The AI for Climate (AICE) Initiative at the University of Chicago builds upon the institution’s deep legacy in weather forecasting and climate science and rapid growth in artificial intelligence (AI) for science. AICE aims to leverage the power of AI and machine learning to accelerate and transform the theory and practice of weather and climate analysis and prediction.
Launched in September 2024, this interdisciplinary effort combines climate physics, computer science, economics, applied and computational math, statistics, public health, and the social sciences to pioneer new models for climate and weather prediction and to inform global efforts on adaptation and resilience.
“AICE is among the first of its kind at a university,” said Pedram Hassanzadeh, Faculty Director of AICE and an Associate Professor of Geophysical Sciences and Computational and Applied Mathematics. “Other programs typically have a narrower scope that’s either limited to one or two disciplines or centered solely on deep learning applications as a new tool. By building a wide coalition of researchers, including those focused on socioeconomic impacts and adaptation and mitigation strategies, our program is helping position the University of Chicago as a leader in this space.”
A joint program of UChicago’s Data Science Institute and the Institute for Climate and Sustainable Growth, AICE has already notched several early successes that have helped advance its mission to revolutionize and democratize climate science and weather forecasting. Here’s a look at some of its recent accomplishments.
Expanding Access to Accurate Forecasts
One of AI’s biggest advantages is its ability to dramatically reduce the computational cost of running climate and weather models.
“Traditional, physics-based models have grown more and more complex and expensive, with long-range forecasts requiring the world’s most powerful supercomputers,” said Hassanzadeh, who helped develop the first working AI-based weather forecast model, NVIDIA’s FourCastNet. “AI is far more efficient, for example, by a factor of 100,000. A properly trained high school student with a laptop can now produce state-of-the-art forecasts, though they would need help interpreting them.”
This shift has profound implications for many around the world, especially in low- and middle-income countries (LMIC), which often lack access to high-performance computing but face acute risks from extreme heat and other climate impacts.
“There is an urgent need for innovative solutions to address the growing impacts of climate change on vulnerable communities,” said Amir Jina, who co-leads AICE with Hassanzadeh. Jina is a development economist at the Harris School of Public Policy. “Our research has demonstrated that countries have an opportunity to better protect their economies from climate uncertainties by improving weather forecasting.” For example, research conducted by Jina and colleagues has shown that improved monsoon forecasts help farmers make better decisions—such as how much to plant, what to plant, or whether to plant at all—ultimately leading to smarter investments and greater resilience. “Through our projects, we aim to empower people with actionable climate information, and ensure they have the tools to adapt and thrive in an increasingly uncertain environment.”
The group has partnered with global efforts to enhance climate resilience, such as Agriculture Innovation Mission for Scale (AIM for Scale)—a joint initiative between the United Arab Emirates and the Gates Foundation, for which Jina chairs the technical panel on weather forecasting.
At COP29, the UN’s most recent climate change conference, AIM for Scale announced a $1 billion package to boost weather services for farmers in LMICs, which included funding to support a research and training program led by AICE.
AICE is developing state-of-the-art AI models and training government officials—initially from 30 countries in the tropics and global south—on how to use them effectively, in partnership with the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in the United Arab Emirates.
Focus on the Long Term
AICE also aims to improve its AI models for subseasonal to seasonal (S2S) weather forecasts. Current AI models outperform the physics-based models for 1- to 10-day forecasts, but like the physics-based models, do not offer much skill beyond these lead times. However, longer range forecasts, from around two weeks to 6 weeks, offer valuable climate insights and decision-making opportunities that shorter ones cannot, such as helping farmers to make strategic decisions on such concerns as crop planning, resource use, and risk management well in advance.
Improving S2S forecasting requires large climate datasets, advanced computing power, and deep expertise in areas such as data science and climate domain science. UChicago’s partnership with Argonne National Laboratory plays a critical role in AICE’s efforts to improve S2S models.
“Both institutions play a key role in developing successful AI for Climate projects. Each has AI and domain science expertise, while Argonne provides the exascale computing resources and the University brings the workforce needed to advance projects,” said Jiwen Fan, co-director of AICE and deputy division director of the Environmental Science Division at Argonne. “By combining our strengths, we can take a leadership role in the development of better S2S forecasts.”
This partnership is also crucial to AICE’s work in climate science. Physics-based climate models currently serve as the gold standard for understanding historical climate patterns, projecting future trends, and testing adaptation, mitigation, and climate engineering strategies. Yet, just like traditional weather forecasting models, they are computationally expensive, as well as slow to run and challenging to scale.
Scientists have recently begun turning to AI emulators to more quickly produce climate projections. These tools learn from decades of physics-based simulations and observational data, allowing researchers to quickly run many “what if” scenarios and explore possible futures more efficiently.
These tools, however, come with challenges. Their accuracy depends heavily on the data they’re trained on, and they may struggle with unfamiliar or extreme situations. In fact, just recently, Hassanzadeh and colleagues showed the challenges of AI models with the rarest, most extreme weather events, so-called “gray swans,” calling for innovation at the intersection of AI, mathematics, and climate physics, one of the major goals of AICE.
With these limitations in mind, researchers at AICE are working to benchmark and better understand AI climate emulators, testing how well their predictions match reality and where they fall short, to ensure they’re both fast and trustworthy. Such work can pinpoint where new innovations are needed to build better physics-informed AI tools for the next-generation of weather and climate modeling.
This work, while in its infancy, holds the potential to democratize access to climate projections much in the same way AI can for weather forecasting.
“If we can demonstrate that these emulators show considerable skill and reliability for rare and extreme events or for weather pattern changes, and build up a group who know how to use them, that could be transformative,” said Tiffany Shaw, Director of AICE and Professor of Geophysical Science. “That’s the real opportunity for AI tools like this – bringing new knowledge to more people who can use it to make strategic decisions while reducing the uncertainties we currently face.”