Key Takeaways
- Google DeepMind and Research have launched WeatherNext 3, the most advanced global weather model.
- The new model can produce hourly forecasts with resolutions as detailed as 5 kilometers.
- WeatherNext 3 uses live satellite and weather station data, improving accuracy in rapidly changing conditions.
Google has unveiled WeatherNext 3, its latest and most advanced global weather model, which promises significantly more accurate and detailed forecasts. Unlike previous AI weather systems that relied on delayed numerical weather prediction data, WeatherNext 3 can learn directly from real-time observations, including satellite and weather station data.
The model generates a new global forecast every hour and can provide predictions at resolutions as detailed as 5 kilometers. Key surface variables such as temperature and moisture can be forecast at 5km resolution, while other surface variables use 10km resolution. Atmospheric variables such as wind speed are forecast at 25km resolution.
Google claims that WeatherNext 3 provides a global weather picture roughly five times sharper than its predecessor, WeatherNext 2, which produced forecasts on a 25km grid every six hours. The system uses a Functional Generative Network mesh transformer that combines hourly geostationary satellite mosaics with traditional historical analysis, allowing it to produce dense weather fields, cyclone tracks, and forecasts for specific weather station locations.
One of the biggest changes in WeatherNext 3 is the data used to generate forecasts. Most AI weather models, including WeatherNext 2, train on information produced by numerical weather prediction models. These physics-based systems rely on supercomputers and can carry a data delay of around six hours. Google says this delay can create problems when forecasting quickly changing conditions such as rain and surface temperatures.
WeatherNext 3 instead uses a continuously updated mosaic of global geostationary satellite observations, allowing it to generate a new forecast every hour using the latest available satellite data. Google says the shorter update cycle can provide earlier and more detailed information when storms, fronts, or precipitation systems develop quickly.
The model also trains directly on sparse weather station observations, allowing it to better capture local differences caused by geography, including coastlines, valleys, and mountain ranges. Google says this could be particularly useful in parts of Latin America, Africa, and Asia-Pacific, where high-resolution regional forecasting has traditionally been limited by the high computing costs of conventional models.
Google has also focused on improving precipitation forecasting, an area where both traditional and AI-based global weather models have struggled. WeatherNext 3 trains on NASA’s Integrated Multi-satellite Retrievals for GPM, or IMERG, alongside Google’s own global precipitation reanalysis based on satellite radar data. According to Google’s evaluations, the model delivered a Continuous Ranked Probability Score improvement of up to 60% against IMERG in medium-range global forecasts. It also improved by up to 30% against MRMS and 10% against rain gauge measurements at early forecast lead times.
WeatherNext 3 can also reproduce sharper boundaries around precipitation systems instead of producing the more blurred estimates seen in previous models. This improvement is crucial for better planning and response to weather events, particularly in regions where accurate and timely weather information is critical.
