Gernot Geppert, Universität Hamburg
Identification of characteristic model-observation deviations for coupled data assimilation
Abstract: Coupled data assimilation should make use of links between model components and utilise observations of one model component to update the states of other components. The update across model components, however, is hindered by the use of instantaneous innovations which carry only limited information about more distant – in a structural sense – states and parameters. This fact has long been acknowledged in model verification where more sophisticated statistics than instantaneous differences are used to detect shortcomings in different model components. In order to improve the efficiency of data assimilation across model components, we transfer model verification techniques to data assimilation. We suggest to use characteristic deviations between model states and observations, here named fingerprints, for data assimilation. An appropriate fingerprint operator will map the fingerprints to new observations suitable for data assimilation.
The choices of fingerprint operators are numerous but obvious choices are established verification statistics like biases, gradients, phase shifts, or conditional differences. To identify appropriate fingerprint operators for coupled land surface-atmosphere models, we use ensembles of simulations with the Icosahedral Non-hydrostatic (ICON) model. We run ICON in a large eddy simulation configuration on a small, limited domain and systematically perturb soil and land-surface parameters and states to produce the ensembles. Subsequently, we test statistics of boundary layer observables to find appropriate fingerprint operators for the update of soil and land-surface variables.