Investigating low-permeability sandstone based on physical experiments and predictive modeling

Abstract

An innovative method is proposed for preparing low-permeability sandstone with different moisture saturation. The permeability of the prepared low-permeability sandstone sample is measured under different confinement and seepage pressures. Based on the experimental results, 10 types of different machine-learning models combined with optimization algorithms are established to predict the permeability of low-permeability sandstone. A comprehensive evaluation and comparison of the 10 types of machine-learning models are conducted to identify the machine-learning model with the best performance. Next, a sensitivity analysis is conducted on the factors influencing the permeability of low-permeability sandstone to elucidate the internal mechanism according to the established machine-learning model. The following conclusions are drawn. With an increase in the confinement pressure, the permeability of low-permeability sandstone with different moisture-saturation levels decreases, and the permeability of low-permeability sandstone decreases with an increase in the moisture saturation. The hybrid particle swarm optimization algorithm-backpropagation artificial neural network (PSO-BPANN) model provides the best results for predicting the permeability of low-permeability sandstone. The established PSO-BPANN model is also reliable for predicting the permeability of low-permeability sandstone from other engineering sites. Among the influencing factors, moisture saturation has the largest effect on the permeability of low-permeability sandstone, followed by the confinement pressure.