Deep learning in the advanced core sample porosity determination with XCT image

X-ray computed tomography (XCT) for core sample porosity analysis has advanced significantly, integrating machine learning with traditional image processing and introducing methods such as Weighted Assignment by CT number calculation. However, several areas still require enhancement, particularly in image preprocessing and model optimization for specific deep learning approaches. Moreover, the application of the Weighted Assignment method to water-sensitive mudstone core samples in this study revealed the inconsistency with the porosity estimation, primarily due to the overemphasis on fracture effects. To address these limitations, this study proposed the ResNet algorithm, a deep learning model developed for XCT-based porosity estimation. ResNet's architecture, incorporating residual learning with shortcut connections, enables effective training of deep networks and enhances feature extraction. The ResNet-based model was trained on XCT images with known porosity values measured by Multi-Sensor Core Logging (MSCL), effectively predicting volume fraction porosity for materials with similar lithological characteristics. Analysis of testing losses, comparison of predicted versus actual porosity values and RRMSE, demonstrates the model's superior performance than traditional methods. This novel application of ResNet to porosity estimation highlights the potential of deep learning to significantly enhance accuracy by 97.82% compared with the weighted assignment method with RRMSE.