A deep learning framework enabling rapid stress field prediction and damage area characterization for double-double laminates

Accurate stress field prediction and failure zone characterization are crucial for the efficient design of laminated composites, particularly when geometric discontinuities such as open holes induce strong stress concentrations. This study introduces a deep learning framework, termed Conditional-Attention U-Net (C-AUNet), for high-fidelity stress field reconstruction and failure zone characterization in double-double (DD) composite laminates. The key innovation lies in embedding the laminate's ABD stiffness matrices, loading force conditions, and geometric configurations directly at the network's latent space. This strategy enables the model to encode mechanical properties without explicitly handling multiple physical parameters. An attention-gate mechanism is further incorporated to suppress irrelevant responses and highlight mechanically relevant features, while a weighted loss function emphasizes regions near hole edges where steep stress gradients and failure initiation occur. Quantitative results show that C-AUNet consistently outperforms the baseline C-UNet across all stress components. The weighted loss function is shown to be effective in enhancing prediction fidelity in stress concentration regions. Attention coefficient analysis reveals a progressive refinement of spatial focus from global to local regions across network layers, aligning with physical stress redistribution mechanisms and enhancing interpretability. The framework demonstrates strong generalization capability for stress prediction under various stacking sequence combinations. C-AUNet achieves substantial gains in different evaluation metrics relative to C-UNet for failure zone prediction while employing TsaiWu and TsaiHill failure criteria. Overall, the proposed method offers a unified, interpretable, and extensible platform for stress field reconstruction and failure analysis in DD composite laminates.