High-Efficiency Transmissive Tunable Metasurfaces for Binary Cascaded Diffractive Layers
Yuetian Jia, Huan Lu, Zhixiang Fan, Bei Wu, Fengzhong Qu, Minjian Zhao, Chao Qian, Hongsheng Chen
Abstract
Recently, there has been a surge of interest in implementing neural networks on wave-based physical platforms due to the superior capability of parallel processing at the speed of light, spawning it complementary to conventional electronic computing. Metasurfaces-based cascaded diffractive layers is one of representative wave-based computing modalities. However, most existing works rely on passive diffractive components, and the enormous network parameters and sufficient computing resources pose a great roadblock. Another dilemma lies in the experimental realization, in which the design of tunable transmissive metasurfaces is arduous. Here, we propose binary cascaded diffractive layers by applying tunable transmissive metasurfaces with opposite phase and transmission efficiency of 96%. Our discrete optimization algorithm reduces computational redundancy during the training process and becomes more suitable for deployment in mobile applications, thus preserving network inference accuracy while avoiding heavy computational tasks and memory consumption. By optimizing the transmission state of each neuron, our system can achieve multiple functions on demand, such as multichannel transmission system and holographic image generation. To validate the feasibility, we fabricated two-layer binary cascaded diffractive layers and tested it using a multichannel transmitter. Our study provides a simple yet viable way for on-site learning due to its high-efficiency metasurface design and low computing requirement.