Estimating Rock Typing in Uncored Wells Using Machine Learning Techniques for Brazilian Pre-Salt Carbonate Reservoir

Abstract Accurate rock typing in uncored wells is essential for enhancing reservoir models, particularly in complex geological formations like the Brazilian pre-salt carbonate reservoirs. This study explores the application of machine learning (ML) techniques to estimate rock types in uncored wells. The research leveraged core data from 11 cored wells to calculate the Rock Quality Index (RQI) and Flow Zone Indicator (FZI), identifying 14 distinct rock types through the Discrete Rock Typing (DRT) method, along with well log data such as gamma ray, density, neutron, permeability, and porosity. Various machine learning models, including Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree, Gradient Boosting, Naive Bayes, XGBoost, and Random Forest were tested, where XGBoost achieved the highest accuracy of 73.3%. Applying XGBoost to all wells resulted in accuracy ranging from 0.6 to 0.91 and the model was subsequently used to estimate rock types in over 20 uncored wells to generate reservoir simulation models. This study highlights the efficacy of machine learning (particularly XGBoost) in addressing reservoir complexities and offering significant improvements in the understanding and development of carbonate reservoirs.