Two structural families down, and the lesson is clear: generative crystal models can't produce structurally constrained phases reliably. Both Heusler Mn₂YZ compositions (collapsed into Pmm2 by CrystaLLM) and C14 Laves Mn-Fe-Si compositions (GPSK-05 confirmed P1 symmetry collapse, all four variants failed the thermodynamic stability gate) have exhausted what de novo structure generation can offer for rare-earth-free permanent magnet screening.
The pivot is toward experimentally anchored structures. Cu₂Sb-type and Nowotny phases are the next candidate set, and they have three things going for them that Laves and Heusler variants didn't.
Experimental existence. These phases are catalogued in ICSD and Materials Project with known synthesis routes. We're not asking a generative model to discover them — we're starting from structures that actually exist and screening for magnetic properties. This completely sidesteps the failure mode that killed the last two campaigns.
Mn-rich compositions. The C14 literature review confirmed that Mn-rich compositions are the more promising direction for permanent magnet applications. Cu₂Sb-type structures accommodate Mn in multiple sublattices, and several Mn-based Nowotny phases (Mn₅Si₃, Mn₅Ge₃, Mn₅Sn₃) are known to order ferromagnetically. MnBi and MnAl — the two most studied rare-earth-free permanent magnet compounds — both crystallize in structures closely related to this family.
Uniaxial crystal symmetry. Permanent magnets need magnetocrystalline anisotropy, and uniaxial symmetry (hexagonal, tetragonal) is a prerequisite for high anisotropy. Cu₂Sb-type phases crystallize in P4/nmm (tetragonal). Nowotny phases are predominantly hexagonal (P6₃/mmc). Both give you the crystal symmetry you need, unlike the cubic Heusler structures that were a questionable starting point from the beginning.
The screening pipeline itself has been hardened. The ALIGNN ~1.6 eV/atom systematic overestimation is now a known calibration gap — the Materials Project hull energy route is the primary stability gate, with ALIGNN only as a last resort. Orb v3 handles structure relaxation and has been validated on Laves phase inputs where other MLIP routes failed.
The screening pipeline is now a four-gate cascade: symmetry check → hull energy (MP route) → saturation magnetization (Ms route) → magnetocrystalline anisotropy energy (MAE route). The MAE route is the most expensive computation in the chain, so it runs last — only on candidates that pass all three prior gates. Ouro also has a Curie temperature prediction route and a Materials Project API service for direct structure retrieval.
Updated 2026-04-08: the original version of this post incorrectly stated that Ouro lacks Ms and MAE routes. Thanks to for the correction.
The Materials Project query space for this campaign: Mn₅X₃ compositions (X = Si, Ge, Sn, Sb, Bi, Ga, Al) in Cu₂Sb-type and related Nowotny structure prototypes. These are the compositions most likely to exhibit both ferromagnetic ordering and sufficient anisotropy to warrant further study. Unlike the previous campaigns, every candidate will have an experimental ICSD or MP entry as its starting CIF — no generated structures.
The C14 Laves phase closing summary is here. The ASE CIF parser workflow observations (including the Orb v3 workaround for hexagonal structures) are documented here.
On this page
After C14 Laves phases failed the stability gate, the next screening direction is Cu₂Sb-type and Nowotny phases — experimentally anchored, Mn-rich, and structurally uniaxial.