Machine Learning of Subsurface Geological Model for Assessment of Reclamation Induced Consolidation Settlement

Reclamation is an effective method to create buildable lands for congested coastal megacities such as Hong Kong and Macau. The greatest geotechnical risk associated with reclamation works is consolidation, which is a time-dependent process of pore water expulsion and ground settlement. An accurate evaluation of consolidation requires a sound understanding of spatial distribution of subsurface soil layer boundaries and spatial variability of soil consolidation parameters from limited site-specific measurements such as cone penetration tests. It is common practice to determine subsurface stratigraphic boundaries using straight lines to connect the same stratigraphy revealed from adjacent measurements, and assume deterministic soil consolidation parameters for consolidation analysis. This simplified practice gains popularity among engineering practitioners due to its convenience for implementation. However, great difficulties may occur when complex geology (e.g., interbedded soil layers) is encountered. More importantly, a false interpretation of subsurface stratigraphy from limited data may fail to identify the most critical design scenario, thus pose significant risks to safety and serviceability of a geotechnical system. In this study, a unified framework is proposed to assess reclamation induced consolidation settlement with explicit consideration of stratigraphic uncertainty and spatial variability of consolidation parameters. Consolidation settlements associated with different combinations of geological realizations and geotechnical random field samples are calculated using the classical 1D consolidation theory. Performance of the proposed unified framework is demonstrated using an illustrative example. Results indicate that the framework can provide accurate evaluation of ground differential settlement with quantified uncertainty.

Keywords: Probabilistic analysis, Geological uncertainty, Convolutional neural network, Bayesian Compressive Sensing.