Tests of foundation interatomic potentials in d⁴, d⁶, d⁷ oxides

Acharya et al., arXiv:2607.08351, 2026
National Laboratory of the Rockies · King’s College London · Daresbury Labs

This paper tests how well foundation machine-learning interatomic potentials (MLIPs) - specifically MACE (two heads: matpes_r2scan and omol) and CHGNet - capture low-temperature electronic and structural order in strongly correlated transition-metal oxides, without any system-specific training.

Current foundation MLIPs are trained primarily on Born-Oppenheimer total energies, forces, and atomic descriptors. Consequently, their fidelity depends entirely on how the underlying electronic instability couples to static structural distortions. But this limited information does not always capture key features in correlated oxides. A foundation MLIP trained only on DFT total energies and forces will reproduce a correlated oxide’s ordering if and only if the electronic instability has already condensed onto a static lattice distortion present in its training data.

The difficulty is made clear in the present study of three isostructural perovskites (ABO₃) containing different d-electron counts: LaMnO₃ (d⁴), LaCoO₃ (d⁶), and NdNiO₃ (d⁷). A hierarchy of complexity emerges, which can be characterized by a single order parameter: a scalar or vector structural parameter, or in the case of LaCoO₃, an electronic order parameter that governs a transition from a low-spin state to a high-spin state.

The Scalar Class: NdNiO₃ (d⁷)

• Undergoes a metal-insulator transition driven by a rocksalt breathing distortion (alternating expansion/contraction of Ni-O bond lengths).

• Order Parameter: Scalar (a single number per site captures the average bond-length deviation).

• MLIPs capture the transition because the structural signature is a straightforward, single-coordinate geometric pattern. The molecularly trained omol potential reproduces the static frozen breathing pattern. matpes_r2scan and CHGNet capture the short-range fluctuations/precursors of the mode.

The Vector Class: LaMnO₃ (d⁴)

• Features cooperative Jahn-Teller (JT) distortions where long Mn-O bonds alternate in the ab-plane in a C-type antiferro-orbital pattern.

• Order Parameter: a vector determining which Cartesian axis carries the elongated bond at each site.

• MLIPs capture the magnitude, but miss the symmetry. CHGNet reproduces the correct JT distortion magnitude at low temperature but locks into the wrong pattern (ferro-orbital, all axes aligned) instead of the experimental C-type antiferro-orbital alternation. omol gets a partial, non-cooperative version. matpes_r2scan misses it almost entirely: yields dynamic JT fluctuations without static spatial ordering.

The On-Site Class: LaCoO₃ (d⁶)

• Undergoes a thermal low-spin (LS) to high-spin (HS) crossover driven by site-local multiplet population shifts.

• Order Parameter: On-site multiplet population (purely electronic with no spatial or structural order parameter).

• Structure-only MLIPs cannot capture this crossover: there is no static, symmetry-breaking lattice fingerprint. matpes_r2scan correctly reflects the absence of structural symmetry breaking, but MLIPs cannot predict the transition itself without local spin/multiplet descriptors.

Conclusions

Foundation MLIPs will reproduce a correlated oxide’s ordering only if trained on the relevant order parameter. Excising MLIPs are trained only on DFT total energies and forces and will capture transitions that depend on scalar structural information, as in NdNiO₃. MLIPs can be improved in different ways, e.g. add per-site spin labels for the nickelate class, Jahn-Teller-active training structures with correct C-type symmetry for the manganite class, and multiplet-resolved auxiliary targets (e.g., from DMFT or vertex-corrected QSGW) for the cobaltite class, since that channel has no geometric signature to learn from at all.