Gravity modelling in sedimentary basins with a mixed Machine Learning-inversion method

We describe a mixed Machine Learning (ML)-Inversion algorithm for sedimentary basin modelling of gravity data. We consider two typical targets of basin modelling, i.e., the basement depth and the diapirs, both being of fundamental importance for energy exploration or tectonic studies. First, Machine Learning is implemented to interpret gravity anomalies along a set of profiles crossing the whole anomaly. Then, the 2D interpreted models are merged into a reference model to constrain the final 3D inversion of the whole set of data. The learning of our Convolutional Neural Network is based on the simplest gravity sources, i.e., faults. Such choice transfers a strong a priori information to the interpretative problem, since gravimetric anomalies due to diapirs, or to the morphology of the basement, can be seen as composed by the constructive interference of anomalies generated by the edges of the sources. We considered up to 221.100 different fault-models with variable parameters (depth, density, thickness and depth-to-top), corresponding to low/ high dip-angle faults and mixed faults. Our workflow was shown suitable for reconstructing different structures in either synthetic or real-data cases: a diapir-shaped source and the basement of a huge sedimentary basin, the Caltanissetta basin (Italy), respectively. In both cases it is shown that our mixed approach produces significant results thanks to the self-constraints extracted during the Machine Learning step.