Fig. 1. Architecture and details of SH-U-Transformer.（a）Architecture of SH-U-Transformer（H and W indicate the size of the feature map. C is the channel’s number）；（b）Architecture of the Embedding Layer（k is the size of the convolution kernel，s is the stride of the convolution，and p is the padding added to all four sides of the input）；（c）Architecture of the Transformer Block（the number of attentional heads in the two multi-headed self-attention mechanisms doubles as the number of channels increases，and the initial number of attention heads is 9）.

Fig. 2. SH-U-Transformer downsampling mode

Fig. 3. Simulated optical path diagram for data generation and wavefront reconstruction

Fig. 4. Comparison of the wavefront reconstruction results of the four methods at four different r₀. The calculated wavefront and its residuals use a different color bar than the original wavefront. The RMS error（RMSE）and the PV error for each residual are given.

Fig. 5. Comparison of RMS errors of residual wavefront under different r₀ for four wavefront reconstruction methods.（a）r₀ ranges from 5 cm to 10 cm；（b）r₀ ranges from 10 cm to 15 cm；（c）r₀ ranges from 15 cm to 20 cm；（d）r₀ ranges from 20 cm to 25 cm. Each stage screen contains 100 datasets.

Fig. 6. Results of wavefront reconstruction at different magnitudes.（a）Reconstruction results from the same wavefront at different magnitudes from top to down，the output wavefronts and their residuals use a different color bar than the original wavefront，and RMS error（RMSE）and the PV error for each residual are given；（b）RMS error distributions of reconstructed wavefronts of different magnitudes are obtained when r₀ is in the range of 5~20 cm. Each stage screen contains 100 datasets.