Study And Parallel Implementation Of High-efficient Optimization Algorithm In Variational Data Assimilation

Numerical weather prediction model is sensitive to the initial value, therefore, the quality of the initial value is vital to the results of the prediction. Data assimilation uses meteorological observations to generate the initial field, it is an important research domain in numerical weather prediction. How to utilize the unconventional observations, supplied by the modern tools such as satellite, radar, is a challenge which data assimilation has to face.Variational data assimialtion is a effective method to solve the problem put forward above. It formulate the data assimilation problem into a unconstrained minimization problem(3D-Var) or a minimization problem with dynamical model as a constraint (4D-Var).The kernel algorithm of the variational data assimilation problem is an optimization algorithm. To meet the demand of Research on Technical Innovation of Numerical Weather Forecasting project,the thesis implements many unconstrained optimization algorithms, analyses application area of each algorithm in detail, and compares the convergence rates,calculating efficiencies and memory space they need, among steepest descent method, non-linear and linear conjugate gradient method, L-BFGS method and descrete truncated-Newton method, points out the merits and demerits of them. Considering the characteristics of operational variational data assimilation system, the thesis gives a scheme for choosing a optimization algorithm. The research results have been applied in GRAPES, a 3D-Var system in our country with independent intellectual property right.The thesis also focuses on the parallelization of the linear CG method and L-BFGS method, designs the parallel algorithms and implementation schemes of the two methods. The analysises of their parallelization efficiencies and the testing results are also given.