Noise-robust alignment method of wide field freeform off-axis three-mirror optical systems based on self-attention mechanism

The alignment accuracy of off-axis three-mirror optical systems with freeform surfaces is limited by the complex nonlinear relationship between the misalignments and wavefront aberrations. To improve the alignment performance, a hybrid deep neural network with feature self-attention (FSA-HNet) is proposed in the paper. In the proposed method, the local feature extraction is combined with global correlation modeling to enhance the prediction accuracy, and a fusion loss function and an exponential moving average strategy are introduced to improve the model's robustness. The simulated alignment verification was conducted to evaluate the performance of the proposed method. Within a misalignment range of ±1 mm in the decenter and ±0.1° in the tilt for the secondary and tertiary mirrors, the FSA-HNet method achieves relatively high precision in a single alignment iteration, and after two alignment iterations, the wavefront RMS value stably approaches the design value of 0.123<i>λ</i> (<i>λ</i> = 632.8 nm) under various noise conditions. Compared with the second-order sensitivity matrix method, the alignment accuracy is improved by approximately 22.4%. In the Monte Carlo simulation, 93.5% of the aligned samples achieve a wavefront RMS better than <i>λ</i>/4, and 82% is better than <i>λ</i>/5. These results demonstrate that the FSA-HNet method could significantly improve the alignment accuracy and efficiency, and increase the yield of complex optical systems under disturbed conditions.