Shack-Hartmann wavefront reconstruction by deep learning neural network for adaptive optics

Abstract

Standard methods of Shack-Hartmann wavefront reconstruction rely on solving a system of linear equations, extracting wavefront estimates from measured wavefront slopes, which are calculated by retrieving centroids from a Shack-Hartmann Wavefront Sensor (SHWFS). As the dimensions of a micro-lens array in the SHWFS increase, the computational cost of processing wavefronts can become increasingly expensive. For applications that require rapid and accurate computations, such as closed-loop adaptive-optic systems, traditional centroiding and the least-squares reconstruction becomes the main bottleneck limiting performance. In this work, we apply a convolutional neural network (CNN) approach to directly reconstruct wavefronts from raw SHWFS measurements, circumventing both bottlenecks. The CNN model utilizes the ResU-Net framework to perform a zonal wavefront reconstruction, and a method for preprocessing the raw data was investigated with the prospect of enhancing the accuracy of this model specifically for the zonal approach to wavefront reconstruction.