[Note: Today’s newsletter is guest edited by John Oldshue. John is an Emmy-winning meteorologist who spent 14 years at Birmingham’s ABC 33/40 as a meteorologist and storm chaser. John is currently one of the founders of the Alabama Weather Network and has agreed to provide technical diligence for this newsletter edition.]

On Wednesday afternoon, with Hurricane Isaias in the Gulf and a hurricane watch up from Bay St. Louis to Indian Pass, the National Hurricane Center did something that wasn't possible two years ago. The forecaster nudged the official track a little east, "closer to the AI-based and corrected-consensus models." The forecaster gave a physical reason. Stronger wind shear near landfall favors a faster, more eastward path, so the AI solution "makes physical sense."

That one paragraph from Forecast Discussion 5 is where hurricane forecasting stands right now. The forecaster had AI guidance and physics guidance side by side, weighed both, and explained the reasoning. By the time you read this, the Isaias track will have changed. The method won't.

How the old way works

Every forecast starts with measurements from satellites, weather balloons, buoys, radar, and the Hurricane Hunter aircraft that were flying through Isaias on Wednesday. Those measurements are uneven and sometimes disagree, so a process called data assimilation blends them with the previous forecast into a best estimate of the atmosphere right now. Forecasters call that estimate the analysis.

Then the model cuts the atmosphere into a three-dimensional grid and solves the equations of fluid motion, heat, and moisture forward in time, one short step after another. The American GFS uses a grid about 13 km wide. The European model uses about 9 km. NOAA's hurricane model, HAFS, zooms a moving 2 km nest onto the storm so it can draw the eyewall. All of this takes a supercomputer and hours. The noon GFS run isn't ready until an hour after the NHC has already issued its forecast.

How the AI way works

An AI weather model skips the equations at forecast time. It's trained on decades of past weather, learns how pressure, wind, temperature, and moisture at one moment relate to the next, then applies those learned relationships step by step into the future. The expensive work happens once, during training. After that, NOAA's AI model finishes a 16-day forecast in about 40 minutes using 0.3 percent of the computing the GFS needs. Google's cyclone model runs 15 days in under a minute on a single chip, which is why Google can run 1,000 versions of a storm where a physics ensemble runs 31 or 50.

The training data comes from physics models. The NHC says AI models "are trained on reanalysis outputs from dynamical models." Reanalysis is a decades-long reconstruction of past weather, built by running data assimilation with a physics model. ECMWF says no observations are used directly in training its AI model (EC-AIFS), and every AI forecast starts from the ECMWF physics model's analysis of the current atmosphere. So the chain runs observations, then physics, then AI. Take out the physics step and the AI has nothing to learn from and nowhere to start.

The 2025 scoreboard

Last year was the first season the NHC used AI models in real-time operations. At 48 hours out, Google's cyclone model missed the storm center by an average of 41 nautical miles. The official NHC forecast missed by 49. The European physics model missed by 65 and the GFS by 79. Google's model was the only one that beat the human forecast, and it did so whether the forecast was made 12 hours ahead or three days ahead.

Three caveats are important regarding the metric above. At five days the NHC was back ahead of Google. Google's model missed the season's first two storms. And over 2023 to 2025 combined, averages of several models were still the best track guidance.

Intensity is where the physics shows through

Track is the easy case for AI. A hurricane goes where the large-scale steering currents push it, and those are the broad patterns a global model sees well. Intensity lives in the core, and the core is where coarse models go blind. Most AI models work on a 28 km grid, so the eye and eyewall fit inside one or two pixels. Training a model to minimize average error rewards smooth fields, which means it would rather draw a wide, weak storm than a sharp, intense one in slightly the wrong spot. And the reanalysis it learned from has storms that run too weak to begin with.

ECMWF put numbers on this in August. Its AI model ran 29 knots too weak on average in 2025, and near 70 knots off on rapidly intensifying storms. A second machine-learning step pulls those errors down to about 2 knots and 23 knots, but it's still a prototype. NOAA says the first version of its own AI model made intensity forecasts worse.

Google's model is the exception. It uses the same 28 km grid and matched the NHC on intensity last year, 12.3 knots of error at 48 hours against 13.1 for the human forecast and 15.5 for NOAA's hurricane model. On Hurricane Melissa, the NHC's post-storm report says the Google model "performed very well" on both intensity and rapid intensification, and for the first time ever the NHC forecast a Category 5 from a storm that was only a Category 1.

Two things set the Google model apart. It was trained on about 5,000 historical storms on top of the usual reanalysis, and it predicts the storm's track, peak wind, and size directly instead of leaving a separate tracker to find the storm in a blurry weather map. Google calls the rest "an open research question." Kate Musgrave, who evaluates AI models for the NHC, told NPR that intensity "is not captured well in the AI models" in general, and Google's did well because it added the hurricane data.

Storms nobody has seen

A physics model doesn't need to have seen a storm before. An AI model does. A University of Chicago team retrained NVIDIA's FourCastNet with every Category 3, 4, and 5 storm removed from its training data, then asked it to forecast a Category 5. It couldn't. A 2026 study in Science Advances found the European physics model beat three AI models on record-breaking heat, cold, and wind at nearly every lead time, even though the AI models did better on ordinary weather.

Those studies tested older AI models, not Google's, and Melissa points the other way. Whether AI can handle a storm outside its training record is an open question. The physics argument is that the equations don't care what year it is.

Not a contest

The agencies aren't picking a side. NOAA's hybrid ensemble pools 31 physics runs with 31 AI runs, and NOAA says the blend beats both parents. Google's model is now one of the models the NHC averages for track, intensity, and wind size. ECMWF is testing a method that steers its physics ensemble toward its AI ensemble, to get the better track without losing the realistic core.

Two things don’t use any AI at the NWS. The forecast cone isn't drawn by any model. It comes from the NHC's own track errors over the past five years, sized so the center stays inside about two-thirds of the time. And storm surge comes from the physics-based SLOSH model, run hundreds of times on the official forecast. No AI model in this story tells you how high the water gets on your street.

Asked whether AI will replace his forecasters, NHC science operations officer Wallace Hogsett gave "a resounding no."

What to do with this on the Gulf Coast

The official forecast already has the AI in it. You don't need to hunt down a Google map, and you shouldn't trust one on its own.

Trust the track more than the peak wind. The NHC's average 48-hour intensity error last year was about 14 knots, roughly the width of a full Saffir-Simpson category. And remember what the cone is. At 48 hours its radius is about 71 miles. Mobile and Pensacola are about 50 miles apart, so two days out the cone can't tell you which side of Mobile Bay the center crosses.

Surge, rain, and tornadoes reach well outside the cone, and water is what kills people in these storms. From 2013 to 2022, freshwater flooding alone caused 57 percent of direct U.S. tropical cyclone deaths.

For local eyes on the storm, John's team at Alabama Weather Network is streaming around the clock. AWN launched in August 2025 as a 24-hour channel covering all 67 Alabama counties and weather camera sensors in almost all 67 counites, with a forecast update every 20 minutes and wall-to-wall coverage when a storm is on the way. It's on Roku, Apple TV, Amazon Fire, and YouTube. AWN consults with the NHC and NWS Mobile and then has a meteorologist explain what the models are doing and why.

Oldshue says “It is kind of a fascinating experiment right now to see if AI or physics will win out. When you take them separately there are some incredible differences in track. At the moment a blend is performing best and that's what we continue using for the track on Hurricane Isaias. ”

The AI models are the fastest new tool the Hurricane Center has. They're built on physics and checked by a person who can still explain the shear. For Isaias, that's the forecast we'll be watching.

LA-AI Foley Meetup Canceled

// community highlights

Know someone who’d enjoy this?

Don’t miss the next issue.