InterPore2026

A Graph Neural Network Framework for Upscaling the Pore Network Modeling Calculations

21 May 2026, 14:20

15m

Oral Presentation (MS15) Machine Learning in Porous Media MS15

Speaker

Mehdi Mahdaviara (Hydrogeology group, Utrecht University)

Description

This study proposes an artificial intelligence (AI)-based framework for upscaling single-phase and two-phase quasi-static simulation results from small subsamples to larger porous media domains. Several simulation methods, including direct numerical simulation (DNS) and pore-network modeling (PNM), are employed to elucidate the transport phenomena within the pore space. While in DNS, the pore space geometry is directly discretized, in PNM, the complex pore morphology is reduced to a simplified network of pores and throats with idealized geometries [1], drastically reducing the computational requirements [2]. Notwithstanding, it remains computationally demanding when applied to very large samples.
To address this challenge, we utilize graph neural networks (GNNs) for upscaling the pore pressure and capillary pressure results from small to large 3D samples. The GNNs are powerful machine learning frameworks capable of directly learning from graph-structured data, such as pore networks [3, 4]. The core principle of a GNN is the iterative aggregation and transformation of information exchanged between interconnected neighboring nodes (pores) [4].
Our framework begins with a binarized tomography of the porous medium, from which both a subsample and the full sample are selected (see Figure). Pore networks are extracted for each, but fluid flow simulations are performed only on the small subsamples to reduce computational expense. The extracted pore network of the subsample is used as input to the GNN, while the node-level fluid flow simulation results serve as the training targets. The GNN is thus trained to predict flow parameters directly from graph data. Once trained, the model is applied to the pore network of the full sample to predict the same flow parameters without additional simulations.
The framework was evaluated using three X-ray tomography images of sandstone samples, including Bentheimer, Castle Gate, and Berea. Results demonstrate that the proposed method achieves high accuracy in upscaling pore pressure and capillary pressure from subsamples to full rock volumes. For instance, the upscaling from the train image dimensions of 2003, 4003, 6003, and 8003 to a validation image of 10003 was conducted, yielding R-squared values of 0.83, 0.91, 0.96, and 0.98, respectively. The training took ~20 seconds, and the upscaling took ~3 seconds, indicating the very computational efficiency of the method. Further assessment indicated the model's ability for transfer learning. While the model was trained on the Bentheimer data, the capillary pressure of the Castle Gate sample is successfully predicted by an R-squared of 0.96.