Researches On Electromagnetic Modeling And Design Based On Machine Learning

Recently,with the rapid development of artificial intelligence,employing the machine learning for electromagnetic(EM)simulation and design has a promising prospect.Machine learning could be a good alternative to EM simulations to construct computer aided design(CAD)models which significantly speed up the EM-based design and save calculation cost by being repeatedly called by an optimization algorithm.To further improve the modeling accuracy and efficiency,this dissertation proposes several models based on machine learning for EM intelligent design,and the main contents are introduced as follows.Chapter 1 mainly gives a summary of EM intelligent design based on machine learning from three aspects of the artificial neural network(ANN),the prior knowledgebased ANN and the transfer function(TF)-based ANN.Chapter 2 mainly introduces the study of an ANN model based on data mining techniques.In this chapter,an ANN model with data mining techniques is proposed.In this ANN method,two data mining techniques,including correlation analysis and data classification,are employed to determine geometrical variable inputs and classify the inputs during the training and testing processes.The validity and efficiency of this proposed method are confirmed with a U-slot patch antenna and two band-notched ultrawide band(UWB)antenna examples.Chapter 3 mainly introduces the study of multi-parameter modeling with ANN.A novel ANN model is firstly proposed to describe the antenna performance with various parameters.In this model,three parallel and independent branches are involved for three different performance parameters.Once the geometrical variables are input,the ANN model can simultaneously obtain S-parameter,gain and radiation pattern from the independent branches.The validity and efficiency of this proposed model are confirmed with a Fabry-Perot resonator antenna example.Then,to employ prior knowledge in ANN for the modeling of finite periodic arrays,a new multi-grade ANN model is also proposed in this chapter.Considering mutual coupling and array environment,the proposed model is designed with two sub-ANNs,element-ANN and array-ANN.Based on the relationship between the geometrical variables and the EM behavior of elements in an array,elementANN is built to provide prior knowledge for the modeling of array-ANN.Then,in review of mutual coupling and array environment,array-ANN is modeled to obtain the EM response of the whole array from the nonlinear superposition of the element responses.Three numerical examples of a linear phased array,a six-element printed dipole array and a U-slot microstrip array are employed to verify the validity and efficiency of the proposed model.Chapter 4 presents the study of the modified extreme learning machine(ELM).In this chapter,an ELM model based on pole-residue-based TF is firstly developed.In this method,to further save the training cost,a modified flower pollination algorithm(FPA)based on the steepest descent method(SDM)is proposed to set the optimal initial weights and thresholds of ELM training to decrease the dependency of the initial weights and thresholds.The validity and efficiency of this proposed method is confirmed by two parametric modeling examples of filter design.Then,a dynamic adjustment kernel extreme learning machine(DA-KELM)with transfer functions is proposed.If satisfactory accuracy has not been obtained,the proposed model,which supports the functionalities of increased learning,reduced learning and hybrid learning,can utilize the overlap between the old training dataset and the new one to achieve the accurate trained results with faster retraining.The validity of the proposed model is confirmed with two examples of a microstrip-to-microstrip vertical transition and a quadruple-mode filter.Finally,based on the DA-KELM,this chapter proposes a semi-supervised learning model lying between supervised learning and unsupervised learning to largely reduce the number of required training samples.The proposed model contains two training processes,the initial training and the self-training.In the initial training process,a small number of training samples from full-wave simulations are used to make the model rapidly converge.Then,in the self-training process,the model produces unlabeled training datasets to train itself till the testing accuracy is satisfied.Two numerical examples of a microstrip-tomicrostrip vertical transition and a dual-band four-pole filter are employed to verify the effectiveness of the semi-supervised learning model.Chapter 5 makes a comprehensive summary of the full dissertation,and it points out the issues needed to be further researched.