Google gives WeatherNext 3 fresh satellite readings for hourly forecasts at five-kilometer resolution
Google DeepMind and Google Research released WeatherNext 3 on Thursday, an AI weather model that takes in fresh satellite readings and produces a new global forecast every hour, with some variables shown at up to five-kilometer resolution.
Justine Calma at The Verge writes that WeatherNext 2 produced forecasts every six hours on a 25-kilometer grid. The new model can show temperature and moisture on a grid five times sharper.
Traditional forecasting starts with supercomputers solving equations that describe how the atmosphere moves. AI weather models learn patterns from old weather records and produce predictions faster. WeatherNext 3 adds the latest satellite observations before making the next hourly forecast.
That turns the satellite feed into the clock. A forecast that refreshes every six hours is a report. One that refreshes hourly can sit inside Search, Maps, and Gemini as a live product layer.
Google says those products are already incorporating WeatherNext 3. The model can also aim a forecast at a specific weather station rather than only averaging conditions across a wide grid, Tim Fernholz at TechCrunch reports.
The company says precipitation forecasts are up to 50 percent more accurate when looking at least a day ahead. TechCrunch reports Google’s rain evaluation improved 60 percent over WeatherNext 2 and that the model has 2.4 times as many parameters. Those are Google’s results and TechCrunch’s account of its evaluation, not a guarantee for the next storm.
Rain is where the hourly satellite input earns its keep. Fast-moving systems change between six-hour runs, and ground gauges are sparse outside the United States and Europe. A fresh view from orbit gives the model another observation before it commits to the next grid.
It is still a hybrid system. WeatherNext 3 learns partly from data produced by physics-based weather models, and weather agencies compare multiple forecasts before issuing warnings. Google says its AI model should work beside traditional forecasting, not replace it.
The old supercomputer output remains in the training set. The new satellite reading moves the next answer closer to now.
Google also designed the model to forecast renewable-energy conditions, including wind speed around turbine height. Better hourly estimates of wind, rain, and cloud cover can change how much power an operator expects to have, not only whether a person carries an umbrella.
The five-kilometer grid is the sharper picture. The hourly satellite intake is the operating change.
—