Physics-Motivated Machine Learning Prediction of Hubbard Interaction Parameters in Transition-Metal Oxides

Accurate Hubbard parameters are essential for beyond-DFT approaches in correlated materials, yet they are often chosen empirically. The constrained random-phase approximation (cRPA) rigorously evaluates Hubbard interactions, but its computational cost hinders large-scale screening. Here, we present machine-learning (ML) models for predicting cRPA-derived Hubbard parameters: effective on-site Ueff, intersite V, and Hund’s coupling J in transition-metal oxides (TMOs). We combine ensemble-learning models with a regression-based brute-force search (BrFS) to achieve predictive accuracy and analytical expressions. Physically motivated features, including TM-d bandwidth and TM-d/O-p band-center separation, capture localization and screening. Our ensemble models achieve MAEs of 0.187, 0.084, and 0.013 eV for Ueff, V, and J, respectively. The analytical forms relate Ueff to electron localization and TM-d/O-p hybridization, suggest that structural family and compactness influence V, and indicate that J depends on elemental descriptors of the TM ion. This study provides an efficient and physically interpretable framework for predicting cRPA-derived Hubbard parameters.