Latent diffusion modeling of porous media informed by spatial statistics

Area of Science:

• Materials Science
• Computational Modeling
• Geophysics

Background:

• Accurate modeling of porous media is essential for understanding fluid transport, mechanical properties, and multiphysics phenomena in various materials.
• Existing methods for generating porous media can be computationally intensive or lack control over structural properties.

Purpose of the Study:

• To develop a novel generative framework for accurate and efficient modeling of porous media structures.
• To leverage latent diffusion models (LDMs) trained on stochastic simulations and experimental data for enhanced generation capabilities.

Main Methods:

• A hybrid approach combining stochastic simulation for generating high-resolution porous samples with latent diffusion models (LDMs).
• Utilized a variational autoencoder for dimensionality reduction and a UNet denoiser for latent-space generation within the LDM pipeline.
• Trained the LDM on a statistically controlled dataset derived from stochastic simulations, preserving spatial correlation statistics.

Main Results:

• The developed LDM successfully learns a data-driven prior over porous structures, preserving key statistical properties.
• Demonstrated unconditional generation of realistic porous media models with high visual diversity and generative fidelity.
• The generated models accurately reflect input spatial statistics.

Conclusions:

• The hybrid generative framework offers a promising pathway for fast, scalable, and controllable porous media generation.
• This approach is particularly beneficial in data-scarce or simulation-heavy research settings.
• The method enhances the ability to study fluid transport and material properties in complex porous systems.