Three-dimensional gravity inversion based on global receptive convolution

Three-dimensional gravity inversion is the process of obtaining the location, shape, and physical parameters of underground anomalies using gravity anomaly data observed at the surface. In recent years, the rapid development of data-driven methods has led to widespread interest in applying Deep Learning (DL) techniques to gravity inversion problems, yielding certain results. Research based on the U-Net network that employs an Attention Feature Fusion (AFF) mechanism for three-dimensional gravity inversion enhances the vertical resolution of the inversion results. However, the reconstruction effect for deeper models still needs improvement. Building on this, the use of Global Receptive Convolution (GRC) to replace certain standard convolutions aims to ensure that the convolution process meets the conditions of spatial position and channel indexing. This approach integrates global and local features into a pixel-wise representation, further learning the original features to improve the reconstruction resolution of deeper models. The network's inversion results show enhanced vertical resolution, and experimental results from the San Nicolás deposit in Mexico demonstrate that the inversion network can clearly predict the basic location and approximate shape of the sulfur deposit, aligning well with known geological data.