DeepMind’s WeatherNext 3: AI Weather Model Jumps to 5 km, Raw-Satellite Forecasting

Google’s AI lab DeepMind has introduced “WeatherNext 3,” a new weather-forecasting model that skips traditional physics-heavy numerical simulations and instead trains directly on raw satellite and observational data. The system reportedly produces hourly forecasts with spatial resolution down to about 5 kilometers, a substantial jump in granularity compared to many operational models.

DeepMind has previously worked on AI forecasters like GraphCast, but WeatherNext 3 is aimed squarely at fast-changing, local phenomena: flash floods, damaging winds, and other extreme events. High‑resolution short‑term forecasts are critical for disaster preparedness, grid planning for wind and solar power, and fine‑tuning airline and logistics routes—all areas where a few hours’ better warning can translate into real economic value.

Researchers and public agencies, however, still worry about the opaque nature of large AI models. When the system is wrong, it can be hard to pinpoint why. European meteorological services and national regulators are now debating how to combine such AI forecasters with established physics-based models and what kinds of validation and oversight will be needed before governments lean on them for critical infrastructure decisions.