Dielectric capacitors are integral to modern energy storage technologies because of their high-power density, rapid charge–discharge capability, and long-term operational stability. However, the dielectric breakdown strength of perovskite-based dielectrics fundamentally limits their energy storage performance, and their experimental evaluation remains labor-intensive and resource-demanding. In this work, we present a data-driven and interpretable machine learning (ML) framework for the prescreening of high dielectric breakdown strength ABO 3 perovskites using a minimal set of physically meaningful descriptors, including electronic (band gap), phononic (phonon cutoff frequency), and structural (lattice parameters) features. Particular emphasis is placed on addressing challenges associated with limited and imbalanced data sets, which are common in dielectric materials research. Advanced resampling strategies are employed to improve the model robustness and predictive reliability under data-scarce conditions. The optimized support vector classifier combined with Synthetic Minority Oversampling Technique (SMOTE) achieves a macro-averaged precision, recall, and F1-score of 0.91, with an overall accuracy of 88%. SHapley Additive exPlanations (SHAP) further provide mechanistic insights into descriptor contributions, highlighting the roles of electronic structure and lattice dynamics in governing dielectric breakdown behavior. Importantly, the proposed framework enables efficient prescreening of candidate materials, reducing experimental trial-and-error and supporting data-driven materials selection. This approach offers a practical pathway toward streamlined material discovery workflows and improved decision-making in the development of high-performance dielectric materials for energy storage capacitor applications.
Sreedev et al. (Thu,) studied this question.