Predicting Effective Diffusivity of Porous Media from Images by Deep Learning

Area of Science:

• Computational physics and materials science.
• Application of artificial intelligence in scientific modeling.

Background:

• Predicting effective diffusivity (D_e) in porous media is crucial for transport phenomena.
• Traditional methods like lattice Boltzmann (LBM) simulations are computationally intensive.
• Machine learning offers a potential alternative for rapid property prediction.

Purpose of the Study:

• To develop and evaluate machine learning models, specifically convolutional neural networks (CNNs), for predicting the effective diffusivity of 2D porous media from structural images.
• To compare the performance of CNN models against traditional empirical models like the Bruggeman equation.
• To investigate methods for improving CNN accuracy, particularly for low diffusivity values.

Main Methods:

• Generating datasets of 2D porous media structures using reconstruction methods.
• Computing effective diffusivity (D_e) for these structures via lattice Boltzmann (LBM) simulations.
• Training and evaluating convolutional neural network (CNN) models on the generated datasets.
• Implementing image processing algorithms to refine pore structure representations for improved CNN predictions.

Main Results:

• Trained CNN models accurately predict effective diffusivity (D_e) with significantly reduced computational cost compared to LBM simulations.
• The optimized CNN model achieves high accuracy (>95% with <10% relative error) for D_e > 0.2 and outperforms the Bruggeman equation, especially for low diffusivity.
• Addressing low diffusivity predictions (<0.1) by incorporating porosity and removing trapped regions improved accuracy, with 70% of predictions having <30% relative error.

Conclusions:

• Deep learning, augmented with domain knowledge (e.g., pore structure analysis), is a powerful tool for predicting porous media transport properties.
• CNN models provide a computationally efficient and accurate alternative to traditional simulation methods for effective diffusivity prediction.
• Further research into integrating physical constraints and advanced image processing can enhance the predictive capabilities of machine learning in porous media science.