Diagnosing the causes of bias in climate models – why is it so hard?

Abstract

The equations of climate are, in principle, known. Why then is it so hard to formulate a bias-free model of climate? Here, some ideas in nonlinear dynamics are explored to try to answer this question. Specifically it is suggested that the climatic response to physically different forcings shows a tendency to project onto structures corresponding to the systems natural internal modes of variability. This is shown using results from complex climate models and from the relatively simple Lorenz three-component model. It is suggested that this behaviour is consistent with what might be expected from the fluctuation–dissipation theorem. Based on this, it is easy to see how climate models can easily suffer from having errors in the representation of two or more different physical processes, whose responses compensate one another and hence make individual error diagnosis difficult. A proposal is made to try to overcome these problems and advance the science needed to develop a bias-free climate model. The proposal utilises powerful diagnostics from data assimilation. The key point here is that these diagnostics derive from short-range forecast tendencies, estimated long before the model has asymptotically settled down to its (biased) climate attractor. However, it is shown that these diagnostics will not identify all sources of model error, and a so-called “bias of the second kind” is discussed. This latter bias may be alleviated by recently developed stochastic parametrisations.