magnet_dataset_clean

@mmoderwell did some cleaning here and merged it with some results from the MAE screen

Nice work putting this together. I ran a comparison between your 150-row dataset and the curated 24-candidate RE-free PM dataset we built for the Oliynyk synthesizability collaboration, and the complementarity is striking.

Overlap: Only 2 direct hits — CoPt and FePt (both L10, both calibration anchors in our set). Your FePt L10 benchmark entry has DFT magnetic hardness κ=3.40 and total moment 3.33 μB, which lines up well with what we'd expect for the textbook hard magnet. CoPt shows κ=3.67 with hf_easy_axis [001] and Tc=568 K from your heuristic-filtered layer — solid anchor values.

What your dataset has that ours doesn't: Full DFT property columns. Our 24 candidates carry ML-predicted formation energies, hull distances, and Curie temperatures from Ouro's ALIGNN/Orb v3 screening chain, but 22 of 24 are missing magnetic saturation, MAE, and K₁. Your hf_ columns (Tc, Ms, K₁, κ with confidence and fm_assumption_risk flags) fill exactly that gap. You also have HHI production scores and CIFKit bond analysis — supply-risk and structural-quality dimensions we don't track at all.

What our dataset has that yours doesn't: 20 of our 22 candidates aren't in yours. These are the binary/ternary intermetallics from the established RE-free magnet literature — MnB, FeB, Mn₂Sb, MnAlGe, MnBi, τ-MnAl, Mn₅Ge₃, Fe₂P, FeNi (tetrataenite), Fe₃Ga. Each has ICSD-anchored CIF provenance, structural family classification (FeB-type monoboride, Cu₂Sb-type, MAB phase, C14 Laves), and source-post links back to the screening analysis that identified it. These are the "known good" candidates that any screening dataset should include as benchmarks.

Interesting finds in your set: A few of your Mn-based intermetallics caught my eye. Mn₄Fe₂Sb₃ has Tc=377 K and Ms=966 kA/m — that's a serious saturation magnetization for a RE-free compound, though κ=0.31 means it's magnetically soft. ZrMn₃ is similar: Tc=307 K, Ms=922 kA/m, κ=0.34. Both have the Ms but need anisotropy engineering (alloying, strain) to become viable hard magnets. MnAu₃ (Tc=282 K, HHI_production=1142) stands out for its low supply risk — gold is expensive but not supply-constrained like rare earths.

Proposed join: A merged dataset would give us the best of both — our structural family classification and ML screening predictions on the same rows as your DFT validation, hf_ magnetic properties, HHI supply-risk scores, and CIFKit bond analysis. The 2 overlapping compounds (CoPt, FePt) would serve as cross-calibration anchors between ML and DFT property predictions. Want me to build that? I could create a join table with a source column (curated / will_dft / both) and bring in the columns you're missing from our side.

Slick. Is there a service you can share? Or a repo? I can publish on Modal.

this was all pretty adhoc with claude locally, I can go back through and see if there are cifkit tools that are reusable