A new insight into digital rock analysis using deep learning: a comparative study on the effect of hyperparameter selection and data augmentation

Convolutional neural networks (CNNs) have proven to be powerful tools in deriving rock properties from micro-computed tomography (micro-CT) images, providing a data-driven alternative to conventional digital rock physics workflows. However, the reliability performance and computational efficiency of CNN-based predictions remain strongly dependent on architectural and training choices that are often determined empirically. This study presents a systematic sensitivity-based evaluation of 3D CNN design and training factors, examining their effects on predictive performance and computational cost across multiple rock properties and dataset sizes. A 3D CNN-based workflow was evaluated for the prediction of five pore-scale properties, including porosity, average pore size, throat size, coordination number, and pore shape factor, from micro-CT images of a carbonate rock. The effects of kernel size progression, network width, number of convolutional layers (network depth), activation function, and learning rate, were quantitatively assessed in terms of predictive accuracy and computational cost. The role of data augmentation was additionally investigated across different dataset sizes. The results showed that (1) despite the high effectiveness of network depth on model accuracy, it has no significant impact on training runtime; (2) Odd-sized kernels with progressive spatial reduction generally outperformed even-sized and aggressively reduced kernel configurations, particularly for complex morphological descriptors such as pore shape factor; (3) data augmentation imporved performance for smaller datasets but provided limited or negative gains for larger datasets; and (4) Computational runtime was influenced primarily by kernel size and network width rather than network depth, learning rate schedules, and activation functions. A comparative sensitivity assessment further revealed that pore shape factor was the most hyperparameter-sensitive property, while porosity showed the highest robustness across different CNN configurations and also network width exhibited the smallest overall influence on predictive performance. Grad-CAM-based analysis provided additional insight into the image regions associated with CNN predictions. Porosity predictions showed the strongest association with pore-space regions, whereas pore radius and throat radius exhibited positive relationships with local pore-space geometry. Overall, the results provide a systematic framework for evaluating CNN hyperparameters, computational cost, and data augmentation in micro-CT-based digital rock analysis.