A fully finite-element based model-space algorithm for three-dimensional inversion of magnetotelluric data

We present a fully finite-element based inversion methodology for imaging 3-Dmagnetotelluric impedance data on unstructured meshes. The inverse problem is formulated using a minimum-structure Gauss-Newton type optimization scheme that minimizes an objective function with respect to the model perturbation. By introducing a rigorous regularization scheme, we derived a Ritz-type variational formulation of the model objective function and designed a face-based finite-element basis function to discretize the model gradient across tetrahedron's inter-element boundaries. The forward modelling engine of our optimization scheme is based on a finite-element solution of the E-field Helmholtz equation that is enforced for the magnetotelluric simulation problem using the appropriate edge-based basis functions and 3D boundary conditions. The optimization algorithm developed here utilizes a message passing interface scheme and uses a direct solver to factorize and store both the regularization matrix and the forward modelling coefficient matrix on the processes working in parallel. Having to do this only once within each Gauss-Newton optimization cycle facilitates both the calculation of the dot product of the model regularization terms with the evolving model perturbation, and computing implicitly the sensitivity-vector products. We validated the methodology and the correctness of the developed algorithm for two test examples (COMMEMI 3Ds) from the literature. Also, by comparing the performance between classes of iterative solvers we demonstrated the superior performance of generalized minimum residual solver in reducing the residual norm of the iterative solver during model updates. Using the algorithm in a geologically realistic scenario, we imaged the anticipated geometry of the Lalor volcanogenic massive sulphide deposit in Canada. The feasibility of the imaging methodology is further evaluated with the survey data, for which, again the algorithm converged to the anticipated model solution reproducing the lithostratigraphic sequence of the ore deposit.