Research On Optimization And Modeling Of Electromagnetic Problems Based On Machine Learning

Recently,machine learning(ML)algorithms have developed rapidly with the advancement of data acquisition,storage,and handling technologies.ML algorithms were first used in computer fields,such as image analysis and pattern recognition.Due to their superior performance,machine learning algorithms are becoming a new method to provide more options for solving complex electromagnetic problems.In this dissertation,two important research directions in wireless communication,i.e.,channel modeling and device design,are taken as the starting point.The related basic throries and key technologies are studied,and efficient solutions based on machine learning algorithms are proposed.The main research contents of this dissertation are as follows:Firstly,according to ITU-R P.1546 recommendation,an electromagnetic field strength prediction model based on digital map is established for outdoor plain wireless communication.The accurate prediction of regional field strength can be realized by using the terrain information provided by digital map.Considering the vast territory and diverse terrain of China,the Deygout model is adopted to deal with the multi peak problem existing in mountainous and hilly areas.Based on the Deygout model,the diffraction loss of signal propagation under undulating terrain is calculated.The two models are reasonably combined to realize the field strength prediction under different terrain conditions.Since the above models are statistical models,a hybrid model combining statistical model and artificial neural network(ANN)is proposed to make up for the shortcomings of poor statistical modeling accuracy.The ANN is trained to represent the relationship between terrain information and prediction errors of field strength to compensate the field strength predicted by the statistical model,thereby effectively improving the prediction accuracy.Next,the channel modeling of the time reversal(TR)communication system is studied by using data-driven machine learning alrorithm.Due to the ability to gather and utilize the information transmitted by the abundant multipath channels of indoor communication,time reversal electromagnetic wave has space-time focusing characteristics,which is benefit to improve the signal-to-noise ratio and enhance the performance of the communication system.A knowledge-based ANN channel modeling method is proposed for the TR indoor communication system.Principal component analysis(PCA)technology is used to reduce the dimensionality of received signal data,and the neural network is used to learn the nonlinear relationship between the receiver position and the dimension-reduced data.During the training process of the ANN,the propagation characteristics of the TR signal are used to guide the optimization of the internal parameters of the ANN model,which improves the performance of the model.A simple indoor wireless communication environment is taken as an example to verify the effectiveness of the proposed method.In the past,TR channel models were established based on statistical modeling and deterministic modeling methods,and it is the first time to use machine learning algorithm to establish TR channel model.In the study of new methods of electromagnetic device design based on machine learning,a multibranch inverse modeling method is firstly proposed,which effectively solves the inverse problem of the antenna array directivity.Different from the electromagnetic forward problem which has the unique relationship,there is nonuniqueness in the inverse problem because there are multiple structural parameters corresponding to the same electromagnetic response.For solving the nonuniqueness problem,the derivative information in the training sample data is obtained by using the adjoint neural network and used to judge whether the nonuniqueness exists.Based on monotonicity,the training samples are divided into multiple groups to ensure the uniqueness of each group.Each group of the training samples is used for the training of an ANN respectively.After the training process,all ANNs are combined to form a whole model.The validity of the proposed model is verified by four numerical examples.Compared with the direct inverse modeling method,the proposed method significantly improves the accuracy of the inverse model.Then,an inverse model(TF-ANN)combining transfer function(TF)and artificial neural network is used to solve the inverse problem of metasurface.By using vector fitting technology to process the electromagnetic response,the poles and residues of TF are fed into the inverse model.Compared with the EM response at discrete points as the input,the input dimension of the TF-ANN inverse model is significantly reduced,which is conducive to reducing the structure complexity and improving the accuracy of the inverse model.A reflective metasurface is taken as an example to verify the effectiveness of the proposed method.Finally,in order to solve the nonuniqueness problem in inverse modeling more efficiently and simply,an improved TF-ANN inverse model is proposed.During the training process,the weights and thresholds of the inverse model are optimized by cascading with two pre-trained forward branches.Taking an electromagnetically induced transparent-like(EIT-like)metasurface as an example,it is verified that the proposed method can have good performance in the case of a small training set.