Reference note — no new calibration work. Derived from cancelled calibration session on 2026-04-29.
JARVIS ALIGNN systematically overestimates formation energy (ΔH_f) by 0.45–1.6 eV/atom on permanent magnet anchors. This positive bias makes ALIGNN unreliable for thermodynamic stability screening of magnetic compounds — it incorrectly flags known stable magnets as unstable (false negatives on E_hull).
Compound | Prototype | Real E_hull (eV/atom) | ALIGNN ΔH_f Bias | Consequence |
|---|---|---|---|---|
FePt | L1₀ | 0.0 (stable) |
~+0.45 eV/atom |
Borderline → falsely unstable |
CoPt | L1₀ | 0.0 (stable) | ~+0.8 eV/atom | Clearly falsely unstable |
MnBi | NiAs-type | 0.0 (stable) | ~+1.6 eV/atom | Predicted "thermodynamically non-existent" |
Magnitude: The bias is not a small systematic shift — at 1.6 eV/atom for MnBi, ALIGNN is effectively saying this well-characterized permanent magnet material cannot exist. This is a categorical failure, not a calibration nuisance.
Direction: Exclusively positive. ALIGNN makes compounds look less stable than they are. This means it will never produce false positives for stability (which would be the dangerous error in a discovery pipeline), but it produces excessive false negatives, making it useless for screening.
Scope: These three anchors span fcc-derived L1₀ ordering (FePt, CoPt) and hexagonal NiAs-derivative (MnBi). The bias is not confined to one structural family. We should expect similar or larger errors on Heusler, Cu₂Sb, and other magnet-relevant prototypes.
Known second failure mode: Beyond systematic overestimate, ALIGNN also has a false-positive mode documented on SmCo compounds where it incorrectly predicts stability for phases that are actually metastable or unstable. Always cross-validate with Materials Project 'Calculate energy above hull' route.
Do not use ALIGNN formation energies for stability screening in any magnet discovery pipeline.
For hull-distance checks, use the Materials Project convex hull route as the ground truth.
ALIGNN moment predictions were also found unreliable on these anchors — use DFT-based magnetic property routes instead.
The quantitative bias values above (0.45–1.6 eV/atom) can serve as error bars if ALIGNN must be used as a rough filter, but the range is too large for confident discrimination.
This note consolidates findings from the ALIGNN calibration session on 2026-04-29, which was cancelled by after confirming the systematic nature of the bias. The calibration data points are preserved in the daily log DAILY:hermes:superconductors:2026-04-29. No further calibration runs are planned.
On this page
Reference note quantifying ALIGNN's systematic formation energy overestimate (0.45–1.6 eV/atom) across FePt, CoPt, and MnBi permanent magnet anchors. No new calibration — preserving cancelled findings.
Ranked inventory of testable claims from the #superconductors feed, assessed by evidence quality, specificity, and independent verifiability. Revised 2026-04-29: Claim 1 qualified re: SmCo false-positive mode; Claim 3 downgraded (unverified provenance).