Non‐stationary return levels of CMIP5 multi‐model temperature extremes

During the period 1979–1992, on average nearly 400 people each year were killed in the United States by excessive heat (NOAA 1995; Kilbourne 1997). In fact, in this period over the United States, excessive heat accounted for more reported deaths annually than hurricanes, floods, tornadoes, and lightning combined (NOAA 1995). Furthermore, agriculture products such as wheat, rice, corn and maize can be significantly reduced by extreme high temperatures at key development stages (NOAA 1980; Hoerling et al. 2013). High temperatures also affect irrigation and evaporation (Sorooshian et al. 2014), drought development (Aghakouchak et al. 2014), energy production and consumption (Tarroja et al. 2014b) as well as greenhouse gas emissions associated with energy production (Tarroja et al. 2014a). Numerous studies indicate that temperature extremes are likely to intensify in the future under different plausible climate scenarios (Alexander et al. 2006; IPCC 2007). Climate model simulations have been widely used to study extreme weather and climate across different spatial and temporal scales. Recently, international climate modeling groups have provided Coupled Model Intercomparison Project Phase 5 (CMIP5) historical and projected climate simulations (Taylor et al. 2012). The scope of CMIP5 also is broader than previous model intercomparison projects (e.g., CMIP3), with carbon emission-driven Earth system model (ESM) experiments now represented along with the typical concentration-driven atmosphere–ocean general circulation model (AOGCM) simulations (Meehl and Bony 2011). Thus, the multi-model gridded CMIP5 Abstract The objective of this study is to evaluate to what extent the CMIP5 climate model simulations of the climate of the twentieth century can represent observed warm monthly temperature extremes under a changing environment. The biases and spatial patterns of 2-, 10-, 25-, 50and 100-year return levels of the annual maxima of monthly mean temperature (hereafter, annual temperature maxima) from CMIP5 simulations are compared with those of Climatic Research Unit (CRU) observational data considered under a non-stationary assumption. The results show that CMIP5 climate models collectively underestimate the mean annual maxima over arid and semi-arid regions that are most subject to severe heat waves and droughts. Furthermore, the results indicate that most climate models tend to underestimate the historical annual temperature maxima over the United States and Greenland, while generally disagreeing in their simulations over cold regions. Return level analysis shows that with respect to the spatial patterns of the annual temperature maxima, there are good agreements between the CRU observations and most CMIP5 simulations. However, the magnitudes of the simulated annual temperature maxima differ substantially across individual models. Discrepancies are generally larger over higher latitudes and cold regions.