A Deep-Learning Enabled Discrete Dielectric Lens Antenna for Terahertz Reconfigurable Holographic Imaging

Abstract

A deep-learning enabled discrete dielectric lens (DDL) antenna with terahertz hologram reconfigurability is proposed. The antenna is constructed by two cascaded discrete dielectric lenses, which are designed based on a diffractive deep neural network (D²NN) with an improved loss function, fed by a static horn. The DDL antenna can achieve dynamic holographic imaging by a simple mechanical translation of the perfect electric conductor (PEC) mask attached to the first lens instead of locally controlling individual meta-atoms through incorporating active elements or phase change materials with complicated feeding networks. The phase profiles for the DDL antenna design are obtained by training four customized input field patterns and corresponding anticipated output target images with the modified D²NN. Dynamic switching of four number images “1, 2, 3, 4” is demonstrated by both full-wave simulated and experimental results. The proposed DDL antenna and design strategy present a new approach to achieve wavefront reconfiguration, especially in the absence of tunable components at high frequencies.