| dc.description.abstract | In recent years, the rapid development of machine learning (ML) technique enables us to address complex electromagnetic (EM) problems such as Radio frequency (RF) signal analysis, antennas, and artificial EM materials design. Although ML can solve many EM problems where conventional methods have failed, developing the most appropriate framework to solve specific EM problems is still an open-ended question. This dissertation investigates the application of ML techniques to solve the problems related to antenna applications. The research work has been conducted in the following three parts: Firstly, a novel Radio frequency fingerprint (RFF) extraction approach based on the Gaussian mixture model (GMM) is employed to achieve antenna classification from its scattering signals. In contrast to conventional statistical feature extraction methods, the proposed model achieves superior performance in terms of classification accuracy. Except for the application in antenna classification, we also use EM signals scattered from human body movement to detect their stand-by emotion state. The approach is verified by applying the deep neural network (DNN) to analyze a large amount of measurement data. Secondly, I propose and demonstrate a novel framework based on an interactive learning to quantify the influence of mutual coupling between meta-atoms in the large scale non-identical metasurface array. By incorporating the deep neural network with optimization algorithm, the proposed architecture can achieve rapid design optimizations on electrically large metasurfaces with disordered meta-atoms. The approach has been validated by several examples for antenna beam optimization and radar cross section (RCS) reduction. Finally, a novel deep learning (DL) based framework is employed to account for the mutual coupling effects in the design of metasurfaces for near-field focusing. The proposed approach is demonstrated via numerical simulations that DL techniques can be used to tackle complex EM problems with a combination of good accuracy from numerical solutions and robustness of analytical approaches. | en_US |
