RIS-Assisted Downlink Pinching-Antenna Systems: GNN-Enabled Optimization Approaches

TL;DR

This work tackles joint optimization in a RIS-assisted multi waveguide PASS downlink by introducing a novel unsupervised three stage GNN that separately learns PA placements, RIS phase shifts, and beamforming vectors. By modeling the system as a graph and enforcing feasibility through specialized layers, the approach achieves fast, real-time inference while maintaining strong performance. Three integration strategies balance optimality and speed, with Strategy II often delivering the best SR and EE and Strategy I offering the fastest latency. The results show notable gains over RIS-only or PA-only baselines, and strong generalization to unseen numbers of users, highlighting the practical potential for scalable RIS-PASS deployments in future wireless networks.

Abstract

This paper investigates a reconfigurable intelligent surface (RIS)-assisted multi-waveguide pinching-antenna (PA) system (PASS) for multi-user downlink information transmission, motivated by the unknown impact of the integration of emerging PASS and RIS on wireless communications. First, we formulate sum rate (SR) and energy efficiency (EE) maximization problems in a unified framework, subject to constraints on the movable region of PAs, total power budget, and tunable phase of RIS elements. Then, by leveraging a graph-structured topology of the RIS-assisted PASS, a novel three-stage graph neural network (GNN) is proposed, which learns PA positions based on user locations, and RIS phase shifts according to composite channel conditions at the first two stages, respectively, and finally determines beamforming vectors. Specifically, the proposed GNN is achieved through unsupervised training, together with three implementation strategies for its integration with convex optimization, thus offering trade-offs between inference time and solution optimality. Extensive numerical results are provided to validate the effectiveness of the proposed GNN, and to support its unique attributes of viable generalization capability, good performance reliability, and real-time applicability. Moreover, the impact of key parameters on RIS-assisted PASS is illustrated and analyzed.

Figures (7)

• Figure 1: Illustration of RIS-assisted PASS for downlink multi-user communications.
• Figure 2: The structure of the proposed three-stage GNN maps the given user locations $\{\bm{\psi}^{\rm U}_k\}$ to a complete solution $\{x^{\rm P}_{n,m}, \bm{\Phi},\mathbf{w}_k\}$ through three sequential stages. Stage 1 (PAGNN) employs $G_1$${\mathbb C}$GCL layers to map the user locations to a set of PA positions $\{x^{\rm P}_{n,m}\}$. Stage 2 (RISGNN) employs $G_2$${\mathbb C}$GCL layers to map the effective channels $\{{\widehat{\bf h}^H_k(\{x^{\rm P}_{n,m}\},\bm{\widetilde{\Phi}})}\}$ computed by \ref{['ec_s2']} to a RIS phase shift matrix $\bm{\Phi}$. Stage 3 (BeamGNN) uses two parallel branches to learn beamforming parameters: after updating the $\{{\widehat{\bf h}^H_k(\{x^{\rm P}_{n,m}\},{\bm\Phi})}\}$ by \ref{['effective_channel']}, the top (bottom) branch employs $G_3$$\mathbb{C}$GAL ($G_5$$\mathbb{C}$GAL) layers and $G_4$$\mathbb{C}$FL ($G_6$$\mathbb{C}$FL) layers to map the obtained $\{{\widehat{\bf h}^H_k(\{x^{\rm P}_{n,m}\},{\bm\Phi})}\}$ to a set of hybrid coefficients $\{\alpha_k\}$ (power allocation $\{p_k\}$). The obtained $\{p_k\}$ and $\{\alpha_k\}$ are then utilized to yield the output beamforming vectors $\{\mathbf{w}_k\}$.
• Figure 3: Illustration of three strategies.
• Figure 4: Impact of number of waveguides on EE and SR, with $L=32$, $M=8$ and $K=4$.
• Figure 5: Impact of number of PAs on EE and SR, with $L=32$, $N=8$ and $K=4$.