AI’s Growing Role in Glaciology

Dartmouth last month hosted the first international academic conference focused on how artificial intelligence can accelerate scientific discovery in glaciology.

Whether it’s from satellite imagery, drone surveys, or ice-core measurements, scientists who study Earth’s glaciers and ice sheets are collecting data faster and in greater quantities than they can fully analyze.

The International Symposium on Artificial Intelligence in Glaciology brought together more than 60 scientists from the United States, Europe, Asia, and Australia on June 23-27 to explore the value of AI in examining and making sense of immense amounts of complex data. Attendees also discussed the need to build trust in these powerful new tools.

In a Q&A with Dartmouth News, Mathieu Morlighem, the Evans Family Distinguished Professor of Earth and Planetary Sciences and one of the event’s lead organizers, discussed the key takeaways from the meeting, which was hosted by the International Glaciological Society and Dartmouth’s Changing Polar Regions academic cluster.

How significant is this symposium in the field of glaciology?

This is the very first international meeting dedicated to AI in glaciology. We had about 60 talks and posters (PDF) spanning the full breadth of the field: ice-sheet dynamics, ocean and sea ice, surface and subglacial hydrology, and more. The symposium captured a field-wide moment, bringing together glaciologists from all around the world. We are at an important turning point: AI is exploding in our field, so the time was right to step back and discuss the current limitations and frontiers, and how to move forward together.

What makes Dartmouth a good host for the meeting, and what is the significance of it being hosted on campus?

I cannot think of a better place than Dartmouth for this symposium. This is where the field of artificial intelligence was born 70 years ago, and Dartmouth has always had a strong focus on polar science, with faculty and researchers across schools on campus. Hosting the symposium here felt very special: it brought the international field to campus and raised Dartmouth’s visibility at the intersection of climate science and AI.

In what ways is AI transforming and enhancing glaciology? What advances and challenges/concerns of using AI were addressed at the symposium?

AI is transforming how we do research across all the sciences. It can handle very large datasets that are sometimes impractical to tackle with classical approaches. It is breaking long-standing computational bottlenecks by training emulators or models that can approximate complex simulations, extracting information from sparse and messy data, automating mapping at scales no human effort could match, and even discovering physical relationships directly from data.

Of course, there is also some healthy skepticism toward these new techniques. It is not yet clear whether fast emulators can be trusted, especially for future projections that involve a warmer climate that the emulators have not necessarily been trained for. The uncertainty associated with these methods remains poorly quantified, and machine learning is often described as a “black box.” It can be hard to tell whether these models are doing the right things for the right reasons. Much of the symposium was dedicated to group discussions aimed at identifying the current frontiers and developing recommendations for the community.

Which topics or presentations at the symposium do you think were especially important to the field and why? Which presentations sparked the most conversation?

There are applications where machine learning is simply necessary, and we have no real alternative. One example is ice delineation—outlining the extent of glaciers from millions of satellite images—to document glacier retreat around the world over the past decades. Doing this manually would be impossible, so it is a great use of AI.

The session that generated the most discussion was on emulating glacier flow as an alternative to classical, physics-based but slower models. Emulators are model mimics that are orders of magnitude faster and can run very long, high-resolution simulations, but there is also a risk that they “hallucinate” on occasion, and it remains unclear how much trust we should place in them.

What was the most exciting or innovative use of AI you saw showcased at the symposium?

What is blowing me away is the performance of emulators. We can now model the flow of an entire ice sheet in a few minutes, whereas classical methods would take days on big supercomputers. These advances make it possible to simulate ice sheets over far longer time periods than ever before, opening the door to new science questions and new hypotheses to test.

How is AI used in your and/or your research group’s work? Are there particular studies or findings that AI made possible that wouldn’t have been before?

Several members of my group are using AI in their research. Mansa Krishna, a fourth-year Guarini graduate student, and Gong Cheng, a former research scientist, now on the faculty at Tongji University, have worked on physics-informed neural networks to infer Greenland’s subglacial topography and other properties we cannot observe directly. They developed a package called PINNICLE that is now used by several groups around the world.

Yinmin Liu, a postdoc who joined the group more recently, is working on coupling machine-learning emulators into the Ice-Sheet and Sea-Level System Model, our ice-sheet model. His work will make it possible to use an emulator for any part of the model, for example, to represent a physical process that is currently missing or poorly parameterized. These projects are all very new and very exciting.

What is the general usage of AI in polar research on campus, and how do you see it growing or evolving in the future?

The use of AI in polar research on campus is growing across modeling, remote sensing, and data assimilation. More students are being trained in these methods, and AI is becoming a standard part of the modeling toolkit rather than “a hobby.” The questions around trust and uncertainty are becoming the central research frontier. We are working hard to better quantify uncertainty, to build confidence in these tools, and to understand when they work and when we should simply not use them.

Relatedly, how do you see AI defining and expanding the research of students and early-career students in the years to come? How do you think they should be prepared for this?

What surprised me at this symposium was the number of students and early career scientists. The new generation of glaciologists is very excited about AI, whereas more senior scientists are sometimes more critical or skeptical about what it can do in our field.

What came up again and again in the discussions is that it is essential for students to become familiar both with traditional methods grounded in mathematics and physics and with machine-learning techniques, so they can bridge the two communities. AI should be part of the toolkit, not an end in itself, so we need to train students well—fluent in established physics-based methods, in how to use machine learning alongside them, and in the limitations of both.

It is an exciting time to be a student!