Climate model Selection by Independence, Performance, and Spread (ClimSIPS v1.0.1) for regional applications

Abstract

As the number of models in Coupled Model Intercomparison Project (CMIP) archives increase from generation to generation, there is a pressing need for guidance on how to interpret and best use the abundance of newly available climate information. Users of the latest CMIP6 seeking to draw conclusions about model agreement must contend with an "ensemble of opportunity" containing similar models that appear under different names. Those who used the previous CMIP5 as a basis for downstream applications must filter through hundreds of new CMIP6 simulations to find several best suited to their region, season, and climate horizon of interest. Here we present methods to address both issues, model dependence and model subselection, to help users previously anchored in CMIP5 to navigate CMIP6 and multi-model ensembles in general. In Part I, we refine a definition of model dependence based on climate output, initially employed in Climate model Weighting by Independence and Performance (ClimWIP), to designate discrete model families within CMIP5 and CMIP6. We show that the increased presence of model families in CMIP6 bolsters the upper mode of the ensemble's bimodal effective equilibrium climate sensitivity (ECS) distribution. Accounting for the mismatch in representation between model families and individual model runs shifts the CMIP6 ECS median and 75th percentile down by 0.43 degrees C, achieving better alignment with CMIP5's ECS distribution. In Part II, we present a new approach to model subselection based on cost function minimization, Climate model Selection by Independence, Performance, and Spread (ClimSIPS). ClimSIPS selects sets of CMIP models based on the relative importance a user ascribes to model independence (as defined in Part I), model performance, and ensemble spread in projected climate outcome. We demonstrate ClimSIPS by selecting sets of three to five models from CMIP6 for European applications, evaluating the performance from the agreement with the observed mean climate and the spread in outcome from the projected mid-century change in surface air temperature and precipitation. To accommodate different use cases, we explore two ways to represent models with multiple members in ClimSIPS, first, by ensemble mean and, second, by an individual ensemble member that maximizes mid-century change diversity within the CMIP overall. Because different combinations of models are selected by the cost function for different balances of independence, performance, and spread priority, we present all selected subsets in ternary contour "subselection triangles" and guide users with recommendations based on further qualitative selection standards. ClimSIPS represents a novel framework to select models in an informed, efficient, and transparent manner and addresses the growing need for guidance and simple tools, so those seeking climate services can navigate the increasingly complex CMIP landscape.