A computational workflow to study particle transport and filtration in porous media: Coupling CFD and deep learning

Abstract

Listen

Translate article

• We employed neural networks for the prediction of 2D porous media properties. • We setup a workflow for the creation of a dataset of CFD simulations on OpenFOAM . • The results of CFD simulations were the training data for the neural networks. • The permeability and the filtration rate were efficiently predicted by the networks. • The neural networks predicted better the results compared to analytical correlations. In this work we developed an open-source work-flow for the construction of data-driven models from a wide Computational Fluid Dynamics (CFD) simulations campaign. We focused on the prediction of the permeability of bidimensional porous media models, and their effectiveness in filtration of a transported colloidal species. CFD simulations are performed with OpenFOAM , where the colloid transport is solved by the advection–diffusion equation. A campaign of two thousands simulations was performed on a HPC cluster, the permeability is calculated from the simulations with Darcy’s law and the filtration (i.e. deposition) rate is evaluated by an appropriate upscaled parameter. Finally a dataset connecting the input features of the simulations with their results is constructed for the training of neural networks, executed on the open-source machine learning platform Tensorflow (integrated with Python library Keras ). The predictive performance of the data-driven model is then compared with the CFD simulations results and with traditional analytical correlations.