Publications

Coordination number, a key measure of pore network connectivity, is crucial for understanding fluid flow and transport properties in reservoir rocks. Traditional methods of estimating coordination numbers are often labor-intensive, relying on experimental data or manual image analysis. This study explores a machine learning approach using a custom-developed Convolutional Neural Network (CNN) and transfer learning to enhance the accuracy and efficiency of coordination number estimation.

A dataset of three-dimensional (3D) digital rock images from sandstone and carbonate formations was obtained from the Digital Rocks Portal. Image processing techniques, including watershed segmentation, were employed to extract pore network features. Various pre-trained CNN models, such as DenseNet201, MobileNetV2, and Xception, were fine-tuned to predict coordination numbers based on these extracted features. Model performance was evaluated using Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and the coefficient of determination (R²).

Results indicate that DenseNet201 outperforms other models, achieving the lowest MAPE (7.27% for sandstone and 13.42% for carbonate) and the highest R² values (0.75 for sandstone and 0.70 for carbonate). Compared to conventional image processing methods, the machine learning-based approach significantly reduces computation time, from 3,092 seconds to just 4 seconds for coordination number estimation. These findings demonstrate the potential of deep learning in reservoir rock characterization, offering a scalable and non-invasive solution for improving hydrocarbon recovery strategies.

For IPA members, this research highlights the potential of combining advanced machine learning techniques with digital rock analysis to improve reservoir understanding and management. Future work will expand the methodology to include other dynamic properties such as permeability and capillary pressure, further enhancing its applications in reservoir optimization.