Parallel Regularization Inversion Of2-D MT Data Based On MPI

With the rapid development of electromagnetic exploration technology, Especially in the second, three-dimensional modeling and inversion technologies constantly updated, More and more large-scale computational problems. The magnetotelluric method of two-dimensional forward there is the need for solving large sparse symmetric coefficient matrix of linear equations, calculate the consumption of a very long time, and reduces the efficiency of forward operations, while reducing the efficiency of two-dimensional inversion, so the use of parallel computing can improve the efficiency of forward and inverse operations.In this paper, the finite element method for two-dimensional magnetotelluric forward modeling, inversion in the objective function by using the regularization method for solving large-scale inversion of linear equations posed, in order to improve the Gauss-Newton the local convergence of the method, the introduction of the damping factor damping gauss-Newton iterative algorithm.the analysis of the MT2D modeling and inversion structure and function of the serial algorithm, program, magnetotelluric modeling and inversion, for the iterative solution of the model parameters, the need to calculate the partial derivative matrix, serial numerical computational experiments show that the partial derivative matrix calculate the most time-consuming. The damping Gauss-Newton iterative algorithm for inversion in the calculation of partial derivatives accounted for the majority of the whole inverse calculation, therefore, only the partial derivative computation parallelization, the remaining part of the reservation serial computing model.Selection of MPI parallel tools, the use of coarse grain parallelism, the parallel strategy is to achieve complete equivalence between the different frequency, analog speed parallel computing in a multi-process, select a different model, the model test to verify the correctness and performance of parallel algorithms. Through the parallel computing model can be achieved to a greater extent of parallelism, and better improve the computational efficiency of the algorithm.