For the first time, the National Hurricane Center in Miami is working with an artificial intelligence company to improve its forecasts of the powerful storms that kill thousands of people globally every year. The Atlantic season has just begun and runs through November.
DeepMind, a Google company based in London, announced on Thursday that it was supplying the government forecasters with a newly enhanced variety of its weather forecasting models. Specialized to focus on hurricanes, the model tracks a storm’s development for up to 15 days, predicting not only its path but also its strength, an ability that earlier A.I. models lacked.
Strength readings can make storm warnings far more accurate. So can reliable predictions of hurricane paths, which are known to zigzag, loop around, slow down, make hairpin turns or come to a complete stop.
The hurricane center is not eliminating its human forecasters. Instead, the Google A.I. program will be used on an experimental basis by those same experts in their existing work. Still, the research partnership is the first time in which the Miami center is drawing on an A.I. company to learn how to better warn of nature’s most destructive storms.
“It’s about helping people protect themselves,” Wallace Hogsett, the center’s science and operations officer, said in an interview. The union of skilled human forecasters and the A.I. tool, he added, has the potential to create “a really powerful partnership.”
Dr. Hogsett said the new accord is structured in what’s known as a cooperative research and development agreement, or CRADA. Originally, CRADAs let Washington spin off federal technologies for industry use. But increasingly, the agreements give the government a window into private-sector innovations.
In its Thursday announcement, DeepMind said its forecasts for hurricane intensity “are as accurate as, and often more accurate than,” traditional methods. That could matter because, for example, some hurricane winds of 75 miles per hour, while dangerous, can be far less consequential compared with explosive blasts of 160 miles per hour, which can shatter homes, uproot trees and knock out power for months.
In addition to the upgraded A.I. model, DeepMind unveiled a computer visualization tool, WeatherLab, that lets users see how the new hurricane forecasts compare with earlier A.I. programs it produced, known as GraphCast and GenCast. Both models made their public debuts last year. Testing showed they outdid traditional forecasts.
The visualization tool shows how the new model and other programs compare in tracking current hurricanes. It can also make comparisons with storms as far back as early 2023. To create a reference point in actual weather rather than computer projections, the tool devotes a separate line to laying out a hurricane’s observed path.
DeepMind said it was planning to make Weather Lab available openly to scientific researchers as well as the general public. In its announcement, the company emphasized that the model represented an experimental tool and that its forecasts for current hurricanes would not be “official warnings.”
The new partnership comes after the Trump administration made hundreds of staff cuts at the National Weather Service, parent agency of the National Hurricane Center. It is now filling a limited number of those vacant posts. The partners said planning for the accord predated the Trump administration, having begun informally a little more than a year ago.
Ferran Alet, a research scientist at DeepMind, said he had experienced “a bit of a cold reception” when he described the earlier A.I. models at specialized meetings of hurricane experts. He added that, six months later, as he outlined the planned upgrade, meeting participants were suddenly “much more welcoming.”
Dr. Alet said he saw the Atlantic hurricane season as “the test” of the upgraded model, adding that the A.I. newcomer looks “very, very strong on both tracks and intensities.”
Last month, the National Oceanic and Atmospheric Administration, the parent agency of the National Weather Service, forecast an “above average” season, with a total of 13 to 19 named storms. An average Atlantic season has 14 named storms, including seven hurricanes and three major hurricanes.
Tom Andersson, a research engineer at DeepMind, said the new model had been schooled on data sets that recorded the life cycles of nearly 5,000 hurricanes over the past 45 years. The data libraries included those of the National Hurricane Center as well as the European Center for Medium-Range Weather Forecasts, a science cooperative for dozens of countries.
Dr. Hogsett of the hurricane center said that its application of A.I. to storm forecasting might seem “new and crazy” but described it as a natural step.
He noted that the model had been trained on a wealth of observational data from networks of balloons, radars, ships, aircraft and satellites that world governments and private companies had built at considerable toil and expense.
“This is a new layer on top of that,” Dr. Hogsett said of the A.I. model. “We’re building tools that combine everything that we’ve invested in over the decades.”
William J. Broad has reported on science at The Times since 1983. He is based in New York.
The post Google and U.S. Experts Join on A.I. Hurricane Forecasts appeared first on New York Times.
