Toward a Simple Probabilistic GCM Emulator for Integrated Assessment of Climate Change Impacts

Friday, 19 December 2014: 2:25 PM
Ian Sue Wing1, Claudia Tebaldi2, Douglas W Nychka3 and Jordan Winkler1, (1)Boston University, Boston, MA, United States, (2)Climate Central, Princeton, NJ, United States, (3)NCAR, Boulder, CO, United States
Climate emulators can bridge spatial scales in integrated assessment in ways that allow us to take advantage of the evolving understanding of the impacts of climate change. The spatial scales at which climate impacts occur are much finer than those of the “damage functions” in integrated assessment models (IAMs), which incorporate reduced form climate models to project changes in global mean temperature, and estimate aggregate damages directly from that. Advancing the state of IA modeling requires methods to generate—in a flexible and computationally efficient manner—future changes in climate variables at the geographic scales at which individual impact endpoints can be resolved.

The state of the art uses outputs of global climate models (GCMs) forced by warming scenarios to drive impact calculations. However, downstream integrated assessments are perforce “locked-in" to the particular GCM x warming scenario combinations that generated the meteorological fields of interest—it is not possible assess risk due to the absence of probabilities over warming scenarios or model uncertainty. The availability of reduced-form models which can efficiently simulate the envelope of the response of multiple GCMs to a given amount of warming provides us with capability to create probabilistic projections of fine-scale of meteorological changes conditional on global mean temperature change to drive impact calculations in ways that permit risk assessments.

This presentation documents a prototype probabilistic climate emulator for use as a GCM diagnostic tool and a driver of climate change impact assessments. We use a regression-based approach to construct multi-model global patterns for changes in temperature and precipitation from the CMIP3 archive. Crucially, regression residuals are used to derive a spatial covariance function of the model- and scenario-dependent deviations from the average pattern. By sampling from this manifold we can rapidly generate many realizations of spatially-coherent fields of temperature and precipitation, conditional on future global mean temperature change. We close by outlining the application of our emulator to characterizing climate change risk to global agriculture.