Deep Markov Models for Data Assimilation in Chaotic Dynamical Systems

Abstract

Recently, the use of deep learning in data assimilation has been gaining traction. One particular time series model known as deep Markov model has been proposed, along with an inference network that is trained together using variational inference. However, the original paper did not address the full capability of the model in data assimilation problem. Therefore, we aim to evaluate the suitability of a deep Markov model and its inference network against a chaotic dynamical system, which often shows up as a problem in data assimilation. We evaluate the model in various generative conditions. We show that when information about part of the target model is known, the model is able to match the capability of a smoothed unscented Kalman filter, even when there are process and observation noise involved.