Estimating Dispersion Coefficient in Flow Through Heterogeneous Porous Media by a Deep Convolutional Neural Network. Issue 18 (25th September 2021)

• Record Type: Journal Article
• Authors: Kamrava, Serveh; Im, Jinwoo; de Barros, Felipe P. J.; Sahimi, Muhammad
• Is Part Of: Geophysical research letters. Volume 48:Issue 18(2021)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2021-09-25
• Subjects: Geophysics -- Periodicals; Planets -- Periodicals; Lunar geology -- Periodicals; 550
• DOI: 10.1029/2021GL094443
• Languages: English
• ISSNs: 0094-8276

Abstract

Abstract: Estimating the longitudinal dispersion coefficient D L in flow through heterogeneous porous media is paramount to many problems in geological formations. Moreover, although it is well‐known that D L is sensitive to the morphology of such formations, it has been very difficult to establish a firm link between the two. We describe a novel deep convolutional neural network (DCNN) for estimating D L . The inputs for training of the network are a large and diverse set of data consisting of three‐dimensional images of porous media, as well as their porosity, and the associated D L values computed by random‐walk particle‐tracking (RWPT) simulations. The trained network predicts D L very rapidly, and its predictions are in excellent agreement with the data not used in the training. Thus, a combination of the DCNN and RWPT simulation provides a powerful tool for studying many flow‐related phenomena in geological formations, and estimating their properties. Plain Language Summary: Measuring or computing the dispersion coefficient DL in flow through porous media, a fundamental characteristic of transport in geological formations and risk analysis, is a time‐consuming endeaver. Moreover, although DL is sensitive to the morphology of a pore space, a direct link between the two has been missing. We proposed a deep convolutional neural network for predicting DL, using 3D images of porous media and their porosities. The accuracy of the predictions for the actual data indicates that the ability of the network for estimating the important flow and transport properties of porous media for new input data. The present work was at the core scale, on the order of the physical sizes of the sandstones used in our study. The same approach may be used at the eld scale. In that case, one generates the input data by following the same procedure, except that the data should be generated for models in which the permeability and porosity of the formation vary spatially, and represent correlated fields. Work in this direction is in progress. Key Points: Dispersion coefficient Convolutional neural network Flow through porous media