Machine learning with hyperparameter optimization applied in facies-supported permeability modeling in carbonate oil...

Area of Science:

• Petroleum Geoscience
• Machine Learning Applications
• Reservoir Engineering

Background:

• Carbonate reservoirs have heterogeneous pore distribution, with low matrix permeability hindering oil flow.
• High-permeability fractures are key conduits, but permeability measurements (core, well tests) are costly and unavailable for many wells.
• Accurate permeability prediction is vital for 3D petrophysical property distribution and efficient field development.

Purpose of the Study:

• To evaluate the performance of six machine learning algorithms for predicting core permeability using conventional well logs.
• To compare data pre-processing techniques and hyperparameter tuning methods for optimizing machine learning models.
• To identify the best-performing machine learning algorithm and configuration for permeability prediction in uncored wells.

Main Methods:

• Utilized a high-quality dataset with multiple well-log inputs (gamma ray, porosity, resistivity, etc.) and core measurements.
• Applied data pre-processing: missing data imputation, scale correction, normalization (log, Box-Cox, NST), and outlier detection.
• Compared random search and Bayesian optimization for hyperparameter tuning; evaluated performance using RMSE, MAE, R², and Adjusted R².

Main Results:

• The XGBoost algorithm, configured with random search, Box-Cox normalization, Z-score outlier detection, and without scale correction, demonstrated superior performance.
• Achieved the lowest prediction errors: 6.9 md (training) and 9.78 md (testing) Root Mean Square Error (RMSE).
• This configuration highlights the importance of specific pre-processing and tuning strategies for accurate permeability prediction.

Conclusions:

• XGBoost is a highly effective machine learning algorithm for predicting permeability in carbonate reservoirs from well logs.
• Optimized data pre-processing and hyperparameter tuning are critical for enhancing prediction accuracy.
• The developed model provides a cost-effective solution for permeability prediction in uncored wells, aiding reservoir management.