Google DeepMind AI Delivers Faster Cyclone Forecasts
Google DeepMind published a Nature paper on 6 August 2026 showing its WeatherNext model beats operational systems at predicting cyclone track, intensity and wind structure. It buys forecasters roughly one extra day of accuracy. A three-day WeatherNext forecast is about as reliable as a two-day forecast from the models it was measured against.
One day sounds small. For evacuation planning, it isn’t.
Google also released the weights. That second part matters more than the accuracy numbers, and most coverage skipped past it.
What the model actually does
WeatherNext produces a full 15-day forecast covering track, intensity and wind structure in under a minute on a single TPU chip. Traditional physics-based systems need supercomputer time to do the same job.
It was trained on close to 20 terabytes of global atmospheric data, paired with IBTrACS, the international archive holding records of nearly 5,000 historical storms.
The architecture uses Functional Generative Networks to spin out up to 1,000 possible outcomes per run. That spread is the point. Rare, dangerous behaviour like sudden rapid intensification shows up in the tail of the distribution, not the average.
The version the National Hurricane Center ran during the 2025 season used a 50-member ensemble.
| Detail | Figure |
|---|---|
| Forecast horizon | 15 days |
| Time to generate | Under a minute on one TPU |
| Training data | ~20TB atmospheric data plus IBTrACS |
| Historical storms in training | ~5,000 |
| Scenarios per run | Up to 1,000 |
| Input resolution | 28km, or 111km for the mini variant |
| Benchmark period | Cyclones from 2023 through 2025 |
| Compared against | ECMWF-ENS for track, HWRF for intensity |
The Hurricane Melissa case
During the 2025 season the model helped the National Hurricane Center call Melissa’s rapid intensification and its landfall in Jamaica. Reports put the Category 5 landfall call five days out. The NHC documented the performance in its 2025 verification report.
That’s the kind of lead time that changes what an emergency management team can physically get done.
Context for the stakes: tropical cyclones have killed more than 700,000 people and caused around $1.4 trillion in economic damage over the past 50 years.
The part that should bother researchers
Here’s what I find genuinely strange about this result, and I haven’t seen anyone push on it properly.
WeatherNext runs on 28km input data. Conventional cyclone models work at resolutions roughly a hundred times finer. The mini variant drops to 111km and still performs respectably.
That shouldn’t work. Cyclone intensity depends on structure at scales far smaller than 28km, and the standard assumption in the field is that you need fine-grained inputs to capture it.
Nobody has a clean explanation yet. Google’s own researchers describe it as an open question.
Which is either the most interesting thing in the paper or a warning sign, depending on how charitable you’re feeling. A model that gets the right answer from inputs that shouldn’t contain enough information is a model whose failure modes you can’t predict.
Forecasters are enthusiastic, not converted
The NHC’s framing is careful. It’s adding AI systems to the toolbox specifically because they fail differently from numerical weather prediction models, not because they’re better across the board.
Science operations officer Wallace Hogsett said forecasters still need to work out why an AI model produced a given result, and understand where it holds up across a full range of storms.
Independent meteorologist Matt Lanza called performance during Hurricane Erin encouraging, good enough to earn a slot in his daily model rotation. He also flagged the obvious caveat. A strong average across a storm’s life cycle doesn’t mean strong performance at every stage of it.
Ryan Torn at University at Albany raised the sharper concern. AI systems trained on historical data tend to smooth abrupt atmospheric changes, and they can miss behaviour that falls outside the record they learned from.
Climate change keeps producing exactly that kind of behaviour.
Why open sourcing changes the picture
Google put code and weights on GitHub under Apache 2.0 across three variants: WeatherNext Cyclones, WeatherNext 2, and WeatherNext 2-mini. The mini runs in a free Colab notebook.
Read that last line again. A national weather service with no supercomputer budget can now run modern cyclone guidance on free cloud hardware.
DeepMind named the target directly, pointing at agencies in the Caribbean, the Pacific Islands and South Asia. Those regions absorb a wildly disproportionate share of cyclone deaths, and they’re the ones that have never been able to afford state-of-the-art physics ensembles.
The larger models still need serious accelerator hardware, and the repository is labelled experimental research code. For official warnings, DeepMind points people back to their national weather agency.
Collaborators on the work included the NHC, the Cooperative Institute for Research in the Atmosphere, and the UK Met Office. Live predictions run through Weather Lab, DeepMind’s public interface.
It arrives during a rough month for Google’s AI division, which restructured its leadership on 5 August and lost several senior researchers. This is the counterweight, and it’s a strong one.
We’re tracking how research models like this feed into commercial products in our coverage of AI trends in 2026, and GuideAITools will follow the 2026 season results as they come in.
