Research On The Maximum Likelihood Ensemble Filter Data Assimilation Technique And Its Parallelization

Atmospheric flow is a nonlinear physical geography system which evolves withthe time and it is sensitive to the initial conditions. Accurate weather forecast is basedon exact intial value of the prediction model. The advancement of data assimilation isimportant for the improvement of numerical weather prediction. Therefore, dataassimilation is one of the most important techniques to improve the quality of weatherprediction. With the increasement of observation data and the improvement ofobservation quality, data assimilation technique develops quickly with new challenges.The quality of data assimilation technique relies on the description of the information ofbackground and observation. The description of background error covariance is moreaccurate from the functional form in3D-Var to the implicated evolution at theassimilation time interval in4D-Var, and to the predicting in the Kalman filter.Unfortunately, the above methods require strict differentiability of the data assimilationsystems, such as differentiable models and observation operators. The MaximumLikelihood Ensemble Filter (MLEF) is one of the data assimilation techniques, whichcombines the ideas used in ensemble filters,3D-var and Hessian preconditioning. TheMLEF method can not only obtain the flow-dependent background error covariance bystatistics of the ensemble samples but also can be applied to non-differentiable systems.It is a new research focus in the current data assimilation field.The paper mainly does research on the Maximum Likelihood Ensemble Filteralgorithm. And then constructs an experimental data assimilation system based on theMaximum Likelihood Ensemble Filter. At last, an optimized system is obtained usingparallel computation. The following are the main conclusions.（1） Based on the study of the theoretical aspects of the Maximum LikelihoodEnsemble Filter algorithm, the paper describes the derivation of generalized gradientand generalized Hessian matrix.（2） Taking the numeric solution of the one-dimensional nonlinear advectionequation and the selecting of the observation operators into consideration, anexperimental Maximum Likelihood Ensemble Filter data assimilation system isestablished. This system takes the analysis solution obtained from the MLEF method asthe intial condition, and then obtains the forecast by a TVD difference scheme of theone-dimensional nonlinear advection equation. Then this forecast is used as thebackground in the next data assimilation cycle.（3） Based on the above studies, the MLEF system has beem verified throughsimulant observation operators: For both differentiable and non-differentiableobservation operators, the analysis error in the MLEF experiment is smaller than in theGRAD experiment. For both differentiable and non-differentiable observation operators, the cost functions and gradients decrease faster in the MLEF experiment than in theGRAD experiment.（4） The paper does research on the parallel algorithm of the MLEF methodand implements a parallel program for the MLEF system using MPI. Firstly, the paperanalyzes the correlation among different components in the sequential MLEF algorithmand the characteristics of the computation. Then a parallel strategy of the MLEF dataassimilation is rised. At last an improved MLEF data assimilation system is obtained byusing MPI. The results show that the parallelized MLEF system can obstain satisfiedspeedups and the parallel efficiency improves varying the increase of the modelresolution and the size of the ensemble.