The Digital Bottleneck in Next-Generation Wireless
As wireless networks transition toward 6G, architectures like massive MIMO and holographic MIMO demand unprecedented spatial resolution to serve multiple users simultaneously. Conventionally, managing this requires complex digital signal processing (DSP) at the baseband, where signals are converted between analog and digital domains via power-hungry radio frequency (RF) chains. As highlighted by Chongwen Huang et al. in their tutorial on holographic MIMO (A Tutorial on Holographic MIMO Communications - Part III: Open Opportunities and Challenges), relying exclusively on digital-domain processing introduces severe bottlenecks in latency, hardware complexity, and energy consumption.
To break this bottleneck, a new paradigm shifts mathematical operations out of the digital processor and directly into the physical environment. By manipulating electromagnetic (EM) waves as they propagate through space, we can perform complex signal processing tasks—such as beamforming, precoding, and interference suppression—at the speed of light.
How Waves Compute
Why can an electromagnetic wave act as a computer? At its core, wave propagation through a medium is a physical manifestation of linear algebra. When multiple waves meet in space, they undergo constructive or destructive interference, which mathematically mirrors addition and subtraction. When a wave passes through a programmable material interface, its phase and amplitude are scaled, mirroring matrix-vector multiplication.
In conventional systems, the environment is treated as a passive channel, forcing all mathematical transformations to happen after signal capture (Stacked Intelligent Metasurface Performs a 2D DFT in the Wave Domain for DOA Estimation). Wave-domain computing flips this script. By routing waves through a series of programmable layers, the physical structure itself executes the required matrix operations. As Marco Di Renzo et al. demonstrate, this slashes processing latency from microseconds (in digital processors) to sub-nanoseconds, bounded only by the natural propagation speed of light (Stacked Intelligent Metasurfaces for Multiuser Beamforming in the Wave Domain).
The Architecture of Stacked Intelligent Metasurfaces
The primary hardware engine enabling this capability is the Stacked Intelligent Metasurface (SIM). Unlike traditional single-layer Reconfigurable Intelligent Surfaces (RIS) that merely reflect signals, a SIM consists of multiple stacked layers of transmissive meta-atoms, as detailed by Ertugrul Basar et al. in Reconfigurable Intelligent Surfaces for 6G: Emerging Hardware Architectures, Applications, and Open Challenges.
Each layer contains an array of subwavelength elements whose electromagnetic properties (such as phase shift) can be individually controlled. Stacking these layers sequentially provides a massive increase in the available degrees of freedom (Hongzheng Liu et al. in Multi-user ISAC through Stacked Intelligent Metasurfaces: New Algorithms and Experiments). As a wavefront passes through the first layer, it is modulated and diffracted toward the second layer, which further shapes the wavefront before passing it to the third, and so on.
This multi-layer cascading effect allows the hardware to perform sophisticated spatial filtering. For instance, the layers can be optimized to physically null out multiuser interference by steering energy peaks toward intended users and deep nulls toward others before the signal ever reaches the wireless channel or the digital receiver (Stacked Intelligent Metasurfaces for Multiuser Beamforming in the Wave Domain; Junjie Fang et al. in Stacked Intelligent Metasurface Assisted Multiuser Communications: From a Rate Fairness Perspective).
Mapping Algorithms to Physical Space
To understand why this works computationally, consider the problem of Direction-of-Arrival (DOA) estimation. Traditionally, estimating where a signal comes from requires sampling the wave at multiple antennas, digitizing each stream, and computing a Discrete Fourier Transform (DFT) in software to find the angular spectrum peaks.
Yong Liang Guan et al. showed that a SIM can be configured to physically approximate a 2D DFT (Two-Dimensional Direction-of-Arrival Estimation Using Stacked Intelligent Metasurfaces). By optimizing the phase shifts of individual meta-atoms across the layers, the physical stack acts as a spatial frequency processor. Incoming waves from a specific angle are naturally focused onto a corresponding energy probe at the output layer. The system identifies the signal's origin simply by detecting which physical probe receives the highest energy, bypassing the need for high-resolution ADCs and digital matrix multiplication (Stacked Intelligent Metasurface Performs a 2D DFT in the Wave Domain for DOA Estimation).
Similarly, this spatial mapping can be extended to complex imaging tasks. For example, Hao Shi et al. integrated a designable metasurface onto an X-ray scintillator, treating the physical layer as an optical convolutional encoder that pre-processes wavefronts to enhance image quality before they hit an electronic sensor (Feature-enhanced X-ray imaging using fused neural network strategy with designable metasurface).
Electromagnetic Neural Networks
Because the layers of a SIM modulate waves sequentially, the architecture functions as a physical implementation of a Diffractive Neural Network or Electromagnetic Neural Network (EMNN). As Haoxian Niu et al. explain, the phase shifts of each meta-atom act exactly like the adjustable weights of a deep neural network (Stacked Intelligent Metasurfaces for Integrated Sensing and Communications).
This insight enables task-oriented semantic communications to occur entirely in the wave domain. In a system designed by Guojun Huang et al., source data (such as image pixels) are mapped directly to the transmission coefficients of the initial metasurface layer (Stacked Intelligent Metasurfaces for Task-Oriented Semantic Communications). As the carrier wave propagates through the subsequent "hidden layers" of the SIM, the physical structure executes the semantic encoding and inference. The output emerges as a directed energy beam tailored for a specific classification or recognition task, eliminating the computational overhead and power consumption of traditional digital encoding.
Hardware Efficiency and the Elimination of RF Chains
Beyond raw speed, the ultimate practical advantage of wave-domain computing is the drastic reduction in hardware complexity. In conventional massive MIMO or near-field beamfocusing systems, scaling up the number of antenna elements requires a proportional scaling of power-hungry RF chains and digital baseband processors (Xing Jia et al. in Stacked Intelligent Metasurface Enabled Near-Field Multiuser Beamfocusing in the Wave Domain).
By shifting beamforming and precoding to the wave domain, a high-frequency base station or a resource-constrained satellite can maintain massive spatial multiplexing capabilities while utilizing significantly fewer RF chains (Shining Lin et al. in Stacked Intelligent Metasurface Enabled LEO Satellite Communications Relying on Statistical CSI).
Furthermore, this principle can be pushed to eliminate traditional upconverters and mixers entirely through direct digital modulation. As explored by Ertugrul Basar et al., a reconfigurable surface can generate signal constellations directly over the air by dynamically altering its reflection or transmission characteristics as an unmodulated carrier wave strikes it (Present and Future of Reconfigurable Intelligent Surface-Empowered Communications [Perspectives]). This mechanism is practically realized by Yifan Xu et al., who demonstrated dynamic beam control and direct image transmission using binary amplitude and phase shift keying (BASK/BPSK) solely by toggling PIN diodes on a metasurface array (Reconfigurable metasurface achieved dynamic beam control and information modulation of L-band beams).
Key Foundational Work
Programmable Metasurface Based Multicast Systems: Design and Analysis (2020) — contributed 2 concept elements to this direction.