Learning Interatomic Forces from Experimental Measurements of Fluid Structure
Abstract
The prediction of interatomic forces from experimental X-ray and neutron scattering data is a historic inverse problem in statistical mechanics. Accurate and robust reconstructions of interatomic forces could be used to create novel force fields for molecular simulations, study many-body effects in statistical ensembles, and advance our understanding of atomic structure and self-assembly of materials. However, there are no existing statistical theories or computational techniques that provide adequate inverse problem solutions for a wide range of physical systems. Furthermore, data-driven methods such as probabilistic machine learning are burdened by the high computational cost of evaluating models for atomic structures. In this study, we developed a local Gaussian process surrogate model that vastly accelerates the estimation of structure data within a molecular simulation framework. We demonstrate that this method provides rapid and accurate evaluation of ensemble fluid structures that enables the application of probabilistic machine learning to optimize force fields from experimental scattering data with uncertainty quantification. These results suggest that machine learning accelerated interatomic force reconstruction from experimental diffraction data is now a viable tool for structural analysis and force field optimization.
