Data-driven predictive model of coal permeability based on microscopic fracture structure characterization
Tianhao Yan, Xiaomeng Xu, Xiaomeng Xu, Jiafeng Liu, Yihuai Zhang, M. Asif Arif, Xiaowei Xu, Xiaowei Xu, Qiang Wang
Abstract
Accurate prediction of coal reservoir permeability is crucial for engineering applications, including coal mining, coalbed methane (CBM) extraction, and carbon storage in deep unmineable coal seams. Owing to the inherent heterogeneity and complex internal structure of coal, a well-established method for predicting permeability based on microscopic fracture structures remains elusive. This paper presents a novel integrated approach that leverages the intrinsic relationship between microscopic fracture structure and permeability to construct a predictive model for coal permeability. The proposed framework encompasses data generation through the integration of three-dimensional (3D) digital core analysis and numerical simulations, followed by data-driven modeling via machine learning (ML) techniques. Key data-driven strategies, including feature selection and hyperparameter tuning, are employed to improve model performance. We propose and evaluate twelve data-driven models, including multilayer perceptron (MLP), random forest (RF), and hybrid methods. The results demonstrate that the ML model based on the RF algorithm achieves the highest accuracy and best generalization capability in predicting permeability. This method enables rapid estimation of coal permeability by inputting two-dimensional (2D) computed tomography images or parameters of the microscopic fracture structure, thereby providing an accurate and efficient means of permeability prediction.