Development of Machine Learning-Based Software for Geotechnical Subsurface Characterization: Case Studies from Singapore

Due to complex spatial variability in subsurface stratigraphy and the sparse and non-uniform nature of site investigation data, traditional approaches often fail to accurately capture the boundary features and spatial distributions of different subsurface strata. Consequently, 3D subsurface reconstruction remains a significant geotechnical engineering challenge. Machine learning and big data provide an opportunity to overcome this long-standing challenge. This paper presents the development of machine learning (ML)-based software for data-driven geotechnical subsurface stratigraphy characterization. This software incorporates two novel ML algorithms developed in-house, namely Multi-scale Meta-learning Model (M3) and Digitization-based Neighborhood Aggregation (DNA), based on data from more than 20,000 boreholes surveyed across Singapore. In this paper, detailed formulations of the two ML algorithms and how the two algorithms collectively provide complete, accurate, and robust 3D subsurface stratigraphy profiles are first described. Additional key functionalities of the software, such as data entry, data visualization, stratigraphy interpretation, visualization of geological features, and result and uncertainty visualization, are also described to demonstrate its capacity for explicit 1D to 3D subsurface characterization, uncertainty quantification, and associated visualization. In addition, this software is designed to be amenable to new geotechnical data, thereby permitting continuous refinement and updating of the algorithms. Notably, this software has already been employed by the Land Transport Authority of Singapore for ongoing underground construction projects and holds the potential to facilitate various geotechnical engineering applications. Success stories are presented as case studies, together with comments from the Land Transport Authority of Singapore. The advantages of the software over existing techniques are also highlighted.