Purpose <p>This study aims to develop an efficient, enhanced deep learning model to replace computationally expensive FEM for stress analysis in composite structures.</p> Methods <p>An enhanced deep learning model CBAM-UNet that integrates convolutional block attention modules (CBAM) into the classical UNet architecture is proposed. Despite the complexity of composite materials, this model can accurately predict stress fields under diverse conditions (loads, microstructures), which paves the way for deducing equivalent elastic constants directly from microscopic features. The model is initially pre-trained on data with fixed fiber numbers and material properties, and later adapted to varying fiber quantities and material combinations through transfer learning. </p> Results <p>Results show that CBAM-UNet accurately predicts the stress field under various loading conditions and outperforms both the standard UNet and other existing improved models, while exhibiting excellent generalization capabilities. Compared to traditional FEM, the prediction efficiency is improved by 2 to 3 orders of magnitude.</p> Conclusion <p>The proposed model achieves fast and accurate stress prediction for composite materials, demonstrating strong generalization ability and showing great potential for engineering applications in rapid composite design and analysis.</p>