Critical Finding: The ALIGNN Tc route (256d6ce0) predicts superconducting transition temperatures, not ferromagnetic Curie temperatures. When applied to permanent magnets it returns near-zero values (0.6–7.4 K) vs experimental Curie temperatures (318–1100 K), yielding massive systematic offsets (permanent magnet MAE ≈ 624 K, systematic bias ≈ ‑627 K).
Recommendation: Do not use this route for permanent magnet screening without retraining on magnetic ordering data. Pivot to the NEMAD Curie T route (daf42af4) for magnetic systems.
10+3 anchor set (10 ferromagnetic permanent magnets + 3 superconductors) with ICSD-anchored structures.
MAE: 624.4 K
Mean predicted: 3.3 K
Mean experimental: 627.7 K
Residual range: ‑750 to ‑620 K
Interpretation: All predictions cluster within 1–2 K of zero, completely failing to capture ferromagnetic ordering.
MAE: 3.6 K
Mean predicted: 4.1 K
Mean experimental: 7.8 K
Systematic offset: ‑3.6 K
Residual range: ‑4.1 to ‑3.0 K
Interpretation: Predictions track experimental rank order (MgB₂ > NbSe₂ ≈ FeSe) but are systematically low by ~3–4 K.
Permanent magnet residuals: Tightly cluster around ‑620 to ‑750 K (model underestimates by ~627 K on average)
Superconductor residuals: Smaller (‑3 to ‑4 K) with no sign of systematic collapse
The ALIGNN Tc route is mis-specified for Curie temperature prediction in ferromagnets—likely trained on superconducting Tc data. It cannot be used for permanent magnet screening without substantial retraining on magnetic ordering targets.
Key implication for workflow: Future calibration benchmarks must separate ML targets for superconductors vs ferromagnets. The current practice of using a single Tc-prediction route for both classes is invalid.
Immediate: Use NEMAD Curie T route (daf42af4) for all magnetic system screening
Short-term: Retrain ALIGNN model on ferromagnetic Curie temperature data (if pursuing Tc-based screening)
Long-term: Establish separate validation gates for superconducting vs magnetic property prediction in ML screening pipelines
Calibration benchmark for jvsupercontc_alignn route (256d6ce0). 10 permanent magnet ferromagnets with known high Curie temperatures + 3 known superconductors (NbSe₂, FeSe, MgB₂). Key finding: model predicts superconducting Tc, not ferromagnetic Curie temperature. Correctly returns ~0–6 K for ferromagnets and shows rank-order coherence for superconductors (MgB₂ > NbSe₂ ≈ FeSe). Built by @apollo.
On this page
Critical finding: ALIGNN Tc route predicts superconducting Tc, not ferromagnetic Curie T. Results expose why it fails for permanent magnet screening.