Performance Evaluation of AI-enabled Refined Beam Prediction with Wide Beam Measurements

5G New Radio (NR) beam management (BM) procedures facilitate the establishment of highly directional beams in millimeter wave (mm-Wave) communications; however, they require extensive measurements that can be cumbersome. Our study introduces machine learning (ML)-aided BM to significantly reduce this measurement burden by predicting the optimal refined beam from primary wide beam measurements. We propose wide beam codebook designs based on circular shifts and random combinations of refined beams, utilizing an ML model to decode these measurements efficiently. Simulation results indicate that our method enhances prediction performance and that our neural network-based decoder surpasses traditional distancebased decoder. Our approach achieves up to $\mathbf{9 9 \%}$ of average cell-throughput and $94 \%$ of 5 -th percentile user equipment (UE) throughput performance compared to exhaustive refined beam scans, demonstrating a substantial reduction in operational demands.