How I use AI in my geology research without losing the science I love

From ScienceMag:

The maps and colors of my geology work were blurring together in my mind, and it was time for a break. I opened my email and saw the long-awaited result of my application to the 2025 U.S. National Science Foundation (NSF) Graduate Research Fellowship Program: honorable mention. When I checked the awardee list, I noticed an overwhelming trend: computer science, quantum physics, machine learning. Even the few natural science proposals awarded funding seemed more focused on the application of artificial intelligence (AI) than on the actual research question. At first, the tilt toward AI felt like a betrayal, as if the fundamental science was no longer “good enough.” But I have since come to see it as an opportunity.

In my undergraduate studies, I had seen how AI was looming large over our future careers, but it hadn’t made much of an appearance in my earth sciences training. I reveled in studying the tangible things around me, quite unlike the layers of abstractions in AI black boxes. I never envisioned myself using those sophisticated algorithms. When I drafted my fellowship proposal, I followed in the strong tradition of observation, mapping, and statistics.

Then, new federal directives prioritized AI research. Those students who had anticipated the shift were successful. Those of us who worked by the old rules found ourselves left behind.

As I began my Ph.D. not long afterward, my adviser—a computational geophysicist who pushed me to think beyond the tangible into the abstract—and I began to collaborate with an aerospace engineering professor who excelled in the use of machine learning. I began to realize those tools were really an application of the math I already knew. Maybe I, too, could learn to use them.

At first I looked for university courses, but most were locked behind heavy computer science prerequisites. My adviser encouraged me to look beyond the classroom; moving from coursework instruction to experiential learning is key to the Ph.D. journey, after all. I started to design projects to “learn by doing,” but I still needed a guide to get started. That’s when I realized: What better tool to introduce me to AI than AI itself? I began my dive into deep learning by using a large language model to guide me as I installed and trained a machine learning model to tackle an interesting problem in earth sciences.

It was jarring to say the least. Rather than enjoying the beauty of the natural world as I gathered data, I was wading through the harsh computer glare to install deep learning environments. You can walk up to a sinkhole, kick a rock over the edge, and dangle your legs as you enjoy your bagged lunch, but I have yet to do that with any part of an AI model. Every error code and missing dependency added to the frustration.

Slowly but surely, however, the pieces began to come together. In a matter of weeks I had my first working deep learning model, which brought insights I never expected. Before long, I was developing AI tools to identify various geologic landforms from publicly available topographic data. Rather than spending weeks mapping, I delineated the very same structures in a few afternoons. Automation not only eliminated the sampling bias of manual mapping, but it freed me to ask deeper questions about the thousands of natural features I was finding.

It’s now been just over a year since I received that email from NSF. That feeling of disappointment has faded, and I feel empowered. There is still much to learn, but I have gained the confidence to learn tools I never imagined I would use. Most important, I have learned that this evolving world doesn’t have to be the end for the science I love. I can embrace the blocks of code without losing the rocks that brought me here in the first place.

Read More