Google DeepMind and Google Research today launched WeatherNext 3, an advanced artificial intelligence model designed to deliver more accurate and higher-resolution global weather forecasts. The model updates hourly and leverages real-time satellite data, a departure from traditional numerical weather prediction (NWP) models that often rely on physics simulations and carry a data lag. This new approach allows WeatherNext 3 to provide a global weather picture that is roughly five times sharper than Google's previous model, WeatherNext 2.
WeatherNext 3 generates forecasts at multiple spatial resolutions. Key surface variables, such as temperature and moisture, are visualized at a 5-kilometer resolution. Other surface variables are available at 10 kilometers, while atmospheric variables like wind speed are forecast at 25 kilometers. In contrast, WeatherNext 2 produced forecasts on a 25-kilometer grid in six-hour increments. Samier Merchant, a Google senior staff engineer, stated that WeatherNext 3 moves beyond the data most global AI models typically train on.
The model’s enhanced capabilities stem from its ability to learn directly from real-time observations, specifically a mosaic of live, global geostationary satellite data. This direct ingestion of raw satellite imagery enables WeatherNext 3 to generate new forecasts every hour, providing more timely and localized predictions for rapidly changing weather events like rain and snow. Traditional AI weather models often depend on datasets from institutions like the European Centre for Medium-Range Weather Forecasts (ECMWF), which can take approximately five hours to compile and are refreshed every six hours. WeatherNext 3 incorporates already-assembled ECMWF output and layers in real-time geostationary satellite imagery to achieve its hourly update cadence. Ilan Price, a DeepMind senior research scientist, noted that the model gains accuracy by using the most recent information without waiting for the next analysis date.
Precipitation forecasting has seen a significant improvement with WeatherNext 3. The model trains on NASA's satellite-based Integrated Multi-satellite Retrievals for GPM (IMERG) dataset and a Google-produced precipitation reanalysis. Google reports that the new model scores up to 60% better on a standard probabilistic accuracy measure for precipitation compared with WeatherNext 2. For longer-term forecasts, users can expect up to 50% more accurate precipitation predictions, particularly in regions where historical forecasts have been less reliable.
Beyond general forecasting, WeatherNext 3 introduces predictions specifically tailored for the renewable energy sector. The model forecasts wind speeds at 100 meters, which is approximately the height of a wind turbine, along with high-resolution cloud cover and solar radiation levels. This data aims to assist grid operators and renewable energy developers in estimating power output from wind and solar assets. The model also provides forecasts for areas in Latin America, Africa, and Asia-Pacific that have historically received less high-resolution forecasting coverage.
WeatherNext 3 is now integrated into various Google products and services, including Google Search, Gemini, Google Maps, and the Google Maps Platform Weather API. Developers and researchers can also access the forecast data through Google Cloud's BigQuery, Earth Engine, and Cloud Storage. Independent evaluations by Brightband on the Operational WeatherBench test indicate that WeatherNext 3 has shown strong performance against other deep learning models from Microsoft, Nvidia, and ECMWF, as well as traditional forecasts from the U.S. National Weather Service.