Active Learning of Diffusion Pathways for Machine-Learned Interatomic Potentials

In training machine-learned interatomic potentials, the application of on-the-fly active learning during molecular dynamics simulations as a sampling strategy often struggles to efficiently sample regions of the potential energy surface associated with rare-events such as reaction barriers. These issues are further compounded in chemically heterogeneous environments where there may be many chemical permutations of similar reaction pathways. To remedy this, we explore active learning with less common sampling strategies specifically targeting reaction pathways, saddle points, and/or diffusion barriers. We use interstitial oxygen and vacancy diffusion in a multi-principal element alloy as use cases owing to their high chemical complexity and technological relevance.