Deep learning-based pore network generation: Numerical insights into pore geometry effects on microstructural fluid flow behaviors of unconventional resources

Pore-scale transport behaviors and mechanisms of rock reservoirs are still not well understood to increase unconventional resource production. This work mainly focuses on proposing a deep learning-based method to rapidly construct optimal pore network with different pore types, and deeply analyze its effects on pore-scale transport behaviors and mechanisms. The pore-scale variables reservoir evaluation indexes are defined to quantitatively evaluate pore geometry effects on the properties and production of rock reservoirs. The two-phase displacement simulations in pore network are conducted to study microstructural flow behaviors and transport mechanisms. Results suggest that the deep learning-based digital labeling algorithm (DL-DLA) has excellent abilities to rapidly construct pore network with errors less than 5%, compared with the classical algorithms. Square pores and circle throats are suggested as the optimal pore network assembly, considering the fluid phase drainage efficiency and production rate. The microstructural transport mechanisms are concluded as the pore-throat drainage, pore-filling, fluid phase mixing and fluid phase equilibrium processes. The novel theoretical relation between fluid phase drainage and microscopic production indexes provides effective tools to estimate rock reservoir production with errors all less than 10%, which are helpful for the technique developments to increase the production of unconventional resources in rock reservoirs.