As climate change accelerates, it is expected to produce stronger and more frequent extreme events, such as droughts. To gain an understanding of how changes in our future climate will impact societally relevant dynamics (e.g., energy, water, agriculture), researchers often integrate Earth System Models (ESMs) with sectoral models (e.g., hydrological models, crop yield models, etc.). To robustly assess the future evolution of extreme events requires large ensembles of stochastic realizations that contain a sufficiently large number of observations to acceptably constrain uncertainty. For this reason, large ESM ensemble products (with 50-100 members) are increasingly being used in climate impacts studies. However, these ensembles often are not bias-corrected and may only have been run for one forcing pathway.

Recent advances in emulation allow for preservation of the spatiotemporal variance and covariance structures, and cross-variable correlation, of key ESM variables such as temperature and precipitation. These capabilities offer a computationally inexpensive means of approximating the output an ESM would have produced had it been run repeatedly for a specific scenario, thus emulating large ensembles of perturbed initial conditions. Here we couple ESM emulation with a global hydrological model to demonstrate the value of Large EMulated ENSembles (LEMENS) for studies of climate impacts on extreme events, using drought as an example. We quantify the statistical advantage that large (versus small) ensembles of ESM emulations offer with respect to projections of extreme (100-year return period) drought events. We also quantify how many emulated ensembles are needed to produce reliable analysis of these extreme drought events, showing that existing large ensembles of 50-100 members may be insufficient in some climate impacts contexts. While drought serves as our primary application, the methodology we apply, and many of the insights we derive, are applicable in other extreme climate impacts contexts that also require multi-model, multi-scenario, bias-corrected extreme characterization within internal variability.