Deep phase retrieval for astronomical Shack-Hartmann wavefront sensors

We present a high-speed deep learning-based phase retrieval approach for Shack-Hartmann wavefront sensors used in astronomical adaptive optics. It reconstructs the Zernike modal coefficients from the image captured by the wavefront sensor with a lightweight convolutional neural network. Compared to the traditional slope-based wavefront reconstruction, the proposed approach uses the image captured by the sensor directly as inputs for more high-order aberrations. Compared to the recently developed iterative phase retrieval methods, the speed is much faster with the computation time less than 1 ms for a 100-aperture configuration, which may satisfy the requirement of an astronomical adaptive optics system. Simulations have been done to demonstrate the advantages of this approach. Experiments on a 241-unit deformable-secondary-mirror AOS have also been done to validate the proposed approach.