TL;DR — Key Takeaways
- NOAA is shifting major weather forecasting workloads from traditional supercomputers to commercial cloud infrastructure, with Google Cloud serving as the primary provider.
- The cloud transition is intended to give NOAA more scalable computing capacity, greater flexibility and faster access to resources as forecasting demands increase.
- Google DeepMind’s AI weather models are also being used to improve tropical cyclone forecasting, potentially giving forecasters more lead time on storm tracks and intensity.
Weather forecasters will be eyeing a cloud in a new way starting next year but it will be a digital one rather than one in the sky. The National Oceanic and Atmospheric Administration (NOAA) is ditching its reliance on supercomputers for weather forecasting to instead rely on Google Cloud Services. The time required to generate a forecast is expected to drop from hours to seconds.
The partnership between NOAA and Google already is paying dividends as researchers at Google DeepMind say its latest AI weather model adds an extra day to hurricane warnings. Today’s quickly-developing extreme weather appears to be more than the average supercomputer can handle when it comes to running large-scale weather simulations. NOAA currently uses two HPE Cray supercomputers in Dogwood, VA and Phoenix, AZ., with each operating at 14.5 petaflops.
NOAA says transferring its forecasting models to the cloud translates into greater flexibility and nimbleness while also directing computing power to where it is most needed. The public can expect earlier and more precise forecasts and warnings for extreme weather events. The cloud transition extends to 122 local weather forecast offices and National Weather Service (NWS) centers.
“We are charting a course to global leadership in cloud-based numerical weather prediction,” said Ken Graham, director of NOAA’s National Weather Service. “This is a significant achievement for NOAA; one that will ensure our critical modeling operations can seamlessly adapt to emerging science and strengthen our mission to protect life and property.”
That’s a “government-ese” acknowledgement that extreme weather events are making weather forecasting more challenging. Extra response time during extreme weather events is critical to avoid casualties. Google DeepMind says its WeatherNext AI model can now predict a hurricane’s (aka cyclone) path three days in advance, up from two. Google DeepMind compares the advance to a decade’s worth of meteorological progress. Google DeepMind and Google Research collaborated with the National Hurricane Center and other agencies to develop the AI program. The model was co-trained on two different sources, one for global weather dynamics and the other a historical database of nearly 5,000 storms, to model extreme weather. Google DeepMind is open sourcing the code and model weights so anyone can build upon them. DeepMind also is releasing a more compact version called WeatherNext 2-mini that can run on a single tensor processing unit (TPU), a Google chip specifically designed for AI applications. The compact version is described as being surprisingly robust despite its lower resolution.
NOAA is using Google DeepMind as the framework for its AI Global Forecast System, which the agency describes as the first suite of weather forecasting models driven by machine learning. The transition from supercomputers to the cloud is scheduled for completion by December 2027. Google Cloud H4D VMs, powered by AMD EPYC processors, are serving as the primary backbone for the cloud-based forecasting system.
