3D Geo-Modeling Framework for Multisource Heterogeneous Data Fusion Based on Multimodal Deep Learning and Multipoint Statistics: A case study in South China Sea

Abstract. Relying on geological data to construct 3D models can provide a more intuitive and easily comprehensible spatial perspective. This process aids in exploring underground spatial structures and geological evolutionary processes, providing essential data and assistance for the exploration of geological resources, energy development, engineering decision-making, and various other applications. As one of the methods for 3D geological modeling, multipoint statistics can effectively describe and reconstruct the intricate geometric shapes of nonlinear geological bodies. However, existing multipoint statistics algorithms still face challenges in efficiently extracting and reconstructing the global spatial distribution characteristics of geological objects. Moreover, they lack a data-driven modeling framework that integrates diverse sources of heterogeneous data. This research introduces a novel approach that combines multipoint statistics with multimodal deep artificial neural networks and constructs the 3D crustal P-wave velocity structure model of the South China Sea by using 44 OBS forward profiles, gravity anomalies, magnetic anomalies and topographic relief data. The experimental results demonstrate that the new  approach surpasses multipoint statistics and Kriging interpolation methods, and can generate a more accurate 3D geological model through the integration of multiple geophysical data. Furthermore, the reliability of the 3D crustal P-wave velocity structure model, established using the novel method, was corroborated through visual and statistical analyses. This model intuitively delineates the spatial distribution characteristics of the crustal velocity structure in the South China Sea, thereby offering a foundational data basis for researchers to gain a more comprehensive understanding of the geological evolution process within this region.