Hurricane forecast: Gabrielle swirls in background; AI steps in forecasting spotlight

Google's DeepMind GenCast AI weather prediction model reportedly outperformed traditional weather forecasting methods in a study.

For the first time in three weeks, there is a named storm in the Atlantic.

Tropical Storm Gabrielle has developed well east of the Lesser Antilles, angling west-northwest at 15 mph with sustained winds of about 50 mph. Gabrielle is no warrior princess, rather a strung-out mess dealing with shear and dry air from a nearby upper low. Expect it to stay feeble through Monday, then possibly strengthen to a hurricane as it trundles out-to-sea while posing no threat to land other than Bermuda.

Elsewhere, a couple of tropical waves crossing the eastern Atlantic over the next 6 to 10 days have marginal chances of development. As I said last week, however, any storms that form east of the Lesser Antilles in late September or October are highly unlikely to ever threaten to the continental U.S.

If one of these waves continues west without developing, it might be worth watching towards the end of September should it reach the western Caribbean. That's a long, long way off, though.

If Gabrielle does become a hurricane, it will do so almost a month later than the average formation date of the season's second hurricane. That means following the longest mid-season gap between named storms since 1992, this year has been quieter to-date than two out of three seasons over the past 75 years.

While the calm first half of the season wasn't a shock, I am pleasantly surprised by the Atlantic's continued ineptitude in the second half of September. Did I expect to be free to install a new washer/dryer or read all of Watership Down this hurricane season? No. But I'll take it. Everyone has to learn the truth about the Bun Wars and the heroic sacrifice of Mr. Bigwig sometime.

Putting the AI in hurricane forecasting

Speaking of the washer/dryer (bear with me), that new unit comes equipped with a something disappointingly not called "washing machine learning," but rather "AI OptiWash."

As the washer peers deeply into my dirty clothes and automatically assigns them bespoke rinse and spin cycles, I am reminded of both Arthur C. Clark's Third Law, which posits that sufficiently advanced technology is indistinguishable from magic, and also the fact that companies will slap "AI" on just about anything as a marketing gimmick.

While my ability to differentiate technology from magic in the realm of laundry is limited at best, meteorology is another domain in which artificial intelligence is making an impact and I do know something about that one.

While machine learning and neural network approaches to weather prediction have been a subject of research for the last 15 years, in 2025 these tools are starting to rival or even outperform traditional physics-based guidance.

How have AI models performed in hurricane forecasting? Early results are promising

This spring, the AI-version of the European Center's global model routinely produced slightly more skillful forecasts for Northern Hemisphere temperatures and jet stream patterns than their stalwart physics-based model, which in turn is a bit more accurate than other global physics-based models including the American GFS.

In June, Google's Deepmind division announced a partnership with the NHC to develop and validate hurricane-specific AI modeling.

While Hurricane Season 2025 has thankfully not afforded the opportunity to put AI forecasts for Atlantic tropical cyclones to the ultimate test, the Deepmind model produced the most accurate 1- to 3-day track and intensity forecasts for Hurricane Erin, besting even the NHC's official forecasts at those lead times.

NOAA's specialized physics-based HAFS hurricane models led the pack 4 and 5 days out for Erin. These results and the fact that the NHC is taking Deepmind output seriously indicate that AI model have something real to contribute to weather prediction, likely stemming from their novel mathematical approach to forecasting.

Traditional physics-based weather modeling uses simplified versions of the fundamental fluid dynamics equations that govern how air and water move and exchange energy to "solve" the gigantic calculus problem of the atmosphere with time.

This approach best approximates how the Earth's atmosphere actually works and has been successful in making the next week of weather reliably predictable, but errors creep in from those simplifications and the fuzzy snapshot of initial conditions. They also require a massive computational resources to run, even with those assumptions.

To understand how AI how can help in forecasting, let's consider Publix Subs

AI modeling is more of a pattern matching exercise, algebra versus calculus. To use an easier example, say that I want to forecast the type of Publix sub on sale each week: let's call our model Prediction Using Binomial Statistics of Underpriced Boarshead Sandwiches (PUBSUBS).

Rather than try to predict sandwich sales using the inscrutable laws of nature, PUBSUBS might be given the last 20 years' worth of Publix fliers and tasked to discover repeating patterns in the data. Have Ultimate, Italian, and Turkey been on sale the last three weeks?

Lucky you, it may well be time for a bargain on the estimable Chicken Tender sub. With enough processing power, the AI model may find patterns too subtle or complicated for a human to pick up on. Likewise, AI hurricane models can suss out the temperature, wind, or moisture content predictors most likely to capture how a storm will move and change.

There are a couple of limitations to this approach. The first is that the AI methods of forecasting are completely dependent on good information about the past and present.

Just as PUBSUBS needs to know the past record of sales to make predictions about future sandwich discounts, AI weather models require both a reliable dataset of historical atmospheric conditions to train the model and an initial snapshot of current weather to make a forecast. Both of those can only be provided by high-quality observational networks, as interpreted through those tricky yet fundamental atmospheric physics equations.

There's no escaping it: weather models need to know their calculus.

Secondly, AI models are only as good as their training sets. Let's say that in response to the Australian-Canadian fusion cuisine craze of the 2030s, Publix puts the new Poutine Vegemite Delight sub on sale every three weeks. Knowing only the sandwiches of 2005-2025, PUBSUBS will be unable to predict this unknown phenomenon, and will need to be retrained for a world in which poutine is routine.

Likewise, AI weather models shine in teasing out fine gradations of typical patterns, but struggle with rare or unprecedented events -- those with the greatest social and economic impacts. Much like large language models can go haywire when confronted with prompts outside of their training set -- such as "Should you eat rocks?" -- AI weather model output may be unreliable when we need it most. (Though their total confidence in making up stuff may be the way in which AI is actually most human of all.)

There's just no taking humans out of the equation

The weather world is awash in a sea of raw data for both the past and present, and it makes sense that a technology that can efficiently identify valuable relationships from history would earn a place in the forecaster's quiver.

It is also encouraging to see research efforts like Deepmind partner with seasoned experts, rather than "disrupt" something that isn't broken. The best hurricane forecast skill comes from ensembles that carefully integrate multiple models with human guidance and insight. The NHC is now ushering AI models into that rigorous framework, where they can do the most good.

The biggest risk of AI weather modeling may be the temptation to chase shiny objects. As the success of HAFS shows, physics-based modeling is still improving via increasing computer power and continued research into more accurate mathematical representations of storms; indeed, the two approaches should complement each other, rather than compete for resources.

As always, I am less afraid of artificial intelligence than natural stupidity. Keep watching the skies, Skynets, and subs.

Dr. Ryan Truchelut is chief meteorologist at WeatherTiger, a Tallahassee company providing forensic meteorology expert witness services and agricultural and hurricane forecasting subscriptions. Visit weathertiger.com to learn more. Email Truchelut at ryan@weathertiger.com.