Upscaling reactive transport models from pore-scale to continuum-scale using deep learning method

Reactive transport modeling of subsurface environments plays an important role in addressing critical problems of geochemical processes, such as dissolution and precipitation of minerals. Current transport models for porous media span various scales, ranging from pore-scale to continuum-scale. In this study, we established an upscaling method connecting pore-scale and continuum-scale models by employing a deep learning methodology of Convolutional Neural Networks (CNNs). We applied Darcy-Brinkmann-Stokes (DBS) method to simulate the fluid flow and reactive transport in pore-scale models, which would act as constituents of a continuum-scale model. The datasets of spatial pore distribution of subvolume samples were used as the input for the deep learning model, and the continuum (Darcy)-scale parameters such as permeability, effective surface area, and effective diffusion coefficient were figured out as outputs (i.e., labels). By applying the trained models of the subvolumes in the entire sample volume, we generated the initial field of porosity, permeability, effective diffusion coefficient, and effective surface area for continuum-scale simulation of a mineral dissolution problem. We took an acid dissolution case as an example to utilize the outcomes of trained deep learning models as input data in the continuum-scale simulation. This work presents a comprehensive upscaling workflow, as bridging the findings of microscale simulations to the continuum-scale simulations of a reactive transport problem.