Deep learning-based methods for segmentation and labelling of clay

Abstract

Recent advancements in x-ray technology have enabled non destructive 3D sub-micron imaging of clay. In this work, a 3D tomography of kaolinite particles is analysed. Conventional segmentation algorithms are used along with a deep learning-aided method, which brings novelty to the clay research area. The clay image is segmented to acquire morphological properties of the material, intended to be used to inform continuum models for clay used at the engineering scale. Imaging clay is especially challenging because of its small particle size and thin aggregated platelets. A small clay dataset will be developed to evaluate the performance of segmentation techniques and to train a machine learning-model called the Segment Anything Model 2 (SAM 2). Advanced contemporary studies in biomedical segmentation show compelling results using SAM 2 and this study is proposing to bring this technique into the area of geomechanics. This study aims to lay a path for future research to strengthen the link between physical relationships and observed clay behaviour by providing information of clay micro-structures.