CNN-Transfer Learning-Based Prediction for Porosity and Absolute Permeability from Carbonate Rock Images

This study is intended to compare the capabilities of three different deep learning-based convolutional neural network models in predicting reservoir rock porosity and absolute permeability from 2D carbonate rock images. We consider a comprehensive evaluation scenario to investigate the performance and training time involved in using the proposed models. These are studied and evaluated using 2D micro-CT images captured at various image resolutions from the four different core samples. The selected core samples demonstrate a wider range of absolute permeability and different levels of heterogeneity. We achieve model variability by adopting the transfer learning framework in two of the three designed models using pre-trained, VGG16, and MobileNetV2 models. Results obtained demonstrate that transfer learning improves model accuracy to predictions at the expense of computational time. With the influence of transfer learning, results show that the accuracy and computational time largely depend on the number of trained parameters being transferred. The proposed models can predict both the rock porosity and absolute permeability within a few seconds compared to numerical simulations and experiments which require larger amounts of time.