Physics-Guided Diffusion Model Boosts Wireless RSRP Prediction by 37%
新智元 · wechat · 2026-09-12
A joint Tsinghua–BUPT team proposed Channel-Diff, a physics-guided diffusion model that frames RSRP prediction as conditional generation: stable large-scale propagation is handled by physics models while random small-scale multipath fading is learned by the diffusion model.
Key design:
- Fuses 10 network parameters with multi-attribute urban maps (terrain + building heights)
- Physics layer encodes path loss (Hata/WINNERII), Fresnel-zone occlusion and multipath geometry
- Teacher stage learns large-scale propagation first; student stage fits real measurements for small-scale residuals, with noise-prior guidance weighting physics vs. data by occlusion level
Results: On two real-world datasets (5.25M+ measurements, 180 radio maps), Channel-Diff beats the runner-up by 37.19% and 25.15% overall; 83% of predictions fall within the ±9.5dB 5G accuracy requirement. Zero-shot cross-dataset transfer still gains 20%+, few-shot (5%-20% data) gains 17%-22%. Ablations confirm the micro-environment feature network (MFEN) matters.
Paper: https://ieeexplore.ieee.org/document/11680013; Code: https://github.com/MichaelTsii/Channel-Diff
More from Research
- FlyGPT lands on Hugging Face: trainable LM wired as a fruit-fly connectome — QuixiAI · 2026-09-14
- FlyGPT trains a language model on a real fruit-fly brain connectome topology — QuixiAI · 2026-09-14
- 4 sink tokens + 64-token window matches distilled linear attention, no training needed — burny_tech · 2026-09-14
- Oak Ridge NL builds fully automated AI system that assembles molecules atom-by-atom for 25+ hours — jiqizhixin · 2026-09-14
- CMU launches Benchmark Radar, a living search engine for AI benchmarks — CarnegieMellonU · 2026-09-14
- Two weeks, two courses: writing a flow matching post by timing what to hide — ariG23498 · 2026-09-14