Forward prediction and inverse design of metasurface via deep neural network integrating multi-task deep learning with genetic algorithm

The traditional design of metasurface mainly relies on complex numerical calculation, which is not only time-consuming and laborious but also requires designers with high professional ability. With the rapid development of artificial intelligence, the deep learning method has blossomed into an effective way to realize rapid design of metasurface. Here, a deep learning architecture that combines multi-task learning neural network and genetic algorithm (MGDNN) is proposed to realize the forward and inverse mapping between the metasurface and its optical response. For a SiN disk arrays metasurface, the forward reflected spectra and inverse structure design are achieved, greatly improving the design accuracy and efficiency. The accuracy rate of forward prediction is 96.7%, and that of inverse on-demand design is 83.9%, overwhelming most of conventional deep neural network (CDNN) thus far. Our proposed method would enable fast and accurate design and optimization of various kinds of metasurfaces with different functionalities, and is expected to play a significant role in the design of future nanophotonics devices.