Evidential deep learning for interatomic potentials

Abstract

Machine learning interatomic potentials have been widely used to facilitate large-scale molecular simulations with accuracy comparable to ab initio methods. To ensure the reliability of the simulation, the training dataset is iteratively expanded through active learning, where uncertainty serves as a critical indicator for identifying and collecting out-of-distribution data. However, existing uncertainty quantification methods tend to involve either expensive computations or compromise prediction accuracy. Here we show an evidential deep learning framework for interatomic potentials with a physics-inspired design. Our method provides uncertainty quantification without significant computational overhead or decreased prediction accuracy, consistently outperforming other methods across a variety of datasets. Furthermore, we demonstrate applications in exploring diverse atomic configurations, using examples including water and universal potentials. These results highlight the potential of our method as a robust and efficient alternative for uncertainty quantification in molecular simulations. The authors introduce an evidential deep learning framework for machine learning interatomic potentials that efficiently provides robust uncertainty quantification, demonstrating its effectiveness across diverse atomic systems.