Adjoint strategies for radiance data assimilation using regional atmospheric system

Large amounts of data are routinely collected by remote sensors, particularly sattelite. The use of these data however is still relatively limited given the magnitude of the information gathered. One way to increase their use is to feed as much of this information as possible into numerical models that are generally lacking much needed initial conditions. This can be done either by first retrieving information and then providing it to numerical models, or by direct assimilation of the data into models. The four dimensional variational (4DVAR) data assimilation method provides framework for generating atmospheric analysis from indirect observations and the complex mesoscale models.

We investigate the strategies to use for adjoint developments when radiative transfer models is controlled by the mesoscale model solution. The ideal way to develop the adjoint of a model is to implement adjoint for all the model options. Actually, automatique differentiation tools permits implemententation in a reasonable time of adjoint modules. Unfortunately, for some sophisticated modules we can not satisfy the cost time or/and memory requirements. So one question is: how complex must the parametrization be for DA? is it necessary to specify an adjoint for all modules?, for example if we have different options for vertical diffusion, which one will be "ideal" for DA, for sensitivity analysis and for identifications.

In our investigation radiative transfer model is controlled by bin microphysics, or the cost time constraint for adjoint computations can not be satisfied for this parametrization (also by using parallel implementation). For this reason we use the bulk microphysics option and we investigate if it's necessary to have a bin option.

We present the developpement of the Regional Atmospheric Modelling System (RAMS) Adjoint, more precisely we: (1) show strategies used for adjoint implementation with automatique differentiation tools, local check-point and justify approximations for parallelization. (2) present sensitivity analysis using different vertical diffusion options. (3) present identifications for physical parameters.