WeatherNext: AI model achieves breakthrough in forecasting cyclones

Google DeepMind reports that WeatherNext Cyclones improves lead time by more than 24 hours for cyclone track, intensity and wind structure: its three-day forecasts match the accuracy earlier systems delivered at two days. The model was evaluated on historical storms from 2023 and 2024, and the work was published in Nature.

It was co-trained on nearly 20 terabytes of atmospheric data and an archive of almost 5,000 storms. A 1,000-member ensemble produces probability maps, while one 15-day forecast takes less than a minute on a TPU. The model uses 28-by-28-kilometer inputs, far coarser than high-resolution approaches.

Google is releasing code and weights for WeatherNext Cyclones, WeatherNext 2 and a compact WeatherNext 2-mini that can run on one TPU.

Cyclone forecasting has long split global steering from fine-scale intensity physics. A single model improving both at coarse resolution could lower the computational barrier for agencies that cannot operate specialized high-resolution systems.

Signal Desk's read: the open release is as important as the headline benchmark. Weather agencies need to test regional bias, calibration and operational reliability themselves. Accessible weights permit that scrutiny and could support local adaptation instead of forcing trust in a hosted forecast.

This is a vendor-authored Google DeepMind account of its own model, so its breakthrough and real-world-impact claims are vendor-reported, although Nature publication and work with forecasting institutions add external scrutiny. The source does not include full basin-level error tables.

The unexplained performance at coarse resolution is both exciting and cautionary. Operational forecasters need to know where a model fails, not merely that its average is better. Ensemble calibration and rare rapid-intensification misses matter more than one overall lead-time figure.

Cyclones caused more than 700,000 deaths and $1.4 trillion in losses over 50 years, according to the source. Track depends heavily on broad atmospheric circulation, while intensity is driven by local processes, historically favoring different model types.

Open distribution can help resolve that uncertainty. Researchers can compare the model with physics-based systems, examine regional performance and test whether the claimed lead-time gain holds under storms unlike those in the historical evaluation set.