Smart design of thermoplastic vulcanizate products: Linking process to performance via interpretable machine learning

Abstract

Thermoplastic vulcanizates (TPVs) are promising materials for various industrial applications due to their lightweight nature, recyclability, design flexibility, and ease of injection molding. A key challenge in designing TPV-based components is controlling and predicting the local stress and plastic strains induced during processing. This study examines the application of machine learning (ML) to identify key injection molding parameters that influence TPV mechanical properties and to predict stress–strain behavior for optimized component design. A full factorial design of experiments (DOE) was conducted to produce 32 TPV plaques under varying conditions. Dumbbell-shaped specimens were extracted from the middle and end regions of each plaque in both transverse and longitudinal flow directions. Cyclic tensile tests were performed to measure two target properties: the stress at 30% strain and the plastic strain, or permanent set, after unloading from the 30% strain cycle, yielding a dataset of 128 samples. ML models of varying complexity and interpretability (decision trees, random forests, gradient boosting, and neural networks) were rigorously evaluated using cross-validation and group-splitting. Given multicollinearity among features, optimal feature sets were selected, simultaneously maximizing accuracy and interpretability. While neural networks achieved the highest predictive performance, decision trees provided full interpretability, a valuable trade-off for industrial adoption. Feature importances and SHapley Additive exPlanations (SHAP) revealed that dosage volume and specimen orientation with respect to the flow direction are the most influential parameters. These findings highlight the potential of data-driven approaches for linking processing parameters to mechanical properties, enabling more efficient TPV-based component design and development strategies.