Bryan Riel , Brent Minchew

Floating ice shelves that fringe the coast of Antarctica resist the flow of grounded ice into the ocean. One of the key factors governing the amount of flow-resistance provided by an ice shelf is the rigidity of the ice that constitutes it. Ice rigidity is highly heterogeneous and must be calibrated from spatially-continuous surface observations assimilated into an ice flow model. Moreover, realistic uncertainties in calibrated rigidity values are needed to quantify uncertainties in forecasts of future shelf flow. Here, we present a physics-informed machine learning framework for inferring the full probability distribution of rigidity values for a given ice shelf, conditioned on surface velocity and thickness fields derived from remote sensing data. We employ variational inference to jointly train neural networks and a variational Gaussian Process to reconstruct surface observations and rigidity values and uncertainties. Application of the framework to synthetic and large ice shelves in Antarctica demonstrate that rigidity is well-constrained in areas where deformation of ice is measurable within the noise level of the observations. Further reduction in uncertainties can be achieved by complementing variational inference with conventional inversion methods. Our results demonstrate a path forward for continuous calibration of ice flow parameters from remote sensing observations.