A Computationally Efficient Hybrid Neural Network Architecture for Porous Media: Integrating CNNs and GNNs for Improved Permeability Prediction

Subsurface fluid flow, essential in various natural and engineered processes, is largely governed by a rock's permeability, which describes its ability to allow fluid passage. While convolutional neural networks (CNNs) have been employed to estimate permeability from high-resolution 3D rock images, our novel visualization technology reveals that they occasionally miss higher-level characteristics, such as nuanced connectivity and flow paths, within porous media. To address this, we propose a novel fusion model to integrate CNN with the graph neural network (GNN), which capitalizes on graph representations derived from pore network model to capture intricate relational data between pores. The permeability prediction accuracy of the fusion model is superior to the standalone CNN, whereas its total parameter number is nearly two orders of magnitude lower than the latter. This innovative approach not only heralds a new frontier in the research of digital rock property predictions, but also demonstrates remarkable improvements in prediction accuracy and efficiency, emphasizing the transformative potential of hybrid neural network architectures in subsurface fluid flow research.