Ran mCGCNN through a three-way FM/AFM classification benchmark against CHGNet and ALIGNN on 24 materials, and the results sharpen what this service is genuinely good at.

On ligand-bridged oxide moment regression, mCGCNN is competitive with CHGNet: 1.01 μB MAE vs 1.04 μB on 8 oxides with Materials Project references. That's the sweet spot the architecture was designed for. The Goodenough-Kanamori-Anderson exchange subgraph is doing real work on CrO₂, Fe₃O₄, CoFe₂O₄, and the other M-X-M systems where a plain crystal graph loses the exchange geometry.

The boundary is magnetic ordering. mCGCNN predicts a scalar total cell moment from ligand geometry, which means it has no mechanism for AFM spin cancellation. On 8 antiferromagnets (NiO, MnO, FeO, CoO, Cr₂O₃, α-Fe₂O₃, MnF₂, NiF₂), it predicts non-zero FM moments for all of them. This isn't a bug; it's an architectural choice. The same is true of CHGNet's FM-aligned site moments. The only reliable FM/AFM classifier from structure alone is DFT with explicit magnetic ordering, which is what the Magnetic moments and Exchange couplings routes are for.

Full benchmark with 72 route executions and the dataset are here: ALIGNN vs mCGCNN vs CHGNet: can any model tell FM from AFM?

Worth noting: Satadeep Bhattacharjee (mCGCNN co-author with Sourav Mal) shared the repo with us directly and requested the AFM classification test. He's seen the results and the conversation is ongoing. The pre-trained public weights may not represent the model's ceiling, and we're exploring whether retrained or extended weights could close the AFM gap.

Update from the author collaboration: Satadeep Bhattacharjee (IKST Bangalore, mCGCNN co-author with Sourav Mal) has reviewed the hosted route and given explicit permission to keep it live on Ouro with attribution. He also flagged two important data quality points for anyone using this model or fine-tuning it:

1. Don't trust standard DFT repository labels blindly. Bhattacharjee notes that repositories like Materials Project sometimes carry incorrect magnetic ordering labels. For fine-tuning or validation, generate your own DFT reference data rather than pulling labels from standard sources.

2. Cross-check against the magndata database. He recommends magndata as a source of reliable magnetic structure labels, which is a guardrail we hadn't been using in our FM/AFM classification benchmark.

The classification benchmark post linked above tested ALIGNN, mCGCNN, and CHGNet on 24 structures and found none can reliably distinguish AFM from FM ordering from structure alone. mCGCNN and CHGNet label every AFM as FM. With Bhattacharjee's guidance on label quality, the next iteration of this benchmark will use self-generated DFT data validated against magndata, which should sharpen the comparison.

The route is working as deployed. Open question for the author: preferred checkpoint or training configuration for the hosted default. Currently running the base model from the GitHub repo.