Modeling of multi-mineral-component digital core based on Res-Unet

As the exploration of oil and gas moves further into less conventional reservoirs, effective methods are required for the fine evaluation of complex formations, particularly digital core models with multiple mineral components. The current technology cannot directly produce digital core images with multiple minerals. Therefore, image segmentation has been widely used to create digital multi-mineral core images from computed tomography (CT) images. The commonly used image segmentation methods do not provide satisfactory CT images of complex rock formations. Consequently, deep learning algorithms have been successfully applied for image segmentation. In this paper, a novel method is proposed to develop an accurate digital core model with multiple mineral components based on the Res-Unet neural network. CT images of glutenite and the corresponding results of quantitative evaluation of minerals by scanning electron microscopy (QEMSCAN) are used as a training dataset for the automatic segmentation of CT core images. The used quantitative metrics show that compared with the multi-threshold and U-Net segmentation methods, the Res-Unet network leads to better results of mineral morphology and distribution recognition. Finally, it is demonstrated that the proposed Res-Unet-based segmentation model is an effective tool for creating three-dimensional digital core models with multiple mineral components.