Abstract: To address the low efficiency of the traditional finite element method (FEM) in predicting the mesoscopic stress fields of solid propellants—a problem arising from the need for dense meshing and nonlinear iterations in high-volume-fraction models—and the limited applicability of existing deep learning approaches, which are primarily focused on fiber-reinforced composites, to particle-reinforced systems, this paper proposes a convolutional neural network prediction method based on the U-Net architecture. First, an improved random sequential addition algorithm was employed to construct a representative volume element (RVE) model with a particle volume fraction of 65%. A dataset of 450 Mises stress fields was then generated using a bilinear cohesive zone model. The designed U-Net model, utilizing an encoder-decoder architecture with skip connections, achieves end-to-end mapping from microstructure grayscale images to the full-field stress distribution. For a 1.5 mm × 1.5 mm model, the predictions yielded a mean absolute error of approximately 5% of the average stress and a coefficient of determination exceeding 0.85, with a computational efficiency gain of over 90% compared to FEM. For large-scale models, a ‘high-resolution input, window sliding, and weighted fusion of overlapping regions’ strategy was developed. Optimal performance was achieved with a step-to-window ratio of 0.125, reducing the average prediction error to 0.034 MPa. The entire process was completed in 604 seconds, representing an efficiency improvement of more than 95% over the FEM simulation time of 14,100 seconds. This method provides an efficient and accurate solution for mesoscale damage analysis of particle-reinforced composites. Furthermore, the proposed sliding window fusion mechanism offers a valuable framework for predicting large-scale stress fields in heterogeneous materials.