A computationally efficient hybrid neural network architecture for porous media: Integrating convolutional and graph neural networks for improved property predictions

Porous media is widely distributed in nature, found in environments such as soil, rock formations, and plant tissues, and is crucial in applications like subsurface oil and gas extraction, medical drug delivery, and filtration systems. Understanding the properties of porous media, such as the permeability and formation factor, is crucial for comprehending the physics of fluid flow within them. We present a novel fusion model that significantly enhances memory efficiency compared to traditional convolutional neural networks (CNNs) while maintaining high predictive accuracy. Although the CNNs have been employed to estimate these properties from high-resolution, three-dimensional images of porous media, they often suffer from high memory consumption when processing large-dimensional inputs. Our model integrates a simplified CNN with a graph neural network (GNN), which efficiently consolidates clusters of pixels into graph nodes and edges that represent pores and throats, respectively. This graph-based approach aligns naturally with the porous medium structure, enabling large-scale simulations that are challenging with traditional methods. Furthermore, we use the GNN Grad-CAM technology to provide new interpretability and insights into fluid dynamics in porous media. Our results demonstrate that the accuracy of the fusion model in predicting porous medium properties is superior to that of the standalone CNN, while its total parameter count is nearly two orders of magnitude lower. This innovative approach highlights the transformative potential of hybrid neural network architectures in advancing research on fluid flow in porous media.