Probabilistic inverse problems using machine learning - applied to inversion of airborne EM data.

This is a Preprint and has not been peer reviewed. This is version 1 of this Preprint.


Download Preprint


Thomas Mejer Hansen


Probabilistic inversion methods allow, in principle, to combine probabilistic geo-information from diverse sources into one consistent statistical model (the posterior distribution) containing all available information. In practice, however, they rely on Monte Carlo sampling methods, which can be extremely computationally demanding.
Here a general, and simple to apply, method is presented, utilizing machine learning, which allows fast direct estimation of properties of the posterior distribution.
The fundamental idea is to construct a training data set that represents all the information represented in the probabilistic formulation of inverse problems. From such a training data set, it is demonstrated how regression and classification type neural networks can be designed, with specific choices of output layer and loss functions, that allows direct characterization of the posterior distribution using regression and classification type
The methodology is demonstrated on probabilistic inversion of airborne electromagnetic data and compared to results obtained computationally more expensive sampling methods.



Earth Sciences, Physical Sciences and Mathematics


Proabilistic Methods, Inversion theory


Published: 2021-03-03 01:02


CC BY Attribution 4.0 International

Add a Comment

You must log in to post a comment.


There are no comments or no comments have been made public for this article.