High-quality weather and climate forecasts are vital to saving lives and managing resources. As climate change accelerates and extreme weather events worsen, accurate forecasting is becoming even more urgent.
Now, researchers at UChicago’s AI for Climate Initiative (AICE) say that they are developing artificial intelligence models that will push the frontiers of weather and climate forecasts. AICE aims to use AI to augment tools for forecasting weather, assessing socioeconomic impacts of climate change, and informing mitigation and adaptation policies.
Traditional, physics-based weather models require multimillion-dollar supercomputers and take hours to produce forecasts. AI could shorten this time frame dramatically, delivering forecasts in minutes on a laptop.
“AI can help quickly zoom in this region, look at what it is, what will it look like, and then have the emergency response,” said Jiwen Fan, the deputy division director of the Environmental Science Division at Argonne National Laboratory and faculty codirector of AICE. “Extreme events prevention and mitigation often need faster response, and that’s where fast simulation is very important.”
AI has already spurred innovations in forecasting, such as Nvidia’s FourCastNet, the first AI-based weather model to produce high spatial resolution forecasts.
AICE, a joint initiative of the Data Science Institute and the Institute for Climate and Sustainable Growth, brings together experts from multiple disciplines across the University, from computer science to physics to economics. Pedram Hassanzadeh, an associate professor in the Department of the Geophysical Sciences who was involved in FourCastNet, is now the faculty director of AICE.
“In most universities, AI people do their own thing,” Hassanzadeh said. “There is no initiative that actually brings people from all these areas together under one umbrella. And I think that’s what we have at AICE.”
Currently, AI weather forecasting relies mainly on historical data to train models. However, physics-free models are not yet equipped to accurately make forecasts as worsening climate change leads to more unprecedented weather events.
One of AICE’s key focuses is to develop a model that integrates AI and climate physics to better forecast rare or unprecedented weather events as historical data become less reliable at predicting future events.
“People have already shown that, when you get to these more extreme events in a dataset, AI models fail, which may not be very important if you are trying to learn cats and dogs, and the outliers wouldn’t matter,” Hassanzadeh said. “But in weather forecasts, in climate science, the outliers are actually the most impactful [and] extreme weather events.”
The AICE team also sees improved integration of physics-based modeling and AI as key to advancing “subseasonal forecasting.” Subseasonal forecasting—predictions ranging from about two to six weeks ahead—has many practical applications in areas like agriculture, water management, and disaster prevention.
This timeframe can be called a “predictability desert” due to the relative difficulty in accurate forecasting compared to short- or long-term projections. This is because subseasonal forecasting relies on both local weather and global climate conditions, making it more complicated to model.
“As you go at longer time scales,” said Amir Jina, an assistant professor at the Harris School of Public Policy and faculty codirector of AICE, “other factors like the ocean heat on the other side of the world might start mattering.” He added that AI makes it easier to simulate those different conditions.
Researchers at AICE also believe that this initiative can help close the forecasting gap between regions. Studies have found large inequalities in forecast accuracy between countries with differences in economic resources.
“Weather forecasting has been an important field, but weather models are becoming bigger and bigger. You need supercomputers, you need highly skilled people to do weather forecasts. So currently, weather forecasting is only affordable by the richest countries in the world—U.S., Japan, China, and then European [countries],” Hassanzadeh said.
“So many people in the other parts of the world have been left out. Maybe they can go pay the European Center [for Medium-Range Weather Forecasts] and buy their forecast, but the forecast by the European Center is best for Europe—that doesn’t work for Africa.”
AI opens up the opportunity to tailor models to user needs, such as incorporating more specific meteorological data from underrepresented regions like the tropics. AICE is also developing a pilot program to train people from meteorology agencies in 30 countries in Africa and Asia on how they can use the AI models to produce their own weather forecasts.
“At the core, the fundamental driver of it is: ‘How can it be more responsive to the actual needs?’” Jina said.
“If you lower the costs of computing—which sounds like a mundane thing to do, but I think that’s totally transformative—suddenly, then you can have a conversation which isn’t bound by constraints.”
While AI models are transforming weather forecasting, their rising energy demands are drawing concern from climate advocates. In 2024, U.S. data centers—which provide the computing power necessary for AI models—consumed about 200 terawatt hours of electricity, which is about one percent of global electricity demand.
Climate activists have warned that advances in AI as a whole can have dangerous effects on the planet. “It’s not like AI is ridding us of the internal combustion engine. People will be outraged to see how much more energy is being consumed by AI in the coming years,” Michael Khoo, a program director at environmental organization Friends of the Earth, told the Guardian.
But researchers at AICE note that traditional weather models are significantly more energy-intensive. GraphCast, an AI model developed by Google DeepMind, is estimated to be about 1,000 times more energy-efficient than traditional forecasting methods.
“Certainly, one of the things we are thinking about is that we don’t want to work on climate while having models that contribute to climate [change],” Hassanzadeh said. “Whether you want to do physics-based modeling or AI-based modeling, there is a cost to do the computation, and that’s something to be aware of and account for.”
