Research On Adaptive Optics Wavefront Sensing Method Based On Deep Learning

The adaptive optics system can correct beam quality degradation in optical systems caused by wavefront aberrations,which is widely used in fields such as astronomical observation,free-space optical communication and biomedicine.As an essential component of the adaptive optics system,wavefront sensors provide the phase information of the distortion,determining the correction accuracy and stability of the system to a large extent.In recent years,deep learning has advanced rapidly and is used in various scientific and industrial fields.The combination of deep learning and adaptive optics is also being extensively and intensively researched and has a wide potential for development.This paper focuses on wavefront detection methods based on deep learning,intending to break through the existing wavefront detection theory and propose practical deep learning wavefront sensors.The content of the full text mainly includes the following part:Firstly,this paper provides an overview of the basic principles of adaptive optics and deep learning.The basic idea,composition,applications and wavefront sensors of adaptive optics systems are presented,and the underlying theory of deep learning and neural networks is explained.Besides,this paper also introduces the current status of the application of deep learning in adaptive optics,clarifies the essentials of deep learning in wavefront detection,and specifies the research direction of the deep learning wavefront detection method.In the second part,the phase diversity deep learning wavefront recovery algorithm is implemented and investigated.A real-time,non-iterative phase diversity method is implemented using a deep neural network.This method uses a neural network to fit the mapping relationship between the focus\defocus images and the wavefront aberration.No iterations are required in the wavefront recovery process,meaning that the proposed algorithm is computationally fast and has high measurement accuracy.Furthermore,numerical simulations were used to analyse the characteristics of the algorithm,and it was found that the accuracy of the algorithm is mainly determined by the wavefront distortion characteristics,the structure of the neural network and the performance of the photodetector.The algorithm’s accuracy for recovering wavefront aberrations with high spatial frequencies is relatively low,which means that the algorithm is not suitable for detecting atmospheric turbulence.Finally,we verified the conclusions of the numerical simulation through experiments.In the third part,we propose a deep learning wavefront recovery algorithm for Hartmann sensors with sparse sub-apertures.Based on the existing research,we propose using the sparse sub-aperture Hartmann sensors as the optical structure and deep learning to achieve wavefront recovery.This method provides a higher measurement accuracy than the phase diversity deep learning method.The number of sub-apertures decreases to 1/10 of the existing Hartmann sensors,breaking the limitation of d/r₀=1 for Hartmann in measuring atmospheric turbulence.Numerical simulations and experiments have proved the above conclusions.The results show that the proposed algorithm can improve the signal-to-noise ratio of the detector signal while achieving high-precision wavefront detection.In the last part,we propose an autoencoder for wavefront detection for Hartmann sensors with sparse sub-apertures.The existing deep learning-based wavefront sensing methods,requiring labelled data to train neural network models,are less promising for engineering applications.Inspired by image autoencoders,we model Hartmann’s optical transmission process in the deep learning framework and implement unsupervised learning of neural network models.The proposed method allows the training of neural network models with unlabelled Hartmann images,breaking through the existing deep learning wavefront detection theory.Numerical simulations investigate the algorithm’s properties.The results show that the proposed algorithm has high measurement accuracy and does not require labelled data,making it the most promising method for deep learning wavefront detection for engineering purposes.In this paper,deep learning-based wavefront detection methods are investigated through numerical simulations and experiments.For deep learning-based wavefront detection methods,the optical structure is optimised,and the prospects for engineering applications are enhanced.In general,the work in this paper provides new ideas for adaptive optics wavefront detection methods.