Terahertz Metasurface On-Demand Design Based On Deep Learning

Metamaterials are artificially designed materials that exhibit electromagnetic properties that differ from those of naturally occurring materials due to their microstructure and arrangement.Electromagnetically induced transparency(EIT)is a nonlinear quantum interference effect that has great potential in various areas such as frequency domain information storage and processing,optical communication,and optical detection.In recent years,the design of metasurfaces has been accelerated by the rise of deep learning,researchers have begun to explore the potential underlying mechanisms between metasurface structures and electromagnetic responses via deep learning.Therefore,the combination of deep learning and metasurface structure design has significant research and application value.This thesis focuses on the on-demand design method of deep learning applied to terahertz EIT metasurfaces,the main works are as follows:1.This thesis proposes numerical encoding matrix as one of the training data sets.This method requires only two pixel-values of 0 and 1 to map the actual EIT metasurface structure into a two-dimensional structure matrix,and it overcomes the limitations of traditional design structure parameters and upgrades it from a single structure parameter dimension to a 2D structure matrix design of the entire metal layer,effectively expanding its parameter design space and search range.Therefore,the more experimental data are covered in a given frequency band,the forward prediction and inverse design effects of the trained network model will be more flexible and accurate.2.A robust integrated framework for EIT-encoded metasurfaces design is constructed using convolutional neural networks and fully connected networks.Upon inputting the structure matrix to the trained forward prediction network,the transmission spectrum corresponding to the EIT metasurface structure can be obtained within a few milliseconds.The inverse design network efficiently learns the inverse mapping between the EIT transmission spectrum and its structure,and the corresponding EIT metasurface structure matrix can be obtained quickly by inputting the target transmission spectrum,which efficiency is unattainable with traditional design methods.The proposed network design framework can also be theoretically extended to the design of metamaterial structures with similar characteristics in the spectrum,such as Fano resonance,owing to the clear characteristic point of the EIT transmission spectrum,which comprises a "W"-shaped transparent window consisting of two troughs and one peak.3.A deep convolutional generative adversarial network(DC-SAGAN)based on a self-attention mechanism has been developed for the inverse design of EIT metasurfaces.DC-SAGAN effectively overcomes the limitations of current deep learning applications in designing EIT metasurface structures,it not simply imitating the training data set,but learning the underlying relationship between the metasurface structure and its electromagnetic response.The structure of the network output not only adheres to the design rules of the EIT metasurfaces,but also shows the robustness of its adaptive generation of the EIT metasurface structure in terms of the network design structure when different target transmission spectra are arbitrarily input,as DC-SAGAN is able to inverse design and produce accurate design solutions beyond the training data,so it provides a reliable design framework for multifunctional EIT device design.