Inverse design of terahertz metagrating based on neural network

In this paper, an inverse neural network based on deep learning is constructed to predict the metasurface structure of the designed terahertz metagrating. The transmittance spectra results from the numerical simulation of the metagrating were used as the input datasets for the inverse neural network, and the output is the corresponding metagrating structure parameters. After training, our inverse network can meet our expectations. The results show that some of the structural parameters predicted by the network are roughly consistent with the actual structural parameters, which indicates that the neural network can predict the corresponding structural parameters by given spectra. This has great application value, for example, it can be used to guide the design of metasurfaces for faster and more convenient purposes.