RETRIEVAL OF TURBULENT COHERENT STRUCTURES IN A CONVECTIVE PLANETARY BOUNDARY LAYER USING A VARIATIONAL FOUR DIMENSIONAL DATA ASSIMILATION TECHNIQUE

Remote sensing techniques, such as radar and lidar, can provide near-real-time two-dimensional or quasi three-dimensional high spatial resolution data of planetary boundary layer (PBL), but these data are limited to radial velocity and reflectivity. On the other hand, large-eddy simulation (LES) technique is able to generate detailed homogeneous PBL flow structures with periodic boundary conditions. A four dimensional data assimilation based on the calculus of variations and optimal control theories (an adjoint model) seems to be able to take advantage of the above two techniques by assimilating real radar and lidar data into a dynamic model without the constraints of homogeneity and periodic boundary conditions. Thus, it has the potential to generate detailed flow structures in complex PBL and provides physical insights into the air-sea and air-land interactions. In this study, we aim to assess the applicability of an adjoint model to turbulent PBL flows.

As a first step, we performed observational system simulation experiments (OSSE) of an adjoint model on the LES databases. The OSSE differs from the identical twin experiments in that the LES databases were generated by a dynamic model different from the one used in the adjoint model. Thus, the retrieval of detailed wind and temperature fields is more challenging and resembles the experiments using real measurement data. The NCAR LES model developed by Drs. Moeng and Sullivan was used to simulate a strongly convective PBL flow with -zi/L=17, where miniature convective rolls, thermals, and vortical structures are observed. The NCAR LES model employs a sophisticated two-part eddy viscosity model which is able to generate the logarithmic velocity profiles near the surface. While in the current adjoint model, a constant eddy viscosity derived from the dimensional analysis is employed. The first experiment utilizes three volumes of "observation" data with a time interval of 30 seconds. (Note that a large-eddy turnover time is about 10 minutes in this flow.) Two fictitious lidars at different locations are assumed to provide two sets of radial velocities. Encouragingly, the results show that large-scale coherent eddies can be derived by an adjoint model and the quality of the optimum solutions is sensitive to the definition of a penalized cost function. A series of retrieval experiments were performed using different time intervals of observations, different numbers of volume scans and different penalties for the cost function. The implications are discussed.