R1-2509021
discussion
AI and ML in 6GR air interface
From NVIDIA
Summary
NVIDIA proposes 12 AI/ML use cases across 6G physical layer, aiming for RAN1 selection with preliminary BLER and spectral efficiency gains demonstrated through neural receivers, digital twin validation, and joint transmitter-receiver learning. The proposals cover DMRS overhead reduction (sparse, superimposed, DMRS-less), RS overhead reduction (SRS/CSI-RS), site-specific learning via RAN digital twin, CSI feedback fusion with SRS, beam management for L1/L2 mobility, multi-modal sensing, link adaptation, interference prediction, and anomaly detection.
Position
NVIDIA proposes studying AI-native 6G development through a three-computer workflow spanning design/training, RAN digital twin simulation (notably Aerial Omniverse Digital Twin with ray tracing), and real-time deployment. For DMRS overhead reduction, they propose three progressive receiver levels: sparse DMRS with neural receivers achieving several dB BLER gain versus LS+LMMSE, superimposed DMRS eliminating dedicated pilot REs with 17% spectral efficiency improvement, and DMRS-less transmission using learned constellations eliminating pilots entirely within ~0.2 dB of perfect-CSI bound. For SRS and CSI-RS, they propose AI/ML-based reconstruction from sub-sampled measurements in frequency/time/port domains. They propose fusing CSI feedback with SRS measurements at network-side decoder, extending two-sided CSI compression from 5G-Advanced with SRS as auxiliary input. For beam management, they propose extending Release-19 beam prediction from intra-cell to inter-cell L1/L2-triggered mobility. For link adaptation, they propose reinforcement learning agents observing ACK/NACK and CQI sequences for MCS selection. They propose AI/ML for interference-plus-noise covariance matrix prediction to feed MMSE equalizers and ducting event identification from multi-cell RSRQ drift patterns. For anomaly detection, they propose ingesting multi-dimensional telemetry including time-series CSI metrics, abrupt SNR drops, HARQ ACK/NACK patterns, and beam-quality indicators to output anomaly/fault scores.
Key proposals
- Proposal 1 (Sec 2): Study AI/ML for site specific learning with RAN digital twin, using a CNN-based channel estimator replacing MMSE with 40% uplink throughput gain.
- Proposal 2 (Sec 3): Study sparse DMRS with AI receivers for channel estimation only or joint channel estimation/equalization/demapping, reducing DMRS density in time/frequency.
- Proposal 3 (Sec 4): Study superimposed DMRS with AI receivers, overlaying known pilot sequences directly on data REs to eliminate dedicated pilot resource elements and increase spectral efficiency by 17%.
- Proposal 4 (Sec 5): Study DMRS-less transmission with AI receivers using learned constellations and joint channel estimation/equalization/demapping, eliminating pilot symbols entirely for ~8% spectral efficiency gain.
- Proposal 5 (Sec 6): Study sparse SRS with AI/ML for channel estimation/prediction/beamforming, sub-sampling SRS in frequency/time/port domains and reconstructing dense channels or precoding weights at network side.
- Proposal 6 (Sec 7): Study sparse CSI-RS with AI/ML for CSI feedback, transmitting on subset of antenna ports or frequency PRBs and reconstructing full channel at UE side.
- Proposal 7 (Sec 8.1): Study AI/ML for fusion of CSI feedback and SRS measurements, combining UE compressed CSI feedback with network-side SRS to improve CSI accuracy or reduce feedback overhead.
- Proposal 8 (Sec 8.2): Study AI/ML for beam management for L1/L2 mobility, extending Release-19 beam prediction from intra-cell to inter-cell for spatial and temporal cases.
- Proposal 9 (Sec 8.3): Study AI/ML for multi-modal sensing using RS-derived tensors and optional non-RF sources to detect objects with range/angle/velocity outputs.
- Proposal 10 (Sec 8.4): Study AI/ML for MCS selection in link adaptation using reinforcement learning on ACK/NACK and CQI sequences.
- Proposal 11 (Sec 8.5): Study AI/ML for interference prediction including covariance matrix prediction for MMSE equalization and ducting interference management using multi-cell observations.