Precipitation generation skips retraining the weather model: one diffusion generator on frozen backbones cuts day-7 Brier score from 0.245 to 0.218

Precipitation generation is split off from the weather foundation model: the backbone supplies only non-precipitation atmospheric states such as temperature, wind and humidity, and an independently trained generator turns those states into observation-aligned 24-hour accumulated precipitation. Weather foundation models are trained on ERA5 reanalysis, whose precipitation carries regional and intensity-dependent biases, so localized extremes are often smoothed or underrepresented; replacing the precipitation target and retraining the whole model is impractical, since training GraphCast took about four weeks on 32 Cloud TPU v4 devices. Across the Pangu-Weather, GraphCast and HRES backbones, one generator is reused without backbone-specific retraining; GraphCast gains 0.103 in ACC and cuts RMSE by 0.623 on average, and also outperforms GenCast.

The generator denoises a precipitation field from Gaussian noise, conditioned on atmospheric evolution compressed by a multi-scale recurrent encoder. Feeding raw high-dimensional atmospheric fields directly into the denoiser is memory-intensive, and sequential states carry substantial spatiotemporal redundancy, so a convolutional encoder first compresses multi-level forcing into a latent representation. Each encoder stage embeds a ConvGRU gate to model temporal change in atmospheric states, and in ablations ConvGRU consistently outperforms ConvLSTM under both conditioning backbones.

At each U-Net resolution the denoiser uses cross-attention to align precipitation features with atmospheric conditions at the matching scale, because rainfall at a location is not explained by same-point, same-time conditions alone. Daily precipitation is the accumulated outcome of atmospheric processes acting across space and time, while predictors such as moisture transport and vertical motion may occur earlier and over surrounding or upstream regions, so nonlocal spatial alignment is needed. In ablations cross-attention beats gated fusion, and the full configuration scores best on the reported metrics, including ACC 0.7015 and RMSE 5.590, indicating complementary gains from temporal extraction and spatial alignment.

The framework targets weather and hydrology readers who need daily accumulated precipitation forecasts, is trained on paired ERA5 and IMERG data, and is evaluated globally, regionally, and on a heavy-rainfall case using observations from 2,288 Chinese meteorological stations. Cross-backbone reuse is validated only for the aligned forecast systems considered here, and different resolutions require separately trained models. The generator samples precipitation conditional on a fixed atmospheric trajectory, and its skill remains dependent on the backbone forecast.

Cross-backbone reuse is empirically validated only for the aligned forecast systems considered here, and transfer to other reanalyses such as JRA and MERRA-2 and to operational products such as GFS still needs evaluation. Training on ERA5 states introduces a distribution shift when conditioning on imperfect forecasts, which the authors probe through noise perturbation and forecast-field fine-tuning experiments. IMERG retrieval uncertainty and the additional cost of diffusion sampling remain further limitations; generated precipitation should be used as decision support and requires regional validation, expert oversight, and clear communication of forecast uncertainty.