Hydraulic Tortuosity is a flow property that describes the degree of twist of pore microstructures in rocks. It is defined as the ratio of true distance travelled by fluid in porous media to length of straight-line path through the rock. Despite being a very simple concept, this property is very difficult to directly estimate in consolidated porous media. Conventionally, tortuosity is estimated via statistical correlations between measured tortuosity and porosity values, or via empirical equations such as the Kozeny-Carman. While these methods are effective for predicting tortuosity of homogeneous and/or unconsolidated porous media, there is no evidence that these methods provide reliable estimates of tortuosity for highly heterogeneous porous media. Such estimates of tortuosity might mislead reactive transport simulations as flow path through porous media influences the retention time, and thus the rates of reactions in the gas, aqueous, and mineral phases occupying the pore space of rocks. To improve the accuracy of reactive transport simulations in heterogeneous porous media, this study employs five machine learning algorithms (Linear Regression (LR), Neural Networks, Random Forest, Support Vector Machine, and Nearest Neighbors) to predict reliable estimates of hydraulic tortuosity for highly heterogeneous porous media. The data employed in this study is obtained from synthetic 3D pore microstructures to ensure that all pore microstructure factors affecting the tortuosity of heterogeneous porous media is measured with perfect accuracy. The 3D pore microstructures were created using four synthetic pore size distribution curves with pore diameter ranging from 6 μm to 30 μm. Spherically shaped pores were assumed for simplicity while the topology of the pores was randomized per sample to produce heterogeneity. The tortuosity for each sample (constructed 3D pore microstructure) was obtained using STAR-CCM+®, a computational fluid dynamics software. Overall, we had 19 features and one target (tortuosity) for each of the 3,275 stochastically generated samples. Our results showed LR to be the best algorithm for our data. Additionally, LR helped capture and rank the most crucial pore microstructural properties in predicting tortuosity. Finally, our results show an improved accuracy in tortuosity prediction compared to existing approaches we attempted and could potentially introduce a simpler way of estimating tortuosity.

Application of machine learning to predict hydraulic tortuosity of highly heterogeneous porous media

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.