Exploring the discrepancies, the extrapolability, and the interpolability of machine learning interatomic potentials

Machine learning interatomic potentials (MLPs) are a promising technique for atomic modeling. While high accuracy and small errors are widely reported for MLPs, an open concern is whether MLPs can accurately reproduce atomistic dynamics and related physical properties in their applications in molecular dynamics (MD) simulations. Here we examine the extrapolability and interpolability of current state-of-the-arts MLPs and reveal a number of discrepancies related to defects and rare events (REs), in comparison to ab initio computation. After revealing current testing for MLPs showing low averaged errors is inadequate, we develop and demonstrate quantitative metrics, which are better indicators for the prediction of atom dynamics and related properties in MD simulations. Leveraging these findings, we develop a RE-enhanced workflow of MLP training that identifies REs, enhances the training set, and optimized MLPs using enhanced performance scores. The MLPs trained by the RE-enhanced workflow are demonstrated to have improved prediction in diffusional properties. Given the identified errors and improved workflow are general to all MLPs, our study provides critical testing and general guidance for future development and improvements of accurate, robust, and reliable MLPs for atomistic modeling.