Google’s AI weather model now uses more raw satellite data
WeatherNext 3 adds satellite observations, hourly forecasting and location-specific physical inputs, potentially strengthening Google’s weather services while exposing accuracy gaps, computational tradeoffs and model anomalies investors should monitor.
Google is a major participant in AI-based weather forecasting, where machine-learning models can approach traditional models’ performance while requiring far less computing power. That efficiency enables more frequent forecasts. WeatherNext 3, described in a Google white paper, now ingests some satellite weather data, reducing the lag between observed conditions and new forecasts. Forecast frequency has increased from every six hours to hourly. Google also increased spatial resolution and model size, while adjusting its processes to limit additional computational demands. Most AI weather models rely on reanalyses: consistent global atmospheric snapshots created by combining numerous observations and estimating conditions where direct measurements are unavailable. Reanalyses can lose information from raw data and are generally produced every six hours. Traditional models often incorporate more raw observations and use physics-based simulations. WeatherNext 3 adds a separate machine-learning model trained on satellite-based precipitation estimates, providing multiple precipitation forecasts. It also uses limited physical information for requested locations—including whether a point is land or ocean and its surface elevation—to calculate surface temperature and dew point. Google says training with weather-station data tagged with those attributes improves predictions. The 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%. Google generally reports that WeatherNext 3 outperforms the European Centre for Medium-Range Weather Forecasts’ AI model on these measures. Performance is not uniformly better. For several variables, WeatherNext 3 trails competing models’ initial six-hours-ahead forecast before leading over the remainder of a 15-day forecast. Some precipitation maps show distinctly hexagonal grid patterns, while sets of surface-temperature forecasts can produce inconsistent global average temperatures. WeatherNext 3 now supplies forecast information across Google Search, Gemini and Maps. Its broader data inputs and lower computing requirements could support more frequent, widely deployed forecasting. However, the model’s computational tradeoffs, visual anomalies and occasional early-horizon weakness remain relevant risks as Google expands AI weather services.