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

In this paper, we consider a flexible-antenna architecture, referred to as a pinching-antenna (PA) system, in which multiple PAs realized by activating small dielectric particles along a dielectric waveguide are jointly employed to serve a single-antenna user. We investigate antenna placement and power allocation optimization in PA-assisted non-orthogonal multiple access (NOMA) systems using a convolutional neural network (CNN). An optimization strategy is developed to determine the PA locations that maximize achievable NOMA performance while satisfying physical and spatial constraints. The proposed method adopts a two-stage structure, combining a user-aware initialization with a gradient-based refinement, enabling near-optimal performance with significantly lower computational cost. A max-min fairness formulation is introduced for power allocation to balance the power budget among users with varying channel strengths, solved efficiently via quasi-linear programming and bisection search. Finally, a CNN-based learning framework is employed to capture the nonlinear mapping between channel conditions and the corresponding optimal power coefficients. This framework can infer near-optimal power allocations for unseen network configurations without retraining, offering scalability and adaptability. Simulation results show that the proposed CNN-based NOMA approach for PA systems improves sum rate and user fairness while reducing computational complexity.