The AIFS: ECMWF’s data-driven weather forecasting system

Machine learning-based models are rapidly transforming medium-range weather forecasting. The European Centre for Medium-Range Weather Forecasts (ECMWF) has developed the Artificial Intelligence Forecasting System (AIFS), a state-of-the-art data-driven model combining a graph neural network encoder-decoder with a sliding window transformer processor. Trained on ECMWF's ERA5 re-analysis and operational numerical weather prediction analyses, AIFS demonstrates exceptional deterministic forecast skill across upper-air variables, surface weather parameters, and tropical cyclone tracks.

Building on this foundation, ECMWF has introduced AIFS-CRPS, a probabilistic extension of AIFS designed for ensemble forecasting. AIFS-CRPS is obtained by training a stochastic model with the Continuous Ranked Probability Score (CRPS) as its loss function. It addresses uncertainties and generates highly skilful probabilistic forecasts. For medium-range timescales, AIFS-CRPS matches or outperforms ECMWF’s physics-based Integrated Forecasting System ensemble across key variables and lead times.

This presentation will highlight recent advancements in deterministic and probabilistic forecasting with AIFS, showcasing its operational readiness and its potential to redefine medium-range forecasting at ECMWF.