Data-driven multi-stage sampling strategy for a three-dimensional geological domain using weighted centroidal voronoi tessellation and IC-XGBoost3D

Conventional site investigation schemes typically use an empirical sampling strategy, which involves equal sampling spacing with regular grid patterns. However, this approach assigns equal importance to the entire site and does not account for subsurface stratigraphic variations and site constraints. In this study, a data-driven multi-stage sampling strategy is proposed to adaptively optimize the borehole locations for a three-dimensional (3D) geotechnical site, considering the subsurface stratigraphic uncertainties and irregular site geometries. The initial sampling plan is determined based on weighted centroidal Voronoi tessellation, which assigns various sampling densities to zones depending on their importance. Measurements obtained in the previous sampling stage are combined with prior geological knowledge to establish and update a 3D geological domain using a physics-informed geological modelling method. The next optimal location for sampling is adaptively determined with the objective of maximizing the reduction in the stratigraphic uncertainty. The proposed method represents the first data-driven 3D site planning method that explicitly considers 3D subsurface stratigraphic variations. The performance of the proposed multi-stage sampling strategy is illustrated using a simulation study. The results indicate that the proposed method efficiently identifies the optimal sampling locations while comprehensively considering the 3D subsurface geological uncertainties and irregular site geometries. © 2023 Elsevier B.V.