Search

A gradient‐information‐enhanced image segmentation method using convolutional neural networks is presented, and then combined with contact angle measurement to establish an automated processing workflow. For three‐dimensional X‐ray images, the segmentation accuracy at interfaces and sparsely distributed small objects directly influences the accuracy of the contact angle measurement. Leveraging reliable gradient information to train the neural network, this segmentation method addresses the issue of inaccurate segmentation of interfaces even at low resolution and with small objects present. Furthermore, memory requirements are reduced by performing analysis on orthogonal two‐dimensional planes. The workflow was tested on water‐wet Ketton limestone, as well as on both water‐wet and mixed‐wet sandstone and a reservoir carbonate. The results from both the segmentation and contact angle measurements underscore the effectiveness of the approach. Notably, the workflow shows considerable generalizability and robustness, even with varying wettability and lithology. Plain Language Summary: We use convolutional neural networks enhanced with gradient information to improve the accuracy of segmentation of three‐dimensional images of porous media and the fluids in the pore space, especially at interfaces and for small objects, even when the images have low resolution. We tested the method on several data sets including water‐wet Ketton limestone, both water‐wet and mixed‐wet Bentheimer sandstone, and a reservoir carbonate. The results demonstrated the effectiveness of our approach compared to traditional methods for segmentation. A particular challenge is the determination of contact angle ‐ the local wettability ‐ at the three‐phase contact loop between two fluid phases and the solid. We show that our new method leads to reliable estimates of the distribution of contact angle for the samples studied. This work allows for more accurate image analysis in multiphase flow in porous media with applications in carbon dioxide and hydrogen storage. Key Points: A gradient information enhanced segmentation method is presented to assist automatic contact angle measurementsThe gradient information method reduces the problems associated with low image resolution and manual segmentation [ABSTRACT FROM AUTHOR]

Porous materials, such as carbonate rocks, frequently have pore sizes which span many orders of magnitude. This is a challenge for models that rely on an image of the pore space, since much of the pore space may be unresolved. In this work, sub‐resolution porosity in X‐ray images is characterized using differential imaging which quantifies the difference between a dry scan and 30 wt% potassium iodide brine saturated images. Once characterized, we develop a robust workflow to incorporate the sub‐resolution pore space into a network model using Darcy‐type elements called microlinks. Each grain voxel with sub‐resolution porosity is assigned to the two nearest resolved pores using an automatic dilation algorithm. By including these microlinks with empirical models in flow modeling, we simulate single‐phase and multiphase flow. By fine‐tuning the microlink empirical models, we match permeability, formation factor (the ratio of the resistivity of a rock filled with brine to the resistivity of that brine), and drainage capillary pressure to experimental results. We then show that our model can successfully predict steady‐state relative permeability measurements on a water‐wet Estaillades carbonate sample within the uncertainty of the experiments and modeling. Our approach of incorporating sub‐resolution porosity in two‐phase flow modeling using image‐based multiscale pore network techniques can capture complex pore structures and accurately predict flow behavior in porous materials with a wide range of pore size. Key Points: A dilation‐based algorithm adds sub‐resolution porosity, quantified by differential imaging, as microlinks to an extracted pore networkEmpirical models are used to characterize the flow properties of microlinksThe tuned multiscale network, using permeability, formation factor, and capillary pressure, matches experimental relative permeability [ABSTRACT FROM AUTHOR]