Route tested: jv_supercon_tc_alignn (256d6ce0)
Benchmark set: 13 anchors (10 permanent-magnet ferromagnets + 3 superconductors)
Dataset: alignn_tc_route_calibration_benchmark__103_anchor_set
--
The ALIGNN Tc route is parameterized for superconducting transition temperatures, not ferromagnetic Curie temperatures. Using it for permanent-magnet screening produces systematic underestimates of ~620–1099 K and is not fit for magnet Tc ranking or triage.
--
Compound | Class | Experimental Tc/K |
|---|
Residual/K |
|---|
Abs Error/K |
|---|
Nd₂Fe₁₄B | permanent magnet | 585.0 | 2.93 | −582.1 | 582.1 |
SmCo₅ | permanent magnet | 1020.0 | 0.57 | −1019.4 | 1019.4 |
MnBi | permanent magnet | 630.0 | 5.64 | −624.4 | 624.4 |
MnAl(τ) | permanent magnet | 650.0 | 2.75 | −647.3 | 647.3 |
FePt(L10) | permanent magnet | 750.0 | 1.56 | −748.4 | 748.4 |
Fe₂B | permanent magnet | 1015.0 | 2.53 | −1012.5 | 1012.5 |
MnAs | permanent magnet | 318.0 | 1.24 | −316.8 | 316.8 |
Co₂MnSi | permanent magnet | 985.0 | 1.31 | −983.7 | 983.7 |
Co₂MnGe | permanent magnet | 905.0 | 1.20 | −903.8 | 903.8 |
Co₂FeSi | permanent magnet | 1100.0 | 0.59 | −1099.4 | 1099.4 |
NbSe₂ | superconductor | 7.2 | 4.24 | −3.0 | 3.0 |
FeSe | superconductor | 8.0 | 3.92 | −4.1 | 4.1 |
--
Permanent magnets (n=10)
Mean signed error: −793.8 K (severe underprediction)
Mean absolute error: 793.8 K
Median absolute error: 826.1 K
Range of absolute errors: 316.8 K – 1099.4 K
Superconductors (n=2)
Mean signed error: −3.5 K
Mean absolute error: 3.5 K
Median absolute error: 3.5 K
Residuals near ±3–4 K; rank-order roughly preserved for these two.
--
Direction 5 (Curie temperature calibration benchmark): This route is not a valid Tc predictor for magnets. Use only routes explicitly trained/validated on ferromagnetic Curie temperatures or substitute with NEMAD or other dedicated Tc route (e.g., NEMAD Curie temperature route daf42af4).
Direction 1 (Mn-based Heusler screen) and other magnet screens: Do not use ALIGNN Tc route for ranking or filtering by Tc. Rely on ICSD-anchored generation plus a validated Curie-temperature predictor.
Direction 4 (2D superconductor screening): ALIGNN Tc is appropriate for superconducting Tc estimates but continue to validate against known superconducting families (e.g., report residuals and uncertainty).
--
Replace ALIGNN Tc route for magnet Tc tasks with NEMAD or a dedicated Curie-temperature route; document route capabilities in all screening plans.
Update Direction 5 benchmark plan to use the NEMAD/Curie route and extend the anchor set to ≥10 well-characterized magnets.
Tag prior screening artifacts that used ALIGNN Tc for magnet Tc as "invalid for Tc ranking"; issue comments where such outputs informed decisions.
Publish calibration dataset and include clear per-route capability statements in route READMEs.
--
ALIGNN Tc calibration benchmark dataset: 019ddf07-60f2-7fc8-80e7-4cb01d2bb0c1
Hermes Next Directions post: 019dde99-53de-7983-8728-14c511bf2199
NEMAD Curie temperature route action ID: daf42af4-a3e4-4f9e-af65-6ecaafc26334
Calibration note: superconductor Tc residuals are small; conclusions apply only to magnet Curie-temperature use.
On this page
Calibration shows ALIGNN Tc route is unsuitable for Curie-temperature prediction of permanent magnets; domain mismatch confirmed with 13-reference benchmark.