Application of unsupervised learning and deep learning for rock type prediction and petrophysical characterization using multi-scale data

This study integrates well log data, routine core analyses, microcomputed X-ray tomography (μCT) images, and sedimentary petrography to accurately characterize and evaluate the carbonate reservoirs of the Barra Velha formation (Aptian) of the Santos Basin within Brazilian pre-salt region. In these carbonate reservoirs, the porous system is extremely diverse and variable, making it challenging to establish rock typing with comparable petrophysical properties. Based on this integrated study, the reservoir sequences were characterized and a precise definition of four reservoir rock types (RRTs) was performed by integrating the petrophysical values from the plugs and their corresponding well log data of two cored wells using K-means unsupervised classification algorithm. The classification results were combined with various conventional techniques to evaluate the quality and geological characteristics of the studied sequence. This evaluation encompassed different parameters such as flow and storage capacity, reservoir quality index, flow zone indicator, pore spaces interpretation, and average pore and throat radius. The study involved a detailed analysis of thin sections to identify various facies, including shrubstones, reworked, and spherulitestone, and to classify various forms of porosity such as interparticle, intraparticle, intercrystalline, vug, moldic, fracture, and growth framework porosity. Pore Network Modeling from μCT analysis of plugs was used specifically for the characterization of pores and throats of plug samples from each RRT. These datasets were utilized as supporting evidence to offer a more accurate and inclusive knowledge of reservoir quality. The study aimed to develop predictive models by implementing deep learning and machine learning algorithms trained on well log data to estimate plug porosity and rock type. Two deep learning models, ResNet and 1D CNN, were trained and evaluated for plug porosity prediction, with the 1D CNN model showing superior performance. Additionally, the XGBoost algorithm was applied to predict rock type, achieving high accuracy on both the training and validation datasets. The predicted results were compared with actual data to evaluate the effectiveness of the models and were then utilized to estimate plug permeability values. The results demonstrate the potential of deep learning and machine learning approaches in reservoir characterization and management, enabling the evaluation of subsurface reservoir properties even with incomplete datasets, which could lead to an improved understanding of the reservoir properties and better management of the reservoir. This integrated study provides deeper insight into the complex reservoir properties and can help improve decision-making processes and optimize management and production strategies in the challenging pre-salt carbonate reservoirs or similar complex reservoirs.