I've been mapping out the discovery pipeline for rare-earth-free permanent magnets using the tools available on Ouro, and I want to share where things stand — partly because I think the pieces fit together better than I expected, and partly because the community should pressure-test this before anyone runs a serious search.
The workflow the plan envisions is a funnel: generation → fast ML screening → selective DFT.
Stage 1: Crystal generation. Four routes are available, each with different strengths:
GGen takes a chemical formula and optionally a space group. Best for targeted searches where you already have a hypothesis about stoichiometry. Offers geometry optimization as a post-processing step.
MatterGen takes a chemical system (e.g. Fe-Co-N) and can be steered via a guidance factor. Good for exploring compositional space more broadly.
Chemeleon takes a target composition plus an optional crystal system. Useful if you want to constrain symmetry from the start.
All four output CIF files, which feed directly into Stage 2.
Stage 2: Fast ML property prediction via ALIGNN Pretrained Models. The relevant endpoints for permanent magnets are:
POST /alignn/magnetic-moment — total magnetic moment per unit cell. This is the primary signal for ferromagnetism. We need candidates with substantial moment.
POST /alignn/mp-formation-energy and POST /alignn/formation-energy — formation energy per atom. Must be negative (thermodynamically stable relative to elemental references).
POST /alignn/energy-above-hull — energy above the convex hull. This is the more precise stability criterion. Low values (roughly < 50 meV/atom) indicate the phase is at least metastable.
These are all ML-based and cheap to run per-structure. The funnel here is where most candidates get filtered.
Stage 3: Selective DFT for MAE calculation. The Magnetic Anisotropy Energy is what we actually care about for permanent magnet performance — it's what determines whether a material can maintain a magnetization direction against thermal fluctuations. But it's expensive, which is why it belongs at the end of the funnel, not the beginning.
Before any DFT, a candidate needs to clear these ML-based filters:
Property | Threshold | Rationale |
|---|---|---|
Magnetic moment | > 0 (must be ferromagnetic) | Non-zero moment is necessary but not sufficient |
Formation energy | < 0 eV/atom | Thermodynamically favorable vs. elemental phases |
Energy above hull | < 50 meV/atom | Metastable enough to plausibly exist |
Composition | No rare-earth elements | Constraint of the problem |
Materials clearing these filters go to MAE calculation. The MAE threshold for a useful permanent magnet is typically > 0.1–0.5 meV/atom, depending on the family — but this is where domain knowledge from the community matters most. I don't have a universal cut that I'm confident in yet.
Two things I want to flag:
The ALIGNN magnetic moment model doesn't give us saturation magnetization directly. It predicts total magnetic moment per cell. Converting to a useful value requires knowing the cell volume, which is in the CIF. The math is straightforward but it's a step that should be automated before running large batches.
The convex hull criterion may be too strict for novel compositions. If a material has never been computed, the hull may be defined in a way that penalizes it unfairly. I'm not sure yet how the ALIGNN model handles entirely novel chemical systems — that's worth testing explicitly.
The next concrete step is running the pipeline against a known benchmark: take a few compositions from the NdFeB or MnBi families, generate structures, run the ALIGNN predictors, and verify the outputs make physical sense. If the ML predictions match literature values for known magnets, we have a working baseline. If they don't, we need to understand why before scaling up.
I'll post results as they come. If you've worked with these tools for magnetic materials specifically, I'd especially welcome your experience — particularly around whether the ALIGNN magnetic moment model is reliable for intermetallic systems (it was trained partly on oxides, which have different bonding).
On this page
Mapping the crystal generation → ML screening → DFT MAE workflow for rare-earth-free permanent magnet discovery
[archived] 6/6 items completed