Deep-learning-based porous media microstructure quantitative characterization and reconstruction method.

Area of Science:

• Computational Materials Science
• Materials Informatics
• Porous Media Research

Background:

• Microstructure characterization and reconstruction (MCR) is crucial for understanding porous media behavior and inverse design.
• Existing MCR methods struggle to balance generation quality, characterization capability, and reconstruction accuracy due to complex variables and multiobjective conditions.

Purpose of the Study:

• To develop an improved generative adversarial network for high-quality, controllable, and accurate 3D porous media microstructure generation.
• To enable end-to-end training using unlabeled data for complex microstructures within practical timeframes.

Main Methods:

• Implementation of an improved 3D Porous Media Microstructure (3DPmmGAN) generative adversarial network based on deep learning.
• Utilizing unlabeled data for training complex, high-randomness microstructures.
• End-to-end training approach for efficient microstructure generation.

Main Results:

• The 3DPmmGAN demonstrates high-quality microstructure generation with enhanced controllability and prediction accuracy.
• The model effectively utilizes unlabeled data for training complex microstructures, reducing time consumption.
• The trained network exhibits adaptability to various geometric configurations and allows quantitative control based on semantic conditions.

Conclusions:

• The 3DPmmGAN is a powerful tool for accelerating the preparation and initial characterization of 3D porous media.
• This method has the potential to significantly enhance the efficiency of porous media design.
• The approach facilitates the discovery of processing-structure-property relations in computational materials science.