AI-generated images based on prompts “tornado in rural Midwestern United States,” “thunderstorm with lightning over city” and “field of corn.” Credit: DALL-E (first two from left) and Microsoft Bing Image Creator, prompted by Everett Hogrefe and Jayme DeLoss
AI helping to unravel complexity of climate, weather and land use, find solutions to climate change
story by Jayme DeLoss
published Aug. 31, 2023
Imagine that we could predict not only severe storms more than a week out but also what the climate will be like in 50 years, and how intervention strategies might lessen the impacts of climate change. Colorado State University researchers are developing ways to do all these things using a powerful tool: artificial intelligence.
Atmospheric Science Professor Elizabeth Barnes uses machine learning, a subset of AI, to disentangle the complexity of climate science. Professor Russ Schumacher, Colorado state climatologist and director of the Colorado Climate Center, led development of a machine learning model that can accurately predict severe weather four to eight days in advance and is now used daily in National Weather Service operations. And a team led by University Distinguished Professor Keith Paustian in the Department of Soil and Crop Sciences will take CSU’s world-renowned greenhouse gas quantification expertise to the next level by combining its strengths with those of machine learning.
Machine learning: The perfect tool for climate science
Barnes’ research group uses machine learning to detect the impacts of climate change, predict weather and climate a few weeks to decades into the future, and explore the potential outcomes of hypothetical climate intervention strategies like geoengineering.
The climate system is incredibly complex, and those who study it rely on massive amounts of data. Barnes said machine learning is the perfect tool for climate scientists.
“We’ve always been using data and trying to pull out all the complexity of the climate system and make it understandable to a human, and now machine learning is allowing us to go in even deeper and find even more complicated relationships,” she said. “The place we’re at right now is still trying to make their predictions understandable to humans.”
Given enough data, a complex enough machine learning model can find patterns among the noise and potentially produce accurate predictions, but Barnes is interested in explainable AI – that is, figuring out how a machine learning model reached the conclusion it did. She likens deciphering the model’s process to solving a maze by starting at the end and working backward.
“We’ve always been using data and trying to pull out all the complexity of the climate system and make it understandable to a human, and now machine learning is allowing us to go in even deeper and find even more complicated relationships. The place we’re at right now is still trying to make their predictions understandable to humans.”
—Atmospheric Science Professor Elizabeth Barnes
“If it does a good job and we can learn why it was able to do it, we then actually learn new climate science,” Barnes said.
Her group is also focused on interpretable AI – sometimes called transparent AI. They are starting from scratch, building machine learning models from the ground up, so the models are understandable to people every step of the way.
“That’s a much slower process and honestly way harder,” Barnes said, “but the result is when it makes a prediction, you don’t have to ask, ‘Why did it make that prediction?’ You already know why.”
Explainability and interpretability are two pieces that can help people trust AI, but a lot of other factors come into play. Barnes, Imme Ebert-Uphoff, a scientist with the Cooperative Institute for Research in the Atmosphere and professor in the Department of Electrical and Computer Engineering, and Computer Science Professor Chuck Anderson are exploring what it will take to create trustworthy AI for studying weather and climate. They are partners in the National Science Foundation-funded Institute for Research on Trustworthy AI in Weather, Climate, and Coastal Oceanography, led by the University of Oklahoma.
Climate scientist Elizabeth Barnes’ group has produced several animations explaining the AI tools they use to study climate.
CSU weather model used in storm prediction
A machine learning model created at CSU has improved forecasters’ confidence in storm predictions and is now used daily by the National Weather Service’s Storm Prediction Center and Weather Prediction Center.
The model, developed in the Department of Atmospheric Science by a team led by Schumacher, is capable of accurately predicting excessive rainfall, hail and tornadoes four to eight days in advance. The model is called CSU-MLP for “Colorado State University-Machine Learning Probabilities.”
Schumacher’s team worked with NWS forecasters over six years to test and refine the model for their purposes. The CSU code is now running on the Storm Prediction Center’s and Weather Prediction Center’s operational computer systems, helping forecasters predict hazardous weather, so people in harm’s way have enough lead time to prepare.
The atmospheric scientists trained the model on historical records of severe weather and NOAA reforecasts, retrospective forecasts run with today’s improved numerical models.
Team member Allie Mazurek, a Ph.D. student, is working on explainable AI for the CSU-MLP forecasts. She’s trying to figure out which atmospheric data inputs are most important to the model’s predictions, so the model will be more transparent for forecasters.
“These new tools that use AI for weather prediction are developing quickly and showing some really promising and exciting results,” Schumacher said. “But they also have limitations, just like traditional weather prediction models and human forecasters have strengths and limitations. The best way to advance the field and improve forecasts will be to take advantage of each of their strengths: the AI for what it’s good at, which is identifying patterns in massive datasets; numerical weather prediction models for being grounded in the physics; and humans for synthesizing, understanding and communicating.”
Schumacher discusses the promise and limitations of AI for weather prediction in more detail in this piece in The Conversation, co-authored by Aaron Hill, a former CSU research scientist who is now a faculty member at the University of Oklahoma.
AI Institute focused on sustainable farms and forests
CSU is partnering with the University of Minnesota, Cornell University and several other universities in a research institute that will use AI to create climate-smart agriculture and forestry practices.
The AI-CLIMATE Institute, which stands for Climate-Land Interactions, Mitigation, Adaptation, Tradeoffs and Economy, was announced earlier this year and is funded by a $20 million grant from the National Science Foundation and the USDA National Institute of Food and Agriculture.
The institute will employ AI techniques like deep learning and knowledge-guided machine learning to improve accuracy and lower the cost of accounting for carbon and other greenhouse gases, providing essential data for carbon offset programs.
Long a leader in measuring and monitoring soil carbon stock changes and soil greenhouse gas emissions, CSU will help create AI-guided methods for model-based predictions, marrying machine learning models with the process-based biogeochemical models developed over decades at CSU.
Paustian is principal investigator of CSU’s AI-CLIMATE team of faculty, postdoctoral researchers and students. Their goal is to figure out how to best manage farm and forest land for maximum carbon sequestration, while reducing other greenhouse gases including nitrous oxide and methane, without harming other ecosystem services like biodiversity or water quality. In other words, they want to optimize land use for all its various purposes, including economic and environmental uses.
CSU scientists involved in the AI-CLIMATE Institute
Keith Paustian, University Distinguished Professor in Soil and Crop Sciences and a senior research scientist at the Natural Resource Ecology Laboratory
Francesca Cotrufo, professor of Soil and Crop Sciences and senior research scientist at NREL
Patrick Keys, assistant professor of Atmospheric Science
Nathan Mueller, assistant professor of Ecosystem Science and Sustainability and Soil and Crop Sciences
Stephen Ogle, professor of Ecosystem Science and Sustainability and senior research scientist at NREL
Sangmi Pallickara, professor of Computer Science
Shrideep Pallickara, professor of Computer Science
Yao Zhang, research scientist at NREL and in Soil and Crop Sciences
“A strength of AI-based models is they’re much better at handling those really complex optimization problems,” Paustian said. “You can use those technologies to get better optimization that can also then feed into decision-making tools.”
AI-CLIMATE’s resulting knowledge-guided machine learning model will assign likelihoods to different potential outcomes so that people can make more informed decisions.
“This project is not only to support better understanding of ecosystems and climate and system management, it’s also to develop information that’s useful for people who are trying to deal with the problems we’re facing,” Paustian said. “If we can know better the consequences of different management as well as all the complexity of the ecosystem factors that are determinative, then we’re going to be able to better design effective solutions and policies.”
AI-CLIMATE will work to understand climate change impacts and explore potential adaptive responses. It also will develop AI-inspired data visualization tools and contribute to scientific workforce development.
AI Research
Artificial intelligence isn’t science fiction anymore. This special report from SOURCE explores the importance of artificial intelligence research and what you really need to know about the potential and impact of this empowering, disruptive and complicated technology. read more