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

Force field parameters for molecular simulations of thermophysical properties are typically optimized iteratively against experimental data. However, this requires extensive computational effort due to multiple parameter evaluations. Although machine learning has been successfully applied in bottom-up derivation of interatomic potentials from quantum chemical data, its use in top-down classical force field parametrization is limited. Moreover, existing machine learning frameworks struggle to build efficient, localized surrogate models while quantifying uncertainties. Here, we address these challenges by presenting a 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 of the refrigerant trans-1,2-dichloroethene using experimental vapor–liquid equilibria data, and achieve convergence with a low number of training points. Beyond reproducing VLE behavior, we employ the calibrated parameters in molecular dynamics studies of liquid density and viscosity, thereby assessing the impact of parameter uncertainties on the simulation results.