Thèse Conception de Récepteurs Bayésiens et Formation de Faisceaux pour les Communications Thz Mimo Multiutilisateurs Utilisant des Modèles Graphiques Probabilistes et des Réseaux de Neurones sur H/F - 75

Établissement : Université Paris-Saclay GS Sciences de l'ingénierie et des systèmes
École doctorale : Sciences et Technologies de l'Information et de la Communication
Laboratoire de recherche : Laboratoire des Signaux et Systèmes
Direction de la thèse : Antoine BERTHET ORCID 0000000205246814
Début de la thèse : 2026-10-01
Date limite de candidature : 2026-04-30T23:59:59

Les communications sans fil dans les bandes sub-THz et THz sont considérées comme un facteur clé pour les réseaux 6G futurs, en raison de l'abondance de spectre inutilisé et du potentiel des systèmes MIMO ultra-massifs (UM-MIMO) capables d'atteindre des débits de données de l'ordre du térabit par seconde et une localisation précise. Cependant, des défis tels que les pertes de propagation importantes, les blocages, les imperfections matérielles et les limites des conceptions de formes d'onde conventionnelles compliquent l'estimation fiable des canaux, l'alignement des faisceaux et les transmissions multi-utilisateurs. Ce projet vise à relever ces défis en exploitant les modèles graphiques probabilistes (PGM) et les réseaux de neurones graphiques (GNN) pour améliorer la conception bayésienne des récepteurs et le beamforming dans les systèmes MIMO THz multi-utilisateurs. En combinant les forces complémentaires des PGM et des GNN pour modéliser les relations entre données et effectuer un passage de messages efficace, la recherche ambitionne de développer des méthodes avancées d'inférence pour l'égalisation conjointe des canaux, le décodage et l'atténuation des interférences, permettant finalement des communications THz robustes et à haute capacité.

Wireless transmissions in sub-THz and THz bands [1] are widely regarded as an enabling technology for future 6G communications, due to the availability of unused spectrum. The short wavelengths enable to pack a large number of antenna elements at the transmitter and receiver side, thus enabling ultra-massive multiple-input multiple-output (UM-MIMO) with the potential of tera-bits per second data rates and precise localization. Besides traditional backhaul links, deploying sub-THz/THz technologies seems relevant for emerging applications such as vehicular [2] and device-to-device communications. Nevertheless, a number of signal processing processing issues remain largely unsolved in order to make affordable sub-THz/THz technologies available. Firstly, MIMO channels in these frequency bands suffer from huge signal-to-noise degradation due to significant path loss and blockage [3], which can partly be compensated using high-gain beamforming. Physical layer waveform design is also an important challenge. It is questionable whether orthogonal frequency division multiplexing (OFDM) used in 4G/5G is still adequate for sub-THz/THz frequency bands due to complexity issues [4,5]. Howewer for the single-carrier (SC) techniques described in [3], the narrowband beamforming model, where signals received over different antenna elements are identical up to a phase shift, may become invalid so that additional time delays need to be compensated [6]. Accurate channel estimation is known to be a difficult task in UM-MIMO settings [7,8] as it usually relies only on pilot symbols to estimate a large number of complex coefficients for growing array sizes. This problem becomes even more complicated in highly dynamic environments induced by user mobility. Finally, low-cost RF implementations incure hardware impairments such as phase noise, I-Q imbalance and amplifier nonlinearities [9], that need proper compensation.

Specific issues will be addressed in this project. Firstly, beam alignment between the transmitter and the receiver is required in order to overcome the overwhelming path loss (otherwise channel estimation would fail). Such mechanisms to establish communication are often time consuming and imperfect [10]. They also do not account for abrupt user appearance/ disappearance caused by blockage. Secondly, accomodating multiuser transmissions requires space division multiple access (SDMA) or beam division multiplexing (BDM) techniques, where groups of users clustered according to the similarity of their angle of departure (AoD) can be resolved [4]. Specific interference cancellation techniques need to be applied in order to cancel inter-group interference [11]. It is worth pointing out that the problem of sub-THz/THz communications with potential beam misalignment or blockage has deep connections with multitarget estimation for direction-of-arrival (DOA) detection and tracking in radar theory [12].

Graphs are a powerful data structure to represent relational data and are widely used to describe complex real-world data structures. Probabilistic graphical models (PGMs) have been well developed in recent years to mathematically model real-world scenarios in compact graphical representations of variable distributions [13,14]. Graphic neural networks (GNNs) are new inference methods developed in recent years and are attracting increasing attention due to their efficiency and
exibility in solving inference and learning problems on graph-structured data, see e.g., [15]. These two powerful approaches have dierent advantages in capturing relationships from observations and how they drive message passing, and they can enrich each other in various tasks. Both GNNs and PGMs are capable of conducting accurate and fast inference on graphs but the major differences between them are how they process relational variables and how they perform message passing in the models.

The possible intersection, and even association, of these two approaches is a promising idea that is spreading in the field of communications theory. PGMs and belief propagation (BP) joint channel equalization and decoding is investigated in e.g., [16,17,18]. In [19,20], GNNs are used for channel decoding. Very recently, [21] has proposed the use of GNNs for joint channel equalization and decoding, one GNN for channel equalization and one GNN for channel decoding. Both GNNs are connected by their variable nodes (VNs) to form a combined factor graph [17] that allows information exchange between equalization and decoding. This is in contrast with [22] where such a combined factor graph is employed in combination with a decision feedback equalizer (DFE) and a channel decoder based on PGMs and BP.

The proposed research project aims to explore when and how GNNs and PGMs combined can improve Bayesian receiver design and beamforming for multiuser MIMO THz communications.

Wireless communications in sub-THz and THz bands are considered a key enabler for future 6G networks due to the abundance of unused spectrum and the potential for ultra-massive MIMO (UM-MIMO) systems that achieve terabit-per-second data rates and precise localization. However, challenges such as severe path loss, blockage, hardware impairments, and the limitations of conventional waveform designs complicate reliable channel estimation, beam alignment, and multiuser transmission. This project focuses on addressing these issues by leveraging probabilistic graphical models (PGMs) and graph neural networks (GNNs) to enhance Bayesian receiver design and beamforming in multiuser THz MIMO systems. By combining the complementary strengths of PGMs and GNNs in modeling relational data and performing efficient message passing, the research aims to develop advanced inference methods for joint channel equalization, decoding, and interference mitigation, ultimately enabling robust and high-capacity THz communications.

The proposed research will adopt a combination of analytical modeling, simulation, and data-driven methods to address the challenges of sub-THz/THz multiuser MIMO communications. First, a comprehensive system model will be developed, capturing channel characteristics, beamforming architectures, hardware impairments, and user mobility effects. PGMs will be formulated to represent the joint distribution of channel coefficients, beam directions, and transmitted symbols, enabling efficient Bayesian inference for channel estimation and multiuser detection.

In parallel, GNNs will be designed to learn the complex dependencies in the multiuser MIMO system, with architectures tailored for joint channel equalization and decoding. Hybrid PGM-GNN frameworks will be explored to integrate the rigorous probabilistic structure of PGMs with the flexibility and learning capability of GNNs, allowing iterative message passing and information exchange between beamforming, channel equalization, and decoding modules.

Performance will be evaluated through extensive numerical simulations, considering metrics such as achievable data rates, beam alignment accuracy, robustness to blockage, and computational complexity.