Google’s AI outperforms NOAA hurricane model in storm-strength study
An independent evaluation found stronger intensity predictions one to five days ahead, but the model’s uncertainty ranges were often too narrow. A newer version was not included.
Loading page…
An independent evaluation found stronger intensity predictions one to five days ahead, but the model’s uncertainty ranges were often too narrow. A newer version was not included.
Listen to this story
The University at Albany’s peer-reviewed test gives forecasters an outside benchmark for Google’s WeatherNext Cyclone, but it does not establish that the system is ready to replace operational judgment. Its strongest result was against NOAA’s model; performance was closer to human forecasts after day one, while intensity bias and overly narrow outcome ranges complicate how its predictions should be used. The study also predates Google’s 2026 WeatherNext version, so its findings are not a verdict on that release.
The evaluation covered 385 tropical cyclones worldwide from June 2022 through May 2026.
WeatherNext Cyclone beat NOAA’s model at every forecast lead time from one to five days.
The AI underestimated storm intensity, with the gap growing at longer lead times; actual intensity often fell outside its predicted range.
Google’s hurricane AI delivered more accurate storm-strength forecasts than NOAA’s state-of-the-art hurricane model in an independent University at Albany study. But the evaluation, published in Geophysical Research Letters, also found a persistent weakness: the AI underestimated storm intensity. The findings give forecasters evidence of both the model’s value and the errors they must account for.
The university described the findings on November 19, following Google’s August 6 announcement of its own cyclone-forecasting research in Nature. That earlier work covered storm tracks, intensity and wind structure. The Albany researchers focused their independent examination on intensity: how well WeatherNext Cyclone predicted the winds storms would actually produce.
The project began in Tang’s graduate course the previous year. Jacob Vile and Nour Elgalad expanded the class project into a full analysis and took it through peer review, working with Tang.
Their evaluation covered 385 tropical cyclones worldwide, spanning June 2022 through May 2026. They checked predicted wind speeds against recorded winds, then compared performance with official forecasts from the National Hurricane Center and Joint Typhoon Warning Center, alongside other commonly used forecasting tools.
WeatherNext Cyclone beat NOAA’s hurricane model at every forecast lead time from one to five days. Against human forecasters, the result was closer: the AI performed on par with them beyond the first day.
The researchers also compared it with two tools designed to flag rapid intensification, defined as a wind-speed increase of at least 35 mph in 24 hours. For that next-day prediction, the AI was about as reliable as the leading comparison tool and produced fewer false alarms.
The model tended to underestimate storm strength, and that gap grew at longer forecast lead times. The study also found weaknesses in where it worked best and how broadly it described possible outcomes:
Tang offered a possible explanation, not a demonstrated cause. The model was trained on broad weather data and basic storm information, he said, so it may lack finer details needed for the strongest, fastest-strengthening storms.
The evaluation does not cover a newer version Google released in 2026. With other AI weather models also in development, Tang wants a shared, transparent way to compare them.
For forecasters, Tang’s emphasis is on knowing when to rely on the model and when to use their own judgment. He sees AI as an additional tool, not a replacement for forecasters.
Loading discussion...
Join the conversation
Explain whether errors in opposite directions should carry equal weight.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.