Tracking the evolution of loess microstructure using microscopic CT 3D reconstruction based on a soil particle perception model

Have you ever wondered how the microstructure of loess changes after being subjected to various external forces? This study, through innovative microscopic CT 3D reconstruction technology combined with deep learning, precisely tracks the microstructural evolution of loess under consolidation pressure, helping us better understand the mechanical properties and engineering behavior of loess.

Background:
As a typical aeolian soil, loess possesses important engineering geological characteristics due to its high porosity, compressibility, and collapsibility. However, the microstructure of loess is the fundamental reason determining these characteristics. Due to the limitations of traditional CT technology in resolving particle details, especially the blurring of particle contours caused by noise, accurate acquisition is difficult. Against this background, how to improve the resolution of microscopic CT images and the accuracy of particle identification has become a research challenge.

Research objectives:
This paper proposes a method based onconvolutional neural networks(CNN) and content-aware image restoration (CARE) theory, aiming to reconstruct a more precise 3D digital model of loess. Through this method, soil particle contours in CT slices can be effectively identified and segmented, further tracking the microstructural evolution of loess under different consolidation pressures.

Research highlights:

1. Innovative experimental design

   : Through the combination of a small consolidation test apparatus with microscopic CT scanning equipment, real-time tracking of microstructural changes in air-fallen loess specimens under consolidation pressure was achieved.

2. Precise particle identification

   : Combined with a deep learning-based soil particle perception model, the contours of loess particles were successfully identified, and particle size distribution curves under different consolidation pressures could be accurately plotted.

3. Mechanical analysis

   : Through precise 3D digital models and the discrete element method (DEM), mechanical parameters such as the mechanical coordination number, contact force distribution density, and pore throat diameter of loess under vertical consolidation pressure were analyzed.

4. 

Materials and methods:

1. Loess microstructure evolution tracking system

   : Including a microscopic CT scanning system and a loading subsystem, high-quality microstructure images can be acquired under different pressure conditions through a systematic experimental scheme.

2. U-Net convolutional neural network

   : The U-Net architecture was used for soil particle segmentation, solving the difficulty of particle identification in high-noise CT slices encountered by traditional methods.

3. Soil particle perception model

   : By adding noise to original CT slices and training the neural network model, the segmentation accuracy of the model under noisy conditions was effectively improved.

4. Pore throat diameter and pore volume fraction analysis

   : For loess samples under different consolidation pressure conditions, the relationship between pore throat diameter and pore volume fraction was analyzed.

5. Normal contact force distribution

   : Based on DEM simulation, the mechanical contact characteristics between particles were analyzed, revealing the mechanical behavior of soil particles under pressure.

6. 

Conclusions:

1. The proposed neural network model training method effectively improved the ability of traditional methods to identify soil particles in high-noise CT slices.
2. Experimental results show that the pore throat diameters calculated based on the rectangular cuboid short-edge model are very close to theMIPtest results, with a deviation of only 2.26 μm, verifying the accuracy of the model.
3. Analysis of the DEM model established from the 3D digital model shows that the normal contact force distribution between particles exhibits a typical butterfly shape, accurately reflecting the biaxial stress state of the specimen in theconsolidation test.

Reproducibility:The innovations of this paper are mainly reflected in the improvement of the particle segmentation method and the subsequent analysis of normal contact forces. By improving CT image quality and particle segmentation accuracy, the evolution process of soil particles under different consolidation pressures can be tracked more precisely.

This study not only provides a new analytical method for the microstructural evolution of soils such as loess, but also lays a technical foundation for future research on similar materials. It has broad application prospects in improving the resolution of CT images and the accuracy of particle identification, and has far-reaching impacts on fields such assoil mechanics, engineering geology, and environmental science.