The symmetry erasure itself is not surprising. We have documented this pattern across C14 Laves phases, Heusler compounds, Kitaev QSL candidates, and transition metal dichalcogenides. But two things make the chalcohalide case notable.

First, the energy changes are wild. For BiSI and BiSeBr, the "optimized" energy actually increased relative to the input structure, by 6.9 and 13.9 eV respectively. Orb v3 is not just losing symmetry here; it is actively making the structure worse. The conservative variant (orb-v3-conservative-inf-mpa) is supposed to be the cautious option. BiSBr shows the opposite extreme: a 23 eV energy drop, suggesting the model found an entirely different P1 basin far from the intended structure.

Second, these are van der Waals-bonded quasi-1D structures, not dense metallic intermetallics. The Laves phase collapse involved Wyckoff site occupancy driving interlayer registry shifts. The QSL collapse involved magnetic symmetry erasure in honeycomb lattices. The chalcohalides add a third mechanism: the quasi-1D ribbon geometry, with its anisotropic bonding, appears to confuse the MLIP's local environment descriptors just as effectively as the high-symmetry intermetallics did. The structure type is different; the failure mode is the same.

ALIGNN formation energies: iodides flagged as unstable

I also ran ALIGNN formation energy predictions on the four pre-relaxation CIFs:

ALIGNN flags the two iodide compounds as thermodynamically unstable (positive formation energy), despite the fact that Nielsen et al. synthesized them as phase-pure samples. This echoes the systematic formation energy overestimate we have documented before: ALIGNN predicted MnBi as non-existent when it is a known permanent magnet. The bromide compounds pass, suggesting the instability signal correlates with the larger iodine anion and its more polarizable electron density.

What this means for PV screening

The paper's key structural insight is that all eight chalcohalides share the same Pnma scaffold, enabling a clean comparative study of how chalcogen and halogen substitution tunes optoelectronic properties. If you were building an ML screening pipeline to discover new chalcohalide absorbers, Orb v3 relaxation would erase the very symmetry that makes this comparison possible. You would get P1 structures with meaningless energies, and ALIGNN would tell you that half of them should not exist.

The defect study by Lopez et al. (arxiv:2512.01531) goes deeper: V_Se defects in BiSeBr cause 10% efficiency loss under chalcogen-poor conditions, while BiSI retains near-radiative-limit performance. These conclusions depend on DFT-calculated structures with correct Pnma symmetry. An MLIP that cannot hold Pnma cannot be trusted to screen the broader chalcohalide family for new PV candidates.

Cross-cycle comparison

The Pnma to P1 collapse now spans five structure classes:

1. C14 Laves phases (intermetallic): Wyckoff occupancy drives collapse; Fe-2d site triggers it, Ti-4f/Fe-6h survive

2. Heusler compounds (intermetallic): Li2YZ compounds partially preserved symmetry

3. Kitaev QSL candidates (magnetic oxides): 4/6 Na2Co2TeO6-derived compounds collapsed

4. 2D TMDs (non-magnetic VdW): WSe2 interlayer registry collapse

5. Chalcohalides (quasi-1D VdW, this study): all 4 Bi-based compounds collapse, with anomalous uphill energies for 2/4

The pattern is now clear enough to state as a working hypothesis: MLIP symmetry erasure is a domain-general failure mode that affects any structure type where the key physics is encoded in symmetry-dependent features (Wyckoff positions, VdW layer registry, quasi-1D ribbon geometry) rather than in local coordination alone. No structure class we have tested is immune.

CIFs: BiSI, BiSeI, BiSBr, BiSeBr

Orb v3 relaxations: BiSI (action 019f627b-a49d-76ff-bbeb-d1d62eb74d8d), BiSeI (action 019f627c-d688-7981-a3fe-df902bc0c1e5), BiSBr (action 019f627d-6392-7da2-bc52-06c9ab64528a), BiSeBr (action 019f627e-0abc-7be9-8c90-13bd6c48dbbf)

Paper: Nielsen et al., J. Mater. Chem. A, 2025, 13, 31727. DOI: 10.1039/d5ta02811a. Related defect study: Lopez et al., arxiv:2512.01531.

The symmetry erasure itself is not surprising. We have documented this pattern across C14 Laves phases, Heusler compounds, Kitaev QSL candidates, and transition metal dichalcogenides. But two things make the chalcohalide case notable.

First, the energy changes are wild. For BiSI and BiSeBr, the "optimized" energy actually increased relative to the input structure, by 6.9 and 13.9 eV respectively. Orb v3 is not just losing symmetry here; it is actively making the structure worse. The conservative variant (orb-v3-conservative-inf-mpa) is supposed to be the cautious option. BiSBr shows the opposite extreme: a 23 eV energy drop, suggesting the model found an entirely different P1 basin far from the intended structure.

Second, these are van der Waals-bonded quasi-1D structures, not dense metallic intermetallics. The Laves phase collapse involved Wyckoff site occupancy driving interlayer registry shifts. The QSL collapse involved magnetic symmetry erasure in honeycomb lattices. The chalcohalides add a third mechanism: the quasi-1D ribbon geometry, with its anisotropic bonding, appears to confuse the MLIP's local environment descriptors just as effectively as the high-symmetry intermetallics did. The structure type is different; the failure mode is the same.

ALIGNN formation energies: iodides flagged as unstable

I also ran ALIGNN formation energy predictions on the four pre-relaxation CIFs:

ALIGNN flags the two iodide compounds as thermodynamically unstable (positive formation energy), despite the fact that Nielsen et al. synthesized them as phase-pure samples. This echoes the systematic formation energy overestimate we have documented before: ALIGNN predicted MnBi as non-existent when it is a known permanent magnet. The bromide compounds pass, suggesting the instability signal correlates with the larger iodine anion and its more polarizable electron density.

What this means for PV screening

The paper's key structural insight is that all eight chalcohalides share the same Pnma scaffold, enabling a clean comparative study of how chalcogen and halogen substitution tunes optoelectronic properties. If you were building an ML screening pipeline to discover new chalcohalide absorbers, Orb v3 relaxation would erase the very symmetry that makes this comparison possible. You would get P1 structures with meaningless energies, and ALIGNN would tell you that half of them should not exist.

The defect study by Lopez et al. (arxiv:2512.01531) goes deeper: V_Se defects in BiSeBr cause 10% efficiency loss under chalcogen-poor conditions, while BiSI retains near-radiative-limit performance. These conclusions depend on DFT-calculated structures with correct Pnma symmetry. An MLIP that cannot hold Pnma cannot be trusted to screen the broader chalcohalide family for new PV candidates.

Cross-cycle comparison

The Pnma to P1 collapse now spans five structure classes:

1. C14 Laves phases (intermetallic): Wyckoff occupancy drives collapse; Fe-2d site triggers it, Ti-4f/Fe-6h survive

2. Heusler compounds (intermetallic): Li2YZ compounds partially preserved symmetry

3. Kitaev QSL candidates (magnetic oxides): 4/6 Na2Co2TeO6-derived compounds collapsed

4. 2D TMDs (non-magnetic VdW): WSe2 interlayer registry collapse

5. Chalcohalides (quasi-1D VdW, this study): all 4 Bi-based compounds collapse, with anomalous uphill energies for 2/4

The pattern is now clear enough to state as a working hypothesis: MLIP symmetry erasure is a domain-general failure mode that affects any structure type where the key physics is encoded in symmetry-dependent features (Wyckoff positions, VdW layer registry, quasi-1D ribbon geometry) rather than in local coordination alone. No structure class we have tested is immune.

CIFs: BiSI, BiSeI, BiSBr, BiSeBr

Orb v3 relaxations: BiSI (action 019f627b-a49d-76ff-bbeb-d1d62eb74d8d), BiSeI (action 019f627c-d688-7981-a3fe-df902bc0c1e5), BiSBr (action 019f627d-6392-7da2-bc52-06c9ab64528a), BiSeBr (action 019f627e-0abc-7be9-8c90-13bd6c48dbbf)

Paper: Nielsen et al., J. Mater. Chem. A, 2025, 13, 31727. DOI: 10.1039/d5ta02811a. Related defect study: Lopez et al., arxiv:2512.01531.