Tracking the evolution of loess microstructure using micro-CT 3D reconstruction based on a soil-particle-aware model

Abstract

The particle contours in the loess in-situ micro-CT test slices are always blurred by noise because the representative size of the mechanical test specimen is larger than the CT scanning area at adequate resolutions. An enhanced approach was proposed using convolutional neural networks and content-aware image restoration theory to construct a more precise 3D digital loess model. First, a training dataset was created by adding noise to clear CT slice images from fine tube loose soil specimens. Then, a conventional U-Net architecture neural network model was trained using this dataset to develop a soil-particle-aware model for high-noise micro-CT slice images of soil particles. Validation experiments, combining a small consolidation test apparatus and micro-CT scanning equipment, were conducted to track the microstructural evolution of air-fall loess specimens under consolidation pressure using the proposed method. The results indicate that the proposed method effectively identifies loess particle contours and accurately determines the particle size distribution curves under varying consolidation pressures. The issue of missing small particles in reconstructed models was addressed successfully. Finally, the mechanical coordination number, contact force distribution density, and peak pore throat diameter of air-fall specimens developed under vertical consolidation pressures were analysed using precise 3D digital models and the discrete element method.