Calibration of E3SMv3 for Low and High Equilibrium Climate Sensitivity Using a Machine Learning Surrogate Model

We use machine learning methods to identify alternative calibrations (parameter tunings) for the Department of Energy's Energy Exascale Earth System Model version 3 (E3SMv3) that yield low or high equilibrium climate sensitivity (ECS). Earth system model (ESM) code is usually released with selected values for parameters that are, in fact, uncertain. We generate a perturbed parameter ensemble of more than 400 members using E3SMv3's atmospheric model, varying 14 uncertain atmospheric parameters through Latin Hypercube Sampling and a subsequent adaptive sampling approach. A high-dimensional surrogate model is then trained and optimized to predict multiple plausible parameter sets that are consistent enough with present-day satellite observations and reanalysis of 46 climate variables. We evaluate the calibrated solutions in E3SMv3's fully coupled model configuration, and show that some atmosphere-only calibrations perform well in the coupled model while also diverging from the default E3SMv3 ECS. Our selected low-ECS and high-ECS calibrations have ECS of 2.69 K and 4.67 K, respectively, compared to the default E3SMv3 ECS of 3.89 K.