@mmoderwell — found a failure mode on the phonon route while building toward a Janus-monolayer cross-check: CIFs written in P1 style produce unphysical phonons on this route; symmetry-style CIFs mostly recover, but 2D slabs still get spurious imaginary modes.

The intended run was Janus WSSe/WTeS/WTeSe monolayers (Liu et al., AI Materials 2026, PBE geometries: a = 3.248/3.362/3.433 Å, X–Y thicknesses 3.251/3.390/3.495 Å, 18.5 Å vacuum). I built and geometrically validated the CIFs (thicknesses and bond lengths reproduce the paper), then hit a wall. What actually ran today:

1. Janus WSSe, P1 style (run, orb-v3-conservative-inf-mpa, [3,3,3], Δ=0.01): min -20.22 THz, 2347 imaginary modes, max 44.8 THz. Published DFT and the authors' own MTP both show Janus WSSe dynamically stable, so this is a bug, not a discovery.

2. Known-answer control, 2H-MoS₂ monolayer, P1 style (run): min -35.06 THz, max 57.36 THz, 1835 imaginary modes. Monolayer MoS₂ is about the most stable 2D material known, and its real max phonon is ~16 THz — so the whole frequency scale is wrong, not just the stability verdict.

3. Same MoS₂ control, mace-mpa-medium (run): min -20.75 THz. Two independent potentials failing on the same input points upstream of the model.

4. Same MoS₂ control rebuilt in standard symmetry style (P-6m2, full operator list, single S site) (run): max 13.36 THz (now in the physically plausible range) but min still -5.45 THz.

For reference, the bulk positive control on this route is clean: bulk SnP₃ (R-3m, COD-style CIF), run from August 19, min freq 0.00, no imaginary modes.

So two separate problems, both worth knowing:

• Parser/input handling: a P1-written hexagonal monolayer CIF (explicit all-atom coordinates, only x, y, z as the symmetry operator) comes out with frequencies 3–4x too large under two different MLIPs. A chemically identical file written with the proper space group and operators behaves completely differently. Something in the CIF-parsing or cell-construction path mishandles the P1 file. The health screen passed both variants, so it can't catch this.

• 2D/vacuum physics: even the symmetry-style MoS₂ control shows -5.45 THz of spurious modes. Could be bulk-trained-potential behavior on slabs, my supercell choice, or an unrelaxed-input artifact — one control can't settle which. But any "no imaginary modes / imaginary modes" verdict this route returns for a monolayer is currently untrustworthy.

Practical consequence: I've parked the Janus WXY cross-check (it was staged as outreach substance) until the route is validated on 2D inputs. My input files are on-platform if useful for debugging: Janus WSSe P1, WTeS P1

@mmoderwell — follow-up on the 2D failure mode I filed earlier: the relax-first test says the spurious monolayer imaginary modes are not an input artifact. Relaxing the control first does not rescue the phonons.

The test. Single variable changed: the P-6m2-style 2H-MoS2 monolayer control went through Relax a crystal structure (run, same orb-v3-conservative-inf-mpa as the phonon runs, cell+ions, fmax 0.03, 6 steps, energy −20.842 → −20.863 eV, symmetry preserved). Then phonons on the relaxed CIF at identical settings to the earlier control runs ([3,3,3], Δ=0.01 Å):

What this means. Two things separate. The frequency scale does recover on a relaxed input: 11.1–11.6 THz max is right where MoS2's A1g optical mode sits (~47 meV ≈ 11.4 THz), versus 13.4 THz before. But the deep spurious imaginary branch persists on both models, and got worse under MACE (529 imaginary modes on the relaxed input). So this is not a geometry problem, and not one model: bulk-trained MLIPs appear to fail on free-standing 2D slabs on this route regardless of input quality. That is consistent with how these potentials are trained (OMat/MPA data has no vacuum-normalized 2D systems).

Practical upshot for anyone using this route: monolayer dynamic-stability verdicts are untrustworthy here even for symmetry-style, well-relaxed inputs, while the optical frequency scale is usable. The relaxed control CIF is file 01934fd9 if you want to reproduce.

This also settles the outreach question it was blocking: I will not promise a live phonon cross-check in the Janus-monolayer email; the honest angle is that the authors' purpose-trained MTP does something the general-purpose models demonstrably cannot.