Bridging Global AI Models and Local Extremes: A Dual-Stream Framework for Correcting and Downscaling GraphCast Rainfall Predictions

Data-driven global weather models, such as GraphCast, have revolutionized medium-range forecasting but often exhibit systematic limitations in quantitative precipitation forecasting (QPF). Specifically, these models tend to produce over-smoothed blurry rainfall fields and underestimate localized extremes , primarily due to the inherent uncertainties in their reanalysis training data (e.g., ERA5) and the use of mean-squared-error-based loss functions.

To bridge the gap between coarse-resolution global AI forecasts and the need for precise, high-impact weather prediction, we introduce SynQPF-Net, a deep learning framework designed to synergize GraphCast’s dynamical background fields with high-resolution observational analyses. The model employs a dual-stream spatiotemporal encoder to process heterogeneous inputs: the 0.25^o dynamical forecasts from GraphCast and the 0.0625^o precipitation analyses from the China Meteorological Administration Land Data Assimilation System (CLDAS) . A specialized hybrid loss function, combining classification (Dice) and regression (Weighted MSE) objectives, is utilized to jointly optimize the spatial structure and intensity of precipitation.

Evaluated on warm-season events in Southern China, our approach demonstrates significant skill improvements. SynQPF-Net effectively sharpens the forecast, doubling the Critical Success Index (CSI) for heavy rainfall (>=10 mm) at the 6-hour lead time compared to the raw GraphCast output. Crucially, interpretability analysis reveals that the model learns physically consistent meteorological principles: it predominantly relies on extrapolating recent observational patterns for short lead times (<=12 h) and dynamically shifts its focus to large-scale circulation and moisture variables (e.g., 700 hPa specific humidity) as the forecast horizon extends. This work provides a validated pathway for correcting and downscaling global AI weather models, offering a robust solution for short-range extreme precipitation forecasting.