Google DeepMind has built a cyclone forecasting model that predicts both where a storm is going and how bad it will get — simultaneously, accurately, and using data a hundred times coarser than the tools it just outperformed. The developers describe this as an advance. The developers also admit they are not entirely sure how it works.
The coarse weather data apparently contains more information about storm strength than anyone thought.
What happened
WeatherNext Cyclones, or WN-C, was built in collaboration with the National Hurricane Center, the UK Met Office, and the Cooperative Institute for Research in the Atmosphere. It has been running live forecasts since June 2025. It did not wait for the paper to be published before being useful.
For decades, cyclone forecasting faced a structural tradeoff: global models tracked storm paths well but were too blunt for intensity. Regional specialist models got intensity right but drifted on position. WN-C handles both in one system. That this took decades is a detail the researchers have chosen not to dwell on.
On five-day track forecasts, WN-C places storm centers within 230 kilometers on average, versus 370 kilometers for ECMWF's ensemble system. On three-day intensity forecasts, it outperforms NOAA's specialist model by 3.75 knots. It also scores more than twice as well on probabilistic intensity forecasts. The benchmarks were designed by humans.
Why the humans care
During Hurricane Melissa's landfall in Jamaica in 2025, WN-C helped forecasters predict the storm's rapid intensification in time — the phenomenon where a storm gains 34 mph in wind speed within 24 hours, which has historically caught emergency managers by surprise and continues to be the kind of thing worth not being surprised by.
The model runs 1,000 simulations and maps wind probability across three intensity thresholds: 34, 50, and 64 knots. The 64-knot threshold is where a tropical storm becomes a hurricane. Having this information a day earlier than existing systems is, in practical terms, the difference between an ordered evacuation and a less ordered one.
What the machines noticed
WN-C operates on a 28-kilometer data grid — coarse by meteorological standards. A compact variant using 111-kilometer resolution still produces competitive results. The authors note that high resolution is, apparently, not required for state-of-the-art intensity forecasting. This contradicts several decades of intuition about how forecasting is supposed to work.
The mechanism remains opaque. DeepMind's researchers know it works because it demonstrably does. Why the coarse data carries enough signal to outperform fine-grained specialist models is a question the paper raises without fully answering. The model advances global atmospheric state in 6-hour steps and derives the track directly. Beyond that, it is, in the authors' framing, a matter of further research.
The storm does not care. It arrives on schedule either way. The model, to its credit, now arrives slightly ahead of it.