Department of Mathematics, Senior Research Scientist, INPE (National Institute for Space Research), Sao Jose dos Campos, Brazil "Neural networks applied to data assimilation process"
11:00 AM, Thursday March 8, 2018, 499 Dirac Science Library
Data assimilation is an essential process in operational prediction centers based on the time integration of partial differential equations. The numerical weather prediction is a very relevant example. The process of data assimilation is an inverse problem of initial condition identification – called ”analysis” –, combining observations with data from the forecasting mathematical model. Historically, the mathematical estimation process was started by using the least square approach. Indeed, all data assimilation schemes in some sense are based on the least square estimator. We are going to show the time evolution from the least square approach up tothe Kalman filter, and beyond - the particle filter, and one step ahead. The variational method is also a generalized least square formulation.
In the talk, results with several data assimilation methods emulated by artificial neural networks will be shown. The algorithmic complexity is reduced by employing neural networks. This is readily realized for 3D atmospheric models. Two atmospheric general circulation models, SPEED and COAPS-FSU, are used to demonstrate the data assimilation speed-up with neural networks emulating the Local Ensemble Transform Kalman Filter (LETKF), and producing a similar analysis.