Perpustakaan Digital ITB

Reservoir characterization is essential in field development planning, particularly in heterogeneous formations where accurate interpretation of well log data strongly influences production decisions. Conventional interpretation methods are often subjective and inconsistent due to reliance on expert judgment. To overcome these limitations, this study proposes a data-driven workflow that integrates unsupervised clustering for reservoir zonation based on similar petrophysical characteristics, which is subsequently used to support supervised modeling. A supervised machine learning framework is developed for cumulative gas prediction using conventional well log data, including gamma ray, resistivity, bulk density, and neutron porosity. Data preprocessing is conducted through normalization to ensure feature scale consistency. Multiple base learners are trained to capture nonlinear relationships between log responses and cumulative gas, and their outputs are integrated using a stacking ensemble with a metalearner to improve predictive robustness. To enhance interpretability, an Adaptive Neuro-Fuzzy Inference System (ANFIS) is incorporated into bridge data-driven learning with rule-based geological reasoning. In addition, Sobol normalized sensitivity analysis is applied to quantify feature contributions and assess model reliability in a consistent interpretability framework. A case study from Field X, Natuna Basin, Indonesia, is used to validate the proposed approach using labeled production and well log data. The results demonstrate that the proposed framework improves cumulative gas prediction while maintaining interpretability and provides a robust data-driven basis for identifying prospective perforation intervals in field development planning.