Speaker
Description
High-resolution X-ray micro-computed tomography (micro-CT) enables pore-scale characterization of rocks, but extracting representative volumes at high resolution is often computationally prohibitive for routine digital rock workflows. In contrast, lower-resolution scans cover larger domains but systematically miss sub-voxel throats and fine-scale connectivity, leading to biased pore-network graphs and large errors in predicted absolute permeability and multiphase flow responses. This work presents a graph-calibration framework that repairs pore networks extracted from low-resolution micro-CT by explicitly introducing and tuning sub-resolution throats using derivative-free optimization, with the goal of matching laboratory-measured absolute permeability.
Starting from a pore network constructed from a low-resolution scan, we generate a set of candidate throats by geometric proximity (k-nearest neighbors in pore coordinate space). Each candidate throat is initialized with sub-voxel hydraulic diameter and a minimum physically consistent length, enabling a controllable “sub-resolution” edge set without altering the pore set. We then formulate a constrained optimization problem where candidate throats are softly activated and their effective diameters adjusted under geometric feasibility limits (e.g., capped by adjacent pore sizes). The objective minimizes the discrepancy between OpenPNM-simulated absolute permeability and the laboratory absolute permeability (Kabs) of the same rock sample, using a log-space misfit and optional sparsity regularization to avoid over-connecting the network.
Optimization is performed with evolutionary optimizers, which are well suited to non-differentiable objectives involving full pore-network. OpenPNM provides the permeability evaluation (single-phase flow) and serves as the basis for subsequent relative-permeability (Krel) analysis. We use paired high- and low-resolution real micro-CT images: the high-resolution scan supports a reference network for geometric/feature comparisons, while the laboratory Kabs provides the calibration target that the optimized low-resolution graph must reproduce under OpenPNM.
Beyond matching a scalar Kabs, we quantify structural fidelity by comparing empirical CCDFs of throat properties (diameter, length, and conductance proxies) between (i) the original low-resolution network, (ii) the optimized low-resolution network, and (iii) the high-resolution reference network. To test whether calibration against laboratory Kabs also improves multiphase predictions, we compare drainage-based Krel curves across these networks and evaluate whether aligning Kabs reduces the discrepancy in Krel trends.
Finally, we investigate a learning-based path to scalability: given the calibrated low-resolution graph, we evaluate graph neural networks (GNNs) as surrogates for predicting permeability-related properties from graph topology and geometric features. The central hypothesis is that calibrating low-resolution graphs to match laboratory Kabs reduces the resolution-induced domain gap, improving downstream GNN generalization and enabling faster screening across rock ensembles.
| References | I'ANSON, Jack M., et al. Using Graph Neural Networks to Predict the Permeability of Porous Media. InterPore Journal, 2025, 2.3: IPJ250825-2.; Gostick, J. T., et al. OpenPNM: A Pore Network Modeling Package. Computing in Science & Engineering (2016); Bennet, P., Doerr, C., Moreau, A., Rapin, J., F. Teytaud, & O. Teytaud. Nevergrad: Black-box optimization platform. SIGEVOlution 14(1), 8–15 (2021). ; Kipf, T. N., & Welling, M. Semi-Supervised Classification with Graph Convolutional Networks (GCN; foundational). (2016). ; Gai, M., et al. Pore-GNN: Permeability prediction of pore network based on graph neural networks. Advances in Geo-Energy Research (2023). |
|---|---|
| Country | Brazil |
| Acceptance of the Terms & Conditions | Click here to agree |
