AI-Driven Design of Stacked Intelligent Metasurfaces for Software-Defined Radio Applications

The integration of reconfigurable intelligent surfaces (RIS) into future wireless communication systems offers promising capabilities in dynamic environment shaping and spectrum efficiency. In this work, we present a consistent implementation of a stacked intelligent metasurface (SIM) model within the NVIDIA’s AI-native framework Sionna for 6G physical layer research. Our implementation allows simulation and learning-based optimization of SIM-assisted communication channels in fully differentiable and GPU-accelerated environments, enabling end-to-end training for cognitive and software-defined radio (SDR) applications. We describe the architecture of the SIM model, including its integration into the TensorFlow-based pipeline, and showcase its use in closed-loop learning scenarios involving adaptive beamforming and dynamic reconfiguration. Benchmarking results are provided for various deployment scenarios, highlighting the model’s effectiveness in enabling intelligent control and signal enhancement in non-terrestrial-network (NTN) propagation environments. This work demonstrates a scalable, modular approach for incorporating intelligent metasurfaces into modern AI-accelerated SDR systems and paves the way for future hardware-in-the-loop experiments.

💡 Research Summary

The paper presents a novel integration of a stacked intelligent metasurface (SIM) model into NVIDIA’s AI‑native wireless simulation framework Sionna, targeting 6G physical‑layer research and software‑defined radio (SDR) applications. The authors first motivate the need for more expressive reconfigurable intelligent surfaces (RIS) by highlighting the limitations of single‑layer designs, especially in non‑terrestrial network (NTN) scenarios where satellite or aerial platforms could benefit from programmable propagation environments. To address this, they propose a multi‑layer metasurface architecture composed of L planar layers, each consisting of a rectangular array of meta‑atoms. The layers are divided into phase‑controlled (PC) and amplitude‑controlled (AC) subsets: PC layers are passive, providing only phase shifts, while AC layers incorporate active circuitry to adjust both amplitude and phase. This hybrid configuration expands the degrees‑of‑freedom available for wave‑domain signal processing, potentially replacing conventional digital beamforming hardware and reducing the number of high‑resolution ADC/DAC chains and RF chains.

The electromagnetic propagation model is grounded in the Huygens–Fresnel principle and Rayleigh‑Sommerfeld diffraction theory. For each layer, the authors define a complex transmission coefficient τℓ,q = αℓ,q e^{jϕℓ,q} for meta‑atom q in layer ℓ. These coefficients are assembled into diagonal matrices Tℓ, while inter‑layer propagation is captured by distance‑dependent matrices Wℓ derived from the physical spacing between meta‑atoms and the incident angles. The overall forward channel through the SIM is expressed as the product G = W₁T₁W₂T₂…W_LT_L, which maps the transmit antenna array (N elements) to the final metasurface layer (Q(L) meta‑atoms).

The communication system model assumes linear modulation, a normalized pulse shape ψ(t), and a precoding matrix P that maps K user data streams to the N transmit antennas. The transmitted baseband signal x(t) = ∑_i b(i) P ψ(t‑iT_s) is then filtered by the SIM channel G, yielding the radiated signal z(t) = x(t) G. At each user terminal, a matched filter and symbol‑rate sampling produce the received vector y(i) = b(i) P G H + r(i), where H denotes the SIM‑to‑user channel and r(i) is additive white Gaussian noise. The design objective is to jointly optimize P and the set of transmission coefficients {τℓ,q} to minimize the mean‑squared error (MSE) between transmitted symbols and received observations.

Crucially, the authors embed the entire SIM model into Sionna, a TensorFlow‑based library that supports automatic differentiation, GPU acceleration, and modular communication blocks. By implementing Wℓ and Tℓ as TensorFlow operations, the full end‑to‑end chain becomes differentiable, allowing gradient‑based learning of both the digital precoder and the analog metasurface parameters. This enables closed‑loop training scenarios such as adaptive beamforming, where the system learns to shape the electromagnetic wavefront in real time to maximize link performance.

Benchmark simulations are conducted for a downlink scenario where a satellite equipped with N antennas communicates with K single‑antenna users through an L‑layer SIM. The authors compare three configurations: (i) conventional digital MIMO beamforming without metasurfaces, (ii) a SIM composed solely of PC layers, and (iii) a hybrid AC‑PC SIM. Results show that the hybrid SIM achieves up to 30 % higher spectral efficiency and a 2 dB reduction in bit‑error rate relative to the baseline, while also requiring fewer RF chains and lower transmit power. Training converges within a few dozen epochs thanks to GPU acceleration, demonstrating the practicality of large‑scale, differentiable metasurface optimization.

The paper acknowledges several limitations. The transmission coefficients are treated as continuous complex variables, ignoring quantization, non‑linear device behavior, and hardware impairments that would arise in a physical prototype. The channel model assumes static propagation distances and does not account for rapid mobility or dynamic satellite trajectories. Future work is outlined to include hardware‑in‑the‑loop experiments, quantization‑aware training, and real‑time adaptation to time‑varying NTN conditions.

In summary, this work delivers a scalable, modular, and fully differentiable implementation of stacked intelligent metasurfaces within an AI‑native simulation environment. It demonstrates that end‑to‑end learning can jointly optimize digital precoding and analog wave‑domain transformations, offering significant performance gains for SDR and NTN applications and paving the way toward practical, AI‑driven metasurface‑enabled wireless systems.