Data-driven construction of Three-dimensional subsurface geological models from limited Site-specific boreholes and prior geological knowledge for underground digital twin

A digital twin is a digital replica or virtual representations of 3D physical entities in the real world. In practice, it is challenging for 3D modelling of subsurface stratigraphy in an underground digital twin due to insufficient site-specific measurements and a lack of efficient 3D spatial prediction tools. In this study, a data-driven and deep learning method, called IC-XGBoost3D, is proposed to build a 3D geological model from limited site-specific boreholes and 2D training images reflecting prior geological knowledge. Anisotropic stratigraphic relationships are learned from two perpendicular 2D training images, and the extracted stratigraphic statistics serve as the input for pre-training a 2D simulation slice. A sequence of 2D simulation slices is then simulated with constraints by site-specific boreholes and subsequently reassembled into a 3D geological model. Note that the parameters of the deep learning method are calibrated with site-specific data and prior training images before being applied to develop the geological model. The model performance is demonstrated and validated using an illustrative example. The proposed method can efficiently generate an anisotropic 3D geological model as a point cloud from two perpendicular training images and sparse boreholes with a high prediction accuracy. More importantly, the proposed method not only infers the most probable 3D geological domain, but also provides a quantitative evaluation of associated 3D stratigraphic uncertainty. Effects of irregular borehole spacing and single training image on the simulation performance are also investigated.