Section: New Results
Towards comparison studies of sequential and variational data assimilation methods for air quality models
The objective is to evaluate the performance of different data assimilation schemes for atmospheric chemistry-transport models (CTM). Contrary to meteorological models, air quality models are non chaotic and the problem is not the control of the initial conditions but of forcing termes. There are recent applications of variational and sequential data assimilation methods based on atmospheric chemistry-transport models, however, comparison studies of the two approaches in the same experimental setting have never been performed mainly due to technical problematics. The recently developed platform Polyphemus (developed by Clime) makes it possible to overcome these technical difficulties for a data assimilation system.
A typical data assimilation system consists of three components, namely data (observation), model (physics), and assimilation algorithms. We adopted object-oriented techniques, such that the developments of the three components are independent from one another. Most of the classical data assimilation schemes for CTMs have been implemented within Polyphemus for comparison studies. The available sequential schemes are Optimal Interpolation (OI), Ensemble Kalman Filter (EnKF), and Reduced Rank Square RooT Kalman filter (RRSQRT). For the variational approach, we have derived the adjoint model for the underlying CTM Polair3D within Polyphemus using automatic differentiation techniques, and the four-dimensional variational assimilation method (4D-Var) is still under testing and will be available in the near future.
For the first application of the data assimilation system, the simple method OI is employed to assess the impact of assimilating EPS troposphere ozone observations. For advanced assimilation methods, say EnKF, RRSQRT, or 4D-Var, one needs the error specifications for CTMs and/or background (a priori) species concentrations. We approximate the model error by perturbing model input data, such as surface emissions, boundary conditions etc. The preliminary results of EnKF and RRSQRT show that the predictions based on assimilated concentrations approach quickly to reference simulations. This is a strong indication that our approximation of model error is not very realistic, and the assimilation suffers from filter divergence.
In summary, we have implemented both classical sequential and variational assimilation methods within Polyphemus platform, and preliminary results are obtained by perturbation methods. For meaningful results, more investigations are needed on (model and background) error modeling for the atmospheric chemistry-transport models.