Throwing the Uncertainty Toolbox at Antarctica: Multi-model Ensemble Simulation, Emulation and Bayesian Calibration of Marine Ice Sheet Instability

Abstract:

Modelling Antarctic marine ice sheet instability (MISI) - the potential for sustained grounding line retreat along downsloping bedrock - is very challenging because high resolution at the grounding line is required for reliable simulation. Assessing modelling uncertainties is even more difficult, because such models are very computationally expensive, restricting the number of simulations that can be performed. Quantifying uncertainty in future Antarctic instability has therefore so far been limited.

There are several ways to tackle this problem, including:

1. Simulating a small domain, to reduce expense and allow the use of ensemble methods;
2. Parameterising response of the grounding line to the onset of MISI, for the same reasons;
3. Emulating the simulator with a statistical model, to explore the impacts of uncertainties more thoroughly;
4. Substituting physical models with expert-elicited statistical distributions.

Methods 2-4 require rigorous testing against observations and high resolution models to have confidence in their results. We use all four to examine the dependence of MISI in the Amundsen Sea Embayment (ASE) on uncertain model inputs, including bedrock topography, ice viscosity, basal friction, model structure (sliding law and treatment of grounding line migration) and MISI triggers (including basal melting and risk of ice shelf collapse). We compare simulations from a 3000 member ensemble with GRISLI (methods 2, 4) with a 284 member ensemble from BISICLES (method 1) and also use emulation (method 3).

Results from the two ensembles show similarities, despite very different model structures and ensemble designs. Basal friction and topography have a large effect on the extent of grounding line retreat, and the sliding law strongly modifies sea level contributions through changes in the rate and extent of grounding line retreat and the rate of ice thinning. Over 50 years, MISI in the ASE gives up to 1.1 mm/year (95% quantile) SLE in GRISLI (calibrated with ASE mass losses in a Bayesian framework), and up to 1.2 mm/year SLE (95% quantile) in the 270 completed BISICLES simulations (no calibration).

We will show preliminary results emulating the models, calibrating with observations, and comparing them to assess structural uncertainty. We use these to improve MISI projections for the whole continent.