Deep learning for diffusion in porous media | JoVE Visualize

Area of Science:

Computational physics
Materials science
Biophysics

Background:

Porous media characterization is crucial for applications in geology, chemical engineering, and biomedical fields.
Predicting effective properties like porosity and diffusion is essential for understanding fluid flow and transport phenomena.
Convolutional neural networks (CNNs) offer a promising approach for analyzing complex microstructures and predicting material properties.

Purpose of the Study:

To investigate the application of CNNs for predicting basic properties of porous media.
To compare the performance of different CNN architectures (C-Net, U-Net) in predicting porosity and effective diffusion coefficients.
To evaluate the transferability of trained models across different porous media types, specifically sand packings and biological tissues.

Main Methods:

Utilized the Lattice Boltzmann Method to generate labeled data for supervised learning.
Developed and modified CNN models, including C-Net and U-Net architectures with self-normalization modules.
Trained and tested models on two distinct porous media datasets: sand packings and biological tissue-mimicking systems.
Applied Archie's law to determine tortuosity from porosity and effective diffusion data.

Main Results:

CNNs trained on geometric analysis achieved reasonable accuracy but lacked transferability between sand and biological media.
The U-Net architecture demonstrated high accuracy in reconstructing concentration fields.
Models trained on one type of porous medium (e.g., sand) generalized effectively to the other (e.g., biological tissue) for concentration map reconstruction.
Tortuosity was successfully derived using Archie's law, linking effective diffusion to porosity.

Conclusions:

CNNs can effectively predict porous media properties, with U-Net showing superior performance and transferability for concentration field reconstruction.
The choice of CNN architecture significantly impacts model generalizability across different porous media types.
The study highlights the potential of deep learning for characterizing complex porous materials in diverse scientific domains.