Prediction of carbonate permeability from multi-resolution CT scans and deep learning

The low-resolution CT scan images obtained from drill core generally struggle with problems such as insufficient pore structure information and incomplete image details. Consequently, predicting the permeability of heterogeneous reservoir cores relies heavily on high-resolution CT scanning images. However, this approach requires a considerable amount of data and is associated with high costs. To solve this problem, a method for predicting core permeability based on deep learning using CT scan images with different resolutions is proposed in this work. First, the high-resolution CT scans are preprocessed and then cubic subsets are extracted. The permeability of each subset is estimated using the Lattice Boltzmann Method (LBM) and forms the training set for the convolutional neural network (CNN) model. Subsequently, the high-resolution images are downsampled to obtain the low-resolution grayscale images. In the comparative analysis of the porosities of different low-resolution images, the low-resolution image with a resolution of 10% of the original image is considered as the test set in this paper. It is found that the permeabilities predicted from the low-resolution images are in good agreement with the values calculated by the LBM. In addition, the test data are compared with the results of the Kozeny-Carman (KC) model and the measured permeability of the whole sample. The results show that the prediction of the permeability of tight carbonate rock based on deep learning using CT scans with different resolutions is reliable.