Skip to main content Skip to secondary navigation
Page Content

How Fast Will Antarctica’s Ice Sheet Melt?

Autonomous drones that collect data based on scientific machine learning models could play a pivotal role in reducing the uncertainty of sea-level rise.

Image
Scenic photo of an iceberg and pieces of ice in Antarctica

Autonomous drones and algorithms designed to tell us where to take measurements could improve our understanding of ice melt in Antarctica. | iStock/anyaberkut

Warming oceans, melting glaciers, and thinning polar ice caps are expected to cause dramatic sea-level changes that threaten coastal communities around the world. Many cities already are thinking about how to cope with chronic flooding that could bring salt and moisture into homes and infrastructure, compromise drinking water and agriculture, and cause extensive damage to ports.

But how much will sea-level rise, and how fast? Many factors are contributing to the outcome, and the predictive models that scientists have developed so far leave a lot of uncertainty in the projections. One question in particular accounts for most of the uncertainty: At what rate will Antarctica’s ice sheet melt? High uncertainty in the contribution from the southernmost continent means that governments around the world must consider infinite scenarios as they try to prepare for the future.

Working at the intersection of science and engineering, HAI Seed Grant recipient Dustin Schroeder and PhD candidate in electrical engineering Thomas Teisberg, and their collaborator Mykel Kochenderfer, an associate professor of aeronautics and astronautics, wanted to reduce the uncertainty in sea-level models to help policymakers get the answers they need to plan ahead. By combining autonomous drone technology with scientific machine learning, they are poised to reinvent how researchers capture ice sheet data, gathering more and better data, to improve our understanding of the forces at play in a warming climate.

“We want to equip policymakers with information to decide how to adapt, but given the difficulty of gathering data from Antarctica, we can’t survey everything,” Schroeder explains. “We need to focus on collecting the most impactful data. The question of where that data is — or how we would know in a formal way — is a hard, technical, AI-rich problem.”

Faster, Smarter Data Collection

Schroeder, an associate professor of geophysics at Stanford’s School of Earth, Energy & Environmental Sciences who heads up the Stanford Radio Glaciology research group, is a leading expert in ice-penetrating radar. Teisberg has experience with unmanned aerial vehicle (UAV) sensing and perception systems. Combining their interests and skills resulted in a two-part approach to the problem.

First, they set out to design a new data collection platform that would use autonomous drones equipped with ice-penetrating radar to take measurements more efficiently. “The current process involves going to Antarctica for months at a time and flying around in World War II airplanes or setting up expensive field camps in the middle of the ice sheet,” Schroeder says. “We envisioned UAVs providing a long-term monitoring approach that’s sustainable and almost entirely automated.”

The second part of their project focused on determining where to find the most valuable data: What if custom algorithms could tell researchers when and where to send the drones to maximize the impact of their work?

Image
an autonomous drone could better capture ice melt in Antarctica

Scientific machine learning models determine where the new drones will find the most helpful data. | Courtesy of Thomas Teisberg

Seeing under the Ice

The list of variables associated with ice sheet melting is long. Researchers want to know how deep the ice bed is, whether it’s frozen or thawed, and the actual temperature of the ice. They also need to understand how tides, seasons, and the passing of time affect the rate of melting. Add to this complexity, the fact that after 50 years of surveying by countries all over the world, some of the closest measurements taken on the continent can be spaced about three to five miles apart, with many places hundreds of miles from the nearest known measurement. The result is a far-from-complete dataset for a region that represents 5.5 million square miles of the globe.

“So far, every time we collect data, we discover something new,” Teisberg says. “We have better models for the ice caps on Mars than we have from Antarctica.”

Since Antarctica is so massive, decisions must be made about prioritizing where and when to collect information that will lower the uncertainty bar. This is where AI comes in. Schroeder and Teisberg are applying scientific machine learning models, also called physics-informed neural networks, to determine where the new drones are likely to find the most helpful data. These initial models incorporate known rules of physics that govern how ice reacts to environmental factors and apply those rules to small datasets. In this way, the algorithms can run fast and produce recommendations quickly. 

Ultimately, this team envisions an iterative cycle, or adaptive surveying process, through which the models process each new batch of data in real time to inform a continuously evolving flight plan. By embracing a data-driven approach, scientists will be able to consider the measurements of the ice bed, the 3D flow of the ice, and the physics that govern its movement — all at once.

Echoes from Mountain Glaciers

The HAI Seed Grant gave Schroeder and his team the kick start they needed to design the proxy machine learning models and begin work on the UAV-based radar system. Getting the drone to find the bottom of an ice sheet will be the first measure of success, according to Teisberg. “We’re pushing fast to do field testing over mountain glaciers that have tens of meters to hundreds of meters of ice, to be sure the system works. From there, we’ll incorporate the ML models into how we plan the flights — something that has never been done before for ice sheet research.”

On the ML side of the project, the team has started focusing on ice shelves, which have interesting dynamics on a smaller scale than the polar ice sheet. If the models can learn and predict how rapid changes cause dramatic impacts with ice shelves, this will be a promising step toward tackling the larger challenge of ice sheets.

Transforming Glaciology

Sea-level rise is not the only factor in climate change, but for now it carries the most uncertainty. Cities will be better prepared to handle sea rises if they can predict the change with accuracy. They can build higher levies, increase the height of ports, or plan for a managed retreat, if needed. 

With their two-part approach, Schroeder and Teisberg hope to improve researchers’ ability to monitor and predict changes to this ice sheet and help these cities prepare for the future. “Over the long term, this work could transform how all glaciologists collect and understand data,” Schroeder says, “but for now, we’re focused on making ice sheet research smarter and more efficient, so that we can have a better understanding of how the warming of Antarctica will affect future sea-level rise.”

Stanford HAI's mission is to advance AI research, education, policy and practice to improve the human condition. Learn more

More News Topics