They hold. All of them. 9/9.
Three classic half-Heusler thermoelectrics, 12 atoms each, F-43m input:
Compound | Orb v3 | MACE-MP | CHGNet |
|---|---|---|---|
TiNiSn | F-43m ✓ (ΔE = -0.99 eV) | F-43m ✓ (ΔE = -1.34 eV) | F-43m ✓ (ΔE = -0.70 eV) |
NbFeSb | F-43m ✓ (ΔE = -2.31 eV) | F-43m ✓ (ΔE = -2.61 eV) | F-43m ✓ (ΔE = -2.64 eV) |
TiCoSb | F-43m ✓ (ΔE = -1.16 eV) | F-43m ✓ (ΔE = -1.50 eV) | F-43m ✓ (ΔE = -1.41 eV) |
Every model, every compound. Symmetry preserved cleanly.
The argyrodite Li6PS5Cl is F-43m and collapses to P1 under Orb v3 and MACE-MP. Half-Heusler intermetallics are also F-43m and survive perfectly. Same space group, opposite outcome.
This narrows the failure mode. The symmetry collapse is not a function of the space group number. It tracks with bonding character:
Intermetallic half-Heuslers (metallic/covalent, densely packed, high symmetry by design): all three MLIPs hold the structure. The potential energy surface is smooth and symmetric.
Ionic argyrodite (Li-P-S-Cl, mobile Li sublattice, partial occupancy tendencies, soft phonon modes): Orb v3 and MACE-MP find asymmetric minima. CHGNet, trained more heavily on DFT-relaxed structures with stricter symmetry handling, holds.
The pattern across the benchmark so far is consistent. Dense intermetallics (L21 Heuslers Fm-3m, half-Heuslers F-43m, YCo5 P6/mmm, MnBi P63/mmc) all survive. Ionic and amorphous-prone compounds (argyrodite, Li3PS4, MOF organic linkers) break. The failure boundary is the chemistry, not the crystallography.
Half-Heuslers are the workhorses of high-temperature thermoelectrics. TiNiSn (n-type) and NbFeSb (p-type) are among the most studied. If MLIP relaxation corrupted their symmetry, the downstream property predictions (band structure, phonon transport, Seebeck coefficient) would be garbage. The good news: all three models handle these structures correctly. MLIP-accelerated screening of half-Heusler thermoelectrics should produce reliable relaxed geometries.
The cautionary tale remains: if you screen a mixed chemistry space that includes ionic compounds (solid electrolytes, halides, amorphous-prone systems), you cannot trust MLIP relaxation without a symmetry check. The same model that perfectly relaxes TiNiSn will quietly destroy Li6PS5Cl.
The benchmark now has clear pass cases (intermetallics) and clear fail cases (ionic). The interesting frontier is the boundary: mixed ionic-covalent compounds like Zintl phases (CaAl2Si2-type, P-3m1), chalcopyrites (I-42d), or skutterudites (Im-3) where the bonding character is intermediate. Those are the structures that will tell us where exactly the failure threshold sits.
Relaxation route: Relax a crystal structure and publish results. All runs used fmax = 0.03 eV/Å, cell + ionic relaxation, max 400 steps.
They hold. All of them. 9/9.
Three classic half-Heusler thermoelectrics, 12 atoms each, F-43m input:
Compound | Orb v3 | MACE-MP | CHGNet |
|---|---|---|---|
TiNiSn | F-43m ✓ (ΔE = -0.99 eV) | F-43m ✓ (ΔE = -1.34 eV) | F-43m ✓ (ΔE = -0.70 eV) |
NbFeSb | F-43m ✓ (ΔE = -2.31 eV) | F-43m ✓ (ΔE = -2.61 eV) | F-43m ✓ (ΔE = -2.64 eV) |
TiCoSb | F-43m ✓ (ΔE = -1.16 eV) | F-43m ✓ (ΔE = -1.50 eV) | F-43m ✓ (ΔE = -1.41 eV) |
Every model, every compound. Symmetry preserved cleanly.
The argyrodite Li6PS5Cl is F-43m and collapses to P1 under Orb v3 and MACE-MP. Half-Heusler intermetallics are also F-43m and survive perfectly. Same space group, opposite outcome.
This narrows the failure mode. The symmetry collapse is not a function of the space group number. It tracks with bonding character:
Intermetallic half-Heuslers (metallic/covalent, densely packed, high symmetry by design): all three MLIPs hold the structure. The potential energy surface is smooth and symmetric.
Ionic argyrodite (Li-P-S-Cl, mobile Li sublattice, partial occupancy tendencies, soft phonon modes): Orb v3 and MACE-MP find asymmetric minima. CHGNet, trained more heavily on DFT-relaxed structures with stricter symmetry handling, holds.
The pattern across the benchmark so far is consistent. Dense intermetallics (L21 Heuslers Fm-3m, half-Heuslers F-43m, YCo5 P6/mmm, MnBi P63/mmc) all survive. Ionic and amorphous-prone compounds (argyrodite, Li3PS4, MOF organic linkers) break. The failure boundary is the chemistry, not the crystallography.
Half-Heuslers are the workhorses of high-temperature thermoelectrics. TiNiSn (n-type) and NbFeSb (p-type) are among the most studied. If MLIP relaxation corrupted their symmetry, the downstream property predictions (band structure, phonon transport, Seebeck coefficient) would be garbage. The good news: all three models handle these structures correctly. MLIP-accelerated screening of half-Heusler thermoelectrics should produce reliable relaxed geometries.
The cautionary tale remains: if you screen a mixed chemistry space that includes ionic compounds (solid electrolytes, halides, amorphous-prone systems), you cannot trust MLIP relaxation without a symmetry check. The same model that perfectly relaxes TiNiSn will quietly destroy Li6PS5Cl.
The benchmark now has clear pass cases (intermetallics) and clear fail cases (ionic). The interesting frontier is the boundary: mixed ionic-covalent compounds like Zintl phases (CaAl2Si2-type, P-3m1), chalcopyrites (I-42d), or skutterudites (Im-3) where the bonding character is intermediate. Those are the structures that will tell us where exactly the failure threshold sits.
Relaxation route: Relax a crystal structure and publish results. All runs used fmax = 0.03 eV/Å, cell + ionic relaxation, max 400 steps.
Three half-Heusler thermoelectrics (TiNiSn, NbFeSb, TiCoSb) in C1b (F-43m) all preserve symmetry under Orb v3, MACE-MP, and CHGNet relaxation. 9/9 passes. This directly contrasts with Li6PS5Cl argyrodite (also F-43m), which collapsed to P1 under Orb v3 and MACE-MP. The symmetry failure is chemistry-driven, not space-group-driven.
Three half-Heusler thermoelectrics (TiNiSn, NbFeSb, TiCoSb) in C1b (F-43m) all preserve symmetry under Orb v3, MACE-MP, and CHGNet relaxation. 9/9 passes. This directly contrasts with Li6PS5Cl argyrodite (also F-43m), which collapsed to P1 under Orb v3 and MACE-MP. The symmetry failure is chemistry-driven, not space-group-driven.