Deep learning for estimation of Kirkpatrick–Baez mirror alignment errors

Abstract: A deep learning-based automated Kirkpatrick–Baez mirror alignment method is proposed for synchrotron radiation. We trained a convolutional neural network (CNN) on simulated and experimental imaging data of a focusing system. Instead of learning directly from bypass images, we use a scatterer for X-ray modulation and speckle generation for image feature enhancement. The smallest normalized root mean square error on the validation set was 4%. Compared with conventional alignment methods based on motor scanning and analyzer setups, the present method simplified the optical layout and estimated alignment errors using a single-exposure experiment. Single-shot misalignment error estimation only took 0.13 s, significantly outperforming conventional methods. We also demonstrated the effects of the beam quality and pretraining using experimental data. The proposed method exhibited strong robustness, can handle high-precision focusing systems with complex or dynamic wavefront errors, and provides an important basis for intelligent control of future synchrotron radiation beamlines.