High-Fidelity Beam Selection in the Upper Mid-Band Using Digital Twin Ray-Tracing and Deep Learning

초록

This paper presents a ray tracing-based simulation framework for evaluating AI/ML-assisted beam management techniques in realistic urban environments. Using Sionna RT, we construct a high-fidelity 3D digital twin (DT) of a dense metropolitan area, enabling accurate channel characterization with spatial consistency. We systematically assess the impact of key system parameters, including carrier frequency, antenna array configuration, downtilt angle, and codebook granularity, on sum-rate performance. In addition, we propose a deep neural network that infers the optimal beam from low-resolution reference signal received power (RSRP) measurements using the simulator, thereby reducing beam training overhead. Simulation results confirm that the proposed AI/ML-based approach achieves meaningful performance improvements compared to existing methods, especially under low codebook resolution settings. This study demonstrates the practicality and effectiveness of combining ray tracing with learning-based beam selection in next-generation wireless networks. © 2025 IEEE.

키워드