As Hurricane Lee tracked slowly westward in the middle of the Atlantic Ocean on Sept. 10, three new weather models developed in the private sector predicted the storm would make landfall in Nova Scotia about a week later. With the storm still thousands of miles from North America, the forecasts turned out to be an astonishingly accurate achievement for technology not long ago considered to be in its infancy.
Generated by artificial intelligence, the models are orders of magnitude faster and cheaper to operate than conventional, government-run weather models. While AI models don’t yet provide all the capabilities needed for operational forecasting, their emergence portends a potential sea change in how weather forecasts are made, and they could signal a new chapter in the weather forecasting rivalry between the United States and Europe.
“There’s an incredible story emerging about the role of AI weather forecasts,” Daniel Rothenberg, an atmospheric scientist with Google sister company Waymo, said in an email. “This is a glimpse into the future of meteorology, and probably on much faster timescales than most folks in the weather enterprise expect.”
AI weather models have made rapid advances in the past 18 months. Google, Microsoft, NVIDIA and China-based Huawei all published academic articles claiming their AI models perform at least as well as the “European model,” widely considered the gold standard in weather modeling. Those claims were recently corroborated by scientists at the European Centre for Medium-Range Weather Forecasts, which operates the European model. Start-ups, including Atmo, Excarta and Zurich-based Jua, are building AI weather models as well.
The European Center began exploring the potential for AI to further improve its forecasts several years ago. Earlier this month, just a day after Tropical Storm Lee developed in the Atlantic, the European Center began publishing forecasts from Google, NVIDIA and Huawei on its website. The models use current conditions from the European model as a starting point to produce a 10-day forecast at six-hour intervals in approximately one minute, according to the European Center.
Advertisement
After predicting Sept. 10 that Lee would make landfall on Nova Scotia, the AI models fluctuated a bit in the following days, but were consistent in forecasting landfall between Cape Cod, Mass., and eastern Nova Scotia. The AI models were “arguably just as good” as the European and American models, Rothenberg said, and were first to accurately hint that Lee could veer close to New England.
AI weather prediction “has suddenly emerged as a legitimate competitor to [conventional models],” Richard James, a meteorologist at Prescient Weather, which provides weather forecast tools for the energy and agriculture industries, wrote in an analysis of AI forecasts for Hurricane Lee.
While James cautions that one storm is too small of a sample to prove AI models are better than conventional models, “given the remarkable pace of innovation in just the past few years … it is not difficult to imagine that [AI] will be able to replace the [conventional] models for at least some applications in the relatively near future,” he wrote.
Advertisement
The improving performance of AI models has not only attracted the attention of the European Center but also the National Oceanic and Atmospheric Administration, which operates the American Global Forecast System model, also known as the GFS. The two agencies are longtime competitors in computer modeling, with the European model demonstrating greater overall accuracy.
NOAA’s Center for Artificial Intelligence was established in 2021, and this week the agency held its fifth annual AI workshop. The NOAA-affiliated Cooperative Institute for Research in the Atmosphere (CIRA) at Colorado State University will soon release a website similar to the European Center’s, which will display forecasts from AI models that start with current conditions from the American model, according to Imme Ebert-Uphoff, machine learning lead at CIRA, who notes the importance of evaluating the models.
“It’s a question of public safety to evaluate these AI-based models carefully. For one, we want to make sure we use all available tools to improve predictions of severe weather events, and AI-based models have the potential to be extremely helpful for that,” Ebert-Uphoff said in an email. “On the other hand, we also need to make sure not to jump onto the newest models too quickly. AI models can have pitfalls that [conventional] models don’t have.”
The difference between conventional and AI weather models
Conventional and AI weather models both use current atmospheric conditions as a starting point for making forecasts, but that’s where their similarities end.
Advertisement
Computer models programmed with complex mathematical equations and operated by the world’s leading government weather agencies have long served as the backbone of forecasts and warnings. The accuracy of these conventional models has steadily improved over many decades, but they are expensive to operate because they require enormous computing power to make trillions of calculations for a single model run.
Share this articleShareAI models are first trained to recognize patterns in vast amounts of historical weather data. They generate forecasts by ingesting current conditions and applying what they learned from the past, a much less computationally intensive process that can be completed in minutes or even seconds on a desktop computer, compared with more than an hour on large supercomputers for conventional models.
Many experts say that AI models probably won’t make conventional ones obsolete. That’s because conventional models are needed to train the AI models and, at least for now, they also feed the AI models with information on the initial state of the atmosphere. However, the speed and efficiency of AI models could transform how weather forecasts are made and enable more accurate and detailed predictions, especially for extreme weather events.
Advertisement
Neil Jacobs, the former acting head of NOAA and chief science adviser for the agency’s next-generation modeling effort, envisions a day when AI models generate the forecast, and conventional models are only used for training. Jacobs points to the potential to run AI models more frequently and at higher resolutions without having to worry about straining computer resources.
“It’s crazy to think what you could do with this once you remove the limitation of high-performance computing off the table,” Jacobs said in an interview. “NOAA can’t afford to buy a system big enough to run the model at the current [highest] resolution you could configure it to. Well, that problem goes away if you’re using an AI-based system.”
The benefits and limitations of AI
One of the most promising applications of AI for weather forecasting is ensemble modeling, which is when the same model is run multiple times, each time starting with slightly tweaked initial atmospheric conditions to represent uncertainties and approximations made by the model. The result is a range of possible outcomes, rather than a single forecast, that meteorologists use to identify the most probable forecast and assess confidence.
Advertisement
Ensemble forecasts from conventional models can miss extreme events, such as excessive rainfall or heat, because they are limited to about 50 simulations due to the time and cost of generating them. AI could enable the generation of much larger ensembles in as little as a few minutes, potentially leading to more useful forecasts and risk assessments for emergency managers, the general public and numerous industries.
“Our hypothesis is we can easily now scale up with AI models to thousands or tens of thousands of ensemble members,” Anima Anandkumar, senior director of AI Research at NVIDIA, said in an interview.
The European Center says it believes ensembles are “key for delivering valuable forecasts for medium-range timescales” and has started a project to create its own AI-based system.
AI models have limitations despite their recent progress. For example, they don’t all yet produce forecasts for a number of key parameters, such as precipitation and clouds. They’ll also need to earn the trust and understanding of forecasters who have spent their careers working with conventional models. But the fast pace of innovation has meteorologists excited about the potential.
“I think it’s the future, particularly for operational forecasting,” Jacobs said.
ncG1vNJzZmivp6x7uK3SoaCnn6Sku7G70q1lnKedZMSmrdOhnKtnYmV%2FdHuPcmZraV%2BdwrO%2ByJyYp51dobKmecCrq6KemZi2oriMoqWtnZyhtqixzZycZp6fp7KkrdKtoKefXw%3D%3D