Google's WeatherNext 3 taps raw satellite data for faster, sharper forecasts

Google has long been a major force in AI-powered weather forecasting, where machine-learning models offer a key advantage: they can match the accuracy of traditional physics-based systems while running on a fraction of the computing resources, enabling more frequent updates.

The company's newly released WeatherNext 3 model represents a significant methodological shift. Instead of relying exclusively on pre-processed global weather snapshots, the updated system now ingests raw satellite data directly—a move that allows forecasts to be generated hourly rather than at fixed intervals.

A shift away from intermediate processing#

Traditionally, AI weather forecasters have operated by feeding machine-learning algorithms into reanalyses—composite models that stitch together diverse observations into a single, globe-spanning atmospheric picture. This approach is computationally neat but lossy: the raw data sources often contain nuances that dissolve during the blending process, and these standardized snapshots typically only refresh every six hours.

Conventional forecast systems built on physics principles have long sidestepped this bottleneck by ingesting raw measurements directly, capturing granular detail about atmospheric conditions. WeatherNext 3 now adopts a hybrid strategy, pulling in satellite observations and enabling hourly forecast cycles as a result.

Additional refinements#

The upgrade includes several complementary changes designed to sharpen predictions:

• Higher spatial resolution allows finer-grained detail across forecasts
• Larger machine-learning model improves the system's ability to identify patterns, though it required optimizations to keep computational costs in check
• Dedicated precipitation module adds a separate machine-learning model trained on satellite-based precipitation data, offering multiple rainfall forecasts

More details are available in a technical white paper published by the team.