The control effect of the pore and fracture characteristics of tectonic coal assisted by the triaxial stress permeability system on permeability

This study systematically investigates the pore and fracture characteristics of tectonic coal and their controlling effects on permeability, which are critical for advancing coalbed methane (CBM) development and geological engineering applications. To achieve this, a triaxial stress permeability system and deep learning (DL) techniques are employed to conduct experiments on tectonic coal samples, enabling the measurement of permeability variations under different stress conditions. Furthermore, high-resolution fracture images are collected, and DL-based methods are utilized for fracture identification and feature extraction. Based on these findings, a permeability prediction model is developed to assess the applicability of DL in predicting the permeability of tectonic coal. The experimental results indicate that, under constant pore pressure, apparent permeability decreases with increasing effective stress. Specifically, when the deviatoric stress is maintained at 50 MPa and the pore pressure at 1 MPa, the apparent permeability declines from 14.846 × 10⁻¹⁸ m² to 8.326 × 10⁻¹⁸ m² as the effective stress increases from 29 MPa to 44 MPa. Moreover, the proposed ShuffleNet V2-Bidirectional Gated Recurrent Unit-Graph Attention Mechanism model demonstrates superior performance compared to existing models, achieving an accuracy of 95.62%, a root mean square error of 4.19, and a mean absolute error of 8.13. These results underscore the model’s effectiveness in accurately predicting the permeability characteristics of tectonic coal, thereby providing valuable technical support for CBM extraction and geological engineering applications. The novelty of this study lies in the integration of the triaxial stress permeability system with DL techniques, which enhances the accuracy of permeability predictions. Additionally, the study offers a novel perspective on fracture characterization in tectonic coal by leveraging the advanced recognition capabilities of DL models, surpassing previous methodologies in the literature.