Google DeepMind has open-sourced an artificial-intelligence weather model that researchers say can forecast tropical cyclones with roughly an extra day of useful lead time compared with leading operational systems.
The model, called WeatherNext Cyclones, predicts not only where a hurricane or typhoon may travel but also how strong it could become and how far its damaging winds may extend. In testing on storms from 2023 through 2025, researchers reported state-of-the-art performance across track, intensity and wind-structure forecasting. The findings were published in Nature on August 6, 2026.
Google is now releasing the code and model weights for WeatherNext Cyclones, WeatherNext 2 and a smaller WeatherNext 2-mini model so meteorological agencies, universities and other researchers can build on the technology. Read Google DeepMind’s WeatherNext announcement
WeatherNext Can Provide About 24 Hours of Additional Forecast Skill
Hurricane forecasting has improved enormously over recent decades, but an additional day of accurate warning remains extremely valuable.
According to Google DeepMind, a three-day WeatherNext Cyclones forecast can achieve approximately the accuracy that previous leading systems reached at two days. Across evaluations of cyclone position, intensity and wind structure, the model produced more than a 24-hour average lead-time advantage. Google described that improvement as roughly equivalent to a decade of historical progress in tropical-cyclone forecasting.
That does not mean every hurricane will now be predicted perfectly one day earlier. Forecast skill varies considerably between storms. A system that performs exceptionally during one hurricane can make larger errors during another.
National Hurricane Center director Michael Brennan consequently described the model as another powerful tool rather than a replacement for the broader forecasting system. Human meteorologists still combine satellites, aircraft observations, ocean measurements and multiple computer models before issuing official warnings.
Hurricane Melissa Became a Major Real-World Test
WeatherNext gained attention during Hurricane Melissa in October 2025.
When Melissa was still a relatively weak system in the Caribbean, conventional forecast models disagreed over its future. Some scenarios kept the storm weak and directed it toward Haiti. WeatherNext instead indicated that rapid intensification was likely and predicted a Category 5 landfall in Jamaica five days ahead with approximately 80 percent confidence.
That probability increased toward 100 percent three days before landfall. Melissa eventually struck Jamaica as a Category 5 hurricane and became the strongest hurricane recorded to make landfall in the country.
The episode was particularly significant because the National Hurricane Center successfully forecast that Melissa would reach Category 5 strength while it was still only at Category 1 intensity—something the agency had not previously accomplished.
Google says WeatherNext contributed to that forecast alongside traditional physics-based systems, satellite data and hurricane reconnaissance. The additional warning time allowed Jamaican authorities to prepare emergency resources and evacuations earlier. See how WeatherNext supported Hurricane Melissa forecasting
Predicting Hurricane Intensity Has Always Been Especially Difficult
Forecasting a hurricane involves solving two related but technically different problems.
The first is predicting its track. Large-scale atmospheric circulation, including high-pressure systems, troughs and prevailing winds, largely determines where a storm travels. Global weather models have become increasingly successful at capturing these steering patterns.
Intensity is harder.
A hurricane’s strength depends on complicated processes occurring around its relatively small inner core. Ocean temperature, humidity, thunderstorms, wind shear and internal structural changes can determine whether a storm weakens or suddenly intensifies.
Traditional forecasting has therefore involved a compromise. Global models provide broad atmospheric context but may lack the resolution needed to reproduce the storm’s inner structure. High-resolution regional models capture more local detail but require far more computing power.
WeatherNext Cyclones appears to bridge some of that gap. It was trained jointly on global atmospheric information and expert-curated records of almost 5,000 historical tropical cyclones. The training included nearly 20 terabytes of atmospheric data.
The AI Works With Surprisingly Coarse Data
One of the study’s most unusual findings is that WeatherNext does not need the extremely fine spatial resolution previously considered essential for good hurricane-intensity forecasting.
The main cyclone model operates with atmospheric inputs representing areas approximately 28 by 28 kilometres. Google says this is around 100 times coarser than the data used by some traditional high-resolution systems.
Even WeatherNext 2-mini, which operates at approximately 111-kilometre resolution, produced surprisingly strong results.
Researchers do not yet completely understand why.
The result suggests that large-scale atmospheric conditions contain more information about future hurricane intensity than scientists previously recognized. WeatherNext may be identifying patterns in those broader conditions that conventional forecasting techniques have not exploited as effectively.
That uncertainty is scientifically interesting but also a reason for caution. AI weather models can identify highly predictive relationships without producing a simple physical explanation for each forecast.
Opening the system to outside researchers could therefore help meteorologists investigate not only whether the predictions work but also what atmospheric signals the model is using.
WeatherNext Can Generate 1,000 Possible Hurricane Futures
Another major advantage comes from speed.
Weather forecasts are inherently uncertain because small differences in atmospheric conditions can eventually produce substantially different outcomes. Meteorologists address this by running ensembles—multiple slightly different simulations showing a range of possible futures.
Traditional numerical weather models are computationally expensive, limiting how many scenarios can practically be produced.
WeatherNext can generate a single 15-day forecast in less than a minute on a Google TPU, according to DeepMind. During the 2025 hurricane season, the system generated 50 scenarios for each storm. It has now been scaled to as many as 1,000 ensemble members.
Those extra simulations become particularly useful when forecasters are searching for low-probability but catastrophic outcomes.
A conventional 50-member ensemble may contain few or no examples of an unlikely rapid-intensification scenario. With 1,000 simulations, unusual pathways have a greater opportunity to appear, potentially revealing risks that should receive additional investigation.
WeatherNext can generate forecasts up to 15 days into the future and produce probability maps showing where tropical-storm and hurricane-force winds could occur.
Open-Sourcing Could Matter Most Outside Wealthy Forecasting Agencies
Google’s decision to release the models could have implications beyond the United States.
Advanced physics-based forecasting systems require massive supercomputers, specialist teams and extensive operational infrastructure. Many countries exposed to cyclones do not possess resources comparable with the United States, Europe or Japan.
A computationally efficient AI model could make sophisticated ensemble forecasting more accessible to smaller meteorological agencies, universities and disaster-response organizations.
Google says it is already working with agencies in regions highly exposed to tropical cyclones, including the Philippines, Taiwan, Indonesia and Vietnam.
The released WeatherNext 2-mini is particularly notable because Google says it can run on a single TPU and is available through a public Colab notebook. Researchers can also inspect and modify the released code rather than relying exclusively on forecasts produced by Google.
Explore the public WeatherNext repository
Open Source Does Not Turn It Into an Official Warning System
WeatherNext forecasts should not be confused with government weather warnings.
Google explicitly labels Weather Lab as an experimental research service. Its predictions are not official evacuation orders, hurricane warnings or public-safety instructions. Anyone facing an active storm should rely on the appropriate national meteorological authority.
That distinction exists because a hurricane forecast involves far more than predicting a track and maximum wind speed.
Forecasters must determine storm surge, rainfall, flooding, tornado potential and local impacts. They also have to communicate uncertainty in a form emergency managers and residents can use.
A mathematically superior model can therefore improve a warning system without replacing the people responsible for deciding what the prediction means on the ground.
AI Weather Forecasting Is Moving From Research Into Operations
WeatherNext is part of a wider transformation in numerical weather prediction.
Earlier Google systems such as GraphCast and GenCast showed that machine-learning models could rival or outperform expensive physics-based forecasts on many global-weather benchmarks. GenCast, for example, generated probabilistic 15-day forecasts in minutes and outperformed the European Centre for Medium-Range Weather Forecasts ensemble across most of the variables and lead times evaluated in its published research.
Hurricanes represented a more difficult frontier because track and intensity require understanding atmospheric processes operating at very different scales.
WeatherNext Cyclones provides evidence that AI models can now contribute meaningfully to both.
The biggest achievement may therefore be less about replacing traditional meteorology and more about giving meteorologists another source of information earlier than previously possible.
An additional day does not sound revolutionary in ordinary life. Before a Category 4 or Category 5 hurricane, it can mean another day to evacuate vulnerable communities, position emergency crews, protect hospitals, close ports and move aircraft or ships away from danger.
Google’s decision to open-source WeatherNext now gives researchers around the world an opportunity to test those claims independently, discover where the system fails and potentially make it better before the next destructive storm arrives.
