Traditionally, training of the critical Lennard-Jones (LJ) parameters of fixed-charge force fields consists of gradient-based optimization of an objective function based on experimental physical property data.  This method has produced improved parameter sets over the years, but requires users to perform equilibrium simulations at each parameter set considered, leading to significant computational expense.  This expense makes thorough optimization or model comparison extremely difficult, hampering efforts to understand parameter dependencies and produce better models.<br/>One potential alternative to repeated equilibrium simulations is the use of surrogate models, which learn simulation observables as a function of force field parameters.  Configuration-based surrogates built using reweighting, or regression-based surrogate models built from simulation outputs have been used successfully to approximate observable surfaces.  We leverage the computational acceleration of regression-based surrogates to implement Bayesian methods for parameter sampling and model comparison.<br/>We use pre-existing surrogate models on the 2-center Lennard Jones plus Quadrupole (2CLJQ) model for simple fluids, to implement Bayesian inference and model selection over the model’s parameters.  Through this strategy, we are able to evaluate whether the model’s quadrupole parameter is useful in reproducing temperature-dependent density, saturation pressure, and surface tension data for simple molecules. Additionally, this process produces parameter probability distributions for each compound, valuable information about the parameter uncertainty and sensitivity. This work demonstrates the utility of Bayesian inference as a tool for model selection and parameterization.<br/>To extend Bayesian inference to more general parameterization problems, we build Gaussian Process (GP) surrogates for physical properties over parameter ranges of interest, based on thousands of individual simulations.  Based on these surrogates, we construct Bayesian posteriors that represent the relative fitness of many parameter sets.  We demonstrate this by producing a full parameter distribution of a set of alcohols, based on density and enthalpy of vaporization measurements.  These parameter posterior distributions identify the best sets of parameters, and provide unique insight into parameter correlation and sensitivity.

Surrogate-enabled Bayesian parameterization of non-bonded force field parameters

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.