Latent space data assimilation by using deep learning

Performing data assimilation (DA) at low cost is of prime concern in Earth system modeling, particularly the era Big Data, where huge quantities observations are available. Capitalizing on ability neural network techniques to approximate solution partial differential equations (PDEs), we incorporate deep learning (DL) methods into a DA framework. More precisely, exploit latent structure provided by autoencoders (AEs) design an ensemble transform Kalman filter with model error (ETKF-Q) space. Model dynamics also propagated within space via surrogate network. This novel ETKF-Q-Latent (ETKF-Q-L) algorithm tested tailored instructional version Lorenz 96 equations, named augmented system, which possesses that accurately represents observed dynamics. Numerical experiments based this particular evidence ETKF-Q-L approach both reduces computational and provides better accuracy than state-of-the-art algorithms such as ETKF-Q.