- WeatherNext 3 uses 12 recent hourly frames from an 11-channel geostationary satellite mosaic alongside ECMWF analysis data. <sup>[1]</sup>
- Google reports up to 60% improvement in precipitation CRPS against IMERG, 30% against MRMS and 10% against rain gauges for early lead times. <sup>[2]</sup>
- The model is available through Google products and cloud services, but Google describes it as experimental and directs users to national meteorological agencies for warnings. <sup>[3]</sup>
Google has released WeatherNext 3, an AI weather model that uses live geostationary satellite observations to refresh forecasts hourly instead of relying only on atmospheric analyses produced every six hours. The model delivers surface predictions at up to 5-kilometer resolution and is being integrated into Search, Gemini, Maps, Google Maps Platform and Earth Engine. [4]
The added inputs are more specific than a generic “satellite data” label suggests. WeatherNext 3 takes 12 recent hourly frames from an 11-channel geostationary satellite mosaic, with the newest imagery arriving in under an hour, and combines them with conventional ECMWF analysis fields. [1]
What Google measured #
Google’s paper evaluates the model with several metrics rather than a single accuracy score. For precipitation, the company reports improvements in Continuous Ranked Probability Score of up to 60% against NASA’s IMERG satellite precipitation product, 30% against the U.S. Multi-Radar Multi-Sensor system and 10% against rain-gauge measurements at early forecast lead times. [2]
The model also predicts temperature and humidity at individual weather-station locations, including stations excluded from training, and produces hourly forecasts for variables such as cloud cover and solar radiation. Google says WeatherNext 3 outperforms its predecessor and ECMWF’s probabilistic AI system on most of its analysis-based tests. [2]
Operational limits #
The release improves the timeliness and geographic detail of automated guidance, especially for fast-changing rain and temperature patterns. But the benchmark results do not establish that every local forecast or severe-weather warning will improve by the same amount; they compare particular variables, lead times and reference datasets. Independent research has also found that AI weather systems can struggle with physical realism and record-breaking extremes, areas where human forecasters and physics-based models remain important. [5]
Google’s developer documentation calls WeatherNext 3 an experimental AI forecasting system, and its product release says users should rely on local meteorological agencies for official forecasts, warnings and public-safety advisories. [3]
Companies mentioned #
Further sources #
The stories that matter, in one email. Free — unsubscribe anytime.