Oxidation-aware machine learning interatomic potential for magnetite Fe3O4

by Dr Libor Vojacek (PSI)

Wednesday 6 May 2026, 15:00 → 16:00 Europe/Zurich

OVGA/200

Magnetite Fe₃O₄ is the oldest known magnetic material, yet the physics of its Verwey transition is still not fully solved due to its coupled structural, charge, and orbital interactions. The difficulty hinges on the multivalent nature of iron (Fe2+/Fe3+), which is a central motif in redox chemistry and functional oxides.Machine learning interatomic potentials (MLIP) trained on DFT+U+V data have recently proven to efficiently tackle the oxidation state of multivalent metallic ions with near ab initio accuracy [1], adding to their intrinsic efficiency enabling the simulations of large and/or complex systems. This is especially relevant for magnetite where the low-temperature monoclinic phase requires a large (224 atom) supercell with a combinatorial number of 64 choose 32 ~ 10¹⁸ possible charge orderings, making ab initio calculations expensive. In this work, we train an oxidation-aware MLIP for magnetite and demonstrate its capabilities side-by-side with a conventional (unaware) fine-tuned foundation model [2]. While the oxidation-aware potential provides the oxidation state directly as a control variable, the conventional potential can capture it intrinsically through the distinct volumes of the octahedral oxygen cages [3].

1. C. Malica and N. Marzari, Npj Comput. Mater. 11 (2025) 212
2. S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, Nat Commun 13 (2022) 2453
3. L. Vojáček and N. Marzari, in preparation (2026)