A data-consistent model of the last glaciation in the Alps achieved with physics-driven AI

25 thousand years ago, the European Alps were covered by the kilometre-thick Alpine Ice Field. Numerical modelling of this glaciation has been challenged by model-data disagreements, including overestimations of ice thickness. We tackle this issue by applying the Instructed Glacier Model, a three-dimensional model enhanced with physics-informed machine learning. This approach allows us to produce 100 Alps-wide and 17 thousand-year-long simulations at 300 m resolution. Previously unfeasible due to computational costs, our experiment both increases model-data agreement in ice extent and reduces the offset in ice thickness by between 200% and 450% relative to previous studies. Our results have implications for better estimating former ice velocities, ice temperature, basal conditions, erosion processes, and paleoclimate in the Alps. This study demonstrates that physics-informed machine learning can help overcome the bottleneck of high-resolution glacier modelling and better test parameterisations, both of which are required to accurately describe complex topographies and ice dynamics.