Composition and Processing Descriptors Govern Electroceramic Property Prediction: A Multi-Modal Interpretable Framework for BaTiO3-Based Ceramics

ABSTRACT Establishing interpretable relationships among composition, processing, and properties in BaTiO 3 -based electroceramics remains a central challenge in functional materials design, where dielectric and piezoelectric performance depend on the complex interplay among alloy chemistry, synthesis conditions, and microstructure. This study presents a multi-modal, explainable learning framework integrating experimental measurements, density functional theory (DFT)-derived compositional descriptors, and literature-extracted synthesis conditions to predict and interpret the piezoelectric coefficient ( d 33 ), dielectric constant ( ɛ r ), and bulk density of BaTiO 3 -based ceramics. A consensus-based feature selection strategy combining Mutual Information, LightGBM, XGBoost, and SHAP reduces 24 candidate descriptors to 15 physically meaningful features, revealing that alloying chemistry and compositional identity are the dominant determinants of dielectric and piezoelectric response, consistent with established structure–property relationships in ferroelectric ceramics. A two-level stacking ensemble with a Ridge meta-learner adaptively weights modality contributions, achieving R 2 = 0.874 ± 0.015 for d 33 , 0.845 ± 0.024 for ɛ r , and 0.687 ± 0.127 for density, while reducing prediction variance by up to 30.4% over naïve feature concatenation. The results confirm the primacy of experimental processing descriptors, demonstrate a complementary contribution of NLP-extracted thin-film synthesis conditions to piezoelectric prediction, and establish a generalizable, interpretable pipeline extensible to broader perovskite functional material families.