Neural network aided diffractive metagratings for efficient beam splitting at terahertz frequencies

Abstract

Abstract The merging of neural networks with metasurfaces is a rising subject in photonics design, which offers an abstract bridge between the geometry of the subwavelength element and the optical response. The commonly involved optical response is the transmission or reflection spectrum, while here we focus on metasurfaces with superwavelength elements and predict multiple diffraction spectra in all the possible orders and orthogonal polarization modes given the geometry. This is achieved by parallel arrangement of several fully connected neural networks with shared input and diverse output diffraction spectra. As an application example, the model is used to find a metagrating as a 1:1 beam splitter in TE mode and 1:1:1 beam splitter in TM mode. The design is taken into fabrication and experimentally tested at 0.14 THz with results that are highly consistent with the prediction.