Abstract Pore network modeling (PNM) with network extracted from 3D images is widely used for studying mass transport in porous media, with its accuracy heavily dependent on the precise determination of hydraulic and diffusive conductance values of local pore pairs. Deep learning (DL) prediction based on true pore geometry offers a promising approach for conductance estimation. However, due to arbitrarily orientation and the lack of explicit inlet and outlet surfaces in 3D extracted pore pair geometries, existing studies [24,25] have relied only on 2D cross section images for hydraulic conductance prediction. A custom finite difference method was recently developed, capable exclusively of estimating diffusion conductance from 3D pore geometry [26]. In this work, a convolutional neural network (CNN) was trained to estimate both diffusive and hydraulic conductance values, with reference ground truth values obtained by applying the Lattice Boltzmann Method (LBM) on 3D pore pair geometries. The predicted pore pair conductance values from the CNN closely matched the LBM results on unseen segmented image data. This study highlights the importance of using the full 3D pore pair geometry for accurate hydraulic conductance evaluation, rather than relying solely on 2D throat cross sections. To evaluate the generality of our method, we calculated the permeability and diffusivity of 3D porous media across a wide range of porosities using PNM with local conductance values predicted by our DL model. Comparisons with results of LBM indicate that while diffusivity was accurately predicted, permeability was underestimated. This underestimation is attributed to the over-segmentation of pore spaces, which resulted in an underestimated flow rate.
Read full abstract