Data Assimilation Networks

Data Assimilation aims at estimating the posterior conditional probability density functions based on error statistics of noisy observations and dynamical system. State art methods are sub-optimal due to common use Gaussian linearization non-linear dynamics. To achieve a good performance, these often require case-by-case fine-tuning by using explicit regularization techniques such as inflation localization. In this paper, we propose fully data driven deep learning framework generalizing recurrent Elman networks assimilation algorithms. Our approach approximates sequence prior densities conditioned log-likelihood cost function. By construction our can then be used for general nonlinear dynamics non-Gaussian densities. As first step, evaluate performance proposed partially observed Lorenz-95 system in which outputs network fitted We numerically show that approach, without any technique, achieves comparable state-of-the-art methods, IEnKF-Q LETKF, across various ensemble size.