A novel method to test non-exclusive hypotheses applied to Arctic ice projections from dependent models.

Olson, R., S.-I. An, Y. Fan, W. Chang, J.P. Evans and J.-Y. Lee
Nature Communications, 10(1), 3016, doi: 10.1038/s41467-019-10561-x, 2019.


A major conundrum in climate science is how to account for dependence between climate models. This complicates interpretation of probabilistic projections derived from such models. Here we show that this problem can be addressed using a novel method to test multiple non-exclusive hypotheses, and to make predictions under such hypotheses. We apply the method to probabilistically estimate the level of global warming needed for a September ice-free Arctic, using an ensemble of historical and representative concentration pathway 8.5 emissions scenario climate model runs. We show that not accounting for model dependence can lead to biased projections. Incorporating more constraints on models may minimize the impact of neglecting model non-exclusivity. Most likely, September Arctic sea ice will effectively disappear at between approximately 2 and 2.5 K of global warming. Yet, limiting the warming to 1.5 K under the Paris agreement may not be sufficient to prevent the ice-free Arctic.

Key Figure

Figure 4. Probability density functions (pdfs) of global mean surface temperature change required for September Arctic sea ice to effectively vanish. a Pdfs for runs with and without interactions (to illustrate the effects of accounting for model interactions). b Pdfs from all runs accounting for interactions, to illustrate the impact of different assumptions. Vertical dotted line: lower desirable warming limit of 1.5° under the Paris agreement. The projections are sensitive to the datasets used, and to the assumptions about the cause of the recent Arctic sea ice decline. There is a distinct probability that keeping global warming below the 1.5° target of the Paris agreement may not be enough to stave off an essential disappearance of summer Arctic sea ice. The figure illustrates the capacity of the method to make predictions of a variable of interest conditioned on a set of non-exclusive hypotheses while accounting for all orders of hypothesis interactions

UNSW    This page is maintained by Jason Evans | Last updated 23 January 2018