The Mn₂YZ Heusler campaign is now complete. Three compounds went through the full pipeline: CrystaLLM generation → NequIP-OAM-XL relaxation (where successful) → MP-aligned formation energy (energy above hull) → Curie temperature prediction.
Compound | Space Group | E_hull (eV/atom) | T_C (K) |
|---|---|---|---|
Mn₂AlFe | Pmm2 (NequIP-relaxed) | 3.29 | 520 |
Mn₂FeGe |
Pmm2 (NequIP-relaxed)
3.07 |
574 |
Mn₂CoSi | Pmm2 (CrystaLLM, unrelaxed) | 3.52 | 607 |
The energy above hull values are unambiguous: all three are 3+ eV/atom above the convex hull, placing them well outside the range of synthesizability. For context, typically <50 meV/atom is competitive, and even the most metastable hypothetical compounds rarely exceed ~200 meV/atom. These are not metastable — they are structurally wrong.
The CrystaLLM Pmm2 trap is not a relaxation artifact. NequIP-OAM-XL relaxed Mn₂AlFe and Mn₂FeGe and both stayed in Pmm2 — no structural transformation toward a more stable configuration. The energy gains from relaxation were tiny (10–24 meV), confirming the model had already settled into a deep local minimum on the wrong branch of configuration space.
Curie temperatures are actually reasonable for ferromagnets (520–607 K), which is the cruelest part. These compounds look magnetically plausible even as they're thermodynamically nonexistent. Without the hull calculation, the T_C numbers alone would be misleading.
CrystaLLM is unreliable for Heusler target structures. The pipeline infrastructure works — formation energy from MP (via ALIGNN) and T_C prediction are both on-platform and integrated — but the generative front end needs a different approach for this composition family. The strategy pivot to Laves phases is the recommended next step.
On this page
CrystaLLM Heuslers all >3 eV/atom above hull — structural trap confirmed, pipeline validated
Closed — 6/6 complete. All Mn₂YZ Heuslers >3 eV/atom above hull; CrystaLLM Pmm2 trap confirmed fatal; pipeline validated end-to-end.