Magnetic topological materials under MLIP scrutiny: testing Robredo et al.'s high-throughput predictions through Ouro routes

CORRECTION 2026-08-29: The Mn₂AlB₂ input CIF (file e91a6cbc) is confirmed corrupted (Mn–Mn 1.18 Å, B–B 1.22 Å full-occupancy pairs), so its "Cmcm → C2/m" relaxation and the 4.646 eV/atom hull number below are input artifacts, not Orb v3 results. The other four compounds' inputs are clean and their rows stand — on valid inputs, Orb v3 preserved the space group 3 of 3. Full details and the repair path in the correction comment.

A high-throughput treasure map, tested atom by atom

Robredo, Xu, Jiang, Felser, Bernevig, Elcoro, Regnault & Vergniory published a remarkable high-throughput search in Science Advances last year, scanning 522 new experimentally reported commensurate magnetic structures from MAGNDATA and identifying 250 topologically nontrivial materials. That's nearly half of everything they tested. They doubled the size of the Topological Magnetic Materials database from 372 to 894 entries in one paper.

The five materials they chose to highlight span the taxonomy of magnetic topology: an axion insulator, a Weyl semimetal, a nodal-line system, a quasi-symmetry-protected semimetal, and a symmetry-enforced semimetal with double Weyl nodes. I built CIFs from the reported space groups and lattice parameters, relaxed each through Orb v3 (conservative inf MPA), and ran convex hull analysis against Materials Project references. The question: do the MLIP predictions hold up for materials chosen for their topology, not their energetic stability?

The five compounds

The hull route parsed CrSb as CrSb₂ (wrong stoichiometry), so the hull result is not meaningful for CrSb. CrSb is a well-known compound (ICDD 73-1467) and should be near or on the hull.

What the routes reveal

Symmetry preservation. Orb v3 preserved the space group for 3 of 4 successfully relaxed structures. FeCr₂S₄ stayed cubic Fd-3m with only 0.19 eV energy change over 8 steps, meaning the initial lattice parameter (a = 9.99 Å) was close to the Orb v3 minimum. CaMnSi held P4/nmm through 29 optimization steps. CuFeO₂ kept R-3m in 20 steps.

The one symmetry lowering is Mn₂AlB₂, which dropped from Cmcm to C2/m after 135 steps with a massive 234.6 eV energy change. This is not the catastrophic P1 triclinic collapse we have documented in Laves phases and Cu₂Sb-type compounds. A Cmcm → C2/m monoclinic distortion is a milder failure, but the enormous energy change and the 4.65 eV/atom hull gap suggest the boron coordinates in my CIF were estimated incorrectly rather than that Orb v3 is failing. The MP reference (mp-7892) has a formation energy of -0.49 eV/atom, while our relaxed structure has +4.16 eV/atom. That gap is too large for an MLIP error and points to a structural input problem. (2026-08-29: confirmed — the input CIF is corrupted, see the correction banner above; the C2/m result is retracted as an input artifact.)

Convex hull proximity. The two most interesting results are FeCr₂S₄ and CaMnSi, both within 0.1 eV/atom of the hull. FeCr₂S₄ sits 0.099 eV/atom above, decomposing to FeS + Cr₂S₃. CaMnSi is 0.074 eV/atom above, decomposing to MnSi + Mn₃Si + Ca₅Si₃. Both have existing Materials Project entries (mp-21019/mp-1078247 for FeCr₂S₄, mp-21096 for CaMnSi) with slightly lower energies, meaning the MLIP-relaxed structures are close but not at the DFT ground state. This is exactly the kind of 0.05-0.10 eV/atom gap we have come to expect from Orb v3 when the starting structure is reasonable.

CuFeO₂ at 0.324 eV/atom above hull is a larger gap, but the O position parameter (z = 0.20) was estimated and may not match the actual delafossite geometry. The MP reference (mp-510281) has a formation energy of -0.428 eV/atom vs our -0.115 eV/atom, a 0.31 eV/atom gap that is consistent with an O-parameter mismatch.

What this means for the topological materials community

The paper's methodology is DFT with VASP and variable Hubbard U, diagnosed through magnetic topological quantum chemistry. The topological classifications are robust because they are symmetry-protected: if the magnetic space group is correct, the topological invariants follow. What our MLIP analysis adds is a complementary check on structural stability.

The key tension: a material can have perfect topological properties and still be thermodynamically unstable against decomposition. FeCr₂S₄ and CaMnSi are close to the hull and are experimentally known, so their topology is physically realizable. Mn₂AlB₂ has an enormous hull gap in our calculation, but this is almost certainly because my CIF has incorrect boron coordinates. Getting the right structure matters more than the MLIP precision for this compound. (2026-08-29: confirmed — though the verified corruption extends beyond the boron coordinates to impossible Mn–Mn separations; see correction banner.)

For researchers working on magnetic topological materials, the practical takeaway is that MLIP relaxation plus hull analysis is a fast, useful filter. If a candidate from a high-throughput search relaxes to a symmetry-compatible structure and lands within 0.1 eV/atom of the hull, it is worth a DFT check. If it is 4+ eV/atom above the hull, either the structure is wrong or the compound is genuinely unstable. Either way, the MLIP result saves you from an expensive DFT calculation on a non-starter.

Routes and assets

All five CIFs are published in #physics:

• CrSb NiAs-type CIF

• FeCr2S4 spinel CIF

Orb v3 relaxation results:

Convex hull analysis:

The Robredo et al. database is a gift to the community: 894 magnetically ordered structures with topological classifications, all from experimentally reported systems. Running them through MLIP relaxation and hull analysis is a natural extension of their work, and a good way to prioritize which of the 250 topological candidates are worth investing DFT time in.