Rapid prediction of microscale stress fields in fiber-reinforced composites from limited finite element data using a deep learning framework

Predicting transverse stress response of fiber-reinforced polymer composites at the microstructural level traditionally relies on finite element method (FEM) simulations of representative volume elements, which are computationally expensive and time-consuming. While deep learning offers a faster surrogate, most existing approaches require hundreds to thousands of FEM-generated training samples, creating a significant data bottleneck. The results demonstrate that accurate stress-contour prediction can be achieved using as few as 20 base representative volume element simulations, expanded to 80 training samples through rotation and reflection augmentation. To address this limited dataset and emphasize mechanically critical stress concentration regions, a two-stage U-Net framework with a dual-output head and a custom attention loss based on red–green–blue (RGB) pixel values was developed. The first stage trains the model on exaggerated stress contours to identify regions of peak stress, while the second refines predictions using standard stress contours to recover realistic stress magnitudes. Trained on 80 augmented samples derived from 20 base simulations, the model reproduced FEM-validated stress contours with a mean absolute error (MAE) of 1.366 ± 0.284 MPa and a structural similarity index measure (SSIM) of 0.943 ± 0.006 . Expanding the base dataset to 85, yielding 340 augmented samples, reduced the MAE to 0.872 ± 0.168 MPa and increased the SSIM to 0.967 ± 0.004 . In both cases, inference took less than one second per sample, compared with 1620 s for a single FEM simulation.