Kalman filtering for nonlinear atmospheric chemistry models : first experiences

The aim of the {sc eu project {sc riftoz is to analyse regional differences in tropospheric ozone over Europe. One of the key activities within {sc riftoz therefore involves recovering ozone concentrations from available measurements. This will be done by running the atmospheric chemistry model {sc lotos over the selected period using a data assimilation technique to incorporate the measurements. A commonly used data assimilation technique is the (extended) Kalman filter. This filter has proved to be very useful in many applications. However, the models involved in these applications are usually only weakly nonlinear, whereas atmospheric models, like {sc lotos, are often highly nonlinear. The paper presents first results on data assimilation with a highly nonlinear test model using the (extended) Kalman filter algorithm. The test model has been designed such that the essential characteristics of the {sc lotos model, including stiff (photo-)chemistry, have been retained. Application of the standard algorithm for Kalman filtering is infeasible because of the huge computational and storage requirements. Instead, a reduced rank approximation of the covariance matrix is used, which reduces the computational burden to an acceptable amount of CPU time. Also attention is paid to reducing the number of noise parameters in the filter algorithm in order to further restrict the number of model evaluations that is required to solve the filtering problem. The results of the tests are very promising and show that Kalman filtering may be successfully applied to atmospheric chemistry models.