Machine learning approaches for inferring large-scale sea ice rheology from Lagrangian and Eulerian data

We describe data-driven approaches for nonparametric estimation of large-scale sea ice rheology, combining kernel methods for machine learning with aspects of dynamical systems theory. We study two training scenarios, where data are provided either by daily, gridded observations of sea ice concentration and velocity at 25 km resolution, or by Lagrangian, floe-resolving data generated by an idealized model of sea ice motion which we have developed to simulate interactions between individual floes. Given such data, our machine learning methodology constructs a dynamical model that predicts the effective rheology, as well as other relevant large-scale observables of sea ice. The advantages and limitations of our approach are illustrated with examples.