3D facies modeling of a carbonate reservoir in the Santos Basin using machine learning

Abstract

Three-dimensional facies modeling is essential for optimizing hydrocarbon exploration and production, particularly in complex carbonate reservoirs. Traditional classification methods rely on manually delineating homologous zones in scatter plots of elastic parameters, introducing subjectivity and limiting predictive accuracy. This study implements a machine learning-based workflow to classify petroelastic facies, flow units, and geomechanical units, minimizing arbitrary thresholding and improving interwell predictions. A supervised learning model was trained using well log data and validated through cross-validation, incorporating pseudo-curves derived from seismic inversion products (e.g., Zp, Zs) to enhance resolution. Data resampling techniques were applied to address class imbalance within the Barra Velha Formation. The resulting classification achieved 70–95 % accuracy and generated 3D models that highlight distinct facies distributions, with reservoir facies concentrated in the top and bottom intervals, while the middle interval exhibited predominantly non-reservoir facies, consistent with petrographic analyses. The facies distribution also revealed elongated reservoirs aligned with NE-SW structural highs, compartmentalized by faults of the same orientation. This improved understanding of facies distribution supports better placement of wells and reduced development risk. By eliminating the need for manual delineation of facies zones, this study presents a novel approach to three-dimensional facies classification from seismic inversion data, offering an objective and efficient alternative to more traditional methods used for reservoir characterization.