Google DeepMind and Google Research have introduced WeatherNext 3, described by the company as its most advanced and accurate global weather model to date, with independent live evaluations by Brightband also ranking it as the most accurate global weather model currently available. Unlike previous AI weather systems that relied mainly on delayed numerical weather prediction data carrying a six hour lag, WeatherNext 3 learns directly from real-time observations, including live geostationary satellite mosaics and weather station data, allowing it to produce a new global forecast every hour rather than every six hours as with its predecessor, WeatherNext 2. The model can provide predictions at resolutions as fine as 5 kilometers for key surface variables like temperature and moisture, roughly five times sharper than WeatherNext 2’s 25-kilometer grid.
The model’s architecture, described by Google as a Functional Generative Network mesh transformer, ingests both live satellite data and traditional historical analysis to output dense gridded fields, discrete cyclone tracks, and station level sparse coordinate predictions natively. This combination allows WeatherNext 3 to maintain physical consistency from broad global wind patterns down to local topography, an area where earlier models struggled to capture variation over just a few kilometers, particularly for communities near coastlines, valleys, or mountain ranges. Google specifically highlighted that this breakthrough could prove especially valuable for regions across Latin America, Africa, and Asia-Pacific that have historically been underserved by high resolution forecasting due to the computing costs of traditional regional models, extending localized forecasting to areas that previously relied on coarser global estimates.
On precipitation specifically, an area where global weather models have historically struggled due to the fast moving, small scale nature of rain and snow systems, WeatherNext 3 was trained on two high quality data sources: NASA’s satellite based Integrated Multi-satellite Retrievals for GPM (IMERG) and Google’s own global precipitation reanalysis based on satellite radar. The result is a Continuous Ranked Probability Score improvement of up to 60 percent against IMERG, 30 percent against MRMS, and 10 percent against rain gauge measurements for early lead times, translating into sharper, more accurate rain and snow predictions rather than the blurry estimates that have plagued earlier AI weather systems. Beyond precipitation, the model introduces forecasts specifically engineered for renewable energy production, including 100-meter wind speed predictions roughly at turbine height alongside high resolution cloud cover and solar radiation estimates, giving grid operators and renewable energy developers more precise data for matching clean energy output with consumer demand.
WeatherNext 3 began powering weather experiences within Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine starting September 3, with Google saying the update delivers up to 50 percent more accurate precipitation forecasts when planning a day or more ahead, with the greatest improvements concentrated in regions where forecasts have historically been less reliable. High resolution forecast data updated hourly is also being made available for researchers, developers, and businesses to query directly through BigQuery and Earth Engine, or bulk download from Google Cloud Storage, without requiring any model setup on their end. Google noted that atmospheric prediction will always retain a degree of unpredictability, but framed WeatherNext 3’s approach of training directly on real-world observations rather than relying solely on traditional modeling constraints as bringing forecasting closer to matching what is actually happening on the ground in real time.
Follow the SPIN IDG WhatsApp Channel for updates across the Smart Pakistan Insights Network covering all of Pakistan’s technology ecosystem.
