Google's AI hurricane model outperforms NOAA in independent test

Дмитрий Лобанов редактор региональной ленты новостей

A week before Hurricane Melissa slammed into Jamaica last fall, forecast models couldn't agree on its path or strength. Google DeepMind's AI model got it right, predicting both the storm's track and its Category 5 intensity. Об этом пишет издание Phys.org.

A new study by University at Albany researchers offers one of the first independent evaluations of Google's AI hurricane model, focusing on intensification. Published in Geophysical Research Letters, the study assessed the WeatherNext Cyclone model's intensity forecasts for every tropical cyclone from June 2022 to May 2026.

Brian Tang, a UAlbany professor who co-led the study, said independent checks matter because they give the public confidence when developers claim improvements, and they offer an objective look at strengths and weaknesses. The team analyzed 385 tropical cyclones worldwide, comparing predicted wind speeds with actual records and against official forecasts from the National Hurricane Center and Joint Typhoon Warning Center.

The AI model was more accurate than NOAA's state-of-the-art hurricane model at every point from one to five days out and performed on par with human forecasters beyond the first day. It was also about as reliable as the leading rapid-intensification tool over 24 hours, with fewer false alarms. Tang noted that rapid intensification is still hard to predict, but AI models may better connect the dots, and Google's model is less likely to predict intensification that never happens.

The study also found weak spots: the model tends to underestimate storm strength, especially further out, and its performance varies by region—best in the North Atlantic, worst in the eastern North Pacific. Its range of possible outcomes was often too narrow, with real intensity frequently outside the model range.

The research began as a class project in Tang's graduate course, expanded by recent graduates Jacob Vile and Nour Elgalad. Vile said he proposed the project to understand the model's biases, and Tang saw its broader value. Elgalad called the publication an incredible opportunity. Google released a newer version of the model this year, not included in the study, and Tang believes forecasters need a shared, transparent way to compare AI tools. "AI is a powerful new tool, but it doesn't replace forecasters," he said.