The model combines the latest developments in meteorology with deep learning techniques. Google said Weather Next 3 will begin feeding into weather information available through Search, Google Maps and Gemini. It will also be made available to users and researchers through Google’s cloud platform.
Weather Next 3 has already emerged as the most accurate model among leading contenders tested on the Operational Weather Bench. Developed by startup Brightband, the utility is designed to compare AI-based weather forecasts using metrics including temperature, wind speeds and humidity.
In testing, Weather Next 3 outperformed other deep learning models developed by Google, Microsoft, Nvidia and the European Centre for Medium-Range Weather Forecasts (ECMWF). It also surpassed traditional forecasts produced by the US National Weather Service and the ECMWF.
Most weather forecasts currently rely on government-owned supercomputers that process mathematical equations designed to represent the physics of weather. Although these systems are accurate, they can be slow and cumbersome.
The release of more than half a century of weather data by the ECMWF enabled researchers to develop training models capable of producing forecasts much faster while maintaining accuracy comparable to government forecasting systems.
Model developers have continued addressing key weaknesses in AI-based forecasting to create systems that can be trusted. However, even these models typically forecast across areas measuring between 15 and 25 square kilometres, are not always effective at predicting rain, and continue to depend on formatted datasets produced by government agencies.
Weather Next 3 improves on some of these limitations by providing forecasts at a resolution of 5 km. Its evaluations for rain have improved by 60% compared with Weather Next 2, while the model can now generate hourly forecasts instead of the standard six-hour forecasting interval.
Weather Next 3 is also a larger model, with 2.4 times more parameters than its predecessors. It tailors the targets for decoder heads to produce more useful answers.
Because the model can ingest weather satellite data in real time on an hourly basis, it is capable of producing forecasts more frequently. Beyond large language models, the transformer revolution in meteorology has also played an important role, with these AI models already being used by European and US weather agencies.
As AI-based forecasting tools advance, DeepMind researcher Alet said higher-resolution forecasts of wind, rain and cloud cover could help make renewable energy projects more dependable. Weather Next 3 therefore represents another step towards faster and more detailed weather forecasting.
About The Author
.jpg)
.jpg)
Comment List