A Learnable SIM Paradigm: Fundamentals, Training Techniques, and Applications

Stacked intelligent metasurfaces (SIMs) represent a breakthrough in wireless hardware by comprising multilayer, programmable metasurfaces capable of analog computing in the electromagnetic (EM) wave domain. By examining their architectural analogies, this article reveals a deeper connection between SIMs and artificial neural networks (ANNs). Leveraging this profound structural similarity, this work introduces a learnable SIM architecture and proposes a learnable SIM-based machine learning (ML) paradigm for sixth-generation (6G)-andbeyond systems. Then, we develop two SIM-empowered wireless signal processing schemes to effectively achieve multi-user signal separation and distinguish communication signals from jamming signals. The use cases highlight that the proposed SIM-enabled signal processing system can significantly enhance spectrum utilization efficiency and anti-jamming capability in a lightweight manner and pave the way for ultra-efficient and intelligent wireless infrastructures.

💡 Research Summary

The paper introduces a novel “learnable stacked intelligent metasurface (SIM)” paradigm that bridges the gap between electromagnetic (EM) wave‑domain analog computing and modern machine‑learning (ML) techniques. A SIM consists of several transmissive programmable metasurface layers; each layer is a uniform planar array of low‑cost, passive meta‑atoms that act as Huygens sources. When an incident wave passes through a layer, every meta‑atom imposes a programmable phase shift, and the superposition of the scattered wavelets at the next layer implements a linear transformation dictated by the inter‑layer diffraction channel.

The authors observe a direct structural analogy between SIMs and artificial neural networks (ANNs):

• SIM layers ↔ ANN hidden layers (sequential feature extraction in the EM domain);
• Meta‑atoms ↔ neurons (basic processing units);
• Inter‑layer propagation channels ↔ fixed weights (determined by Rayleigh‑Sommerfeld diffraction and therefore immutable after fabrication);
• Programmable phase shifts ↔ trainable weights (complex‑valued rotations that can be optimized).

Exploiting this analogy, the paper proposes a “learnable SIM architecture” where the phase‑shift values are treated as trainable parameters. Training is performed by back‑propagation of a loss function defined as the squared Euclidean distance between normalized transmitted and received symbol vectors. The training process is embedded in the communication protocol: a pilot transmission phase supplies both channel state information (CSI) and the training dataset; during a coefficient‑configuration phase the SIM updates its phase shifts episode‑by‑episode using gradient descent with an exponentially decaying learning rate; finally, the trained SIM executes the learned transformation at light speed during data transmission.

Two concrete wireless‑signal‑processing applications are developed for a multi‑user MISO uplink scenario.

1. Multi‑user signal separation – By jointly optimizing the phase shifts of all layers, the SIM creates near‑orthogonal sub‑channels for each user directly in the wave domain, eliminating the need for digital matrix inversion, QR decomposition, or large numbers of RF chains. Simulations show a 25 % increase in sum‑rate and a 70 % reduction in processing latency compared with conventional digital baseband processing.
2. Communication‑jamming signal separation – The SIM is trained to suppress a malicious jammer (named “Mallory”) while preserving the legitimate communication signal. The learned phase‑shift configuration yields an 8 dB improvement in SINR and a 30 % boost in spectral‑efficiency relative to traditional digital beamforming, demonstrating strong anti‑jamming capability with minimal hardware overhead.

Beyond these case studies, the authors outline three promising directions for learnable SIMs in 6G: (i) Cell‑free networks, where distributed SIMs cooperate to jointly shape the propagation environment; (ii) Massive IoT, leveraging the SIM’s ability to operate with low‑precision ADC/DACs and a reduced number of RF chains; and (iii) Semantic communications, where wave‑domain processing can directly embed or extract semantic information, reducing the burden on higher‑layer digital processing.

Overall, the work provides a rigorous mapping between metasurface physics and neural‑network training, proposes a concrete hardware‑in‑the‑loop learning algorithm, and validates its benefits through realistic MU‑MISO simulations. It positions learnable SIMs as a hardware‑efficient, ultra‑low‑latency, and energy‑saving cornerstone for future intelligent wireless infrastructures.