Deep learning for mass transport in porous media

We discuss the convolutional neural networks(CNNs) to predict the basic properties of transportthrough porous media. Two types of transport are considered- fluid flow and diffusion. We use the Lattice-Boltzmann method (LBM) to get numerical data for networktraining; namely, we obtain the fluid flow velocityfields and gas concentration maps at the pore-scale. Westudy how CNNs are effective in predicting macroscopicparameters, such as permeability, porosity, and tortuositybased only on information about the geometry of thesamples. Eventually, we adapt U-Net architecture tostudy the capability of CNNs to predict complex spatialconcentration maps in diffusion phenomena.