Evaluation of mechanical properties of porous media materials based on deep learning: Insights from pore structure

Rock is a complex, porous medium, particularly evident in fine-grained sedimentary rocks such as shale, coal, and sandstone. These rocks are rich in oil and gas resources, and their developmental effectiveness hinges on mechanical properties. The mechanical properties of these sedimentary rocks vary widely, primarily due to their intricate pore structures. This study focuses on porous materials, taking shale as an example, and utilizes scanning electron microscopy to obtain numerous images depicting pore structures. It marks the inaugural application of the HRNet model for identifying pores in these images, achieving accurate and swift identification. This study also benchmarks the recognition accuracy against the U-Net model. Utilizing deep learning for pore recognition, Image J extracts the shape factor and Feret diameter to characterize the pore structure. Subsequently, RFPA software constructs microscopic models with varying pore structure parameters, indicating the pore structure's impact on mechanical properties. The findings revealed that the HRNet significantly outperforms U-Net in recognition accuracy, with a mean intersection over the union of 0.87, a pixel accuracy of 0.99, and an F1-score of 0.83. Employing the deconvolution method, the maximum Feret's diameter of pores primarily concentrates at 19.7, 41.2, and 168.2 nm, with shape factors predominantly at 0.89 and 0.67. The simulation results obtained from the uniaxial compression experiments showed that the content, size, and shape of pores significantly affect the mechanical properties. The research methodologies and outcomes of this study hold considerable potential for advancing the understanding of the pore structure and mechanical properties of porous media materials.