Improved weather prediction using artificial intelligence promises to take a big step forward with the launch of a new European system, which can outperform conventional forecasting methods for up to 15 days ahead.
While tech companies and meteorological offices around the world are already applying AI to the weather, the European Centre for Medium-range Weather Forecasts (ECMWF) said its operational model broke new ground by making global predictions freely available to everyone at any time.
“This milestone will transform weather science and predictions,” said Florence Rabier, director-general of ECMWF, an intergovernmental organisation. “Making the AI Forecasting System operational produces the widest range of parameters using machine learning available to date.”
An experimental version tested over the past 18 months showed the system was about 20 per cent more accurate on key predictions than the best conventional methods, which feed millions of worldwide weather observations into supercomputers and crunch them with physics-based equations.
The new European system could predict the track of a tropical cyclone 12 hours further ahead, giving valuable extra warning time for severe events, said Florian Pappenberger, ECMWF director of forecasts.
The world experienced its hottest temperatures on record in 2024, and Europe has become the fastest warming continent, triggering extreme weather events. The agency has been at the forefront of observations and public awareness about the effects of climate change.
Other medium-range AI forecasting systems under development include GenCast and GraphCast from Google DeepMind, Pangu-Weather from Huawei, FourCastNet from Nvidia and FuXi from Shanghai Academy of AI for Science and Fudan University. All were trained on a database of weather observations compiled by the ECMWF over 40 years.
Comparing the accuracy of competing AI forecasting systems was hard, Pappenberger said, because their relative performance differed according to the variables and timescales assessed. Scores published by the ECMWF give some idea of performance but do not identify an overall champion.
But Pappenberger noted that its system stood out for predicting many more features than standard temperature, precipitation and wind. For example, it also forecasts solar radiation and wind speeds at 100 metres — the height of a typical turbine — helpful for the renewable energy sector.
Although ECMWF forecasts are freely available, the agency does not issue severe weather alerts nor tailor-made predictions to industry users, leaving the specialised forecasts to national or local authorities and private companies.
Climate Capital
Where climate change meets business, markets and politics. Explore the FT’s coverage here.
Are you curious about the FT’s environmental sustainability commitments? Find out more about our science-based targets here
ECMWF and a group of European national met offices have created an open-source technical framework for AI weather systems called Anemoi, after the Greek god of the winds. The underlying machine-learning architecture is based on the same “graph neural network” as Google DeepMind’s forecasting models.
Peter Battaglia, research director at DeepMind, said it was “impressive” to see how the ECMWF had adapted to the AI wave that had reshaped the field in recent years, and the latest open model would add to the pool of knowledge.
The ECMWF plans to improve its system further by increasing its spatial resolution and moving from the present version, which generates one forecast at a time, to “ensemble forecasting” — or creating a collection of 50 forecasts simultaneously with slightly different starting conditions to provide a range of possible outcomes.
In the future, said Kirstine Dale, chief AI officer at the UK Met Office, a mix of physics-based and data-based simulations would be needed for “their combined strengths to provide accurate, fast, reliable and trustworthy forecasts”.
Today the boundaries of reliable day-to-day weather forecasts in Europe are six to seven days ahead for precipitation and wind, and up to 14 or 15 days for temperature, said Pappenberger.
“Machine learning models have a fair chance of extending that because they may be able to extract something out of the data that we may not currently represent well enough in physics-based models.”