← Back to feed

Google DeepMind published WeatherNext in Nature on August 6, 2026, showing that its AI model achieves state-of-the-art accuracy in predicting cyclone track, intensity, and wind structure. The headline result: 3-day forecasts as accurate as what conventional numerical models deliver at 2 days. DeepMind calls this “roughly a decade’s worth of meteorological progress.”

Both WeatherNext 2 and WeatherNext Cyclones — the models used operationally during the 2025 hurricane season — are now available on GitHub under an open-source licence.

What It Does

WeatherNext Cyclones takes global atmospheric conditions as input and iteratively predicts both large-scale weather patterns and fine-grained cyclone tracks up to 15 days in advance. A 1,000-member ensemble run generates localised probability maps of tropical storm-to-hurricane-force winds.

Running that ensemble on a Google TPU takes under one minute for a 15-day forecast — a practical ceiling that had previously required hours on numerical weather prediction infrastructure.

The model was developed with the National Hurricane Center (NHC), the Cooperative Institute for Research in the Atmosphere (CIRA), and the UK Met Office, among other agencies.

Proven in the Field

The model had real-world impact during the 2025 hurricane season. WeatherNext predicted Hurricane Melissa’s rapid intensification and landfall in Jamaica before conventional models did, enabling the NHC to issue an advance warning with enough lead time for ground teams to prepare.

That prediction became a landmark case study in AI-assisted emergency forecasting.

Why Tropical Cyclones Are Hard

Conventional approaches force a trade-off between global-scale weather modelling and fine-scale cyclone structure. Resolving both simultaneously requires enormous computational resources, which limits ensemble size. WeatherNext eliminates that trade-off by handling both within a single model pass, which is what enables the 1,000-member ensemble at practical cost.

Benchmark Numbers

• 3-day track, intensity, and wind-structure accuracy exceeds prior 2-day model performance
• 1,000 scenarios per cyclone for probabilistic decision support
• 15-day TPU forecast runtime: under 1 minute

Open-Source Release

WeatherNext 2 (the general weather model) and WeatherNext Cyclones (the cyclone-specialised variant) are available at github.com/google-deepmind/weathernext. DeepMind is releasing them to support local forecasters, renewable energy planning, and extreme weather preparedness research.

The paper is published in Nature alongside the code release.