Super resolution-assisted pore flow field prediction using neural networks

Direct pore-scale simulations of fluid flow through porous media are computationally expensive to perform for realistic systems. Previous researches have shown that geometry of the microstructure of porous media can be used to predict the velocity fields therein using neural networks. Such trained neural networks, however, perform poorly for unseen porous media, particularly those with a large degree of heterogeneity. In this work, we propose that incorporating a coarse velocity field into the input of neural networks is an effective method for enhancing the prediction accuracy. The velocity field is simulated on a coarsened mesh with a low computational cost. More importantly, it contains global physics information that it can remedy the ill-conditioning produced by the usage of segmented porous media due to the limitation of GPU memory. Numerical results show that incorporating the coarse velocity field greatly improves the prediction accuracy of the fine velocity field when compared to the prediction based on geometric information alone, especially for the porous media with a large interior vuggy pore space. The feasibility of the method is further demonstrated by testing the trained network on real rocks with non-spherical solid grains and much more complex microstructure.