Predicting first-principles Hubbard parameters with equivariant deep learning

Density-functional theory (DFT) extended with Hubbard corrections (DFT+U+V) [1] has shown many successes in predicting the properties of materials containing localized d or f electrons [2], where standard semi-local DFT functionals fail largely due to large self-interaction errors. However, its predictive power relies on having the correct on-site U and inter-site V Hubbard parameters, which should be computed consistently fully from first principles [3]. We have shown that equivariant models using electronic-structure features can predict these Hubbard parameters as obtained from density-functional perturbation theory (DFPT) with almost no loss of accuracy in derived properties at a 100-fold reduction in computational costs [4]. Here, we present an extended approach which leverages both atomic-structure and electronic-structure descriptors for superior generalizability, bringing high-throughput materials design and discovery with first-principles Hubbard functionals within reach.

[1] V. L. Campo and M. Coccocioni, J. Phys. Cond. Matter 22.5, 055602 (2010)

[2] I. Timrov et al., PRX Energy 1, 033003 (2022)

[3] I. Timrov, N. Marzari, and M. Coccocioni, Phys. Rev. B 98, 086127 (2018)

[4] M. Uhrin et al., arXiV:2406.02457 (2024)