Research on residual stress field prediction for multi-riveted samples based on the U-Net network

Riveted joints are widely used in aircraft structures, where residual stresses around rivet holes strongly affect fatigue performance and structural safety. Finite element analysis can provide detailed residual stress fields, but its high computational cost restricts rapid design evaluation and parametric optimization. This study develops a finite-element-based dual-channel U-Net surrogate model for fast residual stress-field prediction in multi-riveted structures. Finite element simulations were conducted for representative rivet configurations and verified using analytical residual stress solutions. The simulation data were converted into structured image-based stress fields, and binary geometric maps and signed distance maps were used as dual-channel inputs to embed rivet layout information and distance-related physical priors. The model was trained using a composite loss function combining mean squared error and structural similarity. Results show that the proposed model achieves stable convergence and outperforms single-channel variants. On unseen testing data, it obtained R ² = 0.888, a Pearson correlation coefficient of 0.944, and an SSIM of 0.934. Compared with representative PINN and GNN models, the U-Net model achieved a better balance between numerical accuracy and structural consistency. The trained surrogate model reduced prediction time from hours to seconds, improving computational efficiency by more than three orders of magnitude. This framework provides an efficient tool for rapid residual stress prediction and design evaluation of riveted aerospace structures.