Oxide perovskites with simultaneously large band gaps and high-static dielectric constants are of considerable interest for advanced microelectronics, dielectric devices, and energy storage applications, yet their discovery remains challenging because electronic insulation, lattice polarizability, and thermodynamic accessibility are strongly coupled and often mutually competitive. Here, we develop a physics-guided multitask learning framework for the joint prediction of the band gap and static dielectric response in chemically constrained single-perovskite oxide ABO3 compounds. To ensure data fidelity and physical comparability, the learning space is strictly restricted to simple oxide ABO3 perovskites from the Materials Project, while mixed-fidelity band gaps, heterogeneous dielectric definitions, and chemically inconsistent samples are excluded. The model integrates role-aware A-/B-site descriptors, perovskite-specific geometric and structural features, multitask prediction of Eg, εtotal, εelectronic, and εionic, explicit physical consistency constraints, auxiliary candidate classification, ranking learning, and reliability-aware screening with uncertainty and out-of-distribution control. Under B-site-grouped cross-validation, the framework achieves 97.4% accuracy, Recall of 96.5%, and an F1 score of 96.1%, while maintaining robust transferability on the independent JARVIS validation set. The results show that high-gap/high-k candidates occupy a chemically non-random subspace governed by B-site-centered electronic–lattice coupling, and that physically consistent multitask learning substantially improves both predictive coherence and candidate enrichment. More broadly, this study establishes a data-consistent, physics-constrained, and transferable paradigm for the intelligent discovery of functional oxide dielectrics.
Y et al. (Mon,) studied this question.