Uncertainty Quantification for Misspecified Machine Learned Interatomic Potentials

Abstract

The use of high-dimensional regression techniques from machine learning has significantly improved the quantitative accuracy of interatomic potentials. Atomic simulations can now plausibly target quantitative predictions in a variety of settings, which has brought renewed interest in robust means to quantify uncertainties on simulation results. In many practical settings, encompassing both classical and a large class of machine learning potentials, the dominant form of uncertainty is currently not due to lack of training data but to misspecification, namely the inability of any one choice of model parameters to exactly match all ab initio training data. However, Bayesian inference, the most common formal tool used to quantify uncertainty, is known to ignore misspecification and thus significantly underestimates parameter uncertainties. Here, we employ a recent misspecification-aware regression technique to quantify parameter uncertainties, which is then propagated to a broad range of phase and defect properties in tungsten via brute force resampling or implicit differentiation. The propagated misspecification uncertainties robustly envelope errors to direct \textit{ab initio} calculation of material properties outside of the training dataset, an essential requirement for any quantitative multi-scale modeling scheme. Finally, we demonstrate application to recent foundational machine learning interatomic potentials, accurately predicting and bounding errors in MACE-MPA-0 energy predictions across the diverse materials project database. Perspectives for the approach in multiscale simulation workflows are discussed.