Use of machine learning to estimate statistics of the posterior distribution in probabilistic inverse problems - an application to airborne EM data.

Abstract

• A machine learning approach for solving probailistic inverse problems by directly estimating properties of the posterior distribution • Allow the use of arbitrarily complex prior and noise models as long as they can be sampled • Exemplified on inversion of airborne electromagnetic data, allowing analysis of more than 100000 1D soundings per second

1D Stochastic Inversion of Airborne Time-Domain Electromagnetic Data with Realistic Prior and Accounting for the Forward Modeling Error

“…Recent strategies exploit the neural networks potential, allowing almost real-time inversions with no significant quality reductions [41]. In addition, in the attempt of retrieving, not merely the conductivity distribution, but, rather, immediately useful pieces of information about the targets and, at the same time, in the effort of supplying reliable assessment of the associated uncertainty, probabilistic petrophysical inversions are becoming more and more common [42].…”