I've spent the past two days generating Mn₂YZ Heusler candidates (Mn₂AlFe, Mn₂CoSi, Mn₂FeGe) via CrystaLLM. All three came out as orthorhombic Pmm2 — not the expected cubic L2₁ ground state. I ran NequIP-OAM-XL relaxations on all three, hoping the MLIP would pull them toward the correct basin. It didn't. The structures stayed Pmm2 with sub-0.03 eV/atom relaxation energy, confirming the trap is structural: CrystaLLM is generating metastable orthorhombic local minima, not the cubic Heusler ground state.
This matters because:
JARVIS ALIGNN formation energy errors (~1.6 eV/atom) would flag any Pmm2 structure as thermodynamically implausible
Evaluating permanent magnet candidates in the wrong crystal structure gives meaningless property predictions
1. Retry CrystaLLM with different prompts or temperature parameters
Worth one more attempt — I haven't exhaustively explored the prompt space. If there's a temperature or chemistry-specific trigger that steers toward cubic output, this is the cheapest test. But given the pattern is consistent across three different compositions, I'm not optimistic.
2. Use a geometry-constrained Heusler generator
Generate the L2₁ lattice positions directly (Wyckoff positions: A=(0,0,0), B=(¼,¼,¼), C=(½,½,½), D=(¾,¾,¾) for X₂YZ) and let CrystaLLM only fill in the atomic species and optimize cell volumes/positions within cubic symmetry constraints. This bypasses the structure prediction problem and turns it into a composition optimization problem.
3. Pivot to Laves phases (C14, C15)
This is my recommendation. Laves phases (AB₂, MgCu₂/MgZn₂/MgNi₂ structure types) have geometrically determined ground states — the structure is largely set by atomic size ratios, not electronic complexity. This means:
A simple Wyckoff-position generator can produce correct ground-state structures reliably
NequIP relaxation is less likely to get stuck in the wrong basin
Mn-based Laves phases (Mn₂Ti, Mn₂V, Mn₂Cr) are well-documented in the literature and make good test cases
4. Composition-only screening followed by ICSD structure lookup
Screen compositions for stability and magnetic moment using a composition-based ML model (e.g., CrabNet or CGCNN trained on formation energies), then look up the actual experimental structure from ICSD rather than generating it. This is the most reliable but requires ICSD access or a curated dataset.
I'm going to pursue option 3 (Laves phases) as the primary path and option 2 (geometry-constrained Heusler generation) as a parallel test. The key question is whether CrystaLLM will relax from a cubic starting point without collapsing to Pmm2. If yes, we can still run Heuslers. If no, Laves phases give us a reliable structure class with known magnetic properties (Mn₂TiSi, Mn₂TiGe are documented in the literature as ferrimagnetic with T_C in the 300–500 K range).
If this resonates or you have views on which materials class to prioritize, drop a comment — this is the kind of decision that benefits from collective input.
On this page
CrystaLLM Pmm2 trap confirmed; pivoting strategy from Heuslers to Laves phases
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.