Research On Design And Application Of Optical Metasurface-based Diffractive Neural Networks

Machine learning not only uses data and algorithms to mimic human learning but also possesses tremendous information processing power,thus receiving widespread attention and becoming a frontier hotspot for current research in many fields.Diffractive neural networks use all-optical systems to perform machine learning,providing broad application prospects for achieving high-speed,low-power artificial intelligence.In addition,the metasurface has a strong ability to manipulate the wavefront,so it becomes an ideal carrier for the diffractive neural network.In this thesis,we mainly focus on two important directions in the diffractive neural network based on optical metasurfaces:The hypersurface design based on machine learning and the application of diffraction neural network in decision and control are studied.The main research contents and innovations of this thesis are as follows:First,there are some problems in the current machine learning-based photonic device design,such as relying on a large amount of training data,high demand for computing resources,etc.By analyzing the characteristics of deep learning and reinforcement learning,this thesis proposes a photonic device design method that combines these two methods.With deep learning,a rough initial structure can be predicted from a small data set,while reinforcement learning can optimize the structure through continuous experimentation and feedback.Thus,combining these two methods can effectively improve the efficiency of photonic device design.As an application example,this thesis uses the method to design efficient metasurface structures for the implementation of diffraction neural networks.Secondly,the current research on diffractive neural networks mainly focuses on simple image classification and object detection,and does not involve any interaction with the environment.This thesis proposes a diffraction neural network implemented using reinforcement learning,which has human-like decision-making and control capabilities and is implemented by the metasurface designed above.By interacting with the environment,this network is able to find the optimal control strategy.The performance of the network was validated on three types of classic games: Tic-Tac-Toe,Super Mario Bros.,and Car Racing.The results show that the network is able to reach the same or even higher levels of performance as human players.This work is expected to drive the shift from simple recognition or classification to advanced decision making and control in diffraction neural networksThe approach to metasurface design,which combines deep learning and reinforcement learning,and the use of diffraction neural networks for decision-making and control proposed in this thesis,provides new ideas and technologies for photonic device design and optical neural network.It also demonstrates its scientific value and potential for practical applications.