InterPore2025

Learning to Simulate Flow through Porous Media with Graph Neural Networks from Experimental Data

19 May 2025, 17:25

15m

Oral Presentation (MS15) Machine Learning and Big Data in Porous Media MS15

Speaker

Linqi Zhu

Description

Fluid flow in porous media plays a crucial role in many environmental and energy sciences applications, including groundwater management, hydrogen and carbon storage, and fuel cells. However, the numerical modeling of such processes remains a very challenging task. Traditional numerical simulation methods often struggle to rapidly and accurately predict pore-scale flow processes in porous media due to high computational costs, complex mesh generation, and difficulty capturing nonlinear flow behavior. To address these challenges, we explore a new modeling framework based on graph neural networks to learn the flow process from experimental data directly. We developed a Pore-scale Graph Network Simulator (Pore-GNS), which embeds pore-structure information in graph construction and introduces soft physical constraints to assist learning. This approach effectively captures the impact of confined spaces on fluid flow while accurately predicting the dynamic evolution of flow fields. By leveraging experimental data of fluid particles for learning, our method can produce particle trajectories for unseen periods and corresponding velocity fields.

Our approaches use real particle motion data obtained from fluid flow in actual porous media as training and validation datasets. The results demonstrate that the Pore-GNS predictions are highly consistent with those from the datasets. Moreover, predicting the single-step trajectories of nearly 1,000 particles takes less than 10 seconds, and generating complete and reliable velocity fields through multi-step autoregressive inference requires only a few minutes, significantly reducing computational overhead. This method shows strong potential for rapid flow prediction in porous media, providing a more efficient solution for geological reservoir simulation, environmental monitoring, and water resource management. As the model scale and multiphysics coupling continue to expand, this approach provides a promising direction for future porous media flow modeling.