Calibrating a Greenland ice sheet model using historical simulations between 1996-2022

Future projections of ice loss from the Greenland ice sheet are subject to large and often poorly quantified uncertainties. These arise both from uncertainties in climate forcing projections and poorly constrained processes in ice sheet models. Ice-sheet model initialisation, in particular, is a major contributor to projection uncertainty. Here, we aim to calibrate a Greenland-wide ice sheet model configuration that replicates the recent trend (1996--2022) in observed changes in ice speed and thickness. First, we generate an ensemble of simulations using the ice-sheet model Úa, each forced with datasets of surface mass balance and ice front positions, and input parameter values sampled from prior probability distributions. This ensemble is then used to train a surrogate model, designed to emulate the temporal- and spatially integrated combined misfit between observed and modelled changes in ice speed and thickness. We then use this emulator for Bayesian inference to determine the posterior model parameter distributions needed to minimise the misfit between observed and modelled quantities of interest and ultimately best replicate the observed trend in Greenland ice sheet mass loss. By calibrating the model in such a way, we can reduce the uncertainties in forward-projections and have confidence in the predictive capabilities of our model.