978.
Strongin R.G. Chislennye metody v mnogoekstremalnykh zadachakh [Numerical Methods in Multi-Extremal Prob-
lems]. — M.: «Nauka». 1978. — 240 p. (in Russian).
Стронгин Р.Г., Гергель В.П., Баркалов К.А. Паралельные методы решения задач глобальной
оптимизации // Известия ВУЗов. Приборостроение. — 2009. — Т. 52. — С. 25-33.
Strongin R.G., Gergel V.P., Barkalov K.A. Parallelnye metody resheniya zadach globalnoy optimizatsii [Parallel meth-
ods of global optimization problems solution] // Izvestiya Vuzov. Priborostroenie. — 2009. — Vol. 52. — P. 25–33 (in
Russian).
Stepanova M.M., Shefov K.S., Slavyanov S.Yu. Multifactorial global search algorithm in the problem
of optimizing a reactive force field // Theoretical and Mathematical Physics. — 2016. —
Vol. 187, Issue 1. — P. 603–617.
Шефов К.С., Степанова М.М. Реализация и применение параллельного алгоритма глобального
поиска минимума к задаче оптимизации параметров молекулярно-динамического потен-
циала ReaxFF // Компьютерные исследования и моделирование. — 2015. — Т. 7, № 3. —
С. 745752.
Shefov K.S., Stepanova M.M. Realizatsiya i primenenie parallelnogo algoritma globalnogo poiska minimuma k zadache
optimizatsii parametrov molekulyarno-dinamicheskogo potentsiala ReaxFF [An implementation of a parallel global
minimum search algorithm with an application to the ReaxFF molecular dynamic force field parameters optimization]
// Kompyuternye issledovaniya i modelirovaniye. — 2015. — Vol. 7, No 3. — P. 745–752 (in Russian).
572