Google DeepMind earns 2026 honor for WeatherNext Cyclones

Google DeepMind is a 2026 science-fair winner for WeatherNext Cyclones, its probabilistic tropical-cyclone forecasting model. A peer-reviewed study published August 6, 2026, reported results against specific forecasting baselines, including a five-day track-error comparison and a separate three-day intensity result.

What WeatherNext Cyclones forecasts

WeatherNext Cyclones forecasts the atmosphere alongside tropical-cyclone track, intensity and wind structure. Its probabilistic forecasts use an ensemble: a set of possible scenarios rather than one predicted path. The model can generate ensembles with up to 1,000 members, forecast in six-hour steps and extend predictions up to 15 days.

Two different performance comparisons

The study measured track and intensity forecasts separately, with different forecast horizons and comparison models. At five days, WeatherNext Cyclones’ ensemble-mean track error averaged 230 km (about 143 miles). ECMWF ENS averaged 370 km (about 230 miles), while GenCast averaged 335 km (about 208 miles). For mean intensity forecasts at three days, WeatherNext Cyclones was 3.75 kt more accurate than HAFS.

How forecast guidance reaches forecasters

Forecasts have been publicly available through Google Weather Lab since June 2025. During the 2025 Atlantic hurricane season, experimental forecasts were supplied to the National Hurricane Center and incorporated into its guidance process.

In its separate account of Hurricane Melissa, Google identifies the system as WeatherNext and reports that forecasts of rapid intensification and landfall in Jamaica reached about 80% confidence five days before landfall and nearly 100% three days before.

Forecast guidance can inform meteorologists, but relevant national meteorological authorities issue official warnings.