AI-CLIMATE Institute: Curbing Climate Change with Artificial Intelligence
Artificial intelligence (AI) is having its moment. With the emergence of publicly-available tools like ChatGPT, the world is abuzz with the possibilities that AI can offer. The recent successes surrounding AI are a result of research and development over 60 years under a variety of names - deep learning, data mining, recommender systems, machine learning, etc. In the last decade, deep learning algorithms surpassed competing technologies in processing text, images, video, and audio with 10-20% better results.
Around the same time as that significant breakthrough, computer scientists at the University of Minnesota began developing unique data mining and machine learning approaches to address one of the world’s most globally pressing issues: climate change. Led by Department of Computer Science & Engineering Professors Vipin Kumar (Principal Investigator) and Shashi Shekhar, the U of M received a five-year $10 million grant from the National Science Foundation’s (NSF) Directorate for Computer and Information Science and Engineering (CISE) Expeditions program in 2010 for their project titled, “Understanding Climate Change: A Data Driven Approach”. This expedition went beyond the physical models traditionally used to understand climate. Instead, it used data mining to analyze global datasets over time from climate models and in-situ and satellite-based sensors to identify patterns that could not be found using traditional climate modeling approaches.
“This line of work started about 25 years ago in collaboration with scientists at the National Aeronautics and Space Administration (NASA),” said Kumar. “We have a long history that dates back to 1999 of building machine learning methodologies to discover new climate patterns, such as El-Nino, and monitor changes to the environment, like deforestation, on a global scale.”
“The 2010 Expedition grant was the biggest investment from the computer science community in climate change research at the time,” said Shekhar. “This study put the University of Minnesota at the forefront of the intersection between computer science and climate change research. By the end of this project, we knew there would be impacts in areas such as agriculture, environment, forests, renewable energy, and water resources.
We also realized that the climate change data violated a key assumption underlying many data-driven methods, mainly that the future data is similar to the past data. Also, deep learning methods did not honor laws of nature - for example, water flows downhill, nearby places are similar, and mass is conserved. Other researchers echoed these concerns in a 2016 national workshop on exploring data science challenges and opportunities in understanding the interactions among food, energy, and water systems facing climate change and population growth. In other words, popular machine learning techniques needed to be generalized to effectively address climate-change-related problems. That really laid the groundwork for our current grant.”
Shekhar is now the director of the new National AI Research Institute for Climate-Land Interactions, Mitigation, Adaptation, Tradeoffs and Economy (AI-CLIMATE). Kumar is the co-principal investigator, and CS&E faculty members Yao-Yi Chiang, Maria Gini, Nikolaos Papanikolopoulos, and Ju Sun are among its key leaders. This five-year $20 million institute is one of seven funded by NSF and the U.S. Department of Agriculture’s (USDA) National Institute of Food and Agriculture (NIFA). The AI-CLIMATE Institute will create and leverage novel AI methods to revolutionize the agriculture and forestry industries, prioritizing climate-smart practices in efforts to curb climate change by not only reducing emissions from, but also absorbing atmospheric carbon into land such as farms, forests, grasslands, and wetlands.
“Even with the progress on electric cars and renewable energy, there are still residual emissions and cases where we have not found a strong alternative to fossil fuels,” said Shekhar. “In order to get to net-zero, we need sectors that are net-negative and absorb more greenhouse gasses than they emit. Agriculture and forestry is one of the more pragmatic areas where we can make some real progress. Right now, agriculture accounts for 10% of carbon emissions in the U.S. There are some practices that can reduce and even negate those emissions - down to negative 4%.”
Farmers and foresters are already using many climate-smart practices, such as forest and fire management, cover crops, nutrient management, biochar, alley cropping, and restoration of forests, grasslands, and wetlands. However, they face a key question: which practice should be used where and when? AI can inform these decisions in many ways.
First, AI improves the availability and quality of data, such as maps of soil health, moisture, and organic carbon. AI also recommends suitable climate-smart practices for specific places, creating a win-win for seasonal and long-term goals such as yield, soil health, water quality, and carbon sequestration. Lastly, AI improves the timing of actions such as watering and fertilizer application to reduce runoffs, erosion, and water quality impacts.
Put simply, AI can help improve data accuracy and optimize land management decisions relating to climate-smart practices, reducing the workload for individual farmers and foresters. With promising preliminary results from their machine learnings, which leverages data from sensors on the ground and satellites, the AI-CLIMATE Institute hopes to improve the accuracy and lower the cost of estimating how much carbon is sequestered in farms and forests. Additionally, implementing strategic climate-smart agriculture could increase revenue for rural areas of the U.S.
Using the knowledge-guided machine learning (KGML) methods pioneered by Kumar and other researchers at the U of M, the Institute is uniquely positioned to create more sophisticated deep learning models that blend the laws of nature with artificial intelligence.
“The U of M has been at the forefront of bringing this new class of machine learning that is grounded in scientific knowledge,” said Kumar. “Traditional machine learning models require large datasets and they are not very good at making predictions in new scenarios they have not encountered before. Because KGML algorithms incorporate scientific knowledge, they can more easily generalize to unseen scenarios and can be trained with much smaller datasets that are typically available in scientific applications. These KGML algorithms are now being used in numerous scientific applications.”
Additionally, the AI-CLIMATE Institute is improving current AI models by adapting them to tackle the challenges of spatial-autocorrelation and spatial variability. The team has developed a spatial-variability-aware neural network where the edge weights are maps rather than numbers, enabling the technology to make more regionally specific recommendations. There is no one-size-fits-all solution for different regions of the planet, and advances in AI can help determine the best course of action using a variety of different factors for each individual farmer.
“We want to develop decision support tools to inform farmers and foresters about which climate-smart practices they should apply in a specific area within a large farm or forest at any given time,” said Shekhar. “Large farms and forests will often have to mix-and-match practices depending on the data collected and analyzed. Our tools will allow them to make better decisions based on their specific situation.”
Over the next five years, the Institute will work to refine their deep learning models, estimate carbon in forests and farms across the U.S., and ultimately create a decision support tool for farmers and foresters. This monumental task will be made possible by key partnerships within the U of M - such as the University of Minnesota - Twin Cities College of Science and Engineering (CSE), Minnesota Robotics Institute, CSE Data Science Initiative, College of Food, Agriculture, and Natural Resource Sciences, and the Office of the Vice President for Research - as well as academic and industry collaborations - including Cornell University, Colorado State University, Purdue University, North Carolina State University, Delaware State University, the International Soil Reference and Information Centre, Indigo, and Land-O-Lakes.
With the ultimate goal of mitigating the effects of climate change, the AI-CLIMATE Institute also aims to improve foundational AI methods with their novel approach, develop and educate a diverse AI workforce, and act as a collaboration nexus for different communities to exchange ideas in order to get more computer scientists involved in climate-smart computing.
“We are extremely grateful for this opportunity, and especially that NIFA and NSF have trusted us with this responsibility,” said Shekhar. “We are also very thankful to the Department, College, and University of Minnesota in helping us get this up and running. If we are successful beyond our wildest imagination, hopefully this project will increase the amount of carbon sequestering in farms and forests in order to speed up our net-zero journey. It could have a huge societal impact. Plus, it could advance foundational AI so it can more effectively address the grand challenges facing our changing planet. We are excited by the opportunity to do something transformative.”
Learn more about the AI-CLIMATE Institute.