Stochastic inversion of time domain electromagnetic data with non-trivial prior

Inversion deals with inferring information about the subsurface (by reconstructing its physical properties), given: 1) observed data (usually collected at the surface) and 2) available forward modelling tools (describing physics of the used geophysical methodology). Inevitably, these forward modelling tools are always characterized by some level of approximation, and, in turn, this inaccuracy, unavoidably, affects the inversion results. This thesis presents, in particular in the context of airborne electromagnetic data, the impact and relevance of quantifying this source of (coherent) error. Specifically, a possible strategy to quantify the modelling error is discussed in the thesis. The adopted strategy for the estimation of the modelling error makes use of prior knowledge about the investigated system. The same prior knowledge is necessary in stochastic inversion frameworks. Stochastic inversion provides a natural way for 1) the assessment of the uncertainty of the final results and 2) for incorporating complex prior information into the inversion, from sources that are not the geophysical observations. Since the assessment of the modelling error is based on prior information that is also used in the stochastic inversion approaches, it is a natural choice to adopt these probabilistic strategies. By taking into account the modeling error, the stochastic inversions can eliminate or, at least, minimize, the effects of the forward approximation in the inversion results. In this thesis, through synthetic and field tests, we discuss the stochastic inversion considering the modeling error. What is called prior in the framework of stochastic inversion is assimilable to the training dataset in the context of Neural Networks: to some extent, in both cases, the final solution is by construction “stationary” with respect to the initially provided ensemble used to feed (or train) the inversion algorithm. Based also on these premises, and in the attempt to find a way to address the “definitive” problem of a fully 3D stochastic inversion, we verify the possibility of extremely efficient Neural Network strategy for the inversion of massive airborne geophysical datasets. Some preliminary, but, still, very promising results on this matter are discussed in the second last chapter of this thesis. Also in this case, the conclusions are drawn based on synthetic and experimental data.