CORDEX-CORE2 Model Selection Framework for Regional Downscaling

The rapid expansion of the CMIP ensemble provides a broader sampling of Earth system behavior, but it also increases the reliance of regional downscaling and applied modeling on transparent and defensible model sub-selection. This challenge is particularly relevant for CORDEX CORE-2, where modeling centers must downscale only a limited number of Earth System Models (ESMs) despite substantial inter-model differences in historical fidelity, structural redundancy and projections uncertainty. Here we present a structured, multi-criteria framework to guide CMIP6 ESM selection for CORDEX CORE-2 downscaling while maintaining cross-domain consistency through a single shared subset of driving models. We evaluate 45 CMIP6 ESMs across nine CORDEX domains using ERA5 as the reference for 1981–2014, applying an extensive set of common and region-specific diagnostics spanning temperature, precipitation, circulation, humidity, sea level pressure, and sea surface temperature. To reduce redundancy, metric contributions are weighted by uniqueness estimated from inter-metric dependence. Model independence is quantified using cosine similarity of standardized performance vectors to identify highly similar models. Model behavior beyond the historical period is characterized using SSP3–7.0 simulations (2015–2100) for the subset of models with the required outputs, assessing regional temperature and precipitation responses to preserve diversity in both thermodynamic and hydrological behavior. We further test alternative aggregation strategies and evaluate ranking robustness using Monte Carlo convergence experiments, showing that stable rankings typically require a large fraction of the full metric suite. Overall, the framework produces reproducible model rankings and provides a practical pathway for selecting ensembles that balance historical credibility, structural diversity, and representative regional responses, thereby improving the interpretability and utility of CORDEX CORE-2 downscaling experiments.