Polar predictability: exploring the influence of GCM and regional model uncertainty on future ice sheet climates

Abstract:

Evaluating uncertainty in GCMs and regional-scale forecast models is an essential step in the development of climate change predictions. Polar-region skill is particularly important due to the potential for changes affecting both local (ice sheet) and global (sea level) environments through more frequent/intense surface melting and changes in precipitation type/amount. High-resolution, regional-scale models also use GCMs as a source of boundary/initial conditions in future scenarios, thus inheriting a measure of GCM-derived externally-driven uncertainty.

We examine inter- and intramodel uncertainty through statistics from decadal climatologies and analyses of variability based on self-organizing maps (SOMs), a nonlinear data analysis tool. We evaluate a 19-member CMIP5 subset and the 30-member CESM1.0-CAM5-BGC Large Ensemble (CESMLE) during polar melt seasons (boreal/austral summer) for recent (1981-2000) and future (2081-2100, RCP 8.5) decades. Regional-model uncertainty is examined with a subset of these GCMs driving Polar WRF simulations.

Decadal climatologies relative to a reference (recent: the ERA-Interim reanalysis; future: a skillful modern GCM) identify model uncertainty in bulk, e.g., BNU-ESM is too warm, CMCC-CM too cold. While quite useful for model screening, diagnostic benefit is often indirect.

SOMs extend our diagnostics by providing a concise, objective summary of model variability as a set of generalized patterns. Joint analysis of reference and test models summarizes the variability of multiple realizations of climate (all the models), benchmarks each model versus the reference (frequency analysis helps identify the patterns behind GCM bias), and places each GCM in a common context.

Joint SOM analysis of CESMLE members shows how initial conditions contribute to differences in modeled climates, providing useful information about internal variability, such as contributions from each member to overall uncertainty using pattern frequencies. In the climatological mean, ensemble members show local temperature anomalies of up to +/- 2 °C relative to the ensemble average. SOM pattern frequency analyses show interquartile ranges for individual patterns of 2-4% across the ensemble. Both are potentially significant sources of melt-prediction uncertainty.