After a hurricane passes, scientists routinely analyze the assorted computer models used to predict its path and power and crown a victor. This year, a surprising new contender has emerged — a forecast model generated by artificial intelligence.
How is it faring? Well, the single best-performing model for last month’s Hurricane Erin was Google DeepMind, a relative AI newcomer in the storm prediction field.
Yes, it’s a small sample size. But although the technology can be over-hyped in some places, it already holds a great deal of promise in the notoriously fickle field of weather forecasting. Some experts say their weather and storm predictions could be — and may already be — faster, more accurate, cheaper and less energy intensive to produce than traditional models.
The National Hurricane Center has already begun to consult AI. This year, forecasters have begun tapping into the DeepMind model — as well as using a new AI-powered tool developed at the University of Miami — to construct their forecasts.
Experts stress there are lots of cautions. For now, these models are simply the shiniest new tool in the toolbox. They’re by no means a replacement for the traditional models, much less the humans that analyze them to make predictions. And like every computer prediction model, including the European and American models, they have their failings and drawbacks.
“No tool, no matter how great it is, is perfect, especially when it comes to weather. I’m confident they’re going to help us, but we’re still in the early stages,” said Wallace Hogsett, science operations officer at NHC, which is based at Florida International University’s campus in West Miami-Dade. “Our hope is that we’ll be more effective carpenters, so to speak, by having all these tools.”
“We might need to buy a bigger toolbox,” he joked.
Chat, is a hurricane coming my way?
AI weather prediction doesn’t mean forecasters are simply asking ChatGPT about what a storm might do next.
Like “large language models” that power popular AI chat apps like Copilot or Claude, these weather models use machine learning to pick up patterns. Except, instead of predicting what word is most likely to come next in a sentence replying to a user’s question, they’re predicting what global weather patterns might do next.
Traditional weather models are run on massive supercomputers that run millions to billions of mathematical equations to simulate the physics that make up the atmosphere — the results of which are predictions about how the world’s weather may shift on an hour-by-hour basis.
These supercomputers spit out estimates for how various changes in weather patterns will affect others around the world, like how a cold front sweeping off the East Coast of the U.S. will push Hurricane Erin on a curving pattern out to sea instead of letting the storm beeline straight into the Caribbean.
AI-based models work a little differently. For one, they don’t “understand” physics the way the traditional models do. Instead, they’re doing what machine learning is best at: pattern matching.
The Wednesday afternoon run of Google DeepMind’s hurricane-specific AI model for a tropical wave in the eastern Atlantic.
These models are based on 40 years’ worth of detailed descriptions of weather around the world and trained to estimate what will happen to current weather features (like a tropical storm) in light of the decades of past observations it has digested.
They’re also less energy-intensive and faster to use; they can be run on laptops instead of hulking supercomputers. Traditional models use “10,000 to 100,000 times more time and energy,” an expert told the University of California.
The results from these new AI models are, perhaps shockingly, pretty accurate.
Google DeepMind released its storm-specific AI model in June. Last month, it got its first big test with Hurricane Erin, which hit Category 5 over open water but largely avoided land-based impacts in the U.S. and Caribbean by executing a hard right turn when it approached the islands.
For the first three day’s of Erin forecast, DeepMind performed better than any other model, including the well-known European and American models. Throughout the five-day forecast period, it was among the best, data from former NHC Branch Chief James Franklin showed.
This image shows the observed track of Hurricane Erin in black, the European weather model predicted track in orange and the DeepMind forecast track in blue. For the first three days, DeepMind’s was the most accurate, but it kept up with the Euro for the rest of the forecast.
“To see that it did very favorably with that is encouraging. It’s a big leap forward right there and instantly adds to the credibility,” said Houston-based meteorologist Matt Lanza, co-founder of a popular blog that tracks tropical weather systems called the Eyewall.
Lanza said he’s been very impressed at the growing skill of AI models and recently added DeepMind to his rotation of five to seven models he checks daily for his own forecasts. However, he cautioned, just because DeepMind performed best overall for Erin doesn’t necessarily mean forecasters should rely entirely on this model going forward.
“That’s the average over the life cycle of the storm. It doesn’t mean it did great at every moment, just that it did great overall,” he said.
That’s how most models are for most storms, Lanza said. Different models perform better for different storms, and there’s no one “best” model that gets it right every time, so forecasters consult several to see what might happen.
“Quite frankly, most of the other models did well, too,” he said. “What it tells you is that these AI models are able to create forecasts that are as good as anything else right now.”
DeepMind’s success with Hurricane Erin, however, also points to another trend forecasters are noticing. These models are getting better at tracking the intensity of a storm, or how strong it is. Previously, they were doing a pretty good job at estimating the storm’s track but weren’t doing as well with intensity.
They’re also less prone to that “windshield wiper” effect, where the spaghetti models for a future storm will whip back and forth over an area as the storm closes in. That’s because the models are optimized to avoid errors and not take risks.
However, that “smoothing effect” means AI models can miss the little wobbles storms make near landfall — small but crucial jogs that can have the worst impacts 50 miles or more this way or that. And because they’re only trained on the last 40 years of data, newly published research has found they’re not great so far at predicting extreme events, the kinds that are set to become more common as climate change messes with the atmosphere.
Hogsett, with the hurricane center, said unlike physics based models that are built on solved math equations, it’s hard to check the work of an AI model and see if the assumptions it made were based on sound logic.
“Because of how these models work, it’s hard to come to the conclusion of how they came to that conclusion,” Hogsett said. “Because they’re new, we still just need to test and learn. We really need a full season of data.”
Hogsett and other forecasters also caution that these new AI-powered tools are not useful without humans to review and analyze them, much less issue timely and accurate warnings ahead of future storms.
“Our forecasters, I’m confident, will always have a role to play because these models are not perfect,” he said. “When it comes to communicating risk and helping people stay safe in the face of hurricane hazards, there will always be a role for a human.”
Like most of the federal government, the National Weather Service had its staff slashed by the Trump administration earlier this year, although the agency has recently begun re-hiring for those empty positions, including some in Florida.
Tracking tropical waves — but better
Beyond the world of AI models that show the weather for the entire globe, there’s an industry of people using AI to create more specialized forecasts.
A homegrown one from University of Miami researchers has also recently been adopted at NHC. It helps scientists spot tropical waves, the small, disorganized spinning clumps of rainclouds that sometimes grow into tropical depressions, storms and hurricanes.
“They are important seedlings for hurricane formation,” said Will Downs, a doctoral student at the University of Miami Rosenstiel School of Marine, Atmospheric, and Earth Science and lead author on the study published about his team’s new wave-tracking tool.
Satellite imagery of three tropical waves identified on June 28th, 2024. The rightmost tropical wave became Tropical Storm Beryl a few hours later, while the leftmost tropical wave became Tropical Storm Chris on June 30th.
Downs and his team used the written descriptions hurricane center forecasters have compiled for every tropical wave they’ve tracked for the last 20 years, fed it into a model and asked it to identify anything similar it spotted in the Atlantic or Pacific.
The results, Downs said, are better than the algorithm the NHC was previously using to spot tropical waves.
“It does pretty darn well at detecting these systems,” he said. “It’s particularly better in the Caribbean at detecting tropical waves, especially weaker ones.”
That’s impressive, because the wind currents in the Caribbean can be pretty turbulent, and picking out a small whorl of could-be storm action from a broader landscape of chaotic spinning isn’t an easy task.
Downs said he thinks the reason this tool is performing better than previous physics-based models at spotting tropical waves is that programming a physics-based model with the lengthy explanations — and multiple exceptions to the rule — of what a tropical wave is can be complicated.
“It’s ‘I know it when I see it,’ but that’s very hard to program explicitly,” he said.
For a machine learning model, it’s much easier to show it a bunch of examples of previous waves and ask it to look for something similar.
However, the tool does have its flaws. For one, Downs said, it’s pretty good at spotting a young, undeveloped tropical wave. But when a wave develops and starts to look more like a tropical storm, the tool suddenly can’t recognize it.
It’s not just hurricanes where AI modeling can be useful. Private companies have leapt to join the race across the country, too.
Andrew Brady, a Georgia-based meteorologist and AI engineer, just launched a product on Thursday that allows users to better pinpoint tornadoes and lightning using machine-learning trained on years of past weather data, called StormNet. His company was snapped up by Open Snow, a firm that used machine learning models to best predict snowfall and help skiiers and snowboarders pick the perfect day to shred.
His goal is to develop a product that can give as much advanced warning of an imminent tornado or lightning strike as possible.
“For tornadoes, for example, we don’t fully understand a lot of elements of tornado genesis, but machine learning models are able to examine the entire atmosphere and pick up on patterns based on elements we don’t even know about,” he said.
However, he acknowledged that, as exciting as recent developments in the AI weather space have been, they’re running into a few ceilings. For one, there’s only 40 years’ worth of high-quality data to train on. Anything older than that isn’t as useful to train AI weather models.
“We don’t have higher resolution data going back that far, like we have the lower resolution data,” he said. “It’s definitely a massive problem.”
But, Brady said, scientists are already working on the next wave of models — a hybrid between the physics-based models and the machine-learning ones — that are already showing a lot of promise.
Who knows, he said. Maybe in 20 to 30 years a 10-day storm forecast could be as accurate as the five day cone, or people could get as much as a day’s warning before a tornado strike.
“The possibilities really are kind of tantalizing,” Lanza said.