Electromagnetic-driven Intelligent Techniques And Their Applications

The development of artificial intelligence(AI)has brought great revolutions to the development of human society,and new applications based on AI has been continuously emerging.In the field of electromagnetism(EM),AI has been used in the acceleration of EM simulation,automated antenna design,inverse scattering imaging,etc.Meanwhile,owing to the powerful capability in controlling EM waves and flexible designs of EM metamaterials,especially the real-time digital configuration abilities of programmable information metamaterials,the AI technologies could easily be brought into the design and intelligent applications of the EM metamaterials.However,current AI technology still has several shortcomings such as overdependence on training data,lack of interpretability and bad generalization,which seriously hamper the development and applications of AI-assisted EM techniques.Inspired by providing physical priori knowledge to AI via computational EM theory,this thesis systematically studies EM-driven AI technology(including machine learning and deep learning algorithms)and its EM applications,in order to reduce the dependence on data and increase the interpretability & generalization ability.Furthermore,this thesis proposes the concept of wave-computing based programmable AI machine(PAIM)and relevant theory & algorithms.Meanwhile,the PAIM is manufactured using programmable metasurfaces,and its various intelligent applications are demonstrated.The main content and innovations of this thesis are summarized as follows:1)A fast scattering-center-extraction method under condition of wide bandwidth &radar-perspective is proposed.Combining fast inverse synthetic aperture radar(ISAR)imaging based on shooting and bouncing ray(SBR)technique and machine-learning optimization technique,this method could extract stable scatter centers in several seconds,by which the radar cross-section(RCS)of wide bandwidth & radar-perspective is able to be recovered to the maximum extent.2)A deep learning network structure inspired by the physical mechanism of wave phase to predict the wide-band reflection phase responses of coding metasurface is proposed,and together with a well-designed machine learning algorithm,a reliable automatic design technique for anisotropic digital coding metasurface with desired wide-band reflection phase responses(8-12GHz)is realized,which is four orders of magnitude faster than using full-wave simulation softwares and can be generalized to other automatic parametric design of antennas.3)Two AI techniques used for holographic imaging are proposed,one is a discrete machine learning optimization algorithm and the other is a physics-assisted unsupervised deep learning network structure.The latter is designed based on the free-space propagation equations of EM wave,and the phase distribution of radiation sources could be immediately obtained once the target holographic imaging is input into the network.The availability of the proposed techniques is proved by experiments on coding metasurface platform,displaying a better imaging quality than canonical Gerchberg-Saxton algorithm.4)An EM-driven unsupervised deep learning network srutcture for inversing imaging is proposed.By strongly combining the contrast source inversion method and generative adversarial network,the proposed technique gets rid of the dependence of coupled training data and could generate high-quality image of target objects when given the EM scattering data,which is unprocurable by canonical inverse imaging algorithms.5)A concept of wave-computing based programmable AI machine(PAIM)and relevant theory & algorithms are proposed.The PAIM is an instantiation device for the integration of EM machenism and convolutional neural network.A PAIM prototype is manufactured by programmable metasurfaces,which could directly modulate EM waves in free space for intelligent computings at light speed.6)Various novel cases of intelligent tasks and wireless communications based on the proposed PAIM are demonstrated,including imaging recognization,muti-beam forming by reinforcement learning and wireless message transmission with coding & decoding &denoising in wave space,which fully explain the enormous potential for PAIM in intelligent EM applications.