Broadband Diffraction-Encoded and Electron-Decoded Neural Networks for Single-Pixel Arrival Information Sensing

With the rapid growth of the Internet of Things and wireless technologies, rapid, low-cost, intelligent arrival information sensing (AIS) devices have emerged as an urgent demand. Traditional array antenna-based AIS technologies usually encounter high costs and slow processing speeds due to complicated hardware architecture and post-processing algorithms. Programmable metasurfaces, as an alternative solution, often suffer from design complexity and time consumption for state switching. As a promising optical computing architecture, diffractive deep neural networks (D<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>NNs) present opportunities for designing rapid, simple, high-performance AIS devices. However, available D<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>NNs-based AIS sensing works employ the multi-pixel reception scheme, which is time-consuming and complicated for the microwave or terahertz band, owing to the limited accessibility of mature detectors that allow simultaneous sensing in multiple locations. To address these issues, we utilized broadband diffraction-encoded and electron-decoded neural networks for single-pixel arrival information sensing. It only requires a passive metasurface and a single-pixel intensity-only detector in terms of hardware architecture and fully connected neural networks at the software level. Leveraging this approach, we realized far-field angle-of-arrival and near-field three-dimensional spatial position estimation tasks. Furthermore, we also provided an adaptive learning strategy to enhance the robustness and enable intelligent machine vision meta-devices. The proposed method allows for a simple and low-cost AIS scheme, potentially facilitating the advancement of AIS meta-devices.