Google Unveils WeatherNext 2 AI Weather Model

Google DeepMind and Google Research announced WeatherNext 2 on Wednesday, unveiling their most advanced artificial intelligence model for weather forecasting. The system promises to deliver global predictions up to eight times faster than previous versions, with hourly resolution that enhances accuracy for variables such as temperature, wind speed, humidity, precipitation, and pressure over the next 0 to 15 days.

The model's core innovation lies in its ability to generate hundreds of possible weather scenarios from a single input of current global conditions. This approach simulates a range of outcomes, including low-probability extreme events like storms or cyclones, which traditional physics-based models on supercomputers would take hours to produce. In contrast, WeatherNext 2 completes each prediction in less than a minute using a single tensor processing unit, or TPU. The technology employs a novel Functional Generative Network, or FGN, which introduces controlled noise into the model architecture. Trained primarily on individual weather elements, known as marginals, the FGN learns to forecast interconnected systems, or joints, such as regional heat impacts or wind farm...