The United States National Hurricane Center (NHC) began using Google DeepMind’s tool last year, and is currently collaborating with the tech giant to develop a new model.
These efforts have taken on particular significance this year, driven by an exceptionally strong El Niño that has triggered a record number of hurricanes in the Pacific, and by Hurricane Isaias which struck the southeastern United States last month.
But according to Ferran Alet, a DeepMind researcher specializing in cyclones and lead on the study, Google has not produced a tool specifically tailored to hurricanes, thus forcing scientists to rely on models used for routine weather forecasting.
“The biggest challenge for us”, he told AFP, “is extreme weather phenomena”. “Once that is addressed (…) everything else will follow”.
1,000 Possible Scenarios
Before the advent of AI, the NHC used several types of models to forecast tropical cyclones. The models referred to as “dynamic” rely on supercomputers to solve the equations of fluid dynamics and thermodynamics that govern the atmosphere.
As for the statistical models, they lean on historical relationships between storm behavior and variables such as location and date, whereas the so-called “hybrid” models combine physics and statistics.
AI-generated models, for their part, learn from vast data sets, much like statistical models, but at a far larger scale and with the capacity to detect far more complex patterns.
According to a paper published last August by DeepMind in the scientific journal “Nature,” AI-generated models offer, on average, forecasts of an extra day. This means that today’s three-day forecasts are as reliable as two-day forecasts provided by older models.
DeepMind, for example, calculated that Isaias could have become a tropical cyclone a week before it formed.
The model can also generate a thousand possible scenarios for each cyclone and forecast precipitation as well as the track and intensity of the storm.
“NHC began integrating AI models into its work by the end of June 2025. Google’s DeepMind models were added to some NHC tools in 2026 to assist calculations on the track and intensity of storms,” explains Dan Brown, NHC operations chief to AFP.
The NHC also uses AI models developed by the European Centre for Medium-Range Weather Forecasts and the National Weather Service, and has reached an agreement with Google to collaborate on developing new models.
“Excellent Results”
AI-generated models delivered “excellent results during the Isaias event,” notably “in terms of the track,” says Dan Brown.
According to him, DeepMind as well as the NHC’s own models are the most effective at calculating hurricane trajectories.
However, this technology has its limits. According to Dan Brown, models can “vary their trajectory or intensity forecasts significantly from one cycle to the next”, thereby forcing specialists to rely on their experience and judgment to determine which models best reflect a given storm, and to combine them.
The role of researchers remains “more important than ever,” the meteorologist asserts. “An algorithm can indicate the probability of a track, but it does not account for real-time data from aircraft and satellites to determine the structure of the storm.”
