Deep Learning-Based Reconstruction of 3D Morphology of Geomaterial Particles from Single-View 2D Images

Area of Science:

• Geology
• Materials Science
• Artificial Intelligence

Background:

• Particle morphology contains critical information about formation environments.
• Accurate three-dimensional (3D) reconstruction of particle morphology is essential for detailed analysis.
• Traditional methods for 3D reconstruction can be time-consuming and complex.

Purpose of the Study:

• To develop and validate an artificial intelligence (AI) based method for reconstructing 3D particle morphology from 2D images.
• To assess the model's generalizability across naturally formed, manufactured, and digitally generated particles.
• To provide a rapid and effective tool for generating 3D numerical representations of geological particles.

Main Methods:

• A deep learning approach utilizing voxel representation and multi-dimensional convolutional neural networks was employed.
• Over 100,000 particles from diverse sources (desert sand, lunar soil simulant, digital particles) were used for training and testing.
• The reconstructed morphologies were evaluated using multi-scale morphological descriptors and compared against real 3D particles.

Main Results:

• The AI model successfully reconstructed the 3D morphology of particles from 2D images across all tested types.
• Statistical properties of morphological descriptors from reconstructed particles were consistent with real 3D particles.
• The model demonstrated validity and generalizability for diverse particle samples.

Conclusions:

• The proposed AI method provides a rapid and effective means for 3D particle morphology reconstruction from 2D images.
• This technique facilitates in-depth analysis of particle properties, including mechanical behavior and transport characteristics.
• The study offers a valuable approach for generating 3D numerical models of geological particles for further research.