Advancing Interatomic Force Prediction with Machine Learning: Accelerated Methods for Extracting Force Fields from Experimental Scattering Measurements
Abstract
The prediction of interatomic forces based on experimental scattering data represents a significant and longstanding inverse problem in statistical mechanics, both from a theoretical and computational standpoint. However, one of the primary hurdles faced by data-driven solutions to such inverse problems lies in the computational burden associated with evaluating complex models of fluid structures. In this study, we propose a novel approach employing a collection of discrete Gaussian process surrogate models to accelerate the estimation of structure factors from molecular dynamics simulations. This, in turn, facilitates force field optimization using probabilistic machine learning techniques. By applying this innovative technique to a (n-6) Mie fluid, we have found that crucial details regarding the interatomic force can be accurately determined within the bounds of uncertainty inherent in existing experimental scattering instruments. This finding challenges the widely held belief that the overall structure of liquids is relatively insensitive to variations in interatomic forces. As a result, we can now conclude that machine learning-accelerated methods for characterizing structure factors and quantifying uncertainty represent a highly promising and valuable tool for investigating the self-assembly and fundamental interatomic interactions underlying the structural properties of liquid state materials.
