“Exploring AI’s Role in Environmental Science”

College of Coastal Georgia student Damian Elmore took a deep dive into the world of artificial intelligence (AI) through a hands-on internship that blended cutting-edge technology with real-world problem solving. This summer, Elmore explored how AI can be used in scientific research and gained insight into one of today’s fastest-growing industries through the WaterSoftHack Fellows Program.

The WaterSoftHack initiative is focused on building a skilled workforce to develop and use advanced digital tools that support water science research and education. It promotes the use of cyberinfrastructure and machine learning, such as AI, to improve how water science is taught and studied. This program was a great for fit Elmore, who is earning a bachelor’s in environmental science with concentrations in marine science and chemistry. He also has interests in science research, physical oceanography, geospatial analysis, and climate chemistry.

Elmore was in the process of applying for jobs when Dr. James Deemy, associate professor of environmental science, learned about the fellowship program and encouraged him to apply. At that time, Elmore was working on a project for the Okefenokee National Wildlife Refuge, which is under threat from proposed local mining. He was exploring whether AI could help address some of the environmental concerns. He used that project in his application to WaterSoftHack and was accepted. Out of more than 250 applications, Elmore was one of 18 accepted into the program—and the only undergraduate student among graduate students, postdoctoral researchers, and professors.

The fellowship was entirely virtual. The first week focused on scientific communication training, the basics of machine learning, and an introduction of different AI models. Participants then divided into groups to develop project ideas.

“We bounced around a few ideas and landed on a project that none of us had ever worked on,” Elmore said. “It was entirely start to finish—what do we want to do? We had to make our own models, make a product, and then present it with a position paper—all within a span of three and a half days.”

Elmore described his group members as “phenomenally talented people.” He was grouped with two postdoctoral students and one doctoral candidate. One group member, Meghana Nagaraj, a postdoctoral research scholar at the University of Central Florida, was instrumental in helping Elmore understand more about AI models. For their project, the group explored how AI can be used to predict and alleviate drought conditions, focusing on the upper Colorado River Basin and its catchment areas. A catchment area is a region of land that drains water into a particular body of water, such as a river, lake, or ocean. The U.S. Geological Survey (USGS) monitors catchments across the country and measures factors such as river flow, height, and water chemistry. Elmore’s group trained two different AI models—a transformer model and a long short-term memory model (LSTM). A transformer model is effective at understanding and generating language, and is the foundation of tools like ChatGPT, Google Translate, and many voice assistants. LSTM helps computers understand and make predictions based on sequences of data, such as time series, or speech.

“I coded the entire transformer model with the assistance of Meg. Meg did the LSTM, and we compared which one worked better. Hers worked a little bit better than mine, just by nature of the model,” Elmore said. “We used different environmental attributes, like how much water is in the soil and other specific data from the USGS, and predicted drought with it.”

The group also examined how much water should be released or saved from dams in the upper Colorado River Basin, such as the Taylor Park Reservoir, based on predicted drought conditions.

“We back dated everything. We made our model predict now, minus four years, as its future. Then, with those conditions, we modified how much water is released in the basin and how much water would be saved knowing about droughts ahead of time,” Elmore said. “We saved them about 6%, which is millions of gallons of water over four years—in theory.”

Elmore and Nagaraj focused on coding the AI model, while the other group members worked on their 26-page position paper. At the conclusion of the two-week program, the fellowship hosted a virtual symposium where the groups presented their projects. Elmore’s group won the Best Project award. Although the fellowship was virtual, Elmore will meet some of his group members in person at the upcoming American Geophysical Union conference in December in New Orleans. He’s excited to thank them face-to-face.

“We’re also hoping to continue the work, and one day, get published,” he said.

Rethinking AI

Participating in the fellowship has reshaped the way Elmore views and uses AI. One of his goals going into the program was to share what he’s learned with his peers and address their concerns. He’s already given two talks to students in a sustainability course about AI and water usage. There has been growing concern about how much water it takes to cool the servers that power AI computations, as well as the amount of water needed to run the facilities that house AI servers. Elmore said AI uses much less water than people think.

“One of the previous models, GPT-3 (Generative Pre-trained Transformer 3), uses a set amount of water per prompt. If that ramps up linearly, by 2027, they are still only using 8 milliliters of water per prompt,” he said. “You’re going to need 60 prompts for a normal glass of water.”

Now that Elmore has peeked behind the curtain of how AI works, he views it more as a helpful tool that people can learn to use more effectively. Many are opposed to AI due to its use of others’ creative work and the potential for students to use it to cheat on assignments. Elmore believes that’s not what AI is designed for.

“Don’t use AI for what it’s not good at, like writing a paper. It has so much data to pull from that it’s going to mix things. It’s like if you read 17 books on the same subject and were told to only write a paper based on one of them—you’re going to mix it up. It’s bad at writing papers, and also, you should write your own papers,” he said. “I use AI as a search engine. I’ll say, ‘Find me this source.’ Then I go read the source and pull out my own conclusions. Or, if I spend a lot of time on computers using applications, I can say, ‘Help me figure out how to click these buttons that I’m looking for.’ Don’t do it for me—help me to teach myself. Also, it’s really good at coding and is made of code, so you can say, ‘Help me revise this code.’ I instruct people to think of AI as a tool, not another person to do your work. If you approach it that way, you’re going to get better results. If you know how to do the thing you’re asking AI to do, then that combo is going to produce a better product.”

The fellowship was also an opportunity to understand the direction of science research. Elmore was fascinated by how data science—using data, statistics, and computer models—is progressing to not just analyze what’s happened, but also to predict what might happen in the future.

“I can use what I’ve learned to predict numbers the same way GPT is used to predict text,” Elmore said. “The first transformer model was announced in a paper in 2018. All of this AI progression has been since 2018. It’s crazy to think about. A few years ago, it was horribly bad at writing anything, and look how much it’s progressed until now. In the future, it’s going to be better at these things,” he said.

Elmore’s favorite part of his internship experience was the hours he spent getting it wrong. Like many people, Elmore doesn’t like to ask others for help. However, he came to realize that part of the process of learning and growing is admitting when you need help.

“I reached out to Meg over a couple of Zoom calls where we were trying to figure it out and it wasn’t working—then it started to work. That had to be my favorite part—the hours of getting it wrong and having to ask questions. It’s a big lesson that I’ve taken away,” he said. “To a certain degree, everyone is on the same playing field, but being the only undergraduate was very daunting. Meg had already coded an LSTM before and she offered to do it, and used a different version this time. She continued her knowledge, but I came in with nothing, which helped me grow a lot and made me comfortable with asking questions of people who already know better, because that’s what science is. It’s asking questions and furthering everyone’s knowledge. That takeaway was my favorite part. I’ve been to a few conferences where it’s just people in suits bullying people that don’t know better. I definitely didn’t experience that here.”

Elmore asked the principal investigator of the fellowship what they looked for in the applicants. She told him that they were looking for individuals who were willing to get things wrong and be okay with it. By nature of the fellowship, they expected people to fail—but learn from the process. Elmore said that the environmental science program at the College adopts the same philosophy.

“Be okay with failing. That has been instilled in our program. You’re not going to get far without failing—which is powerful,” Elmore said.

Damian Elmore and his classmates during their trip to Sapelo Island this summer.

Photo provided by Damian Elmore

Participating in the fellowship is also a great foot in the door for a future career involving AI. Because AI labs are still new, many applications ask for someone to have an interest in AI—not necessarily experience. For Elmore, this internship can help his résumé stand out if he decides to pursue work in an AI lab. That’s why he encourages students to just apply to opportunities that come their way. You never know what experience you can gain just from the process of trying.

“I am ever thankful for the fellowship, outside of the networking opportunities and the people that I’ve met,” he said.

Elmore is set to graduate in fall 2026 and is looking forward to furthering his education in graduate school. With his career interests involving geospatial climate chemistry and oceanography, he hopes to find a graduate program near the water.

“If I do climate chemistry, it will be in the ocean. I do love the geospatial aspect of science. That fascinates me,” he said.

Along with the WaterSockHack Fellowship, Elmore also participated in the scientific diving course this summer and presented his research at the Summer Science Symposium. He worked with students Ollie Mercer and Emma Robison on a project about aquatic vegetation predictions at Phil Foster Park in Palm Beach, Florida. Elmore shared that he had to leave the symposium early to work on more coding for the fellowship project.

Elmore is currently not conducting any independent research, but is helping other students with their projects. He welcomes questions from students about AI, and is a strong advocate for digital safety.