Google DeepMind's WeatherNext 3 Outperforms Government Forecasts

Google DeepMind and Google Research have released WeatherNext 3, a deep-learning AI model designed to redefine global weather forecasting. The system outperforms traditional government forecasts and competing AI models in both accuracy and resolution, marking a shift toward data-driven meteorology.

The model achieves a spatial resolution of 5km for key surface variables, including temperature and moisture. This is a significant leap over previous AI models, which typically operated at resolutions of 15-25 square kilometers. To achieve this precision, WeatherNext 3 is 2.4 times larger in terms of parameters than its predecessor, WeatherNext 2.

Beyond resolution, the model improves the frequency of updates. By ingesting real-time geostationary satellite data, WeatherNext 3 provides hourly forecasts, moving past the industry-standard six-hour prediction interval. In performance testing on Operational WeatherBench, the model beat forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF), the US National Weather Service, and AI-driven models from Nvidia and Microsoft. Regarding precipitation, evaluations against baselines showed a Continuous Ranked Probability Score (CRPS) improvement of up to 60% against IMERG satellite-based precipitation data.

The shift to data-driven physics

Traditional weather forecasting has long relied on massive government supercomputers to solve complex physics equations. However, the landscape changed in 2018 when the ECMWF released vast amounts of historical weather data, enabling researchers to train deep-learning models. These AI systems identify patterns in data rather than relying solely on mathematical physics.

Ferran Alet, a staff research scientist manager at DeepMind, noted that machine learning is uniquely suited for this task because it targets the problem of approximating "noisy physics from incomplete information and finite compute."

Industry and consumer impact

The transition to higher-resolution, more frequent AI forecasts has tangible real-world implications. In developing countries, more precise weather data can significantly improve crop yields for agriculture. For the energy sector, better predictions of cloud cover, rain, and wind make renewable energy projects more dependable and easier to integrate into the grid.

For the general public, Google is integrating these capabilities directly into consumer products. The model will power weather data in Google Search, Maps, and Gemini, while also being made available to enterprise users via Google Cloud. Samier Merchant, a senior staff engineer at Google, stated that this represents the first time these core variables will directly power a wide array of Google products.

What to watch

While Google positions WeatherNext 3 as a breakthrough in incorporating raw observations for high-resolution global forecasts, the claim is not without dispute. The startup WindBorne has challenged this assertion, citing its own WeatherMesh 6 model. As AI continues to encroach on traditional meteorology, the industry will likely see further competition between big tech firms and specialized startups to determine which architecture provides the most reliable early warnings for extreme weather events.