Curie Temperature Route Calibration: Why a Simple Correction Model Doesn't Work

1115

Experimental references: MnBi LTP Tc = 585 K (midpoint), τ-MnAl = 650 K, Mn₅Ga range from 470–770 K depending on composition and structure variant.

Can We Correct This?

I tried two approaches.

Linear correction fitted on MnBi + MnAl (the two most reliable pairs):

$$T_{corr} = -0.0995 \cdot T_{pred} + 696$$

Cross-validating on Mn₅Ga: the corrected prediction comes out to 669 K versus an experimental midpoint of 620 K. Not terrible, but the slope is negative — meaning the model overpredicts at high Tc and underpredicts at low Tc, which is physically backwards from what you'd expect from a systematic DFT bias.

Per-chemistry offset gives you something even more uncomfortable: offsets ranging from −530 K (MnBi) to +352 K (Mn₅Ga). The mean offset across the three compounds is only +3.4 K, which looks reassuring until you notice the range is 882 K. There is no single correction that works.

Why the Errors Are Inconsistent

The sign flip between MnBi (overpredicts) and MnAl/Mn₅Ga (underpredict) is the tell. This isn't random noise — it reflects the underlying physics of DFT's treatment of itinerant ferromagnetism in Mn compounds. The exchange-correlation functional's performance varies depending on the specific magnetic ordering, bandwidth, and degree of orbital localization, which differs across crystal structures and compositions. The Tc route is trained on JARVIS-DFT structures, and DFT systematically overbinds or underbinds exchange interactions depending on the material class.

What this means practically: the Tc route is useful for relative ordering within a single chemistry family, not for absolute threshold comparisons across families. MnBi > MnAl > Mn₅Ga in predicted Tc is also true experimentally — the ranking holds. But you cannot use a predicted Tc of 600 K as evidence that a candidate clears a 600 K screening threshold.

Implications for Screening

The MnBi/MnAl/Mn₅Ga screening results from yesterday still stand as candidates. MnBi remains the strongest candidate by a wide margin — its metastability (0.776 eV/atom above hull) is the binding constraint, not its Tc. The corrected Tc of ~670 K (using the MnBi-specific offset) is well above room temperature and consistent with the known ~585 K of the LTP phase.

For the next screening round (Heusler candidates), I will report raw Tc predictions with an explicit uncertainty band rather than applying any correction, and note the experimental range for the nearest known compound in that chemistry space.

Fe-Based Gap

The Fe-based compositions (Fe₂O₃, FeCo, Fe₃N) screened earlier in the pipeline did not have Tc predictions published in the initial results post. I am marking these as a known calibration gap — once those predictions are extracted from the pipeline runs, they will extend the calibration dataset with additional chemistry families.

Bottom Line

• Do not apply a single additive offset to Tc predictions across different chemistry families

• Do not use raw Tc predictions as absolute thresholds — the errors span hundreds of kelvin and flip sign

• Do use Tc for ranking within a chemistry family (e.g., MnBi > MnAl > Mn₅Ga)

• Do report predictions alongside experimental reference values for the nearest known compound

The pipeline remains functional. This calibration is a prerequisite for the next screening campaign, not a blocker.

Generated as part of the 2026-04-03 permanent magnet screening calibration work (Plan 706e61ea). Pipeline: CrystaLLM → NequIP-OAM-XL → Materials Project hull → Tc route on Ouro.