Google’s WeatherNext AI Improves Cyclone Forecasting
Google’s WeatherNext AI Improves Cyclone Forecasting
Google DeepMind and Google Research have reported a major advance in tropical-cyclone forecasting with WeatherNext Cyclones, an AI weather model designed to predict not just where a cyclone will travel, but also how strong it will become and how large its damaging wind field may grow.
The results were published in Nature on August 6, 2026. In tests covering tropical cyclones from 2023–2025, WeatherNext Cyclones provided, on average, a day or more of additional useful forecasting lead time compared with leading operational models. Researchers describe that improvement as roughly comparable to about a decade of historical progress in cyclone forecasting.
What is WeatherNext?
WeatherNext is a family of AI-based weather forecasting systems developed by Google DeepMind and Google Research.
Traditional numerical weather prediction models simulate atmospheric physics by solving enormous systems of mathematical equations. These models are extremely powerful, but they can also require substantial computing resources and time.
WeatherNext takes a different approach. It learns patterns from enormous amounts of historical atmospheric data and then uses those learned relationships to predict how weather conditions are likely to evolve.
The latest WeatherNext family can forecast variables including:
- atmospheric pressure
- temperature
- precipitation
- wind speed
- wind direction
- humidity
- tropical cyclone tracks
- cyclone intensity
- cyclone size and wind structure
Google says WeatherNext 2 can generate hundreds of possible weather scenarios very rapidly and is significantly faster than the previous generation.
Why cyclone forecasting is particularly difficult
Cyclones—called hurricanes in the Atlantic and eastern Pacific and typhoons in parts of the western Pacific—are unusually challenging because forecasters need to answer several different questions simultaneously.
1. Where will the cyclone go?
This is known as the track forecast.
A hurricane can move hundreds or thousands of kilometers, and relatively small errors in atmospheric steering currents can shift its eventual landfall significantly.
A forecast might therefore correctly predict that a cyclone will become extremely powerful but still incorrectly predict which city or coastline it will hit.
2. How powerful will it become?
This is the intensity forecast.
Forecasters need to predict maximum sustained winds and changes in the storm's central pressure.
This is especially difficult when rapid intensification occurs—for example, when a cyclone suddenly gains enormous strength over warm ocean water in only a day or two.
3. How large will it become?
Wind speed at the cyclone's center is only part of the danger.
Forecasters also need to know how far outward tropical-storm-force and hurricane-force winds will extend.
WeatherNext Cyclones attempts to forecast track, intensity and wind structure within the same AI system.
The major WeatherNext breakthrough
One of the most important findings is the improvement in forecast lead time.
Google describes the result this way conceptually:
A WeatherNext forecast made three days before an event can have roughly the accuracy that previous leading models achieved only about two days beforehand.
In other words, forecasters may receive approximately 24 additional hours of useful information.
That extra day can be extremely important during a major hurricane.
It could mean another day for governments to:
- organize evacuations,
- close vulnerable roads,
- move emergency equipment,
- prepare hospitals,
- protect ports,
- reposition rescue teams,
- secure electrical infrastructure,
- evacuate offshore workers,
- warn airports and airlines,
- and inform residents about an approaching storm.
The peer-reviewed study found an average improvement of a day or more in lead time for cyclone track, intensity and wind-radius predictions.
How WeatherNext Cyclones works
A particularly interesting part of the system is that it combines information about the entire atmosphere with specialized knowledge about tropical cyclones.
Google trained it using two major types of information.
Global atmospheric data
The model learned from almost 20 terabytes of atmospheric information, according to Google.
This allows it to recognize relationships involving large-scale weather systems such as:
- high-pressure areas,
- low-pressure systems,
- jet streams,
- atmospheric troughs,
- wind patterns,
- temperature,
- humidity,
- and other circulation patterns.
These large systems often determine where a cyclone travels.
Historical cyclone observations
WeatherNext was also trained using the International Best Track Archive for Climate Stewardship, or IBTrACS.
This global dataset contains information on almost 5,000 historical tropical cyclones.
Combining ordinary atmospheric information with specialized cyclone records allows the system to learn not only general weather behavior but also the distinctive behavior of extreme storms.
It doesn't make only one prediction
Another important feature is its use of ensemble forecasting.
Instead of saying:
“The hurricane will definitely follow this exact route.”
the model can generate many possible futures.
Imagine that WeatherNext generates 1,000 simulations.
Perhaps:
- 650 scenarios take the cyclone toward one coastline,
- 220 scenarios move it farther east,
- 100 show it remaining offshore,
- and 30 show an unusual but potentially catastrophic path.
Forecasters can then examine the entire probability distribution rather than relying on a single prediction.
This is particularly important because weather is inherently uncertain.
Up to 1,000 possible cyclone scenarios
Google says WeatherNext Cyclones can now produce ensembles containing as many as 1,000 members.
Traditional operational ensembles frequently contain roughly 50 members.
Moving from 50 toward 1,000 scenarios can help reveal low-probability outcomes that might otherwise be missed.
This becomes especially valuable for extreme events such as rapid intensification.
A 50-member ensemble might contain only one or two examples of an unlikely scenario. A much larger ensemble gives forecasters a clearer picture of the probability distribution around such outcomes.
The Nature paper specifically notes that 1,000-member ensembles can better represent rare events than conventional 50-member ensembles.
Forecasts extending up to 15 days
WeatherNext Cyclones generates scenarios extending as far as 15 days into the future.
That does not mean a precise 15-day hurricane landfall forecast should be treated as certain.
Accuracy naturally decreases as forecasting range increases.
Instead, the extended range helps meteorologists identify developing risks earlier—for example, that atmospheric conditions are increasingly favorable for a tropical cyclone to approach a particular region.
Google's Weather Lab can display experimental predictions of cyclone:
path → intensity → structure → size → development
up to 15 days ahead.
A surprising finding: WeatherNext doesn't require ultra-high resolution
This may be one of the scientifically most interesting findings.
Conventional thinking has generally held that accurate hurricane-intensity prediction requires extremely high-resolution models.
That's reasonable because a hurricane's core contains relatively small-scale processes involving:
- thunderstorms,
- convection,
- eyewall formation,
- heat transfer,
- moisture,
- ocean-atmosphere interactions,
- and powerful localized winds.
Yet Google says WeatherNext Cyclones operates using atmospheric information at roughly 28 × 28 kilometer resolution.
That is dramatically coarser than specialized regional hurricane models.
Nevertheless, the AI achieved highly competitive intensity forecasts.
Researchers say this suggests coarse global atmospheric data may contain more information about future hurricane intensity than scientists previously realized.
Exactly how the neural network extracts that information remains an active research question.
Hurricane Melissa provided a major real-world test
One of the most striking examples came during Hurricane Melissa in October 2025.
Melissa ultimately became a devastating Category 5 hurricane affecting Jamaica.
When the storm was still much weaker, its future strength was highly uncertain.
Google reports that WeatherNext's ensemble predicted a Category 5-strength Jamaican landfall approximately five days ahead with about 80% probability, rising close to 100% approximately three days before landfall.
This wasn't simply a retrospective experiment.
The U.S. National Hurricane Center was receiving Google DeepMind guidance operationally.
NHC discussions from the storm explicitly referenced the Google DeepMind ensemble. For example, on October 24, 2025, the NHC noted that the majority of the Google DeepMind ensemble members were reaching Category 5 intensity.
Another NHC discussion reported that 41 of 50 Google DeepMind ensemble members showed Melissa reaching Category 4 or Category 5 strength.
The National Hurricane Center's Melissa forecast was historic
The official NHC forecast subsequently became notable in its own right.
According to NOAA's 2025 verification report, the NHC provided almost three days of advance notice that Melissa would make landfall in Jamaica as a Category 5 hurricane.
It was the first time the NHC had forecast a cyclone to eventually reach Category 5 intensity while it was only Category 1 at the time the forecast was issued.
The NHC's 2025 season verification report also found that the Google DeepMind ensemble mean showed slightly better short-range track performance than official forecasts during the season, although official NHC forecasts generally performed better than most individual dynamical and consensus models overall.
That's an important distinction: WeatherNext isn't intended simply to replace meteorologists. It provides another powerful source of guidance that human forecasters can combine with satellites, aircraft observations, conventional numerical models and professional judgment.
Traditional forecasting vs. WeatherNext
| Feature | Traditional numerical models | WeatherNext Cyclones |
|---|---|---|
| Main approach | Physics-based equations | Machine learning |
| Cyclone track | Strong capability | State-of-the-art results |
| Cyclone intensity | Historically difficult | Major improvement |
| Wind structure | Modeled | AI predicts it directly |
| Forecast horizon | Multiple days | Up to 15 days |
| Ensembles | Often dozens of members | Up to 1,000 members |
| Computation | Potentially very expensive | Extremely fast inference |
| Resolution requirement | Often high for cyclone intensity | Strong results around 28 km |
| Rare-event analysis | Limited by ensemble size | Large ensemble improves probabilities |
The two methods should not necessarily be seen as competitors.
A future forecasting system could combine physics-based models + AI models + observations + expert meteorologists.
Why speed matters
Weather forecasting operates under severe time pressure.
A traditional high-resolution numerical model must repeatedly calculate atmospheric physics across millions or billions of grid points.
AI models do most of their expensive computation during training.
Once trained, producing a new forecast can be dramatically faster.
Google says WeatherNext Cyclones can produce a 15-day forecast in less than one minute on a TPU.
This makes huge ensembles feasible.
Instead of spending all available computing power generating a handful of expensive simulations, meteorologists could potentially generate hundreds or thousands of AI scenarios.
What does “one extra day of accuracy” really mean?
This claim can easily be misunderstood.
It doesn't mean hurricanes are now perfectly predictable one day earlier.
It means that when researchers compare forecast errors at different lead times, WeatherNext reaches an accuracy level at an earlier point that competing operational models generally reach closer to the event.
A simplified example might look like this:
| Forecast time | Older/leading operational model | WeatherNext |
|---|---|---|
| 5 days before | Large uncertainty | Lower uncertainty |
| 4 days before | Improving | Similar to competitor's 3-day quality |
| 3 days before | Good | Similar to competitor's 2-day quality |
| 2 days before | Very good | Better still |
This is why researchers describe the improvement in terms of lead-time advantage.
WeatherNext's three major cyclone improvements
Track
Where will the storm go?
WeatherNext predicts large-scale atmospheric circulation that steers the cyclone.
Intensity
How strong will it become?
The model forecasts changes in maximum wind speed and can help identify rapid intensification.
Wind structure
How far will dangerous winds extend?
This matters because two Category 3 hurricanes can have dramatically different geographical footprints.
The Nature results indicate a lead-time advantage across all three areas rather than track prediction alone.
Why predicting intensity is such a big deal
Hurricane track forecasting has improved enormously over several decades.
Intensity forecasting has been harder.
Rapid intensification remains especially dangerous.
For example, consider a storm expected to arrive with:
120 km/h winds
versus one that rapidly strengthens to:
250+ km/h winds.
Those represent dramatically different risks to:
- buildings,
- hospitals,
- electricity networks,
- telecommunications,
- ports,
- airports,
- emergency shelters,
- coastal communities,
- and evacuation planning.
Even if the track forecast were identical, getting the intensity wrong could have enormous consequences.
That's why WeatherNext's progress on intensity prediction is arguably as important as its track improvements.
WeatherNext 2 and WeatherNext Cyclones are being opened to researchers
Google announced that it is open-sourcing WeatherNext 2 and WeatherNext Cyclones, including code and model weights.
Researchers can therefore study, test and build upon the technology rather than having access only to Google's published forecasts.
Google has also introduced WeatherNext 2-mini, a smaller version designed to be practical enough to run on a single TPU, including through a public Colab environment.
Potential users include:
- universities,
- national meteorological agencies,
- climate researchers,
- emergency-management organizations,
- humanitarian organizations,
- energy companies,
- agricultural researchers,
- transportation organizations,
- and disaster-response groups.
But WeatherNext does not replace official hurricane warnings
This point is crucial.
WeatherNext's forecasts remain a forecasting input, not an official emergency warning.
Google itself tells users to rely on their national meteorological agencies for official forecasts and warnings. Weather Lab also labels its cyclone predictions as experimental.
Human meteorologists still incorporate information from:
- hurricane-hunter aircraft,
- weather satellites,
- ocean buoys,
- radar,
- ships,
- surface stations,
- conventional numerical models,
- AI models,
- historical climatology,
- and their own meteorological expertise.
The strongest forecasting system is therefore likely to be AI working alongside scientists, rather than AI operating entirely independently.
Why this development matters
The importance of WeatherNext is larger than simply producing a more accurate map.
Cyclones can produce several deadly hazards simultaneously:
Extreme wind
Damages homes, electricity networks and infrastructure.
Storm surge
Ocean water pushed inland can inundate coastal communities.
Extreme rainfall
Slow-moving storms can produce catastrophic flooding.
Landslides
Mountainous regions may face major landslide risks after extreme precipitation.
Large waves
Shipping, ports and coastal infrastructure can be threatened far from the storm's eye.
A forecast that gives governments and communities even 12–24 additional hours of reliable information can influence emergency decisions.
WeatherNext's roughly one-day lead-time improvement is therefore potentially highly significant.
The bigger picture: AI is changing weather forecasting
WeatherNext is part of a broader shift toward AI-based meteorology.
Historically, weather prediction was dominated almost completely by numerical simulations of physical equations.
The emerging architecture looks different:
Observations → AI models + physical models → thousands of possible scenarios → human forecaster → public warning
AI provides speed and pattern recognition.
Physics-based models provide scientifically grounded atmospheric simulation.
Observational systems provide information about what is actually happening.
Meteorologists interpret the combined evidence.
Together, they could enable forecasts that are both faster and more accurate.
In simple words
Think of conventional cyclone forecasting as a meteorologist asking a handful of extremely sophisticated computer models:
“What do you think this storm will do?”
WeatherNext allows scientists to ask something closer to:
“Based on everything learned from decades of global weather and thousands of previous cyclones, generate hundreds or thousands of plausible futures and tell me how likely each dangerous outcome is.”
That probabilistic approach is particularly valuable for hurricanes because the most important question isn't always what is most likely to happen.
Sometimes emergency planners also need to know:
“What is unlikely, but could be catastrophic if it happens?”
WeatherNext's very large ensembles are designed to help answer exactly that question.
Bottom line
Google's WeatherNext Cyclones represents an important milestone because it improves track, intensity and wind-structure forecasting simultaneously, can produce forecasts as far as 15 days ahead, runs extraordinarily quickly and can generate ensembles containing up to 1,000 possible storm scenarios.
Peer-reviewed results published on August 6, 2026 indicate an average forecasting lead-time advantage of at least about one day over leading operational systems—an improvement the researchers compare with roughly a decade of historical progress in cyclone forecasting.
Its performance during Hurricane Melissa in 2025 also demonstrated that AI cyclone forecasts can provide useful operational guidance, including information about extreme intensification that was incorporated into National Hurricane Center decision-making.
The technology therefore isn't simply about knowing tomorrow's weather more accurately. Its most important potential benefit is giving people, emergency services and governments more time to prepare before an extreme storm arrives.