Thinner Ice, Faster Flow: AI Redraws Alps’ Ice Sheet Dynamics

How a Swiss team slashed computational time by 99% to reveal the Alps’ thinner, faster-flowing ice sheet—and why it matters for predicting sea level rise.

Research: A data-consistent model of the last glaciation in the Alps achieved with physics-driven AI. Image Credit: FamVeld / Shutterstock

Scientists at the University of Lausanne (UNIL) have used physics-driven AI to massively speed up computer calculations and simulate the last ice cover in the Alps. The new results, which are much more in line with field observations, show that the ice volume was 35% lower, and the thickness mismatch with geological evidence improved by 200–450% compared to earlier models. This innovative method opens the door to countless new simulations and predictions linked to climate upheavals. The research is published in Nature Communications.

Twenty-five thousand years ago, the Alps were covered by a layer of ice up to two kilometers thick. For almost fifteen years, 3D digital models based on climate reconstructions, thermodynamics, and ice physics have put this glaciation into perspective. However, these models have sparked debate in the scientific community. Until now, there has not been a full correspondence between these simulations and the physical traces - rocks, moraines, etc. - found in the field, particularly trimlines (erosion markers indicating past ice surfaces), which bear witness to past ice thicknesses.

A team of scientists from the University of Lausanne (UNIL) have just solved this persistent problem. For the first time, they have used artificial intelligence trained on ice physics principles to massively boost their new glacial evolution model, running 100 simulations with varied parameters to generate a large series of results of unprecedented accuracy: they correspond much more closely to the physical traces left on the ground. The study confirms that trimlines reflect true ice surface elevations, resolving a long-standing debate about their origin. Their results show a 35% reduction in total ice volume compared to previous reference simulations, with ice thickness discrepancies reduced by up to 450%. Model resolution has been increased from two kilometers to 300 meters, and it is only thanks to this precision that it is possible to describe the complex topography of the Alps numerically.

Here, the simulation is compared against the Parallel Ice Sheet Model (PISM) 2 km simulation by Jouvet et al., and displayed by showing modelled Last Glacial Maximum (LGM) ice thickness (a, b), ice surface velocity (c, d), and basal topography (e, f) fields in the main Rhein valley (Alpenrhein). Ice thickness and velocity fields are plotted above a 30 m digital elevation model of the local topography, also shown in g. h Indicates the location and direction of the three-dimensional view shown in other panels. More versions of this figure are available for other regions of the Alps in Supplementary information. All panels were produced using ArcGIS Pro 3.2.2 (Esri), and panels a–d, and g, h include the 30 m digital elevation model from ALOS World 3D DEM (version 4.1). h Features geographical data from OpenStreetMap.

The simulations also reveal localized ice flow acceleration in narrow valleys due to the "venturi effect," which previous models missed. This explains the thinner ice cover in key regions like the Rhein and Garda valleys. Based on field observations, this breakthrough, which is in line with the current state of scientific knowledge, shows, for example, that certain peaks, such as the Hörnli nunatak (near the Matterhorn), were clearly protruding from the ice during the Ice Age.  

The research is significant in more ways than one. Firstly, the ability to correctly model the glacial past is essential to understanding our environment. For over 2 million years, the Earth has experienced alternating glacial and warm cycles, which have profoundly shaped the landscape in which we live. The new model now corresponds much more closely to the evidence left on the ground following the retreat of the glaciers and makes it possible to better quantify many natural phenomena, such as glacial erosion, which has contributed mainly to sculpting the relief of the Alps.

On the other hand, the innovative methodology used in this research marks a new era in numerical modeling. "By using recent technology and applying it to the last major glaciation in the Alps, we can finalize 17,000-year simulations at very high resolution (300 m) in 2.5 days, whereas such spatial resolution would have taken 2.5 years to calculate using traditional methods, which are also extremely costly and energy-intensive", explains Tancrède Leger, researcher at UNIL's Faculty of Geosciences and Environment (FGSE), and first author of the study.

With this approach, the model first learns about the physics of ice flow using physics-informed deep learning methods. It then receives data on the climate of the period (temperature, precipitation, etc.) to calculate ice supply and melt.

Deep learning calculations are then performed not by the traditional central processing unit (CPU), but via a GPU (or graphics processing unit). This enables numerous operations to be performed in parallel, boosting the computer's computing power phenomenally.

"We’ve gone from large supercomputers to a single GPU, achieving resolutions once deemed impossible," illustrates Guillaume Jouvet, FGSE professor behind the AI model and co-first author of the study. "By exploring 100 parameter combinations, we ensured our results are robust and capture the Alps’ complex ice dynamics."

This progress will enable new research. In particular, a new SNSF-funded project is about to begin using this revolutionary method to better predict the impact of the melting Greenland and Antarctic ice sheets on the rise of the global sea level.