Deep Data Assimilation: Integrating Deep Learning with Data Assimilation

In this paper, we propose Deep Data Assimilation (DDA), an integration of (DA) with Machine Learning (ML). DA is the Bayesian approximation true state some physical system at a given time by combining time-distributed observations dynamic model in optimal way. We use ML order to learn assimilation process. particular, recurrent neural network, trained dynamical and results process, applied for purpose. At each iteration, function that accumulates misfit between forecasting DA. Subsequently, compose model. This resulting composition includes features process can be used future prediction without necessity fact, prove DDA approach implies reduction error, which decreases iteration; achieved thanks training very useful cases when are not available steps cannot reduce error. The effectiveness method validated examples sensitivity study. technology two different applications: Double integral mass dot Lorenz system. However, algorithm numerical methods proposed work other physics problems involve equations and/or variables.