On February 25, 2025, the European Centre for Medium-Range Weather Forecasts (ECMWF) made its Artificial Intelligence Forecasting System (AIFS) operational - the moment AI weather forecasting moved from research papers into the daily production pipeline of one of the world’s leading forecasting centers. ECMWF described it as the first fully operational, open machine-learning weather model spanning a wide range of parameters.
AIFS combines a graph neural network encoder and decoder with a sliding-window transformer processor, and is trained on ECMWF’s ERA5 reanalysis and operational analyses. It ingests the same initial conditions as the traditional physics-based Integrated Forecasting System (IFS) - around 60 million quality-controlled observations, mostly from satellites - so the two can be compared head to head. ECMWF reported that AIFS outperforms its physics-based forecasts on many measures, including up to a 20 percent improvement in tropical cyclone track forecasts, while using on the order of 1,000 times less energy to produce a forecast.
An ensemble version, AIFS ENS, generating 51 perturbed scenarios, followed into operations on July 1, 2025. The move signaled that machine-learning models had matured from impressive benchmarks to trusted operational forecasting infrastructure.