Machine learning and UNet++ based microfracture evaluation from CT images

The analysis of densely laminated shale comprising microfractures, pores, matrix, and kerogen at the pore scale is challenging due to the demanding resolution requirements. In this study, stacked micro-Computed Tomography (CT) images were analyzed using machine learning and UNet++ deep learning to investigate porosity and permeability features. The results indicated that random forest machine learning was more effective at identifying weakly connected fractures with low-intensive gray color, and comparably efficient in detecting microfracture boundaries. In addition, UNet++ is shown limited ability to spot small objects and pixels with low color strength compared to machine learning. 3D digital rock analysis revealed that only ∼1%–∼2% of pores were connected, with porosity primarily residing in the disconnected state. This finding contributes to microfracture permeability and plays a crucial role in the initial production of shale reservoirs. Flow-through experiment simulations showed that permeability differed significantly horizontally and vertically due to the intense lamination sequence, emphasizing the need to create a complex fracture network for successful hydrocarbon production from such shale reservoirs. The study highlights the potential of machine learning and UNet++ deep learning in overcoming the demanding resolution requirements of shale analysis. The findings underscore the importance of understanding the pore-scale characteristics of shale for efficient hydrocarbon production.