AI/ML Will Dominate 6G, Starting With The PHY
Artificial intelligence (AI) and machine learning (ML) will impact the complete sixth-generation (6G) cellular ecosystem. AI developments are being evaluated for addressing the complexities of radio beam management, optimizing circuit design, facilitating more efficient traffic flows, reducing power consumption in small circuits and large systems, and predictive maintenance. These optimizations rely on ML models trained on technical data from networks, circuits, and even on synthesized data from simulation and emulation tools. With an AI-driven physical layer (PHY), for example, 6G will benefit from AI optimization at the media access control (MAC) and above layers.
Expect PHY developments in the AI/ML air interface, building on three work items in 3GPP Release 19 5G-Advanced: channel state information (CSI) prediction, beam prediction, and AI-based positioning accuracy improvements, as well as a 3GPP Release 20 5G-Advanced item: CSI compression. Enhancements to these use cases will be studied as well as new promising use cases, such as joint source and channel coding for CSI and channel estimation with AI/ML.
How Could CSI Reporting Differ in 6G?
Successful enhancements have demonstrated more accurate CSI reporting with AI/ML in the air interface compared to traditional approaches. The tradeoff is that engineers must run a lot of simulations to train very complex neural networks and then communicate that information back to the simulator. Part of the complexity at this point stems from the need to evaluate different neural networks to determine benefits. This step will be necessary until those networks are defined in the standard and the industry knows what neural networks to use for the user equipment (UE), base station, etc.
Research also is exploring the benefits and possibilities of implementing AI in the device for CSI compression, beam management, and positioning. Adapting the transmission to the channel conditions results in better radio link performance, thus improving the multiple-input multiple output (MIMO) performance of a given radio channel. This modification requires channel estimation based on the uplink sounding reference signal (UL SRS) at the base station or downlink channel state information reference signals (DL CSI-RS) at the UE.
How Can AI Improve CSI Compression in 6G?
For UE-reported CSI, design engineers must typically make a tradeoff between using detailed information and less overhead. CSI compression helps to obtain better CSI information at the network side with the same overhead as traditional methods, or reduce the overhead without losing valuable CSI information, as shown in the figure. For massive MIMO, CSI reference signals are typically sent out on each antenna to characterize the channel. With the number of antennas rising, however, training data consumes a lot of resources. 6G is expected to actively reduce or increase the amount of CSI RS.
Figure. One 3GPP study topic looks to replace traditional CSI with a compressed version enabled by AI.
Researchers are examining the variability of the demodulation reference symbols inside the physical downlink shared channel (PDSCH) or physical uplink shared channel (PUSCH) with AI/ML. If more data is not being used for reference signals, you can use it for actual user data, such as voice packets, video streams, and application data. In this way, you can get more data through the channel. Using CSI compression (a two-sided model) allows more detailed CSI information to be sent without increasing the UL traffic overhead. With this approach, the base station can get a more complete understanding of the channel and determine a pre-coding matrix optimized for better throughput.
To optimize the PHY with AI/ML, Keysight has used simulation tools as well as emulation and instrumentation. By comparing current 5G signals and channel models, the team showed that AI/ML provided performance improvements over traditional 5G approaches to CSI feedback, positioning accuracy, and beam management. Note that research currently uses the same signals as 5G, but simulation and prototyping requires the integration of some new features like alternate modulation schemes. Another finding was that a new workflow was needed for the radio-frequency (RF) communications and AI/ML teams to work together, so a collaboration tool was implemented: a new generation simulator with an AI-native platform that allows communications and data scientists to work on the same platform.
Engineers also are investigating some of the 5G-Advanced AI/ML work items ahead of 6G. Other studies are exploring how to train AI/ML models and validate AI/ML solutions. With AI/ML procedure test in the Keysight UXM Wireless Test Platform, you can develop AI/ML models, run them in a real-time inference engine, and test key performance indicators (KPIs) for end-to-end performance.
To find out more about how the 6G PHY is evolving and the key use cases, tradeoffs, and technologies influencing it, download the eBook, 6G Research and Innovation: From 0 to PHY and Beyond. For more information on 6G trends and developments, check out the following resources.