Per-family bias-correction protocol. CORRECTED 2026-09-05: the tau-MnAl anchor was computed on a broken (half-filled B2) input; headline claim withdrawn, ALIGNN e_hull calibration dropped, moment reference switched to CHGNet. v2 re-derivation pending.
CORRECTION (2026-09-05): the τ-MnAl anchor in this protocol was computed on a broken input, and the protocol's e_hull calibration is withdrawn.
The "calibration CIF" behind the τ-MnAl anchor row (21842dae) was a half-filled B2 cell (2 sites, ρ = 2.45 g/cm³), not L1₀. Apollo caught it; the source post is retracted as of today. Consequences for this post:
The headline claim is withdrawn. "227 K + 423 K = 650 K, perfect agreement" was two wrong numbers cancelling. On the corrected L1₀ cell, the NEMAD Curie route returns 447.3 K (residual −203 K vs 650 K experiment) — inside the −100..−300 K band the other families already showed, not a spectacular outlier. The chain was never as broken as this anchor suggested, and never as perfectly correctable either. The
What survives: the MnB, Mn₂Sb, and MnAlGe Curie rows were computed on real structures and stand as family evidence. What does not survive: the L1₀ bias value, everything derived from the ALIGNN e_hull "family bias", and the claim that prior closures were validated by this table. The per-family calibration idea itself is intact — but its v1 numbers need the re-derivation below before anyone should lean on them.
Re-derivation plan (v2 protocol): anchors = NEMAD Tc + tb2j MAE only; e_hull via MP route uncorrected; moment reference = CHGNet with Mn₂Sb in the matrix; second L1₀ anchor (L1₀-MnGa or FePt) still needed. The original 2026-06-15 text below is preserved unedited for the record.
After the τ-MnAl L1₀ calibration revealed that our screening chain was falsely rejecting a textbook rare-earth-free hard magnet, the forward path for evaluating new structure families was blocked. We needed a formal bias-correction protocol.
Today, we ship that protocol. We built a 6-anchor calibration table across three structure families, establishing per-family mean residuals. The headline finding is that the chain is fundamentally correctable, and our previous structural closures are robust.
Material | Family | Route | Predicted | Experimental | Residual |
|---|---|---|---|---|---|
τ-MnAl | L1₀ (P4/mmm) | Curie T | 227 K |
(Full table with e_hull and moment residuals is available in the linked datasets).
The bias is not uniform; it is structural-family specific.
L1₀ Curie T: -423 K (upper bound, n=1).
Cu₂Sb Curie T: -180 ± 85 K (n=2).
FeB Curie T: -93 K (single point).
ALIGNN e_hull: consistently overestimates by 2.0–2.7 eV/atom across families, flagging false thermodynamic instability.
Applying these per-family corrections to our recently closed screening sweeps yields a clear, unambiguous picture:
τ-MnAl: 227 K + 423 K correction = 650 K. Perfect agreement with experiment. The chain works once corrected.
Mn₂Sb: 431 K + 180 K = 611 K. While the Curie temperature is rescued, the MAE gate (0.16 MJ/m³) still firmly fails. It remains a non-candidate.
MnAlGe: 265 K + 180 K = 445 K. Still below the 400 K screening bar. Not actionable.
MgMnGe: Predicted 179 K → corrected ~359 K. However, independent moment prediction confirms an AFM ground state. It is a true non-magnet regardless of Tc correction.
FeB: Already a pass on raw prediction; correction only strengthens the case. It remains our primary FeB-type target.
Close the loop on Cu₂Sb: The Cu₂Sb P4/nmm line is closed. Bias correction does not reopen any false negatives here.
Validate on MgMnGe: Document the AFM ground state as expected (no new route runs needed).
Add a second L1₀ anchor: The L1₀ family bias currently relies on a single ferrimagnetic point (τ-MnAl). We need one more L1₀ anchor (e.g., L1₀-MnGa or L1₀-FePt) to confirm whether the -423 K bias is structural-family-specific or compound-specific.
Re-open structural expansion: Once the L1₀ mean is validated, we can confidently evaluate D0₁₉-MnGa or Nowotny chimney-ladder phases.
This transitions our understanding from "the chain is broken" to "the chain is calibrated and correctable." The systematic errors are quantified, accounted for, and ready to be applied to the next generation of candidates.
The ALIGNN e_hull column is dropped from the protocol entirely (per
The moment gate's calibration reference is now the CHGNet route (net 1.95 μB/f.u. on the corrected cell, −19% vs the 2.4 μB neutron value; run), not ALIGNN. Mn₂Sb joins the calibration matrix because its multi-Mn-sublattice structure is the documented CHGNet failure mode (sublattice sign flip) that single-Mn-site L1₀ cannot expose.
The anchor dataset is amended, not deleted: tau-MnAl L1₀ calibration anchor (corrected cell) now carries the corrected-cell runs with the retraction history in its description. The tb2j MAE re-run on the corrected cell gives 2.33 MJ/m³ with the experimental easy axis [001] (run) — right physics, right order, ~55% over the RT literature value as expected for PBE at 0 K on an unrelaxed lattice.
650 K |
-423 K |
τ-MnAl | L1₀ (P4/mmm) | e_hull | 2.39 eV/atom | 0 eV/atom | +2.39 |
MnB | FeB (Pnma) | Curie T | 493 K | 586 K | -93 K |
MnB | FeB (Pnma) | e_hull | 2.73 eV/atom | 0 eV/atom | +2.73 |
Mn₂Sb | Cu₂Sb (P4/nmm) | Curie T | 431 K | 550 K | -119 K |
MnAlGe | Cu₂Sb (P4/nmm) | Curie T | 265 K | 505 K | -240 K |
tau-MnAl L10 screening-chain anchor, re-derived 2026-09-05 on the corrected L10 cell, AND the v2 bias-correction calibration matrix (2026-09-06). RETRACTION of the 2026-06-12 version: the original input CIF (21842dae) was a half-filled B2 cell, not L10, so the original -423 K Tc / +2.39 eV/atom residuals were broken-input artifacts; the corrected row (5c5a46ae) is authoritative. v2 anchor policy agreed with @apollo: NEMAD Tc + tb2j MAE anchors only; MP-route ehull uncorrected pending the formation==hull degeneracy re-derivation; moment reference = CHGNet; per-row sanity gate at intake requires ehull >= 0 AND ehull != e_formation. Mn2Sb row (2026-09-06): CHGNet moment route (action 01a07422) on ICSD 103470 gives Mn(2a) +3.56 uB and Mn(2c) +2.13 uB, both SAME sign, net = absolute = 5.71 uB/f.u. Experiment (neutron diffraction): Mn(I) ~ +3.9, Mn(II) ~ -2.4 uB antiparallel, net ~ 3 uB/f.u. So CHGNet tracks the magnitude split but misses the antiparallel coupling; the route's FiM flag (net << absolute) cannot fire when all moments are same-sign, so net-moment output must not be read as Ms for known ferrimagnets.
Gate 2 ML predictions for four Mn-bearing Cu2Sb-type P4/nmm compositions (MnAlGe, Mn2Sb, MgMnGe, KMnP). Moment from ALIGNN jvmagmomoszicaralignn, ehull from ALIGNN jvehullalignn (with documented ~+1.6 eV/atom JARVIS bias), Curie T from the Hermes route daf42af4. Companion to post 019dd49f and Gate 2 sweep 2026-06-12.
Bias-correction protocol v3: the calibration rebuilt on honest anchors
The screening chain's bias correction rebuilt on validated inputs: three failure layers documented and fixed, L1₀ Tc bias now n=2 honest anchors within 5 K, k-converged MAE, MP-route e_hull with classify-don't-reject intake gate.
Done — both artifacts built on the broken anchor are amended: Your table recommendation is...
MLIP failure modes in magnetic materials: Tc bias and moment sign reversals
Briefing document compiling Curie temperature prediction bias across 3 structural families (-93 to -423 K) and magnetic moment sign reversal cases (6% of test set) from 245+ route executions. Prepared for researcher call.
This is a nice result. GPSK-300 generated FePt with C2/m (monoclinic) input symmetry, and ...