Speaker
Description
Modeling the process of multiphase flow through porous media is an essential task as it is involved in many mitigation technologies, including CO2 geological storage, hydrogen storage, and fuel cells. To study these multiphase flow processes, scientists can utilize micro-CT scanners with synchrotron sources to image the rock pores in a nanometer-scale spatial resolution while recreating the in situ condition of gas flowing through the porous medium. State-of-the-art research facilities can now generate 3D imaging data with billions of voxels with temporal resolution in the order of seconds, creating a unique opportunity to advance our understanding of fluid flow physics.
However, despite the advancement in experimental capability to obtain high-resolution experimental data, the modeling of pore-scale multiphase flow in porous media remains very challenging. Current modeling approaches generally fall into three major classes: lattice/particle-based models, continuum models, and pore-network models. The former two approaches solve governing equations on the pore scale but are often extremely computationally expensive. Pore network models simplify porous media into interconnected pores and throats, offering computational efficiency but usually lacking accuracy due to the complexity of the physics involved.
Machine learning-based approaches are emerging in recent pore-scale modeling studies. However, most existing studies aim to learn fluid flow physics from simulation data, which suffers from significant drawbacks. Due to the aforementioned computational challenges in pore-scale modeling, training datasets are often incapable of reflecting the accurate behavior of real-world multiphase flow. They are often simulated with highly simplified pore geometry (e.g., packed spheres) or simplified physics (e.g., negligible viscous effects and assumptions of steady-state flow). In addition, many previous approaches were designed based on convolutional neural networks (CNNs) that are optimized for grid-like structures. As a result, they struggled with the irregular geometries of complex pore structures.
In this study, we address this challenge using a graph neural network-based approach and directly learn pore-scale fluid flow using micro-CT experimental data. We used graph structure to represent irregular pore structures, where nodes represent fluid volume and edges represent fluid flow paths. Unlike traditional CNNs, the graph-based representation can handle complex geometries by passing messages only between connected nodes, ignoring nodes separated by solid rock grains. Learning the fluid flow dynamics using a GNN-based architecture leverages inductive biases to focus on local interactions between fluid and pore structures. This allows for zero-shot generalization to other boundary conditions or bigger input fields through section-based training. During inference, given an initial state, the model can autoregressively predict the evolution of the multiphase flow process over time.
Learning directly from experimental data allows us to replicate realistic flow phenomena, including those that remain beyond the reach of simulations due to incomplete understanding. Our results demonstrate that GNN can effectively capture complex fluid behaviors and generalize well across varying boundary conditions, providing a promising direction to leverage the extensive micro-CT experimental data to study pore-scale physics.
| Country | UK |
|---|---|
| Acceptance of the Terms & Conditions | Click here to agree |
