Perpustakaan Digital ITB

Cement bond evaluation is essential for well integrity assurance, typically assessed through post-job acoustic logging tools such as the Cement Bond Log (CBL). However, incomplete or missing CBL data is a common limitation in older wells. This study aims to predict post-job CBL values based solely on pre-job well logs using a machine learning framework. Rather than assuming which specific log features determine cement bond quality, this study adopts a data-driven hypothesis testing approach. The result is used not only to make predictions but also to identify which pre-job petrophysical logs consistently correlate with CBL amplitude. The approach involves training XGBoost on three datasets: a global dataset covering the full logged interval, and two lithology-specific subsets representing Alluvium and Granodiorite. A total of nine pre-job features were used, including porosity, density, gamma ray, caliper, and resistivity logs. Model performance was evaluated using R² and RMSE, with additional SHAP-based interpretation to assess feature influence. A minimal feature set consisting of just GR, NPHI, RXOZ, and HCAL still preserved considerable predictive power, achieving an R² of 0.602 and RMSE of 8.845 mV. This result aligns with the earlier SHAP interpretation in the model overview.