Force Field Development for Molecular Simulations with Structure Optimized Potential Refinement
Abstract
One of the fundamental applications of molecular simulations is to predict the atomic structure and self-assembly of fluid-state materials. However, recent advances in experimental neutron and X-ray diffraction are revealing that many widely used force fields do not predict atomic structures that are consistent with experimental data. Directly reconstructing interatomic potentials from structure, the so-called inverse problem of statistical mechanics, is a possible method to improve and develop force fields that accurately model experimentally observed structures. While inverse techniques such as iterative Boltzmann inversion and relative entropy minimization have been successfully applied to coarse-grain model systems, there are no existing techniques that provide satisfactory solutions for real physical systems. In this study, we present a machine learning assisted structure refinement technique, called structure optimized potential refinement, that can reconstruct transferable force fields from a single neutron diffraction measurement in monatomic fluids. Applying this technique to the noble gas series (Ne, Ar, Kr, and Xe), we recover transferable force field parameters that provide excellent agreement to experimental structure and reproduce macroscopic thermodynamic properties. This result is the first demonstration that experimental scattering data contains sufficient information to predict underlying force parameters at the sub-angstrom length scale. Applications of structure optimized potential refinement include force field optimization for structure-based applications, modeling fluids in extreme environments, and predicting quantum many-body effects in fluid systems.
