Customized Design of Frequency-Selective Surfaces Using the Vector-Graph-Feature-Extraction Deep-Neural-Network (VGFE-DNN) Method

In this communication, we propose the vector-graph-feature-extraction (VGFE) approach to accurately and concisely characterize details of electromagnetic metasurface (EM), and then train a progressive deep neural network (DNN), which can realize accurate prediction of complex electromagnetic spectrums from precise EM structures, or vice versa. Compared with the traditional neural-network-assisted design of metasurface structures, the VGFE-DNN method ensures comprehensive coverage of graphical features of metasurface and simplifies the number of feature extraction parameters at the same time, which can cover the complex and diverse surface structure designs with a smaller number of parameters. It offers strong adaptability and versatility by flexibly altering the electromagnetic spectrums and parameter combinations according to the needs of design objectives. The multifrequency point FSS design was realized as an example. A DNN was built up and trained by 6000 random combinations of parameters with an average error of 4%. After VGFE-DNN training, a variety combination of different spectral widths, different numbers, and location of frequency points can be easily customized, while maintaining angular stability and high light transmittance. The FSS structures with one to five frequency points were designed, respectively, and two of them were experimentally fabricated and tested, which successfully verified the accuracy of the VGFE-DNN method.