Using deep learning to capture gravel soil microstructure and hydraulic characteristics

Area of Science:

• Geotechnical Engineering
• Computational Geoscience

Background:

• Gravel soil hydraulic properties are complex, influenced by fine particle content and microstructure.
• Accurate analysis of these properties is vital for geotechnical applications like seepage and stability assessments.
• Traditional methods may struggle to capture the intricate pore characteristics influencing hydraulic behavior.

Purpose of the Study:

• To apply Wasserstein Generative Adversarial Networks (WGANs) for reconstructing 3D digital gravel soil samples.
• To generate specific microstructure realizations and analyze their hydraulic properties.
• To validate the WGAN model's ability to represent actual soil behavior and improve upon existing machine learning models.

Main Methods:

• Utilized the WGAN with Gradient Penalty technique for digital sample reconstruction.
• Employed µ-CT scanning to create a training dataset from Guilin gravel soil samples.
• Evaluated reconstructed samples based on porosity, correlation functions, specific surface, Euler characteristics, and permeability.

Main Results:

• Reconstructed gravel soil models showed high consistency with original samples in microstructural parameters.
• Permeability evaluations confirmed that reconstructed realizations accurately represented the soil prototype.
• The WGAN model effectively captured hydraulic property variations due to coarse and fine particle size differences, unlike previous models.

Conclusions:

• WGANs provide a powerful tool for analyzing gravel soil hydraulic properties by reconstructing realistic microstructures.
• The developed model accurately represents pore characteristics and predicts seepage behavior.
• This technique offers a significant advancement in modeling gravel soil hydraulics, particularly concerning particle size influences.