Machine learning assisted geophysical characterization of deep-seated upper jurassic carbonate deposits in Penobscot Field, Nova Scotia

Abstract

Carbonate rocks hold significant potential for economic activities, including mineral and hydrocarbon exploration, highlighting the importance of detailed characterization. This study focuses on delineating and characterizing deep-seated Upper Jurassic carbonate deposits (Abenaki Formation) in the Penobscot Field using model-based inversion (MBI) and machine learning (ML) algorithms. The MBI model achieved 94.3% correlation with a relatively lower error of 605.8 m/s∗g/cm³ in impedance estimation. The ML models were evaluated using metrics such as precision, recall, F1 score, accuracy, and misclassification rates. The XGB model consistently outperformed the RF, ANN, and SVM models, achieving the highest precision, recall, F1 score, and accuracy across shale, sand, and carbonate facies classifications. It recorded an overall accuracy of 0.927 and a misclass rate of 0.073, surpassing SVM (accuracy: 0.901, misclass: 0.099), RF (accuracy: 0.866 & misclass: 0.134), and ANN (accuracy: 0.838 & misclass: 0.162). Furthermore, the effective porosity volume was predicted with a correlation of 85.75% and a mean absolute error of 0.02. It was found that the Artimon Member (∼85 m) includes an upper porous carbonate reservoir unit (∼35 m) with impedance 11,500–15,000 m/s∗g/cm³, carbonate probability 70–80% and porosity 12–15%, and a deeper siliciclastic unit (∼45 m) with impedance 9500–13,000 m/s∗g/cm³, carbonate probability 20–30% and porosity nearly 3–4% possibly during a significant transgressive phase of sea-level rise. The underlying Baccaro Member (∼260 m) predominantly comprises thick carbonate facies with impedance 12,500–16,000 m/s∗g/cm³, carbonate probability 80–90% and porosity nearly 6–9%. This quantitative study examines how depositional environments, mineralization, and diagenesis shape the distribution of carbonate facies in the Penobscot Field. By integrating advanced seismic inversion with machine learning, it refines the characterization of deep-seated carbonate facies, offering insights for identifying potential carbonate hydrocarbon bearing zones worldwide.