Stylolites are complex geological features that usually appear as irregular surfaces. Forming pressure-driven dissolution, they are commonly filled with insoluble materials compared to surrounding host rocks. In underground reservoir flow (e.g., hydrogen, carbon dioxide, or natural gas storage) stylolites can influence reservoir fluid flow. They can act either as flow barriers or as flow pathways in reservoirs. Therefore, identifying stylolites in porous rocks is an essential task for the correct prediction of underground fluid flow. In this study, we develop a hybrid deep-learning workflow to automatically detect and classify stylolites in slabbed whole-core images. First, the “You Only Look Once” object-detection architecture is used to locate corer rock boundaries in raw whole core images containing multiple cores. These core images are then subdivided into 1,827 smaller image patches. We employ two convolutional neural network architectures of ReNet-50 and ResNeXt-50 for the final classification task. As the dataset, 150 m cores from three wells of a carbonate reservoir are used. The core image patches are first classified into five different image classes of stylolites, fractures, vertical plugs, horizontal plugs, and intact rock, manually. As the dataset is imbalanced with respect to the number of images in each class, data-augmentation techniques such as flipping, cropping, rotation, and adjustments to brightness and contrast are implemented. This expands the dataset images to almost 16,000 images in total. Both networks are pre-trained on the ImageNet dataset and fine-tuned on the augmented dataset. To control overfitting, additional regularization techniques like dropout and adaptive learning-rate scheduling are used. Results show that ResNeXt-50 achieves the best classification performance of 92% on previously unseen whole-core images. • Stylolite detection is crucial for fluid flow model predictions. • YOLO architecture was used for automatic preprocessing stage to identify core images. • Deep convolutional neural networks are used for classification of whole core images. • Fine-tuned ResNeXt-50 with transfer learning provided most accurate classification results. • Dropout and scheduled learning rate were effective to mitigate overfitting.
Sadeghnejad et al. (Sun,) studied this question.