Evaluating CMIP6 Model Fidelity at Simulating Non-Gaussian Temperature Distribution Tails

Under global warming, changes in extreme temperatures will manifest in more complex ways in locations where temperature distribution tails deviate from Gaussian. Confidence in global climate model (GCM) projections of temperature extremes and associated impacts therefore relies on the realism of simulated temperature distribution tail behavior under current climate conditions. This study evaluates the ability of the latest state-of-the-art ensemble of GCMs from the Coupled Model Intercomparison Project phase six (CMIP6), to capture historical global surface temperature distribution tail shape in hemispheric winter and summer seasons. Comparisons with a global reanalysis product reveal strong agreement on coherent spatial patterns of longer- and shorter-than-Gaussian tails for both sides of the temperature distribution, suggesting that CMIP6 GCMs are broadly capturing tail behavior for plausible physical and dynamical reasons. On a global scale, most GCMs are reasonably skilled at capturing historical tail shape, exhibiting high pattern correlations with reanalysis and low values of normalized centered root mean square difference, with multi-model mean values generally outperforming individual GCMs in these metrics. A division of the domain into sub-regions containing robust shift ratio patterns indicates higher performance over Australia and an overestimation of the degree to which tails deviate from Gaussian over southeastern Asia in all cases, whereas model skill over other regions varies depending on season and tail of the temperature distribution. For example, model performance during boreal winter indicates robust agreement (\u3e85% models) with reanalysis for shorter-than-Gaussian warm tails over the Northern Hemisphere, whereas cold-tail shape is generally mischaracterized by GCMs over western Russia. Although there is spatial and model variability, overall, results highlight the capability of the CMIP6 ensemble in capturing seasonal temperature distribution deviations from Gaussianity, boosting confidence in model utility and providing insight into the complexity of future changes in temperature extremes