Inverse design of generic metasurfaces for multifunctional wavefront shaping based on deep neural networks

Abstract

Learning based on deep neural network brings new opportunities for intelligent design of metasurfaces as well as exploration of powerful optical functionalities. Inverse design of the metasurface geometry given the target optical response is a challenging ill-conditioned problem, which is usually solved by simplifying the input response and reducing the number of output geometry neurons. In this study, a convolutional neural network is employed to inverse design metasurface unit cells with nonintuitive generic patterns described by 100 binary pixels according to a group of target spectra including the real and imaginary transmission curves in TE and TM polarization. After careful tuning of the hyperparameters and comparison of the classification/regression output model, the network is used to design unit cells with user-defined phase delay at different frequencies and different polarization states. Two metasurfaces for holographic imaging at 0.8 THz and 1.2 THz are designed from the same network with the later one showing different images in orthogonal polarization states, which are validated by full-wave simulation. The large degree of freedom in the input and output of the neural network opens enormous searching space to meet various requirement for multifunctional wave manipulation.