Dependence of Inferred Climate Sensitivity on the Discrepancy Model

TL;DR

This study examines how different temporal error models affect the estimation of climate sensitivity from earth system models, finding that AR(1) is often appropriate and that error structure choice impacts uncertainty estimates.

Contribution

It introduces a comparative analysis of various discrepancy models in climate sensitivity inference, highlighting the importance of selecting suitable temporal error structures.

Findings

ECS estimates vary with discrepancy model choice.

AR(1) often provides a better fit than IID or Gaussian processes.

Uncertainty in Gaussian process parameters is notably high.

Abstract

We consider the effect of different temporal error structures on the inference of equilibrium climate sensitivity\footnote{ECS is defined as the realized equilibrium surface warming---globally-averaged surface air temperature---for a doubling of CO$_2$}(ECS), in the context of an energy balance model (EBM) that is commonly employed in analyzing earth system models (ESM) and observations. We consider error structures ranging from uncorrelated (IID normal) to AR(1) to Gaussian correlation (Gaussian Process GP) to analyze the abrupt 4xCO$_2$ CMIP5 experiment in twenty-one different ESMs. For seven of the ESMs, the posterior distribution of ECS is seen to depend rather weakly on the discrepancy model used suggesting that the discrepancies were largely uncorrelated. However, large differences for four, and moderate differences for the rest of the ESMs, leads us to suggest that AR(1) is an appropriate discrepancy correlation structure to use in situations such as the one considered in this article. Other significant findings include: (a) When estimates of ECS (mode) were differrent, estimates using IID were higher (b) For four of the ESMs, uncertainty in the inference of ECS was higher with the IID discrepancy structure than with the other correlated structures, and (c) Uncertainty in the estimation of GP parameters were much higher than with the estimation of IID or AR(1) parameters, possibly due to identifiability issues. They need to be investigated further.