Interference-Aware Complex Phasor Message Passing Network for Joint Beamforming Optimization in Multi-User IRS-Assisted 6G Systems

Abstract Intelligent Reflecting Surface (IRS)-assisted communication has emerged as a key enabler for next-generation 6G wireless networks, offering enhanced spectral efficiency through programmable propagation environments. However, joint optimization of active transmit beamforming and passive IRS phase shifts remains a challenging non-convex problem due to the coupled amplitude–phase interactions and multi-user interference. In this paper, we propose an Interference-Aware Complex Phasor Message Passing Network (CPMPN) for joint beamforming optimization in multi-user IRS-assisted systems. The proposed framework introduces a novel complex-valued graph representation that preserves the intrinsic amplitude and phase structure of wireless channels, enabling accurate modeling of signal superposition and interference dynamics. An interference-aware message passing mechanism is developed to adaptively weight channel interactions based on propagation strength, while a feasibility-preserving phase output layer ensures physically valid IRS configurations. In addition, a joint prediction module simultaneously generates transmit beamforming vectors and IRS phase shifts within a unified learning framework. The proposed method is evaluated using the DeepMIMO ray-tracing dataset under realistic millimeter-wave propagation conditions. Simulation results demonstrate that CPMPN achieves an average achievable sum rate of 26.3 bps/Hz, outperforming state-of-the-art optimization and learning-based methods including deep unfolding, Transformer-based beamforming, and graph neural networks. Furthermore, the proposed framework approaches the semidefinite relaxation upper bound with a small optimality gap of approximately 3%, while significantly reducing computational complexity through single-pass inference. Comprehensive ablation and statistical analyses confirm the effectiveness and robustness of the proposed design. These results highlight the potential of physics-aware complex graph learning for scalable and real-time beamforming optimization in future 6G wireless systems.