DL-FDTD Algorithm Research And EDA Software Development

Machine learning has become a cornerstone of modern science,and it is widely used in all scientific fields.However,its combination with computational electromagnetics(CEM)algorithms remains to be investigated.In this thesis,we combine a deep learning(DL)based machine learning approach with a traditional FDTD(Finite-Difference Time-Domain)algorithm to investigate the principles and implementation techniques of the DL-FDTD algorithm and apply them to computational electromagnetics.This thesis first proposes a DL-based FDTD algorithm and elaborates on the FDTD algorithm using two different neural network methods,the recurrent convolutional neural network(RCNN)and the multilayer perceptron(MLP),and compares it with the existing work.Finally,by comparing the relative error of the algorithm in two numerical arithmetic cases,it can be seen that the proposed algorithm has been effectively improved on the basis of the existing work,and the relative error of the existing work has been reduced.In this thesis,we also propose a DL-based absorption boundary condition for PML.The absorption boundary condition is based on the LSTM neural network.By sampling and training the data set of PML at higher levels,the absorption boundary condition of PML based on LSTM network is obtained with only one layer,and the reflection coefficient is smaller than that of the traditional PML boundary condition with one layer thickness,and the correctness of the algorithm is verified by numerical examples.The thesis then analyzes how to develop a general-purpose electromagnetic simulation software,from engineering requirements,system architecture,and core functional modules,and shows part of the core module code written in the latest C++language standard,and proves the effectiveness of the code using numerical examples.Finally,as an outlook,this thesis analyzes examples of using machine learning to solve real engineering problems,such as using machine learning to solve electromagnetic inverse problems and calculating radar cross-sectional area(RCS).We believe that combining deep learning with traditional computational electromagnetics methods will definitely be a hot topic in the future development of computational electromagnetics.