Metals: CHGNet MAE 0.20 μB. Ligand oxides: mCGCNN competitive (wins SrRuO3/CrO2/EuO). Not a wrapper bug — domain mismatch.
Available
Two magnetic-moment routes on the same CIFs, split into chemistries where each model should win:
Predict total magnetic moment (μB) — mCGCNN (Mal & Bhattacharjee, 2026): scalar DFT total cell moment
Comparable quantity: total magnetic moment per unit cell (μB).
Side-by-side total magnetic moment (μB/cell): mCGCNN vs CHGNet on metallic permanent magnets (ABACUS DFT) and ligand-bridged oxides (Materials Project / mCGCNN sample labels).
Subset | Ref. | CHGNet MAE | mCGCNN MAE | Winner |
|---|---|---|---|---|
Metallic magnets (5 w/ DFT) | ABACUS | 0.20 μB | 3.15 μB | CHGNet |
Ligand oxides / nitrides (8) | MP / sample labels | 1.08 μB | 1.01 μB | mCGCNN (tie-break) |
Independent MP oxides only (5) | Materials Project | 1.04 μB | 1.32 μB | CHGNet MAE, but mCGCNN wins 3/5 |
mCGCNN is not broken — it is just the wrong tool for elemental Fe/Co. On ligand-bridged magnets it is competitive and sometimes clearly better (SrRuO₃, CrO₂, EuO).
Same six known-magnet CIFs from Ouro DFT on known magnets. Fe DFT was on the 1-atom primitive and scaled ×2 to the 2-atom conventional CIF. SmCo₅ has no DFT (missing Sm Dojo NC-FR PP).
Material | Atoms | DFT | CHGNet | mCGCNN | CHGNet |err| | mCGCNN |err| |
|---|---|---|---|---|---|---|
Fe bcc | 2 | 4.68 | 5.01 | 9.34 | 0.33 | 4.66 |
Co hcp | 2 | 3.09 | 3.35 | 7.14 | 0.26 | 4.05 |
FeCo B2 | 2 | 4.45 | 4.50 | 9.12 | 0.05 | 4.67 |
FePt L1₀ | 2 | 3.33 | 3.12 | 4.34 | 0.21 | 1.01 |
MnBi | 4 | 7.28 | 7.10 | 8.63 | 0.18 | 1.35 |
SmCo₅ | 6 | — | 7.80 | 8.13 | — | — |
MAE vs DFT (5 materials): CHGNet 0.20 μB, mCGCNN 3.15 μB.
CHGNet also returns (MAE vs DFT 0.09 T on these magnets). mCGCNN does not emit Ms.
These are closer to mCGCNN’s training domain (ternary/quaternary compounds with O/N bridges and GKA-style M–X–M geometry). References are Materials Project total_magnetization for five independent entries, plus three labels from the mCGCNN repo sample set (flagged — may overlap training).
Material | Atoms | Ref. | CHGNet | mCGCNN | CHGNet |err| | mCGCNN |err| |
|---|---|---|---|---|---|---|
CrO₂ (mp-19177) | 6 | 4.00 | 4.99 | 4.71 | 0.99 | 0.71 |
Fe₃O₄ (mp-19306) | 14 | 28.00 | 28.18 | 26.23 | 0.18 | 1.77 |
EuO (mp-21394) | 2 | 7.00 | 6.62 | 6.82 | 0.37 | 0.18 |
SrRuO₃ (mp-22390) | 20 | 8.00 | 5.08 | 8.33 | 2.93 | 0.33 |
CoFe₂O₄ (mp-753222) | 14 | 26.00 | 25.29 | 22.38 | 0.71 | 3.62 |
CaFeO₃† | 20 | 16.00 | 16.68 | 15.30 | 0.68 | 0.70 |
NdTiFe₁₁N† | 14 | 23.78 | 25.62 | 23.63 | 1.83 | 0.16 |
GdO† | 2 | 6.04 | 6.97 | 5.40 | 0.93 | 0.65 |
† mCGCNN sample-dataset label — treat as soft evidence.
Oxide MAE (all 8): mCGCNN 1.01 μB, CHGNet 1.08 μB.
Independent MP only (5): CHGNet 1.04, mCGCNN 1.32 — but mCGCNN wins head-to-head on CrO₂, EuO, and especially SrRuO₃ (0.33 vs 2.93 μB).
Same scalar target. mCGCNN regresses DFT total cell moment (μB). It does not predict moment direction / spin vectors. FM vs AFM is a separate classification head in the paper.
Paper MAE is already ~2 μB. Holdout test MAE is 2.02 μB vs CGCNN 2.54. Expecting CHGNet-level (~0.2 μB) accuracy on metals was never realistic.
Inductive bias is ligand exchange. The magnetic stream encodes M–X–M angles (GKA). Elemental Fe/Co have empty ligand sets → OOD. Fe/Co/FeCo predictions (7–9 μB) also sit near the checkpoint normalizer mean (8.08 μB).
Wrapper is fine. Authors’ preprocess.py + inference.py on our Fe CIF reproduces 9.342 μB exactly. Primitive 1-atom Fe still predicts ~4.93 μB (~2.3× MP/DFT), so the ~2× Fe/Co pattern is the model, not a conventional-cell bug.
CHGNet is a different animal. Site-projected MLIP moments, FM-aligned for Ms — excellent on itinerant metals, weaker when the MP total-moment label encodes a different magnetic solution than FM-aligned local sums (SrRuO₃ is the clearest example here).
SrRuO₃ (largest mCGCNN win):
Predict total magnetic moment per unit cell (μB) and Ms / μ₀ Ms from a CIF. Use for ligand-bridged magnets — oxides, nitrides, and other M–X–M exchange systems where Goodenough–Kanamori–Anderson geometry matters. Prefer CHGNet (or similar) for elemental metals and alloys without bridging ligands.
Infer per-site magnetic moments with CHGNet and estimate saturation magnetization assuming collinear ferromagnetic alignment of those local moments. Outputs Site moments (µB) with element labels Net vs absolute cell/formula-unit moments (near-zero net + large absolute ⇒ AFM/FiM-like cancellation) Estimated Ms / Js in A/m, T (µ₀ Ms), emu/cm³, emu/g, and µB/ų This is a fast local-moment screen, not a magnetic-ordering solver. Pair with Curie-temperature prediction for a fuller magnet dossier.
Fe (largest mCGCNN miss):
Predict total magnetic moment per unit cell (μB) and Ms / μ₀ Ms from a CIF. Use for ligand-bridged magnets — oxides, nitrides, and other M–X–M exchange systems where Goodenough–Kanamori–Anderson geometry matters. Prefer CHGNet (or similar) for elemental metals and alloys without bridging ligands.
Infer per-site magnetic moments with CHGNet and estimate saturation magnetization assuming collinear ferromagnetic alignment of those local moments. Outputs Site moments (µB) with element labels Net vs absolute cell/formula-unit moments (near-zero net + large absolute ⇒ AFM/FiM-like cancellation) Estimated Ms / Js in A/m, T (µ₀ Ms), emu/cm³, emu/g, and µB/ų This is a fast local-moment screen, not a magnetic-ordering solver. Pair with Curie-temperature prediction for a fuller magnet dossier.