Assimilation of radar observations of a supercell storm using 4DVar: Parameter retrieval experiments

In this study we examine the utility of a 4DVar adjoint method for assimilating Doppler radar data of a supercell thunderstorm into a numerical model for the purposes of analysis and forecasting. The storm we have chosen for study is the 17 May 1981 Arcadia, Oklahoma tornadic supercell which was observed by two Doppler radars.

The 4DVar adjoint method attempts to find the best fit of a numerical model solution to a time series of radar data, as well as a background field. The numerical model solves the anelastic equations of motion and uses a simple warm rain microphysical scheme. By running the numerical model forward and its adjoint backwards in time, a cost function which measures the difference between the model solution and the data/background can be minimized. During this minimization, some of the unobserved fields such as crossbeam velocity, temperature and cloudwater can be retrieved in the initial condition. It has been found that the retrieval and minimization process can be improved by adding penalty terms (such as spatial/temporal smoothness constraints) to the cost function.

A number of 4DVar experiments have been run using observations of the Arcadia supercell from a single Doppler radar. The results are compared with independent observations from the second radar, as well as observations from a 444 meter tower that the supercell passed over. The sensitivity to various parameters in the cloud model and variational scheme has been examined. A qualitative comparison is also performed with results from another data assimilation scheme, the Ensemble Kalman Filter. Finally, forecasts experiments using the initial conditions retrieved by the 4DVar method have been performed, along with an examination of the sensitivity to variations in the environmental sounding.