3D gravity inversion using cycle-consistent generative adversarial network

Gravity inversion is an important approach for obtaining the spatial structure and physical properties of underground geological bodies based on surface information. Owing to the recent advances in deep learning, neural network-based methods have been widely used for gravity inversion. However, convolutional neural networks (CNNs) require a large number of labeled samples, and the generation of such datasets for all considered geological bodies, requiring gravity forward modeling, is expensive in terms of time and storage space. To reduce the dependence on labeled samples, a three-dimensional gravity inversion method based on a cycle-consistent generative adversarial network (Cycle-GAN) is proposed herein. This network comprises two parts: generator subnetworks and discriminator subnetworks. The generator subnetworks generate gravity forward and inversion data, while the discriminator subnetworks mainly ensure the consistency of the distribution between the generated and real data. We compared the results obtained on synthetic and real data. The findings suggest that Cycle-GAN outperforms CNNs in the inversion of underground geological bodies when using a small number of labeled samples. Furthermore, the results obtained using the proposed method on real data from the San Nicolas deposit in central Mexico are consistent with previously reported results.