Evaluating the performance of an optimized tree-ensemble learning algorithm for predicting sonic logs in the Gandhar CO2 EOR project

Abstract

Predicting unrecorded well-logs is essential for improving subsurface characterization, particularly for CO₂ storage in geological formations. This study presents a novel and optimized tree-ensemble learning algorithm for predicting compressional (P)-sonic logs in the Gandhar oilfield of India. Specifically, we developed a Gradient Boosting Regressor (GBR) by optimizing hyperparameters using a cross-validated grid search technique to enhance prediction accuracy and uncertainty quantification. Optimal input features (gamma-ray, neutron-porosity, resistivity, and density logs) were selected based on their significant correlation with the target (P-sonic log) measurements and their contribution to minimizing prediction error. The optimized GBR model was trained and tested on wells containing the optimal input features and the target measurements, and then applied to predict unrecorded P-sonic logs in three blind wells. Results showed that the algorithm predicted the overall trend and amplitude of the actual P-sonic log with high prediction accuracy and effectively captured lithological variations. Compared to the empirical methods, GBR demonstrated superior performance with lower mean absolute error and root mean square error. Prediction errors stabilized beyond 20,000 data points, suggesting further improvement depends on more representative lithological data. Prediction intervals highlighted lower uncertainty (higher model confidence) in zones with narrower intervals and abundant training data, particularly in the Hazad sands. Conversely, wider intervals reflected greater uncertainty in underrepresented lithological zones. Predicted logs were successfully utilized to model CO₂-saturated velocities in the Hazad sands. This approach provides a robust, scalable machine learning algorithm with optimized hyperparameters for enhanced well-log prediction and uncertainty quantification, supporting reliable, risk-informed reservoir characterization.