Optimizing formulation of raw materials, polymer type, filler size, and process temperature in wood polymer composites using deep learning-enhanced digital twin

Abstract

An innovative predictive surrogate framework based on deep learning and particle swarm optimisation (PSO) has been used to efficiently integrate wood or other lignocellulosic reinforcements into wood-plastic composites (WPCs). A total of 300 high-throughput data points for the loading level of wood filler, particle size, cellulose/lignin ratio, coupling agent content, polymer matrix type, and processing temperature were selected to train the MLP model. While the model considers 7 core engineering parameters, it expanded these parameters to 24 neurons in the structural input layer, using one-hot encoding instead of the usual label encoding to avoid artificial ordinal bias. To avoid data leakage from the inter-paper experiments, a study-wise cross-validation system (GroupKFold split) was applied using the 42 independent source papers. Based on these rigorous validation parameters, the model shows high stability, with cross-validated coefficients of determination ( R ² ) of 0.94 for tensile strength, 0.92 for flexural strength (MOR), 0.95 for flexural modulus (MOE), and 0.89 for 24-h water absorption. To understand why neural predictions were made, Shapley additive explanations (SHAP) were used to identify the main factors affecting structural integrity: wood loading and CA concentration. Based on this analysis, a multi-objective PSO was applied in the fine design space to solve the problems and set the Pareto-optimal wood incorporation parameter to 45 wt.%. This critical balance maximises wood utilisation while minimising degradation of mechanical properties. Furthermore, the optimisation process has been consistent with the industry standard range of 3.0–3.5 wt.% CA and operating temperature of 185 °C to minimise severe hemicellulose decomposition. This paper will help optimise the use of raw materials and process temperatures in WPC production.