10
Loading video...
By Marie Boran
Technology Reporter
0
Share
Newsweek is a Trust Project member
See more of our trusted coverage when you search.Prefer Newsweek on Googleto see more of our trusted coverage when you search.
A weather model driven by artificial intelligence has stunned meteorologists by predicting Hurricane Milton's landfall within 7 miles on average—outperforming traditional models by almost 100 miles.
Artificial Intelligence/Integrated Forecasting System (AIFS) forecast Milton's landfall location within 7 miles of the actual point, five days before it hit Florida near Siesta Key. Other leading models' maximum errors were up to 100 miles from landfall, while AIFS' maximum error didn't exceed 13 miles.
"While it is difficult to draw conclusions from one case, it seems like AIFS provided somewhat better track forecast for Milton one to two days before landfall. From October 2 on, the AIFS consistently predicted that Milton would make landfall in Florida," a spokesperson for the AIFS team told Newsweek.
Read More on Tech
Writing on X (formerly Twitter), Bryan Bennett, a meteorologist who closely monitored the storm, described the model's "unbelievable accuracy." He emphasized the model's early and precise predictions: "It predicted a landfall near Siesta Key more than five days in advance. The model knew that it wouldn't make landfall in Pinellas County or anywhere else in Florida."
Bennett noted that the advanced weather model has consistently delivered precise forecasts for other storms this season, including Hurricane Helene in September. "This is the future," he wrote. "Modelers are likely taking notice and will attempt to hop on board the AI weather modeling in the coming years."
The Tech Behind AIFS
AIFS stands for Artificial Intelligence/Integrated Forecasting System, a nod to ECMWF's well-known Integrated Forecasting System (IFS).
In 2017, the IFS model accurately predicted the path of Hurricane Irma about a week in advance, the 2019 heatwave in Europe, and the conditions that contributed to the bushfires in Australia in 2019 and 2020.
Unlike traditional numerical weather prediction models that rely heavily on physics-based equations, AIFS uses machine-learning techniques.
The model is trained on decades of weather data, allowing it to recognize complex patterns and relationships in atmospheric conditions. This enables it to generate forecasts rapidly and with significantly less computational power compared to conventional models.
The AIFS team spokesperson told Newsweek that they are currently training the model on historical reanalysis but that "the magnitude of extremes (i.e., tropical cyclone intensity) is limited by what can be captured by the reanalysis."
"Although we trained on several decades of historical data, the challenge/question is always whether the samples for extreme (and therefore rare) events such as tropical cyclones are large enough to make good predictions," they added.
The AIFS is inspired by GraphCast, an open-source machine learning architecture developed by Google DeepMind, which uses graph neural networks (GNNs). Neural networks are programs that process data, while graph neural networks process data that can be represented as graphs, such as weather patterns over geographical grids.
Andrew Brady, a weather prediction expert at AtmoAnalytics, explained on X how AIFS functions: "These models aren't weather models designed by AI. They use a somewhat similar underlying architecture but do not use human-like logic in any way. 'Machine learning' is probably an easier way to understand these models than the term 'AI.'"
...
The AIFS model goes through three key stages: encoding initial weather data onto a grid, processing this data through multiple layers where grid points exchange information, and decoding this processed data back onto a global grid to generate forecasts. This process allows AIFS to produce a six-hour forecast rapidly, which can be extended by feeding the output back into the model over and over again.
Why Meteorologists Are Excited
The success of AIFS signifies a major advancement in accurately predicting large-scale weather events like hurricanes. Meteorologists are particularly enthusiastic about the model's ability to provide consistent and reliable forecasts over extended periods.
Matt Lanza, a meteorologist based in Houston, Texas, highlighted the model's consistency, stating that even the best traditional models have not demonstrated such reliability in their projections.
Noting how the AI weather model worked to predict the path of Hurricane Francine, which hit the Gulf Coast in September, Lanza said on X: "Insane. I mean, INSANE consistency with Francine."
"This is a major victory for AI modeling. The consistency is just incredible. Even the best performing traditional models I don't believe were this consistent."
This research between meteorological institutions and technology companies could accelerate the development of even more sophisticated forecasting tools. ECMWF's code is open source.
Enthusiasts eager to explore this AI weather prediction system can access it through ECMWF's open data initiative. The organization has made it possible for users to run AI models themselves.
The ECMWF team said in September: "We are now training a higher-resolution version of AIFS. An important next step will be to extend AIFS to create ensemble forecasts. Here, different avenues are possible, such as probabilistic training or exploiting generative AI architectures."
Update 10/14/24, 12:05 p.m. ET: This story was updated with comment from ECMWF's AIFS team.
View 0 Comments
Submit Correction
View Editorial & AI Guidelines