An analysis of 4D variational data assimilation and its application (Q1014343): Difference between revisions

Data assimilation belongs to the large scale constrained optimization and inverse problems. The purpose of 4D data assimilation is to estimate the the unknown parameters like model's initial and boundary condition by fitting the model with observations distributed in time and in three dimensional space. Data assimilation is a part of the challenge for every forecasting problem, like weather prediction and environmental forecasting. The authors present the 4D data assimlation problem as a partial differential equation-constrained optimization problem in a functional analysis setting. The main focus is the solution of the nonlinear sensitivity equations. Because they consist of diffusion, convection and reaction parts, the symmetrical Marchuk-Strang operator splitting is applied to solve them. For a numerical solution of the optimal control problem, gradient and congugate gradient based iterative methods, Newton and quasi-Newton type methods are discussed. The authors propose a quasi Newton-type method because it is faster than the other gradient based descent algorithms and requires less function evaluations. It is shown that the limited BFGS (Broyden-Fletcher-Goldfarb-Shanno) method which is suitable for large scale problems is globally convergent. As an example for 4D data assimilation problem, the solution procedure for an atmospheric chemical transport model is presented. For numerical results they refer to other papers.