By: Coco Xu

GraphCast, an A.I. model, just made a better hurricane forecast in minutes than what a group of meteorologists and a room of supercomputers managed to create in hours.

On July 4, 2024, members of the European Center for Medium-Range Weather Forecasts assembled data from international planes, buoys, and satellites on Hurricane Beryl, a Category 5 storm that was brewing in the Caribbean. Supercomputers then analyzed the data and concluded that the hurricane would most likely make landfall in Mexico.

The same day, the experts consulted the A.I. model, GraphCast, which predicted that the hurricane would land in Texas. As it turns out, the A.I. was right: On July 8, Hurricane Beryl surged onto Texan soil, killing 36 unsuspecting citizens and leaving millions more with no electricity.

“This is a really exciting step,” said Matthew Chantry, an A.I. expert at the European Center for Medium-Range Weather Forecasts. The A.I., which was created in London by DeepMind, a Google company, managed to surpass the agency’s best forecasting model in accuracy in over 90 percent of test cases.

Currently, creating a good forecast is far from easy. Scientists must carefully analyze a dynamic whirlwind (no pun intended) of observations detailing air temperature, wind speed, the locations of clouds, and more. Using the compiled data, modelers then create a digital model of the Earth. Finally, the data is fed to supercomputers, which generate predictions in a series of calculations that would have given the wittiest mathematician a headache. This whole process can take up to several hours.

To make matters worse, the Earth’s atmospheric status is extremely unpredictable. A prediction that was accurate a second ago could be the farthest thing from reality a minute later. This is what leads to the annoyingly familiar frustration of walking into a rainstorm in a pair of shorts.

That’s where A.I. comes in. As seen in the Hurricane Beryl case, A.I. can produce weather forecasts much more accurate than the most advanced supercomputers. That has to do with machine learning, the method it uses to learn to form predictions. It involves scanning a set of provided data for similarities, which enables A.I. to notice patterns that humans cannot. That, in turn, makes it much more perceptive of the Earth’s constantly changing weather patterns.

But accuracy is not A.I.’s only advantage over traditional forecasting methods. Its usage can also save scientists hours of time and effort. The best models can finish in seconds what a supercomputer can complete in hours. Even better, A.I. consumes much less space than supercomputers, as it can be run on a simple desktop and sometimes even a laptop.

But if you clicked on this article expecting it to rant about how meteorology could be yet another sector in which A.I. could replace humans in, you’re in luck, because Dr. Molina of the University of Maryland says that A.I. usage will most likely be complemented with the work of experts because neither is completely flawless.

“All models are wrong to some extent,” she said, adding that the A.I. “might get the hurricane track right but what about rain, maximum winds and storm surge?” The human role is still very much present, as a trained scientist is still needed to consider the more specific aspects of weather forecasting. The future of meteorology is counting on the next generation of aspiring scientists – and artificial intelligence – to lead it.