Deep learning-assisted finite element simulation for characterizing the microstructure-dependent stress field of HTPB propellant

This paper proposes an end-to-end deep learning framework based on the U-Net architecture to directly map the microstructure of Hydroxyl-terminated polybutadiene (HTPB) propellant to its mesoscopic von Mises stress field. A comprehensive dataset of representative volume element (RVE) images paired with corresponding stress fields is generated using the random sequential adsorption algorithm, covering a range of particle volume fractions and spatial distributions, and is augmented through various transformations to improve model generalization. The U-Net model, with its encoder–decoder structure and skip connections, effectively captures fine microstructural details and predicts stress fields under both linear and nonlinear tensile loading conditions. Experimental results show that the model achieves high prediction accuracy and demonstrates excellent extrapolation capability for untrained volume fractions. Compared to traditional finite element simulations that require hours per RVE, the proposed approach reduces computation time to seconds while maintaining accuracy, offering a powerful tool for the digital design and mesostructural optimization of HTPB propellants.