Learning Geometry-Aware Radio Propagation Models for Channel Knowledge Map Construction in 6G Networks

Sixth Generation (6G) wireless networks are expected to integrate non-terrestrial platforms and operate over high-frequency spectra, superseding the limitation in Beyond-Fifth Generation (B5G) systems. In such networks, the high-frequency radio propagation behavior is increasingly sensitive to the surrounding physical environment. This thesis is grounded in the observation that channel behavior is not solely influenced by transmitter–receiver separation distance but is inherently dependent on environmental structures. This work investigates how channel knowledge can be learned directly from the communication environment using Machine Learning (ML) techniques. Since conventional channel modeling approaches often fail to provide environment-adaptive solutions suitable for real-world deployment, this thesis develops learning-based methodologies that are environment-aware, interpretable, and data-efficient, particularly for Unmanned Aerial Vehicle (UAV)-assisted communication networks. The thesis proposes a sequence of learning frameworks that move from statistical channel generation toward spatially grounded, representation-level modeling and cross-modal knowledge transfer strategies. These models offer a variety of computationally efficient alternatives to accurate deterministic methods while preserving adaptability and generalization. These contributions support the development of environment-aware channel models for Channel Knowledge Map (CKM) construction. The first contribution is the design of an explainable generative channel modeling framework that learns statistical distributions of multipath channel parameters directly from data. A generative neural architecture captures complex distributional patterns of channel characteristics while conditioning on transmitter–receiver spatial distance, generating synthetic channel characteristics for multipath components. A key novelty is the integration of the eXplainable Artificial Intelligence (XAI) method into the training process, where meaningful features are weighted to guide gradient updates, improving predictive accuracy and data efficiency of the model. The second contribution is an environment-aware probabilistic framework for channel state modeling. Rather than treating this task as a deterministic classification problem or focusing on the estimation of a single channel state probability, the thesis formulates channel state prediction as a unified probabilistic learning task. A comprehensive set of geometrical descriptors is extracted from real urban environments around each receiver location and applied to train several state-of-the-art supervised ML models that simultaneously estimate the probabilities of Line-Of-Sight (LOS), Non-Line-Of-Sight (NLOS), and Blocked states. Extensive evaluation across multiple city sections with diverse building profiles demonstrates satisfactory generalization of ML models to previously unseen environments. Furthermore, explainability analysis is employed to identify the dominant geometrical factors governing each channel state probability output, confirming the physical consistency of the learned relationships. Furthermore, through the use of important features derived from the explainability analysis, we augment a benchmark empirical LOS probability model by reparameterization, yielding a geometry-sensitive formulation with improved accuracy for heterogeneous urban scenarios. The third and main contribution addresses large-scale signal attenuation modeling through a novel Path Loss (PL) prediction framework based on cross-modal transfer learning. A three-dimensional (3D) convolutional neural network is introduced that directly processes the volumetric representation of urban environments surrounding the receiver, learning spatial geometric features to estimate PL. To exploit complementary information from heterogeneous data modalities, including geometric and channel-domain data, which illustrate different aspects of the communication environment and may not all be available during deployment, a cross-modal distillation strategy is proposed in which the 3D student model obtains valuable information from teacher models trained on these different data modalities. Both output-level and intermediate-layer distillation strategies are investigated, demonstrating significant improvements in the prediction accuracy of the student model. Building upon this framework, the thesis further introduces an adaptive dual-teacher distillation mechanism in which the contributions of multiple teachers are dynamically weighted using a softmax-inspired strategy based on their temporal loss evolution. In addition, a novel latent-space distillation strategy based on Kullback–Leibler divergence is proposed to constrain the internal latent representation of the student to follow the posterior distributions of the teachers, resulting in enhanced training stability and superior generalization. One direct application of the proposed models is the generation of region-based channel knowledge for the construction of 3D CKMs, required for environment-aware operation in future wireless networks.