Fast imaging for the 3D density structures by machine learning approach

Residual Bouguer gravity anomaly inversion can be used to imaging for local density structures or to interpret near-surface anomalous mass distribution. The reasonable prior information is the crucial recipe for obtaining a realistic geological inversion result, especially for the ill-posed geophysical inversion problem. The conventional strategies introduce the prior constraints or joint multidisciplinary information in object function as regularization, and then use some optimization algorithm to minimize the object function. This process is called model-driven approach and is usually time-consuming. In recent years, the rapid development of machine learning technology has provided new solutions for solving geophysical inversion problems. Machine learning methods can reduce the dependence on prior information in the inversion process through setting special training datasets, and the time consumption of an inversion process executed by the trained model can be shortened by several orders of magnitude, which is conducive to fast inversion for the same type of application scenarios. In this study, we were inspired by the U-net model and develops the GV-Net (Gravity voxels inversion network) model using the convolutional neural network for the inversion of residual gravity anomalies. We first discussed the effects of different loss functions on the convergence speed of model training and prediction accuracy. Then, we analyzed the robustness of our model by changing noise levels of the datasets. At last, we employed this model in a real scenario. The results have demonstrated that the GV-Net model has the ability to deal with specific inverse problems by predefined training datasets.