Non-Newtonian viscous fluid models with learned rheology accurately reproduce Lagrangian sea ice simulations

Abstract

Polar sea ice is crucial to Earth's climate system. Its dynamics also affect coastal communities, wildlife, and global shipping. Sea ice is typically modeled as a continuum fluid using a model proposed almost 50 years ago, which is moderately accurate for packed ice, but loses its predictive accuracy outside of the central ice pack. Discrete element methods (DEMs), which are commonly used for modeling granular media, offer an alternative by resolving the behavior of individual ice floes, including collisions, frictional contact, fracture, and ridging. However, DEMs are generally too costly for large-scale simulations. To address this, we present a framework for inferring rheological behavior from DEM velocity data. We characterize isotropic constitutive laws as scalar functions of the principal invariants of the strain-rate tensor. These functions are parameterized by neural networks trained on DEM data. By combining machine learning and finite element methods, we incorporate the governing partial differential equation (PDE) into the training, requiring to solve a PDE-constrained optimization problem for the network parameters. We focus on unidirectional parallel shear flows, which allow us to infer the effective shear viscosity. We find that, over a wide range of ice concentrations, the velocity fields observed in a complex sea ice DEM can be captured by a nonlinear rheology. Depending on the ice concentration, a shear-thinning or a shear-thickening behavior is observed. Moreover, the effective shear viscosity is found to increase by several orders of magnitude with changes as small as 5% in the sea ice concentration. We show that the learned rheology generalizes to different forcing scenarios, time-dependent problems, and settings in which compressibility is not a dominant factor.