PiDF: Physics-informed descriptor framework for stability screening of multicomponent oxide ABO3 perovskites

Stability assessment of multicomponent oxide ABO3 perovskites is important for computational materials screening, but remains challenging because mixed A-site or B-site occupation enlarges the composition space and introduces local ionic-size disorder and nominal charge imbalance. These effects are not fully described by conventional single-composition geometric descriptors, such as the Goldschmidt tolerance factor. In this work, we propose the Physics-Informed Descriptor Framework (PiDF), a physics-informed descriptor framework for interpretable stability prediction and screening of multicomponent oxide ABO3 perovskites. PiDF extends the classical tolerance factor by using composition-weighted effective ionic radii for mixed crystallographic sites and further incorporates disorder-aware descriptors and a charge-mismatch descriptor to quantify radius dispersion and nominal valence imbalance in mixed-site compositions. Using a deduplicated benchmark of 254 multicomponent oxide ABO3 compositions, PiDF achieves a ROC-AUC of 0.871 ± 0.043 and a PR-AUC of 0.935 ± 0.027 over 30 repeated random splits, outperforming logistic-regression baselines and remaining competitive with stronger nonlinear learners. Ablations further show that radius-disorder descriptors provide the main improvement in predictive performance, while the charge-mismatch descriptor supplies complementary chemical information. These results indicate that PiDF provides a compact and physically interpretable front end for stability screening and candidate prioritization of multicomponent oxide perovskites.