Climate models are extremely sophisticated computer programs that encapsulate our understanding of the climate system and simulate, with as much fidelity as currently feasible, the complex interactions between the atmosphere, ocean, land surface, snow and ice, the global ecosystem, and a variety of chemical and biological processes.
The complexity of climate models—the representation of physical processes like clouds, land surface interactions, and the representation of the global carbon and sulphur cycles in many models—has increased substantially since the IPCC First Assessment Report in 1990, so in that sense, current Earth System models are vastly ‘better’ than the models of that era. This development has continued since the Fourth Assessment, while other factors have also contributed to model improvement. More powerful supercomputers allow current models to resolve finer spatial detail. Today’s models also reflect improved understanding of how climate processes work—understanding that has come from ongoing research and analysis, along with new and improved observations.
Climate models of today are, in principle, better than their predecessors. However, every bit of added complexity, while intended to improve some aspect of simulated climate, also introduces new sources of possible error (e.g., via uncertain parameters) and new interactions between model components that may, if only temporarily, degrade a model’s simulation of other aspects of the climate system. Furthermore, despite the progress that has been made, scientific uncertainty regarding the details of many processes remains. An important consideration is that model performance can only be evaluated relative to past observations, taking into account natural internal variability. To have confidence in the future projections of such models, historical climate—and its variability and change—must be well-simulated. The scope of model evaluation, in terms of the kind and quantity of observations available, the availability of better coordinated model experiments, and the expanded use of various performance metrics, has provided much more quantitative information about model performance. But this alone may not be sufficient. Whereas weather and seasonal climate predictions can be regularly verified, climate projections spanning a century or more cannot. This is particularly the case as anthropogenic forcing is driving the climate system toward conditions not previously observed in the instrumental record, and it will always be a limitation.
Quantifying model performance is a topic that has featured in all previous IPCC Working Group I Reports. Reading back over these earlier assessments provides a general sense of the improvements that have been made. Past reports have typically provided a rather broad survey of model performance, showing differences between model-calculated versions of various climate quantities and corresponding observational estimates. Inevitably, some models perform better than others for certain climate variables, but no individual model clearly emerges as ‘the best’ overall. Recently, there has been progress in computing various performance metrics, which synthesise[sic] model performance relative to a range of different observations according to a simple numerical score. Of course, the definition of such a score, how it is computed, the observations used (which have their own uncertainties), and the manner in which various scores are combined are all important, and will affect the end result.
Final Draft (7 June 2013) Chapter 9 IPCC WGI Fifth Assessment Report