Microstructure reconstruction of 2D/3D random materials via diffusion-based deep generative models.

Area of Science:

• Materials Science and Engineering
• Computational Materials Science
• Artificial Intelligence in Materials Design

Background:

• Establishing process-structure-property (PSP) relationships is vital for material design.
• Existing generative models like VAEs and GANs have limitations in learning complex data distributions.
• Accurate microstructure reconstruction is a key challenge in materials informatics.

Purpose of the Study:

• To develop and apply advanced generative models for high-fidelity material microstructure reconstruction.
• To explore the capabilities of denoising diffusion probabilistic models (DDPM) beyond traditional methods.
• To enable conditional generation of microstructures with controlled properties, including 3D structures and permeability.

Main Methods:

• Utilized denoising diffusion probabilistic models (DDPM) to learn probability distributions of high-dimensional microstructure data.
• Employed denoising diffusion implicit models (DDIM) for microstructure randomness regulation and gradient material generation via latent space interpolation.
• Extended 2D reconstruction to 3D, integrating permeability via feature encoding and validating with the Lattice Boltzmann Method.

Main Results:

• Successfully reconstructed diverse microstructures (inclusion, spinodal, chessboard, fractal noise) with high fidelity.
• Achieved quantitative control over microstructure randomness and generated gradient materials.
• Enabled conditional generation of 3D porous microstructures with targeted permeability, validated by Lattice Boltzmann simulations.

Conclusions:

• DDPM offers a powerful framework for learning complex microstructure distributions and enabling accurate reconstruction.
• The developed methods provide novel approaches for material reverse design and property-guided microstructure generation.
• This work paves the way for advanced computational material design by bridging generative modeling and physical property prediction.