Electromagnetics analysis based on the machine learning methods

Abstract

This thesis conducts a systematic study on the novel computational electromagnetics (CEM) analysis methods using the machine learning (ML) techniques. Nowadays, conventional electromagnetic methods are facing various challenges, including efficiency, computation complexity, etc. ML technologies are becoming very attractive to modeling sophisticated problems that are too complex to develop an accurate explicit models. This thesis is attempting to explore whether machine learning methods could be used to innovate the conventional CEM, and what benefits or drawbacks they could bring to fundamental CEM algorithms. According to our knowledge, this thesis is among the very first efforts to apply ML into the fundamental CEM studies. Firstly, a novel idea is proposed by rethinking the method of moments (MoM) solving process into a machine learning training process. Its most important significance is to make the MoM process compatible to ML process. Based on the artificial neural network (ANN), the conventional linear algebra MoM solving process becomes training a machine learning using the system construction data. So far this might be the only ML solution for CEM MoM, according to our literature reviews. Secondly, a new perfectly matched layer (PML) absorbing boundary condition (ABC) for the finite difference time domain (FDTD) method is proposed and developed based on ML models. The resultant ML based models could be integrated into FDTD solving process with satisfactory accuracy. This model is the only ML based PML idea that has ever been proposed and published. Thirdly, electromagnetic inverse (EMIS) problems are solved by deep learning (DL) approaches. The newly proposed methods are based on the deep complex-valued convolutional neural networks (ConvNets). Compared with conventional EMIS methods, the designed deep ConvNets can successfully reconstruct the high contrast target with better accuracy. Fourthly, a new source reconstruction method (SRM) based on deep ConvNet is proposed to reconstruct the equivalent sources of the target. It could be a convenient approach feasible for the EMC and EMI analysis. The proposed method can be conveniently extended to make the direction of arrival (DOA) estimation and the target permittivity measurement. Another related effort is to utilize DL approach for the far-field sub-wavelength imaging of the near-field resonant metalens at microwave frequencies. In the last effort, the deep ConvNet idea is used to recover the layer thickness and temperature of snow from the passive microwave remote sensing (PMRS). From our studies, new algorithms and ideas have been proposed and tested for CEM. What we found is that the ML methods seem to be more compatible to inverse EM problems than forward EM problems. ML methods have the advantage of simplifying complex arbitrary target process and providing impressive noise tolerance to the data. Hence, our proposed DL approaches for inverse EM problem can present better performance over many conventional methods. But the accuracy control of ML model made them not a strong presence in the forward EM solving. Because it is still the beginning of ML for CEM, more profound ideas could be further developed in the future on top of the findings of this thesis.