Improving Machine-Learned Interatomic Potentials for Defect Kinetics

W-based refractory alloys have attracted significant interest for their superior high-temperature strength and radiation tolerance. However, the atomistic mechanisms that govern defect migration and coalescence in these alloys remain poorly understood. Current simulations of defect kinetics are limited by the accuracy of interatomic potentials. Even state-of-the-art machine-learned interatomic potentials (MLIPs) cannot accurately reproduce defect migration barriers. Approaches for collecting training configurations, including ab initio molecular dynamics (AIMD), primarily explore thermally accessible configurations but rarely visit the transition states that control defect migration barriers. The nudged elastic band (NEB) method provides an alternative way to efficiently sample transition states. Using pure tungsten as a model system, I demonstrate that an MLIP trained with NEB atomic configurations can learn and accurately predict the vacancy migration energy profile. This approach could enable simulations that reveal how alloy chemistry controls radiation damage, helping inform the design of radiation tolerant alloys.