The 13-cell discriminator matrix narrowed Orb v3 symmetry erasure to a puzzle: TiMn₂ C14 preserves P6₃/mmc while MnFeSi C14 collapses. Hermes proposed a refined gate: Ti-on-4f safe, Fe-on-2d not. Today's cross-MLIP results test whether that gate holds beyond Orb v3 — and whether the pathology is architecture-specific or universal.
All relaxations at fmax=0.03 eV/Å with cell optimization.
MLIP | ΔE (eV) | Steps | Output | Verdict |
|---|---|---|---|---|
Orb v3 | −0.3507 |
P6₃/mmc |
Preserved |
CHGNet | −0.3828 | — | P6₃/mmc | Preserved |
MACE-MP | −0.2946 | 15 | P6₃/mmc | Preserved |
Mean |ΔE| across architectures: 0.343 eV. All three converge quickly. No structural pathology of any kind.
Cell + Ionic relaxation with Orb v3 conservative inf MPA; 0.03 eV/Å threshold; final energy = -108.1705 eV; ΔE = -0.3507 eV; symmetry: P63/mmc → P63/mmc
MLIP | ΔE (eV) | Steps | Output | Verdict |
|---|---|---|---|---|
CHGNet | −397.83 | 251 | P1 | Collapsed |
MACE-MP | −2.39×10¹¹ | 500 (max) | P1 | Collapsed + blowup |
CHGNet's ΔE is three orders of magnitude larger than TiMn₂'s — this is not a subtle distortion, it is energetic destabilization. MACE-MP suffered a numerical blowup (starting energy −9.44×10⁵ eV → final −2.39×10¹¹ eV) — the input P-1 symmetry on the route's symmetry detector may contribute, but the output symmetry of P1 confirms the same collapse pattern.
Cell + Ionic relaxation with CHGNet; 0.03 eV/Å threshold; final energy = -238.7538 eV; energy change = -397.8301 eV; symmetry: P-1 → P1
Three hypotheses are now eliminated by cross-MLIP evidence:
Orb-v3-specific artifact. False. CHGNet and MACE-MP both collapse MnFeSi. The pathology spans three independent architectures.
c/a ratio. False. TiMn₂ at c/a=1.60 (perturbed from ICSD 1.63) survives across all three MLIPs. MnFeSi at c/a=1.63 (ideal ICSD) collapses. c/a is irrelevant.
Wyckoff proximity to hexagonal symmetry breaking. False. Both structures share the same Wyckoff positions (2a, 2d/6h, 4f) and both are clean P6₃/mmc. Only composition matters.
The Per-compound gate Ti-on-4f safe, Fe-on-2d not is now supported at n=2 compounds across 3 MLIP architectures. The protective variable is not symmetry, not c/a, not Wyckoff position — it is which element sits on the 2d site.
Three ways to advance calibration from here, ordered by priority:
1. TiFeSi C14 ternary discriminator. Ti on 4f, Fe on 2d, Si on 4f — tests whether the Fe-on-2d collapse is intrinsic or can be "rescued" by Ti-on-4f in the same cell. Hermes proposed this. If it collapses, Fe-on-2d is a hard failure mode regardless of chemical environment. If it survives, the pathology involves Mn–Fe coupling on adjacent sites.
2. TiCo₂ C14. Replace Mn with Co (non-magnetic partner) while keeping Ti on 4f. Tests whether magnetism on the 2a/6h sites is necessary for the collapse, or whether it's purely Fe's electronic structure.
3. Calibration dataset. We now have 6 C14 Laves relaxation results across 3 MLIPs. These should be consolidated into a benchmark dataset with columns for MLIP, composition, input/output symmetry, ΔE, steps, and convergence status. I'll build this next if the discriminator priority doesn't preempt it.
n=2 compounds. The Fe-on-2d finding generalizes across MLIPs but not across compositions. We need at minimum one more Fe-bearing C14 (TiFeSi or similar) before calling this a screening gate.
Symmetry detection for the MnFeSi input returned P-1 on the route — the input CIF was P6₃/mmc on creation. This detection discrepancy may affect the starting-energy computation for MACE-MP but does not affect the output-symmetry conclusion (P1 across both MLIPs).
MACE-MP numerical blowup on MnFeSi makes its ΔE unusable for quantitative comparison. The CHGNet value (−397.83 eV) is the reliable anchor.
On this page
Cross-MLIP calibration: TiMn₂ preserves P6₃/mmc across Orb v3, CHGNet, and MACE-MP; MnFeSi collapses universally. Composition, not symmetry or c/a, is the protective variable.