We got a glimpse of what the Google DeepMind (GDM) model could do during last hurricane season, and we’ve been watching all of the AI models this season, too, but Hurricane Isaias provided the first real test for it this Atlantic hurricane season.

The model began to hint at tropical development in the Bay of Campeche as many as 12 days before the system actually formed, but it did lose the signal for several days. It and other models brought back development chances just over three days before the tropical depression that became Isaias formed. 

(MORE: Hurricane Isaias recap)

This storm then brought an interesting conundrum several days before landfall: Do we trust the models that we have been using for years such as the European model and GFS American model or the HAFS and HMON hurricane models — or do we trust the new kid on the block, this Google DeepMind AI model? 

The two camps of models were decidedly split several days before landfall, and it happened that the Google model was on the right eastern side of the models. The GDM was frequently painting a landfall over Florida with many ensemble members showing a strong to major hurricane while the GFS, the westernmost model suite, often showed a landfall over Mississippi or Louisiana a couple of days in advance. 

Any extremes in models are typically taken with a grain of salt because they’re more likely to be wrong. The National Hurricane Center took a stance early with Isaias, from early on Oct. 7, two and a half days before landfall: 

Get More of Our News on Google

Set The Weather Company as a 'Preferred Source' to get quicker access to the news you value.

As mentioned in the previous discussion, there remains quite a bit of along and across-track spread in the track guidance, with the AI-model guidance remaining on the east side, and physics-based models on the west side of the guidance envelope. The NHC track forecast continues to favor more of the eastward solutions, but not quite as far east as the latest HFIP Corrected Consensus Approach (HCCA) or Google DeepMind ensemble mean (GDMI).

That bet paid off big.

Notice how much empty space there is to the left (west) of the various runs of the GDM. That space was there because of the legacy global and hurricane models. As we got closer to landfall, the NHC continued to favor, rightfully, the GDM and other in-house corrected guidance on the right (east) side of its initial forecasts.

(HURRICANE TRACKER: Watch the latest storms from Hawaii to the Atlantic)

If you take a look at the forecasts from Google and the National Hurricane Center, below, you can immediately see how well GDM did when everything was over.

The GDM even outperformed the NHC at several near-term timesteps for track forecasts.

The Google DeepMind model did quite well with intensity too, even several days out. Isaias reached a peak intensity of 120 mph and made landfall at 105 mph. 

(PHOTOS: See the damage left behind by Hurricane Isaias)

And we don’t want to steal all the thunder for the GDM. NOAA’s HMON and HWRF models, two other models specializing in hurricanes, did very well for both intensity and track forecasts. 

Model trouble spots

There was one sore spot for the DeepMind suite where the hurricane models did a bit better: The sting jet (or band of very strong winds) that developed on the western side of Isaias near landfall. That's a feature that isn't typical of hurricane landfalls.

Experimental runs of the HAFS hurricane-specific models showed some stronger winds on the western side of Isaias closer to landfall. Winds over 100 mph occurred there in reality. 

This does come with the caveat that DeepMind is not (yet) meant to pick out such small-scale features. 

And to give credit where credit is due, the rapid transition from Category 3 hurricane to a deflated post-tropical storm in around 12 hours was extraordinarily unusual.

Jonathan Belles has been a digital meteorologist for weather.com for 10 years. His favorite weather is tropical weather, but also enjoys covering high-impact weather and news stories and winter storms. He's a two-time graduate of Florida State University and a proud graduate of St. Petersburg College.