Google DeepMind’s WeatherNext Cyclone model predicted Hurricane Melissa’s track and Category 5 strength a week before landfall in Jamaica last fall, while conventional models disagreed on both path and intensity.
A University at Albany study published in Geophysical Research Letters now provides one of the first independent evaluations of the AI system, testing it against every tropical cyclone worldwide from June 2022 through May 2026.
“We decided to examine Google’s AI hurricane model as it has quickly become one of the most talked-about tools in weather forecasting,” said Brian Tang, professor of atmospheric and environmental sciences at the University at Albany, who co-led the study. “Independent checks matter because they give the public more confidence when a model’s developers claim it has improved. They also offer an objective look at both strengths and weaknesses.”
Tang, with recent graduates Jacob Vile and Nour Elgalad, analyzed 385 tropical cyclones. They compared the AI model’s predicted wind speeds against actual observations and benchmarked performance against official forecasts from the National Hurricane Center and Joint Typhoon Warning Center, other operational tools, and two specialized rapid-intensification products. Rapid intensification is defined as a wind-speed jump of at least 35 mph (56 kilometers per hour) in 24 hours.

The AI model was more accurate than NOAA’s state-of-the-art hurricane model at every lead time from one to five days out and performed on par with human forecasters beyond the first day. It matched the leading rapid-intensification tool for 24-hour reliability while issuing fewer false alarms.
“Rapid intensification is still really hard to predict, especially when and how extreme it will be, because it depends on a lot of factors lining up that we don’t fully understand or observe well,” Tang said. “AI models may be better at connecting those dots. In particular, we found Google’s model is less likely to predict rapid intensification that never happens.”
The study identified weaknesses: the model systematically underestimates peak intensity, with the gap widening at longer lead times; skill varies by basin, highest in the North Atlantic and lowest in the eastern North Pacific; and its forecast spread is often too narrow, so observed intensities frequently fall outside the predicted range.
“The model was trained on fairly big-picture weather data and basic storm information, so it may be missing some of the finer details needed to predict the strongest, fastest-strengthening storms,” Tang said.
The project began in Tang’s graduate course. “When I began this project almost a year ago, I had no idea it would become a published paper,” said Vile, who graduated in the spring. “This model kept coming up in our forecasting discussions. I proposed the project just to better understand some of its biases, but Professor Tang saw how useful the results could be to the broader scientific community.”
Google DeepMind, founded in the UK in November 2010 and acquired by Google in 2014, merged with the Google Brain division in April 2023 to form the current AI research laboratory. The organization first drew wide attention in 2016 when its AlphaGo program defeated world champion Lee Sedol, and again in 2020 when AlphaFold achieved breakthrough accuracy in protein structure prediction, eventually releasing over 200 million predicted structures covering virtually all known proteins.
Elgalad added: "The project tackled an important, ongoing topic in tropical cyclone forecasting, so it exposed me to the research happening in the field right now." Having our work published is an incredible opportunity.”