Deep Learning Modeling And Optimization For Specific Electromagnetic Structures

With the rapid development of information technology,the operating frequency and integration of devices continuously improve,accompanied by the increasingly complex electromagnetic(EM)environment.The efficient manipulation of EM waves is facing new challenges.The artificial periodic structure performs macroscopic EM properties that do not exist in nature through geometric design and further achieves breakthroughs in the size and performance of EM devices,which becomes a research hotspot.However,traditional design methods rely on experience and are time-consuming and labor-intensive.It is urgent to research efficient inverse design methods.Owing to the significant improvement of computing power and the modeling ability of deep neural networks for complex nonlinear problems,deep learning methods provide a new way to solve EM-related problems.But in terms of inverse problems,there are still many key problems that need to be solved: First,the constrained modeling problem of EM structures,i.e.,how to generate structures that satisfy certain constraints under high degrees of freedom; second,the quantification,prediction,and inverse design problems of EM eigenmodes,i.e.,how to make the neural networks efficiently learn and deal with EM pattern information; third,the inverse analysis problem of EM eigenmodes under complex boundary conditions,i.e.,how to use deep learning to discover reasons from the phenomenon.Aiming at these bottlenecks and challenges,this thesis studies modeling and optimization problems for specific EM structures with deep learning techniques in three aspects: inverse design,inverse optimization,and inverse analysis.The research contents are as follows:1.Aiming at the inverse design of metasurfaces,a criss-cross cell structure is designed to achieve phase control of the two orthogonal polarization components of the inci-dent wave.Based on a few parameter scan results,a deep neural network model is constructed to realize the rapid inverse design of bifocal metalens.Aiming at key problem 1,an arbitrary hexagonal metasurface element is proposed to realize the constrained modeling of the structure,as well as the simultaneous control of the amplitude and phase of the two polarization components of the transmitted wave.The accuracy of the inverse design is verified by the simulation results.2.Aiming at the inverse design and band structure optimization of topological pho-tonic crystals and key problem 2,a method to quickly identify eigenmodes is pro-posed to realize the inverse design of photonic crystal eigenmodes by the deep neu-ral network.A pseudospin photonic topological insulator with a core-shell structure and a hybrid neural network model with both classification and regression are pro-posed,and the inverse optimization of the band gap and the inverse design of the band structure inversion are realized.The topological properties of the designed structure are verified by the simulation of the edge state.A double-checked tan-dem model is proposed to solve the gradient vanishing problem,and the method is verified by the band structure design of photonic crystals.It provides a new ap-proach for the research and design of higher-order topological states of topological photonic crystals.3.Aiming at the inverse analysis of higher-order modes in resistively loaded mono-cone TEM cell and key problem 3,the mode analysis theory and solution of mono-cone is obtained.Fast inverse prediction of the coefficients of modes at broadband is achieved by deep learning methods.Furthermore,the mechanism of standard field deviation caused by higher-order modes is revealed in the time domain by Fourier transform.The results are verified by simulation and experiment.It provides theo-retical guidance for the design of suppressing higher-order modes in TEM cells and provides references for EM inverse analysis methods related to deep learning.