Inverse Design of Metasurface Holography on Deep Residual Networks

The electromagnetic properties of meta-atoms determine the degree of freedom of the transmitted electromagnetic waves manipulated by metasurfaces. However, there are singular point mutations in the mapping between the electromagnetic properties and geometric structure of meta-atoms, which increases the difficulty of design. In the paper, we present a meta-atoms inverse design approach based on deep residual networks combined with electromagnetic properties classification. We designed the network to predict the geometry of meta-atoms from the amplitude and phase of the transmission coefficients. By combining electromagnetic property classification, the network effectively solves the problem of inaccurate predictions caused by singular point mutations and reduces training data. The mean square error of all parameters in the data set is less than 0.23. Furthermore, we also designed the holographic imaging metasurfaces to verify network reliability. The network proved to be a new tool for the inverse design of meta-atoms.