Fast high-frequency porosity characterization from computer tomography images and deep learning

This work presents a methodology for combining deep learning with whole-core computed tomography (CT) imaging and laboratory data for fast and high-frequency estimation of porosity logs. The proposed method trains a convolutional neural network (CNN) model to predict laboratory-scale porosities by using core-scale CT images as input. Despite the different scales, the model works well to predict the porosities measured at laboratory scales using the textures present in core-scale CT images. The main contribution of this work is the data-driven methodology for fast and high-frequency characterization of porosity at the core scale. The proposed methodology can provide a much more efficient porosity characterization than the traditional workflow based on the qualitative and subjective evaluation of experts. The method was evaluated in a well-based cross-validation process in 26 wells from 3 different Brazilian pre-salt fields, using two distinct CNN architectures of varying complexity. The results show that the models can make good predictions for new wells with acceptable margins of error. The proposed method can be easily integrated into digital rock workflows as it is based on data available for traditional petrophysical analyses and deep learning models available in open source.