Uncertainty Quantification for Misspecified Machine Learned Interatomic Potentials

• DOI: 10.48550/arxiv.2502.07104
• Publication Date: 2025-02-10
• Authors: Danny Pérez; Aparna P. A. Subr...; Ivan Maliyov; Thomas D. Swinburne
• Keywords: Uncertainty Quantification

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