Fusion of Limited Site-Specific Borehole Logs and Geophysical Data from a Different Site for Three-Dimensional Subsurface Geological Modeling Using Multiscale Generative Adversarial Network

Three-dimensional (3D) modeling of subsurface stratigraphy is a challenging task because site-specific measurements (e.g., boreholes) for ground investigation (GI) are often sparse and limited. Geophysical surveys provide massive data for stratigraphic interpretation, which can be leveraged to supplement site-specific GI data. However, existing studies are rare that integrate geophysical data with borehole data for development of site-specific 3D subsurface geological models in geotechnical practice, especially when geophysical data and borehole data are from different sites. In this study, a data-driven and generative machine learning method, called multiscale generative adversarial network (MS-GAN), is adopted to integrate limited borehole logs with geophysical data from a nearby site for developing 3D subsurface geological models with quantified uncertainty. The geophysical data from a nearby site, or a site with similar geological settings, are used to construct a 3D training image for MS-GAN to learn stratigraphic patterns and generate 3D subsurface geological models conditioned on limited boreholes from a specific site. An enhancement is also made to MS-GAN for handling incomplete borehole logs, which is often encountered in practice. The method is demonstrated and validated through a case study in Hong Kong. The results show that MS-GAN can effectively integrate limited borehole logs with geophysical data from a different site to generate 3D subsurface geological models with high accuracy and quantified uncertainty.