Data Assimilation by Neural Network for Ocean Circulation: Parallel Implementation

Haroldo F. de Campos Velho, Helaine Cristina Morais Furtado, Sabrina B. M. Sambatti, Carla Barros Osthoff Ferreira de Barros, Maria E. S. Welter, Roberto Pinto Souto, Diego Carvalho, Douglas de O. Cardoso

Abstract: Data assimilation (DA) is an essential issue for operational prediction centers, where a computer code is applied to simulate physical phenomena by solving differential equations. The procedure to determine the best initial condition combining data from observation and previous forecasting (background) is carried out by a data assimilation method. The Kalman filter (KF) is a technique for data assimilation, but it is computationally expensive. An approach to reduce the computational effort for DA is to emulate the KF by a neural network. The multi-layer perceptron neural network (MLP-NN) is employed to emulate the Kalman in a 2D ocean circulation model, and algorithmic complexity to KF and NN is presented. A shallow-water system models the ocean dynamics. Synthetic measurements are used for evaluating the MLP-NN for the data assimilation process. Here, a parallel version for the DA procedure by the neural network is described and tested, showing the performance improvement for a parallel version of the NN-DA.