(TNS) – After a hurricane passes, scientists regularly analyze the various computer models used to predict their paths and power and crown winners. This year, an astounding new competitor has emerged – predictive models generated by artificial intelligence.
How is it operated? Well, last month's single best performance model for Hurricane Erin was Google Deepmind, a relative AI rookie in the Storm forecasting field.
Yes, it's a small sample size. But while the technology can be over-deferred in some places, there are already big promises in the infamous, whimsical weather forecasts. Some experts say their weather and storm predictions could be faster, more accurate, cheaper and more intensive to produce, and could already be.
The National Hurricane Centre has already begun consulting with AI. This year, predictors began tapping on the DeepMind model to build forecasts using a new AI-powered tool developed at the University of Miami.
Experts emphasize that there is a lot of attention. For now, these models are simply the newest tools that shine in the toolbox. They are by no means an alternative to traditional models, and far fewer people analyze them and make predictions. And, like all computer predictive models, including European and American models, they have failures and drawbacks.
“No matter how great it is, there's no perfect tool, especially when it comes to weather. I'm sure they're trying to help us, but we're still in the early stages,” said Wallace Hogsett, NHC's Science Operations Officer, based on the Florida International University campus in West Miami-Dade. “Our hope is to become, so to speak, a more effective carpenter by having all these tools.”
“You might need to buy a bigger toolbox,” he joked.
Chat, will hurricanes do my path?
AI weather forecasts do not mean that predictors are simply asking ChatGpt about what the storm will do next.
Like “large language models” that power popular AI chat apps like Copilot and Claude, these weather models use machine learning to pick up patterns. However, instead of predicting the words that are most likely to come next in a sentence that responds to a user's question, they are predicting what global weather patterns will do next.
Traditional weather models run on large supercomputers that run millions to billions of mathematical equations to simulate the physics that make up the atmosphere. The results are predictions of how the world's weather moves every hour.
These supercomputers spit out estimates of how the cold front off the US east coast straightens the Caribbean in a curved pattern, but how it will affect others around the world, like how it drives Hurricane Erin out into the ocean.
AI-based models work a little differently. One is not to “understand” physics like traditional models do. Instead, they do what machine learning is best: pattern matching.
These models are based on 40 years of detailed explanations from around the world and are trained to estimate what will happen to current weather functions (such as tropical storms) in light of observations from past decades that have been digested.
It is also not energy-intensive and is fast to use. Instead of hugging a supercomputer, you can run it on a laptop. The traditional model “uses 10,000-100,000 times more time and energy,” an expert told the University of California.
The results of these new AI models are probably shocking and quite accurate.
Google Deepmind released an AI model unique to Arashi in June. Last month I took my first major test at Hurricane Erin. Hurricane Erin hit Category 5 with open water, but barely avoided land impacts in the US and Caribbean by performing a hard right turn when approaching the island.
In the first three days of Erin's prediction, Deep Mind performed better than other models, including the famous European and American models. Throughout the five-day forecast period, it was one of the best data from former NHC branch chief James Franklin.
“It's encouraging to make sure it's become so positive. It's a huge leap right away and it's quickly becoming more reliable,” said Houston-based meteorologist Matt Lanza, who is the co-founder of a popular blog that tracks tropical weather systems called the Eye Wall.
Lanza said he was very impressed with the growing skills of AI models, and recently added a deep mindset to the rotation of five to seven models he checks daily for his own predictions. But he warned. Just because Deepmind performed the best overall for Erin doesn't necessarily mean that predictors should rely entirely on this model in the future.
“It's an average beyond the lifecycle of a storm, and that doesn't mean it's great at every moment.
That's the presence of most models in most storms, Lanza said. Different models perform better against different storms and there is no “best” model that will make it right every time, so predictors will consult some to see what happens.
“Frankly, most of the other models worked well,” he said. “What that says to you is that these AI models can make predictions as good as others for now.”
However, the success of deep minds in Hurricane Erin also points to another trend that predictors are aware of. These models have improved tracking the intensity of the storm, or how strong it is. Previously, they did a pretty good job estimating storm trucks, but they didn't do that either.
It also has the effect of a “windshield wiper” effect, and future Storm spaghetti models will move between areas as the storm closes. This is because the model is optimized to avoid errors and not take risks.
However, “smoothing effect” means that AI models may miss nearby wobbling storms. And they found that, as they are only trained on data from the last 40 years, newly published studies have not been very large so far to predict extreme events.
Unlike physics-based models that are built on resolved mathematical equations using the Hurricane Center, Hogsett is difficult to see the work of AI models and see if the assumptions it made are based on sound logic.
“It's difficult to reach a conclusion on how these models come to that conclusion because of how they work,” Hogsett said. “They're new so we still need to test and learn. We really need seasonal data.”
Hogsett and other predictors warn that these new AI-powered tools are not useful for humans to review and analyze them.
“Our predictors have a role that I should always play, with confidence, because these models are not perfect,” he said. “There's always a human role when it comes to communicating risks and helping people stay safe in the face of hurricane risks.”
Like most of the federal government, the National Weather Service cut staff by the Trump administration earlier this year, but recently began rehiring for empty positions, including some in Florida.
Tracking tropical waves – but better
Beyond the world of AI models showing weather around the world, there is an industry of people using AI to create more specialized forecasts.
Homemade researchers from the University of Miami researchers were also recently hired by the NHC. Scientists help to find tropical waves, small, turbulent spinning chunks of rain rows.
“It's a key seedling for the hurricanes,” said Will Downs, author of Atmospheric and Geosciences, Miami Rosenstiel University, of the published research on his team's new wave tracking tool.
Downs and his team have asked them to compile and supply models for all the tropical waves that Hurricane Center predictors have tracked over the past 20 years, and to identify similar ones discovered in the Atlantic or Pacific Oceans.
According to Downs, the results were superior to the algorithms that NHC used to find tropical waves previously.
“It works pretty well to detect these systems,” he said. “In the Caribbean, it's particularly good at detecting tropical waves, especially weak waves.”
It's impressive as Caribbean's winds are rather turbulent and choosing small swirls of storm action from a wider landscape of chaotic spinning is not an easy task.
Downs said he believes the reason the tool performs better than previous physics-based models in finding tropical waves is to program a physics-based model with multiple exceptions to the long explanations and rules that can make tropical waves complicated.
“It's 'I know it when I see it,' but it's very difficult to explicitly program,” he said.
With machine learning models, it is much easier to show you a lot of previous wave examples and ask them to look for something similar.
However, this tool is flawed. For one, it's pretty good to find the young, undeveloped tropical waves, Downs said. But once the waves develop and appear to be like a tropical storm, the tool is suddenly unrecognizable.
Hurricanes aren't the only things that AI modeling can help. Private companies are also jumping on to participate in races across the country.
Georgia-based meteorologist and AI engineer Andrew Brady just launched the product Thursday, allowing users to better identify tornadoes and lightning using machine learning training trained with past weather data called Stormnet. His company jumped to Open Snow, a company that best predicted snowfall and used machine learning models to help skiers and snowboarders choose the perfect day.
His goal is to develop products that can allow for advanced warnings of impending tornadoes or lightning strikes.
“We don't fully understand that in a tornado, for example, there aren't many Genesis elements in tornadoes, but machine learning models can look at the entire atmosphere and pick up patterns based on factors that you don't know,” he said.
But he admits that it is just as exciting as AI's recent development in weather spaces, but they've hit several ceilings. For one thing, there are only 40 years of high quality data to train. Something older than that doesn't really help train AI weather models.
“There's no data that is very high resolution, just like the resolution data is low,” he said. “That's definitely a big problem.”
But Brady said scientists are already working on the next wave of models: a hybrid of physics-based models and machine learning models.
Who knows, he said. Over 20-30 years, a 10-day storm forecast can be as accurate as a 5-day cone.
“The possibilities are really appetizing,” Lanza said.
© 2025 Miami Herald. Distribution Tribune Content Agency, LLC.
