Prediction of diffusional conductance in extracted pore network models using convolutional neural networks. (May 2022)

• Record Type: Journal Article
• Authors: Misaghian, Niloo; Agnaou, Mehrez; Sadeghi, Mohammad Amin; Fathiannasab, Hamed; Hadji, Isma; Roberts, Edward; Gostick, Jeff
• Is Part Of: Computers & geosciences. Volume 162(2022)
• Publication Date: 2022-05
• Subjects: Pore network modeling -- Diffusive conductance -- Deep learning model -- Convolutional neural network; Environmental policy -- Periodicals; 550.5
• DOI: 10.1016/j.cageo.2022.105086
• Languages: English
• ISSNs: 0098-3004

Abstract

Abstract: Pore network modeling (PNM) based on networks extracted from tomograms is a well-established tool for simulating pore-scale transport behavior in porous media. A key element of this approach is the accurate determination of pore-to-pore conductance values, which is a complex task that greatly affects the accuracy of flow and diffusive mass transport studies. Classic methods of conductance estimation based on analytical solutions and shape factors only apply to simple pore geometries, whereas real porous media contain irregular-shaped pores. Although direct numerical simulations (DNS) can accurately estimate conductance considering pores' real morphology, it has a high computational cost that becomes infeasible for large tomograms. The present work remedies this problem using a deep learning (DL) approach, with a specific focus on diffusional transport which has received less attention than hydraulic conductance. A convolutional neural network (CNN) model was trained to estimate diffusive conductance of PNM elements from volumetric images of porous media. The developed framework estimates the diffusive conductance by analyzing individual pore-to-pore 3D images isolated from the tomogram to fully capture the topology and shapes. A key outcome of the present work is that only images of the pore regions are used as input data, avoiding excessive preprocessing time for data preparation. The results of the diffusive conductance prediction show good agreement with the test data obtained by DNS method, with 0.94 R 2 prediction accuracy and a speedup of 500x in prediction runtime. Highlights: A deep learning model was developed to estimate diffusive conductance of pore network elements. Direct numerical simulations of diffusive transport was used as ground truth for training. The deep learning model uses only images of pore to pore regions as input. The trained model can predict conductances in seconds within a high accuracy. Pore networks using the deep learning-based conductance predict formation factor twice as accurately.