Abstract
Capillary pressure (Pc) and relative permeability (kr) functions are fundamental to characterizing multiphase fluid displacement in porous media, yet accurately predicting them across scales remains a major challenge. Conventional empirical correlations fail to capture discrete pore-scale dynamics, such as Haines jumps, while classical pore network modeling (PNM) is computationally prohibitive for large-scale domains. To bridge this gap, we present a hybrid deep learning framework combining a Graph Neural Network (GNN) and Multilayer Perceptron (MLP) that integrates micro-scale pore geometry and topology with macroscopic descriptors to predict full Pc and kr curves. Leveraging the inherent size invariance of GNNs, this framework enables the direct spatial upscaling of digital rock physics simulations. The architecture was trained and validated on an extensive dataset of 124,876 3D pore networks extracted from micro-computed tomography images (2003 and 4003 voxels) of five distinct sandstones. Topological analysis demonstrates that explicitly embedding the pore network structure captures critical micro-scale displacement fluctuations that are typically obscured by core-scale averaging and traditional empirical models. Furthermore, the framework successfully upscales simulations to 10003-voxel domains, achieving high predictive accuracy with R2 values exceeding 0.97 for Pc, 0.96 for non-wetting relative permeability (krnw), and 0.94 for wetting relative permeability (krw). Blind evaluations on unseen sandstone and highly complex carbonate architectures confirm the robust generalization capability of the model. Finally, the framework delivers orders-of-magnitude computational speedups, reducing kr and Pc simulation times for a 10003-voxel sample from 103 to 16 seconds and 70 to 8 seconds, respectively.

This publication has 67 references indexed in Scilit: