Improved Computer Vision Framework for Mesoscale Simulation of Xiyu Conglomerate Using the Discrete Element Method

Abstract

The complex mechanical characteristics of the Xiyu conglomerate significantly influence the resistance and deformation features of its caverns’ surrounding rock, thereby constraining the construction of related water diversion tunnels. This paper introduces an improved SegFormer framework developed for the detection of mesoscale geomaterial structures. Computerized tomography (CT) scan images of the Xiyu conglomerate were employed to establish a high-precision numerical model. From the results of segmentation, the proposed algorithm outperformed UNet, HRNet, and the original SegFormer neural network. The segmentation results were used to calculate the porosity, and biaxial compression numerical simulation experiments based on the real structure were carried out using the particle flow code (PFC). We observed the failure process of the model and obtained the shear strength of the Xiyu conglomerate. We explored the causes and influencing factors of the anisotropy of the Xiyu conglomerate from the microstructure perspective and provide a micro-observation basis for establishing an anisotropic mechanical model.

Keywords

• Xiyu conglomerate
• machine learning
• digital image
• discrete element method
• mechanical properties