Wavefront Estimation from a Shack-Hartmann Wavefront Sensor Using a Convolutional Neural Network

Images of astronomical objects and orbiting satellites captured by ground-based telescopes often suffer from blurring caused by atmospheric turbulence. Adaptive optics (AO) systems mitigate this issue by employing wavefront sensors to detect distortions and deformable mirrors to correct the distortions in real-time. The Shack-Hartmann wavefront sensor (SHWFS) is particularly effective, consisting of an array of identical lenslets positioned in the telescope's pupil plane. Each lenslet forms a low-resolution image of the incoming wavefront, resulting in a grid of spots corresponding to different subapertures. Traditionally, wavefront aberrations in the SHWFS are estimated using the Least Squares (LS) method, relying on matrix operations. However, with advancements in machine learning, this paper explores the application of a convolutional neural network (CNN) to predict wavefront aberrations directly from noiseless SHWFS data, expressed in terms of Zernike coefficients. Numerical results show that the CNN not only outperforms the traditional LS method in wavefront reconstruction but also demonstrates resilience to some levels of atmospheric turbulence and noise.