Update to Google’s AI weather model improves forecast accuracy
Google’s WeatherNext 3 adds satellite observations, hourly forecasts and location-specific physical data, improving accuracy versus its predecessor and generally versus ECMWF’s AI model, although early-horizon weaknesses and unusual grid patterns remain.
Google has released WeatherNext 3, the latest version of its artificial-intelligence weather model. AI systems can approach traditional forecast performance while requiring far less computing power, enabling more frequent updates. Most AI weather models rely primarily on reanalysis: a consistent global atmospheric snapshot that combines observations and estimates conditions where direct measurements are unavailable. Reanalyses are generally produced every six hours and can lose information from underlying data. WeatherNext 3 adds some satellite observations, reducing the lag between current conditions and a new forecast, and increases forecast frequency to hourly. Google also increased spatial resolution and enlarged the machine-learning model, while adjusting processes to limit higher computational demands. A separate model trained on satellite-based precipitation estimates provides multiple precipitation forecasts. The system adds limited physical information for requested locations, including whether a point is land or ocean and its surface elevation. Google says training with weather-station data tagged with those characteristics improves surface-temperature and dew-point predictions. The accompanying white paper reports roughly a 5% improvement in upper-atmosphere accuracy versus WeatherNext 2, equivalent to about six additional hours of accurate forecast lead time. Location-specific surface-temperature calculations improved accuracy by up to 30%. WeatherNext 3 generally outperformed the European Centre for Medium-Range Weather Forecasts’ AI model on these measures. Performance is not uniformly superior. For several variables, the model trails competing systems’ initial six-hours-ahead forecast before pulling ahead over the remainder of a 15-day forecast. Some precipitation maps display distinctly hexagonal grid patterns, and multiple surface-temperature forecasts can produce inconsistent global average temperatures. For investors, the update signals Google’s effort to expand AI capability while retaining computing efficiency. WeatherNext 3 now supplies forecast information across Google Search, Gemini and Maps. Its broader significance will depend on whether satellite inputs, higher resolution and hourly updates deliver reliable value without undermining efficiency or consistency.