Details
- Google DeepMind says WeatherNext is a new AI weather model published in Nature that improves cyclone track and intensity forecasting and adds about 24 hours of average lead time for preparation.
- The company says 3-day forecasts from WeatherNext now match the quality of prior models at 2 days out, which it describes as a decade’s worth of progress in one step.
- WeatherNext was trained on years of global atmospheric data plus a curated set of nearly 5,000 historical cyclones.
- The model can generate each 15-day probabilistic forecast scenario in under a minute on a TPU, and DeepMind says it can produce 1,000 probabilistic predictions per storm.
- During Hurricane Melissa, DeepMind says WeatherNext predicted a Category 5 landfall five days ahead with 80% confidence, and those forecasts are being made available through WeatherLab.
- DeepMind says it is open sourcing the code and model weights on GitHub so researchers and forecasters can adapt the system for academic work, operations, or localized models.
Impact
WeatherNext pushes AI weather forecasting further into operational territory, with a concrete gain in lead time that matters for emergency response and evacuation planning. The open release of code and weights also lowers the barrier for researchers and national weather services to test, adapt, and localize the model. In a field where rivals such as Huawei, NVIDIA, and university labs have been advancing data-driven forecasting, DeepMind is signaling that frontier AI can now compete on both accuracy and deployability.
