Ensemble consistency testing using causal connectivity

Understanding differences in climate model
output can be challenging in light of the chaotic nature
of the climate system and its inherent variability.
While progress has been made in automatically detecting
small changes, making quantitative assessments of when
changes truly affect the climate state in the model -
and thus indicate a potential problem - has remained
elusive. We present a first step in this direction based on
evaluating changes in the connectivity structure among
key model variables. We use a collection (ensemble) of
climate model simulations to obtain a graphical model
of the relationships of 15 key climate variables and
then build a statistical model to probabilistically describe
the occurrence of these relationships. This statistical
model forms the basis of a test to evaluate new runs.
We illustrate our methodology using data from a large
publicly available ensemble of climate model simulations.