This visualization showcases a six-day prediction of wind speed and sea level pressure, from the powerful HR-Stormer model at a 30-kilometer resolution.
AI is popping up in more and more contexts these days, including in weather forecasting. At Argonne National Laboratory, in a collaboration with UCLA, researchers are tapping a novel AI approach to potentially deliver more accurate forecasts at a fraction of the computational cost of traditional models.
For decades, weather forecasting has relied on complex equations, crunched by supercomputers, to model atmospheric thermodynamics and fluid dynamics. This new AI-driven approach, however, takes a different tack. It employs “foundation models” trained on “tokens” — small pieces of information that also are at the heart of large language models like OpenAI’s ChatGPT and Antrhopic’s Claude. But rather than breaking words into tokens, the novel weather-focused AI uses tokens that are patches of weather charts depicting atmospheric conditions like humidity, temperature, and wind speed.
Tokenizing weather data for AI models
“Instead of being interested in a text sequence, you’re looking at spatial-temporal data, which is represented in images,” explained Argonne computer scientist Sandeep Madireddy, in a press release. “When using these patches of images in the model, you have some notion of their relative positions and how they interact because of how they’re tokenized.”
The team has found that even low-resolution data can yield surprisingly accurate predictions. The conventional wisdom of weather forecasting has assumed that higher resolution is a must for better forecasts. But the tradeoff of using more precise physics models is computational cost, said Argonne atmospheric scientist Rao Kotamarthi, in a press release. Kotamarthi noted that the research team is finding that they can “actually get comparable results to existing high-resolution models even at coarse resolution.”
Potentially more accurate climate-change modeling
The implications are significant. Some AI models are already outperforming current methods for predictions beyond seven days, potentially extending the window for accurate forecasts. While the immediate focus is on weather, researchers see broader applications in climate modeling.
One challenge of modeling weather is the fact that historical data is losing its relevance as the climate changes over time as a result of the additional carbon in the atmosphere. “With the climate, we’ve gone from what had been a largely stationary state to a non-stationary state,” said Argonne environmental scientist Troy Arcomano, in a press release. The growing amounts of carbon are “also changing the Earth’s energy budget,” Arcomano said. ”It’s complicated to figure out numerically and we’re still looking for ways to use AI.”
Wielding supercomputing power
Argonne’s supercomputer, Aurora, which is one of the fastest in the world, will be instrumental in this endeavor, enabling researchers to train massive AI-based models at high resolutions. “We need an exascale machine to really be able to capture a fine-grained model with AI,” Kotamarthi stated. R&D World featured the Aurora as a top semiconductor breakthrough in the first half of 2024.
The potential benefits of this research extend across numerous sectors. Aviation, shipping, and other industries reliant on accurate weather forecasts stand to gain significantly from improved long-term predictions. The ability to make more informed decisions based on these forecasts could have far-reaching economic implications.
The study, funded by Argonne’s Leadership-Directed Research and Development Program, received the Best Paper Award at the “Tackling Climate Change with Machine Learning” workshop in Vienna.
Outstanding. It will enable earlier advance warning to the masses resulting in saved lives, reduce in damage and lost. More effort should be invest in this area to make it even better and allow for improve in earlier warning.