Estimating two-dimensional physical parameters of digital rocks using deep learning

Abstract This research focused on estimating the physical parameters of porous rocks crucial in hydrocarbon exploration using deep learning algorithms. Laboratory measurements have limitations such as time, cost, and core sample limitations, so digital rock models have emerged as a powerful alternative. Digital rock technology involves creating high-resolution images of rock samples using techniques such as micro-CT scanning for the detailed analysis of rock structures and calculation of physical parameters through image processing and numerical simulations. In this work, the CNN architectures included custom-developed models, and transfer learning was applied using pre-trained models DenseNet201, ResNet152, MobileNetV2, InceptionV3, and Xception to estimate physical parameters such as permeability, absolute porosity, effective porosity, tortuosity, and average grain size. Both CNN A and CNN B were good models for estimating permeability with CNN B being the best model for estimating tortuosity, Xception the best model for estimating absolute porosity and effective porosity, and DenseNet201 the best model for estimating average grain size. These results underscore the potential of deep learning in enhancing the efficiency and accuracy of physical parameter estimation in digital rock analysis.