3D gravity inversion based on deep learning

Gravity inversion is a typical geophysical inversion method that obtains the underground density distribution by analyzing the gravity anomaly. Normally, it can be divided into geophysics-based and deep learning based inversion. The 3D geophysics-based inversion is a time- and memory-consuming method, so 3D inversion is not routinely implemented in practical data interpretation. Here, we propose a deep learning method to transfer the 3D inversion problem to a multiple layers 2D mapping problem by decomposing the 3D target into four 2D images, including the horizontal location, vertical center, thickness and density distribution. This method is denoted as “decomposition network”. By implementing synthetic experiments with regular and complex models, and comparing with the 3D U-Net inversion, the proposed network has proved can reconstruct underground targets with high accuracy and high efficiency.