Method for optical adjustment with deep learning to quantitatively predict misalignment in optics

We developed a method for optical adjustment using a deep learning model to quantitatively predict misalignment of optical components. The proposed model predicts the misalignment parameters using only through-focus images of a point source, while conventional methods require specialized measurements or extensive manual analysis. There is no need for special preparation for measurements, and quantitative prediction will reduce the cost of optical adjustment. A distinctive aspect of our method is that the training dataset is not obtained through measurements but generated using ray-tracing simulation, which produces through-focus images with various type of aberrations. By applying the method to a simple parabolic mirror and a reflecting telescope, we demonstrated its prediction accuracy. The through-focus images obtained from simulated optics, according to the predicted misalignment parameters, matched the measured images. We adjusted two optics and confirmed that the measured images after adjustment were in good agreement with the simulation images of the designed optics.