Experiment-Driven Performance Prediction of Fused-Filament-Fabricated Shape Memory Polymer Composites With a Data-Divide-Assisted Neural Network

Abstract

Fused filament fabrication (FFF) of shape memory polymer composites (SMPCs) offers potential for aerospace applications requiring mechanical robustness and reliable shape-memory performance. However, performance-oriented process optimization remains challenging. Therefore, this study establishes an experiment-driven process-property database and develops a framework to predict tensile strength, shape fixation ratio, and shape recovery ratio. Tensile and thermomechanical data are obtained through full-factorial and orthogonal experiments, and analysis of variance is used to evaluate the effects of the experimental factors on the corresponding properties. On this basis, a data-divide-assisted genetic algorithm-optimized backpropagation neural network (DDGA-BP) is developed. The model retains the trained network and its training record obtained during genetic optimization and uses the retained network for final prediction, thereby improving prediction stability and reproducibility. To evaluate reliability and stability, DDGA-BP is compared with multiple linear regression, random forest regression, and conventional GA-BP over ten repeated carbon-content-stratified data splits. DDGA-BP achieves the highest mean $R^2$ values of 0.836, 0.942, and 0.931 for the three properties, together with the lowest mean MAE and RMSE and lower variability than conventional GA-BP. The proposed framework enables accurate and reproducible property prediction and supports performance-oriented design of 4D-printed SMPC components.