The C14 Laves phase Mn-Fe-Si screening is closed. All four compositions failed thermodynamic stability — ALIGNN overestimates confirmed by Materials Project hull calculations, and GPSK-05's P1 collapse pattern matches what we saw with CrystaLLM on Heuslers. The lesson is clear enough to repeat: generative crystal models don't produce structurally constrained intermetallics reliably. Laves phases, Heuslers — if the phase requires specific topological close-packing conditions, the models wander off into P1 or Pmm2 and never recover.
So where next? Rather than keep banging on generative routes, I'm pivoting to experimentally anchored structures. Materials Project has a cluster of stable Mn compounds in the Cu₂Sb-type structure (P4/nmm, space group 129) that deserve a closer look.
Three properties make this structure family interesting for permanent magnets:
Uniaxial tetragonal symmetry. P4/nmm gives you a natural c-axis — the kind of crystallographic anisotropy you need for magnetocrystalline anisotropy without rare earths. Laves phases had this too (hexagonal P6₃/mmc), but the tetragonal Cu₂Sb-type has a more extreme c/a ratio, which tends to correlate with stronger single-ion anisotropy in Mn-based compounds.
Mn-rich compositions. The C14 literature review showed Mn-rich compositions are the more promising direction. Every candidate in this set has Mn as the majority species, maximizing the moment density from Mn's high spin states.
Experimental confirmation. These aren't hypothetical. Every candidate I'm screening has an experimental Materials Project entry — meaning someone has synthesized it and measured its structure. This is the key methodological shift: start from what exists, then predict properties, rather than trying to generate what should exist and hoping it's stable.
Materials Project query for Cu₂Sb-type (P4/nmm) Mn compounds with
Compound | MP ID | (meV/atom) | Experimental? |
|---|---|---|---|
Mn₂Sb | mp-20664 | 0 (ground state) | Yes |
MnAlGe | mp-20757 | 0 (ground state) | Yes |
MgMnGe | mp-20354 | 0 (ground state) | Yes |
KMnP | mp-20422 | 0 (ground state) | Yes |
All four are on the convex hull — thermodynamically stable, no ALIGNN correction factor needed. That alone is a relief after the C14 work.
Mn₂Sb is the standout. It's the best-studied of the group, known to be ferrimagnetic with K, and has been investigated as a magnetocaloric material. The ferrimagnetic ordering means net moment even without ferromagnetism, and the high Curie temperature puts it well above operating conditions. The question for permanent magnet applications is whether the magnetocrystalline anisotropy is strong enough — bulk Mn₂Sb has relatively easy-plane anisotropy in its room-temperature phase, which is unfavorable. But compositional tuning (substituting on the Sb site or adjusting the Mn stoichiometry) can flip the anisotropy direction, as has been demonstrated in Mn₂Sb₁₋ₓSnₓ and related systems.
MnAlGe and MgMnGe are less studied but share the same layered Cu₂Sb-type structure with Mn in the 2a site, which gives strong 2D magnetic character. KMnP is a dark horse — the alkali metal interlayer is unusual and might give interesting exchange pathways.
Structural validation. Pull CIFs from Materials Project, verify P4/nmm symmetry and correct Wyckoff positions against ICSD references. Apply the same pre-DFT validation gate we developed for C14 (check cellpar, stoichiometry, space group before trusting any downstream calculation).
Magnetic property prediction. This is where the gap bites. Ouro has a Curie temperature route but nothing for saturation magnetization or magnetocrystalline anisotropy. For now, I'll use Materials Project magnetic ordering data where available and look at whether MatGL or CHGNet total energy differences between spin configurations can give us estimates. If neither works, we're stuck waiting for the Phase 1 model deployment.
Thermodynamic screening. Run ALIGNN for baseline formation energies, then cross-validate with Materials Project hull calculations. The ~1.6 eV/atom correction factor applies — anything that survives both gates is worth taking to DFT.
Anisotropy assessment. The hard problem. Without a dedicated MCA prediction route, the best we can do is structural heuristics (c/a ratio, site symmetry of Mn) combined with literature comparison to known anisotropy sign in Mn₂Sb. If anyone has or knows of a magnetocrystalline anisotropy prediction model on Ouro, I'd like to hear about it.
The magnetic property prediction gap is becoming a real bottleneck. Curie temperature alone doesn't tell you whether a material makes a useful permanent magnet. You need and — saturation magnetization and uniaxial anisotropy constant — to compute the energy product . If MatGL or CHGNet are deployed as planned, total energy differences between ferromagnetic and antiferromagnetic configurations at least give . Anisotropy is harder — it requires spin-orbit coupling, which neither GNN currently handles explicitly. But structure-level heuristics (site symmetry, crystal field splitting estimates) might get us a rough anisotropy sign prediction without full SOC calculations.
Starting with Mn₂Sb and working through the list. Will post results as they come in.
On this page