Artificial intelligence techniques for the physical layer of non-terrestrial networks

The upcoming 6th Generation (6G) of wireless communication systems aims to extend the capabilities of 5G by enabling immersive, intelligent, and ubiquitous connectivity. Building on 5G, 6G envisions a paradigm shift driven by two key technological enablers: Artificial Intelligence (AI) and Non-Terrestrial Networks (NTNs). Deep Learning (DL) represents a fundamental departure from traditional model-based wireless design, introducing the ability to learn, infer, and optimize directly from data. This potential has been recognized by 3GPP, leading to a Release 18 study dedicated to the application of AI to the New Radio air interface. These efforts mark the beginning of a transition toward AI-native radio access networks, where Physical Layer (PHY) algorithms can self-optimize and adapt to diverse deployment scenarios. In parallel, NTNs extend the communication infrastructure beyond terrestrial boundaries, enabling seamless global coverage and integration across heterogeneous network layers. Their dynamic topologies, long propagation delays, and hardware constraints, however, present new challenges for PHY design and optimization. In this context, the application of AI to the PHY of NTNs emerges as a natural convergence point between these paradigms. By leveraging DL’s capacity for pattern recognition, prediction, and decision-making, PHY algorithms can be tailored to the unique characteristics of non-terrestrial channels. This thesis aims at investigating how AI can enhance PHY functionalities in NTNs, with a particular focus on improving throughput, reliability, and spectral efficiency under the constraints imposed by non-stationary environments and limited computational resources. Three main research directions are addressed: 1) demodulation of Non-Orthogonal Multiple Access (NOMA) schemes; 2) channel prediction for dynamic and resource-constrained NTN scenarios; and 3) equalization of Faster-than-Nyquist (FTN) signaling for bandwidth-efficient transmissions. Each topic is approached through the design and evaluation of DL architectures, emphasizing their ability to learn from realistic channel statistics and approximate optimal PHY behavior while maintaining feasible computational complexity.