Integrated machine learning workflow for 3D pore type modeling in the pre-salt carbonates of the Tupi Field, Santos Basin

Abstract

The discovery of pre-salt carbonates in Brazil has significantly reshaped the country's oil and gas industry, now accounting for over 78 % of national production. However, the complex heterogeneity of these reservoirs, characterized by unique mineral compositions, intricate pore systems, and diagenetic processes, poses significant challenges for reservoir characterization. This study introduces a comprehensive methodology to address these challenges, focusing on the carbonate-bearing Barra Velha Formation in the Tupi Field, Santos Basin. The workflow integrates seismic inversion, well log data, and advanced machine learning models, including the Gradient Boosting Classifier, Random Forest Regressor, and Gradient Boosting Regressor. These models are used to classify petroelastic facies, predict petrophysical properties, and generate 3D pore type volume maps for compliant, reference, and stiff pores. Key findings reveal significant heterogeneity in pore distributions, with the lower Barra Velha Formation exhibiting greater variability. Stiff pore volumes display an inverse relationship with reference pores, forming distinct “stiff corridors” in certain regions, while compliant pores are localized around transition zones. The results demonstrate strong correlations between well log data and seismic-scale predictions, highlighting the methodology's reliability. These 3D pore type volume models provide valuable insights into reservoir heterogeneity, aiding in the identification of high-quality reservoir zones and supporting improved exploration and production strategies. This study underscores the importance of incorporating geological, petrophysical, and diagenetic factors into reservoir characterization workflows and emphasizes the adaptability of the proposed methodology to different geological settings.