Eingang zum Volltext in OPUS
Hinweis zum Urheberrecht
Dissertation zugänglich unter
Calibrating an earth system model using the adjoint method
Kalibrieren eines Erdsystemmodells mit der adjungierten Methode
Dokument 1.pdf (9.712 KB)
Freie Schlagwörter (Englisch):
climate model , model calibration , data assimilation , adjoint method
38.81 , 38.82
Stammer, Detlef (Prof. Dr.) , Köhl, Armin (Dr.)
Tag der mündlichen Prüfung:
Kurzfassung auf Englisch:
This thesis investigates the potential of the adjoint method for calibrating a climate model. The adjoint method is applied to optimize process parameters on climate timescales to reduce model biases. The difficulty that must be overcome is the limited assimilation window in the adjoint method. Extending the assimilation window longer than the characteristic period of the fastest growing mode, will lead to the occurrence of secondary minima accompanied by an exponential increase of the adjoint sensitivities, and the gradient-descent minimization algorithm is likely trapped into local minima. With a long assimilation window such as climate timescales, the adjoint model cannot provide useful gradients for the optimization. To overcome the limited assimilation window problem, synchronization which is implemented as nudging technique is exploited to regularize the fast-growing modes of the nonlinear system and hence extend the feasible assimilation window for parameter estimation.
Firstly, the performance of this method was investigated based on Lorenz (1963) model. It was shown that: by using a finite nudging coefficient which is strong enough to push the positive Lyapunov exponents to negative values, the feasible assimilation window can be
extended arbitrary and the control parameter can be efficiently and reasonably retrieved. Performance of this method depends on synchronization efficiency which is influenced by observation noise, observation frequency, variables chosen for nudging and nudging strength. With noisy and sparse observations, an optimal nudging coefficient which best recovers true signal can be predefined and benefits the parameter estimation.
Secondly, this method was applied to an intermediate earth simulation model, the PlanetSimulator (PlaSim). I closely examined the usefulness of the adjoint model generated by an automatic differentiation tool TAF. Then identical twin experiments were performed with two different configurations, with and without moisture parameterizations (the ‘maximal’ and ‘minimal’ configurations, respectively). The optimization successfully retrieved the default values of the control parameters for both the two configurations with assimilation window of 2-month and 1-year. At last, the ’maximal’ configuration was used optimize process parameters by assimilating the ERA-Interim data. A number of assimilation experiments using 4,7,16 control parameters and using different observations in the cost function were conducted. The contributions of each parameter to the model state variables were studied in detail. By optimizing two parameters controlling absorptivity (longwave) of clouds and water vapor, the global mean bias of net long wave radiation at the surface and at the top of the atmosphere can be significantly reduced. The global mean bias of short wave radiation at the surface and at the top of the atmosphere can be efficiently reduced by optimizing parameters tuning cloud optical properties. The air temperature is also considerably improved. Then, the estimated parameters were tested with the free model (without nudging terms). The improvements in the radiative fluxes and the air temperature are similar to that in the assimilation experiments which further validate the usefulness of the method. Other model states such as covective precipitation and surface latent heat flux show both improvement and deterioration. However, the specific humidity is hardly improved which is likely due to model deficiency. This study demonstrates that by using synchronization, the adjoint method can be applied to estimate process parameters on climate timescales efficiently. The method overcomes difficulties of parameter estimation in chaotic models and provides a promising way for tuning process parameters in coupled climate models.