While most technology has advanced faster than anyone could have anticipated, one area is still particularly challenging: weather forecasting.
And with good reason. The technology for this science has improved significantly over the years, but weather forecasting remains very hard due to seemingly endless variables.
Check out any local forecast and you’ll see just 10 days ahead. But what if these forecasts could see a month or even several months in advance? What if we could project how a community’s or region’s climate will change years, or even decades, from now? What if natural disasters such as forest fires and tornadoes could be spotted before they start, giving people more time to act?
These are the incredible possibilities available through AI.
“We want to use the largest class of AI foundation models to help initialize our largest climate prediction models so that we can save a lot of computational time,” said Peter Thornton, director of the Climate Change Science Institute at ORNL.
By feeding historical weather data with current climate factors into foundation models, ORNL researchers have scaled their AI foundation model with 100 billion parameters, or connections, on ORNL’s Frontier supercomputer using 24,000 graphics processing units (the preferred computer chips for training AI models). It’s through this capability that scientists can predict potential weather events and advise on how to protect homes and save lives.
“We typically have to run these models in a time frame of hundreds or even thousands of years to get them to a point where they’re ready to launch prediction experiments within the model,” said Thornton. “That’s very computationally expensive, and we think this largest class of AI model is the perfect tool for jumping to that starting point quickly instead of having to grind through a lot of numeric calculation.”