Transient calibration of the Amundsen Sea Embayment: a model and methodology comparison

It is important that modellers be able to make reasonable projections of ice-sheet loss over the next 50-100 years so that costal planners can make informed decisions. Despite many innovations in modelling and satellite observing, several recent studies show that the agreement between models and the observational record remains poor -- raising concerns about their ability to predict responses to changes in climate on decadal to centennial time scales. A common means to address model-data misfit is via assimilation of satellite data, in which poorly constrained parameters are chosen in a way to minimize this misfit. Data assimilation has been employed extensively, including by many of the models participating in the ISMIP6 intercomparison. However, they are limited to observations at a single point in time (a "snapshot"). Such inversions do not take advantage of the time series of velocity and altimetry observations currently available -- largely due to the complexity and computational expense. The use of Automatic Differentiation makes such approaches possible. The method -- termed "transient assimilation" or "transient calibration" -- provides a physically consistent, time-dependent model which agrees with time-resolved observations.

But with such a development, questions arise: how does transient calibration impact future modelled behaviour? Which types of observations most strongly constrain models? Are results consistent across different models? Here we apply transient calibration to the Amundsen Sea Embayment using two independent models, the Ice-sheet and Sea-level System Model (ISSM) and the MITgcm STREAMICE model, using time-resolved velocities and surface altimetry from 2004 to 2017. We assess the performance of transient compared to snapshot calibrations in terms of capturing past and current trends in speed change, thinning, and grounded ice loss; and we additionally run 50-year projections using the calibrated models. While overall 50-year mass loss is not strongly dependent on assimilation strategy, the rates of mass loss vary greatly, suggesting greater differences on the century time scale or longer. Moreover, we see that the runs calibrated with altimetry: (i) agree best with recent mass loss rates; (ii) show the greatest conformity between models; and (iii) show the largest mass loss rates at the end of the 50-year runs. The results likely have implications for optimal data needs and assimilation strategies in the next generation of ice-sheet models.