AI-POWERED DIGITAL BEAMFORMING FOR 6G COMMUNICATIONS: TOWARDS GENERALIZED ADAPTIVE BEAM SELECTION

Abstract

The evolution of wireless communication networks toward sixth-generation (6G) systems presents significant challenges in achieving efficient and adaptive beamforming, particularly in millimeter-wave (mmWave) and sub-terahertz (THz) frequency bands. Traditional beamforming techniques rely on exhaustive search-based methods and predefined codebooks, which suffer from high computational complexity, latency, and limited adaptability in dynamic environments. Machine learning (ML) offers a promising alternative by leveraging multimodal data-driven approaches to improve beam selection accuracy, generalization, and real-time adaptability. This thesis explores the integration of ML-based techniques for beamforming, proposing two novel approaches: a deep learning-based beamforming model and a federated transfer learning (FTL) framework. The first approach employs convolutional neural networks (CNNs) and gated recurrent units (GRUs) to process multimodal sensory data, including radar, LiDAR, and GPS signals, for enhanced beam selection in dynamic environments. The second approach introduces a Maturing Federated Transfer Learning (MFTL) framework, which enables privacy-preserving and distributed learning while addressing data heterogeneity across network nodes. The proposed framework significantly improves beam selection performance without requiring centralized data collection by incorporating edge-based training and adaptive aggregation techniques. Extensive experiments were conducted using real-world and simulated datasets, including DeepSense6G and 5G MIMO, to evaluate the proposed models against stateof- the-art (SOTA) methods. Performance metrics such as Top-1, Top-3, and Top-5 accuracy, along with Distance-Based Accuracy (DBA) scores, validate the effectiveness of the proposed ML-based beamforming strategies. Results demonstrate substantial improvements in beam selection accuracy, computational efficiency, and adaptability to dynamic and heterogeneous wireless environments. The findings of this research contribute to the advancement of ML-driven beamforming methodologies, paving the way for more intelligent, scalable, and energyefficient beam selection mechanisms in 6G networks. Future work will explore reinforcement learning-based beam optimization, as the dynamic nature of the problem makes it well-suited for RL. Additionally, the federated learning framework requires further enhancements to effectively address the system's data heterogeneity. A metalearning approach is also a promising option to improve model adaptability and achieve better generalization across diverse scenarios.