Joint Beam Scheduling and Precoding for LEO Satellites Via Physics-Aware GNN

Abstract

Joint beam scheduling and precoding in beam-hopping (BH) low earth orbit (LEO) satellite systems is formulated as a mixed-integer nonlinear programming (MINLP) problem. Conventional alternating optimization (AO) algorithms suffer from high computational complexity due to iterative solving, failing to meet the real-time requirements of large-scale phased array satellites. To address this, we propose a physics-aware graph neural network (GNN) framework for joint optimization. Aiming at maximizing the system sum-rate subject to per-slot power budget and quality-of-service (QoS) constraints, we design a bipartite graph architecture that models the antenna-user interaction topology. A sequential scheduling module is developed to construct beam illumination patterns by jointly considering channel quality and inter-user channel correlation. A distributed edge-based beam-forming module synthesizes beamforming vectors from latent graph embeddings. Simulation results demonstrate that the proposed framework achieves more effective beam scheduling and improved sum-rate across different system configurations compared to conventional AO benchmarks, while reducing computation time by two orders of magnitude to the millisecond level.