Improving training data for magnetic property prediction
Materials Project magnetic moments come from DFT+U, with the Hubbard U fitted to reproduce formation energies, not moments. Near a spin-state crossover the moment turns on that choice of U. A different but equally defensible value flips the ion between a high-spin and a low-spin solution, and the database records one of them with no sign that the other exists. I wanted to ask whether models trained on those labels inherit the ambiguity, and whether theory can flag the fragile labels before a single calculation is run. My first hypothesis, d-electron count, was wrong. The one that held is about the ligand field rather than the count. What matters is whether, at an ion's actual charge state, it competes with Hund's coupling. Co³⁺ reaches that regime, Co²⁺ does not, despite sharing d⁷ with Ni³⁺.
Then, trying to keep up with the state of the art interpretability research in large networks, I decided to open the model up to understand whether CHGNet already represents what it gets wrong. It does, though the circuit I first found for it, dissolved under a control I should have run at the start.
Read the report- −0.34skill on the worst held-out physics stratum, where a random split reports +0.12 and hides it
- δ = 0.49Cliff's delta separating Co³⁺ from Co²⁺ errors in CHGNet, across 360 materials
- 66 meVgap between two self-consistent spin states in Li₃CoNi₃O₈, seeded deliberately
- 0.907AUC that element fractions alone reach on the feature I had built a circuit around, and the whole network adds only 0.009