Permeability prediction for natural porous rocks through feature selection and machine learning

The relationships between macroscopic properties and microstructural characteristics are of great significance for natural porous rocks, based on which transport properties can be directly predicted from measurable pore microstructures. However, the explicit establishment of such microstructure-property mappings appears to be difficult, due to the intricacy, stochasticity and heterogeneity of pore network systems. In this paper, a data-driven framework is developed to explore the inherent microstructurepermeability linkage, where multiple techniques are integrated together, including stochastic reconstruction, microstructure characterization, pore-scale simulation, feature selection and machine learning. A large number of digital rock samples are generated from micro-CT imaging and the stochastic reconstruction algorithm. The pore microstructures are quantitatively characterized by using various morphological descriptors. High-fidelity lattice Boltzmann simulations are conducted to evaluate the permeability values of these samples. The optimal set of morphological descriptors that best represents permeability is identified through a performance-oriented feature selection process, and then a machine learning-based surrogate model is constructed to implicitly link permeability with microstructural features. This model can efficiently predict permeability with a broad range of values, and it exhibits great superiority over commonly used empirical/analytical relations in terms of predictive accuracy, general applicability and data-driven evolvability. Besides, deep insights into the underlying microstructure-permeability correlation can also be obtained from the feature selection results, thanks to the good interpretability of morphological descriptors.