Uncertainties in climate projections exist due to natural variability, scenario uncertainty, and model uncertainty. It has been argued that model uncertainty can be decreased by giving more weight to those models in multimodel ensembles that are more skillful and realistic for a specific process or application. In addition, some models in multimodel ensembles are not independent. We use a weighting approach proposed recently that takes into account both model performance and interdependence and

more »

... apply it to investigate projections of summer maximum temperature climatology over North America in two regions of different sizes. We quantify the influence of predicting diagnostics included in the method, look at ways how to choose them, and assess the influence of the observational data set used. The trend in shortwave radiation, mean precipitation, sea surface temperature variability, and variability and trend in maximum temperature itself are the most promising constraints on projections of summer maximum temperature over North America. The influence of the observational data sets is large for summer temperature climatology, since the observational and reanalysis products used for absolute maximum temperatures disagree. Including multiple predicting diagnostics leads to more similar results for different data sets. We find that the weighted multimodel mean reduces the change in summer daily temperature maxima compared to the nonweighted mean slightly (0.05-0.45 ∘ C) over the central United States. We show that it is essential to have reliable observations for key variables to be able to constrain multimodel ensembles of future projections. Key Points: • Model weighting slightly reduces summer warming signal over central North America • More than one predicting diagnostics should be used to inform the weighting • Shortwave radiation trend, mean precipitation, and SST variability are possible constraints on projections of summer maximum temperature Supporting Information: • Supporting Information S1 Correspondence to: R. Lorenz, ruth.lorenz@env.ethz.ch Citation: Lorenz, R., Herger, N., Sedláček, J., Eyring, V., Fischer, E. M., & Knutti, R. (2018). Prospects and caveats of weighting climate models for summer maximum temperature projections over North America. in the CMIP5 archive encounter too frequent phases with small evapotranspiration and high temperatures. Excluding those models from the CMIP5 multimodel mean resulted in reduced biases and lower absolute values in extreme temperature projections. Hence, models that overestimate the influence of the land surface on the atmosphere are more likely to have positive absolute temperature and variability biases in the historical period. Christensen and Boberg (2012) found a connection between positive temperature biases in the warm period of the year and larger temperature change into the future. Therefore, models with larger positive biases in absolute temperature and variability in the historical period are more likely to overestimate future temperature change. Hence, by constraining temperature biases in the historical period, we can potentially be more confident in our projections into the future. More generally, uncertainties in climate projections originate from three main sources: (1) natural variability, (2) model uncertainty, which is a combination of the fact that models can never completely represent the reality and differences in how the models approximate it, and (3) scenario uncertainty (Hawkins & Sutton, 2009) . It is not possible to reduce the uncertainty due to natural variability (Deser et al., 2012) on time scales more than a few years, but we can estimate how large the natural variability itself is and can therefore account for it. The scenario uncertainty is also difficult to reduce since societal decisions, innovation, and technological progress do not follow physical laws but are largely choices that society makes. In the Intergovernmental Panel for Climate Change (IPCC) Assessment Reports this is taken into account by creating several possible scenarios which span a range of plausible economic pathways. Based on the corresponding greenhouse gas (and other) emissions, climate models make projections. This leaves us with model uncertainty as our best starting point, if we want to reduce the uncertainty in future projections and increase the reliability thereof. Figure 3. Linear regressions for tasmaxCLIM in the North America region. Grey lines and R 2 correspond to Ordinary Least Squares regression, while black lines and R 2 correspond to Theil-Sen regression method. The red lines (first two columns) and the red triangles (third and fourth columns) are the observational and reanalysis estimates (lines because we do not have observations for the delta terms). and 8e). Including more than two diagnostics generally suggests a larger ΔtasmaxCLIM in the northern part of the domain and a smaller ΔtasmaxCLIM in the southern part of the domain in the weighted multimodel mean. However, only when using MERRA2 is I 2 smaller than 1 in these cases.