NHC verification finds Google DeepMind’s GDMI leading individual hurricane guidance in 2025

On 12 June 2025, Google DeepMind and Google Research launched Weather Lab and began sharing live predictions from an experimental tropical-cyclone model with NHC forecasters. The system generated 50 possible cyclone scenarios extending as far as 15 days and attempted to predict formation, track, intensity, size and structure within one probabilistic model. DeepMind’s launch evidence was retrospective and developer-controlled, so the 2025 season was the first important prospective operational test.

NHC later described 2025 as the first season in which it incorporated AI-based models into real-time operations. Its preliminary report called GDMI “very useful,” while warning that several AI systems remained under development and were not always delivered reliably enough for routine use. The completed seasonal verification subsequently showed that GDMI was not merely an interesting experimental output: it had become competitive with, and under defined conditions better than, established operational guidance.

In homogeneous Atlantic comparisons, meaning forecasts were compared only where every included model was available for the same cases, GDMI slightly outperformed the NHC official track forecast from 12 through 72 hours. Its Atlantic intensity forecasts had skill comparable to the official NHC forecast and placed it among the strongest individual intensity models at several lead times.

Its strongest relative result came in the eastern North Pacific. From 48 through 120 hours, GDMI performed considerably better than the other individual track models and beat both the official NHC forecast and every tested consensus aid. That is especially notable because consensus aids combine several strong models and are normally more difficult for one individual system to outperform consistently.

DeepMind later summarized the merged Atlantic and North Pacific results by saying WeatherNext was the season’s top-performing individual model for both track and intensity. That summary is useful, but it remains a Developer / Vendor Claim and compresses substantial variation between basins, forecast ranges and metrics. The more granular NHC findings above should therefore define the public VyDex claim.

Hurricane Melissa provides the clearest operational example. On 24 October, while Melissa was still a relatively weak tropical storm, every DeepMind ensemble member projected Category 4 strength or higher, and NHC said its track and intensity forecasts were being blended with GDMI guidance. By the following day, roughly four-fifths of the 50 members projected Category 5 strength, and an NHC discussion described GDMI as the best-performing intensity guidance to that point in the season. On 26 October, 48 of the 50 members still reached Category 5.

Melissa ultimately struck Jamaica on 28 October with estimated sustained winds of 185 mph and a central pressure of 892 millibars. NHC had issued almost three days of advance notice of a Category 5 Jamaican landfall—the first time it had forecast Category 5 intensity while a storm was still only Category 1.

This result does not mean GDMI beat NHC or the best consensus systems everywhere. Performance depended on basin, forecast horizon and metric. The strongest verified superiority was in eastern North Pacific track forecasting at 48–120 hours. Atlantic track advantages were smaller and concentrated at shorter ranges, while Atlantic intensity performance was described as comparable to the official forecast rather than universally superior.

NHC’s official forecast remained the stronger overall decision product in several of the season’s hardest situations. Across Atlantic rapid-intensification cases, NHC achieved a higher probability of detection and critical success index than every real-time model. For Melissa specifically, its official intensity forecast outperformed every model at nearly every lead time. GDMI contributed valuable evidence to the human forecast; it did not independently produce the official warning or replace the observations, aircraft reconnaissance, physical models, consensus systems and forecaster judgment behind it.

GDMI was also unavailable during the early portion of the season. The homogeneous comparisons excluded the first two Atlantic storms and the first five eastern North Pacific storms, so this was not a complete all-storm evaluation. NHC separately noted that some experimental AI systems were not consistently available in time for routine operational use.

One season is important prospective evidence, but it cannot establish performance across every possible hurricane environment. The Atlantic had fewer forecast cases than normal, while an unusual concentration of Category 5 storms and rapid-intensification episodes made intensity prediction particularly difficult. Future seasons could expose weaknesses that the 2025 sample did not.

The system’s ability to generate scenarios extending 15 days should also not be confused with verified, high-accuracy hurricane intensity guidance across that entire horizon. The central operational comparisons in the NHC report extend through five days. The 15-day horizon describes what the model can generate, not a demonstrated 15-day intensity-forecasting threshold.

The supported conclusion is that AI hurricane forecasting crossed a meaningful operational threshold in 2025. Earlier global AI weather models could provide useful track guidance but generally smoothed away the strongest winds and failed badly at intensity. GDMI became a system that forecasters could consult in real time for both dimensions and that subsequently achieved leading individual-model performance under formal NHC verification.

The result should not be interpreted as hurricane forecasting being solved or human forecasters becoming unnecessary. It shows something narrower and more consequential: an AI model overcame a central documented limitation of the previous model generation well enough to become useful inside one of the world’s highest-stakes operational forecasting workflows.