Quantifying uncertainty in climate predictability using perturbed physics ensembles and climate model emulation

Climate models are essential tools for understanding and predicting Earth system processes and feedbacks, but uncertainties in their future projections remain challenging to characterize. Improvements in physical process realism and the representation of human influence arguably make models more comparable to reality, but also increase the degrees of freedom in model configuration leading to parametric and structural uncertainties in projections. Perturbed physics ensembles sample the uncertainty space through different choices of parameter settings. Climate model emulators can be a computationally efficient method of producing large ensembles of climate model output, in order to study different sources of uncertainty. In this work we use a machine learning algorithm to build an emulator for the land surface component of a climate model. Using a perturbed physics ensemble of model simulations, we train the emulator to predict model output given a set of parameter values as input. We optimize parameter values by comparing emulated model output with observations across multiple relevant metrics, including global carbon and water flux benchmarks. We also account for structural and observational uncertainty through a novel Bayesian calibration approach. By sampling the resulting posterior distributions and running future climate simulations, we can then estimate the contribution of land model parameter uncertainty in future projections of climate change.