Workshop on Physics-based AI for Wireless Communications
Tuesday, 3 June 2025, 14:00 – 17:30, room 0.B
Organisers:
- Máximo Morales Céspedes (Univ. Carlos III de Madrid, ES)
Motivation and Background
In the architectural domain, there is a need for full integration and interoperation between satellite, aerial and terrestrial network components, merged in a unique dynamic‐adaptive network infrastructure denoted as the 3D network. Within this architecture, the evolution of mobile communications needs a combination of several innovative and complementary advances at the physical layer (PHY), medium access control (MAC), and radio resource management (RMM) that may be optimised with the use of Artificial Intelligence (AI) and Machine Learning (ML). This workshop will take place in the framework of PASSIONATE project, whose goals aims to unlock ML for wireless by customising and accounting-by-design the unique properties (“physics-based”) of the networks they are applied to. Physics-based ML is, in addition, the suitable approach to ensure the scalability, generalisation, reliability, and user trust of ML, enabling ML solutions that are technically robust and possibly explainable-by-design.
The PASSIONATE project will pave the way to a complete change of paradigm in the way AI/ML is applied to wireless communications, developing physics-based native AI tools and applying them to the design of innovative PHY, MAC, and RRM techniques.
Structure
14:00 – 14:10: Opening
- Máximo Morales Céspedes
14:10 – 15:00: Keynote
- Towards Resilient Wireless Industrial Communication Networks, Nurul Huda Mahmood, University of Oulu
- General Introduction: 6G, HRLLC (5 mins)
- Industrial wireless communications: intro, challenges (10 mins)
- Conceptual overview of different technological solutions to the challenges, including AI methods (10- 15 mins)
- Examples of AI based solutions with results (5-10 mins)
- Conclusions + QnA (5 mins)
15:00 – 15:30: Technical Presentations
- Data and model-driven machine learning for wireless communications and sensing, Markku Juntti and Nhan Nguyen from University of Oulu
In wireless communications signal processing, conventional optimization-based algorithms often suffer from high computational complexity and long run times, while black-box deep learning models lack interpretability and require extensive training. In this talk, we provide an overview of our research on developing lightweight and efficient machine learning (ML) solutions that enable rapid and stable execution with performance gains and low complexity. We emphasize the advantages of model-based ML, which integrates principled domain knowledge with advanced deep learning techniques, and knowledge distillation, which facilitates effective model compression. Our discussion will cover emerging topics, including hybrid beamforming, reconfigurable intelligent surface (RIS) control, integrated sensing and communications (ISAC), and multi-modal sensing. - Deep-learning-aided ISAC with Superimposed Training in OFDM, Lianet Méndez-Monsanto Suárez, Kun Chen-Hu, María Julia Fernández-Getino García, Ana García Armada, Universidad Carlos III de Madrid
Integrated Sensing and Communications (ISAC) is expected to provide a broad range of novel services and applications in sixth-generation (6G) mobile networks. Achieving this requires a physical layer capable of supporting the coexistence of both target sensing and broadband communications. Current standardization efforts offer this functionality through the use of dedicated pilots, referred to as positioning reference signals (PRS), at the expense of reduced data rates. To address this limitation, the use of superimposed training (ST) has emerged in the literature. ST involves superimposing a pilot signal onto the existing communication waveform, thereby achieving zero time-frequency resource overhead while enabling high-resolution sensing. However, existing methods either rely on multiple antennas for pilot-data separation or require computationally intensive algorithms. To address this, we present a ST-based ISAC method with orthogonal frequency-division multiplexing (OFDM) waveform, where a dual-function pilot signal serving both channel estimation and sensing is superimposed onto the OFDM data signal, combined with a simple matched filtering process at the receiver for simultaneous channel estimation and target detection. To avoid the limitations of traditional threshold-based criteria for tap detection, we introduce a deep learning (DL)-aided target detection method, eliminating the need for prior statistical knowledge. Simulation results demonstrate the effectiveness of the proposed approach, which, when trained offline, achieves high tap detection accuracy and enables efficient simultaneous channel estimation and target localization.
Coffee break
16:00 – 16:50: Keynote
- An AI/ML-defined Radio Interface for Wi-Fi: Benefits and Challenges, Szymon Szott, AGH University of Krakow
Wireless interfaces, previously consisting of fixed, embedded firmwares are now available as fully flexible software-defined radios. The next stage in this evolution will be an AI/ML-defined radio, i.e., a radio architecture specifically designed to support AI/ML-based optimization and decision-making in communication functions. We depict the promised benefits and potential challenges of AI/ML-defined radios and discuss a potential roadmap for their development and adoption. To better introduce the AI/ML-defined radio concept, we focus on its applicability to Wi-Fi and show the benefits of adopting such radios.
16:50 – 17:25: Technical Presentations
- Advancing Wireless PHY Layer with Model-Based AI, Iñigo Bilbao, Eneko Iradier, Alejandro Gonzalez, Jon Montalban, Iñaki Eizmendi, Pablo Angueira, University of the Basque Country (UPV/EHU)
The present moment is characterized by an exponential increase in the use of Artificial Intelligence (AI) in wireless communications systems. This development has led to a situation in which data has become indispensable. The data quality is a pivotal factor in determining the performance of these systems. However, the prevailing data-driven AI models are not equipped to facilitate a comprehensive understanding of their implementation, nor do they permit the coherent adaptation of AI algorithms according to the characteristics of the system and the underlying physics. Consequently, there is a growing interest in AI models based on models and system physics. This paper undertakes a comprehensive study of various wireless communication PHY layer components, proposing model-based AI solutions that offer advantages in areas such as performance, explainability, and complexity. - AI-based blockage prediction for optical wireless communications based on realistic datasets, Máximo Morales Céspedes, Alejandro López Barrios, Juan Carlos Torres Zafra, Universidad Carlos III de Madrid
The accuracy of Artificial Intelligence (AI) in predicting blockage within optical wireless communications (OWC) heavily depends on the datasets used for training. In this work, we compare the performance of AI models trained on measured (realistic) data against those trained on synthetic data. Our analysis underscores the importance of easily accessible parameters, such as illumination and daytime, for effective blockage prediction using AI algorithms. We present both our measurement methodology and the design of these algorithms, with results indicating that blockage can be detected approximately 0.5 seconds beforehand through the use of simple illumination sensors.
17:25 – 17:30: Workshop close.
