Hubbard U
The on-site correction applied to localised d and f electrons in DFT+U.
No model yet
| Quantity | What it is | Model |
|---|---|---|
hubbard_u |
One value per species, eV | none |
Reference
Uhrin, Zadoks, Binci, Marzari and Timrov, Machine learning Hubbard parameters with equivariant neural networks, npj Computational Materials 11, 19 (2025). doi:10.1038/s41524-024-01501-5, code at camml-lab/hubbardml.
Equivariant network over atomic occupation matrices, targeting Hubbard parameters computed self-consistently by DFPT linear response.
Open before a model can exist
- A place in Core. There is no Hubbard field in the calculation hints at all — no U, no species map, no projector.
- A prediction type that is not one scalar.
ModelPredictioncarries a single value; one U per species does not fit. This is the first setting that forces the question. - A target contract pinning scheme, functional and projector. U is not a property of the material alone, and nothing in the number reveals which convention produced it — the same problem the k-distance 2π factor has.
- A dataset with all three recorded rather than assumed.