R1-2509302
discussion
Use cases for AI/ML in 6GR interface
From KT
Summary
This document from KT Corp. proposes five new AI/ML use cases for the 6G radio interface study, comprising three extensions to 5G-Advanced beam management and two entirely new use cases for CSI and DMRS overhead reduction, with a mix of UE-sided and NW-sided model considerations.
Position
KT proposes studying beam prediction during initial access using a UE-sided model to predict optimal SSB beams from a measured subset, extending Rel-19 spatial beam prediction concepts of Set A and Set B, and to infer narrow beams from wide-beam measurements to reduce latency. They propose studying inter-cell/TRP beam prediction with both UE-sided and NW-sided models, where the UE measures a subset of beams from multiple cells to infer top-K beams across those cells for spatial and temporal prediction. They propose studying beam failure prediction and candidate beam prediction for BFR to anticipate failures and dynamic candidate beam configuration, reducing RS measurement overhead. KT further proposes studying spatial/frequency domain CSI prediction to reduce CSI-RS overhead by enabling the UE to perform full-port channel estimation when the base station selectively deactivates antenna ports. They also propose studying AI/ML-based DMRS design for dynamic and sparse DMRS configurations to reduce signaling overhead based on bandwidth and channel conditions.
Key proposals
- Proposal 1 (Sec 2.1.1): Study the beam prediction during initial access with UE-sided model as an extension of 5GA BM use case.
- Proposal 2 (Sec 2.1.2): Study the inter-cell/TRP beam prediction with both UE-sided and NW-sided models as an extension of 5GA BM use case.
- Proposal 3 (Sec 2.1.3): Study the beam failure prediction and candidate beam prediction for BFR with both UE-sided and NW-sided models as an extension of 5GA BM use case.
- Proposal 4 (Sec 2.2.1): Study the spatial/frequency domain CSI prediction for CSI-RS overhead reduction with UE-sided model.
- Proposal 5 (Sec 2.2.2): Study the AI/ML based DMRS design for DMRS overhead reduction with both UE-sided and NW-sided models.