A Bayesian optimized structural force field for noble gases enabled by a radial distribution function surrogate model
Abstract
A Bayesian optimized (n-6) Mie force field for noble gases benchmarked to experimental neutron diffraction measurements of the radial distribution function is presented. The proposed computational approach constructs a Gaussian process surrogate model of molecular simulation derived radial distribution functions at randomly sampled parameter values. The surrogate model approach is found to accelerate Bayesian optimization to experimental radial distribution function data from ~1000 CPU years to < 1 CPU hour. Transferability of the Bayesian optimized (n-6) Mie force field is confirmed with grand Canonical Monte Carlo simulations of vapor-liquid equilibrium in the noble gases Ne, Ar, Kr, and Xe with competitive accuracy to independently optimized vapor-liquid equilibrium force fields. This study demonstrates that simulated structure accuracy at the sub-angstrom length-scale is achievable with machine learning potentials and further that structure-optimized potentials are transferable to the prediction of emergent thermodynamic behavior.
