Towards of workflow for ESM Tuning using Perturbed Parameter Ensembled: Some perspectives from NASA/ GISS and LEAP

Perturbed parameter ensembles (PPEs) are increasingly acknowledged as an essential tool for tuning Earth System Models (ESMs) owing to the fact that they, in conjunction with machine learning emulators, allow for the estimation of dozens of parameters with some semblance of computational affordability. However, there are few examples of successful use of PPEs to tune large sets of ESM parameters, and there is certainly no consensus on how ESM PPEs should be constructed, what emulators should be built from them, and how observations should be used to ultimately select plausible parameters. Additionally, there has yet to be clear answers to whether and how process-level insights and prior knowledge from high-resolution models can be unified with tuning using global observations and ESM simulations. We present the methodology used to tune 45 parameters in the NASA GISS ModelE3 using 36 global satellite metrics as constraint. We explain the motivation behind choices made, some lessons learned and potential improvements to our methodology, and the important role of accounting for observational uncertainties. We will motivate the community to progress from PPEs to Calibrated Physics Ensembles (CPEs): ensembles of model parameters whose uncertainty has been constrained and quantified by common sets of satellite observations with each member close to radiative balance. We argue that CPEs provide a more meaningful basis upon which to compare CMIP model uncertainty than PPEs. Finally, we highlight the potential use of PPE tuning methodologies to drive satellite mission design by quantifying information gain on uncertain physics components of ESMs – a hitherto unrealized capability.