Prediction of Properties of Polymer Composite Formulations Using Ensemble Models With Feature Generation

The development of polymer composites remains largely dependent on trial‐and‐error experimentation, despite increasingly complex performance requirements in industrial applications. An interpretable machine‐learning framework is presented for the simultaneous prediction of nine mechanical, thermal, electrical, and physicochemical properties of polymer composites using an industry‐derived dataset comprising 112 commercial products. Unlike most prior studies that focus on a single property or limited material systems, the proposed approach models multiple properties across a broad range of commercially relevant polymers, fillers, and additives. Multiple feature set configurations are systematically evaluated within an ensemble learning framework incorporating Shapley Additive exPlanations (SHAP)‐based feature selection and interpretation, demonstrating that composition‐based feature generation capturing nonlinear interactions yields the most robust predictive performance. The resulting models achieve external test set coefficients of determination ( R 2 ) exceeding 0.90 across all properties. SHAP‐based interpretation confirms physical consistency with established composite mechanisms, while composition–property contour analyses provide practical guidance for identifying formulation trade‐offs among competing properties. The models are implemented in the PolyMIDAS platform, enabling composition‐based property prediction and multi‐property pre‐screening prior to experimental compounding. Overall, this work establishes a practical and interpretable AI‐driven framework that supports rational polymer composite design and reduces reliance on costly iterative experimentation.