Google's WeatherNext 3 AI model now incorporates raw satellite data and basic physical properties like land/ocean and elevation to improve forecast accuracy. The model shows significant improvements over its predecessor and the ECMWF AI model, though some oddities in predictions persist.
The advancements in Google's AI weather model could lead to more accurate and timely weather forecasts, impacting various sectors from agriculture and transportation to disaster preparedness and daily planning. The integration of AI with physical data represents a step towards more robust and reliable predictive capabilities in meteorology.
Google's WeatherNext 3 artificial intelligence model for weather forecasting has been updated to utilize more raw satellite data and incorporate basic physical properties. Unlike traditional models that simulate physical processes based on location-specific data, machine-learning models like WeatherNext 3 are trained on historical patterns. The new version adds information about whether a location is land or ocean, and its surface elevation, to refine calculations for surface temperature and dewpoint.
The team behind WeatherNext 3 reports significant improvements in forecast accuracy. They noted approximately a 5 percent increase in upper atmosphere condition accuracy compared to their previous model, which translates to about six additional hours of reliable forecast lead time. The integration of specific location surface temperature calculations led to accuracy improvements of up to 30 percent. These metrics generally show WeatherNext 3 outperforming the European Centre for Medium-Range Weather Forecasts (ECMWF) AI model.
However, the white paper acknowledges some anomalies. In initial six-hour forecasts for certain variables, WeatherNext 3 sometimes performs worse before surpassing other models for the remainder of a 15-day forecast. Larger-scale pattern predictions also exhibit peculiarities, such as visible hexagonal shapes in precipitation maps, likely stemming from the model's grid structure. Additionally, the method used to generate multiple surface temperature forecasts to represent outcome ranges can result in global average temperatures that are inconsistently higher or lower than expected.
Despite these quirks, Google states that WeatherNext 3 represents a significant advancement in AI-based weather prediction by moving beyond purely analytical data to include information-dense, low-latency observational data. The updated model now serves as the primary source for forecast information across Google services, including Search, Gemini, and Maps.