Abstract. Multi-model ensembles (MME) are key ingredients for future climate projections and the quantification of their uncertainties. Developing robust protocols to design balanced and complete computer experiments for MME is a matter of active research. In this study, we take advantage of a large-size MME produced for Greenland ice sheet contributions to future sea level by 2100 to define a series of computer experiments that are closely related to practical MME design decisions: what is the added value of including a specific set of members in the projections, i.e. either adding new models (Regional Climate Model, RCM, or Ice Sheet Model, ISM) or extending the range of some parameter values? We use these experiments to build a random-forest-based emulator, whose predictive capability to assess Greenland sea level rise contributions in 2100 proves very satisfactory for low and high levels of warming but less effective for intermediate levels. On this basis, we assess the changes in the emulator's predictive performance, both in terms of prediction accuracy and uncertainty, and the emulator-based probabilistic predictions, in terms of changes in the 17th, 50th and 83rd percentiles, for given temperature scenarios, compared to the reference solution built using all members. For the considered MME, several aspects are outlined: (1) the highest impact of removing the most selected RCM, i.e., MAR, due to the large number of simulations available; (2) the significant impact of excluding the SSP5-8.5 scenario for high temperature scenarios, and of the Community Ice Sheet Model (CISM) for low temperature scenarios leading to absolute changes up to 30 % of the high and low percentiles respectively; (3) the non-negligible impact of having a MME designed with a unique ISM or a unique RCM, i.e., CISM or MAR model in our case, leading to percentile absolute changes ranging between 10 % and 20 % compared to the reference solution; (4) the lesser importance of the choice in the range of the Greenland tidewater glacier retreat parameter. These results point to the size of the training set as the key driver of the changes, which supports the need for large ensembles to develop accurate and reliable emulators, hence encouraging large participation to projects such as the Ice Sheet Model Intercomparison Project ISMIP. We also expect our recommendations to be informative for the design of next generations of MME, in particular for the next Ice Sheet Model Intercomparison Project in preparation (ISMIP7).
Rohmer et al. (Tue,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: