Google DeepMind's WeatherNext 2 Forecasts 8x Faster at Hourly Resolution

Google DeepMind and Google Research have released WeatherNext 2, a forecasting model the company calls its most advanced and efficient to date, capable of generating forecasts 8x faster than its predecessor with resolution down to a single hour.

The model can produce hundreds of possible weather outcomes from a single starting point, with each prediction taking less than a minute on a single TPU. Google says the same computation would take hours on a supercomputer running physics-based models.

WeatherNext 2 surpasses the previous state-of-the-art WeatherNext model on 99.9% of variables — temperature, wind, humidity — and lead times across the 0-15 day range, according to Google. The company published a Continuous Ranked Probability Score (CRPS) comparison between WeatherNext 2 and WeatherNext Gen as part of the release.

The performance gain rests on a new AI architecture Google calls a Functional Generative Network (FGN). The approach injects noise directly into the model architecture, keeping the forecasts it generates physically realistic and interconnected rather than producing statistically plausible but physically incoherent output — a known weakness of earlier generative approaches to weather simulation.

Trained on marginals, predicting joints

The architecture has an unusual property. The model trains only on "marginals" — individual, standalone weather elements such as the precise temperature at a specific location, wind speed at a certain altitude, or humidity. Yet from that training alone, it learns to forecast "joints": large, interconnected systems that depend on how all the individual pieces fit together.

Google says this joint forecasting underpins its most useful predictions, such as identifying entire regions affected by high heat or estimating expected power output across a wind farm. Ensembles capturing the full range of possibilities, including worst-case scenarios, are what meteorologists need for planning — and what Google says it has already tested with weather agencies through its experimental cyclone predictions.

From lab to product

Google is moving the research into production across its stack. WeatherNext 2's forecast data is now available in Earth Engine and BigQuery. The company is also launching an early access program on Google Cloud's Vertex AI platform for custom model inference.

On the consumer side, WeatherNext technology now powers upgraded weather forecasts in Search, Gemini, Pixel Weather, and Google Maps Platform's Weather API. In the coming weeks, it will also help power weather information in Google Maps itself.

The commercial stakes are considerable. Weather drives decisions across global supply chains, flight paths, and daily commutes — use cases Google explicitly cites — and AI-based forecasting that runs on a single TPU instead of a national supercomputer lowers the cost barrier for agencies and businesses that could never afford traditional numerical weather prediction infrastructure.

Google says it will continue integrating new data sources and expanding access to the model, and positions the open forecast data as a way to accelerate scientific discovery for researchers, developers, and businesses. Alongside the paper and developer documentation, the company points to Google Earth, Earth Engine, AlphaEarth Foundations, and Earth AI as entry points to its broader geospatial AI work.