Estimation And Prediction Of Atmospheric Turbulence Aberration Based On Deep Learning

Adaptive Optics(AO)is an optical technology for real-time measurement and correction of wavefront distortion,which is widely used in ground-based telescopes,laser communications,and biological imaging et al.However,with the continuous development of AO,there are also some new and challenging issues,such as how to obtain wavefront information with high speed and how to reduce time delay error in AO systems,etc.With the development of machine learning,especially deep learning techniques,some of these complex inverse problems can be solved.There are three main popular directions in deep learning for AO,including estimating distorted wavefront aberration from optical modulation images of wavefront sensor or intensity images,prediction of future wavefront with historical multi-frame wavefront information and reconstruction of high-resolution images from fuzzy images.This paper mainly focuses on the two major branch directions: intelligent wavefront sensing and intelligent wavefront prediction for simulation and experimental research,to provide a feasible solution for the research of intelligent adaptive optics system.The main research contents of this paper are as follows:Firstly,the main working principle and related theory of intelligent wavefront sensing based on phase diversity are described,laying a theoretical foundation for the simulation in the following text.Then,a phase estimation model based on Convolution Neural Network(CNN)model is proposed,and the average residual wavefront Root Mean Square(RMS)after estimation and reconstruction accounts for about 6% of the original wavefront RMS,indicating that the phase information can be effectively estimated by the proposed model.Several factors affecting the results of the phase estimation were analyzed,including the Zernike polynomial order,turbulence intensity,and noise.The analysis results show that when the order of Zernike polynomials increases,the recovered residual wavefront RMS will also increase accordingly,indicating that better estimation results can be obtained by using fewer Zernike polynomials within the allowable range of truncation error;Similarly,with the increase of turbulence intensity,the recovered residual wavefront RMS will also increase slightly.By studying the effect of estimating other turbulence intensity cases under one case of turbulence intensity,it is shown that the models trained under strong turbulence often have better generalization ability and can effectively estimate phase information under weak turbulence conditions;Finally,the robustness of the model is studied by adding Gaussian noise with different signal-to-noise ratios to intensity images.The results indicate that the model trained in the noise-free case still has a good estimation effect for noisy images with a signal-tonoise ratio higher than 10 d B.It shows that the model has a certain degree of robustness.Secondly,The main principles and related theories of intelligent wavefront prediction based on temporal Zernike coefficients are presented,and a single variable prediction model based on deep learning is proposed.This model uses the Long Short Term Memory(LSTM),which is often used for temporal data prediction.Taking the prediction results of the second and 10 th order Zernike coefficients represent tilt and higher order terms,respectively.The prediction results indicate that whether it is a tilt term or a higher order term,the accuracy after prediction has been improved by 70%compared to without prediction.Then,several factors that affect the prediction results are analyzed in detail,including the length of the input sequence,sampling frequency,and turbulence intensity.The analysis results show that as the input length increases,the prediction accuracy will improve,but not linearly.That is,it is possible to choose a shorter input frame number while ensuring the prediction accuracy,so that in actual systems,prediction can be made by collecting less phase information,reducing system requirements,The results show that when the input length is 4,a relatively ideal prediction accuracy can be obtained;The analysis of the sampling frequency shows that with the decrease of the sampling frequency,the prediction accuracy will also decrease;The analysis results of turbulence intensity show that for single order prediction,the turbulence intensity does not affect the prediction results.Finally,based on the conclusions drawn from univariate prediction research,we studied the multivariable prediction to be done in actual systems,proposed three basic models for multivariable prediction based on deep learning,namely,prediction models based on LSTM,CNN,and Transformer.The prediction results of the three models are compared,and the results show that the prediction effect based on CNN model was the best and the prediction time was the shortest,The average residual wavefront RMS predicted by CNN accounts for3.47% of the unpredicted average residual wavefront RMS,indicating that the prediction can effectively reduce the delay error of the system.Finally,an experimental platform based on Hartmann wavefront sensors was built to generate experimental data.The entire experimental system structure was introduced,and the entire experimental process and main functions of the experimental equipment used were described.Experimental data under five turbulence intensities were collected and analyzed briefly,including the continuity,stationarity,wavefront RMS distribution,and power spectra density of the experimental data,indicating the availability of the experimental data.Finally,the CNN model and parameters with good prediction results in the simulation study are used to verify and analyze the experimental results show that the RMS of the predicted residual wavefront is lower than that of the residual wavefront before the prediction under different turbulent situations,indicating that the prediction is effective.