RUSH: A Novel Fully AI-driven Framework for Seamless Integration of Observations and Global AI Forecasts in Short-term Weather Prediction

Short-term weather forecasting, especially for extreme events, remains challenging due to the need to effectively combine recent observations with numerical weather predictions. To tackle this challenge, we present RUSH (Rapid Update Short-term High-resolution forecast), an innovative framework designed to provide high-resolution (1 km) precipitation forecasts on a national scale with lead times up to 24 hours. RUSH follows the recent attempts to create fully AI-driven kilometer-scale forecasting systems that completely replace traditional numerical modeling with a combination of machine learning and observational data. Our system employs a Latent Diffusion Model architecture to seamlessly blend information from multiple data sources, including radar composites, satellite observations (SEVIRI bands), and ECMWF's AI-based global forecasting system (AIFS). 

The model is conceptually designed to transition from observation-driven predictions in the first few hours to a sophisticated spatial and temporal downscaling of AIFS forecasts at longer lead times. This approach aims to leverage the strengths of both data sources: the high spatial and temporal resolution of observational data for immediate forecasts, and the physically consistent evolution provided by AIFS for longer horizons. By utilizing an end-to-end AI architecture from global to local scale, RUSH not only addresses the computational constraints typically associated with traditional numerical weather predictions but also explores the potential for a new generation of fully data-driven weather forecasting systems. 

Our framework processes multi-source input data at different spatial and temporal scales, including radar-derived 30-minute precipitation accumulations, key SEVIRI channels, and selected AIFS forecast fields at 25km resolution. The model's sequence-to-sequence architecture allows for flexible spatial domain handling and probabilistic precipitation forecasting through multiple realizations. 

We will present preliminary results from two experimental implementations over different European domains (Italy and Belgium), demonstrating the model's capability to generate rapid-update forecasts and discussing its potential for operational implementation in weather services. The evaluation will focus on precipitation prediction skills across different intensity thresholds and temporal scales, with particular attention to extreme event forecasting. A preliminary comparison with operational limited area models (COSMO-2I and ALARO-AROME) over selected case studies will assess the competitiveness of this fully AI-driven approach against high-resolution numerical weather prediction systems.