This opinion article was written by GWSC Human-Environmental Analyst Penny Beames.
Scientific American published a story in early January on the use of AI in weather forecasting and, as one might imagine, it’s a topic of great interest to us.
Artificial intelligence is making waves in a number of fields, and atmospheric science is among them. AI forecast models like GraphCast and MetNet have proven they can successfully predict weather phenomena faster than traditional models, and they require less raw computing power to do so.
While speed and computer power are important facets of weather prediction, accuracy is paramount. That’s where AI models falter in a few crucial ways.
AI models rely on the statistical analysis of historic training data to predict the future rather than a deep understanding of atmospheric physics. As such, they excel at predicting events they’ve seen before. But rare, extreme events are becoming more commonplace with climate change.
Kim Wood is an associate professor of atmospheric science and hydrology at the University of Arizona. She told Scientific American that AI models struggle with the kinds of rare events “that can change people’s lives forever.”
Moreover, the models require high-quality training data if they’re going to nail the basics.
GWSC’s Director, Mike Gremillion, is cautiously optimistic about the use of AI in forecasting. He was a weather forecaster for the U.S. Department of Defense for almost 30 years, so he is keenly aware of the importance of good data.
“A model’s output is only as good as the data that feeds it,” he said. “We’ve made some incredible advances over the years in how we measure weather conditions and interpolate data, but there’s still a long way to go.”
Thankfully, atmospheric scientists all over the world continue to improve how we collect, distribute, and interpret global weather data. Those kinds of improvements can contribute to both traditional and AI models.
Will AI models completely replace traditional models and teams of human forecasters? It’s unlikely. Human forecasters are still better equipped to interpret model results through the lens of their own regional or topical expertise.
But they are very likely to become powerful tools that accompany and even progress physics-based forecasting.