Prioritizing CMIP6 Models and their Ensemble Members for Hydroclimate Downscaling over the Conterminous United States

Because of the coarse (~100 km) resolution of CMIP6 models, projections of local changes in temperature and precipitation require sub-grid process information that can come from downscaling. Given the mismatch between the substantial personnel and computational resources required per model to implement either statistical or dynamical downscaling solutions and the sheer number of models and ensemble members for each experiment in CMIP6, only a small fraction of the archive can be downscaled. A judicious and parsimonious approach or set of approaches is needed to ensure that the models and ensemble members that are downscaled ultimately produce a fair measure of the range of projected hydroclimate states. Specifically, we test the assumption that the selection of a single ensemble member for downscaling from each model is sufficient for capturing multi-model ensemble statistics. We calculate the RMSE of initial condition (IC) ensemble members relative to historical observations and multi-model ensemble averages and compare the ratio of between-model to within-model variance within this metric over the Conterminous United States (CONUS) and National Climate Assessment (NCA) regions. For historical simulations, this ratio is much greater than one, but for end-of-century simulations, this ratio falls to less than one across the CONUS and NCA regions, especially for higher emissions scenarios such as SSP585. These results indicate that more ensemble members per model need to be downscaled for higher emissions scenarios to achieve an unbiased multi-model ensemble for hydroclimate downscaling. Regionally-resolved Taylor diagrams reveal where IC ensemble members differ the most and, consequently, where more IC ensemble member downscaling efforts must be focused.