Infocube

PDAF - Parallel Data Assimilation Framework

Data assimilation is the technique to combine real observations with numerical models of geophysical systems. The goal is to obtain improved estimates of the model state, fluxes in the system, or indications where the model misrepresents reality. A common usage for data assimilation is the initialization of model states for numerical whether prediction.

During the recent years there have been developments of advanced data assimilation algorithms. Most of these algorithms base on the Kalman filter. However, these algorithms utilize an ensemble of model states that is integrated with the full numerical model of the geophysical system in order to represent the best estimate of the state as well the estimate of its error. These algorithms have been classified as ensemble-based Kalman filters.

Data assimilation with these advanced algorithms together with large-scale models is computationally extremely demanding. Thus, high-performance computers need to be used and the assiilation system needs to be parallelized. In addition, the data assimilation system has to be implemented, typically on the basis of complex existing models that were not programmed with data assimilation in mind.

The Parallel Data Assimilation Framework PDAF provides an environment and well-defined interface to set up a parallel data assimilation system. The framework is attached to the model code with only minimum changes in general parts of the model code. Then the model can generally be executed as before using its data assimilation extension. PDAF provides fully implemented, parallelized, and optimized filter algorithms. In addition, the parallelization support for the integration of an ensemble of model states is provided. For the full implementation of the assimilation system a user will only need to implement a set of routines that are particular to the model and the observations to be assimilated. With this the implementation of a data assimilation system suited for supercomputers is significantly simplified.

PDAF is supported through BremHLR and is developed and hosted at the Alfred Wegener Institut for Polar and Marine Research, Bremerhaven, Germany, where more information about PDAF is provided.

© BremHLR 2017