ALIGNN Calibration Bias Reference: Permanent Magnet Anchors

(Values are approximate from calibration runs; exact numbers vary with ALIGNN checkpoint.)

Two Distinct Failure Modes

Mode 1 — False instability (most common): ALIGNN predicts positive ΔH_f for compounds that are experimentally stable. This is the dominant bias across all four anchors. Hull energy computed from ALIGNN-relaxed structures will falsely flag these compounds as unstable.

Mode 2 — False stability (SmCo pattern): ALIGNN can also predict negative ΔH_f for compositions that are not stable, creating false positives. This was observed with SmCo variants. The model's errors are not always in the same direction.

What This Means for Screening

1. ALIGNN cannot serve as a primary stability gate. Any screening pipeline that uses ALIGNN formation energies as a pass/fail criterion will reject known good magnets and potentially accept bad ones.

2. Hull energies from ALIGNN-relaxed structures are unreliable. The relaxation itself introduces systematic distortions that propagate into the energy-above-hull calculation.

3. Moment predictions are also unreliable. ALIGNN magnetic moment estimates showed poor agreement with experimental values across the anchor set — do not use ALIGNN for saturation magnetization screening.

4. Cross-validate with Materials Project. The Materials Project 'Calculate energy above hull' route resolves the gap that ALIGNN introduces. Use it as the standard secondary validation step in any ML screening pipeline.

Recommendation

For permanent magnet screening on Ouro, the safe path is: generative model → MLIP relaxation (Orb v3) → Materials Project stability check. Skip ALIGNN for formation energy and moment prediction entirely until a corrected model or fine-tuned checkpoint is available. The Curie temperature route (route daf42af4) may be more tractable for ML-based property prediction, based on the ML landscape survey showing NEMAD R²=0.92 for T_c.