Abstract
Abstract. Initial-condition large ensembles with ensemble sizes ranging from 30 to 100 members have become a commonly used tool for quantifying the forced response and internal variability in various components of the climate system. However, there is no consensus on the ideal or even sufficient ensemble size for a large ensemble. Here, we introduce an objective method to estimate the required ensemble size that can be applied to any given application and demonstrate its use on the examples of global mean near-surface air temperature, local temperature and precipitation, and variability in the El Niño–Southern Oscillation (ENSO) region and central United States for the Max Planck Institute Grand Ensemble (MPI-GE). Estimating the required ensemble size is relevant not only for designing or choosing a large ensemble but also for designing targeted sensitivity experiments with a model. Where possible, we base our estimate of the required ensemble size on the pre-industrial control simulation, which is available for every model. We show that more ensemble members are needed to quantify variability than the forced response, with the largest ensemble sizes needed to detect changes in internal variability itself. Finally, we highlight that the required ensemble size depends on both the acceptable error to the user and the studied quantity.
Highlights
To quantify how effective the separation of forced response and internal variability is, we show the root-meansquare error (RMSE) of ensemble means for different ensemble sizes compared to the 200-member mean
If the acceptable error (RMSE) is 0.1 ◦C, 10–30 ensemble members are sufficient over the tropical ocean, while more than 50 ensemble members are required over most land regions
Sampling from a pre-industrial control simulation to estimate the required ensemble size has two advantages: this can be done before producing a large ensemble for the model and is based on a simulation that is available for every climate model in CMIP5 and CMIP6
Summary
Single model initial-condition large ensembles (SMILEs) are a valuable tool for cleanly separating a model’s forced response from internal variability and improving our understanding of the observed trajectory of the climate system in the past, as well as its projected future evolution (Zelle et al, 2005; Deser et al, 2012a; Rodgers et al, 2015; Kay et al, 2015; Maher et al, 2019; Branstator and Selten, 2009; von Känel et al, 2017; Kirchmeier-Young et al, 2017; Frankignoul et al, 2017; Stolpe et al, 2018). Thompson et al (2015) argue that the pre-industrial control simulation can be used to provide a robust estimate of internal variability and represent future internal variability, implying that a single ensemble member for each model may be sufficient. Some studies argue that the pre-industrial control simulation is sufficient to quantify internal variability and no large ensemble is required. This approach only works if the internal variability does not change over time. Realisations 101–200 were added later and use the same configuration as the first 100 realisations but are initialised from different years of the preindustrial control simulation
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
