final_program_abstracts[1]

11 IMSC Session Program
The CMIP multi model ensemble and IPCC: Lessons learned
and questions arising
Thursday - Plenary Session 1
Reto Knutti
Institute for Atmospheric and Climate Science, ETH Zurich, Switzerland
Recent coordinated efforts, in which numerous general circulation climate models
have been run for a common set of experiments, have produced large datasets of
projections of future climate for various scenarios. Those multi-model ensembles
sample initial condition, parameter as well as structural uncertainties in the model
design, and they have prompted a variety of approaches to quantifying uncertainty in
future regional climate change. International climate change assessments like IPCC
rely heavily on these models and often provide model ranges as uncertainties and
equal-weighted averages as best-guess results, the latter assuming that individual
model biases will at least partly cancel and that a model average prediction is more
likely to be correct than a prediction from a single model. This is based on the result
that a multi-model average of present-day climate generally out-performs any
individual model. However, there are several challenges in averaging models and
interpreting spread from such ensembles of opportunity.
Among these challenges are that the number of models in these ensembles is usually
small, their distribution in the model or parameter space is unclear and the fact that
extreme behavior is often not sampled when each institution is only developing one or
two model versions. The multi model ensemble should probably be interpreted as a
set of ‘best guess’ models from different institutions, all carefully tuned to the same
datasets, rather than a set of models representing the uncertainties that are known to
exist or trying to push the extremes of plausible model response.
Model skill in simulating present day climate conditions is often weakly related to the
magnitude of predicted change. It is thus unclear how the skill of these models should
be evaluated, i.e. what metric should be used to define whether a model is ‘good’ or
‘bad’, and by how much our confidence in future projections should increase based on
improvements in simulating present day conditions, a reduction of intermodel spread
or a larger number of models. Metrics of skill are also likely to depend on the
question and quantity of interest.
In many probabilistic methods, the models are assumed to be independent and
distributed around the truth, which implies that the uncertainty of the central tendency
of the ensemble decreases as the number of models increases. Because all models are
based on similar assumptions and share common limitations, this behavior is unlikely
to be meaningful at least for a large number of models. Indeed the averaging of
models and the correlation structure suggest that the effective number of independent
models is much smaller than the number of models in the ensemble, and that model
biases are often correlated.
The bottom line is that despite of a massive increase computational capacity and
despite of (or maybe because of) an increase in model complexity, the model spread
in future projections is often not decreasing. Even on the largest scale, e.g. for climate
sensitivity, the range covered by models has remained virtually unchanged for three
decades. Probabilistic projections based on Bayesian methods that determine weights
Abstracts 227