Google DeepMind and Google Research presented a new deep learning model, WeatherNext 3. It became the first global meteorological AI system capable of generating detailed hourly weather forecasts and working directly with primary satellite data in real time.
Google DeepMind and Google Research presented a new deep learning model, WeatherNext 3. It became the first global meteorological AI system capable of generating detailed hourly weather forecasts and working directly with primary satellite data in real time.
As TechCrunch reports, citing the developers' announcement, the new development has already shown the best results in the Operational WeatherBench test from startup Brightband. The model surpassed in accuracy calculations from Microsoft, Nvidia, the European Center for Medium-Range Weather Forecasts (ECMWF), and the US National Weather Service.
According to Google senior engineer Samier Merchant, this is the first time that key meteorological variables from an AI model will directly underpin the company’s mainstream products — Search, Google Maps, and Gemini Assistant. The model is also available to businesses through BigQuery, Earth Engine, Google Maps Platform, and Google Cloud Storage.
The main advantage of WeatherNext 3 is the level of detail and speed of data updates. The system architecture is an ensemble model, which contains 2.4 times more parameters than its predecessor WeatherNext 2. The model is trained to process raw satellite images directly, bypassing the pre-processing stage by supercomputers. This has allowed to increase the accuracy of precipitation predictions by 60% compared to the previous version and to issue forecasts every hour, instead of every six hours, as was previously the case.
The resolution of the calculations depends on the type of parameters: surface data at weather stations (in particular temperature and humidity) are predicted with an accuracy of up to 5 km, while other indicators (for example, wind) are predicted with a resolution of 10 km.
In addition to consumer services, the developers focused WeatherNext 3 on industrial applications, in particular for renewable energy. The model calculates the level of solar radiation and cloudiness, which helps operators of solar and wind power plants more efficiently manage generation. The technology has already been tested by the US National Hurricane Center to determine the trajectory and predict the landfall of Hurricane Melissa in Jamaica.
«The idea behind many AI applications is to perform tasks as comprehensively as possible, from start to finish. Adding the ability for the model to now also predict what the weather station will record each hour brings the forecasting task closer to its essence,» said Daniel Rothenberg, an atmospheric scientist at Brightband.
Experts note that the introduction of such transformers into meteorology allows to significantly reduce the cost of calculations compared to traditional supercomputer mathematical models. This makes accurate forecasts available to regions that do not have the resources for their own supercomputers, and also optimizes the work of agriculture and the energy sector.
