At the 2024 Paris Olympics, the city’s unpredictable weather has taken center stage. Starting off with a downpour during its opening ceremony that soaked more than 10,000 participating athletes, subsequent rainfall has postponed matches in skateboarding and tennis. Now, the summer games are bracing for a heat wave across both Paris and the city of Marseilles, where competitions in soccer and sailing will be held.
A handful of meteorological services and universities are using this year’s Olympics to advance weather forecasting research. These teams include a group of researchers incorporating A.I. tools to aid forecasts at the University of Texas (UT) at Austin. Led by UT Austin professor Dev Niyogi and postdoctoral fellow Manmeet Singh, the research lab uses machine learning to focus on weather conditions at a smaller “city-scale” of 1 kilometer by 1 kilometer or less, taking data from large models typically providing weather at a scale of around 25 kilometers by 25 kilometers. “We are trying to represent cities, and neighborhoods from cities, into weather forecasting,” Niyogi told Observer.
In addition to measuring traditional indicators like rain chance, wind speed, temperature and humidity in their daily forecasts, the UT Austin researchers are using their lab’s global building heights dataset to pursue more specialized outputs like thermal comfort, which measures heat conditions while taking into account factors like shade provided by nearby buildings and trees. The team’s efforts, which need significant computing power, are supported by computing systems at the university’s Texas Advanced Computing Center.
Alongside around a half dozen forecasts from organizations including the national weather services of the U.S., the U.K., Germany and Sweden, the UT Austin team’s daily forecasts are sent to Olympic officials as part of a research demonstration project supported by the World Meteorological Organization (WMO). Supplementing a regular forecast from Météo-France, the official French meteorological association, the project participants are not providing official forecasts. Still, their results can be shared with organizers who are able to use information as they see fit.
The project’s goal is to understand “what kinds of methods or approaches are working or not working, such that they will help us in the future in developing models that could perhaps be used in the next Olympics,” said Niyogi. We are using the games as a means to improve our understanding—this is machine learning helping actual learning.”
Weather forecasting has increasingly shown potential as an area demonstrating rapid advances with the help of A.I. Led by Google (GOOGL), a new model known as Neural GCM combined A.I. and conventional physics-based models to develop forecasts faster than traditional weather forecasting models and is more accurate than purely A.I.-models, according to a recent study from Nature.
In recent years, tech companies around the globe have poured resources into developing A.I. tools for weather forecasting. Last year, Google, Microsoft (MSFT), Nvidia (NVDA) and China’s Huawei said that the performance of their A.I. weather models is on par with a well-respected European model run by the European Centre for Medium-Range Weather Forecasts (ECMWF), a claim confirmed by the ECMWF. And earlier this month, Google DeepMind’s forecasting system GraphCast correctly predicted that Hurricane Beryl would land in Texas.
According to Niyogi, the future of weather forecasting will include a hybrid of A.I.-based models and physics and dynamics-based models, the latter of which remains the “DNA and backbone” of weather forecasting. “The combination of dynamics-based models with A.I. is where the field is going,” he said. “In the future, it is not going to be one or the other; it’s going to be one with the other.”
