A Study Of MT And CSAMT 2-D Joint Inversion And Parallel Agorithm

The inversion effects of magnetotelluric (MT) method is poorer than Controlled source audio-frequency magnetotelluric(CSAMT) When the target is in the shallow earth and the inversion effects of CSAMT method is poorer than MT When the target is in the deep stratum. In order to improve the inversion result of MT in the shallow earth and the result of CSAMT in the deep stratum the paper adopts MT and CSAMT joint inversion. We get the ideal inversion result through the joint inversion of MT and CSAMT. Because the running time of joint inversion is long, we implement the MPI algorithm in the joint inversion. The speed of the parallel algorithm is very high.Nonlinear conjugate gradient method (NLCG) is an efficient method for the magnetotelluric inversion. The method can avoid computing Jacobi matrix and update the model through one forward and several pseudo-forward. The method saves much running time and improves the efficiency. When the grid of model is big and several frequency need to be computing, the computing speed needs to improve. The Magnetotelluric(MT) is a frequency domain method and different frequency can explore the target for different depth. Different frequency data need to do forward modeling and pseudo-forward modeling in the inversion separately, and they can be allocated for several processes. Apply MPI parallel method into magnetotelluric nlcg 2D inversion. We need to solve the problem of reading and writing a file for several processes. Applying gatherv method to gather data is very flexible. The program is easy to code and maintain. The program is very strong. The speedup ratio is very well through computing and comparision. The bigger is, the size of 2D, the longer is the running time of inversion. In order to utilize the efficient feature of NLCG, we apply MPI parallel algorithm in the NCLG inversion for big grid. Many scholars has applied one parallel algorithm (such as MPI,GPU) in a serial program, but the paper study the mixture parallel algorithm for the NLCG inversion. Both of the parallel algorithm (MPI and GPU) is applied in the mixture parallel program. The equations of MT with different frequency need to be solved and they are separately, so the task of solving equation for different frequency is allocated to different processes in the MPI algorithm. At the same time, we solve the equation with GPU algorithm.Apply MPI parallel method into magnetotelluric "NLCG" 2D inversion. We need to solve the problem of reading and writing a file for several processes. Applying "gather" method to gather data is very flexible. The program is easy to code and maintain. The speedup ratio is very well through computing and comparing with serial algorithm. We applied two methods in GPU algorithm. One is coding with CUDA language, the other is using "Cula Sparse" library. If we want to use the "Cula Sparse" library, we need to change the format of the matrix. The efficiency of the transformation for the matrix is very important. We can save the total time through optimize the transformation algorithm. The efficiency of the MPI+CULA Sparse is not stable through applying the "Cula Sparse" and MPI to NLCG algorithm, so we applied MPI+CUDA in NLCG agorithm.Finally, we have developed the MPI+GPU mixture algorithm for the 2D MT NLCG inversion and the MPI parallel algorithm for MT and CSAMT joint inversion.