Climate models contain closure parameters which can act as effective
“tuning handles” of the simulated climate. These appear in physical
parameterization schemes where unresolved variables are expressed by
predefined parameters rather than being explicitly modeled. In the
current climate model tuning process, best expert knowledge is used to
define the optimal closure parameter values, based on observations,
process studies, large eddy simulations, etc. Questions arise, however:
“Do the parameters represent observable quantities?”, or “Do the
parameters depend on the model discretization, such as grid interval?”
In fact, closure parameters span a low-dimensional space, and the
parameter probability densities can be objectively estimated
simultaneously for all relevant parameters.
Several estimation methods can be applied. Here, adaptive parameter
estimation, based on the Monte Carlo Markov Chain (MCMC) method, is
introduced. Initial experimentation with the ECHAM5 climate model is
presented. The results obtained so far show that:
(a) Adaptive MCMC is technically relatively easy to implement,
(b) It is computationally quite expensive, and
(c) The key question for successful estimation in case of ECHAM5, or any
climate model for that matter, is the definition of a so-called
objective function.
The adaptive MCMC sampling tends to converge towards an equilibrium
probability density corresponding to a chosen Bayesian objective
function: this is a “criterion” against which the short climate
simulation is assessed. A simple objective function can be, for
instance, the global net radiation which should be close to zero in a
“good” climate simulation. This alone does not necessarily constrain the
parameters sufficiently: near zero net radiation can be obtained by
using an unrealistic parameter combination. Additional terms must be
added therefore to the objective function, constraining the simulation
more tightly to the observed regime. [This research is funded by the
Academy of Finland for 2010-2013.]
17.02.2010
13:30 Uhr