Fast prediction method of displacement front in heterogeneous porous media using deep learning and orthogonal design

Heterogeneity plays a vital role in the micro-flow through porous media, but experimentally observing the flow details is time-consuming. This study provides a fast method for displacement front prediction in various heterogeneous porous media based on deep learning and orthogonal design. It uses the orthogonal design to generate displacement schemes with different permeability contrasts, capillary numbers, and viscosity ratios and utilizes the lattice Boltzmann simulation to obtain the datasets of displacement front at breakthrough. The prediction network is then established based on the U-Net structure. Finally, the displacement fronts of porous media with various heterogeneities are predicted. Compared to training with a dataset generated by random sampling, this method can halve the time required to establish the dataset and network training without compromising accuracy. Three orders of magnitude reduce the time necessary for network prediction compared to the lattice Boltzmann simulation. The results indicated that the total water saturation decreases as the permeability contrast increases and the water saturation and front position ratios rise. As the permeability contrast grows, the influence of the capillary number and viscosity ratio on the water saturation and front position ratios becomes more pronounced. And the influence of the viscosity ratio on total water saturation is more significant in low permeability contrast porous media. This research is helpful for the study of microscopic channeling and remaining oil distribution and further guides reservoir development.

Automated summary

This study presents a fast method for predicting displacement fronts in different heterogeneous porous media using deep learning and orthogonal design. By generating displacement schemes with different permeability contrasts, capillary numbers, and viscosity ratios through orthogonal design, and obtaining datasets of displacement fronts through lattice Boltzmann simulation, a prediction network is established based on the U-Net structure. The results show that this method significantly reduces the time required for dataset establishment and network training, and greatly reduces the time needed for network prediction compared to lattice Boltzmann simulation.