應用研究
WeatherNext 3 Directly Ingests Satellite Observations, Relaunching Global Probabilistic Weather Forecasts Every Hour
Google DeepMind’s new model no longer merely imitates reanalysis data produced every six hours; it incorporates low-latency satellite imagery into forecast initial conditions. It improves precipitation and station-level predictions, but evaluations in mountainous and polar regions and for heavy rainfall still expose limitations caused by sparse data and imperfect metrics.
Google DeepMind and Google Research have introduced WeatherNext 3, advancing global AI weather forecasting from “learning analysis fields generated by numerical models” to directly assimilating observational data. Every hour, the system ingests mosaicked geostationary satellite imagery with 11 channels at 0.1-degree resolution, combined with ERA5 and ECMWF HRES analysis data. Whereas WeatherNext 2 is initialized every six hours, the new model can relaunch forecasts hourly, providing an average of roughly two to three hours of additional effective lead time for rapidly developing precipitation events.
The architecture remains centered on a Functional Generative Network, but its shared icosahedral-grid Transformer has been expanded from 24 layers and 768 dimensions to 32 layers and 1,024 dimensions. Data sources with different resolutions and update frequencies have their own encoders and decoders before being mapped onto a shared processing grid. Training resolution is progressively increased from 1 degree to 0.25 and then 0.1 degrees. In addition to gridded fields, dedicated output heads can natively predict precipitation, tropical cyclone tracks, and near-surface temperature and dew point at arbitrary coordinates, eliminating the need for a separate conventional downscaling pipeline.
In tests using 2024 data, the paper reports that probabilistic forecast scores for mid- and upper-atmospheric variables improved by an average of about 5% relative to WeatherNext 2—equivalent to roughly six additional hours of forecast lead time at the same accuracy. For stations excluded from training, short-range two-meter temperature CRPS improved by as much as 30%. Google has integrated the model into Search, Gemini, Maps, the Google Maps Platform Weather API, and Earth Engine.
These figures still require cautious interpretation. Some comparisons use HRES-derived fields as ground truth, and the interpolation method may favor the higher-resolution model. Heavy-rainfall samples are relatively scarce, with most gains concentrated near the rain/no-rain threshold. The paper also acknowledges biases in observation-sparse areas such as the Andes, the Himalayas, and polar oceans; discontinuities in some outputs at six-hour boundaries; and a tendency for the cyclone ensemble to underestimate uncertainty. The next question is whether independent real-time evaluations will confirm that these advantages persist for extreme weather and in regions outside the training data distribution.