The CENTRIC project aims to develop a sustainable AI-native Air-Interface for 6G networks, revolutionizing wireless communication through user-centric design. By leveraging advanced AI techniques, the project focuses on user objectives and application-specific requirements, creating a user-centric AI Air Interface (AI-AI). The project is utilizing advances in machine learning (ML) to enable the development and discovery of efficient waveforms, custom modulations, and transceivers for the physical layer as well as customized lightweight communication protocol and sustainable RRM techniques for the MAC and RRN layers, respectively.

To ensure practical implementation, CENTRIC explores innovative hardware computing substrates with energy-efficiency properties, including neuromorphic computing and mixed analog-digital platforms. The project also explores AI-driven innovation and addresses critical challenges in achieving its 6G AI-AI vision. In this paper, we present CENTRIC’s vision and contributions to AI-native physical layer techniques for 6G networks.

End-to-End Learned Waveforms and Modulation

Over the past decade, machine learning (ML) has significantly impacted various engineering fields, including signal processing in wireless communications. This revolution is evident in the 3GPP work group on ML for 5G Advanced (Rel. 18) and the integration of neural network hardware accelerators into 5G modems by chip manufacturers. Despite its non-newness, the research driving this current wave of ML adoption began less than a decade ago. ML is often used due to model deficits or algorithmic deficits, indicating the lack of reliable mathematical models or effective algorithms.

CENTRIC emphasizes the importance of well-established models and engineering insights in AI-native solutions. Leveraging domain knowledge is instrumental to developing ML-enhanced algorithms that can be generalized with minimal training data. Techniques like MIMO detection and channel estimation are common examples, with “deep unfolding” being a common technique. Another research line aims to replace physical layer algorithms with neural networks, interpreting the transmitter, channel, and receiver as a single neural network or autoencoder. This concept, also known as “End-to-End (E2E) learning”, optimizes the entire communication system from E2E with respect to a chosen loss function. E2E learning can lead to new codes, waveforms, and modulation schemes that are spectrally efficient and hardware-friendly due to lower peak-to-average-power ratio (PAPR) than existing solutions.

Figure 1: Centralized AI models supports sensing
aided beam management in MU-MIMO.

To fully realize the benefits of E2E learning, CENTRIC considers the following research directions as crucial:

• Integration of model-driven and end-to-end learning for new waveforms for the sub-THz band and transmission of short packets are essential components of 6G AI-AI. Existing works have shown potential for data-driven optimization of waveforms, but this approach has not been applied for THz channels and hardware due to phase noise and non-linear effects. With CENTRIC’s E2E learning framework, it is now possible to train the entire communication chain as a single neural network.
• Symbol modulation and demodulation are essential components of the PHY layer. In 5G NR physical downlink shared channel, modulation schemes are used in combination with channel coding to determine data transmission spectral efficiency. Currently used fixed modulation types do not adapt to specific channel conditions or RF impairments. In CENTRIC, deep learning techniques are used to overcome the downsides of fixed modulations and to design a flexible constellation mapper that adapts to real-world channel conditions by dispensing with channel probability distribution models.

AI-empowered MIMO communications

For the last two decades, MIMO communication has been one of the main drivers for boosting the spectral efficiency of modern mobile communication systems. Looking towards the future, the path to further development of MIMO processing is plagued with old and new challenges that AI techniques seem particularly well-suited to overcome: (i) scaling issues in performance and computational complexity resulting from increasing MIMO dimensions; (ii) infeasibility of MIMO precoding due to increased pilot overhead for CSI acquisition and computational burden; and (iii) directional beamforming and beam-based operations required at high frequencies to overcome pathloss and blockage vulnerabilities.

The technical innovations that CENTRIC proposes in the area of MIMO processing to overcome the difficulties described above include:

• 6G AI MU-MIMO neural receiver: MIMO detection has been extensively studied, but current solutions often struggle with realistic channel models or require dedicated neural networks for different system parameters. CENTRIC has now developed a novel neural network-based receiver for MU-MIMO detection which is compliant with 5G NR PUSCH signal while also able to support beyond 5G functionalities including pilotless communications and custom constellation learning. The neural receiver has been evaluated and measured on realistic 3GPP channel and verified in the lab via hardware-in-the loop experiments. Further enhancements to the receiver to short packet transmissions and waveform learning are being explored in CENTRIC. Transfer learning techniques, which transfer knowledge from training models to new configurations or tasks, are also being explored for adapting neural network-based receivers to diverse system parameters.
• CSI compression and prediction: The acquisition of CSI at the transmitter and receiver is a significant challenge for future wireless networks due to the high pilot overhead. CENTRIC has developed novel ML techniques for CSI compression and prediction based on DNN and autoencoders. Learning frameworks for efficient use of ML models for CSI feedback without sharing large data sets are also being developed.
• Sensing aided Beam management in 5GNR focuses on selecting and retaining a proper beam pair between transmitter and receiver for good connectivity. Without sufficient measurement reports, signal blockage may interrupt service, leading to the degradation of quality. Some works propose beam prediction using additional information like position or LIDAR (Light detection and ranging), focusing on localization of individual users for better service. CENTRIC focuses on AI methods for user-centric, sensing aided beam operations in mmWave networks.

Figure 2: Neural network replaces LMMSE equalizer,
demapping and channel estimation blocks in traditional receivers.

Further information

• CENTRIC project website: https://centric-sns.eu/