Deep Learning Enables Multifunctional Metasurfaces Design With Mutual Coupling Estimation

Exploring a universal approach to enhancing the performance of metasurfaces by quantifying the influence of mutual coupling between unit cells is an appealing topic for many researchers, yet achieving it within the scope of conventional methods is extremely challenging. In this article, we present a versatile neural network model for analyzing the actual electromagnetic (EM) responses of unit cells including their transmission amplitude, phase, and near-field radiation pattern, with the aim of improving the efficiencies of metasurfaces. By incorporating prior knowledge into deep neural networks (DNNs) through physics-driven neural networks, an ultra-small dataset is sufficient to train the model with faster convergence speed and lower computational costs. With the assistance of optimization algorithm, this framework empowers the researchers to accurately shape wavefronts of metasurfaces by taking into account the coupling effects in both near-field and far-field regions. Finally, two far-field multiple beam metasurfaces and two near-field holographic metasurfaces, as examples, are employed to demonstrate the efficiency improvement achieved in various metasurfaces design.