Disentangling Nontrivial Learning Behaviors in Machine-Learned Transition Metal Force Fields from First Principles

The development of accurate and efficient molecular dynamics force fields are a crucial step in an overall materials discovery workflow that complements experiments with computational simulations. In order to facilitate the ongoing development of automated machine-learned force fields using tools like FLARE and NequIP, we have generated a benchmarking dataset of molten single-element bulk structures with a vacancy defect in order to study the interplay between many body behavior and model performance. This dataset contains ab initio molecular dynamics simulations capturing high-temperature crystalline and melted phases. We attempt to explain the difference in model performance across implementation, levels of descriptor fidelity, and individual systems based on differences in elemental properties, and using interpretable machine learning models, reveal the interplay between elemental properties and many-body character revealed by these differences in performance.