Inversion of real gravity data from geological faults using a generative neural network model

Normalising flows is a novel generative neural network model, which can be applied to Bayesian parameter inference. When gravity inversion is reformulated as a probabilistic inference problem, stable results can be obtained that naturally incorporate the inherent uncertainties and noise from the source background and the instrument. As opposed to some standard methods, Bayesian gravity inversion does not default to a single solution in an ill-posed problem, but informs the user about all possibilities that are consistent with the gravimetry survey of interest. It has been demonstrated that the normalising flow method can provide accurate results for a simulated data set, even when applied to high-dimensional data. Once the network is trained, the results can be obtained within seconds and it can be reused, without retraining, for multiple gravimetry surveys that are consistent with the training data set. Here, improvements on the previous work are presented, where the method is applied to a more realistic and complex geophysical problem; the inversion of gravity measurements to infer parameters of geophysical faults. The normalising flow network is trained and tested for fault models with various complexities, and finally the method is applied to the inversion of airborne gravimetry data.