Machine learning models for wireless channel modelling

Wireless communication systems are rapidly evolving to support increasingly diverse services and environments, ranging from personal and industrial connectivity to intelligent transportation and non-terrestrial networks. The design and deployment of such systems fundamentally rely on accurate characterization of the radio propagation channel, which governs how electromagnetic waves travel from the transmitter to the receiver while interacting with the surrounding environment. Traditional radio propagation modeling approaches include empirical and deterministic techniques. Empirical models, such as Okumura–Hata and COST-231, provide computationally efficient estimates derived from large-scale measurements but capture only averaged trends. Conversely, deterministic methods like Ray Tracing (RT) offer site-specific predictions by explicitly modeling reflection, diffraction, and scattering phenomena, yet they require extensive geometric information and entail high computational cost, making them less suitable for large-scale or dynamic scenarios. The emergence of 5G, 6G, and the Internet of Things (IoT) reinforces the need for fast, flexible, and accurate propagation modeling tools. In this context, Machine Learning (ML) represents a promising paradigm capable of learning complex relationships between environmental features and radio channel behavior directly from data. Once trained, ML models can deliver accurate predictions at significantly lower computational cost than RT, with improved generalization across environments. This thesis develops and evaluates ML methodologies to address electromagnetic problems in wireless systems, including regression tasks for received power, path loss, shadowing, fast fading, and delay spread, as well as classification tasks such as Line-of-Sight detection and object recognition via backscattered signals. The study investigates tabular-data models (XGBoost, Multilayer Perceptron) and image-based deep architectures (1D CNNs, UNet). Ground-truth datasets are generated through deterministic RT simulations in industrial and urban environments, enabling systematic performance comparison against classical empirical formulations. Results highlight the potential of ML-based models as scalable, computationally efficient, and accurate tools for wireless channel characterization.