Speaker
Description
RUSH (Rapid Update Short-term High-resolution) is an AI-native, rapid-update nowcasting framework that turns heterogeneous inputs into probabilistic 0–24 h forecasts at 30-min steps on a 1-km grid over Belgium. Methodologically, RUSH couples two ingredients: (i) observation-led evolution learned from 30-min radar accumulations and 15-min SEVIRI channels, and (ii) large-scale dynamical context from ECMWF’s AIFS. A latent-diffusion sequence-to-sequence model performs conditioning and super-resolution in latent space, avoiding manual blending and enabling seamless transition from nowcasting to short-range guidance. The system is designed for fast cycling as new observations arrive and exposes uncertainty via ensembles. We will present the framework design, training/ingestion pipeline and provide an update on current work which now includes longer forecast times and large-scale predictions downscaling.
