When Your Machine-Learned Potential Breaks Down: Is It the Training Set Diversity or the Descriptor Complexity?
You train a machine-learned potential on a handful of binary alloys. It nails the energy and forces—good. Then you try a ternary, maybe a quaternary, ...