Machine learning of three-dimensional subsurface geological model for a reclamation site in Hong Kong

Land reclamation from ocean is a major solution to deal with land shortage in coastal megacities such as Hong Kong. The primary geotechnical risk associated with land reclamation is consolidation of fine-grained materials, e.g., soft marine deposit, and a sound understanding of spatial distribution of three-dimensional (3D) subsurface soil layer boundaries, or interfaces, and their stratigraphic connectivity to surrounding drainage boundaries is a prerequisite for an effective reclamation design. In practice, accurate delineation of 3D subsurface stratigraphic boundaries is challenging due to a lack of effective tools for building 3D subsurface geological domains from limited site-specific data while taking full account of geological uncertainty. In this study, a novel stratigraphic modelling and uncertainty quantification method, called 3D iterative convolution eXtreme Gradient Boosting (IC-XGBoost3D), is adopted for automatically developing 3D subsurface geological domains from limited measurements. IC-XGBoost3D roots in deep learning and learns typical stratigraphic features from a pair of perpendicular training images reflecting local prior geological knowledge. The method is physics-informed and data-driven and can efficiently build and update subsurface geological models from limited site-specific data with quantified uncertainty. The method is applied to develop the 3D subsurface geological domain of a reclamation site in Hong Kong. The model performance is evaluated statistically using leave-one-out cross-validation. Results indicate that complex depositional stratigraphic patterns of fine-grained materials at the reclamation site can reasonably be replicated. Effects of measurement data number on the model performance are investigated, and useful insights are gained for developing subsurface geological domains of reclamation sites.