Param Ga

Journal of Molecular Liquids 337 (2021) 116579
Contents lists available at ScienceDirect
Journal of Molecular Liquids

journal homepage: www.elsevier.com/locate/molliq
Efficient parameterization of intermolecular force fields for molecular

dynamics simulations via genetic algorithms
Abdullah Bin Faheem a, Jong-Yun Kim b,c, Sang-Eun Bae b,c, Kyung-Koo Lee a,⇑
a
Department of Chemistry, Kunsan National University, Gunsan, Jeonbuk 54150, Republic of Korea
b
Nuclear Chemistry Research Team, Korea Atomic Energy Research Institute, Daejeon 34057, Republic of Korea
c
Radiochemistry & Nuclear Nonproliferation, University of Science & Technology, Daejeon 34113, Republic of Korea
a r t i c l e i n f o a b s t r a c t
Article history: The development of van der Waals (vdW) parameters for molecular dynamics force fields typically
Received 6 April 2021 requires fitting the parameters to reproduce experimental physical properties, such as the density or
Revised 14 May 2021 the heat of vaporization. These fitting procedures depend on a limited number of conformations, involve
Accepted 24 May 2021
expensive DFT calculations, or need hand-tuning of the resulting parameters. Such methods have little to
Available online 27 May 2021
no support for the simultaneous fitting of the parameters, are time-consuming, and neglect coupling
effects since each parameter is fitted individually rather than a common set of calculations where all
Keywords:
parameters are derived simultaneously. Hence, in this study, we present a genetic algorithm (GA) that
Genetic algorithm
Molecular dynamics simulation
automates this fitting process. In this manner, the GA-based optimization can fit all the vdW terms at
Force field the same time and does not require any physical intervention. Once executed, the GA will automatically
Parameterization determine the optimum parameter set based on the properties included in the fitness function. The GA
Acetonitrile reported herein has been successfully applied to acetonitrile, a common organic solvent, to reasonably
reproduce both the thermodynamic and dynamic properties at room temperature as well as in a wide
temperature range. The use of GA not only serves as a powerful optimization tool but also opens up
new avenues in the development of accurate force fields that can provide valuable molecular insights.
Ó 2021 Elsevier B.V. All rights reserved.
1. Introduction [15,16], the intramolecular parameters include the equilibrium

bond lengths, bond and dihedral angles, along with the corre-
Classical molecular dynamics (CMD) is a powerful tool for sponding force constants and rotational barriers (for dihedral
studying the structure and dynamics of molecules at the atomic terms), while the intermolecular interactions are normally
level. Simulations based on classical force fields permit the user described by atom-centered partial charges and van der Waals
to work with large systems and long simulation times. Although (vdW) parameters. Ideally, general force fields should be complete
they cannot be used to describe electron transfer or bond forma- and transferable [17]. The term complete implies that the general
tion and dissociation, they are incredibly useful in investigating force field should have full parameter sets for most compounds,
the time dependence of systems containing thousands or millions while transferable means that the force field should be transferable
of atoms. Hence, CMD has become invaluable in studying systems between different molecules. For example, the parameters for the
that include but are not limited to biomolecules [1–3], surface methyl group in n-hexane, 1-hexene, or 1-hexanol should be iden-
science [4–6], and liquid-phase calculations [7–9]. tical. These force fields employ very similar methodologies in
The success of the CMD calculations is determined by their abil- developing parameters, which utilize extensive experimental or
ity to accurately reproduce experimental properties. However, the quantum calculation data sets for different types of molecules. As
accuracy of CMD simulations is highly dependent upon the force such, general force fields are usually applicable to a large variety
field, i.e., the functional form for the potential energy, and the of systems.
parameters used to define the intra- and intermolecular interac- For example, the intramolecular force field parameters in
tions in the system. In traditional all-atom force fields such as AMBER, OPLS, and CHARMM are normally fitted against DFT calcu-
GROMOS [10], AMBER [11,12], OPLS [13,14], and CHARMM lations at the appropriate level of theory [13,16,18]. Typically, it is
preferred that the level of theory used is not computationally
⇑ Corresponding author.
expensive and produces the least error. For the bond and angle
E-mail address: kklee@kunsan.ac.kr (K.-K. Lee).
parameters, the harmonic potential for the bonds and angles may
https://doi.org/10.1016/j.molliq.2021.116579
0167-7322/Ó 2021 Elsevier B.V. All rights reserved.
A. Bin Faheem, Jong-Yun Kim, Sang-Eun Bae et al. Journal of Molecular Liquids 337 (2021) 116579
be simply obtained by scanning a set of structures by varying the liquid-state properties such as densities and heats of vaporization
bond and angle values and plotting the energy profiles obtained [13,16,18]. Nevertheless, because parameters in general force fields
from the different structures. The equilibrium values correspond are optimized purely based on condensed phase properties and at
to the bond or angle value that gives the lowest energy. The asso- room temperatures, they may not produce valid results in the gas
ciated force constant is obtained by fitting a quadratic function or solid phase or at different temperature conditions. For example,
around this minimum. The equilibrium bond and angle values Kaminski et al. observed that the OPLS force field is unable to pre-
and their force constants can also be obtained from experimental dict the gas-phase dimerization energies of methanethiol and etha-
data such as infrared, microwave, or X-ray and neutron diffraction nethiol even though their liquid-state properties were in very good
studies and are compared to values obtained from CMD agreement with the experimental results [13]. In addition, pre-
calculations. sently available general force fields are not consistent in their per-
Although the dihedrals are described by a more complex func- formance for different types of molecules. This has been
tion for potential energy, they are parameterized in a similar way investigated extensively by Wang et al., where the General AMBER
to the bond and angle terms. A scan of energies of many possible Force Field (GAFF) was used to perform CMD simulations of differ-
torsion angles is conducted [13,16,18]. The obtained energies are ent molecules [26,27]. As pointed out by Wang et al., while as a
fitted to a truncated Fourier series. Usually, the performance of whole the average percentage error can be quite reasonable, the
the dihedral terms is evaluated by comparing the rotational scan GAFF parameters perform well in some cases and poorly in others.
about the dihedral of interest obtained from CMD to quantum For example, the diffusion coefficients obtained for proteins in
calculations. aqueous solutions reproduced experimental data to a much better
The nonbonded interactions consist of two main components, extent in comparison to organic solvents [26]. These deviations
namely, the electrostatic and vdW forces. The electrostatic interac- from experimental results were mainly attributed to the presence
tions are described by Coulomb’s law, where the interaction energy of unoptimized vdW parameters. Similar results have been
is proportional to the partial charges of the species. In contrast, the reported in previous literature as well, where researchers have
vdW interactions are most commonly defined using the Lennard- either tested out different parameter sets or fine-tuned a set of
Jones (L-J) potential function. Although different interatomic vdW parameters according to their needs [28–35]. As such,
potentials for the vdW interactions have been reported, such as because the intermolecular vdW parameters heavily influence
the Buckingham potential, the Born–Meyer–Huggins potential, the physical properties of the system, the need for parameter sets
the Tersoff potential, or the embedded-atom potential, the L-J that can reproduce experimental data is crucial.
potential remains the most widely used intermolecular potential Methods for deriving parameters during force field develop-
in classical many-body simulations. The reason for this is mainly ment employ a limited number of molecular conformations to
because of its simplicity and low computational cost; the use of a determine force field parameters from DFT calculations. Usually,
simple functional form with a fewer number of optimizable the parameters obtained from quantum mechanical calculations
parameters makes the L-J potential easy to use and computation- require further hand-tuning to match experimental physical prop-
ally inexpensive. Therefore, in commonly available force fields erties [18,29]. Such methods have little to no support for the simul-
such as AMBER, OPLS, and CHARMM, the vdW terms are repre- taneous fitting of the parameters, resulting in a large number of
sented by the L-J potential. Because the L-J terms are tightly cou- calculations and neglect coupling effects because each parameter
pled to the electrostatic terms, different force fields adopt is fitted individually rather than a common set of calculations
different approaches in parameterizing the intermolecular terms. where all parameters are derived simultaneously. This is particu-
For instance, in OPLS and CHARMM, the vdW parameters are para- larly the case for the vdW parameters, which are tightly coupled
metrized alongside the atomic partial charges. In this case, the with one another. As the hand-tuning approach involves fixing a
nonbonded parameters are optimized to reproduce ab initio inter- certain parameter and adjusting the remaining parameters, the
action energies and are validated against liquid-state experimental required calculations become increasingly expensive as the num-
properties, such as the heat of vaporization and density. The partial ber of parameters increase.
charges are often subjected to further optimization using the pop- For small molecules with a few parameters, hand-tuning the
ular ChelpG scheme, which fits the atomic charges to the quantum vdW terms might be successful. However, for large molecules, this
mechanical electrostatic potential (ESP) surrounding a molecule might not be possible due to the multidimensional parameter
[19–23]. This is in stark contrast to AMBER, where the partial space and the inherent coupling of those parameters. Nevertheless,
charges are obtained on a case-by-case basis and are derived to this can be long and tedious even for small molecules because one
reproduce the ab initio molecular ESP using the RESP approach has to rely on one’s chemical intuition to guess which set of param-
[24]. Then, the vdW terms are parameterized separately and vali- eters might provide an acceptable description of the system and if
dated against experimental properties. needed, manually vary each parameter to search for a set that most
As mentioned previously, the accuracy of CMD calculations is accurately describes the physics of the system. Such a hand-tuning
dependent upon the empirical parameters that define the potential approach requires an intimate knowledge of the chemical struc-
energy of the system. While the intramolecular parameters influ- tures of the molecules simulated, which might not always be pos-
ence the structure of the molecule, the intermolecular parameters sible [29].
describe the interactions between different molecules. Therefore, Recently, genetic algorithms (GA), an example of a very efficient
the intermolecular interactions heavily influence the physical optimization algorithm, have been widely used by chemists to
properties of the system, i.e., density, heat of vaporization, diffu- optimize MD force field parameters. GAs have been successfully
sion coefficients, etc. Nonbonded parameters require special atten- used to optimize dihedral terms [17,18], vdW parameters [25],
tion to define accurately the physics of any system. For assigning atomic charges [24] and polarizability [36], and reactive [37] and
partial charges, methods such as RESP and ChelpG are already well coarse-grain [38] force fields. GAs are a class of evolutionary algo-
established. On the other hand, a universal or widely accepted rithms that attempt to simulate the process of biological evolution
method for deriving vdW parameters does not exist. After the through steps such as selection, mutation, and crossover. Just as in
determination of the partial charges, the vdW terms are adjusted nature where organisms fight to survive, the genes that produce
to reproduce high level ab initio or experimental data [25]. the best individuals survive and reproduce. The power of the GA
The vdW parameters in commonly available force fields such as lies in its ability to efficiently deal with multidimensional or non-
AMBER, OPLS, and CHARMM are typically refined to reproduce linear problems and complex and poorly understood search spaces
2
[24]. In GAs, first an initial set of chromosomes or the initial pop- parameterization that can be directly compared to the GA’s perfor-
ulation is generated. In the context of GAs, the aforementioned mance [29,35]. From the results reported herein, the use of GA
genetic material or a chromosome can be any set of parameters proved to be incredibly efficient in finding the global minimum
that give a solution or result. The initial population is then evalu- and thus a more practical solution to parameter optimization prob-
ated via a scoring function, or in other words a fitness function, lems in comparison to the traditional hand-tuning approach.
where a better fitness value indicates a superior performance of
the parameter set. Next, in the selection operation, chromosomes 2. Methodology
with high fitness get to propagate to the next generation, and those
having low fitness perish. The probabilistic selection of organisms 2.1. Force field
for mating introduces additional variety into the population and
prevents falling into a local minimum. New chromosomes are then In this study, we employ the following all-atomic force field,
generated through the crossover and mutation steps. The new which describes the intra- (Uintra) and intermolecular (Uinter) inter-
genes extend the reach of the algorithm outside the pool of values actions as follows:
from the previous generation. X X
2
While GAs have been used by previous researchers in optimiz- U intra ¼ K b ðb b0 Þ þ K h ðh h0 Þ2
ing or developing parameter sets for CMD, most studies have Bonds Angles
focused on developing intramolecular parameters. For example, X Vn

þ ð1 þ cos½n£ dÞ ð1Þ
Wang et al. introduced ‘‘Parmscan,” an automatic search engine 2
Dihedrals
that utilized GAs to optimize the torsional parameters for the
AMBER force field [17]. In the aforementioned study, the Fourier ( " 12 6 # )
X
N1 X
N
rij rij qi qj
terms for dihedral parameters in the AMBER force field were devel- U inter ¼ 4eij þ þ ð2Þ
oped to reproduce the relative energies of 11 different hydrocar- i ¼ 1 j ¼ iþ1
rij r ij r ij
bons obtained from quantum calculations. Similarly, the GA
module ‘‘Paramfit,” was developed to optimize the bond, angle, where b, h, n, and £ are the bond length, angle, dihedral multiplicity
and torsional parameters for CMD by fitting the parameters against and dihedral angles, respectively. The b0, h0, and d values represent
energy calculations at the quantum level [18]. The software mod- the equilibrium values of the bond, angle, and dihedral terms; and
ule was tested by developing the dihedral parameters for alanine Kb, Kh, and Vn are the associated force constants. The L-J parameters
tetrapeptide and lipid tails. In a study by Courcot et al., the in Eq. (1), eij and rij, are obtained from the Lorentz–Berthelot mixing
intramolecular parameters for polymolybdate and polytungstate rules in Eq. (3). The nonbonded electrostatic interactions are
clusters were developed by comparing the equilibrium structures described by the last term in Eq. (2), where qi and qj are the partial
from CMD to X-ray diffraction studies [39]. However, since inter- charges of atoms i and j, respectively, and rij represents the distance
molecular interactions can heavily influence the physical proper- between those atoms.
ties of the system, protocols for parameterizing the vdW
pffiffiffiffiffiffiffiffiffi rii þ rjj
interactions require special attention. Nevertheless, to the best of eij ¼ eii ejj ; rij ¼ ð3Þ
2
our knowledge, only a limited number of GA studies have been
conducted for the development of vdW parameters. For instance, As this study uses GA to optimize the L-J parameters for ACN,
Wang et al. employed the use of GA for developing the vdW this leads to 8 adjustable terms. These are the eii and rii values of
parameters for pure organic liquids, where the optimization 4 different atom types (see Fig. 1).
scheme used experimental densities and ab initio interaction ener- For the intramolecular terms, the equilibrium bond lengths,
gies to fit the parameters [25]. However, such a methodology does angles, and torsional angles for ACN were determined from DFT
not take into account the dynamic properties of the system, such as calculations at the MP2/aug-cc-pvtz level via Gaussian 09 [41],
the diffusion coefficient or the viscosity. Furthermore, because the while the force constants were extracted from the GAFF parameter
densities were estimated from intermolecular interaction energies set. Subsequently, the partial charges of the molecules studied
and were verified only at certain temperatures, it is not evident were obtained from the ESP of the optimized structures, using
whether the performance of the parameters optimized by GA is the RESP method available in the AMBER 12 package [42]. Previ-
consistent at different temperatures. In another study by Chen ously developed general force fields have employed the use of
et al. [40], the vdW parameters for metal–organic frameworks the HF/6-31G* level of theory to obtain the ESP. However, it is
(MOFs) were optimized for predicting the CO2 and CH4 adsorption well-known that this method uniformly overestimates molecular
isotherms in MOFs. The parameters were optimized to reproduce polarity. Indeed, the HF/6-31G* theory level yields dipole moments
ab initio interaction energies and the CMD-obtained adsorption that are approximately 10% larger than those observed in the gas
isotherms were validated against experimental data. Therefore, it
is apparent that GA studies on the development of intermolecular
vdW parameters are scarce. As discussed previously, since the vdW
parameters heavily influence the physical properties of the system,
work on the development of optimization protocols for the vdW
parameters requires significant attention.
Hence, in this study, we investigate the use of GAs and their
application in the parameterization of the intermolecular L-J
parameters for classical force fields. Using acetonitrile (ACN) as
the case study in this work, we present an optimization scheme
that is not only able to reproduce experimental physical properties
for a wide temperature range but also the microstructural charac-
teristics obtained from highly accurate first-principles MD (FPMD)
and experimental data. In this work, the reasons for choosing ACN
are twofold: (1) it is a very common organic solvent with a wide Fig. 1. Structure of ACN with atomic labels. Colors: nitrogen, blue; carbon, gray;
range of applications and (2) is the subject of extensive manual hydrogen, white.
3
Table 1 obtained from the CMD simulations are used to calculate the fit-
Optimized L-J parameters obtained from GA. Refer to Fig. 1 for atom-type definitions. ness values of the population. Following the CMD simulation and
Atom-type r (Å) e (kJ mol1) the determination of the fitness values, the offspring are produced
N 3.1106 0.5600 after a sequence of steps, namely, selection, crossover, and muta-
C1 3.5422 0.5312 tion. In the selection step, a mating pool is generated, which is used
C2 3.3199 0.4530 to create the next generation. The GA operators, crossover and
H 2.4574 0.0188 mutation, are then applied to the individuals in the mating pool
to obtain the offspring. The GA used in this study employs elitism,
where the best parents are introduced into the offspring popula-
phase [12,43]. This difference is exploited in condensed phase sim- tion without any changes (described later). These steps are
ulations because it takes into account the implicit polarization of repeated until the desired termination criteria is reached.
the molecule. Before starting the GA-based optimization, the appropriate pop-
Alternatively, polarization effects can be included in the vdW ulation size should be determined to minimize the number of eval-
parameters as well. For example, Nikitin et al. used DFT calcula- uations. The number of evaluations is obtained by multiplying the
tions at the MP2 level of theory to reproduce experimental dipole total number of generations taken for the GA’s convergence by the
moments for ACN and then optimized the L-J parameters against population size. In general, the optimum population size can vary
experimental physical properties [44]. Similarly, since the GA significantly and is highly dependent on the optimization problem.
reported in this study takes into account the polarization effects In an analysis of previous studies where GAs were used to optimize
with the optimized L-J terms, the ESP was calculated using the parameters for CMD force fields, the population sizes used were
B3LYP/6-311G++(d,p) level of theory. Such a method was chosen observed to vary from 12 to 128 [18,24,47–49]. For example, Leo-
because the calculated dipole moment (4.08 D) was much closer narski et al. reported that increasing the population size from 128
to the experimental value (3.92 D) than that obtained from HF/6- to 8192 did not result in better fitness values [38]. In another study
311G+(d) (4.24 D) or MP2/6-311G++(d,p) (4.30 D) [45,46]. by Marques et al., increasing the population size from 12 to 60
while keeping the number of evaluations the same, did not
2.2. Van der Waals parametrization improve the final results [48]. This implies that if the population
size is greater than a certain value, better parameters and a
The classical representation of genes in GAs is in the form of a decrease in the total number of evaluations cannot be achieved.
string of bits, i.e., 0s and 1s. Such types of GAs are also referred Hence, we have opted to use a population size of 12, the smallest
to as binary-coded GAs. The bit strings in such GAs undergo cross- reported population size for force field optimization. From the
over and mutation steps to produce the offspring and are con- results described herein, it is evident that such a population size
verted to real values to express the genes. In this manner, is sufficient for converging to an optimum solution in a reasonable
binary-coded GAs intuitively mimic biological evolution. For the number of generations.
case of force field optimization, the bit strings are converted to real To generate the initial population, we used the parameters from
values, which in turn are used as parameters to run CMD simula- GAFF. At the beginning of the optimization, an initial population is
tions. The trajectories obtained with the parameters are then eval- generated by randomly selecting parameters within 10% variance
uated using the GA algorithm. Binary-coded GAs have been of the GAFF parameters. The parameters obtained at each genera-
previously used in the development of force fields. In a study by tion are used to run CMD simulations. From these, the physical
Angibaud et al., binary-coded GAs were used to optimize the reac- properties, i.e., the densities and diffusion coefficients, are deter-
tive force field for silicon [47]. However, for bit strings, changing a mined and used to calculate the fitness value using Eq. (4). Individ-
single bit can result in a large change in its real value counterpart. uals with a low fitness function value encode better solutions (are
Thus, binary-coded GAs are suitable for exploring a wide parame- more fit) and are better able to reproduce experimental densities
ter space because of the characteristics of the binary representa- and diffusion coefficients. The density and diffusion coefficient
tion of genes. However, because CMD simulations are sensitive to were used in the fitness function to replicate the experimental
the values of the force field parameters, a large change in the equilibrium and dynamic characteristics of the system.
parameters from one generation to the next can introduce unreal- 1
0 calc exp
istic solutions into the population, making binary-coded GAs qcalc qexp D D
unattractive for force field optimization. Fitness ¼ 100 @ þ A ð4Þ
An alternative to binary-coded GAs is using real values directly
qexp Dexp
to create the offspring, also called real-coded GAs. Real-coded GAs
hold a particular advantage over the binary representation of In Eq. (4), the qcalc and qexp represent the calculated and experimen-
genes, where it is possible to limit the range within which the off- tal densities, respectively, while Dcalc and Dexp are the calculated and
spring population is obtained. In other words, the offspring do not experimental diffusion coefficients, respectively. A point to be noted
contain genes that are substantially different from those of their is the exclusion of any adjustable weighting factors in Eq. (4). Nor-
parents. For the case of force field optimization, the starting mally, the fitness function includes more than one term against
parameters for GA are already predefined from quantum calcula- which the CMD parameters are fitted and usually includes weight-
tions or general force fields. Since parameters for CMD contain ing factors [25,38,50]. By adjusting the weighting factors, one can
the physical properties of the target molecules, it is crucial to consider the contribution of one term to be more than the other
search the parameter space that is physically meaningful and large terms during parameter optimization. However, as there is no clear
changes between the parameters of subsequent generations are link between the MD parameters and a given physical property, the
not necessary and might give unrealistic parameters. Therefore, use of such a method might produce a force field that is biased only
in this study, we employ the use of real-coded GAs, where the real toward a particular physical property, which is undesirable. Hence,
values were used directly to create offspring. we employ a more general-purpose fitness function that does not
The flow chart in Fig. 2 displays an overview of the main steps include weighting factors. In Eq. (4), the absolute difference (for
performed by the GA used in this work. In the beginning, an initial example, |Dcalc – Dexp|) is divided by the experimental value, which
population is created. Then, the parameters in the initial popula- normalizes the values of each term in Eq. (4) and allows equal con-
tion are used to run CMD calculations, after which, the properties tribution from all terms. In this work, the density and diffusion
4
Fig. 2. Flowchart of the genetic algorithm used herein.
coefficient of ACN at 298 K, i.e., 0.7766 g cm3 and 4.34 105 cm2 over, after selecting a random cut point along the sequence of
s1, respectively, were used as the experimental values in the fit- the two parents, the children are obtained by swapping the param-
ness function [29,35,51,52]. eters among the two parents [48]. However, for real-coded GAs,
The GA used herein adopts elitism, where the fittest individuals discrete recombination operators cannot introduce new values
are guaranteed a place in the next generation. Individuals with into the next generation; they just rearrange existing ones. This
high fitness values can be lost if they are not selected to reproduce results in a poor sampling of the search space and may require
or if crossover or mutation operators change them. Elitism can sig- large population sizes (as discussed previously). An alternative to
nificantly improve the GAs performance since it can be used to discrete recombination operators is the use of arithmetic crossover
eliminate the chance of any undesired loss of information during operators, where the children are produced from the weighted
the crossover or mutation stages [53–55]. Therefore, based on average of the parents. Such crossover methods can introduce
the fitness function (Eq. (4)), the parents are ranked and the top new values into the offspring population and possess a greater
two fittest parents are propagated into the offspring population search power in comparison to the discrete recombination opera-
without alterations. tors. The simulated binary crossover (SBX) is a well-known exam-
The selection operator selects suitable candidates for the mat- ple of an arithmetic crossover operator and has achieved good
ing pool from the parent generation. We employed the use of the results in several problems of varying difficulty and dimensionality
tournament-type selection because this selection method was [48,54]. An illustration of the SBX crossover is presented in Fig. 3
observed to work well with the crossover operator (discussed (c). In the SBX crossover, the offspring (Child 1 and Child 2) are
later) used in this study [48]. The tournament-type selection step obtained from the parents (Parent 1 and Parent 2) in the following
consists of picking two random individuals from the parent popu- manner [48]:
lation and comparing their fitness values. The individual with a
higher fitness value is selected and added to the mating pool. This (1) select a random value, l, that is between 0 and 1
process is repeated until the desired population size is reached (12 (2) calculate b, where.
in this case). A schematic of the tournament-type selection is illus-
trated in Fig. 3(a). 1
For the crossover operator, care should be adopted, as it is the b ¼ ð2lÞn þ 1 ; if l 0:5 ð5Þ
main part of the GA algorithm that is mainly responsible for
exploring the parameter search space. An example of a common n þ1 1
1
crossover method is shown in Fig. 3(b), termed the one-point b ¼ ; if > 0:5 ð6Þ
crossover. This crossover operator is considered to be among the 2ð1 - lÞ
class of discrete recombination operators. In the one-point cross-
5
the results in Section 3 indicate a favorable GA performance and

thus the used methods are sufficient.
2.3. Classical molecular dynamics
The starting configuration for performing CMD simulations

with GA-derived parameters at each generation was generated
using the GAFF parameters. Usually, at the beginning of the CMD,
simulations in the NPT ensemble are conducted to equilibrate the
simulation box. Such calculations may require simulation times
in the range of several tens of nanoseconds. However, since the
starting structure used herein is pre-equilibrated by GAFF param-
eters, computational overheads from these long simulations in
the NPT ensemble are eliminated, thus substantially reducing the
overall time taken for each GA generation.
To obtain the starting pre-equilibrated structure, first, an initial
configuration was generated via Packmol [49]. The systems simu-
lated contained 512 ACN molecules that were placed in a cubic
box. Then, the simulation box was heated from 0 to 800 K for
1 ns. This was followed by a 1 ns cooling step from 800 to 298 K.
Both the heating and cooling steps were performed in the NVT
ensemble. After which, a 2 ns equilibration run was performed in
the NPT ensemble. In order to obtain the physical properties from
CMD to evaluate the fitness values of each GA generation, the
pre-equilibrated box was subjected to a 1 ns NPT simulation to
equilibrate the box with the L-J parameters from GA. This was then
followed by a 1 ns production run at NPT to calculate the required
physical properties. To examine the effect of the simulation time
on the calculated densities and diffusion coefficients of pure ACN,
1 and 20 ns MD simulations (production run) were used to calcu-
late the physical properties. It was observed that increasing the
Fig. 3. Schematic illustration of the (a) tournament-type selection, (b) one-point time of the production run from 1 to 20 ns lead to no statistically
crossover, (c) SBX crossover, and (c) mutation operator. The letters, i.e., ‘‘A,” ‘‘B,” or relevant changes in the values of the calculated densities and dif-
‘‘C,” represent the individuals in the parent population. fusion coefficients
All CMD simulations were performed with the LAMMPS [56]
package at 298 K (unless mentioned otherwise) and 1 atm under
(3) obtain children periodic boundary conditions. The temperature and pressure were
controlled by the Nosé–Hoover thermostat and barostat with
Child 1 ¼ 0:5½fð1 þ bÞ Parent 1g fð1 bÞ Parent 2g ð7Þ relaxation times of 0.2 and 1.0 ps, respectively [35,57,58]. Equa-
tions of motion were integrated with a time step of 1 fs. The Parti-
Child 2 ¼ 0:5½fð1 þ bÞ Parent 1g fð1 bÞ Parent 2g ð8Þ cle–Particle Particle–Mesh algorithm was employed for long-range
coulombic interactions and a cutoff of 12.0 Å was used for non-
where n is a nonnegative real number. Large values of n create off-
bonded interactions. The atomic coordinates were saved every
spring that are close to their parents, while a small n allows distant
0.1 ps for analysis.
values to be selected for the children [48]. In this manner, selecting
a large n value produces offspring that lie within a reasonable range
2.4. First-principles molecular dynamics
of possible solutions. A value of n = 50 was selected for this study
and a crossover probability of 100% was employed, implying that
FPMD calculations were conducted using the hybrid Gaussian
all the children were created through the SBX crossover method.
and Plane Wave density functional implemented in the CP2K/
In GAs, the mutation operator causes a random jump from the
Quickstep program [59]. The valence electrons were described by
original value (Zorg) in a solution set to a mutated one (Zm). This
the molecularly optimized double-f basis set (MOLOPT-DZVP-SR-
is in stark contrast to the crossover operator, where two parents
GTH) and the Perdew–Burke–Ernzerhof functional, whereas the
are used to generate the offspring, as seen in Eqs. (7) and (8). How-
core electrons were treated with the Goedecker–Teter–Hutter
ever, for developing CMD force fields, if the difference between Zm
(GTH) pseudopotentials. A total of two parallel FPMD runs were
and Zorg is large, the offspring population may contain unreason-
conducted. Initial configurations for the FPMD simulations were
able values. Therefore, the mutation operator used herein changes
taken from random CMD snapshots using the GAFF parameters. A
the value (Zorg) of a random parameter in the offspring population
simulation box consisting of 60 ACN molecules was used for all
by ±10%. The probability of an individual to undergo mutation was
FPMD simulations.
set to 5%.
All FPMD simulations were run using the NPT ensemble at 298 K

Z m ¼ Z org þ Z org Nð-0:1; 0:1Þ ð9Þ and 1 atm under periodic boundary conditions. The temperature
and pressure were controlled via the Nosé–Hoover thermostat
where N(0.1, 0.1) represents a random value between 0.1 and and barostat with time constants of 0.2 and 1 ps, respectively.
0.1. The cutoff of the auxiliary plane wave was set to 500 Ry. Grimme’s
It is important to note that the choice made for the crossover DFT-D3 model was also used to take into account vdW interactions
and mutation probabilities may be further optimized by conduct- [60]. A total of 100 ps of FPMD trajectory was obtained for each
ing a rigorous evaluation of all possible combinations. However, FPMD run, where a time step of 0.5 fs was employed and the coor-
6
dinates were saved every 0.01 ps. Because the energy (energy vs nent of the pressure tensor (ab = xy, xz, yz; as a – b) at time t and is
time) exhibited changes in the first 10 ps, the trajectory during calculated via the following equation [63]:
the first 10 ps was discarded and the remaining trajectory was !
used for further analysis. 1 XN
pia ðtÞpib ðtÞ X N
Pab ¼ þ rija ðtÞf ijb ðtÞ ð16Þ
V i¼1
m i i <j
2.5. Physical property calculations
pi(t) is the center of mass momentum of molecule i with mass mi,
2.5.1. Density rij = ri – rj is the intermolecular distance, and fij is the force on mole-
The average bulk density (q) can be calculated from the average cule i by molecule j, and the summation is over all N molecules. The
volume of the simulation box, hVi, using Eq. (10) [25,27], where Nres viscosity was calculated using an average of the Pxy, Pxz, and Pyz val-
is the number of molecules in the simulation box, M is the molar ues. To obtain statistically reliable viscosity values, additional 3 ns
mass of ACN, and NA is the Avogadro constant. simulations in the NVT ensemble were carried out.
Nres M
q¼ ð10Þ
N A hVi 3. Results and discussion
To verify the ability of the GA algorithm in optimizing the vdW

2.5.2. Diffusion coefficient
parameters, the GA procedure described herein was used to
The translational motion of a molecular species can be
develop the L-J parameters of ACN. There are two major reasons
described by calculating the mean square displacement (MSD).
for selecting ACN to test the performance of the GA. First, ACN is
This manner of analysis provides insights into the rate at which
a common organic solvent with a wide range of applications, and
molecules move. The diffusion coefficient (D) can be obtained from
second, it has been the subject of extensive manual parameteriza-
the MSD via the following Einstein relation [26,29]:
tion which can be directly compared to the GA’s performance
1 d [29,35].
D ¼ lim MSD ð11Þ
6 t!1 dt In this section, we present the efficiency of our GA in parame-
terizing the L-J terms and then discuss the performance of the opti-
1 XN
mized L-J parameters in calculating the physical and
MSD ¼ hðx x0 Þ2 i ¼ ðxn ðtÞ xn ð0ÞÞ2 ð12Þ
N n¼1 microstructural properties. To best our knowledge, this is the first
study where GA has been used to derive parameters for ACN that
where xn(0) is the initial position of the molecule; xn(t) is the are able to not only reproduce experimental physical properties
position of the molecule at a time t, and N is the total number of for a wide temperature range but also structural characteristics
molecules. from highly accurate FPMD and neutron scattering data.
2.5.3. Heat of vaporization

3.1. GA force field parameterization
The heat of vaporization (Hvap) was determined using the same
approach as that of Wang et al. [25], and is illustrated as follows:
As discussed in Section 2, the experimental density and diffu-
Hv ap ðTÞ ¼ Epotential
gas ðTÞ Epotential
liquid ðTÞ þ RT þ C ð13Þ sion coefficient were used to fit the L-J parameters for ACN. Fig. 4
illustrates the evolution of the fitness, density (q), and diffusion
1 coefficient (D) of the individual that has the best fitness value in
Epotential
gas ðTÞ ¼ Eminimized
gas ðTÞ þ RTð3N atom 6 Ncons Þ ð14Þ the generation. Initially, the starting parameters are generated by
2
randomly selecting parameters within 10% of the GAFF parameters
where Epotential
gas and Epotential
liquid are the potential energies calculated and give densities and diffusion coefficients that are far away from
in the gas and liquid phases, respectively; Epotential can be directly the experimental values. This is indicated by the fitness value of
liquid
the fittest individual in Fig. 4(a), which is 27.6 at the beginning.
calculated from liquid-phase MD simulations and Epotential
gas can be
However, as the GA progresses, the fitness value decreases and
estimated via Eq. (14); Eminimized
gas is the energy after energy mini- eventually converges to a value of 0.42 at generation 28. Because
mization in the gas phase; and Natom and Ncons are the number of the fitness value does not change for 20 or so generations, it can
molecules in the simulation box and the number of constrained be said the L-J parameters have reached their optimum values
degrees of freedom, respectively. The correction term C in Eq. and the GA has converged. The q and D values of the fittest individ-
(13) accounts for the difference between the vibration energies cal- ual at each generation in Fig. 4(b) and (c) exhibit similar behavior.
culated quantum mechanically and classically, as well as the polar- In the beginning, the q and D deviate by approximately 13.4% and
ization and nonideal gas effects. However, since in most cases C is 14.2% from their experimental values (dashed horizontal lines), but
very small, it can be neglected and was not used in this work as the GA progresses iteratively, they approach values almost sim-
[25,27]. ilar to the experimental measurements. The L-J parameters
obtained from the GA optimization are listed in Table 1.
2.5.4. Shear viscosity To determine quantitatively the reliability of the L-J parameters
The shear viscosity (ƞ) can be determined from the following obtained by GA, the physical properties (q, D, Hvap, and ƞ) obtained
Green–Kubo expression where ƞ is obtained from the autocorrela- using the GA-derived parameters are compared with the previ-
tion function of the off-diagonal elements of the pressure tensor ously developed force fields using the mean absolute percentage
[61–63]: error (MAPE) as follows:
Z1 N
V 1 X
Ai F i 100
g ¼ Pab ðtÞPab ð0Þdt ð15Þ MAPE ¼ ð17Þ
kB T N i¼1 Ai
0
where V is the volume of the simulation box, kB is the Boltzmann where Ai and Fi are the experimental and calculated properties,
constant, and T is the absolute temperature. Pab(t) is the ab compo- respectively, and N is the total number of properties (4 in this case).
7
Fig. 4. (a) Fitness value, (b) density (q), and (c) diffusion coefficient (D) of the fittest individual in each generation. Experimental values are indicated by the dashed lines.
Table 2
Comparison between the physical properties and mean absolute percentage error (MAPE) of the force field obtained from GA optimization and the models developed in previous
work (at 298 K).
Model q D Hvap ƞ MAPE

(g cm3) (10–5 cm2 s1) (kJ mol1) (mPa s) (%)
Experiment [29] 0.7766 4.34 32.94 0.3410 –
Grabuleda et al. [64] 0.7112 4.09 30.63 0.2805 9.73
Vanommeslaeghe et al. [65] 0.7860 3.08 34.39 0.4631 17.62
Koveraga et al. [29] 0.7639 4.35 30.67 0.3188 3.81
Orhan et al. [66] 0.7555 4.75 29.79 0.2807 9.84
Kowsari et al. [35] 0.7758 3.82 32.26 0.3356 3.93
From GA 0.7799 4.34 31.13 0.3180 3.16
From Table 2, it can be observed that the GA-derived model pro- Table 2 shows the ability of the GA reported herein to optimize
vides the best overall results with the lowest MAPE. Therefore, the the vdW parameters.
GA-optimized L-J terms can reproduce the experimental properties To test the reproducibility of the GA, additional GA optimiza-
to a better extent overall in comparison to previously reported tions were also conducted, and the statistics are presented in the
force fields. The performance of the GA force field is similar to supporting information. We found that, in general, there is no sta-
the results reported by Koveraga et al. [29] and Kowsari et al. tistically significant difference in achieving convergence with dif-
[35], where the L-J terms of ACN were parameterized by manually ferent initial parameters. This can be seen in Table S1, where the
changing the parameters to match the experimental physical prop- standard deviation for the physical properties among all the GA
erties. A point to note is that the heat of vaporization and viscosity, runs was less than 0.5% of the average. This indicates the reason-
which were not used in the fitness function for GA optimization, is able reproducibility of the physical properties by different GA runs.
slightly underestimated. However, the GA-derived parameters can However, because of the random generation of the initial parame-
reasonably estimate the experimental results. The comparison in ters, each different GA run might converge to a slightly different
8
parameters. The underestimated densities by GAFF highlight the

disadvantages of using general force fields discussed in Section 1.
Therefore, the development of force optimization methodologies
is crucial for MD studies. In comparison with GAFF, the CHARMM
force field (Vanommeslaeghe et al. [65]) gives densities that are
much closer to the experimental results in comparison with GAFF.
However, it does not give reasonable diffusion coefficients (which
will be discussed in the following sections). From Fig. 5, the GA
force field and the parameters developed by Kowsari et al. [35]
exhibit the best performance at different temperatures, whereas
the force fields reported by Koveraga et al. [29] and Orhan et al.
[66] underestimate the densities higher temperatures. In conclu-
sion, the calculated densities with the GA parameters well repro-
duce the experimental results; both the magnitudes and the
overall trend for the temperature range.
The diffusion coefficients and viscosity are key physical proper-
ties because they represent the dynamics of the system. Such prop-
erties are useful in studying the motions of molecules with respect
to one another and shed light on the transport mechanisms that
Fig. 5. Density of pure ACN at different temperatures calculated with the GA force govern the system. Therefore, to evaluate the performance of the
field (red circles) and the different force fields developed by Grabuleda et al. [64] GA force field in determining the dynamic properties of the system,
(green triangles), Vanommeslaeghe et al. [65] (blue diamonds), Koveraga et al. [29] diffusion coefficients and viscosities were calculated at different
(cyan triangles), Orhan et al. [66] (magenta diamonds), and Kowsari et al. [35] temperatures using the GA-derived parameters and compared
(purple triangles), in comparison with experimental measurements (solid dashed
line) [67].
with previously developed force fields and experimental results
(Fig. 6). The experimental diffusion coefficients and viscosities
were obtained from Hurle et al. [67] and Dymond et al. [68],
fitness value (Table S1). As such, several GA optimizations were respectively.
performed and the parameters that gave the best fitness values From Fig. 6(a), although all of the investigated force fields
among the different runs were reported. reproduce approximately similar experimental trends, differences
in the performances of the different force fields are apparent. As
3.2. Molecular property calculations mentioned previously, while the CHARMM force field is able to
estimate the experimental densities adequately, it significantly
3.2.1. Temperature dependence of density, diffusion coefficient, and underestimates the diffusion coefficients (by more than 25%). Sim-
viscosity ilarly, previously developed force fields either reasonably repro-
Fig. 5 compares the temperature dependence of the density of duce densities or the diffusion coefficients. This means that none
the GA force field with previously reported force fields and exper- of the previously reported force fields can predict both the densi-
imental data from the results by Hurle et al. [67]. In general, all the ties and the diffusion coefficients simultaneously. In comparison,
force fields follow experimental trends, with the density of ACN the GA-derived parameters can reasonably reproduce both the
showing a decreasing trend at increasing temperatures. However, experimental densities and diffusion coefficients. In addition to
the performances of the different force fields vary significantly. the experimental trends, because the diffusion coefficients signify
For example, while the force field by Grabuleda et al. [64] follows the translational dynamics of the molecules, the diffusion coeffi-
the experimental trends, the density is 7–10% lower than cients at different temperatures can be used to calculate the activa-
experimental results. Since this is a version of the GAFF force field tion energy needed for ACN to leave its local coordination
for ACN, this error is most likely due to the unoptimized L-J environment. The activation energy was calculated from the Arrhe-
Fig. 6. (a) Diffusion coefficient and (b) viscosity of pure ACN at different temperatures obtained from the GA force field, previously developed models, and experimental data.
The colors and symbols are the same as in Fig. 5.
9
nius plots of Fig. 6(a) and is listed in Table S2. As seen in Table S2, (AE) and the relative error (RE) (see supporting information). The
the GA-derived force field produces an activation energy that is values for AE and RE are summed over the temperature range
closest to the experimental measurements. This shows the efficacy shown in Figs. 5 and 6 and are listed in Table S3. Among the inves-
of the GA-derived force field in being able to provide accurate tigated force fields, although certain physical properties are better
insights into the transport mechanisms affecting ACN molecules. predicted by other force fields and give smaller errors, our GA-
Further validation of the dynamic properties using the calcu- derived parameters exhibit the best overall performance with
lated and experimental viscosities is shown in Fig. 6(b). It can be small AE and RE values. In general, when using the GA force field,
observed that even though the calculated viscosities follow similar it is possible to reasonably reproduce both the thermodynamic and
trends as the experimental values, the studied force fields are dynamic properties at room temperature as well as in a wide tem-
unable to accurately reproduce experimental values. Almost all perature range (Figs. 5 and 6). This signifies the efficacy of the GA
the force fields in Fig. 6(b) underestimate the viscosities, while methodology described in this study, which eliminates the need
the CHARMM force field overestimates the experimental viscosi- for manually varying the L-J parameters to reproduce the experi-
ties. In general, as the degree of intermolecular interactions mental measurements.
increases, the viscosity of the system rises, and vice versa. At
higher viscosities, the molecules interact more strongly with one 3.2.2. Radial distribution functions
another and move more slowly, resulting in a lower diffusion coef- The GA-derived parameters are further evaluated by analysis of
ficient. Therefore, because the diffusion coefficient is inversely pro- the microstructures present in the system. The most obvious way
portional to the viscosity, this explains why the diffusion to study the local organization of any system is the analysis of
coefficient with the CHARMM force field is lower than for the other the radial distribution functions (RDFs). Although the RDFs
force fields (Fig. 6(a)). However, even though the viscosities between various atom pairs have been investigated in previous
obtained from the GA-derived force field underestimate the exper- work [29,35], these studies offer no concrete verifications of the
imental results, this difference is comparable to previously shape and magnitudes of the RDFs illustrated and are only com-
reported force fields. Thus, it can be stated that the model pro- pared with other CMD models for ACN. Hence, to verify the validity
posed in this work gives a better overall performance in predicting of our model’s ability to accurately represent the microstructural
physical properties than previously reported force fields, especially characteristics of ACN, we compare the CMD results with highly
those by Koverga et al. [29] and Kowsari et al. [35], where the L-J accurate FPMD and neutron scattering data [69]. Fig. 7 illustrates
parameters were derived manually. This is further elaborated on RDFs calculated from both CMD and FPMD, while Figs. S1 and S2
in the following sections. In conclusion, the ability of the GA- are the experimental neutron scattering results obtained from
derived parameters in predicting macroscopic experimental prop- the work by Humphreys et al. [69].
erties indicates the usefulness of the GA-based optimization From Fig. 7, it is evident that the CMD can reproduce the FPMD
reported herein. and experimental results to quite a reasonable extent. The N–N
To quantitatively compare the performances of the previously RDF in Fig. 7(a) obtained from CMD shows a double peak at
reported force fields and the GA-derived parameters reported in approximately 4.25 and 5.45 Å. This directly coincides with the
this study, we employ the use of 2 metrics: the absolute error FPMD results where the first and second peaks are located at about
Fig. 7. Site–site RDFs (at 298 K) from CMD (GA-derived parameters) and FPMD for the (a) N–N, (b) N–C1, (c) N–C2, (d) C1–C1, and (e) N–H atom pairs in ACN.
10
4.25 and 5.55 Å, respectively. These results are similar to experi- Table 3
mental measurements seen in Fig. S1 as well, where the first N– Calculated reorientational correlation times obtained from the GA force field,
previously developed force fields, and experimental measurements.
N peak is located at 4.1 Å. However, a point to be noted is the
absence of a second N–N peak in Fig. S1. As the second peak in Model s1 (ps) s2 (ps)
Fig. 7(a) is quite small, it is likely that it is not observed in the Experiment 3.28a, 3.68a 1.02a, 1.30b
experimental diffraction data. The broad C1–C1 peak in Fig. 7(b) Grabuleda et al. [64] 3.35 1.12
at 4.35 Å (for CMD) overlaps with the first N–N peak, similar to Vanommeslaeghe et al. [65] 4.10 1.39
Koveraga et al. [29] 2.74 0.90
what can be seen from the FPMD (at 4.45 Å) and experimental Orhan et al. [66] 2.84 0.93
measurements (at 4.5 Å) (Fig. S2). RDFs obtained from CMD indi- Kowsari et al. [35] 2.77 0.88
cate that the probability of finding a methyl group near the nitro- From GA 2.83 0.92
gen is higher in comparison to its counterparts. This is apparent in a
Ref.[72]
Fig. 7(c), where the peak for the N–C2 RDF is the highest among all b
Ref.[70]
the RDFs in Fig. 7 for CMD (at 3.45 Å), FPMD (at 3.45 Å), and Fig. S1
(at approximately 3.4 Å). This has been linked to the strong inter-
actions between the positively charged hydrogen and the nega- surements studies [70,72,73]. Table 3 summarizes the reorienta-
tively charged nitrogen [35]. Previous MD studies on ACN have tional correlation times obtained from the GA force field,
also pointed towards the presence of weak hydrogen bonds previously developed force fields, and experimental measurements.
between the nitrogen and the methyl hydrogen atoms [29,35]. In Fig. S3 illustrates the evolution of the C(t) function calculated using
previous studies, it has been reported that the location of the first the GA force field.
N–H peak was reported shortest among all pair interactions in ACN From Table 3, it can be observed that the reorientational corre-
molecules. This is true for the MD simulations conducted in this lation times calculated using the GA-derived parameters underes-
study as well where the first peak for the N–H RDF in Fig. 7(e) is timates the experimental values. Nevertheless, the performance of
at 2.65 Å and 2.75 Å for CMD and FPMD, respectively. The N–H the GA force field is comparable to the previously developed
RDF also indicates the presence of a secondary peak in both the parameters by Koveraga et al. [29], Orhan et al. [66], and Kowsari
CMD and FPMD, at 4.05 Å. The calculated N–H RDF peaks are sim- et al. [35] On the other hand, the CHARMM force field overesti-
ilar to the neutron scattering measurements, where the first and mates the experimental data. Surprisingly, the reorientational
second N–H peaks are located at approximately 2.7 and 3.8 Å, times calculated using GAFF provide the best results. However, as
respectively (Fig. S1). Thus, it is clear from the results herein that mentioned earlier, because GAFF parameters are unable to reason-
our GA force field can reasonably predict the intermolecular inter- ably replicate experimental physical properties, its use may be lim-
actions in ACN, as confirmed by RDFs obtained by FPMD and exper- ited in conducting CMD studies.
imental neutron scattering data. Hence, the use of GA to optimize the parameters is an attractive
The use of highly accurate FPMD simulations and experimental strategy for future studies in force field development. From the
measurements to validate the intermolecular structures in ACN has results reported in this study, it is encouraging that the GA force
previously not been investigated and as such we believe the field for ACN achieved excellent overall performances. Although
methodology adopted in this study provides a more reliable way certain physical properties, such as the density or the diffusion
to verify the structural properties from CMD simulations. From coefficient, are better predicted by other force fields, the GA-
the RDFs in Fig. 7, it is apparent that the force field optimized using derived parameters show better performance for different physical
GA is not only able to reproduce the physical properties but also properties both at room temperature and in a wide temperature
the microstructures present in liquid ACN. range. This emphasizes the significance of the optimization
methodology reported herein, where GAs can be used to optimize
3.2.3. Reorientational dynamics CMD force fields to provide useful molecular insights. The use of
The diffusion coefficients and viscosities investigated in previ- GA eliminates the need for manually adjusting the force field
ous sections are properties that measure the translational dynam- parameters as it does not require any human intervention. While
ics of the system. However, the rotational or reorientational the force field developed in this study (for ACN) might not be trans-
dynamics of a molecule can also be considered to be important. ferrable to different molecules, the use of GA offers an attractive
Similar to the translational motion, the rotation of the molecule alternative to manually optimizing the vdW parameters. In princi-
is directly affected by its local environment. Therefore, the rota- ple, this approach can be used to further optimize already available
tional or reorientational kinetics for a certain molecule is affected parameters, such as those present in GAFF or CHARMM, or develop
by the molecules in its first coordination shell. Such an analysis new parameters for proprietary systems.
is useful in investigating the changes that take place in a mole-
cule’s coordination environment [70,71].
To test the ability of the GA-derived parameters in predicting 4. Conclusion
the reorientational dynamics of the system, the reorientational
correlation times s1 and s2 were obtained by fitting an exponential CMD simulations have attracted significant attention in recent
decay function in Eq. (19) to the curve obtained from the reorien- decades as they are incredibly useful in providing insights into a
tational time correlation function in Eq. (18) [72]. wide variety of systems at the atomic level. However, the accuracy
! ! of CMD simulations is highly dependent upon the force field, i.e.,
CðtÞ ¼ hP l e ðt Þ e ð0Þ i ð18Þ the functional form for the potential energy, and the parameters
used to define the intra- and intermolecular interactions in the sys-

t tem. To properly define any system, force field parameters that can
CðtÞ ¼ A exp - ð19Þ accurately replicate experimental results are essential. Thus, the
sl
success of the CMD calculations is determined by their ability to
!
where e ðt Þ represents the unit vector along the axis of the ACN accurately reproduce experimental properties. Although general
molecule, Pl is the lth-order Legendre polynomial. For the purposes force fields such as AMBER, OPLS, CHARMM, and GROMOS are
of this study, the CN bond was chosen to define the unit vector. available, their use is limited as their performance varies from
The calculated value can be directly compared to NMR and IR mea- one molecule to another. Because the intermolecular vdW interac-
11
tions are known to heavily influence the physical properties, previ- Appendix A. Supplementary material
ous work has involved intensive parameterization of the vdW
parameters in order to reproduce experimental physical proper- Supplementary data to this article can be found online at
ties. Such methods have little to no support for simultaneous fit- https://doi.org/10.1016/j.molliq.2021.116579.
ting of the parameters, resulting in a large number of
calculations and neglect coupling effects since each parameter is
fitted individually rather than a common set of calculations where References
all parameters are derived simultaneously. This is particularly the
case for the vdW parameters, which are tightly coupled with one [1] A.Y.L. Sim, P. Minary, M. Levitt, Modeling nucleic acids, Curr. Opin. Struct. Biol.
22 (2012) 273–278, https://doi.org/10.1016/j.sbi.2012.03.012.
another. Because the hand-tuning approach involves fixing a cer-
[2] M. Karplus, J.A. McCammon, Molecular Dynamics Simulations of Biomolecules,
tain parameter and adjusting the remaining parameters, the num- Nat. Struct. Mol. Biol. 9 (2002) 646–652. https://doi.org/https://doi.org/
ber of required calculations becomes increasingly expensive as the 10.1038/nsb0902-646.
[3] F. Vitalini, F. Noé, B.G. Keller, Molecular dynamics simulations data of the
number of parameters increases.
twenty encoded amino acids in different force fields, Data Br. 7 (2016) 582–
Hence, we report a GA-based optimization scheme for the 590, https://doi.org/10.1016/j.dib.2016.02.086.
parameterization of intermolecular vdW terms and use ACN, a [4] J. Vatamanu, Z. Hu, D. Bedrov, C. Perez, Y. Gogotsi, Increasing energy storage in
common organic solvent, to assess the GA’s performance. By electrochemical capacitors with ionic liquid electrolytes and nanostructured
carbon electrodes, J. Phys. Chem. Lett. 4 (2013) 2829–2837, https://doi.org/
including both the density and the diffusion coefficient into the fit- 10.1021/jz401472c.
ness function, we were able to optimize the vdW terms for the L-J [5] R. Jorn, R. Kumar, D.P. Abraham, G.A. Voth, Atomistic modeling of the
potential to reproduce the experimental physical properties over a electrode-electrolyte interface in Li-ion energy storage systems: Electrolyte
structuring, J. Phys. Chem. C. 117 (2013) 3747–3761, https://doi.org/10.1021/
wide temperature range. The calculated density, diffusion coeffi- jp3102282.
cients, and viscosities at different temperatures with the GA- [6] P.Y. Yang, S.P. Ju, H.S. Hsieh, J. Sen Lin, J.Y. Hsieh, Electrolytic molecule in-pore
derived parameters provided better overall results in comparison structure and capacitance of supercapacitors with nanoporous carbon
electrodes: A coarse-grained molecular dynamics study, Comput. Mater. Sci.
with previously developed force fields. The use of FPMD simula- 166 (2019) 293–302, https://doi.org/10.1016/j.commatsci.2019.05.010.
tions and experimental neutron scattering results proved incredi- [7] M. Takeuchi, Y. Kameda, Y. Umebayashi, S. Ogawa, T. Sonoda, Ishiguro, M.
bly useful in verifying the important interactions in ACN. In Fujita, M. Sano, Ion-ion interactions of LiPF6 and LiBF4 in propylene carbonate
solutions, J. Mol. Liq. 148 (2009) 99–108, https://doi.org/10.1016/
addition to the translational dynamics, our GA-derived parameters j.molliq.2009.07.003.
were also able to reasonably predict the reorientational dynamics [8] S. Lee, S.S. Park, Dielectric properties of organic solvents from non-polarizable
as well. While previously developed L-J parameters involved man- molecular dynamics simulation with electronic continuum model and density
functional theory, J. Phys. Chem. B. 115 (2011) 12571–12576, https://doi.org/
ually adjusting each parameter and ignore the coupling effects
10.1021/jp207658m.
between them, the GA reported herein needs no such physical [9] J.K. Shah, J.F. Brennecke, E.J. Maginn, Thermodynamic properties of the ionic
intervention and can fit all the L-J terms simultaneously. Once liquid 1-n-butyl-3-methylimidazolium hexafluorophosphate from Monte
run, the GA will automatically determine the optimum parameter Carlo simulations, Green Chem. 4 (2002) 112–118, https://doi.org/10.1039/
b110725a.
set based on the properties included in the fitness function. [10] C. Oostenbrink, T.A. Soares, N.F.A. Van Der Vegt, W.F. Van Gunsteren,
Future studies will extend the methodology presented in this Validation of the 53A6 GROMOS force field, Eur. Biophys. J. 34 (2005) 273–
work to other molecular systems and to more complex potential 284, https://doi.org/10.1007/s00249-004-0448-6.
[11] D.S. Cerutti, W.C. Swope, J.E. Rice, D.A. Case, Ff14ipq: A self-consistent force
energy functions with a greater number of optimizable terms, field for condensed-phase simulations of proteins, J. Chem. Theory Comput. 10
opening up new avenues in the development of force fields that (2014) 4515–4534, https://doi.org/10.1021/ct500643c.
can accurately describe the physics of a system and provide useful [12] C.I. Bayly, K.M. Merz, D.M. Ferguson, W.D. Cornell, T. Fox, J.W. Caldwell, P.A.
Kollman, P. Cieplak, I.R. Gould, D.C. Spellmeyer, A Second Generation Force
molecular insights. Field for the Simulation of Proteins, Nucleic Acids, and Organic Molecules, J.
Am. Chem. Soc. 117 (1995) 5179–5197, https://doi.org/10.1021/ja00124a002.
[13] G.A. Kaminski, R.A. Friesner, J. Tirado-Rives, W.L. Jorgensen, Evaluation and
reparametrization of the OPLS-AA force field for proteins via comparison with
CRediT authorship contribution statement accurate quantum chemical calculations on peptides, J. Phys. Chem. B. 105
(2001) 6474–6487, https://doi.org/10.1021/jp003919d.
[14] W.L. Jorgensen, J. Tirado-Rives, The OPLS Potential Functions for Proteins.
Abdullah Bin Faheem: Investigation, Writing - original draft, Energy Minimizations for Crystals of Cyclic Peptides and Crambin, J. Am.
Formal analysis, Software. Jong-Yun Kim: Writing - review & edit- Chem. Soc. 110 (1988) 1657–1666, https://doi.org/10.1021/ja00214a001.
ing. Sang-Eun Bae: Writing - review & editing. Kyung-Koo Lee: [15] A.D. Mackerell, M. Feig, C.L. Brooks, Extending the treatment of backbone
energetics in protein force fields: Limitations of gas-phase quantum
Writing - review & editing, Supervision. mechanics in reproducing protein conformational distributions in molecular
dynamics simulation, J. Comput. Chem. 25 (2004) 1400–1415, https://doi.org/
10.1002/jcc.20065.
[16] A.D. MacKerell, D. Bashford, M. Bellott, R.L. Dunbrack, J.D. Evanseck, M.J. Field,
Declaration of Competing Interest S. Fischer, J. Gao, H. Guo, S. Ha, D. Joseph-McCarthy, L. Kuchnir, K. Kuczera, F.T.
K. Lau, C. Mattos, S. Michnick, T. Ngo, D.T. Nguyen, B. Prodhom, W.E. Reiher, B.
Roux, M. Schlenkrich, J.C. Smith, R. Stote, J. Straub, M. Watanabe, J.
The authors declare that they have no known competing finan- Wiórkiewicz-Kuczera, D. Yin, M. Karplus, All-atom empirical potential for
cial interests or personal relationships that could have appeared molecular modeling and dynamics studies of proteins, J. Phys. Chem. B. 102
to influence the work reported in this paper. (1998) 3586–3616, https://doi.org/10.1021/jp973084f.
[17] J. Wang, P.A. Kollman, Automatic parameterization of force field by systematic
search and genetic algorithms, J. Comput. Chem. 22 (2001) 1219–1228,
https://doi.org/10.1002/jcc.1079.
[18] R.M. Betz, R.C. Walker, Paramfit: Automated optimization of force field
Acknowledgement parameters for molecular dynamics simulations, J. Comput. Chem. 36 (2015)
79–87, https://doi.org/10.1002/jcc.23775.
This research was supported by the mid- and long-term nuclear [19] T. Li, P.B. Balbuena, Theoretical studies of lithium perchlorate in ethylene
carbonate, propylene carbonate, and their mixtures, J. Electrochem. Soc. 146
research and development program through the National Research
(1999) 3613–3622, https://doi.org/10.1149/1.1392523.
Foundation of Korea (NRF-2017M2A8A5014716), funded by the [20] L.B. Silva, L.C.G. Freitas, Structural and thermodynamic properties of liquid
Korean Ministry of Science and ICT and the National Research ethylene carbonate and propylene carbonate by Monte Carlo Simulations, J.
Foundation of Korea (NRF) grant funded by the Korea government Mol. Struct. THEOCHEM. 806 (2007) 23–34, https://doi.org/10.1016/j.
theochem.2006.10.014.
(MSIT) (NRF-2019R1A4A102980111 and NRF- [21] J.N.C. Lopes, A.A.H. Padua, Molecular Force Field for Ionic Liquids III:
2020R1I1A3066503). Imidazolium, Pyridinium, and Phosphonium Cations; Chloride, Bromide, and
12
Dicyanamide Anions, J. Phys. Chem. B. 110 (2006) 19586–19592, https://doi. Sagui, V. Babin, T. Luchko, S. Gusarov, A. Kovalenko, and P.A. Kollman (2012),
org/10.1021/jp063901o. AMBER 12, University of California, San Francisco.
[22] J.N.C. Lopes, A.A.H. Pádua, Molecular force field for ionic liquids composed of [43] F.Y. Dupradeau, A. Pigache, T. Zaffran, C. Savineau, R. Lelong, N. Grivel, D.
triflate or bistriflylimide anions, J. Phys. Chem. B. 108 (2004) 16893–16898, Lelong, W. Rosanski, P. Cieplak, The R.E.D. tools: Advances in RESP and ESP
https://doi.org/10.1021/jp0476545. charge derivation and force field library building, Phys. Chem. Chem. Phys. 12
[23] M.L.P. Price, D. Ostrovsky, W.L. Jorgensen, Gas-phase and liquid-state (2010) 7821–7839, https://doi.org/10.1039/c0cp00111b.
properties of esters, nitriles, and nitro compounds with the OPLS-AA force [44] A.M. Nikitin, A.P. Lyubartsev, New Six-site Acetonitrile Model for Simulations
field, J. Comput. Chem. 22 (2001) 1340–1352, https://doi.org/10.1002/ of Liquid Acetonitrile and its Aqueous Mixtures, J. Comput. Chem. 28 (2007)
jcc.1092. 2020–2026, https://doi.org/10.1002/jcc.20721.
[24] M.V. Ivanov, M.R. Talipov, Q.K. Timerghazin, Genetic algorithm optimization of [45] H.V.T. Nguyen, K. Kwak, K.K. Lee, 1,1-Dimethylpyrrolidinium tetrafluoroborate
point charges in force field development: Challenges and insights, J. Phys. as novel salt for high-voltage electric double-layer capacitors, Electrochim.
Chem. A. 119 (2015) 1422–1434, https://doi.org/10.1021/acs.jpca.5b00218. Acta. 299 (2019) 98–106, https://doi.org/10.1016/j.electacta.2018.12.155.
[25] J. Wang, P. Cieplak, J. Li, Q. Cai, M.J. Hsieh, R. Luo, Y. Duan, Development of [46] P. Alston Steiner, W. Gordy, Precision measurement of dipole moments and
polarizable models for molecular mechanical calculations. 4. van der waals other spectral constants of normal and deuterated methyl fluoride and methyl
parametrization, J. Phys. Chem. B. 116 (2012) 7088–7101, https://doi.org/ cyanide, J. Mol. Spectrosc. 21 (1966) 291–301, https://doi.org/10.1016/0022-
10.1021/jp3019759. 2852(66)90152-4.
[26] J. Wang, T. Hou, Application of molecular dynamics simulations in molecular [47] L. Angibaud, L. Briquet, P. Philipp, T. Wirtz, J. Kieffer, Parameter optimization in
property prediction II: Diffusion coefficient, J. Comput. Chem. 32 (2011) 3505– molecular dynamics simulations using a genetic algorithm, Nucl. Instruments
3519, https://doi.org/10.1002/jcc.21939. Methods Phys. Res. Sect. B Beam Interact. with Mater. Atoms. 269 (2011)
[27] J. Wang, T. Hou, Application of molecular dynamics simulations in molecular 1559–1563, https://doi.org/10.1016/j.nimb.2010.11.024.
property prediction. 1. Density and heat of vaporization, J. Chem. Theory [48] J.M.C. Marques, F.V. Prudente, F.B. Pereira, M.M. Almeida, A.M. Maniero, C.E.
Comput. 7 (2011) 2151–2165, https://doi.org/10.1021/ct200142z. Fellows, A new genetic algorithm to be used in the direct fit of potential energy
[28] C. Park, M. Kanduč, R. Chudoba, A. Ronneburg, S. Risse, M. Ballauff, J. Dzubiella, curves to ab initio and spectroscopic data, J. Phys. B At. Mol. Opt. Phys. 41
Molecular simulations of electrolyte structure and dynamics in lithium–sulfur (2008), https://doi.org/10.1088/0953-4075/41/8/085103.
battery solvents, J. Power Sources. 373 (2018) 70–78, https://doi.org/10.1016/ [49] L. Martínez, R. Andrade, E.G. Birgin, J.M. Martínez, PACKMOL: A package for
j.jpowsour.2017.10.081. building initial configurations for molecular dynamics simulations, J. Comput.
[29] V.A. Koverga, O.M. Korsun, O.N. Kalugin, B.A. Marekha, A. Idrissi, A new Chem. 30 (2009) 2157–2164, https://doi.org/10.1002/jcc.21224.
potential model for acetonitrile: Insight into the local structure organization, J. [50] A. Mondal, J.M. Young, T.A. Barckholtz, G. Kiss, L. Koziol, A.Z. Panagiotopoulos,
Mol. Liq. 233 (2017) 251–261, https://doi.org/10.1016/j.molliq.2017.03.025. Genetic Algorithm Driven Force Field Parameterization for Molten Alkali-
[30] P. Kumar, S. Yashonath, Ionic conductivity in aqueous electrolyte solutions: Metal Carbonate and Hydroxide Salts, J. Chem. Theory Comput. (2020), https://
Insights from computer simulations, J. Mol. Liq. 277 (2019) 506–515, https:// doi.org/10.1021/acs.jctc.0c00285.
doi.org/10.1016/j.molliq.2018.12.090. [51] K. Hickey, W.E. Waghorne, Viscosities and volumes of dilute solutions of
[31] V.S. Sambarivarao, O. Acevedo, Development of OPLS-AA Force Field formamide in water + acetonitrile and for formamide and N, N-
Parameters for 68 Unique Ionic Liquids, J. Chem. Theory Comput. 5 (2009) dimethylformamide in methanol + acetonitrile mixed solvents: Viscosity B-
1038–1050, https://doi.org/10.1021/ct900009a. Coefficients, activation free energies for viscous flow, and partial molar, J.
[32] K. Xu, X. Ji, C. Chen, H. Wan, L. Miao, J. Jiang, Electrochemical double layer near Chem. Eng. Data. 46 (2001) 851–857, https://doi.org/10.1021/je0003647.
polar reduced graphene oxide electrode: Insights from molecular dynamic [52] E. Hawlicka, R. Grabowski, Solvation of Ions in Acetonitrile-Methanol Solutions
study, Electrochim. Acta. 166 (2015) 142–149, https://doi.org/10.1016/ of Sodium Iodide, Berichte Der Bunsengesellschaft Für Phys. Chemie. 94 (1990)
j.electacta.2015.03.101. 486–489, https://doi.org/10.1002/bbpc.19900940413.
[33] R. Wang, S. Bi, V. Presser, G. Feng, Systematic comparison of force fields for [53] P. Sharma, A. Wadhwa, K. Komal, Analysis of Selection Schemes for Solving an
molecular dynamic simulation of Au(111)/Ionic liquid interfaces, Fluid Phase Optimization Problem in Genetic Algorithm, Int. J. Comput. Appl. 93 (2014) 1–
Equilib. 463 (2018) 106–113, https://doi.org/10.1016/j.fluid.2018.01.024. 3, https://doi.org/10.5120/16256-5714.
[34] D. Van Der Spoel, P.J. Van Maaren, H.J.C. Berendsen, A systematic study of [54] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist multiobjective
water models for molecular simulation : Derivation of water models genetic algorithm: NSGA-II, IEEE Trans. Evol. Comput. 6 (2002) 182–197,
optimized for use with a reaction field A systematic study of water models https://doi.org/10.1109/4235.996017.
for molecular simulation : Derivation of water models optimized for use with a [55] A. Konak, D.W. Coit, A.E. Smith, Multi-objective optimization using genetic
reaction fie, 10220 (2011) 10220–10230. https://doi.org/10.1063/1.476482. algorithms: A tutorial, Reliab. Eng. Syst. Saf. 91 (2006) 992–1007, https://doi.
[35] M.H. Kowsari, L. Tohidifar, Systematic evaluation and refinement of existing org/10.1016/j.ress.2005.11.018.
all-atom force fields for the simulation of liquid acetonitrile, J. Comput. Chem. [56] S. Plimpton, Fast parallel algorithms for short-range molecular dynamics, J.
39 (2018) 1843–1853, https://doi.org/10.1002/jcc.25337. Comput. Phys. 117 (1995) 1–19, https://doi.org/10.1006/jcph.1995.1039.
[36] J. Wang, P. Cieplak, J. Li, T. Hou, R. Luo, Y. Duan, Development of polarizable [57] S. Nosé, A unified formulation of the constant temperature molecular
models for molecular mechanical calculations I: Parameterization of atomic dynamics methods, J. Chem. Phys. 81 (1984) 511–519, https://doi.org/
polarizability, J. Phys. Chem. B. 115 (2011) 3091–3099, https://doi.org/ 10.1063/1.447334.
10.1021/jp112133g. [58] W.G. Hoover, Canonical dynamics: Equilibrium phase-space distributions,
[37] P. Pahari, S. Chaturvedi, Determination of best-fit potential parameters for a Phys. Rev. A. 31 (1985) 1695, https://doi.org/10.1103/PhysRevA.31.1695.
reactive force field using a genetic algorithm, J. Mol. Model. 18 (2012) 1049– [59] J. Vandevondele, M. Krack, F. Mohamed, M. Parrinello, T. Chassaing, J. Hutter,
1061, https://doi.org/10.1007/s00894-011-1124-2. Quickstep: Fast and accurate density functional calculations using a mixed
[38] F. Leonarski, F. Trovato, V. Tozzini, A. Leś, J. Trylska, Evolutionary algorithm in Gaussian and plane waves approach, Comput. Phys. Commun. 167 (2005)
the optimization of a coarse-grained force field, J. Chem. Theory Comput. 9 103–128, https://doi.org/10.1016/j.cpc.2004.12.014.
(2013) 4874–4889, https://doi.org/10.1021/ct4005036. [60] S. Grimme, J. Antony, S. Ehrlich, H. Krieg, A consistent and accurate ab initio
[39] B. Courcot, A.J. Bridgeman, Optimization of a Molecular Mechanics Force Field parametrization of density functional dispersion correction (DFT-D) for the 94
for Polyoxometalates Based on a Genetic Algorithm, J. Comput. Chem. 32 elements H-Pu, J. Chem. Phys. 132 (2010), https://doi.org/10.1063/1.3382344.
(2011) 240–247, https://doi.org/10.1002/jcc.21610. [61] E.M. Kirova, G.E. Norman, Viscosity calculations at molecular dynamics
[40] L. Chen, C.A. Morrison, T. Düren, Improving predictions of gas adsorption in simulations, J. Phys. Conf. Ser. 653 (2015), https://doi.org/10.1088/1742-
metal-organic frameworks with coordinatively unsaturated metal sites: Model 6596/653/1/012106.
potentials, ab initio parameterization, and gcmc simulations, J. Phys. Chem. C. [62] J. Ma, Z. Zhang, Y. Xiang, F. Cao, H. Sun, On the prediction of transport
116 (2012) 18899–18909, https://doi.org/10.1021/jp3062527. properties of ionic liquid using 1-n-butylmethylpyridinium tetrafluoroborate
[41] Gaussian 09, Revision D.01, M. J. Frisch, G. W. Trucks, H. B. Schlegel, G. E. as an example, Mol. Simul. 43 (2017) 1502–1512, https://doi.org/10.1080/
Scuseria, M. A. Robb, J. R. Cheeseman, G. Scalmani, V. Barone, B. Mennucci, G. A. 08927022.2017.1321760.
Petersson, H. Nakatsuji, M. Caricato, X. Li, H. P. Hratchian, A. F. Izmaylov, J. [63] P.E. Smith, W.F. van Gunsteren, The viscosity of SPC and SPC/E water at 277
Bloino, G. Zheng, J. L. Sonnenberg, M. Hada, M. Ehara, K. Toyota, R. Fukuda, J. and 300 K, Chem. Phys. Lett. 215 (1993) 315–318, https://doi.org/10.1016/
Hasegawa, M. Ishida, T. Nakajima, Y. Honda, O. Kitao, H. Nakai, T. Vreven, J. A. 0009-2614(93)85720-9.
Montgomery, Jr., J. E. Peralta, F. Ogliaro, M. Bearpark, J. J. Heyd, E. Brothers, K. [64] X. Grabuleda, C. Jaime, P.A. Kollman, Molecular Dynamics Simulation Studies
N. Kudin, V. N. Staroverov, T. Keith, R. Kobayashi, J. Normand, K. Raghavachari, of Liquid Acetonitrile: New Six-Site Model, J. Comput. Chem. 21 (2000) 901–
A. Rendell, J. C. Burant, S. S. Iyengar, J. Tomasi, M. Cossi, N. Rega, J. M. Millam, 908, https://doi.org/10.1002/1096-987X(20000730)21:10<901::AID-JCC7>3.0.
M. Klene, J. E. Knox, J. B. Cross, V. Bakken, C. Adamo, J. Jaramillo, R. Gomperts, R. CO;2-F.
E. Stratmann, O. Yazyev, A. J. Austin, R. Cammi, C. Pomelli, J. W. Ochterski, R. L. [65] K. Vanommeslaeghe, E. Hatcher, C. Acharya, S. Kundu, S. Zhong, J. Shim, E.
Martin, K. Morokuma, V. G. Zakrzewski, G. A. Voth, P. Salvador, J. J. Darian, O. Guvench, P. Lopes, I. Vorobyov, A.D. MacKerell, CHARMM General
Dannenberg, S. Dapprich, A. D. Daniels, O. Farkas, J. B. Foresman, J. V. Ortiz, J. Force Field: A Force Field for Drug-Like Molecules Compatible with the
Cioslowski, and D. J. Fox, Gaussian, Inc., Wallingford CT, 2013. CHARMM All-Atom Additive Biological Force Fields, J. Comput. Chem. 31
[42] D.A. Case, T.A. Darden, T.E. Cheatham, III, C.L. Simmerling, J. Wang, R.E. Duke, R. (2010) 671–690, https://doi.org/10.1002/jcc.21367.
Luo, R.C. Walker, W. Zhang, K.M. Merz, B. Roberts, S. Hayik, A. Roitberg, G. [66] M. Orhan, Dielectric and transport properties of acetonitrile at varying
Seabra, J. Swails, A.W. Götz, I. Kolossváry, K.F.Wong, F. Paesani, J. Vanicek, R.M. temperatures: A molecular dynamics study, Bull. Korean Chem. Soc. 35
Wolf, J. Liu, X. Wu, S.R. Brozell, T. Steinbrecher, H. Gohlke, Q. Cai, X. Ye, J. Wang, (2014) 1469–1478, https://doi.org/10.5012/bkcs.2014.35.5.1469.
M.-J. Hsieh, G. Cui, D.R. Roe, D.H. Mathews, M.G. Seetin, R. Salomon-Ferrer, C.
13
[67] R.L. Hurle, L.A. Woolf, Self-diffusion in liquid acetonitrile under pressure, J. [70] K. Yuan, H. Bian, Y. Shen, B. Jiang, J. Li, Y. Zhang, H. Chen, J. Zheng, Coordination
Chem. Soc. Faraday Trans. 1 Phys. Chem. Condens. Phases. 78 (1982) 2233– number of Li+ in nonaqueous electrolyte solutions determined by molecular
2238. https://doi.org/10.1039/F19827802233. rotational measurements, J. Phys. Chem. B. 118 (2014) 3689–3695, https://doi.
[68] J.H. Dymond, M.A. Awan, N.F. Glen, J.D. Isdale, Transport properties of org/10.1021/jp500877u.
nonelectrolyte mixtures. IX. Viscosity coefficients for acetonitrile and for [71] J. Boisson, G. Stirnemann, D. Laage, J.T. Hynes, Water reorientation dynamics in
three mixtures of toluene+acetonitrile from 25 to 100°c at pressures up to 500 the first hydration shells of F and I, Phys. Chem. Chem. Phys. 13 (2011)
MPa, Int. J. Thermophys. 12 (1991) 433–447, https://doi.org/10.1007/ 19895–19901, https://doi.org/10.1039/c1cp21834d.
BF00502360. [72] P.J. Gee, W.F. Van Gunsteren, Acetonitrile revisited: A molecular dynamics
[69] E.K. Humphreys, P.K. Allan, R.J.L. Welbourn, T.G.A. Youngs, A.K. Soper, C.P. study of the liquid phase, Mol. Phys. 104 (2006) 477–483, https://doi.org/
Grey, S.M. Clarke, A Neutron Diffraction Study of the Electrochemical Double 10.1080/00268970500473450.
Layer Capacitor Electrolyte Tetrapropylammonium Bromide in Acetonitrile, J. [73] A. Sugitani, S. Ikawa, S. Konaka, Effect of temperature on the infrared band
Phys. Chem. B. 119 (2015) 15320–15333, https://doi.org/10.1021/acs. shapes and reorientational and vibrational relaxation of liquid acetonitrile,
jpcb.5b08248. Chem. Phys. 142 (1990) 423–430, https://doi.org/10.1016/0301-0104(90)
80037-X.
14

Param Ga

Uploaded by

Copyright:

Available Formats

Param Ga

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Param Ga

Uploaded by

Copyright:

Available Formats

Journal of Molecular Liquids 337 (2021) 116579

Contents lists available at ScienceDirect

Journal of Molecular Liquids

Efficient parameterization of intermolecular force fields for molecular

1. Introduction [15,16], the intramolecular parameters include the equilibrium

focused on developing intramolecular parameters. For example, X Vn

Fig. 2. Flowchart of the genetic algorithm used herein.

the results in Section 3 indicate a favorable GA performance and

2.3. Classical molecular dynamics

The starting configuration for performing CMD simulations

To verify the ability of the GA algorithm in optimizing the vdW

2.5.3. Heat of vaporization

Model q D Hvap ƞ MAPE

parameters. The underestimated densities by GAFF highlight the

You might also like