近日，东南大学崔铁军团队研究了实现可重构衍射神经网络的多功能可动型编码超表面。相关论文于2026年2月26日发表在《光：科学与应用》杂志上。

光学计算因其低延迟、高能效和多维度处理能力，在众多应用领域展现出重要前景。然而，当前衍射神经网络普遍缺乏可重构性，这限制了光学计算系统的可扩展性。

受活字印刷技术启发，研究组提出可移动式编码超表面，能够实现电磁计算、全息成像和传感探测等多功能应用。通过级联多层超表面，他们进一步开发了可移动式可重构衍射神经网络。通过替换最后隐藏层中的超原子，该网络可从手写数字分类任务无缝切换至字母分类任务。此外，单层可移动式编码超表面可通过模块化超原子重排，重构实现电磁全息成像与多人生命体征传感。所提出的可移动式编码超表面具有重构简便、高度灵活和模块化可扩展等特性，为多功能可复用电磁计算、全息成像与传感应用提供了新途径。

附：英文原文

Title: Multifunctional movable-type coding metasurface enabling reconfigurable diffractive neural networks

Author: Yu, Zhicai, Li, Xinyu, Gu, Ze, Chen, Long, Su, Jianlin, Cai, Zixuan, Yu, Xinyi, Qin, Shilong, Zhang, Lei, Ma, Qian, You, Jian Wei, Cui, Tie Jun

Issue&Volume: 2026-02-26

Abstract: Optical computing holds significant promise across diverse applications due to its low latency, power efficiency, and multidimensional processing capabilities. However, current diffraction neural networks (DNNs) generally lack reconfigurability, limiting the scalability of the optical computing systems. Inspired by movable-type printing technology, here we propose a movable-type coding metasurface to enable multiple functionalities such as electromagnetic (EM) computing, holography, and sensing. By cascading multiple layers of the proposed metasurfaces, we further develop a movable-type reconfigurable DNN (MT-RDNN). It can be seamlessly adapted from handwritten digit to letter classification tasks by replacing the meta-atoms in the last hidden metasurface layer. Moreover, a single-layer movable-type coding metasurface can be reconfigured to perform EM holography and multi-person vital sign sensing through modular meta-atom rearrangement. Featuring simple reconfiguration, high flexibility, and modular scalability, the proposed movable-type coding metasurface enables versatile and reusable EM computing, holography, and sensing applications.

DOI: 10.1038/s41377-026-02216-6

Source: https://www.nature.com/articles/s41377-026-02216-6