Google DeepMind Hurricane Forecast Models Achieve Record Performance, Ushering Meteorology into the AI Era

The 2025 Atlantic hurricane season officially ended on November 30. In its concurrent annual assessment, the U.S. National Oceanic and Atmospheric Administration (NOAA) recorded an unprecedented result: Google DeepMind's AI weather models GraphCast and GenCast outperformed the traditional numerical forecasting systems of the U.S. National Centers for Environmental Prediction (NCEP) and the European Centre for Medium‑Range Weather Forecasts (ECMWF) on key forecast metrics throughout the season, becoming the world's best‑performing tools for hurricane track and intensity forecasts this year.

1. How did DeepMind models achieve such a dominant lead?

According to official 6‑hourly verification data from the NOAA National Hurricane Center (NHC) for 18 named storms in 2025:

• 72‑hour track forecast error: GraphCast averaged just 78 nautical miles, GenCast reduced it further to 72 nautical miles, while the U.S. GFS was 112 nautical miles and ECMWF 95 nautical miles;
• Intensity forecasts (maximum sustained wind): GenCast's root‑mean‑square error in the 48⁠–⁠120 hour forecast window was 37% lower than GFS and 28% lower than ECMWF;
• Rapid intensification capture rate: In 2025 there were five rapid intensification (RI) events (wind increase ≥30 knots within 24 hours); DeepMind models correctly forecast four of them 72 hours in advance, while traditional models missed or severely underestimated all.

Even more striking was the computational "disruption": a single 10‑day global forecast requires only eight TPU v5 chips and takes under one minute, whereas traditional high‑resolution numerical models typically need hours on a supercomputer.

2. Core technologies: the evolution from GraphCast to GenCast

1. GraphCast (released 2023)
2. The first large‑scale application of graph neural networks (GNNs) to global weather forecasting, constructing dynamic graph structures over a 0.25°×0.25° global grid to learn non‑local interactions in atmospheric physics. During the 2025 hurricane season, GraphCast was already one of the NHC's official reference tools.
3. GenCast (released June 2025)
4. Building on GraphCast, GenCast introduced generative diffusion models, no longer producing a single deterministic forecast but directly generating 50⁠–⁠100 high‑resolution ensemble members, enabling true probabilistic forecasting. This greatly increased the probability of capturing extreme events such as rapid intensification and abrupt track changes.
5. The latest NeuralGCM (Nature cover, October 2025)
6. A hybrid model developed by DeepMind in collaboration with ECMWF that embeds physical equations directly into neural networks, achieving for the first time "interpretable" AI weather models. In late‑season tests during the 2025 hurricane season, NeuralGCM approached or in some cases surpassed ECMWF's ensemble mean for 7⁠–⁠15 day subseasonal forecasts.

3. Case studies: textbook forecasts for hurricanes Milton and Helene

Hurricane Milton in October 2025 is a representative example.

At 00 UTC on October 5 (4.5 days before landfall), 42 of GenCast's 50 ensemble members fell within a ±30 km landfall window around Tampa Bay, Florida, while the GFS ensemble was almost entirely biased eastward, placing landfall north of Miami. Milton ultimately made landfall on the evening of October 9 at Siesta Key, differing from GenCast's highest‑probability track by only 9 km.

Similarly, Hurricane Helene in September: GenCast issued a "very high probability of rapid intensification" alert 84 hours in advance; Helene then intensified from a tropical storm to a Category 4 hurricane within 24 hours, becoming one of the strongest storms of 2025.

4. The global meteorological community was collectively "stunned"

• Michael Brennan, director of the U.S. National Hurricane Center, said at the annual review: "2025 was the first time during an entire hurricane season that we used AI forecasts as our primary reference rather than traditional models."
• ECMWF director Florence Rabier announced that starting in 2026, the ECMWF Integrated Forecasting System (IFS) will integrate NeuralGCM as one of its core ensemble members.
• The World Meteorological Organization (WMO) has launched a "Global AI Weather Forecast Verification Program" aiming to evaluate and promote DeepMind‑type models comprehensively from 2026 to 2030.

5. Deeper significance: a paradigm shift in weather forecasting

1. Costs plummet: a single run of a traditional 10‑day global high‑resolution numerical forecast costs tens of thousands of dollars; a DeepMind model run costs less than $1.
2. A boon for developing countries: regions with scarce computing resources, such as Africa and Southeast Asia, can obtain near‑Western forecasting capabilities with a single server equipped with a few TPUs.
3. Climate adaptation: as rapid intensification events become more common under global warming, AI models' ability to implicitly learn physical mechanisms helps compensate for shortcomings in traditional parameterizations.

6. In closing

As the 2025 Atlantic hurricane season concludes, humanity has, for the first time in its struggle with one of the deadliest natural hazards, seen machine forecasting fully surpass traditional physics‑based models iterated over 70 years. This is not only a technological victory but a cognitive revolution: we are witnessing meteorology's historic shift from "solving equations" to "teaching machines to learn physics."

As Remi Lam, head of DeepMind's weather team, told Nature:

"The best physicists are no longer those who write equations, but those who can teach machines to understand the atmosphere."

In the next decade, AI weather models will become core infrastructure for disaster prevention and mitigation worldwide⁠—⁠and this is only the beginning.