Smart Sensing of Subsurface Soil Liquefaction Potential Using Machine Learning and Virtual Reality

Description

Soil liquefaction is a phenomenon in which underground loose saturated soil suddenly loses strength and stiffness during an earthquake or under dynamic loading. Subsurface soil liquefaction often results in catastrophic failure of buildings and civil infrastructures resting on liquefied ground, as frequently observed during devastating earthquakes, leading to tremendous economic losses and severe casualties. Soil liquefaction-induced damage to buildings and infrastructures is strongly influenced by the 3D spatial distribution of subsurface liquefiable soils. Therefore, it is crucial to properly delineate the spatial distribution of soil liquefaction potential to assess and mitigate liquefaction-related hazards. The liquefaction potential of subsurface soil is usually evaluated using in-situ tests, such as cone penetration tests (CPTs). However, within a site, CPTs are only conducted at limited locations, owing to time constraints, cost, and limited access to the subsurface, resulting in sparse CPT soundings. This poses a great challenge on how to accurately delineate the 3D spatial distribution of soil liquefaction potential from sparse CPT soundings. Furthermore, evaluation of soil liquefaction potential involves several uncertainties, such as uncertainties in seismic loading, model uncertainties associated with the CPT-based liquefaction assessment method, and spatial variability of subsurface soil. These uncertainties and variabilities greatly affect the assessment of liquefaction potential and its 3D spatial delineation. The proposed project aims to develop a machine learninga (ML)-based smart sensing strategy and virtual reality (VR) system with software for real-time evaluation and visualization of the 3D spatial distribution of subsurface soil liquefaction potential from sparse CPT soundings. The developed methods quantitatively model the associated uncertainties/variabilities and utilize data-driven Bayesian supervised learning to analyze sparse CPT data in real time for delineating the 3D spatial distribution of soil liquefaction potential and quantifying its uncertainty. Based on the quantified uncertainty, the methods recommend subsequent CPTs and their optimal locations in an automatic and self-adaptive manner for the real-time updating of liquefaction potential assessment results. The PI has performed preliminary studies on 2D cases and obtained promising results. A VR system with software will also be developed to dynamically visualize the 3D results and facilitate applications of the proposed methods. Leveraging on recent unprecedented advances in ML and VR, this project will develop innovative methods and immersive geotechnical tools for assessing and mitigating liquefaction-related hazards. The research outcomes will be beneficial to geotechnical practice worldwide and in Hong Kong, where many reclamations constructed before 1970 might be susceptible to liquefaction under dynamic loading.