Thermally controllable diffractive processor using hybrid metasurfaces

Abstract All-optical diffractive neural networks (DNNs) have garnered notable attention in artificial intelligence because of their unprecedented ability to perform complex tasks at the speed of light. However, current DNNs designs mainly rely on non-tunable diffractive components with fixed functionality. Here, to introduce active diffraction units into DNNs at terahertz (THz) spectrum, we report a hybrid metasurface-based thermally controllable diffractive processor that mimics task-switchable DNNs using thermally reconfigurable meta-neurons composed of vanadium dioxide (VO 2 ) resonators and gold resonators. The optical response of the network layer of this processer can be dynamically switched based on ambient temperature to meet dynamic task requirements. We experimentally validate the efficacy of the processer through classification tasks involving single and double handwritten digits as well as fashion products. Moreover, we propose an information encryption framework that leverages ambient temperatures as decryption keys. This temperature-dependent all-optical processor with multiplexing capability paves the way for the next-generation high-speed optical computing, dynamic optical systems, and secure communications.