Title:Evidential Deep Learning for Interatomic Potentials

Abstract:Machine learning interatomic potentials (MLIPs) have been widely used to facilitate large scale molecular simulations with ab initio level accuracy. However, MLIP-based molecular simulations frequently encounter the issue of collapse due to decreased prediction accuracy for out-of-distribution (OOD) data. To mitigate this issue, it is crucial to enrich the training set with active learning, where uncertainty estimation serves as an effective method for identifying and collecting OOD data. Therefore, a feasible method for uncertainty estimation in MLIPs is desired. The existing methods either require expensive computations or compromise prediction accuracy. In this work, we introduce evidential deep learning for interatomic potentials (eIP) with a physics-inspired design. Our experiments demonstrate that eIP consistently generates reliable uncertainties without incurring notable additional computational costs, while the prediction accuracy remains unchanged. Furthermore, we present an eIP-based active learning workflow, where eIP is used not only to estimate the uncertainty of molecular data but also to perform uncertainty-driven dynamics simulations. Our findings show that eIP enables efficient sampling for a more diverse dataset, thereby advancing the feasibility of MLIP-based molecular simulations.