Dynamic Hybrid Beamforming for Near-Field ISAC: From Model-Based Optimization to GNN-Based Learning

Abstract

The phenomena of distance-dependent degrees of freedom (DoFs) and spherical wavefronts in the near-field region open new avenues for the development of integrated sensing and communication (ISAC). However, most existing works that employ hybrid beamforming architectures lack the flexibility to accommodate spatially varying DoFs and thus do not fully exploit near-field characteristics. In this paper, we consider a near-field ISAC system with a hybrid analog/digital architecture, where the number of active radio frequency (RF) chains is dynamically adjusted to balance the communication rate, sensing accuracy, and power consumption. Specifically, we formulate an energy efficiency maximization problem under the transmit power and sensing beampattern constraints. We develop a model-based optimization framework that jointly optimizes the hybrid beamforming matrices and RF chain selection. To reduce the computational complexity and enhance scalability, we develop a graph neural network with constraint-preserving reparameterization (GNN-CPR), which leverages a knowledge-guided learning mechanism to learn the beamforming matrices through graph-structured feature aggregation while satisfying the constraints. Simulation results demonstrate that the proposed model-based optimization and GNN-CPR algorithms achieve higher energy efficiency than the baseline schemes. While the proposed model-based optimization algorithm provides better performance, the proposed GNN-CPR algorithm offers a significant advantage in computational efficiency.