Deep Learning for 3-D Inversion of Gravity Data

Abstract

Three-dimensional (3-D) gravity inversion obtains the density distribution of subsurface geological bodies through observed gravity anomalies. Recently, data-driven methods based on deep neural networks (DNNs) have received considerable attention for geophysical inverse problems. They proved to be superior to physics-driven methods that suffer from nonuniqueness issues and high computational costs. Several deep-learning inversion strategies have been developed for geophysical modeling and are mainly applicable for two-dimensional (2-D) subsurface imaging. Despite their effectiveness, deep-learning inversions suffer from appropriate generalization to new case scenarios. In this study, a novel 3-D gravity inversion method based on encoder–decoder neural networks is proposed. The network has a gravity field encoder for 2-D gravity field feature extraction, a dimension transformation module for feature space transformation, and a density structure decoder for 3-D density structure reconstruction. A highly random dataset is constructed to enlarge the feature distribution space of the samples and perform hyper-parameter experiments to improve the accuracy and generalisability of the network. Numerical examples using synthetic gravity data show that the accuracy of the network can reach 97% in the entire area and 81.5% in the area of density structures while reducing the computational time by 20 times. In the experiments on real data of the Vinton salt dome, the results are in good agreement with the known geological information. Our method significantly improves the inversion accuracy and generalisability of 3-D gravity inversion and can be applied in real case scenarios with less computational time.