Hybrid LBM and machine learning algorithms for permeability prediction of porous media: A comparative study

• Published in: Computers and geotechnics 2024-04, Vol.168, p.106163, Article 106163
• Main Authors: Kang, Qing, Li, Kai-Qi, Fu, Jin-Long, Liu, Yong
• Format: Article
• Language: English
• Subjects: Machine learning; Microstructural characterization; Permeability prediction; Pore-scale modelling; Porous media; Stochastic reconstruction
• ISSN: 0266-352X; 1873-7633
• DOI: 10.1016/j.compgeo.2024.106163

Investigation of porous media’s permeability is vital for underground resource extraction. Due to the disordered internal structures, it remains a challenge to accurately evaluate the permeability of natural porous media. This study introduces a novel hybrid model that combines the lattice Boltzmann method and different machine learning algorithms to predict porous media’s intrinsic permeability. Firstly, a database containing 1000 sets of 3D digital microstructures of random porous media and corresponding permeability values is compiled. Three machine learning models, i.e., convolutional neural network, adaptive boosting and artificial neural network, are employed to build data-driven predictive models. Results show that the goodness of fit (R2) values from the three machine learning models are greater than 0.9, and the AdaBoost possesses the highest prediction accuracy and best generalization capacity. The machine learning model can improve prediction accuracy by 15% compared to empirical formulas. This study proposes a framework for evaluating the permeability of porous media, based on which practitioners can efficiently assess the permeability of porous media by inputting images or data (pore structure information). Furthermore, it provides references for rapid predicting the permeability of porous media, which opens new pathways to the investigations of underground seepage.