Dealing with Uncertainty in AEM Models (and Learning to Live with it)

Abstract

Interpreting inversions and modelling airborne electromagnetic (AEM) data is ambiguous. Assessment on the degree of certainty of how representative a selected model is, always reassuring. Geophysicists assessing AEM models are often faced with the conundrum of determining a single ‘best’, ‘right’ and geologically sensible model from of all the possible solutions. This paper explores the characteristics of several acceptable models, without being concerned with details of any particular one.

Geoscience Australia’s reversible jump (trans-dimensional) Markov chain Monte Carlo (rj-MCMC) is a stochastic algorithm which has enabled the sampling of thousands of plausible models that fit the data at each individual location. Through the statistical analysis of these ensembles of models, a measure of uncertainty and a probable distribution of conductivities at that depth can be derived.

On most occasions, single ‘best’ solutions from deterministic inversions are found to be reasonable representations of the whole suite of models recovered by the MCMC. But the importance of exploring multiples models and their limitations resides on trying to extract what information can actually be determined from the data, information which often cannot be given by a single best model.