Four Million Years and Counting: Designing and analysed the RAPIT GCM Ensemble
As the concentration of greenhouse gases in the atmosphere grows the risk of catastrophic events, such as the collapse of the AMOC increases.
To guide policy we need to estimate the size of these risks. Estimating climate risks needs a combination of physics and statistics. In this paper
we discuss the methods and results from the RAPIT project looking at the risk of AMOC slow down or collapse. We use a very large perturbed parameter
(or physics) ensemble to investigate the effect of anthropogenic climate change on the strength of the AMOC. The model we use is HADCM3, and we
run a very large (o20000) ensemble on volunteers’ computers using climateprediction.net.
Each ensemble member, with one duplicate, is run with a
different set of values for the uncertain model parameters. We now have almost 5 million years simulated in our ensemble. Even with such a large
ensemble, there are important questions of how we design the experiment. We discuss these issues and why we chose the design we did. Our design
gives us a very wide range of possible HADCM3 climates most of which are not realistic. Using emulators and a measure of model discrepancy, we
use a method called ‘history matching’ to combine the model output and data. Unlike calibration where we try to find the ‘best’ parameter values,
history matching rules out implausible climates. We thus restrict the range of not implausible HADCM3 models. All extra data further reduce the volume
of Not Ruled Out Yet (NROY) space. I will discuss our initial results and show how calibration may allow us to produce better climate models.