Neural networks possess incredible capabilities for extracting abstract features from data. Electromagnetic computing harnesses wave propagation to execute computational operations. Metasurfaces, composed of subwavelength meta-atoms, are capable of engineering electromagnetic waves in unprecedented ways. What happens when combining these three cutting-edge technologies? This question has sparked a surge of interest in designing physical neural networks using stacked intelligent metasurface (SIM) technology, with the aim of implementing various computational tasks by directly processing electromagnetic waves. SIMs open up an exciting avenue toward high-speed, massively parallel, and low-power signal processing in the electromagnetic domain. This article provides a comprehensive overview of SIM technology, commencing with its evolutionary development. We subsequently examine its theoretical foundations and existing SIM prototypes in depth. Furthermore, the optimization/training strategies conceived to configure SIMs for achieving the desired functionalities are discussed from two different perspectives. Additionally, we explore the diverse applications of SIM technology across the communication, sensing, and computing domains, presenting experimental evidence that highlights its distinctive advantages in supporting multiple functions within a single device. Finally, we identify critical technical challenges that must be addressed to deploy SIMs in next-generation wireless networks and shed light on promising research directions to unlock their full potential.

翻译：神经网络具备从数据中提取抽象特征的卓越能力。电磁计算利用波传播来执行计算操作。由亚波长超原子构成的超表面能够以前所未有的方式调控电磁波。将这三项尖端技术相结合会产生什么？这一问题激发了利用堆叠智能超表面技术设计物理神经网络的研究热潮，旨在通过直接处理电磁波来实现多种计算任务。SIM为电磁域中高速、大规模并行且低功耗的信号处理开辟了一条令人振奋的新途径。本文首先从演进历程出发，对SIM技术进行了全面综述。随后，我们深入探讨了其理论基础与现有SIM原型。此外，我们从两个不同视角讨论了为配置SIM以实现预期功能而设计的优化/训练策略。同时，我们探索了SIM技术在通信、感知和计算领域的多样化应用，并通过实验证据凸显了其在单一设备中支持多重功能的独特优势。最后，我们指出了在下一代无线网络中部署SIM必须解决的关键技术挑战，并阐明了释放其全部潜力的前瞻性研究方向。