Learning and retrieval for warm-starting charge-self-consistent DFT+DMFT

Abstract

Charge-self-consistent (CSC) DFT+DMFT delivers quantitative correlated-electron physics one configuration at a time, making ensemble sampling dependent on reliable warm starts for its expensive fixed-point iteration. Here we compare retrieval from the most similar converged structure with an E(3)-equivariant model that predicts a physics-structured self-energy and Fermi level. Because the full CSC loop refines both initializations, every production result remains a converged DFT+DMFT solution. Across metallic Fe, correlated FeO, and Mott-insulating NiO, learning reduces the typical iterations to sustained convergence from 8 to 3, 8 to 3, and 6 to 1. Retrieval matches this median speed when a dense same-state archive is available, but several poor transplants reveal that structural similarity does not predict transplant quality. In a pre-registered volume window excluded from training and donor pools, learning retains its speed, whereas retrieval starts $3.4\times$ farther from the fixed point and approaches cold-start cost. Either initialization can select a distinct near-degenerate branch on a rugged CSC landscape. Applied end to end, the workflow generates over one thousand correlated energy and force labels for iron at Earth's-core conditions and trains an equivariant interatomic potential. Solid--liquid coexistence with 9216 atoms gives $T_m = 6225 \pm 42\,\mathrm{K}_{\mathrm{stat}}$ at 330~GPa, consistent with recent experiments. A 50-configuration DFT+DMFT audit resolves the potential's energy calibration but does not justify a corrected melting temperature, because four-atom cells cannot realize a liquid. The resulting regime map favors retrieval within dense coverage, amortized learning at and beyond its boundary, and solver refinement throughout.