Reliable 1D magnetotelluric probabilistic inversion considering modelling assumption violations

Summary

A comprehensive understanding of the sources of uncertainty is essential in stochastic inversion workflows of magnetotelluric data. Input uncertainty related to the electromagnetic noise and measurement biases can be reliably estimated statistically during processing. Uncertainties related to limitations and oversimplifying assumptions made on the physics and geometry by the forward solver employed are usually lumped into error floors in magnetotelluric inversion workflows.

Here we propose a workflow for using 1D trans-dimensional Markov chain Monte Carlo samplers for estimating subsurface conductivity and its associated uncertainty. Our methodology replaces error floors with site-specific likelihood functions which are calculated using a machine learning algorithm trained on a set of synthetic 3D conductivity training images. The learning method quantitatively compensates for the bias caused by the 1D earth assumption. This is achieved by exploiting known dimensional properties of the magnetotelluric phase tensor.

We apply this workflow to synthetic data to quantify the improvement in reliability compared to classical 1D probabilistic inversion.

• magnetotellurics
• probabilistic inversion
• machine learning