Guillaume Cordonnier
University of Zurich, Inria Université Côte d'Azur, ETH Zurich, University of Alaska Fairbanks - guillaume.cordonnier@inria.frG Jouvet, G Cordonnier, B Kim, M Lüthi, A Vieli and A Aschwanden
Modeling glaciers, icefields or ice sheets requires a complex coupled simulation of the ice dynamics, transport and mass balance. We will present novel, learned surrogate models that estimate the ice flow by a Convolutional Neural Network, which is trained from data generated with state of the art ice flow models. By doing so, the most computationally demanding model component is substituted by a cheap emulator. Once trained with representative data, we demonstrate that out model permits to model mountain glaciers up to 1000 × faster than Stokes one with fidelity levels above 90% in terms of ice flow solutions leading to nearly identical transient thickness evolution. Addtionally, we will show recent results that demonstrate how variational formulations can help to bridge the gap between learned models and classical solvers for viscous fluids.
Guillaume Cordonnier is a research scientist in Computer Graphics, in the Graphdeco research group at Inria d’Université Côte d’Azur. He received his Ph.D. from Université Grenoble Alpes on multidisciplinary topics at the edge between computer science and geomorphology, which was granted the best Ph.D. award from the CNRS GdR IGRV. His research interests cover the inverse control of the simulation of natural phenomena, which brings several interdisciplinary collaborations between computer science, especially machine learning and VFX, with geology, botany, CFD or paleontoglogy.