Research On Wavefront Reconstruction Methods Based On Deep Learning

Adaptive optics system(AOS)consists of three main components: wavefront sensor,wavefront controller and wavefront corrector.Among them,the wavefront sensor is the "eye" of the AOS,which provides the benchmark for the AOS correction,and its performance determines the "ceiling" height of the AOS.The most commonly used wavefront sensor in AOS is the Shack-Hartmann wavefront sensor(SHWFS),but SHWFS only uses the subaperture slope as the feature vector in the wavefront reconstruction process,which is greatly affected by mode coupling and mode confusion errors.The number of reconfigurable modes is usually 0.7～0.8 times of the subaperture number,which is much lower than its theoretical limit,i.e.,2 times of the subaperture number.In addition to SHWFS,phase-diversity wavefront detection is also one of the key technologies for aberration measurement in AOS.Phase-diversity wavefront sensor(PDWFS)uses an iterative algorithm to establish the link between focal plane images and defocused plane images and aberration wavefront,but the iteration time of the algorithm is long and the operation efficiency is slow.It is difficult to meet the demand for real-time AOS closed-loop control,and is usually used only for correction of AOS non-co-optical static aberration or image post-processing.Therefore,how to design a wavefront reconstruction algorithm to meet the real-time requirements of AOS while improving the wavefront reconstruction accuracy is the current problem to be solved.This topic will explore the possibility of applying deep learning algorithm to wavefront reconstruction to address the limitations of the application of traditional wavefront detection techniques,and improve the accuracy and real-time performance of wavefront reconstruction by using the powerful fitting ability of deep learning algorithm.The main research work and innovation are as follows.1.The research background and significance of wavefront reconstruction methods based on deep learning are briefly analyzed,and the research status of wavefront reconstruction methods based on deep learning is elaborated.The basic principles of convolutional neural network(CNN),methods of inference acceleration and deep learning framework are briefly introduced,and the advantages and difficulties of deep learning applied to wavefront reconstruction are clarified.2.The corresponding 3rd to 15 th order Zernike mode coefficients are reconstructed directly from the focal images and defocused images of PDWFS using Phase-diversity convolutional neural network(PD-CNN),and the inference speed of PD-CNN reaches sub-millisecond level.This method will be expected to break the PD limitations of the application of wavefront detection technique and promote it as a higher performance wavefront sensor for AOS.3.The super-resolution wavefront reconstruction method is implemented using Shack-Hartmann convolutional neural network(SH-CNN),i.e.,the number of reconstructed modes exceeds the number of SHWFS subapertures by a factor of 2.This method fully exploits the composite characteristics of subaperture spot,reduces the mode confusion and mode coupling errors and improves the wavefront reconstruction accuracy.Simulation experiments were conducted based on 100-unit SHWFS,and the 299 th order Zernike mode coefficients were reconstructed from the SHWFS images using SH-CNN,and the inference time also reaches sub-millisecond level.In addition,the performance of SH-CNN in detecting higher-order aberrations was tested on a 1.8m telescope 241-unit Deformable Secondary Mirror(DSM)AOS,and the closed-loop control of the 66 th to 135 th order Zernike modes was achieved to further verify the generalization of SH-CNN.Through the study of this paper,super-resolution SHWFS reconstruction is achieved by using deep neural networks,i.e.,the number of reconstructed modes is larger than the theoretical limit value of the slope-based wavefront reconstruction method,which reduces the mode confusion and mode coupling errors caused by the slope-based method and solves the problem of limited reconstruction accuracy of SHWFS.Second,a deep neural network is used to achieve high-speed phase recovery,which solves the problem that the PD is time-consuming and easily falls into local optimum.In addition,the inference acceleration of the deep neural network models proposed in this paper are performed using NVIDIA’s Tensor RT,and the inference speed after acceleration all reaches sub-millisecond level,which further demonstrates the advantages of deep learning algorithm applying wavefront reconstruction and provides some new solutions for the wavefront detection needs of AOS.