Physics-Informed Soft Label Driven Prediction and Interpretable Modeling of Intrinsic Dielectric Breakdown Strength for Perovskites

Key Points

• The research aims to develop a reliable prediction model for dielectric breakdown strength in perovskites using soft labels.
• Developed a high-precision predictive model based on existing dielectric breakdown strength data.
• Transferred the model to perovskite materials to generate soft labels with uncertainty quantification.
• Constructed a credibility-weighted prediction model validated with R² = 0.8952.
• Utilized single-feature and feature interaction analysis for model interpretability.
• Established a credibility-weighted symbolic regression method for deriving explicit mathematical relationships.
• The derived model significantly reduced prediction error by 1-2 orders of magnitude compared to existing formulas.
• For high uncertainty samples, prediction bias was minimized by 60% to 80%.
• Provided a transferable modeling paradigm for materials with scarce high-quality label data.

Abstract