AI beats meteorologists: DeepMind was ahead in the hurricane season
The 2025 hurricane season was a litmus test for artificial intelligence in weather forecasting – and Google DeepMind passed.
Hurricane Melissa caused severe damage in Jamaica
(Image: Institute for Research in the Atmosphere at Colorado State University, NOAA)
Can Artificial Intelligence (AI) overshadow existing weather models? This year's North Atlantic hurricane season, which is currently ending in November 2025, should show. And the AI model Google DeepMind has delivered: According to a report by Ars Technica it provided the most accurate predictions of all tested systems. The US tech blog refers to evaluations by researcher Brian McNoldy from the University of Miami. For the first time, AI systematically surpassed the forecasts of human experts.
The US weather model GFS (Global Forecast System) performed particularly poorly this year. For the 5-day forecast, the deviation in storm path was around 360 nautical miles, while DeepMind was significantly closer to reality with 165 nautical miles.
Human forecasts surpassed
However, Google's AI even surpassed the forecasts of human experts at the National Hurricane Center. The AI also excelled in predicting storm intensity, even though events occurred, such as Hurricane Melissa, which caused severe devastation in Jamaica, that had not occurred in decades.
Videos by heise
In addition to the good performance of the forecasts, there are further advantages of an AI model. It does not require an expensive supercomputer and is therefore considered particularly efficient. At the same time, it is significantly faster in calculations than traditional models. And neural networks can also learn from mistakes and correct themselves during operation. The results are all the more astonishing, as the technology is still very new and further progress is expected.
Poor performance of GFS puzzling
Why the US model GFS performed so poorly is still unclear. It is conceivable that staff and equipment cuts by the new US administration in the course of the DOGE campaign have created data gaps. However, a failed modernization of the model core is also a possible explanation.
Experts expect meteorologists to rely heavily on AI models in the future. Traditional physics-based models would thus lose importance. Overall, there were 13 named storms in this Atlantic season.
(mki)
