Developing machine learned potentials for high temperature applications

Machine learned potentials inherently perform worse on structures outside of the training set than structures directly trained to. This presents a challenge in developing stable, transferable potentials for high temperature applications as there is plenty of energy for the atoms to move away from ground state and other expected configurations. We will summarize lessons learned while developing the W-ZrC SNAP potential to study material behavior under fusion reactor conditions (≥1000K). We will discuss progress and challenges in developing a W-ZrC-H ACE potential, which will enable the study of plasma-material interactions.