AI-Driven Optimization of Wave-Controlled Reconfigurable Intelligent Surfaces

TL;DR

This work tackles the challenge of configuring a wave-controlled RIS without explicit channel state information by adopting a data-driven pipeline that links biasing standing-wave amplitudes $\boldsymbol{W}$ to the RIS radiation pattern via a neural network (NN). The NN architecture is optimized with a genetic algorithm (GA), while offline beamforming is refined by simulated annealing (SA) that uses NN outputs as feedback; successful configurations are cached in a lookup table for fast real-time RIS operation. The approach accounts for hardware non-idealities and element coupling by learning from data rather than relying on imperfect physics-based models, and it demonstrates that SA driven by the NN can achieve SLNR performance comparable to a full physics-based simulator, even for interpolated angles. Collectively, the methodology enables CSI-free, scalable RIS control with potential for rapid beam steering and pattern synthesis in complex environments.

Abstract

A promising type of Reconfigurable Intelligent Surface (RIS) employs tunable control of its varactors using biasing transmission lines below the RIS reflecting elements. Biasing standing waves (BSWs) are excited by a time-periodic signal and sampled at each RIS element to create a desired biasing voltage and control the reflection coefficients of the elements. A simple rectifier can be used to sample the voltages and capture the peaks of the BSWs over time. Like other types of RIS, attempting to model and accurately configure a wave-controlled RIS is extremely challenging due to factors such as device non-linearities, frequency dependence, element coupling, etc., and thus significant differences will arise between the actual and assumed performance. An alternative approach to solving this problem is data-driven: Using training data obtained by sampling the reflected radiation pattern of the RIS for a set of BSWs, a neural network (NN) is designed to create an input-output map between the BSW amplitudes and the resulting sampled radiation pattern. This is the approach discussed in this paper. In the proposed approach, the NN is optimized using a genetic algorithm (GA) to minimize the error between the predicted and measured radiation patterns. The BSW amplitudes are then designed via Simulated Annealing (SA) to optimize a signal-to-leakage-plus-noise ratio measure by iteratively forward-propagating the BSW amplitudes through the NN and using its output as feedback to determine convergence. The resulting optimal solutions are stored in a lookup table to be used both as settings to instantly configure the RIS and as a basis for determining more complex radiation patterns.

Figures (11)

• Figure 1: Wave-controlled RIS composed of two physical layers 10742896. Top layer: $M$ elements in each row along the $x$ direction, where each element is connected to a varactor diode. Bottom layer: $N$ BSWs are excited on each TL to create the biasing voltages sampled at each RIS element. Each row is controlled via the connection on the left where a signal with $N$ adjustable frequency components is injected by a waveform generator. Adjacent metallic patches and varactors on the top layer, and adjacent DC voltage outputs on the bottom layer are uniformly separated by distance $d_x$ in the $x$ direction and $d_y$ in the $y$ direction.
• Figure 2: RIS unit cell geometry. Each rectangular metallic conductor is biased by the sampled voltage through a via and connected to a grounded varactor. The reverse-biased varactors act as tunable capacitors to polarize the incident electric field along the $y$ direction.
• Figure 3: Geometry of the biasing TL. The length $L_\mathrm{p}$ of the TL path for one unit cell and the distance $d_x$ between two adjacent rectifier circuits are detailed. These circuits, which rectify the BSWs 10742896, are located at the bottom, with one assigned to each RIS element. The voltage $w(x_m,t)$ is extracted from the biasing TL at the location $m$, while the rectified voltage $w(x_m)$ provides the bias to the varactor of the $m$th element. $w(x_m,t)$ is rectified using the diode $D_\mathrm{r}$ and by following its envelope or peaks through the $RC$ circuit shown, with a carefully chosen time constant to minimize voltage drops in $w(x_m)$ due to capacitor discharge.
• Figure 4: Analytical model of the RIS unit cell including an RLC model of the varactor to calculate the reflection coefficient as a function of frequency or varactor biasing voltage. The equivalent impedance $Z_{\text{RIS}}$ is used to find the reflection coefficient $\Gamma$.
• Figure 5: Equivalent resistance and capacitance for the SMV1231-040LF varactor as functions of the biasing voltage across the varactor model shown in Fig. \ref{['fig:circuit_ris']}.