Advances in modeling high-dimensional uncertainty in ice-sheet dynamics
Please login to view abstract download link
The evolution of the Greenland and Antarctic ice sheets impact sea-level change, with potentially profound consequences for coastal communities and infrastructure. Accurate projections of ice-sheet mass balance require not only sophisticated dynamical models but also a rigorous treatment of uncertainties arising from observational data and model formulations. In this presentation, we explore state-of-the-art methods for calibrating Greenland and Antarctic ice sheet models through the inversion of high-dimensional model parameters. These methods utilize large-scale Partial Differential Equation (PDE)-constrained optimization techniques and employ Bayesian inference [1] to efficiently approximate the posterior distribution of inferred parameters. We detail the computational aspects of our scalable and portable implementation, which is based on algorithmic differentiation (AD), adjoint methods and fast solvers. Additionally, we address challenges related to the proper characterization of uncertainty, discussing various choices of prior and likelihood distributions and their effects on the posterior distribution and the forward propagation of uncertainty. Finally, we present results from uncertainty quantification studies enabled by efficient sampling of the parameters' posterior distribution.
