Google DeepMind's WeatherNext Cyclones Gives Forecasters an Extra Day to Save Lives
Google DeepMind's WeatherNext achieves a decade of forecasting progress in one leap, now open-sourced with 1,000-scenario ensemble predictions per storm
- Nature paper published: WeatherNext Cyclones achieves state-of-the-art accuracy in cyclone track, intensity, and wind structure — equivalent to a decade of forecasting progress.
- +24 hours of lead time: 3-day forecasts now match the quality of prior models' 2-day forecasts, on average across all three forecast dimensions.
- Open-sourced: Code and model weights for WeatherNext Cyclones, WeatherNext 2, and WeatherNext 2-mini are freely available on GitHub.
- 1,000-scenario ensembles: Each 15-day forecast scenario runs in under a minute on a TPU; this season forecasters get 1,000 probabilistic scenarios per storm via WeatherLab.
- Hurricane Melissa proof point: WeatherNext predicted Category 5 landfall in Jamaica 5 days in advance — the NHC's first-ever forecast of a storm going from Cat 1 to Cat 5 before landfall.
- Resolution surprise: The model achieves top intensity accuracy at 28km resolution, 100x coarser than traditional models, challenging a core assumption in the field.
Tropical cyclones have killed more than 700,000 people and caused $1.4 trillion in economic losses over the past 50 years. The core problem for forecasters has always been time: every hour of additional warning translates directly into lives saved. Google DeepMind just published a paper in Nature showing that its WeatherNext Cyclones model achieves state-of-the-art accuracy in predicting a storm's track, intensity, and wind structure , and is now open-sourcing the code and model weights for anyone to build on.
A decade of progress, compressed into one model
On average, WeatherNext gives forecasters an extra day's worth of predictive accuracy: three-day forecasts are as good as what prior models could provide for only the next two days. This scale of improvement corresponds roughly to a decade's worth of meteorological progress. That's not a marginal gain , it's the kind of leap that changes how emergency managers plan evacuations.
The benchmark numbers back this up. FGN's ensemble mean track position error is significantly lower than GenCast's, offering approximately a 24-hour advantage in accuracy at 3-5.5 day lead times. And this holds across all three dimensions of a cyclone forecast:
- Track: where the storm goes
- Intensity: how strong it gets (measured in knots)
- Wind structure: the spatial extent of damaging winds
The NHC's annual verification report confirmed WeatherNext as the top-performing individual model for both track and intensity, supporting expert decision-making throughout the hurricane season.
The problem it actually solves
For decades, cyclone forecasting forced a painful trade-off. Models consistently excelled at predicting a storm's path or its intensity, but rarely both simultaneously. Larger global models proved adept at tracking, yet lacked the granular detail to assess the small-scale atmospheric processes driving rapid intensification, while high-resolution local models, capable of capturing intensity changes, often struggled with broader contextual accuracy.
WeatherNext bridges this gap with a single model. It achieves this through a unique combination of its training, architecture, and approach to low-resolution inputs , predicting both global atmospheric patterns and fine-scale cyclone behavior end-to-end.