Microstructural tracking of undisturbed loess during compaction and collapse using micro-CT and soil-particle-aware model

The collapsibility of loess significantly increases the risk of rainfall-induced engineering geological disasters in loess regions. The unique microstructure of loess is the primary factor determining its macroscopic collapsibility. Combined mechanical testing and Micro-CT imaging enable the correlation between microscopic and macroscopic deformation within the same specimen. However, the micro-CT images of representative-size specimens contain significant noise, making accurate segmentation of particles and pores in CT-based reconstructions challenging. To address these limitations, we developed a small consolidation device compatible with micro-CT scanning. A soil-particle-aware model based on the artificial neural network was used to identify particle contours in CT images. The results demonstrate that during compaction, loess exhibited a densified particle arrangement with slight particle inclination. During the collapse, the reorientation of particles from a vertical to a horizontal orientation caused large deformation of the specimen. The collapsed macropores are transformed into multiple isolated pores, resulting in a homogeneous and dense structure for the specimen. For the same loess specimen undergoing compaction and collapse processes, the proportion of particles with a 0°−30° inclination gradually decreased from 19% to 16% and then to 13%. At the same time, the volume proportion of pores with diameters less than 20 μm increased from 5% to 8% and then to 25%. Therefore, compaction densifies the particle arrangement, and the overall structure remains stable. In contrast, collapse induces large deformation through significant particle reorientation and pore collapse, ultimately triggering structural failure in the undisturbed loess framework.