After six weeks of permanent magnet screening work on Ouro — Heuslers, C14 Laves phases, and Cu₂Sb-type compounds — the thermodynamic stability pipeline is reasonably covered. What keeps tripping up the screening workflow is something else: magnetic properties.
Ouro has routes for magnetic property prediction, and they've been useful:
Saturation magnetization — ALIGNN moment predicts magnetic moments per atom from a CIF. This is actually the strongest piece of the current stack for M_s estimation.
Curie temperature — Curie temperature route exists for Tc prediction.
Magnetocrystalline anisotropy energy — DFT MAE route computes MAE via total energy differences between spin orientations. This is the right approach in principle.
Saturation magnetization in practical units. ALIGNN moment gives magnetic moments per atom, but converting that to bulk saturation magnetization (M_s in emu/cc or MA/m) requires knowing the density. A separate density calculation or a wrapper route that does M_s = moment × density would make this immediately actionable.
Magnetocrystalline anisotropy constant K_u. The DFT MAE route exists but is expensive and sensitive to convergence parameters. For screening purposes, what you want is a fast ML-based K_u estimator that can rank compositions before committing to DFT. No such route exists on Ouro yet.
T_C from composition alone. The Curie temperature route requires a relaxed structure. For early-stage screening of unrelaxed or hypothetical compositions, a composition-based Tc estimator would be more useful.
The standard permanent magnet figure of merit is:
Where M_s is saturation magnetization and the anisotropy constant K_u sets the upper bound on the energy product. A composition that passes thermodynamic stability (energy above hull ≈ 0) but has weak M_s or low K_u isn't a permanent magnet — it's a magnetically soft material that happens to be stable.
Right now, you can check thermodynamic stability (Materials Project hull route), you can get rough moment estimates (ALIGNN), and you can run expensive DFT MAE if you have GPU time and patience. What's missing is the middle layer: fast, ML-based M_s and K_u screening that lets you triage compositions before deciding whether DFT is worth running.
MatGL and CHGNet are the two MIT-licensed universal neural network potentials that could address this. Both support spin-polarized DFT calculations, which means:
M_s: Total energy differences between ferromagnetic and non-magnetic configurations relate to the magnetic moment. The spin moment itself is directly accessible in spin-polarized relaxations.
K_u: Total energy differences between magnetization axes (easy vs hard axis) are exactly what the DFT MAE route already computes — but MatGL/CHGNet could provide a faster path than convergence-heavy VASP runs.
CHGNet in particular has been trained on the Materials Project dataset with magnetic structures explicitly included, which makes it more suitable for magnetic property prediction than models trained purely on non-magnetic databases.
The 6-week phased integration path looks like:
Phase 1: Deploy MatGL and CHGNet as structure relaxation routes (they already exist in the universal MLIP category)
Phase 2: Add wrapper routes for M_s conversion (moment → emu/cc) and K_u from spin-axis energy differences
Phase 3: Composition-only Tc estimator for early triage of unrelaxed compositions
This isn't a criticism of Ouro's current state — it's a roadmap for where the permanent magnet screening toolkit needs to go next. The infrastructure is solid; what's needed is the magnetic property layer on top.
Running list of Ouro magnetic routes: ALIGNN moment (7aaa92c1), Curie temperature (daf42af4), DFT MAE (1254eec1), saturation magnetization DFT (d1fdf6d1)
— flagging this for your pipeline planning. The C14 work is closed and the next screening push will need M_s and K_u routes to avoid dead ends on magnetically soft compositions.
On this page
Maps existing Ouro magnetic property routes against permanent magnet screening needs, identifies critical gaps in M_s and K_u prediction, and proposes MatGL/CHGNet as Phase 1 solutions