A Grand Design for Future Ice Sheet Projections

Abstract:

If we are to make robust projections of the probability of sea level rise from the ice sheets, these must be founded in both glaciological theory, represented by dynamical ice sheet models, and statistical inference, i.e. formal uncertainty quantification (UQ). No such studies yet exist for either the Greenland or Antarctic ice sheet. Therefore projections risk being physically implausible, difficult to interpret, or both.

More optimistically, ice sheet models have many advantages over climate models with respect to uncertainty quantification. They are computationally cheaper, simpler to understand, and have fewer input parameters and output variables. It is relatively straightforward to switch between different model structures, such as physics approximations, basal drag laws, and resolution. Moreover the ice sheet modelling community is relatively small and is not constrained methodologically or culturally by the legacy - and pitfalls - of the CMIP multi-model "ensemble of opportunity".

These advantages present us with a golden opportunity to create a new vision for policy-relevant sea level projections: we can design a grand ensemble that quantifies multiple modelling uncertainties in a statistically rigorous and efficient way. Such an ensemble systematically samples model parameters and structures, initial and boundary conditions, in the most informative way given the available resources and also allows statistical inference.

I will review some of the UQ steps that have been taken in ice sheet modelling, such as Bayesian calibration of parameters and projections, history matching (statistically-formalised ruling out of poor parameter values), and emulation (statistical surrogates of numerical models). I will then describe how to implement these and other techniques in a multi-model ensemble design, including: sequential experimental design, which is more efficient than the usual Latin Hypercube; quantifying uncertainty in Full Stokes and high resolution models with the help of cheaper models; how to calibrate models with observations; and how emulation can improve all of these. The proposed design will not only generate more robust and meaningful sea level projections but also provide thorough sensitivity analyses to prioritise future model development and observational campaigns.