Google’s AI Weather Model Nailed Its First Major Storm Forecast

While generative AI tools that primarily amount to slop generators grab most of the attention in the artificial intelligence space, there are occasionally some actually useful applications of the technology, like Google DeepMind’s use of AI weather models to predict cyclones. The experimental tool, launched earlier this year, successfully managed to provide accurate modeling of Hurricane Erin as it started gaining steam in the Atlantic Ocean earlier this month.

As Ars Technica first reported, Hurricane Erin—which reached Category 5 status and caused some damage to the island of Bermuda, parts of the Caribbean, and the East Coast of the United States—provided Google DeepMind’s Weather Lab with the first real test of its capabilities.

According to James Franklin, former chief of the hurricane specialist unit at the National Hurricane Center, it did quite well, outperforming the National Hurricane Center’s official model and topping several other physics-based models during the first 72 hours of modeling. It did ultimately fall off a bit the longer the prediction effort ran, but it still topped the consensus model through the five-day forecast.

While Google’s model was impressively accurate in the first days of modeling, it’s the latter ones that are most important to experts, per Ars Technica, as days three through five of the model are the ones that officials count on to make decisions on calls for evacuation and other preparatory efforts. Still, it seems like there may be some promise in the possibility of AI-powered weather modeling—though the sample size here is pretty small.

Most of the current gold standard modeling techniques used for storm prediction use physics-based prediction engines, which essentially try to recreate the conditions of the atmosphere by factoring in things like humidity, air pressure, and temperature changes to simulate how a storm might behave. Google’s model instead pulls from a massive amount of data that it was trained on, including a “reanalysis dataset that reconstructs past weather over the entire Earth from millions of observations, and a specialized database containing key information about the track, intensity, size and wind radii of nearly 5,000 observed cyclones from the past 45 years.”

According to Google, it tested its model on storms from 2023 and 2024, and found that its five-day prediction managed to predict the path of a storm with more accuracy than most other models, coming about 140km or 90 miles closer to the ultimate location of the cyclone than the European Centre for Medium-Range Weather Forecasts’ ensemble model, which is considered the most accurate model available. Now it can point to a storm that it tracked in real-time as proof of concept, though there is no reason to think AI tools like this will completely displace the other approaches at this stage.