Learning-based Enhancements for Sub-THz Communications: Robust Transceivers and Beam Management
Thesis event information
Date and time of the thesis defence
Place of the thesis defence
OP auditorium (L10), Linnanmaa
Topic of the dissertation
Learning-based Enhancements for Sub-THz Communications: Robust Transceivers and Beam Management
Doctoral candidate
Master of Science (Technology) Dileepa Marasinghe
Faculty and unit
University of Oulu Graduate School, Faculty of Information Technology and Electrical Engineering, CWC - Radio Technologies
Subject of study
Communications Engineering
Opponent
Professor Jyri Hämäläinen, Aalto University
Custos
Professor Nandana Rajatheva, University of Oulu
Learning-based Enhancements for Sub-THz Communications: Robust Transceivers and Beam Management
This thesis focuses on designing data-driven, learning-based techniques to address three critical challenges in enhancing sub-terahertz (sub-THz) communications: the increasing phase noise (PN) necessitating PN-robustness, the modest output power of power amplifiers (PA) calling for constraints on the peak to average power ratio (PAPR), and resource-intensive beam management (BM) procedures demanding predictive measures.
The first part of the thesis focuses on designing PN-robust and low-PAPR single-carrier (SC) waveforms and neural network (NN)-based receiver techniques. First, PAPR-constrained geometric constellation shaping under residual PN is explored, based on an SC transceiver. To further improve PAPR gains, the trainability is extended to pulse-shaping filters alongside an adjacent channel leakage ratio (ACLR) constraint to control spectral leakage with a given excess bandwidth. At the receiver, state-of-the-art PN-robust demapping techniques and a novel NN demapper are explored for optimized signal detection. Furthermore, a novel PN-resilient neural transceiver incorporating a deep NN receiver with a trainable pilot scheme is presented, reducing the pilot overhead and thus improving the spectral efficiency while ensuring reduced PAPR. The waveform optimization problems are formulated end-to-end, where the objectives are converted to their augmented Lagrangian form, and a backpropagation-inspired technique is employed to obtain numerically robust designs for PN while adhering to constraints. The method’s efficacy is substantiated through simulation results, demonstrating gains in the block error rate (BLER), spectral efficiency (SE), and PAPR reductions. The practical applicability of the solutions is then verified on a proof of concept (PoC) system operating with commercial-grade sub-THz hardware.
The second part of the thesis focuses on reducing resource-intensive periodic measurements for BM in mmWave/sub-THz systems. A novel machine learning (ML) model and a predictive BM scheme are proposed. Leveraging past beam decisions and user tracking from a light detection and ranging (LiDAR) system, the model predicts the next beam from a predefined codebook. Results demonstrate significant RF resource savings while maintaining the accuracy of the beam decisions.
The first part of the thesis focuses on designing PN-robust and low-PAPR single-carrier (SC) waveforms and neural network (NN)-based receiver techniques. First, PAPR-constrained geometric constellation shaping under residual PN is explored, based on an SC transceiver. To further improve PAPR gains, the trainability is extended to pulse-shaping filters alongside an adjacent channel leakage ratio (ACLR) constraint to control spectral leakage with a given excess bandwidth. At the receiver, state-of-the-art PN-robust demapping techniques and a novel NN demapper are explored for optimized signal detection. Furthermore, a novel PN-resilient neural transceiver incorporating a deep NN receiver with a trainable pilot scheme is presented, reducing the pilot overhead and thus improving the spectral efficiency while ensuring reduced PAPR. The waveform optimization problems are formulated end-to-end, where the objectives are converted to their augmented Lagrangian form, and a backpropagation-inspired technique is employed to obtain numerically robust designs for PN while adhering to constraints. The method’s efficacy is substantiated through simulation results, demonstrating gains in the block error rate (BLER), spectral efficiency (SE), and PAPR reductions. The practical applicability of the solutions is then verified on a proof of concept (PoC) system operating with commercial-grade sub-THz hardware.
The second part of the thesis focuses on reducing resource-intensive periodic measurements for BM in mmWave/sub-THz systems. A novel machine learning (ML) model and a predictive BM scheme are proposed. Leveraging past beam decisions and user tracking from a light detection and ranging (LiDAR) system, the model predicts the next beam from a predefined codebook. Results demonstrate significant RF resource savings while maintaining the accuracy of the beam decisions.
Last updated: 12.2.2025