Physics-informed and Interpretable Machine Learning of Reclamation-induced Consolidation of Soil Using Sparse Site Investigation and Field Monitoring Data

Description

Machine learning (ML) has achieved tremendous successes in many fields and attracted increasing interest in geotechnical applications, e.g., analysis of soil consolidation induced by reclamations. However, ML development and applications in geotechnical practice are challenging, because extensive training databases are a key to the success of ML, but not available in geotechnical practices. Geotechnical data are often project-specific and limited, leading to a difficulty of developing a suitable database required for training ML. In addition, ML is frequently criticized for its black-box nature. A lack of suitable database and model interpretability are two major obstacles for ML development and applications in geotechnical engineering. This proposed project will address these two obstacles and develop physics-informed and interpretable ML methods for analysis of soil consolidation induced by reclamation. In the proposed ML methods, a project-specific database is generated for each reclamation project through finite element analyses. This database encompasses both project-specific data collected at various project stages (e.g., sparse site investigation data, construction sequences such as preloading) and domain knowledge (e.g., soil mechanics, numerical analysis principles, and engineering judgment). Subsequently, the project-specific database is combined with field monitoring data under a sparse dictionary learning framework to provide accurate, real-time, and interpretable predictions of soil consolidation responses at locations without monitoring data or for future time steps. The PI has developed a preliminary ML method using settlement monitoring data and obtained promising results. In the proposed project, the ML methods will be further developed to incorporate multi-source monitoring data and consider various soil constitutive models and parameters. The interpretability of the ML methods will be enhanced using a game theory-based approach. The proposed methods will be calibrated using real data from reclamation projects, and a software package will be developed to facilitate their applications. This proposed project aims to accelerate a paradigm shift in geotechnical practices towards digital transformation and intelligence. The research outcomes will provide effective and interpretable ML methods and software that may revolutionize geotechnical (e.g., reclamation) practice worldwide. Reclamation-induced soil consolidation responses are spatiotemporally varying, and their accurate prediction is critical for successful delivery of reclamation projects, particularly for shortening project duration and saving cost. This will have tremendous economic and societal impacts, particularly in Hong Kong, where several large-scale reclamation projects are in progress, including the “Lantau Tomorrow Vision” project, which aims to create approximately 1700 hectares of new lands to address the pressing land shortage issue.