Transdimensional Monte Carlo Inversion of AEM Data

Abstract

A new approach for the 1D inversion of AEM data has been developed. We use a reversible jump Markov Chain Monte Carlo method to perform Bayesian inference. The Earth is partitioned by a variable number of nonoverlapping cells defined by a 1D Voronoi tessellation. A cell is equivalent to a layer in conventional AEM inversion and has a corresponding conductivity value. The number and the position of the cells defining the geometry of the structure with depth, as well as their conductivities, are unknowns in the inversion. The inversion is carried out with a fully non-linear parameter search method based on a transdimensional Markov chain. Many conductivity models, with variable numbers of layers, are generated via the Markov chain and information is extracted from the ensemble as a whole. The variability of the individual models in the ensemble represents the posterior distribution. Spatially averaging results is a form of ?data-driven? smoothing, without the need to impose a specific number of layers, an explicit smoothing function, or choose regularization parameters. The ensemble can also be examined to ascertain the most probable depths of the layer interfaces in the vertical structure. The method is demonstrated with synthetic time-domain AEM data. The results show that an attractive feature of this method over conventional approaches is that rigorous information about the non-uniqueness and uncertainty of the solution is obtained. We also conclude that the method will also have utility for AEM system selection and investigation of calibration problems.