Inversion of 3D Basin Basement and Density Using a Vision Transformer Deep Learning Network

Summary Geophysical gravity inversion is of great significance in revealing basin structures during exploration. However, when traditional gravity inversion methods are applied to basins, the accuracy and resolution are often limited by factors such as incomplete data coverage and the volumetric effect of the gravity field. In contrast, deep learning methods offer significant potential in handling complex geological data because of their powerful feature-extraction capabilities. Yet, conventional deep learning inversion approaches typically require extensive time for parameter tuning, making it challenging to achieve optimal training outcomes. In this study, we introduce a Vision Transformer (ViT) model for predicting both the three-dimensional (3D) basement undulation and the density factor of the overlying layer in sedimentary basins. By leveraging the self-attention mechanism, ViT can effectively capture global dependencies in geophysical gravity data, thereby providing more accurate insights into basin structures. Experimental results demonstrate that ViT achieves higher accuracy than traditional methods in predicting both the basement interface and the density factor of the overlying layer.