Detecting and identifying the impact of parameter interaction on climate model outputs based on two Perturbed Parameter Ensembles (PPEs)

Climate model Perturbed Parameter Ensembles (PPEs) help inform how the parameters affect climate model outputs, and are used to train surrogate models for purposes such as parameter estimation. It is common and relatively easy to identify important model parameters from analyzing the PPEs, but how the parameters jointly (i.e., parameter interaction) affects the model outputs (e.g., cloud forcing and precipitation) is not always fully explored. In this work, we apply a simple additive emulator method to analyze two sets of PPEs from two climate models, i.e., CESM2 CAM6 and GISS ModelE3, with a focus on the role of parameter interaction on the model outputs. We identify and list parameter groups that greatly affect the model outputs averaged at different temporal and spatial scales. Among the parameters within such groups, some of them could seem deceivingly unimportant if their interaction with others are not considered, which is not uncommon in previous studies. We find that the model outputs at different spatial and temporal scales respond differently to the parameter interaction. Such information could help determine more effective metrics for climate model parameter estimation. We also gain insights on PPE design from this work. For example, the number of ensemble members in a set of PPE matters. With less than 200 ensemble members, the impact of parameter interaction will be a lot more difficult to capture compared to working with 500 ensemble members. Our work provides a paradigm for studying parameter interaction in climate model PPEs. It allows more information to be extracted from the analyzed PPE.