MACHINE LEARNING-DRIVEN PREDICTION OF BOND INDEX FROM CEMENT BOND LOGS FOR ENHANCED WELL INTEGRITY

Abstract

Ensuring wellbore integrity through reliable cement zonal isolation is critical for safe and sustainable hydrocarbon and geothermal operations. Conventional interpretation of Cement Bond Logs (CBLs) and the derived Bond Index (BI) often depend on manual waveform analysis, which introduces subjectivity and delays in decision-making. In this study, we developed a machine learning framework to automate BI prediction directly from raw wireline acoustic data and well parameters. Data were obtained from well 15/9-F-11 B in the decommissioned Volve field, which provided extensive CBL recordings in a geologically heterogeneous setting. The raw CBL data in DLIS format were processed, cleaned, and standardized to generate a structured dataset suitable for supervised learning. Four regression algorithms—Random Forest, Gradient Boosting, Support Vector Regression, and Linear Regression—were trained using an 80:20 depth-preserving split and evaluated with MAE, RMSE, MAPE, and R². Ensemble methods delivered superior performance, with Random Forest achieving MAE = 0.00027, RMSE = 0.00048, and R² = 0.999996. Gradient Boosting also produced highly accurate predictions, while Linear Regression and SVR showed significantly lower performance. The findings demonstrate that tree-based ensembles can capture complex non-linear relationships between acoustic signals and cement bond quality with near-perfect accuracy. Integrating such models into logging workflows enables faster, more objective assessment of zonal isolation, supporting timely interventions and long-term well integrity.