The moving target of simulating ENSO: A timeline of Community Earth System Model version 3 (CESM3) Development

Abstract. The process of developing Earth System Models (ESMs) varies across modeling centers globally, but the overarching goal is largely the same – to improve the representation of physically-based processes such that biases in the mean climate state and its variability are minimized. Developers face a number of scientific and technical hurdles to ensure the best use of limited computational resources, storage space, and time. These challenges are particularly pronounced for modes of climate variability that are characterized by high internal variability, like the El Niño Southern Oscillation (ENSO). In this study, we leverage the development of the Community Earth System Model version 3 (CESM3) as a case study to illustrate some of these difficulties as well as to highlight new model developments and their impacts on ENSO.

ENSO is a complicated indicator of model performance in that it can be characterized by a wide number of metrics to assess how well (or poorly) it is simulated. This is at odds with the model development process as a whole, which requires that a manageably small number of metrics be selected for model analysis as teams conduct hundreds of simulations to arrive at a single "best" model configuration. Selecting too many metrics runs the risk of dramatically slowing progress in developing a coupled ESM. As a result, we discuss a minimal set of ENSO metrics here that we consider indicative of overall model performance. We find that biases in the spatial extent of ENSO events persisted throughout the development cycle, with sea surface temperature anomalies (SSTa) that extend too far into the West Pacific for all simulations. Other metrics are more sensitive to model changes, including ENSO amplitude and duration. Such metrics are, however, prone to significant internal variability. This is confirmed by temporally sub-sampling long pre-industrial control simulations of previous model versions (CESM1 and CESM2), which also adds critical context to the evaluation of CESM3; any new model version should ideally not be markedly worse than past iterations. Ultimately, CESM3 produces a reasonable ENSO in comparison to previous model versions and relative to observations, but it remains difficult to attribute changes in its representation to individual model changes due to the significant internal variability.