Perpustakaan Digital ITB

This study applies supervised machine learning to estimate key geomechanical parameters—static and dynamic Young’s modulus, Poisson’s ratio, and minimum horizontal stress gradient—using only triple-combo logs from a shaly sandstone reservoir in the ITB Field, North-West Java. Two predictive strategies were evaluated: a direct approach using triple-combo logs with predicted compressional and shear sonic logs, and an indirect approach using only triple-combo logs. Five machine learning algorithms—random forest, XGBoost, LGBM, k-nearest neighbors, and feedforward neural network—were trained and tested on data from two wells with complete logging and geomechanical measurements. Direct random forest delivered the best performance across most targets: compressional sonic log (R² = 0.9173), shear sonic log (R² = 0.9286), dynamic poisson’s ratio (R² = 0.7755), and static poisson’s ratio (R² = 0.7746). Direct XGBoost provided the best prediction for static young’s modulus (R² = 0.9418) and minimum horizontal stress gradient (R² = 0.9344). Direct LGBM provided the best prediction of dynamic young’s modulus (R² = 0.9399). Although indirect models showed strong performance, direct predictions were generally more stable due to predicted sonic log. SHAP analysis revealed that direct models were dominated by predicted compressional and shear sonic logs inputs, while indirect models rely on depth, density, resistivity, and porosity, aligning with petrophysical principles. A practical and explainable machine learning framework is introduced based solely on triple-combo logs, employing both direct and indirect approaches, and integrates SHAP analysis to ensure petrophysical interpretable geomechanics predictions in datalimited shaly sandstone reservoirs.