Design of metasurface skin cloak based on a low-complexity monitoring model and deep learning

Abstract

This work proposes a research scheme to speed up the design of metasurface skin cloak through low-complexity phase monitoring model and deep learning. This skin cloak conceals a three-dimensional arbitrarily shaped object by complete restoration of the phase of the reflected light at specific wavelength. And the possibility of realizing spectral prediction by deep learning is analyzed. During the study, a phase monitoring system was designed in which the detector, the light source and the monitored nano-antenna were sequentially distributed at equal distances from the emitted wavelength of the light source, so that the monitored phase amount was exactly equal to the phase change before the reflected wave, thus eliminating the need for multiple monitors to measure and calculate the phase change before and after the reflection. The traditional metasurface design is usually constructed by manual library construction based on the phase distribution and the relationship between phase variation and dimensional variation of the cell structure, so this work combines the aforementioned monitoring model with deep learning to generate the database required for modeling. The two variable parameters of device length and width were first defined, and the reflected wavefront phase change used as the optical response, and we reprocessed the original data and finally build and trained an artificial neural network model for forward prediction of optical response. This network can obtain its MSE below 0.001 for the test set after the training is completed. Thus the scheme can replace the role of simulation software to some extent, and its prediction process can be completed in a few milliseconds, improving the efficiency of the design metasurface process.