In a sensitivity-based evaluation of three-dimensional convolutional neural networks for micro-computed tomography digital rock physics, network depth significantly enhanced prediction accuracy for pore-scale rock properties while showing no significant impact on overall training runtime. Researchers evaluated this 3D deep learning workflow using micro-CT imagery of a carbonate rock to examine how core architectural choices and training configurations govern both model accuracy and computational cost.
The study quantitatively assessed model performance across five distinct pore-scale descriptors: porosity, average pore size, throat size, coordination number, and pore shape factor. To determine how model design influences outcomes, the investigation tested the progression of kernel sizes, network width, the number of convolutional layers, choice of activation function, and learning rate schedules across varying dataset sizes.
Kernel configuration proved critical for complex structural predictions. The results demonstrated that odd-sized convolutional kernels utilizing progressive spatial reduction consistently outperformed configurations based on even-sized kernels or aggressive spatial reduction strategies. This architectural advantage was particularly pronounced when predicting complex morphological descriptors such as the rock’s pore shape factor.
The impact of data augmentation varied substantially depending on the volume of training data available. While applying data augmentation techniques improved predictive performance when training on smaller datasets, the practice yielded limited or negative performance gains when applied to larger dataset sizes, highlighting clear limits to standard augmentation strategies in digital rock characterization.
Analyses of computational efficiency revealed that training runtime was primarily driven by kernel size progression and network width, whereas network depth, learning rate schedules, and activation functions exerted negligible influence on execution time. Furthermore, a comparative sensitivity assessment established that pore shape factor was the most hyperparameter-sensitive property, whereas porosity demonstrated the highest robustness across different network configurations, and network width showed the smallest overall effect on predictive performance.
Interpretability assessments using Grad-CAM highlighted the specific micro-CT image regions governing neural network predictions. The analysis showed that porosity predictions exhibited the strongest association with segmented pore-space regions, while pore radius and throat radius predictions demonstrated positive relationships with local pore-space geometry, establishing a systematic framework for evaluating deep learning designs in digital rock analysis.
In conventional digital rock physics workflows, determining rock properties from micro-computed tomography images has increasingly relied on convolutional neural networks as a data-driven alternative. However, the reliability, predictive performance, and computational efficiency of these deep learning models frequently depend on architectural structures and training parameters that practitioners select empirically. To resolve these empirical uncertainties, researchers Sara Shirafkan, Mohammad Ahmadi, and Reza Yousefzadeh conducted a systematic sensitivity-based evaluation examining how core 3D CNN design choices and training factors influence both predictive accuracy and computational cost across multiple rock properties and dataset sizes.
The experimental workflow applied 3D convolutional neural networks directly to micro-CT imagery of a carbonate rock to evaluate five specific pore-scale properties: porosity, average pore size, throat size, coordination number, and pore shape factor. The investigation quantitatively tested the influence of kernel size progression, network width, the number of convolutional layers representing network depth, activation functions, and learning rates. In parallel, the researchers analyzed data augmentation across different training dataset sizes, assessing whether synthetic sample expansion provided consistent benefits across varying training volumes.
The comparative sensitivity assessment highlighted distinct variations in how structural descriptors respond to model architecture. Complex morphological descriptors proved especially demanding, with pore shape factor emerging as the most hyperparameter-sensitive property tested. Odd-sized kernels combined with progressive spatial reduction delivered superior accuracy over even-sized or aggressively reduced designs for these intricate features. Conversely, porosity demonstrated the highest robustness across configurations, while network width exhibited the smallest overall impact on predictive performance despite driving computational runtime alongside kernel dimensions.
Grad-CAM visualizations offered critical interpretability by pinpointing image regions linked to neural predictions. The analysis confirmed that porosity estimates correlated most strongly with segmented pore-space regions, whereas pore radius and throat radius aligned positively with local pore geometry, establishing a comprehensive evaluation framework for micro-CT digital rock analysis.
Subject of Research: Sensitivity evaluation of 3D convolutional neural network hyperparameters in micro-CT digital rock analysis
Article Title: A new insight into digital rock analysis using deep learning: a comparative study on the effect of hyperparameter selection and data augmentation
Article References: Shirafkan, S., Ahmadi, M., & Yousefzadeh, R. (2026). A new insight into digital rock analysis using deep learning: a comparative study on the effect of hyperparameter selection and data augmentation. Journal of Petroleum Exploration and Production Technology. https://doi.org/10.1007/s13202-026-02240-7
Image Credits: AI Generated
DOI: 10.1007/s13202-026-02240-7
Keywords: convolutional neural networks, digital rock analysis, micro-CT imaging, carbonate rock, porosity prediction, pore shape factor, throat size, coordination number, average pore size, hyperparameter selection, data augmentation, kernel size progression, network depth, network width, computational runtime, Grad-CAM analysis, pore-scale properties, digital rock physics
Cite Scienmag News
Cassandra Pierce. (November 19, 2026). Network Depth Improves 3D CNN Digital Rock Accuracy Without Increasing Training Runtime. Scienmag. https://scienmag.com/network-depth-improves-3d-cnn-digital-rock-accuracy-without-increasing-training-runtime/
Cassandra Pierce. "Network Depth Improves 3D CNN Digital Rock Accuracy Without Increasing Training Runtime." Scienmag, 19 November 2026, https://scienmag.com/network-depth-improves-3d-cnn-digital-rock-accuracy-without-increasing-training-runtime/. Accessed 19 November 2026.
Cassandra Pierce. "Network Depth Improves 3D CNN Digital Rock Accuracy Without Increasing Training Runtime." Scienmag. November 19, 2026. https://scienmag.com/network-depth-improves-3d-cnn-digital-rock-accuracy-without-increasing-training-runtime/








