The CMIP ensembles represent the most comprehensive source of information available to decision-makers for climate adaptation, yet it is clear that there are fundamental limitations in our ability to treat the ensemble as an unbiased sample of possible future climate trajectories. There is considerable evidence that models are not independent, and increasing complexity and resolution combined with computational constraints prevent a thorough exploration of parametric uncertainty or internal variability. Although more data than ever is available for calibration, the optimization of each model is influenced by institutional priorities, historical precedent and available resources.
The resulting ensemble thus represents a miscellany of climate simulators which defy traditional statistical interpretation. Models are in some cases interdependent, but are sufficiently complex that the degree of interdependency is conditional on the application. Configurations have been updated using available observations to some degree, but not in a consistent or easily identifiable fashion. This means that the ensemble cannot be viewed as a true posterior distribution updated by available data, but nor can observational data alone be used to assess individual model likelihood.
We assess recent literature for combining projections from an imperfect ensemble of climate simulators. Beginning with our published methodology for addressing model interdependency and skill in the weighting scheme for the 4th US National Climate Assessment, we consider strategies for incorporating process-based constraints on future response, perturbed parameter experiments and multi-model output into an integrated framework.
We focus on a number of guiding questions: Is the traditional framework of confidence in projections inferred from model agreement leading to biased or misleading conclusions? Can the benefits of upweighting skillful models be reconciled with the increased risk of truth lying outside the weighted ensemble distribution? If CMIP is an ensemble of partially informed best-guesses, can we infer anything about the parent distribution of all possible models of the climate system (and if not, are we implicitly under-representing the risk of a climate catastrophe outside of the envelope of CMIP simulations)?