CSU climate researchers tied to new $20 million NSF AI center

Imme Ebert-Uphoff and Libby Barnes
Imme Ebert-Uphoff of the Department of Electrical and Computer Engineering and CIRA at CSU and Libby Barnes in the Department of Atmospheric Science

As partners in a five-year, $20 million NSF-funded program led by the University of Oklahoma, CSU will work to greatly expand how AI is used in environmental research, with a critical focus on making sure that the answers we get are not only accurate and fast, but trustworthy.

Research Professor Imme Ebert-Uphoff, working in both the Department of Electrical and Computer Engineering and CIRA at CSU, is leading CSU’s delegation in the new NSF AI Institute for Research on Trustworthy AI in Weather, Climate, and Coastal Oceanography announced Wednesday. Ebert-Uphoff is joined in the project by Atmospheric Science Associate Professor Elizabeth Barnes, Professor Chuck Anderson of the Department of Computer Science, and other CSU researchers, focusing on AI algorithm development, environmental applications, and workplace education and advocacy.

Artificial intelligence is everywhere under the skin of our modern world – our cars have sensors that detect traffic patterns and avoid collisions. Our phones recognize our faces and automatically unlock. Modern retail relies on artificial intelligence for stocking and logistics – everything from medical supplies to the tomatoes in the grocery store are now governed by algorithms that recognize patterns and guide purchasing and shipping decisions. And opportunites to apply AI in research fields abound.

The importance of artificial intelligence is such that the National Science Foundation introduced for the first time a $100 million program to fund institutes to study and develop newer and better AI algorithms – by far and away, the most significant direct commitment to AI that the NSF has ever provided. Five such institutes were funded under the program, including the OU-led program that CSU has partnered with.

Piecing together a complicated relationship

Visible (left), infrared (middle), and microwave (right) image of Hurricane Isaias from July 31, 2020. Microwave imagery is extremely helpful in determining internal structure of tropical storms, but has inconsistent coverage of individual storms. AI techniques are allowing the use of consistent visible and IR images to recreate the microwave imagery. (credit: RAMMB/CIRA/CSU)
Visible (left), infrared (middle), and microwave (right) image of Hurricane Isaias from July 31, 2020. Microwave imagery is extremely helpful in determining internal structure of tropical storms, but has inconsistent coverage of individual storms. AI techniques are allowing the use of consistent visible and IR images to recreate the microwave imagery. (credit: RAMMB/CIRA/CSU).

Hurricanes are tricky creatures to forecast. Complex structures, made up of dozens of growing or decaying thunderstorm cells, rotating at a high rate of speed around a central low pressure system, all of which is interacting with ocean temperatures, winds at all levels of the atmosphere, and low- and high-pressure systems, sometimes hundreds of miles away, make for a challenging forecast problem to solve.

One tool that helps untangle this knotted mess is satellite remote sensing. Especially useful are observations taken from the microwave spectrum of light, which can tell us a lot about a storm’s dynamics invisible to the naked eye, but satellites that can provide this imagery are few and far between. Forecasters might get one or two pictures of a particular storm each day, hours apart. Other satellites can track these storms down to the minute, 24/7 – but these satellites use visible or infrared imagery and don’t carry the special microwave imagers forecasters really want.

This is where artificial intelligence steps in – helping to piece together the (often-complex) relationships between what a storm looks like in microwave imagery and in visible or infrared light, ‘recognizing’ what a microwave image would look like for a storm by only looking at that storm’s infrared imagery. What’s more, by not needing to explicitly solve those complex relationships between the two kinds of imagery, getting the microwave ‘answer’ also becomes fast – fast enough that forecasters can work on it in real time, making AI an increasingly valuable forecast tool.

Chris Slocum, a research scientist working with NOAA’s National Environmental Satellite, Data, and Information Service, developed just such an algorithm while working at CSU’s Cooperative Institute for Research in the Atmosphere (CIRA).

Atmospheric scientists, like Slocum are just starting to tap AI’s potential. Others, like Barnes, are using neural networks to study and predict climate. A neural network is a series of algorithms that can efficiently sort through data and recognize patterns. Barnes’ group trains neural networks to identify changes in Earth’s environmental systems, compared to natural variability. They also apply this machine-learning technique to prediction. Neural networks can sift through a sea of data to create skillful forecasts on the challenging subseasonal-to-decadal scale.

Advances in AI capability

Long a research topic in computer science, the algorithms and theories underlying artificial intelligence have seen remarkable growth in complexity and capability. With research stretching back to 1980s, Professor Chuck Anderson of the Department of Computer Science at CSU has helped guide this growth over the decades. “Our research in AI over 30 years has taught me that the key to collaboration is understanding how to translate AI methods into the language of the discipline” says Anderson. “In the current collaboration of the CSU team this common ground has and will continue to lead to custom AI methods that reveal novel answers to key questions in environmental science.”

The use of artificial intelligence in the environmental sciences, however, has taken longer to catch on. To the uninitiated, artificial intelligence can seem like a ‘black box’ – one where observations go in, answers come out, and the workings in-between a mystical and even inscrutable process. Even to scientists, who generally delight in figuring out the in-between as part of what science means, AI can seem suspicious. Which is where education and support come in.

Applications for weather and climate

Ebert-Uphoff has long championed the view that atmospheric scientists and AI researchers need to collaborate closely to achieve meaningful AI algorithms for weather and climate applications. “It is crucial for the atmospheric scientist to have an intuitive understanding of how the AI algorithm derives its output, just as much as the AI expert needs to have an intuitive understanding of the application, in order to derive innovative and robust solutions.” This is particularly important for the field of physics-guided machine learning which seeks to merge physical knowledge and AI methods to gain the best of both worlds.

The new institute will also create more opportunities to develop and promulgate AI techniques and products. At CSU, researchers will lead two ‘use-case’ studies, seeking to better understand tropical storms and prediction of severe weather events at short- and long-term time scales.

Most importantly, however, the center will look at how the algorithms are used and understood, using social science tools developed by partner agencies to provide context. AI tools that provide good answers are the starting point – AI tools that inspire confidence in methods and procedures are the goal.

With partners from the University at Albany, the University of Washington, North Carolina State University, Texas A&M Corpus Christi, Del Mar College, the National Center for Atmospheric Research, Google, the IBM Weather company, NVIDIA Corporation, Disaster Tech, and NOAA, CSU will have plenty of opportunities to engage with, learn from, and advocate for sophisticated artificial intelligence use in environmental research.

“In recent years, AI has been proven to be a potential game changer for many applications in the environmental sciences,” said Ebert-Uphoff.  “Now is the time to focus on how to use AI responsibly in these areas, namely to build on our current work using explainable AI for weather and climate to develop methods of trustworthy AI that are dependable and whose reasoning can be explained to the stakeholders.”

With hurricane season upon us, the impacts of prolonged drought and severe weather impacting wildfires nationwide, and a growing need to understand climate, weather, and our role in both, the need for these tools is increasingly apparent. And they will surely become even important in the years to come. And researchers at CSU stand at the forefront of making these tools more useful and accessible to all.