Pore-scale study of mineral dissolution in heterogeneous structures and deep learning prediction of permeability

Abstract

Reactive transport processes in porous media with dissolution of solid structures are widely encountered in scientific and engineering problems. In the present work, the reactive transport processes in heterogeneous porous structures generated by Monte Carlo stochastic movement are simulated by using the lattice Boltzmann method. Six dissolution patterns are identified under different Peclet and Damkohler numbers, including uniform pattern, hybrid pattern, compact pattern, conical pattern, dominant pattern, and ramified pattern. Particularly, when Peclet and Damkohler numbers are larger than 1, the increase in the heterogeneity rises the chance of preferential channel flow in the porous medium and thus intensifies the wormhole phenomena, leading to higher permeability. The pore-scale results also show that compared with the specific surface area, the permeability is more sensitive to the alteration of the structural heterogeneity, and it is challenging to propose a general formula between permeability and porosity under different reactive transport conditions and structural heterogeneity. Thus, deep neural network is employed to predict the permeability–porosity relationship. The average value of mean absolute percentage error of prediction of 12 additional permeability–porosity curves is 6.89%, indicating the promising potential of using deep learning for predicting the complicated variations of permeability in heterogeneous porous media with dissolution of solid structures.