Practical 3D inversion of large airborne time domain electromagnetic data sets

Abstract

In this paper we show that 3D inversion of large airborne time domain EM data, which is traditionally considered impractical, can be rapidly carried out by using a thoughtful workflow. In our 3D inversion algorithm, the number of cells in the mesh and the number of soundings are two factors that slow down the inversion. Therefore, we develop a strategy of adaptive mesh and sounding refinement to minimize the number of cells and the number of soundings required by the inversion. At the beginning, a coarse mesh and a few soundings are used to quickly build up a large-scale model. Then the mesh is refined and more soundings are added based upon their data misfit. At each iteration of the inversion, a certain number of soundings are randomly selected, and we change the data selection from iteration to iteration. This allows us to down-sample the field data without much loss of information. Once the large-scale model is obtained, we carry out some tile inversions that focus on smaller areas with a locally refined mesh to better resolve the small-scale features. The workflow is demonstrated by a synthetic example with 2121 transmitters that takes about 10 hours to be solved compared to about 150 hours if we had started the inversion on a fine mesh and used all of the transmitters. The methodology of speeding up the inversion by adaptive mesh and data refinement can also be applied to other EM surveys.