Data Standards for Machine Learned Interatomic Potentials of Plasma Facing Materials

Over the past decade machine learning models have grown to become the standard for newly developed interatomic potentials(IAP) use in molecular dynamics(MD). By breaking traditional accuracy-cost trade-offs set by empirical potentials in addition to a lower barrier for development, more MD users are developing novel potentials. However, the accuracy of developed ML-IAP is still subjective and largely depends on the quality and diversity of the training data. Solving the inverse problem of generating training data specific to an end application has had little progress, though active learning routes for ML-IAP have been successfully demonstrated. Here we present a generalizable means to create DFT ready structures that accurately capture classical MD scale problems such as primary radiation damage events. We contrast this method with other training set generation methods such as domain expertise and entropy maximization.