Design and experiment of terahertz unidirectional transmission structure based on neural network

In metasurface design, massive meta-atoms have to be optimized to produce the desired properties, which are time-consuming and sometimes prohibitive. In this paper, neural network model composed forward modeling and inverse design network is proposed to realize rapid, efficient, and automatic metasurface design of THz unidirectional transmission. Specifically, massive metasurface samples are generated by Python-CST co-simulation to train the neural network model and make it robust to predict the THz responses and vice versa. The consistency of simulation and prediction results imply that the model can accurately capture the nonintuitive complex relationship between structural parameters and electromagnetic spectra. To further verify the model, retrieved structural parameters for unidirectional transmission are deduced due to inverse design network and double L-shape arrays metasurface is experimental fabricated. It is shown that the experimental asymmetric transmission responses by retrieved structural parameters are identical with the simulations by design parameter. Such results effectively verify the design and model, providing a paradigm for functional metasurface design and the development of applications in optoelectrical fields.