NOMA Assisted Downlink Power Allocation in Pinching Antenna Systems Using Convolutional Neural Network

In this paper, we address the complex resource allocation problem in a flexible-antenna architecture, referred to as a pinching-antenna (PA) system. This system utilizes a dielectric waveguide and dynamically activated small dielectric particles to serve the multiple single-antenna users via non-orthogonal multiple access (NOMA). To address the challenges posed by dynamic operating environments that necessitate real-time reconfiguration, we develop a low-complexity framework for PA placement and power allocation optimization, as traditional iterative methods are computationally prohibitive. We first formulate the optimization problem as a mixed-integer non-linear programming problem to maximize the sum rate performance under physical constraints. Due to the inherent complexity and the need for rapid inference, a two-stage solution is introduced: a user-geometry aware initialization followed by a gradient-based refinement, achieving near-optimal performance efficiently. We then model the power allocation challenge as a max-min fairness problem via quasi-convex programming and low complexity, bisection-based algorithms that achieve globally optimal solutions using only simple scalar evaluations. To enable real-time application, we employ a convolutional neural network (CNN)-based learning framework to capture the highly complex, non-linear mapping between instantaneous channel conditions and optimal power coefficients. This work is motivated by the need to achieve near-optimal performance with significantly lower inference complexity than conventional optimization-based methods. By leveraging the generalization capability of CNNs, the proposed approach enables fast inference and can predict near-optimal power allocations for unseen network configurations without retraining. Simulation results demonstrate that this integrated CNN-based NOMA approach for PA systems delivers enhanced performance in terms of sum rate and user fairness.