عنوان مقاله [English]
Accurately predicting subsurface flow properties holds immense significance across various domains, ranging from water resource management to the petroleum industry. In this study, recognizing the computational intensity and time constraints associated with digital rock analysis for petrophysical property calculations, we introduce a workflow that leverages deep learning to swiftly and precisely estimate these properties from micro-CT images, obviating the need for resource-intensive computational methods. Specifically, a Convolutional Neural Network (CNN) was employed to train and predict multiple physical properties of porous media using micro-CT scan images as input data. The micro-CT scan images, derived from a carbonate rock sample, were divided into 9,261 images, each with dimensions of 100x100x100, for network training. Key parameters such as porosity, throat size, pore size, connection number, and pore shape factor for each image were computed using network extraction algorithms. The designed network›s performance was evaluated, considering factors like the number of layers and learning rate. Subsequently, when tested on a separate dataset, the network exhibited impressive coefficients of determination for the mentioned parameters, namely 99% for porosity, 90.2% for Avg.throat size, 94.5% for Avg.pore size, 93.6% for Avg.connection number, and 75.3% for Avg.pore shape factor. Furthermore, the average relative error percentage for each property remained below 4%. These results signify a strong agreement between the predicted values and the actual properties, affirming the efficacy of this approach in swiftly and accurately estimating petrophysical properties from micro-CT images.