R1-2509354
discussion
AI/ML in 6GR Interface
From CEWiT
Summary
This CEWiT contribution proposes 17 AI/ML-driven use cases for the 6G radio interface across five major areas: CSI overhead and DMRS-based enhancements, integrated sensing and communication (ISAC), network energy saving (spatial/power adaptation and DTX/DRX), and inter-cell beam management. The document provides simulation results for CSI spatial overhead reduction demonstrating significant AI/ML gains and proposes specific study directions for each area.
Position
CEWiT proposes studying AI/ML-based CSI spatial overhead reduction where a UE-sided or NW-sided model predicts full-port CSI from sparse CSI-RS port measurements, with simulation results at 1/4 port reduction showing NMSE of -17.5 dB (AI) versus 13.56 dB (non-AI baseline 2D spline interpolation) for 512 ports. They propose model selection-based AI/ML CSI spatial overhead reduction for a set of sparse CSI-RS port pattern designs including interleaved-based and subpanel-based configurations. For DMRS overhead reduction, they propose studying Sub Use Case A (sparse DMRS across time/frequency) and utilizing some user data symbols with known location and modulation order as pilots to improve channel estimation accuracy. They propose initiating a study on AI/ML-based inter-cell beam prediction and consider location data as essential information, showing prediction accuracy of ~66% for Top-1 beam RSRP. For network energy saving, they support AI/ML enhancement to spatial/power domain adaptations including adapting number of ports/antenna elements/power offsets applied to both UE-specific and cell-specific signals, plus AI/ML-based cell DTX/DRX aligned with UE-DRX. They also propose studying AI/ML-based ISAC framework support for single-sided and two-sided models.
Key proposals
- Proposal 1 (Sec 2.1): Consider AI/ML based CSI overhead reduction in spatial-domain as a potential use case for 6G AI/ML air interface.
- Proposal 5 (Sec 2.2): Study on CSI overhead reduction including the following sub-use case:
- Proposal 6 (Sec 2.3): Consider model selection-based AI/ML CSI spatial overhead reduction for a set of sparse CSI-RS port pattern designs.
- Proposal 9 (Sec 2.4): Study on DMRS-assisted CSI prediction where explore methods to use DMRS signals to enhance or predict CSI.
- Proposal 10 (Sec 2.5.1): We Propose to study Sub Use Case A for the defined configurations:
- Proposal 12 (Sec 2.5.1): We propose to utilize some user data—i.e., symbols whose location and modulation order are known to the receiver—either independently or in combination with DMRS, to improve channel estimation accuracy.
- Proposal 13 (Sec 3.1): Study AI/ML based ISAC framework support for single sided and two-sided model.
- Proposal 14 (Sec 4.1): 6G should support AI/ML Enhancement to spatial and power domain adaptations for energy savings, considering adapting number of ports/antenna elements/power offsets, application to both UE specific and cell specific signals, optimal CSI framework and reporting for various adaptations, and UE assistance information.
- Proposal 15 (Sec 4.2): 6G should support AI/ML based cell DTX/DRX, the enhancement should consider alignment with UE-DRX and User Assistance information.
- Proposal 16 (Sec 5.1): We propose that RAN1 initiates a study on AI/ML-based inter-cell beam prediction to evaluate feasibility of AI/ML models for predicting target cell beam identifiers and parameters, potential overhead reduction and latency improvements compared to conventional approaches, and requirements for dataset exchange and model training in a multi-cell environment.
- Proposal 17 (Sec 5.1): We propose that location data be considered as an essential information in AI/ML-based inter-cell beam prediction studies.
- Proposal 2 (Sec 2.1): Consider UE sided model for AI/ML based CSI-RS spatial overhead reduction, with channel measurements as input.
- Proposal 4 (Sec 2.1): For the evaluation of CSI overhead reduction in spatial-domain consider interleaved based and subpanel based sparse configurations.
- Proposal 7 (Sec 2.3): Consider building specification impact for AI/ML CSI overhead reduction based on the AI/ML use-cases considered in Rel.19.
- Proposal 11 (Sec 2.5.1): We propose to study all possible ways of including channel estimation, interpolation and equalization for implementing AI/ML based receiver using sparse DMRS pattern across time and frequency.