Computer-aided alignment method based on deep learning

Abstract

The optical system will always be different from the design value after been processed. It is necessary to adjust the relative positions between the optical elements to improve the imaging quality of the system. However, if the elements are adjusted randomly, the alignment will be inefficient, so the computer-aided alignment method came into being. In this article, for the alignment of large aberration systems, a new fully-connected network computer-aided alignment (Fc-Net CAA) method is proposed. The systems’ wavefront errors (WFEs) are described by the Zernike polynomials which have a huge advantage in describing system aberrations and we proposed a Fc-Net model for predicting systems’ misalignment. The Fc-Net model is trained with the WFEs of thousands of randomly misaligned instances of the lens system that are modeled in the optical design software, so as to establish the relationship between the system aberrations and the amount of misalignment. In this way, the proposed Fc-Net CAA can achieve the computer-aided adjustment process for systems with large aberrations without a complicated iterative process. The off-axis three-mirror system with aspheric surfaces was simulated and adjusted. During the simulation, a single round of adjustment can make the optical system close to the design wave aberration values, and the average of the five field-of-view WFEs is enhanced from 2.4λ (RMS; λ=550nm) to 0.0764 λ (average). The simulation results verify that the improved algorithm can solve the large initial alignment error of the offaxis reflective optical system with aspheric surfaces.