Reconstructing bulk residual stress in 2XXX aluminum-based thick plates from contour method measurements using attention U-Net

• This work developed an Attention U-Net neural network framework for rapid reconstruction of bulk residual stress in 2XXX aluminum-based thick plates. • The proposed method enables high-fidelity reconstruction of bulk residual stress within seconds, based on contour method measurement data. • The attention gating mechanism effectively enhances the accuracy of the reconstructed residual stress in both magnitude and spatial distribution pattern. • The model demonstrates strong robustness and generalization across plates with different geometries and heat-treatment conditions. The thick plates made of 2xxx aluminum alloy and corresponding composite materials are endowed with strong bulk residual stress when subjected to heat treatment. The traditional method uses finite element (FE) models to predict bulk residual stress, then selects specific cross-sections for comparison and calibration with contour method measurements. This process is time-consuming and requires extensive testing of thermophysical parameters. In this work, an Attention U-Net model is employed to directly reconstruct the bulk residual stress from the residual stress data measured from representative sections via contour method. Our findings demonstrate that the Attention U-Net model can accurately reconstruct the bulk residual stress. On the test set, it achieves a mean absolute error (MAE) of 1.443 MPa and a structure similarity index (SSIM) of 0.910. The model delivers accurate results within seconds, substantially reducing the computational time compared to traditional FE methods. Moreover, the model maintains high accuracy for thick plates with different geometries, with an MAE of 1.70 MPa and an SSIM of 0.90 in cross-geometry reconstruction tests. Providing an efficient framework for rapid assessment of bulk residual stress under different heat treatment parameters, accelerating low-stress process design for aerospace components.