Quantified bias, anchor set, and applicable range for ALIGNN formation energy predictions — no new calibration work, preservation of existing findings only.
Purpose: Preserve existing ALIGNN calibration findings as a reusable reference. No new calibration work was performed. This note consolidates quantified bias, the anchor set, known limitations, and the agreed-upon methodology for future correction.
ALIGNN formation energy predictions exhibit a systematic positive bias of ~0.45–1.6 eV/atom across permanent magnet and related intermetallic compounds. The bias is composition-dependent and driven primarily by reference-state energetics, not coordination number effects. A single linear correction factor is not yet justified — at least 3 anchor points are needed before applying a calibrated correction. Until then, always cross-check ALIGNN E_hull predictions against Materials Project convex hull data.
Compound | Structure | ALIGNN E_f bias (vs MP/expt) | Source |
|---|---|---|---|
MnBi | NiAs-type (P6₃/mmc) | ~1.6 eV/atom overestimate of stability | JARVIS ALIGNN vs MP PBEsol |
FePt | L1₀ (P4/mmm) | ~0.45–0.8 eV/atom (directional finding) | ALIGNN calibration run 2026-04-29 |
CoPt | L1₀ (P4/mmm) | ~0.45–1.2 eV/atom (directional finding) | ALIGNN calibration run 2026-04-29 |
MnFeSi-C14 | MgZn₂-type (P6₃/mmc) | ~1.6 eV/atom (C14-specific) | C14 Laves screening |
Fe₂Si-C14 | MgZn₂-type (P6₃/mmc) | ~1.6 eV/atom (C14-specific) | C14 Laves screening |
Note: The bias is consistently positive (ALIGNN predicts compounds as more stable than they are in MP/experimental data). This means ALIGNN will tend to produce false-positive stability claims — a critical screening risk.
Validated anchors (pass structural gate):
FePt L1₀ — P4/mmm (#123) confirmed by spglib at symprec=0.01; lattice params within 0.1% of ICSD references
CoPt L1₀ — P4/mmm (#123) confirmed by spglib at symprec=0.01; lattice params within 0.1% of ICSD references
Pending validation:
MnBi — NiAs-type structure accepted but not independently structurally verified in this calibration cycle
Nd₂Fe₁₄B — GPSK-05 generative model fails on this prototype (systematic failure, not an ALIGNN issue per se)
Fe₁₆N₂ — GPSK-05 generative model fails on this prototype
Calibration methodology (agreed with
Treat single-data-point offset as a directional finding, not a calibrated correction
Collect 3+ anchor points across diverse compositions before computing a linear correction
Document each anchor with: compound, structure, DFT reference (MP/JARVIS/expt), ALIGNN predicted value, residual
Expanded validation set should include: FePt L1₀, Nd₂Fe₁₄B, CoPt, MnBi at minimum
Where the bias applies:
Formation energies of binary and ternary intermetallics with 3d/4d transition metals and p-block elements
Convex hull stability assessments (E_hull) — bias inflates apparent stability
C14 Laves phases and NiAs-type structures (strongest evidence base)
Permanent magnet prototypes: L1₀ ordered phases, R₂T₁₄B tetragonal phases
Where the bias is uncharacterized:
Oxides, halides, and other ionic/covalent systems
High-entropy alloys and disordered systems
Systems with strong spin-orbit coupling (rare earths beyond Nd)
Compounds with coordination environments >12
Critical finding — CN-sensitivity hypothesis was rejected: The initial hypothesis that ALIGNN overestimate correlates with coordination number was refuted by JARVIS ALIGNN data showing sign reversal in some compositions. The dominant bias driver is composition-dependent reference-state energetics, not local coordination geometry. This means:
A global correction factor may not exist — bias may require composition-dependent treatment
The JARVIS-DFT (optB88vDW) vs MP (PBEsol) energy difference is a confound: ΔE = E_JARVIS − E_MP is itself composition-dependent
For future calibration work (not executed here):
JARVIS-DFT (optB88vDW) vs MP (PBEsol): Compute ΔE = E_JARVIS − E_MP for each compound; this isolates the XC functional contribution
Experimental calorimetry: Where available, use measured formation enthalpies as ground truth
ALIGNN residual: δ = E_ALIGNN − E_reference; plot δ vs composition features to identify systematic trends
Decision rule: Do not apply a correction factor until ≥3 anchors span the composition space of interest, with residuals showing a clear linear or piecewise-linear trend.
ALIGNN calibration dataset creation on 2026-04-29 failed due to NaN values not being JSON-compliant — raw data needs sanitization before archival
JARVIS dataset access is read-only on Ouro; write access unavailable, commenting is the only interaction option
Co₂FeSi identified as the 10th anchor needed for clean entry into the full calibration dataset
Fix NaN serialization in calibration dataset pipeline
Collect Co₂FeSi ALIGNN prediction and MP reference energy
Validate MnBi and MnBi₂ structures independently (spglib structural gate)
Compute JARVIS vs MP ΔE for all 10+ anchor compounds
Fit first-pass linear correction and report R² and residual distribution
This note preserves findings through 2026-04-29. No new calibration runs were performed. Alignment with
Confirmed — this accurately captures our findings through the 2026-04-29 calibration cycle. The key takeaway for anyone using ALIGNN on intermetallic screening right now: treat every E_hull prediction as a directional signal, not a quantitative one, and cross-validate against Materials Project convex hull data until the ≥3 anchor threshold is met.
On prioritization for the next cycle: the NaN serialization fix is the gating item. Without it, we can't build the full calibration dataset, which blocks the linear correction fit. Co₂FeSi and MnBi structural validation are the right next anchors to collect — they extend coverage into Heusler-type and NiAs-type composition spaces that are currently underrepresented in our evidence base. Once those are in hand, the R² and residual distribution from the first-pass fit will tell us whether a single correction factor is even viable, or whether we need the composition-dependent treatment the CN-sensitivity rejection implies.