In brief
- WeatherNext 2 generates hundreds of global forecasts in under a minute, enabling more frequent scenario updates than conventional models.
- Google is already using the system across Search, Gemini, Pixel Weather, and Maps, with broader rollout planned.
- A new modeling approach, Functional Generative Networks, boosted accuracy on key measures, including extreme wind and cyclone tracking.
Google DeepMind introduced a new AI-powered weather-forecasting system on Monday, capable of generating global weather predictions eight times faster than traditional tools, it said.
Dubbed WeatherNext 2, the system is being positioned as a tool to help agencies prepare for severe conditions more quickly, as the world continues to grapple with frequent natural disasters spurred by an increasingly warming climate.
To do this, it generates hundreds of possible scenarios from a single starting point, each computed in under a minute on a single Tensor Processing Unit, a specialized chip developed by Google to accelerate machine learning and AI workloads.
“We rely on accurate weather predictions for critical decisions-from supply chains to energy grids to crop planning,” Google DeepMind research scientist Peter Battaglia wrote on X. “AI is transforming how we forecast weather.”
Deployment across Google products
WeatherNext 2 forecast is already running in Search, Gemini, Pixel Weather, and the Google Maps Weather API, with broader support coming at a later date.
“We’re working with the Google teams to integrate WeatherNext into our forecasting system,” WeatherNext 2 product manager Akib Uddin said in a statement. “Whether you’re on search, Android, or Google Maps, weather affects everyone, and so by making better weather predictions, we’re able to help everyone.”
Conventional models can take hours, limiting how often scenarios can be refreshed, DeepMind said. By using advanced AI, WeatherNext 2 outperformed its earlier operational model, WeatherNext Gen, the company claims.
“It’s about eight times faster than the previous probabilistic model that we released last year, and in terms of resolution, it is six times greater,” Battaglia said in a statement. “So instead of making six-hour steps, it takes one-hour steps. It outperforms the previous weather next gen on 99.9% of the variables that we tested.”
In practical terms, that means the new system produced more accurate forecasts of temperature, wind, humidity, and pressure almost everywhere and at nearly every point in the 15-day window.
DeepMind attributed the gains to a new modeling approach described in a June research paper on Functional Generative Networks, or FGN, which changes how the system represents uncertainty and generates forecast variations.
A new modeling approach
FGN is trained only on single-variable forecasts, or “marginals,” such as temperature, wind, or humidity at a specific location, according to Google.
Despite this, the model learns how those variables interact, allowing it to predict broader, interconnected patterns, such as regional heat events and cyclone behavior.
Google said FGN matched GenCast on extreme two-meter temperature forecasts and exceeded it on extreme ten-meter wind forecasts, depending on the variable.
The model also showed stronger calibration across lead times and better performance when forecasts were evaluated over larger regions rather than individual points.
Using the Continuous Ranked Probability Score—a standard accuracy metric that checks how closely a model’s full range of predicted outcomes matches what actually occurred—the paper reports average improvements of 8.7% for average-pooled CRPS and 7.5% for max-pooled CRPS compared with GenCast.
Cyclone forecasting performance
FGN also improved tropical cyclone forecasts.
Compared with historical tracks from the International Best Track Archive for Climate Stewardship, the ensemble-mean predictions reduced position errors by about 24 hours of lead time between three- and five-day forecasts.
A version of FGN run at 12-hour timesteps showed higher error than the six-hour version but still outperformed GenCast at lead times beyond two days.
Track-probability forecasts showed higher Relative Economic Value across most cost-loss ratios and lead times.
DeepMind said experimental cyclone-prediction tools built with this technology have been shared with weather agencies.
“You get more accurate forecasts, and you get them faster, and that helps everyone make the right decisions, especially as we start seeing more and more extreme weather,” Uddin said. “I think there’s a whole spectrum of applications for better weather forecasting.”
Generally Intelligent Newsletter
A weekly AI journey narrated by Gen, a generative AI model.
