Machine learning accelerated methods to predict interatomic forces from experimental structure measurements
Abstract
The prediction of interatomic forces from experimental scattering data is a historic inverse problem in statistical mechanics from both a theoretical and computational perspective. However, one of the biggest challenges for data-driven solutions to inverse problems is the computational expense of evaluating expensive models of fluid structures. Here a set of discrete Gaussian process surrogate models is proposed to accelerate the estimation of structure factors, enabling force field optimization to experimental scattering data with probabilistic machine learning techniques. Applying this technique to a (n-6) Mie fluid, we find that details of the interatomic force can be determined accurately within the uncertainty of existing experimental scattering instruments, challenging the widely held view that structure is insensitive to interatomic forces. We conclude that machine learning accelerated methods for structure factor characterization and uncertainty quantification are an attractive tool to study self-assembly and structural properties of materials and fundamental interatomic interactions.
