Seeing permeability from images: fast prediction with convolutional neural networks

Area of Science:

• Geosciences
• Computational Science
• Materials Science

Background:

• Pore-scale modeling of porous media is crucial for understanding fluid flow.
• Traditional methods for permeability prediction are computationally intensive.
• Image recognition neural networks present a novel approach for rapid analysis.

Purpose of the Study:

• To develop and validate a framework for fast permeability prediction using convolutional neural networks (CNNs).
• To compare the performance of physics-informed CNNs against regular CNNs and traditional methods.

Main Methods:

• Generation of diverse porous media samples.
• Fluid dynamics simulations to compute ground-truth permeability.
• Training of CNNs on simulated data.
• Validation of CNN predictions against simulation results.
• Incorporation of physical parameters into physics-informed CNNs.

Main Results:

• CNNs achieved excellent predictive performance for permeability across various porosities and pore geometries.
• The approach demonstrated significant reduction in computational time compared to fluid dynamics simulations.
• Physics-informed CNNs outperformed regular CNNs, particularly for complex heterogeneities.
• Conventional Kozeny-Carman approach was insufficient for heterogeneous media with dilated pores.

Conclusions:

• Image recognition neural networks provide a powerful and efficient tool for pore-scale permeability prediction.
• Physics-informed CNNs enhance predictive accuracy by integrating domain knowledge.
• This method holds great potential for accelerating research and applications in porous media analysis.