Phase-to-pattern inverse design for a fast realization of a functional metasurface by combining a deep neural network and a genetic algorithm

Metasurface provides an unprecedented means to manipulate electromagnetic waves within a two-dimensional planar structure. Traditionally, the design of meta-atom follows the pattern-to-phase paradigm, which requires a time-consuming brute-forcing process. In this work, we present a fast inverse meta-atom design method for the phase-to-pattern mapping by combining the deep neural network (DNN) and genetic algorithm (GA). The trained classification DNN with an accuracy of 92% controls the population generated by the GA within an arbitrary preset small phase range, which could greatly enhance the optimization efficiency with less iterations and a higher accuracy. As proof-of-concept demonstrations, two reflective functional metasurfaces including an orbital angular momentum generator and a metalens have been numerically investigated. The simulated results agree very well with the design goals. In addition, the metalens is also experimentally validated. The proposed method could pave a new avenue for the fast design of the meta-atoms and functional meta-devices.