Google rolls out WeatherNext 2 AI model; What does it mean for future of climate prediction?

Google on Monday launched an upgraded artificial intelligence model, WeatherNext 2, that the company says can generate weather forecasts eight times faster and at hourly resolution.

The model, now available to users outside research labs, can produce hundreds of possible weather scenarios in under a minute on a single TPU processor, compared with hours on traditional physics-based supercomputers, Google said in a blog post.

WeatherNext 2’s forecast data is being integrated into Google’s Earth Engine and BigQuery platforms, and the company has launched an early-access programme on Google Cloud’s Vertex AI for organisations that want to run customised model inference.

The system is also being used to upgrade forecasts in Google Search, Gemini, Pixel Weather and Google Maps Platform’s Weather API, with Maps itself to follow in the coming weeks.

The new model surpasses the capabilities of the earlier WeatherNext system across 99.9% of variables, including temperature, wind and humidity, and for lead times of up to 15 days. It uses a new architecture called a Functional Generative Network that injects noise directly into the model to produce physically consistent forecasts, Google said.

WeatherNext 2 is trained only on individual weather variables, or “marginals”, but the company said it can still generate accurate “joint” forecasts of complex, interconnected systems, such as identifying regions facing extreme heat or estimating expected power output across an entire wind farm.

Google said the technology has already helped weather agencies test multi-scenario cyclone predictions, and that broader access could support researchers, developers and businesses in making decisions related to supply chains, energy planning and climate risk.

The company added it is exploring new data sources and expanding access to advance the model’s capabilities further.