From Computational Fluid Dynamics to Structure Interpretation via Neural Networks: An Application to Flow and Transport in Porous Media

Abstract

The modeling of flow and transport in porous media is of the utmost importance in many chemical engineering applications, including catalytic reactors, batteries, and CO2storage. The aim of this study is to test the use of fully connected (FCNN) and convolutional neural networks (CNN) for the prediction of crucial properties in porous media systems: the permeability and the filtration rate. The data-driven models are trained on a dataset of computational fluid dynamics (CFD) simulations. To this end, the porous media geometries are created in silico by a discrete element method, and a rigorous setup of the CFD simulations is presented. The models trained have as input both geometrical and operating conditions features so that they could find application in multiscale modeling, optimization problems, and in-line control. The average error on the prediction of the permeability is lower than 2.5%, and that on the prediction of the filtration rate is lower than 5% in all the neural networks models. These results are achieved with at least a dataset of ∼100 CFD simulations.