3 days ago

Google Upgrades AI Weather Model, Promising More Accurate Forecasts

Google has introduced WeatherNext 3, an upgraded artificial intelligence weather-forecasting model that the company says delivers more accurate and higher-resolution global forecasts by incorporating real-time satellite observations directly into its prediction system.

The new model can generate forecasts every hour, compared with the six-hour update cycle of WeatherNext 2, while providing atmospheric predictions at resolutions of up to 5 kilometres. Google says this makes the system roughly five times sharper than its predecessor, which operated on a 25-kilometre grid.

Real-Time Satellite Data

A key change in WeatherNext 3 is the use of a mosaic of live global geostationary satellite data. Earlier versions, including WeatherNext 2, were primarily trained using data produced by numerical weather prediction models.

Google says directly incorporating observations allows WeatherNext 3 to maintain a more continuously updated representation of atmospheric conditions and respond more quickly to rapidly developing weather systems such as storms and precipitation events.

The model also uses detailed observational data from weather stations, helping it account for regional characteristics such as topography that can influence temperature, humidity and other surface conditions.

Precipitation has traditionally been one of the more difficult variables for global weather models to predict accurately.

Google trained WeatherNext 3 using precipitation datasets including NASA’s Integrated Multi-satellite Retrievals for GPM (IMERG) and a Google global precipitation reanalysis based on satellite radar data.

The company says evaluations show improvements in precipitation forecasting of up to 60% against IMERG and 30% against MRMS, while early lead-time comparisons against rain-gauge measurements showed an improvement of up to 10%.

Google also says that for forecasts made at least a day ahead, users could see up to 50% more accurate precipitation forecasts, with the largest improvements expected in regions where forecasting has historically been less reliable.

Applications Beyond Daily Weather

WeatherNext 3 is also designed to support applications that require detailed forecasts of atmospheric conditions.

The model includes 100-metre wind forecasts intended to help estimate wind-turbine output, along with high-resolution cloud and solar-radiation forecasts that could assist solar-power operators in estimating electricity generation.

These capabilities could make AI-based forecasting useful for renewable-energy planning, agriculture, logistics and disaster preparedness, particularly in regions where high-resolution conventional forecasting is expensive or limited.

WeatherNext 3 is being integrated into Google Search, Gemini, Google Maps, Google Maps Platform and Google Earth Engine. Google has also made WeatherNext 3 forecasts available through its Weather Lab platform for researchers and developers.

The development reflects a broader shift in meteorology toward combining traditional physics-based forecasting with AI. Government weather agencies continue to rely heavily on numerical weather prediction and high-performance computing, while AI models can generate forecasts considerably faster and with lower computational requirements.

Google says WeatherNext 3 does not eliminate the inherent uncertainty of weather forecasting, but its combination of real-time observations, higher spatial resolution and hourly updates is intended to improve predictions of rapidly changing local weather conditions.

By Subramanya Joshi