Full-Jacobian Gauss-Newton 3D marine magnetotelluric inversion

Abstract

In the past, when solving 3D magnetotelluric (MT) inverse problems, computing the Jacobian (sensitivity matrix) at each inversion iteration has been avoided due to the relatively large memory and complexity requirements. We have developed a model-space Gauss-Newton 3D marine MT inversion scheme where the full Jacobian is computed at each inversion iteration. The memory and complexity requirements are sufficiently modest to allow the inversion of 3D data sets on a small computer cluster. For a real data example from the Barents Sea, our scheme successfully converged starting from a homogeneous half-space model. This is in contrast to a previous quasi-Newton 3D MT inversion scheme we developed, which for the same data set failed to converge starting from homogeneous half-space models. In addition, testing revealed noticeable improvements in terms of the required number of inversion iterations for convergence and receiver-imprint related artefacts and when going from a quasi-Newton based scheme to our new Gauss-Newton inversion scheme.