 |
| Titel |
Using gradient descent with a perfect model of the rotating annulus |
| VerfasserIn |
Roland Young, Roman Binter, Falk Niehörster, Leonard Smith, Peter Read |
| Konferenz |
EGU General Assembly 2010
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 12 (2010) |
| Datensatznummer |
250044397
|
|
|
|
|
|
| Zusammenfassung |
| The existence of shadowing trajectories, model trajectories consistent with a sequence of past
observations of a system, is a desirable property of any predictive model. Techniques for
finding such trajectories are, to some extent, understood in low-dimensional systems such as
the Lorenz equations, and there is significant interest in their application to high-dimensional
systems such as general circulation models (GCMs). Gradient descent of indeterminism is
one such technique, and it is well-established for finding shadowing trajectories
of low-dimensional analytical systems. It has also been used for state estimation
in some limited experiments with an idealised quasi-geostrophic model (Judd et
al. 2004, Physica D, 190, 153) and with the NOGAPS weather model (Judd et al.
2008, J. Atmos. Sci., 65, 1749). Its ability to find model trajectories that shadow
observations for a long time in high-dimensional models of real systems is not yet well
explored.
We apply gradient descent to a model of the thermally-driven rotating annulus laboratory
experiment, a system intermediate in model complexity and physical idealisation between
low-dimensional analytical systems and high-dimensional GCMs. We report on some
preliminary results from applying gradient descent to the annulus under the controlled
conditions of the perfect model scenario. Starting from a sequence of noisy observations
combined with a model, our demonstration shows that a sequence of states close to the true
system trajectory is recovered.
The laboratory setting allows investigation of the properties of gradient descent using a
real physical system and a non-idealised model in a situation where the complexity of the
flow can be controlled, the experiments can be repeated, and where there is potential for
long-range observations under laboratory conditions. We hope to extend this work to
laboratory data, which will allow us to determine how well models of the rotating annulus
reproduce observations, and hence improve our fundamental understanding of this system.
We also hope to inform the use of gradient descent with high-dimensional GCMs by
determining whether it is a feasible method for finding shadowing trajectories in a real
physical system. |
|
|
|
|
|
|