Update to Google’s AI weather model improves forecast accuracy

Google’s latest AI forecasting system, WeatherNext 3, injects a small amount of physical context into a predominantly machine-learning approach. Unlike traditional numerical models that simulate physical processes from first principles, WeatherNext 3 checks whether a queried location is land or ocean and uses its surface elevation when computing surface temperature and dew point. The training process pairs those location tags with historical weather station data to teach the model how those physical factors relate to observed conditions. In results presented in a white paper, Google reports measurable gains compared with its prior WeatherNext 2 model and with an ECMWF AI model. The team cites about a 5 percent increase in accuracy for upper-atmosphere conditions relative to their previous model, which they say corresponds to roughly six additional hours of reliable forecast lead time. Adjustments to how the model calculates surface temperature for specific points produced accuracy improvements of up to 30 percent. The paper also notes some anomalies and limitations. For several variables, WeatherNext 3 performs worse than peers in the first six hours of forecasts before overtaking them over the remainder of a 15-day window. Forecast maps show artifacts from the model grid in some fields—for example, precipitation patterns that appear as distinct hexagonal blobs—and the ensemble method used for surface temperature occasionally yields snapshots where the global average temperature is noticeably higher or lower than expected. Google frames WeatherNext 3 as moving beyond purely analysis-driven AI forecasting by incorporating dense, low-latency observational inputs. The company has shifted its public forecast outputs to WeatherNext 3, making it the source of weather information across Google Search, Gemini and Maps.