Data-Driven Development of Three-Dimensional Subsurface Geological Model from Limited Boreholes and Prior Geological Knowledge for Site Characterization

The delineation of subsurface stratigraphy is an indispensable task in geotechnical site characterization, which benefits subsequent geotechnical design and construction. In recent years, three-dimensional (3D) subsurface geological modelling has received increasing attention because it can realistically represent subsurface stratigraphic variations at a specific site. Geotechnical site investigation data (e.g., boreholes) are necessary for developing 3D subsurface geological models. Due to time, budget, or technical constraints, measurements are commonly limited and sparse in engineering practice, leading to a long-lasting challenge in properly developing 3D subsurface geological models. To tackle this challenge, a generative and data-driven machine learning method called multi-scale generative adversarial network (MS-GAN) is adopted in this study. This method can properly learn 3D stratigraphic information from a 3D training image representing prior geological knowledge for automatically generating 3D subsurface geological models conditioned on limited boreholes in a data-driven manner. The performance of this method is illustrated using a simulated data example. The results indicate that MS-GAN not only accurately and effectively delineates stratigraphic variations in 3D space, but also quantifies stratigraphic uncertainty from multiple 3D realizations.