Benchmarking FastIce, a new massively parallel thermomechanical ice flow solver

This is the authors' abstract. We don't add key points for this paper.

Efficient modeling of ice sheets involves considering multiple coupled physical processes, including thermomechanical interactions. While using a Full-Stokes model for ice flow in Greenland and Antarctica provides most accurate results, it is can be extremely costly on a large scale with existing software, justifying the use of various reduced models. Some important features of ice sheets, such as ice streams, are inherently three-dimensional. Accurate stress distribution, particularly around topography features comparable to ice thickness and near a grounding line, can only be achieved with a full stress tensor. In addition, a reference Full-Stokes solver for regional to ice sheet scale simulations can be a valuable tool for calibrating reduced models. We introduce FastIce, a novel ice flow model for massively parallel architectures, written in Julia. Leveraging GPUs (Nvidia, AMD) and supporting distributed computing, FastIce includes a thermo-mechanically coupled Full-Stokes ice flow model and a novel conservative energy formulation for describing thermal effects. FastIce is written to be easily extensible, and its core is fully differentiable, enabling data assimilation pipelines using adjoint sensitivities and automatic differentiation (AD). We validate FastIce through ISMIP-HOM benchmark tests and assess the coupled thermomechanical solver using the method of manufactured solutions. Our results showcase the thermo-mechanical instability arising from the non-linear interaction between temperature-dependent viscosity of ice and shear heating, reproducing existing analytical results. We present the performance testing of FastIce in single-node and distributed scaling benchmarks on LUMI, the largest European supercomputer.