Fast Approximate Data Assimilation for High-Dimensional Problems

Currently, real-time data assimilation techniques are overwhelmed by data volume , velocity and increasing complexity of computational models. In this paper, we propose a novel data assimilation approach which only requires a small number of samples and can be applied to high-dimensional systems. This approach is based on linear latent variable models and leverages machinery to achieve fast implementation. It does not require computing the high-dimensional sample covariance matrix, which provides significant computational speed-up. Since it is performed without calculating likelihood function, it can be applied to data assimilation problems in which likelihood is intractable. In addition, model error can be absorbed implicitly and reflected in the data assimilation result. Two numerical experiments are conducted, and the proposed approach shows promising results.