Written by Nandini Budithi
The Midwest is well known for experiencing all four seasons in a single month, and nowhere is this more apparent than in Madison, Wisconsin. One week in October, everyone wears T-shirts and flip-flops, and the next, hooded puffers and sweatshirts are all everyone wears. Well, except for that one guy in shorts. However, this phenomenon isn’t exclusive to the Midwest. Due to climate change, regions across the world have been experiencing drastic changes in weather patterns over the last decade. Even forecasters struggle to describe the recent weather on the morning news.
How do we model storms that are bigger? How do we model precipitation on such a small scale?
– Michael Morgan
Not only do these changes affect your ability to plan for a week, but they can also impact a country’s ability to prepare for storms and natural disasters. Traditional weather models are struggling to keep up with the shifting tides. Fortunately, UW-Madison’s own Atmospheric and Oceanic Sciences Department is developing AI powered models with faster weather forecasting.
Traditional models rely on atmospheric physics and fluid dynamics to predict weather events and characteristics. First, the models must define constraints using concepts such as Newton’s second law, thermodynamics, and conservation of mass. Next, they derive equations to determine temperature, pressure, and wind speeds.
However, these processes only work for a specific size model. Michael Morgan, a professor in the Atmospheric and Oceanic Sciences department at UW-Madison, explains, “In 10×10 km chunks, we can observe thunderstorms that are smaller, but how do we model storms that are bigger? How do we model precipitation on such a small scale?” While the ability to define constraints makes physics-based weather models reliable, it also limits the magnitude of storms and precipitation that we can model.
Overcoming these limitations, Morgan and his lab are developing and using AI-based models to help understand and predict organized storms, like cyclones, hurricanes, and other tropical storms. They utilize either raw data from satellites on campus or organized data from the European Center for Medium-Range Weather Forecast (ECMWF) analyses. This data is then used to generate and train machine learning models to predict severe weather events. These new models can generate a prediction much faster than a traditional, physics-based model without deriving a single equation.
One of these AI models currently being developed is called Prob Severe. Prob Severe deduces the likelihood of a severe weather event occurring within the hour and updates every two minutes with a new prediction. As Morgan describes, “The first flashes of lightning might not have even occurred yet. But through the training of data, they’re able to predict over the next hour where lightning most likely occurs. They’ve actually shown considerable skill in that.”
Not only are these models swift, but they are precise and accurate. Although they have faced skepticism in the past, machine learning models are outperforming other models such that major weather agencies around the world have begun to use them for regional weather forecasting. ECMWF, a leading figure in weather forecasting, has begun operational use of its AI model, the AIFS. This model was used to predict snowfall during the winter. Tech companies like Google are also capitalizing on this technology. Google DeepMind’s forecasting model is operated by the National Oceanic and Atmospheric Administration (NOAA) to forecast hurricanes, such as Hurricane Melissa that passed over Jamaica.
Of course, more testing and operational use needs to be done before they fully replace physical models, but their speed and accuracy are very promising for the future. Unlike most generative AI models, they don’t display any signs of “hallucinating” predictions, and their operational use is proving to be beneficial for affected communities. Training these models is resource-intensive, with data centers using significant energy. Yet on the flip side, according to Morgan, the runtime post training is incredibly fast and uses only a few CPUs, compared to the regular global forecast that requires hundreds. Any decent-sized laptop could generate forecasts, saving tons of resources and computing power in the long term.
The accessibility of models could lead to their widespread adoption in the future, letting every country, city, and person have access to on-demand forecasts. Perhaps then we can plan for a safer future.
Graphic Design by Shreya Venkatesh
View this article and the rest of Spring 2026‘s issue in our GALLERY!
