Gravel soil faces significant engineering challenges such as leakage erosion and soil flow due to its complex composition and susceptibility to groundwater effects. This study integrates the entire machine learning process, including pre- and post-processing of images, WGAN implementation, and validation of hydraulic and morphological properties. Obtaining intact gravel soil samples is difficult and costly due to their erodible nature in the Li River, China. A μ-CT scanning series is employed to capture detailed images with three microstructural characteristics of gravel soil, forming the basis for training datasets using WGANs. This approach allows the generation of similar 3D realizations that replicate the microstructural characteristics and hydraulic behaviors of a prototype of gravel soils. Through computational fluid dynamics (CFD) simulations, the effectiveness of the realizations in hydraulic behavior within reconstructed porous structures is verified. This process indirectly validates the consistency between the realization′s microstructure and the prototype. This integrated methodology not only enhances understanding but also aids in the optimization of engineering designs and applications in geotechnical and materials science disciplines.
Read full abstract