 |
| Titel |
"Supermodelling" by Adaptive Synchronization of Different Climate Models |
| VerfasserIn |
Gregory Duane, Frank Selten, Noel Keenlyside, Wim Wiegerinck, Juergen Kurths, Ljupco Kocarev |
| Konferenz |
EGU General Assembly 2011
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 13 (2011) |
| Datensatznummer |
250057748
|
|
|
|
|
|
| Zusammenfassung |
| Climate models of the class used by the IPCC give divergent results
in regard to the magnitude of average temperature change in response
to increased greenhouse gas levels, and in regard to regional projections.
One would like to surpass the skill of averaged model outputs by using
the best features of each model in run time. The supermodelling approach
extends the "interactive ensemble" previously used for ENSO prediction,
by introducing a set of connections between select pairs of corresponding
variables in a small suite of different models. With weights on these
connections obtained by training on 20th century data, the models tend
to synchronize with one another as well as with reality.
Supermodelling has been tested with simple systems of ODE's and with
quasigeostrophic models. It was previously shown that quasigeostrophic
models with forcing in different sectors could be fused as a supermodel
of a "real" system with a combination of forcings. Here we show that
the weights needed to produce the correct supermodel can be found adaptively
using a "real" training set. The advantage of the approach is that only
a small set of trainable connections need be considered (as compared to
the large number of parameters in each model) but such connections can
be selected in a naive manner.
The approach is controversial in part because, even for a simple set of ODE's,
the connections found by the learning procedure may define a skillful supermodel
with no clear physical interpretation. That situation arises
from a simple degeneracy: Even if one model has a perfect equation for
a given dynamical variable, the learning scheme may choose to combine other
models to reproduce the "perfect" behavior. A learning scheme that favors
binary-valued weights can generate a more interpretable supermodel. The
tradeoff between flexibility and the possibility of physical interpretation
is studied.
The approach is also controversial because it is not clear that training
on 20th century data will define connections that are optimal for 21st
century greenhouse gas levels. It is shown with simple models that connections
found by the learning procedure tend to be robust against
large variations in the parameters of individual models and resulting
bifurcations. Also, it is demonstrated that the approach applies to systems with
widely different intrinsic time scales, such as atmosphere-ocean models.
Supermodelling will therefore be useful for at least a significant
fraction of the 21st century, will counter tribalism in our
field, and will produce a verifiable consensus on details of climate change. |
|
|
|
|
|
|