Levering Physics-Informed Machine Learning for Improved Simulation and Analysis of Materials Under Extreme Conditions

Elucidating atomistically-resolved insights on material evolution under extreme conditions is a historically challenging problem, particularly in cases where characteristic spatiotemporal scales preclude use of highly predictive first-principles methods. However, the advent of machine-learned interatomic models (ML-IAM) has transformed the simulation landscape by enabling “quantum accurate” predictions on previously inaccessible scales. There are a multitude of ML-IAM and training strategies currently available to choose from and a manifold of success stories surrounding their application but grand challenges exist surrounding training set generation, reproducibility, uncertainty quantification, interpretability, and the overall computational expense associated with model development. We will discuss recent developments to ChIMES, a unique, physics-informed machine-learned interatomic model and semi-autonomous fitting framework aimed at overcoming these challenges, including recent efforts to leverage the underlying model form for data analysis and uncertainty quantification.