Sequential Bayesian Force Field Calibration of Lennard-Jones Parameters with Experimental Data

Intermolecular parameters in force fields for molecular simulation studies of thermophysical properties are typically derived through an iterative fitting process to experimental data. However, this approach suffers from extremely high computational cost because of the necessity to perform molecular simulations for a multitude of parameter values. While machine learning approaches have emerged in recent years in bottom-up derivation of interatomic potentials from quantum chemical simulations, their application in top-down methods for parameterizing classical force fields is much less explored. Moreover, the few available machine learning approaches for top-down parametrizations are not yet capable of constructing efficient and localized surrogate models while simultaneously quantifying uncertainties. Here, we overcome these issues by presenting a new sequential design strategy, based on Gaussian Process surrogate models, directly tailored for a Bayesian calibration task. We employ our approach to calibrate force field parameters for the refrigerant trans-1,2-dichloroethene (R-1130(E)) with experimental vapor-liquid phase equilibria (VLE) data and achieve convergence with a low number of training points. In addition to utilizing the calibrated model for simulation studies on the VLE of R-1130(E), we also apply it for predictive molecular dynamics studies on its liquid density and viscosity, and thereby assess the impact of parameter uncertainties on the simulation results.