Multiscale Machine Learning of Subsurface Stratigraphy from Limited Site-specific Measurements and Prior Geological Knowledge using Iterative Convolutional Neural Networks (CNN)

Description

Subsurface stratigraphy depicts the occurrence and spatial distribution of different subsurface soil/rock strata as an image and provides indispensable information for design and construction of geotechnical structures. It is developed from site investigation data and often supplemented by prior geological knowledge and engineering judgement. However, site investigation data from a specific site are often quite limited (e.g., a few boreholes) in engineering practice, and the obtained stratigraphy might be inaccurate, leading to the most frequently encountered problem in geotechnical construction: unexpected site conditions. For example, unexpected subsurface granite was encountered during the construction of the Guangzhou-Shenzhen-Hong Kong Express Rail Link project commissioned in 2018, which delayed the project by more than 2 years and increased the cost by HK$20 billion. This problem is exacerbated for Building Information Modelling (BIM) or digital twins of civil infrastructures in which highresolution 3D stratigraphy is required to represent the subsurface soils and rocks. The goal of this project is to resolve this long-standing problem of developing accurate subsurface stratigraphy from sparse measurements by recognizing that stratigraphy is an image and can be leveraged using an image-based deep machine learning method called Convolutional Neural Networks (CNN). However, direct application of CNN to subsurface stratigraphy is not feasible because millions of images are required to train a CNN model and abundant measurements are required for predictions using the trained CNN model. Neither millions of images nor abundant measurements are available in geotechnical practice. In this project, a novel iterative CNN model is proposed to tackle these difficulties. In the proposed model, only one training image is needed for iteratively extracting its spatial stratigraphic patterns from large to small scales progressively. The extracted patterns at different scales are used together with limited site-specific measurements to iteratively develop the stratigraphy. The PI has developed a preliminary iterative CNN method for 2D geological cross-sections and obtained promising results. This project will enhance the developed 2D method, systematically extend it to 3D cases, compile a database of training stratigraphy with different geological origins, and develop computer software to facilitate its application to engineering practice. The output from this project will be a game changer that transforms the current engineering practice of manually developing highly simplified subsurface stratigraphy into automatic generation of high-resolution subsurface stratigraphy with uncertainty quantification. The research outcomes will benefit geotechnical practice worldwide and in Hong Kong specifically, where many large-scale infrastructure projects are currently in progress.