Multi-model machine learning framework for predicting material properties of BaTiO3-based ferroelectric ceramics

Predicting multiple functional properties of ferroelectric ceramics simultaneously remains challenging due to complex nonlinear interactions between composition, processing conditions, and microstructural evolution. This study presents a unified machine learning framework for predicting four key properties of BaTiO 3 ceramics: piezoelectric coefficient ( ), dielectric constant ( ), dielectric loss tangent ( ), and density ( ). The framework integrates an in-house experimental dataset from ten independent fabrication runs, augmented to 10,000 instances using Conditional Tabular Generative Adversarial Networks (CTGAN), with statistical fidelity validated through KL divergence and Wasserstein distance analyses. Domain-specific features comprising compositional ratios, dopant-site descriptors, and processing parameters are synthesized to capture nonlinear structure–property relationships. A consensus-based feature selection strategy combining Random Forest importance and SHAP values confirms that all fourteen engineered descriptors carry meaningful predictive information. Seven models comprising three baselines (Linear Regression, Random Forest, SVR) and four gradient boosting algorithms (AdaBoost, XGBoost, LightGBM, CatBoost) are benchmarked using Bayesian hyperparameter optimization and five-fold cross-validation. CatBoost achieves the highest performance across all targets, with of 0.922 for , 0.933 for , 0.855 for , and 0.674 for . Ablation studies confirm that domain-specific feature synthesis contributes approximately 38 percentage points of improvement, establishing a reproducible and interpretable pipeline for multi-property prediction in ferroelectric ceramics.