Interpretable machine learning-assisted discovery of BaTiO3-based ceramics with enhanced dielectric constant

Abstract

Barium titanate (BaTiO₃)-based ceramics are widely used as dielectric materials in electronic devices, but their performance optimization remains challenging due to complex composition–property relationships. Herein, based on experimental data, we present an interpretable data-driven approach that integrates machine learning (ML), SHapley Additive exPlanations (SHAP) and efficient global optimization (EGO) to efficiently explore BaTiO₃-based ceramics with enhanced dielectric constant. The SHAP analysis reveals that, for both the A-site and B-site, the most critical factors influencing the dielectric constant are associated with the polarizability of the element and further identifies the favorable feature value ranges for achieving a high dielectric constant. Guided by these insights, we constructed a candidate compositional space and explored it using the EGO algorithm, leading to the identification of several novel and promising BaTiO₃-based compositions. Remarkably, with only a single round of experimental validation, one of the ML-guided compositions achieved a room-temperature dielectric constant exceeding 7050, representing a 32% enhancement over the maximum value in the initial dataset, while maintaining a low dielectric loss below 0.01. This result underscores both the effectiveness and efficiency of the proposed strategy. Moreover, this work not only provides an effective ML-assisted strategy for navigating vast compositional spaces, but also reveals the key factors governing the dielectric constant in BaTiO₃-based materials.