Google DeepMind open-sources WeatherNext, adding a day of cyclone forecast lead time
Google DeepMind open-sourced code and weights for WeatherNext Cyclones alongside a Nature paper with the National Hurricane Center, CIRA, and the UK Met Office: its three-day forecasts of cyclone track, intensity, and wind structure match what prior systems delivered at two days, roughly a decade of meteorological progress. The release includes WeatherNext 2 and a compact 2-mini that runs on a single TPU, putting frontier cyclone forecasting in reach of any national agency or research group without supercomputer-scale physics models. WeatherNext Cyclones hits that accuracy on 28x28km data, 100x coarser than traditional models — a result the authors say remains unexplained.
Source: deepmind.google ↗
Our three-day forecasts are as good as what prior models were able to provide for only the next two days.
Why this matters
- → Cyclone forecasts gain an extra day of warning — potentially saves lives in vulnerable regions.
- → Open-source model with 28km resolution outperforms traditional physics models at 100x coarser data.
- → Compact 2-mini version runs on single TPU, democratizes hurricane prediction for any national agency.