Efficient three-dimensional inversion of magnetotelluric data using approximate sensitivities

Abstract

SUMMARY A method for accurately approximating sensitivities is introduced for the efficient 3-D inversion of static-shifted magnetotelluric (MT) data. Approximate sensitivities are derived by replacing adjoint secondary electric fields with those computed in the previous iteration. These sensitivities can reduce the computation time, without significant loss of accuracy when constructing a full sensitivity matrix for 3-D inversion, based on the Gauss–Newton method. Additional reduction of computational cost can be attained by modifying the inversion scheme to run on a parallel computing platform. The effectiveness of approximate sensitivities is tested by inverting both synthetic and field data obtained in Pohang, Korea, and Bajawa, Indonesia. The accuracy of approximate sensitivities is validated by sensitivity analysis of synthetic data. To make the inversion of static-shifted MT data more stable, a weighting coefficient for static-shift parameters is added to the objective function and is updated at each iteration. Approximate sensitivities are calculated much faster than exact sensitivities, and are accurate enough to drive an iterative inversion algorithm.