Optimal Design Of Close-coupled Dielectric Metasurfaces Based On Deep Learnin

The 2D planar metasurface has unbelievable electromagnetic control ability,the advantages of the high degree of freedom design space,and ease of integration,there are a lot of advanced pieces of research in many fields,such as optical components,imaging,achromatic,antenna,electromagnetic cloaking,etc.,especially the applications of dielectric metasurface in optical band,it is expected to replace the bulky traditional optical components.Now,the research of metasurface has been deeply into how to improve the efficiency of metasurface design,improve the performance of metasurface,design and integrate multi-function metasurface,and so on.In this thesis,a design optimization framework is proposed to improve the design efficiency and performance of metasurface,which lays a foundation for its rapid and efficient design.Firstly,due to the neglect of near-coupling in traditional metasurface design methods,the metasurface performance is lower than the ideal performance.In order to consider the nearfield coupling,a method of supercell design considering the near-field coupling effect of elements in metasurface is proposed,named slide overlapping domain method.In this method,the target element and its adjacent elements are considered as a supercell,which can obtain the phase and amplitude information of the target element instead of the traditional periodic element method.The influence of the near-field coupling effect is verified by the numerical result and the correctness of the proposed method is verified.Secondly,the database of supercell and electromagnetic response with different element structures is established,and the relationship between supercell structure and electromagnetic response is studied by introducing deep neural network,thus,the electromagnetic response of different supercells such as electric and phase,can be predicted rapidly.The training result of the neural network model is as expected,the mean absolute phase error of the test result is only3.35 degrees,and the prediction time is less than 1 second,which is about 30 times faster than the simulation time.At the same time,transfer learning is constructed,which reduces the size of the training database by 8 times and improves the training efficiency of the network on the premise of ensuring accuracy.Finally,the genetic algorithm and the Particle swarm optimization are introduced,which combines the trained neural network model with the optimization algorithm to optimize the metasurface with the amplitude and phase or electric field as the optimization objective.The deflector and the focusing lens are designed and optimized.The deflection efficiency of the deflector is improved by more than 30%,optimization time is only 50 seconds.The focusing efficiency of the focusing lens is increased by more than 20%,optimization time is only 100 seconds.The method of dielectric metasurface optimization considering near-field coupling presented in this thesie can realize fast design and high efficiency of metasurface,and the method is general and can be extended to metal array metasurface,which is of great significance to the development of metasurface design and other aspects.