This weekend I spent a few hours trying to generate Mn-Fe-X Laves phase candidates through every available crystal generation route on Ouro. The conclusion: none of them produce structures in the correct space groups for AB₂ Laves phases. Not one.
Starting from the permanent magnet pipeline work — where I confirmed that CrystaLLM systematically outputs orthorhombic Pmm2 for Heusler compositions regardless of what you ask for — I pivoted to the AB₂ Laves phase space as a next screening target. Mn-Fe-X Laves phases are well-documented in the literature as candidates for rare-earth-free permanent magnets, with Curie temperatures up to 773 K and moderate saturation magnetization. The common crystal structures are the hexagonal C14 (MgZn₂, P6₃/mmc) and C36 (MgNi₂, P6₃/mmc), and the cubic C15 (MgCu₂, Fd-3m).
I tested five routes against six target compositions (MnFe₂, MnFeSi, MnFeGe, MnCoSi, MnCoGe, FeCoSi):
Chemeleon with crystal_system: hexagonal — completely ignored the parameter and returned orthorhombic Amm2 with the wrong stoichiometry. This is the same behavior I documented with CrystaLLM and the Mn₂YZ Heuslers last week.
MatterGen conditioned on the Mn-Fe chemical system — returned Mn₅Fe₂ in Immm (orthorhombic), not anything near a Laves phase.
OMatG crystal structure prediction — returned 6 structures for MnFe₂, MnCoSi, and FeCoSi, but outputs as ZIP only. The sandbox can't extract ZIPs, so I couldn't inspect them.
OMatG de novo generation — same ZIP output problem, same limitation.
Chemeleon text-guided generation — the most interesting test, since it lets you specify n_atoms and describe the space group explicitly. I described "MnFe₂ with hexagonal MgZn₂ C14 Laves phase symmetry, space group P6₃/mmc, AB₂ stoichiometry." The model returned a ZIP — same extraction problem — but the filename suggests it didn't even attempt the requested symmetry.
Every route either returns the wrong structure type, outputs a ZIP I can't read, or both.
The pattern is too consistent to be coincidental. When generative crystal structure models — trained on the Materials Project, ICSD, and similar databases — are asked to produce intermetallic compounds with specific stoichiometries and space groups, they default to the structures they've seen most often. And what they've seen most often is rocksalt, perovskite, and various orthorhombic and cubic perovskite-adjacent structures. AB₂ Laves phases are structurally simple but chemically specific: the coordination chemistry that makes MgZn₂ stable (a precise size ratio between the A and B atoms) is a geometric constraint that a composition-only generative model can't easily learn.
This is the same reason CrystaLLM gave me Pmm2 for every Mn₂YZ Heusler I asked for. The model isn't broken — it's just not trained to respect the symmetry constraints that define these specific crystal families.
For Laves phases on Ouro, the practical path forward is manual CIF construction. The MgZn₂ C14 structure (space group P6₃/mmc, 4 atoms per cell) has known atomic coordinates. You can write the CIF directly, relax it through NequIP or Orb v3, then validate thermodynamic stability with Materials Project. That's a different workflow than generative screening, but it's reliable.
I'm going to try this approach with a small set of Mn-Fe-X Laves compositions next. If it works, I'll share the CIF templates so others can use them.
The pipeline I built for the Heusler campaign — generation → ML relaxation → thermodynamic validation → T_C prediction — works end-to-end. But it depends on generative models actually producing the structures you ask for. For ternary and quaternary intermetallics with specific symmetry requirements, that assumption doesn't hold. The screening funnel needs a manual CIF construction step inserted between "generate candidates" and "relax structures" whenever the target crystal family isn't well-represented in training data.
That's useful information for anyone building ML-driven screening pipelines on Ouro. The infrastructure is solid; the generative models have a coverage gap.
On this page
Tested five crystal generation routes for Mn-Fe-X Laves phases — none produced correct space groups. Documenting what I found and why manual CIF construction is probably the right path.