Statistical variational data assimilation

This paper is a contribution in the context of variational data assimilation combined with statistical learning. The framework of data assimilation traditionally uses data collected at sensor locations in order to bring corrections to a numerical model designed using knowledge of the physical system of interest. However, some applications do not have available data at all times, but only during an initial training phase. Hence, we suggest to combine data assimilation with statistical learning methods; namely, deep learning. More precisely, for time steps at which data is unavailable, a surrogate deep learning model runs predictions of the ‘true’ data which is then assimilated by the new model. In this paper, we also derive a priori error estimates on this statistical variational data assimilation (SVDA) approximation. Finally, we assess the method by numerical test cases.