Sun et al. (2025) Fusion of multi-source precipitation records via coordinate-based generative models

Identification

• Journal: Nature Communications
• Year: 2025
• Date: 2025-12-29
• Authors: Sencan Sun, Congyi Nai, Baoxiang Pan, Wentao Li, Lu Li, Xin Li, Efi Foufoula-Georgiou, Yanluan Lin
• DOI: 10.1038/s41467-025-67987-9

Research Groups

• Ministry of Education Key Laboratory for Earth System Modeling, Department of Earth System Science, Tsinghua University, Beijing, China.
• Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China.
• State Key Laboratory of Hydrology-Water Resources and Hydraulic Engineering, Hohai University, Nanjing, China.
• School of Atmospheric Sciences, Sun Yat-sen University, Guangzhou, China.
• Institute of Tibetan Plateau Research, Chinese Academy of Sciences, Beijing, China.
• Department of Civil and Environmental Engineering and Department of Earth System Science, University of California Irvine, USA.

Short Summary

The study introduces PRIMER, a coordinate-based diffusion model framework that fuses heterogeneous precipitation data from gauges, satellites, and reanalysis. The model successfully overcomes the trade-offs between spatial coverage and local accuracy, providing a unified tool for bias correction, downscaling, and optimal interpolation.

Objective

• To develop a generative Bayesian framework (PRIMER) capable of learning an informative prior from imperfect, multi-source precipitation records to produce high-resolution, physically consistent, and bias-corrected precipitation fields.

Study Configuration

• Spatial Scale: East Asia (20–45° N, 100–125° E), focusing on mainland China, with a target resolution of 0.1° (~10 km).
• Temporal Scale: Hourly resolution; training utilized data from 2000–2020, with testing on 2016 events and future projections for 2050.

Methodology and Data

• Models used: PRIMER (Precipitation Records Infinite MERging), a coordinate-based probabilistic diffusion model utilizing Neural Operators, Sparse Convolutional Residual Blocks (SparseConvResBlocks), and a U-Net architecture.
• Data sources:
  • Reanalysis: ERA5 (0.25° resolution).
  • Satellite: GPM IMERG (0.1° resolution).
  • Observations: In situ gauge records from over 30,000 Automatic Weather Stations (AWS) in China.
  • Forecasts/Scenarios: ECMWF HRES operational forecasts and CMIP6 (CAM-MPAS-HR) future scenario simulations.
• Training Strategy: A two-stage process involving pretraining on gridded products (ERA5/IMERG) followed by fine-tuning with sparse gauge observations to achieve "climatological jailbreak" (local refinement without losing global coherence).

Main Results

• Error Reduction: PRIMER significantly reduced Mean Absolute Error (MAE) and Continuous Ranked Probability Score (CRPS) at gauge sites compared to raw ERA5 and IMERG. For a specific Meiyu event, ensemble-mean $\Delta$MAE improved from 0.46 mm/h to 0.14 mm/h for ERA5-based posterior samples.
• Physical Realism: The model accurately reproduced the frequency distribution of heavy precipitation tails and corrected spatial anisotropy (orientation and focal length) of precipitation systems.
• Generalization: The framework demonstrated "zero-shot" capability by successfully bias-correcting unseen ECMWF HRES forecasts and downscaling CMIP6 future climate scenarios (2050) without retraining.
• Data Fusion: Integrating a subset of gauges (20%) during posterior sampling (GaugeFusion) further enhanced accuracy, mimicking operational optimal interpolation.

Contributions

• Coordinate-Based Representation: Unlike traditional grid-fixed models, PRIMER treats precipitation as a continuous spatial field, allowing the direct integration of irregular gauge locations without destructive interpolation.
• Unified Bayesian Framework: Establishes a "plug-and-play" prior that can be applied to diverse tasks (downscaling, bias correction, gap-filling) by simply modifying the likelihood function.
• Foundation Model Potential: Demonstrates that generative AI can distill multi-source, imperfect Earth system data into a robust prior that generalizes across different models and climate states.

Funding

• National Natural Science Foundation of China (Grant 42130603).
• National Key R&D Program of China (Grant 2024YFF0809004).
• US National Science Foundation (Grants IIS2324008 and RISE CAIG 2425748).
• Computational support from ColorfulClouds Tech.

Citation

@article{Sun2025Fusion,
  author = {Sun, Sencan and Nai, Congyi and Pan, Baoxiang and Li, Wentao and Li, Lu and Li, Xin and Foufoula-Georgiou, Efi and Lin, Yanluan},
  title = {Fusion of multi-source precipitation records via coordinate-based generative models},
  journal = {Nature Communications},
  year = {2025},
  doi = {10.1038/s41467-025-67987-9},
  url = {https://doi.org/10.1038/s41467-025-67987-9}
}

Generated by BiblioAssistant using gemini-3-flash-preview (Google API)