Predicting Material Properties with Machine Learned Interatomic Potentials

Abstract

Machine learning interatomic potentials (ML-IPs) have emerged as a promising approach for bridging the gap between quantum electronic structure calculations (QM) and large scale classical molecular modeling simulations and have shifted the development of these many-body force fields to become predominantly data-driven. Although these machine learning methods have been successfully used as specialized models for specific tasks, their performance and accuracy are often assessed by simple metrics during the training phase, while there has been little attention devoted to how data efficient and transferable these methods are in unexplored conditions. Here, we aim to investigate how well state-of-the-art machine learning interatomic potentials generalize beyond their intended systems and tasks. We focus on the Spectral Neighbor Analysis Potential (SNAP) and Message-Passing Graph Neural Networks (GNNs), compare their accuracy and data efficiency and examine their stability during long time integration of classical molecular dynamics. We extract thermodynamic properties connected to chemical reaction dynamics, kinetics and transition barriers of physical processes that are not in the learned phase space. We find that GNNs outperform SNAP as more robust ML-IPs and connect this to the importance of including out-of-domain applications as an extensive set of benchmarks for assessing the effective performance of machine learning architectures. Finally, we discuss the necessity of incorporating an active learning framework as a method to generate robust machine learning reactive potentials.