Histogram-Free Reweighting with Grand Canonical Monte Carlo: Post-simulation Optimization of Non-bonded Potentials for Phase Equilibria

Abstract

Histogram reweighting (HR) is a standard approach for converting grand canonical Monte Carlo (GCMC) simulation output into vapor–liquid coexistence properties (saturated liquid density, ρliq sat, saturated vapor density, ρvap sat, saturated vapor pressures, P vap sat, and enthalpy of vaporization, Δ H v). We demonstrate that a histogram-free reweighting approach, namely, the Multistate Bennett Acceptance Ratio (MBAR), is similar to the traditional HR method for computing ρliq sat, ρvap sat, P vap sat, and Δ H v. The primary advantage of MBAR is the ability to predict phase equilibria properties for an arbitrary force field parameter set that has not been simulated directly. Thus, MBAR can greatly reduce the number of GCMC simulations that are required to parameterize a force field with phase equilibria data. Four different applications of GCMC-MBAR are presented in this study. First, we validate that GCMC-MBAR and GCMC-HR yield statistically indistinguishable results for ρliq sat, ρvap sat, P vap sat, and Δ H v in a limiting test case. Second, we utilize GCMC-MBAR to optimize an individualized (compound-specific) parameter (ψ) for 8 branched alkanes and 11 alkynes using the Mie Potentials for Phase Equilibria (MiPPE) force field. Third, we predict ρliq sat, ρvap sat, P vap sat, and Δ H v for force field j by simulating force field i, where i and j are common force fields from the literature. In addition, we provide guidelines for determining the reliability of GCMC-MBAR predicted values. Fourth, we develop and apply a post-simulation optimization scheme to obtain new MiPPE non-bonded parameters for cyclohexane (ϵCH2 , σCH2 , and λCH2 ).

Systematic Optimization of a Fragment-Based Force Field against Experimental Pure-Liquid Properties Considering Large Compound Families: Application to Saturated Haloalkanes

“…However, to date, the refinement against experimental data has mainly relied on manual procedures, rendering the task of force-field development extremely laborious and time-consuming. Recent attempts at automating the fitting against condensed-phase observables include in particular the POP scheme , and the ForceBalance scheme ,− ,− (see also refs − for related semiautomated approaches). However, until now, these have only been applied to parameter optimization for one single compound as opposed to a collection thereof. Automated protocols and workflows for the simulation setup, property calculation, and data analysis, which aim at facilitating and standardizing the calculation procedures.…”

Systematic optimization of a fragment-based force field against experimental pure-liquid properties considering large compound families: application to oxygen and nitrogen compounds

“…The main steps of the scheme are: This scheme borrows from earlier work on isomer enumeration [79][80][81][82] and topology construction, [83][84][85][86][87][88][89][90][91][92][93][94][95] as well as on automated single-compound force-field optimization approaches such as the POP scheme, 96,97 the ForceBalance scheme, 46,[98][99][100][101][102][103][104][105] and the related schemes. [106][107][108][109][110][111] One key feature of CombiFF is that once the time-consuming task of target-data selection/curation has been performed, the optimization of a force field is entirely automatic, given access to a sufficient number of processors, and only requires a few days of wall-clock computational time. The CombiFF also represents an ideal framework for assessing the impact of specific functional-form decisions on the accuracy of a force field at an optimal level of parametrization.…”

gmak: A Parameter-Space Mapping Strategy for Force-Field Calibration

“…Due to the complexity of FF development and to the different interests and backgrounds of FF developers, there is no single universally adopted parametrization workflow. For example, the search for optimal (single-objective) or efficient (multiobjective) parameter sets may be carried out manually by trial and error or by relying on well-established optimization algorithms, − such as hill-climbing, simplex minimization, ,, the Levenberg–Marquardt method, − the L-BFGS method, evolutionary algorithms, multiobjective evolutionary algorithms, , multiobjective Particle Swarm Optimization, and multiobjective Genetic Algorithms . These parameter-calibration workflows also frequently rely on training surrogate models ,− ,− , (SMs), such as in Gaussian process regressors (GPRs), multistate Bennett’s acceptance ratio − (MBAR), pair correlation function rescaling (PCFR), radial basis functions , (RBFs), and thermodynamic reweighting, ,, to replace the simulations involved in the estimation of the target properties, thus reducing the underlying computational cost.…”

“…For example, the search for optimal (single-objective) or efficient (multiobjective) parameter sets may be carried out manually by trial and error or by relying on well-established optimization algorithms, − such as hill-climbing, simplex minimization, ,, the Levenberg–Marquardt method, − the L-BFGS method, evolutionary algorithms, multiobjective evolutionary algorithms, , multiobjective Particle Swarm Optimization, and multiobjective Genetic Algorithms . These parameter-calibration workflows also frequently rely on training surrogate models ,− ,− , (SMs), such as in Gaussian process regressors (GPRs), multistate Bennett’s acceptance ratio − (MBAR), pair correlation function rescaling (PCFR), radial basis functions , (RBFs), and thermodynamic reweighting, ,, to replace the simulations involved in the estimation of the target properties, thus reducing the underlying computational cost. Other strategies with similar goals include the use of the statistical-mechanical fluctuation formula, ,− whereby the derivatives of the (mechanical) target properties with respect to the parameters are obtained from an analytical expression (which requires a single simulation per derivative estimate), as opposed to based on finite-difference approaches (which require multiple simulations per derivative estimate).…”

“…The fourth choice concerns the fitting of the parameter set against the chosen observables. This problem is very frequently − represented as the minimization of an objective function that combines (possibly with different weights attributed to each property) the errors of the simulated estimates of the target properties relative to the experimental data as a function of the parameter set, e.g., in the form of a weighted absolute error function. Less frequently, − a multiobjective perspective is employed in the calibration, where the errors of the target properties relative to the reference data are taken simultaneously as a vector function, instead of combined as a real function.…”