AI and ML for 6G networks

6G artificial intelligence and machine learning

Even though artificial intelligence is one of the ten main 6G research areas, it is not a standalone research area. It still plays into all of the other areas however, such as cell-free massive MIMO, full-duplex communication and intelligent reflecting surfaces. The performance of every single example can be enhanced by data-driven, trained systems in 6G networks, increasing energy efficiency and therefore also sustainability at the same time. Using trained machine learning models for signal processing tasks like channel estimation, equalization and demapping will further optimize the air interface compared to present-day 4G LTE and 5G NR networks.

Rohde & Schwarz supports research activities across Europe, Asia and the US and works as a partner in projects like the 6G-Access, Network of Networks, Automation & Simplification (6G-ANNA) lighthouse project. This project aims to develop a design for 6G that includes end-to-end architecture and simplifies the interaction between humans, technology, and the environment using new sensors and algorithms to detect human movements.

Your AI challenge for 6G networks

Establishing an AI-native air interface for 6G networks means replacing blocks in the signal processing chain on the physical layer with trained machine learning models. The first step in this process is to replace individual processing blocks but ultimately combine tasks that logically belong together in one trained machine learning model. Such tasks are channel estimation, channel equalization and demapping. These tasks are combined and replaced with one single trained ML model known as a neural receiver.

However, the signal processing for the 6G air interfaces is just one area where the use of ML may provide an advantage. Another area is the linearization of power amplifiers or the entire RF Frontend used in today’s mobile devices and base stations. AI or ML can be applied to 6G for the air interface and RF Frontend during several different phases:

Phase 1: Initially, ML may replace today’s deterministic software-algorithm-based linearization models for power amplifiers with ML. Research already started in this field in 2020 and is primarily driven by universities. Key industry players also have conducted studies on this subject however. This process is also set to be applied to the entire RF Frontend (= antenna system and transceiver).

Data accessibility is a clear challenge when it comes to artificial intelligence for 6G. This is because access to data sets is required to train a neural network. The RF Frontend is typically designed by one vendor. This means that all necessary data for training neural networks is in the hands of one single vendor – making it easier to realize this phase.

Phase 2: This phase focuses on receiver aspects, applying the concept of a neural receiver, by replacing signal processing blocks such as channel estimation, channel equalization and demapping with a trained ML model.

Phase 3: This is where end-to-end (E2E) optimization comes in. ML is used to jointly optimize TX, RX, and baseband processing. The ultimate goal during this phase is to adapt the transmission to the underlying application (voice call, web browsing, XR, etc.) and deployment scenario of the impact of the transmission channel with the ML designs being a part of 6G PHY / MAC itself. A first step towards E2E learning is the replacement of the modulation mapper with a custom, learned constellation that perfectly adapts to the imperfections of the transmitter, receiver and the impact of the wireless channel. Custom modulations allow a pilotless transmission and thus further improves the performance of the overall system.

Such highly adaptive physical layer implementations require extensive verification prior to deployment in the field. This verification requires models operating reliably – even under the rare conditions observed in the field. However, trained AI/ ML models are only as good as the training data they have been trained with. This is where AI/ML model lifecycle management (e.g. model training, selection, exchange, activation and monitoring) comes in, as frequent collaboration between the user devices and base station/network is expected. Testing and measurement must verify smooth interoperability between the components provided by different vendors.