Deep learning-enhanced multifunctional THz metasurfaces via conditional variational autoencoder

Abstract The conventional process for developing multifunctional metasurfaces relies on trial-and-error simulation-based optimization methods. This approach is not only time-consuming, but also requires a high level of expertise from the designers and often does not produce the desired results. Deep learning can automate the design process, enabling rapid and intelligent design of multifunctional metasurfaces. However, there are still limitations in deep learning-assisted modeling and design, such as limited design freedom and single-function realization. In this paper, a deep learning framework based on a conditional variational autoencoder (CVAE) network is proposed for the design of multifunctional terahertz (THz) metasurfaces. We introduce vanadium dioxide (VO 2 ) phase-change material into the design of metasurfaces with random patterns to achieve switchable functionality between absorption and polarization conversion. The adopted network architecture incorporates a conditional discriminative mechanism on the basis of a variational autoencoder, enabling the network to more accurately identify spectral characteristics and generate corresponding metasurface structures. The absorption and the polarization conversion rate (PCR) of the generated structures have been shown to closely match the target spectra. With the development of intelligent design technology, the practical applications of metasurface absorbers and polarization converters in fields such as stealth technology and optical communication will be significantly accelerated.