During a real-world test, the system successfully predicted that Hurricane Melissa would strike Jamaica as a Category 5 storm five days before landfall, while it was still only a Category 1.
This increased accuracy is critical for disaster response, as officials from the National Hurricane Center state that even a few hours can determine the success of large-scale evacuations and the staging of emergency supplies.
Unlike traditional numerical models that are limited by available computing power, WeatherNext can generate 1,000 different potential scenarios for a single storm to account for the "butterfly effect"—a phenomenon where tiny atmospheric changes lead to vastly different outcomes.
This allows forecasters to see a much broader range of possibilities than the 50 scenarios typically produced by existing systems.
Technically, the model operates as a "black box," meaning researchers do not yet fully understand the specific mechanism the AI uses to identify intensity signals within low-resolution weather data.
DeepMind open-sourced the technology to encourage the broader scientific community to investigate these underlying physics and refine the tool further.
While the AI offers a significant leap in forecasting capability, experts emphasize that human specialists remain essential for interpreting these machine-learning outputs to communicate specific risks to the public.