WeatherGenerator project aims to recast ML for Earth system modelling (www.ecmwf.int)

The European Centre for Medium-Range Weather Forecasts (ECMWF) is spearheading a groundbreaking initiative called WeatherGenerator, funded by the EU Horizon program, which aims to innovate machine learning applications for weather forecasting and Earth's system modeling. Set to launch in February 2025, this project unites 16 European organizations, including national meteorological services and research institutes specializing in high-performance computing and machine learning. WeatherGenerator will serve as a versatile foundation model, capable of supporting 22 diverse applications ranging from weather and climate prediction to renewable energy and flood forecasting. What sets WeatherGenerator apart is its ability to integrate multiple datasets, including reanalysis, observational inputs from satellites, and climate forecasts, to model complex Earth system processes. Utilizing a technique known as masked token learning, the tool will learn to fill in missing data and correlate various input streams, enhancing its resilience and predictive capabilities. By covering a broad spatial and temporal range, from local to global scales, WeatherGenerator promises to transform how machine learning is applied in environmental science while supporting the EU's Destination Earth initiative, potentially leading to improved climate resilience and resource management on a global scale.