Machine Learning and Topological Approaches for Predicting Permeability of Porous Media

Fluid flow through porous media is critical in many industrial and natural processes, such as oil recovery and groundwater filtration. Permeability is a widely-used measure of how easily fluid flows through a porous medium, so developing fast and accurate methods to estimate it is important. In this project we investigate whether machine learning (ML) models informed by geometric, topological, and network-based descriptors can accurately predict permeability.

We generate synthetic 3D porous structures using PuMA software and compute their permeability using flow simulations, from which we extract structural descriptors and two-point correlation functions. We also reduce the 3D datasets to pore-scale network representations using PoreSpy, and we utilize computational tools from topological data analysis to calculate topological measures of the porous structures. The combined set of descriptors is used to train ML models for permeability prediction.

Our results show that ML models trained on this combination of features are able to predict permeability with high accuracy. In particular, models that include topological information consistently perform better than those using geometric or network data alone. These findings suggest that topological features can serve as effective descriptors for fluid flow in porous media.