Google has released WeatherNext 3, an updated AI weather model that now incorporates raw satellite data. By moving beyond reliance on reanalysis data, the model can generate forecasts on an hourly basis, significantly reducing the time lag compared to previous versions that relied on six-hour snapshots.
The new model features increased spatial resolution and a larger architecture, requiring process adjustments to manage computational demand. Additionally, Google introduced a separate machine-learning model specifically for precipitation estimates and integrated physical data, such as surface elevation and land-ocean classification, to improve temperature and dewpoint accuracy.
Performance data indicates a 5 percent improvement in upper atmosphere accuracy compared to WeatherNext 2, which the company equates to six additional hours of forecast lead time. Furthermore, surface temperature accuracy for specific locations has seen gains of up to 30 percent, outperforming the ECMWF AI model in these metrics.
Despite these improvements, the model exhibits some anomalies, such as hexagonal patterns in precipitation maps and inconsistent global average temperatures across different forecast snapshots. The system also shows lower performance in initial six-hour forecasts before improving over the remainder of the 15-day period.
WeatherNext 3 is currently deployed across Google services, including Search, Gemini, and Maps. The development team describes the integration of low-latency observation data as a significant advancement for AI-driven meteorological forecasting.
