Surrogate permeability modelling of low-permeable rocks using convolutional neural networks

By Jianwei Tian, Chongchong Qi, Yingfeng Sun, Zaher Mundher Yaseen. 2020

50 citations recorded by the source corpus.

Abstract Permeability is an important mechanical property of low-permeable rocks, the prediction of which plays a significant role in forecasting many engineering problems like gas production and CO2 geo-sequestration. However, the ultra-low permeability and strong heterogeneity of low-permeable rocks hinder the accurate determination of permeability experimentally and numerically. In this study, 3D convolutional neural networks (CNNs) were trained to rapidly evaluate the permeability of low-permeable porous media. An improved quartet structure generation set considering pore size distribution (PSD) acquired by experiment and anisotropy (QSGS) was proposed to generate low-permeable porous media. Together with the permeability calculation using D3Q19 model of Lattice Boltzmann Method (LBM), the whole dataset with a size of 4,500 was prepared for CNN modelling. Moreover, a parametric study was conducted to investigate the influence of CNN architecture and training epochs on its performance. The performance discrepancy corresponding to the magnitude of absolute permeability was also discussed. The modelling results indicate that the CNNs architecture has an important effect on its performance. The best architecture achieved an average validating loss of 1.20 × 10−3 while the worst architecture achieved an average validating loss of 3.60 × 10−3. The CNN modelling suffered a little from over-fitting when the epoch was over 20. The optimum 3D CNN achieved remarkably high accuracy of permeability prediction with high correlation coefficient (R=0.996) and low errors (MAE=0.017, RMSE=0.03) on the testing set. Furthermore, it was found that the testing error of the lower permeability groups was relatively larger than that of larger permeability groups.

Subfields: Computational Mechanics, Mechanics of Materials, Ocean Engineering

View Paper: https://www.semanticscholar.org/paper/da00915862429f55f3e4151524855a7aef62942f