Determination of transport properties of rock cores based on neural network algorithms

In this work, a method for determining the transport properties of rock cores based on neural network segmentation of microtomographic images is proposed. The use of deep convolutional networks, particularly the U-Net architecture with a pretrained ResNet34 encoder, made it possible to improve the accuracy of pore structure delineation under conditions of limited volume of labeled data. The proposed approach includes an automatic segmentation stage, followed by the construction of a three-dimensional model of the pore space and geometric calculation of permeability characteristics, including through porosity and minimum cross-sectional area of channels. Application of data augmentations and dice loss function ensured model resistance to overfitting and high-resolution boundary delineation between phases. The results demonstrate the effectiveness of integrating modern machine learning methods into digital petrophysics tasks, opening up opportunities for analysis and prediction of filtration properties of rocks without using traditional numerical simulation techniques.