Machine Learning-Guided Electronic Configuration Design for Dielectrics With High Energy Storage Performance.
TL;DR
An interpretable machine learning framework based on Shapley Additive exPlanations is developed to guide the compositional design of K0.5Na0.5NbO3-based relaxor ferroelectrics, leading to Sc as the optimal dopant, outperforming the Bi5/6In0.5Sn0.5O3-doped system.
Abstract
Dielectric capacitors possess ultrahigh power density and are promising for advanced energy storage and pulsed power applications, yet conventional trial-and-error doping strategies limit design efficiency and performance optimization. Here, an interpretable machine learning framework based on Shapley Additive exPlanations is developed to guide the compositional design of K0.5Na0.5NbO3-based relaxor ferroelectrics. Key descriptors governing polarization behavior are identified from 18 features, leading to Sc as the optimal dopant, outperforming the Bi5/6In0.5Sn0.5O3-doped system. The designed 0.85(0.96K0.48Na0.52NbO3-0.04BaZrO3)-0.15Bi5/6Sc0.5Sn0.5O3 is predicted to exhibit a large polarization difference of 26.3 µC cm-2. First-principles calculations and experimental validation show that Sc3+ forms more ionic Sc-O bond than In3+, giving less distorted and markedly more rigid polar units. These rigid units act as a restoring force during polarization switching and suppress the remnant polarization, while the strong Sc─O bond reduces oxygen vacancy concentration and enhances the dielectric breakdown strength. As a result, an ultrahigh energy storage density of 6.91 J cm-3 and efficiency of 91.9% are achieved, representing a 66.1% improvement over the In-based system. This work demonstrates an interpretable machine learning-guided electronic configuration strategy and highlights the critical role of dopant electronic structure and local bonding in optimizing energy storage performance.