Machine learning modeling of permeability in 3D heterogeneous porous media using a novel stochastic pore-scale simulation approach

• Permeability can be estimated from stochastic pore-scale simulations. • Stochastic permeability estimations are closer to measured permeability. • Machine learning replicates permeability estimates from direct pore-scale simulations. • Machine learning significantly reduces the computational time of pore-scale simulations. Accurate predictions of rock permeability is critical for resource exploration and environmental management. To improve on existing approaches to permeability prediction, this study employed a stochastic pore-scale simulation approach. The petrophysical properties needed for the implementation of this approach are porosity and pore size distribution (PSD) of rock samples which can be obtained easily from mercury injection capillary pressure measurements. The approach was tested on four carbonate and five siliciclastic rock cores. To consider a wide range of possible pore connectivity scenarios that can be associated to the same PSD and porosity, the employed stochastic pore-scale simulation approach involves the generation of hundreds of 3D pore microstructures of the same PSD and porosity but different stochastic pore connectivity. Permeability is calculated by averaging the permeability distribution obtained from pore-scale flow simulations through the generated 3D pore microstructures. Permeability estimations were closer to measured permeability with this approach than with five deterministic empirical model equations. Machine learning was used to reduce the required number of pore-scale simulations by 157 times and reproduced permeability estimated from pore-scale flow simulations with a mean absolute percentage error of 10%.