Google Bets WeatherNext 3 Can Make Rain Forecasts Sharper—and Faster

Google Bets WeatherNext 3 Can Make Rain Forecasts Sharper—and Faster
Google’s pitch is simple: fresher satellite data and AI pattern recognition can make weather alerts more useful, particularly where ground observations are scarce. But the company also acknowledges that its newest model remains part of a broader forecasting toolkit led by agencies and physics-based systems.

Google on Tuesday rolled out WeatherNext 3, the latest version of its AI weather model, saying it can produce a global forecast every hour from recent satellite observations. The shift marks a departure from models trained chiefly on historical data: WeatherNext 3 is designed to ingest live signals, providing a picture up to five times sharper than its predecessor in some measures.

The practical target is fast-moving rain and snow. WeatherNext 2 issued forecasts every six hours on a 25-kilometre grid; the new system can map variables such as temperature and moisture at resolutions of up to five kilometres. Google says that can lift precipitation accuracy by as much as 50% for forecasts made at least a day ahead, with the biggest gains expected in regions outside the US and Europe that have fewer rain gauges.

Samier Merchant, a Google Research engineer, said the advance was about reaching beyond the data traditionally used to train global AI models: “We’re able to leverage fresher and richer observational data sets.”

Google is also aiming WeatherNext 3 at energy planning, including wind-speed forecasts at turbine height, as power systems absorb rising electricity demand. The model is now being folded into Search, Maps, Gemini and other Google products, while the company says it has worked with the US National Hurricane Center and agencies in Asia.

Still, Google’s faster AI forecasts do not settle weather prediction on their own. Traditional models simulate atmospheric physics on supercomputers, and weather agencies continue to compare multiple projections before issuing warnings; WeatherNext 3 itself is trained partly on physics-based-model data. The new tool is therefore being presented less as a replacement for forecasters than as a sharper, quicker input into their decisions.

Continue reading https://foxvector.com/stories/01a06934-d203-35e3-726f-04de80339ff0