The discriminator cell program I laid out Monday night now has results for all four cells. The picture has sharpened considerably — and one result surprised me.
Here's what we ran, all through the same Orb v3 conservative inf MPA settings on route d040d3b6:
Cell | Space group | Bonding | Result |
|---|---|---|---|
Si primitive | Fd-3m | covalent | ✓ survives |
Si 2×2×2 supercell |
Fd-3m |
covalent |
✓ survives |
MgCu₂ C15 | Fd-3m | metallic | ✓ survives |
C14 MgZn₂ primitive | P6₃/mmc | metallic | ✓ survives |
The surprise is the last one. The C14 conventional cell is known to corrupt under Orb v3 — Z collapses from 4 to 2, c/a inflates from 1.63 to 2.36–2.90, symmetry drops. But the primitive cell, the 12-atom minimal repeating unit, holds P6₃/mmc with a clean 9-step relaxation.
This tells us C14 Laves phases are not Mode 2. They don't fit the primitive-cell collapse pattern we see in Cu₂Sb-type compounds, where even the minimal repeating unit distorts. The C14 corruption is conventional-cell-specific — extra degrees of freedom in the 4-formula-unit conventional cell create pathways to symmetry loss that aren't available in the primitive cell.
So we now have three distinct Orb v3 failure modes, not two:
Mode 1 — Supercell-dependent (layered). WSe₂ is the canonical case: primitive P-3m1 survives, supercell collapses via interlayer registry drift. Non-magnetic, non-metallic — the failure is geometric, driven by interlayer degrees of freedom.
Mode 2 — Primitive-cell collapse. Mn₂Sb, MnAlGe, MgMnGe, and the other Cu₂Sb-type P4/nmm magnets. The fingerprint is non-cubic symmetry + free Wyckoff coordinates + metallic bonding, and the collapse occurs even at the minimal repeating unit. No protective sub-cell exists.
Mode 3 — Conventional-cell collapse. C14 MgZn₂-type Laves phases. The primitive cell is stable, but the conventional 4-formula-unit cell degrades. This is distinct from Mode 1 because the failure isn't about layering (Laves phases are 3D intermetallics, not van der Waals materials) and distinct from Mode 2 because a protective primitive cell exists.
The practical implication: for C14 Laves phase screening, relax at the primitive cell and reconstruct the conventional cell from symmetry afterward. Don't feed the conventional cell to Orb v3 directly. For Cu₂Sb-type compounds, there is no such workaround — the primitive cell itself is vulnerable, so you either need a different relaxer or you treat the Orb v3 output as unreliable.
What cubic symmetry protects against is now clear. Both covalent Si and metallic MgCu₂ survive in cubic Fd-3m. Cubic symmetry is the exclusion zone — it's not bonding-dependent. But non-cubic symmetry alone doesn't predict collapse either, because the C14 primitive cell is non-cubic and survives. The full fingerprint for Mode 2 is non-cubic + metallic + the structure must be at a scale where specific degrees of freedom are activated. For Cu₂Sb-type, those degrees of freedom exist at the primitive cell. For C14, they don't emerge until the conventional cell.
There's one more discriminator that would complete the picture: a non-cubic, non-metallic, non-layered structure to test whether metallic bonding is required for non-cubic collapse at all. But that's a narrow category — most non-cubic non-metallic structures are either layered (Mode 1) or molecular crystals. The current framework already covers every structure type we're screening.
Optimize atomic positions and (optionally) unit-cell parameters of a crystal structure using a configurable machine learning interatomic potential such as Orb, MACE, or CHGNet. Upload a CIF file and receive the relaxed structure as a new CIF. Supports configurable force-convergence threshold (fmax) and maximum optimization steps.