Application of the Ensemble Kalman Filter for Gravity Wave Parameter Estimation

SCHEFFLER, G AND M. PULIDO

Foz Iguazu (Brazil)

Conferencia; The Meeting of the Americas, American Geophysical Union; 2010

American Geophysical Union

 At present, free parameters of physical parameterizations on climate models are tuned arbitrarily in order to reproduce the observed general circulation. Data assimilation provides a conceptual frame for objective parameter estimation. Nevertheless, physical parameterizations usually operate in a completely non-linear way, as a function of their parameters so that they represent a challenge to data assimilation techniques which are in general based on the linear model hypothesis. In this work, we describe the development and application of an efficient parameter estimation technique based on the Ensemble Kalman Filter (EnKF) for a gravity wave drag scheme. We implemented a non-deterministic EnKF, with simple modifications like covariance inflation to avoid filter divergence and analysis smoothing to damp rapid variations of model parameters. Early tests showed that a very low sensitivity in some parameters was found for single time frames. The developed technique proposes the use of independent time frames and therefore the parameterization is used for different wind and temperature profiles. In this case, the physical parameterization is evaluated in different regimes so that a higher sensitivity of the parameters is obtained. Identical twin experiments, in which model-generated synthetic ´data´ are assimilated into the model, and used to evaluate the technique´s potential for accurate parameter estimation and the convergence of the technique to the known parameters. The developed data assimilation technique pretends to overcome, in some way, the drawbacks associated with the nonlinearities of the parameterization. Moreover, we perform a comparison of our results with a standard variational technique. The variational technique does not converge towards the true parameters for arbitrary guess parameters because the cost function is not convex. On the other hand, EnKF does require explicit linearization for obtaining the optimal Kalman gain matrix so that under this hypothesis the minimization problems that occurs in variational assimilation are avoided. To examine if the estimation is robust we performed a series of experiments with different true parameters, covering a physically realistic range of parameters in the control space, we found some dependendencies between errors in estimation and parameters values because nonlinearities in the gravity wave scheme are apparently stronger in certain regions of the parameter space.