Google AI Meteorological model nailed its first major storm forecasts

Although generative AI tools that are mainly equivalent to solp generators attract most of the attention in the artificial intelligence space, there are sometimes really useful applications of technology, such as the use by Google Deepmind of IA weather models to predict cyclones. The experimental tool, launched earlier this year, managed to provide precise modeling of Hurricane Erin while it was starting to gain steam in the Atlantic Ocean earlier this month.

As Ars Technica reported for the first time, Hurricane Erin – which reached category 5 status and caused damage to Bermuda Island, in some parts of the Caribbean and the East Côte du United States – provided the Google Deepmind weather laboratory with the first real test of its capacities.

According to James Franklin, former head of the specialized unit of hurricanes of the National Hurricane Center, it did very well, outperforming the official model of the National Hurricane Center and at the top of several other models based on physics during the first 72 hours of modeling. He finally dropped a little more the prediction effort worked, but he nevertheless exceeded the consensual model through the five -day forecasts.

Although the Google model was impressive impressively in the first days of modeling, it is the latter who are the most important for experts, according to ARS Technica, because the days three to five are those on which those responsible count to make decisions on calls for evacuation and other preparatory efforts. However, it seems that there may be a promise in the possibility of meteorological modeling powered by the AI ​​- although the size of the sample here is quite small.

Most current techniques of modeling gold standards used for the prediction of the storm use prediction engines based on physics, which essentially try to recreate the conditions of the atmosphere by affecting things such as humidity, air pressure and temperature changes to simulate how a storm could behave. The Google model rather removes a massive amount of data on which it has been formed, including a “set of reanalysis data which reconstructs the weather conditions spent all over the earth from millions of observations, and a specialized database containing key information on the track, intensity, size and wind rays of almost 5,000 cyclones observed in the past 45 years.”

According to Google, he tested his model on the storms of 2023 and 2024, and found that his five -day prediction managed to predict the path of a storm with more precision than most other models, coming at around 140 km or 90 miles closer to the ultimate location of the cyclone than the European model of the center for average weather forecast. Now, he can indicate a storm that he has followed in real time as proof of concept, although there is no reason to think that AI tools like this will completely move other approaches at this stage.