Parallelized 3D Inversion of CSEM Data Based on Finite Element Method With Unstructured Mesh

Presently, 3D modeling technique is becoming even popular, especially during the last decade, and it begins to play a more important role in data interpreting. In this abstract, we will present our recent developments on parallelized 3D inversion algorithm for frequency domain CSEM data interpreting. By combining the edge-based FEM direct forward modeling and the Gauss-Newton optimization we form the inversion framework. In order to decrease the singularity of the sources we take the secondary field formulation. We discretize the study domain by the flexible unstructured tetrahedral elements to better modeling the complex geological model and topography. For 3D forward modeling and adjoint problem, we implemented the parallel direct solver MKL Pardiso. Considering the frequencies for forward modeling and sensitivity calculation in frequency domain are independent, we further parallelized the algorithm over frequencies on MPI to speed up the solver. In the view that the iterative solver LSQR is on the one hand more numerically stable and on the other hand much more memory saving compared to the traditional CG method, we impose LSQR to the Gauss-Newton inversion process to solve the inversion normal equations. Besides, we proposed a new roughness operator for unstructured discretization which works well for CSEM inversion problems. By applying our algorithm to the land CSEM model and topography marine CSEM model, we validated the effectiveness and robustness of our inversion scheme.