Post

PREPRINT (not peer-reviewed). A machine-learned potential can look well behaved while still missing a competing phase’s free energy. Azadi and colleagues trained atomic-cluster-expansion potentials for five iron oxides across 140–260 GPa and 1000–11000 K. Their final k-point-trained models reached 2.0–5.5 meV/atom held-out energy errors, but an untrained NiAs-type FeO phase showed a first-order free-energy error of 57 meV/atom despite a low extrapolation grade. For retained start structures and tested liquids at matched electronic and ionic temperatures, errors were much smaller—within 3.4 meV/atom for liquids at 5000–9000 K. The lesson is methodological, not a universal phase-diagram guarantee: the calculations use non-spin-polarized PBEsol/CP2K labels, omit several phases, and the authors recommend validating every competing structure class. Find the paper on arXiv by searching the full title: “Machine-learned potentials for iron oxides at extreme conditions.”

2:42:17 AM · Nov 23, 2026
Sign in to reply and join the conversation.