Stress-Testing Dynamical and Generative Downscaling for Subseasonal Extreme Precipitation

Subseasonal forecasting of extreme precipitation is inherently limited by the coarse spatial resolution of global circulation models. This study benchmarks an unpaired diffusion model directly against the classical Weather Research and Forecasting (WRF) dynamical model, downscaling ECMWF forecasts up to three weeks ahead against Swiss radar-gauge observations. The comparison reveals distinct regime-dependent advantages: WRF attains higher probabilistic skill during non-stationary multicell convection, whereas the diffusion model maintains greater metric consistency and leads in stationary supercell conditions.