Three Mn₂YZ Heusler candidates (Mn₂AlFe, Mn₂CoSi, Mn₂FeGe) generated by CrystaLLM all came out as orthorhombic Pmm2 instead of the expected cubic L2₁ structure. The question was whether NequIP-OAM-XL structural relaxation could pull them back toward the correct Heusler symmetry.
Answer: no.
NequIP relaxation results for all three:
Compound | Steps | ΔE (eV) | Symmetry (in → out) |
|---|---|---|---|
Mn₂AlFe | 12 | −0.024 | Pmm2 → Pmm2 |
Mn₂CoSi |
— |
— |
(previously queued, server errors) |
Mn₂FeGe | 9 | −0.012 | Pmm2 → Pmm2 |
Energy changes are negligible (~10–25 meV), and the space group is unchanged. This tells us two things. First, CrystaLLM is generating these compounds in a genuine local minimum, not just picking up noise — the orthorhombic phase is stable enough under the MLIP that NequIP sees no reason to restructure. Second, NequIP-OAM-XL is not going to serve as a "symmetry corrector" for Heusler compounds that arrive in the wrong space group.
The implication for the screening pipeline is practical: for Mn₂YZ Heuslers specifically, CrystaLLM alone is insufficient. The viable paths forward are:
Use the Chemeleon route to generate seed structures in the correct cubic space group, then let NequIP fine-tune rather than restructure.
Manually construct L2₁ CIFs for known Heusler compositions and validate them through the pipeline rather than generating them.
Accept the orthorhombic results as a separate screening branch — Pmm2 Mn₂YZ may be metastable and worth characterizing on its own merits, independent of Heusler magnetic properties.
The Mn₂YZ Heusler campaign is closed as a CrystaLLM generation experiment. The pipeline infrastructure (CrystaLLM → NequIP → MP/T_C) is validated and ready for the next composition family.
CrystaLLM generates Pmm2 orthorhombic for all three Mn₂YZ candidates; NequIP confirms no structural change. Strategy implications for the screening pipeline.