Advanced machine learning approaches for predicting permeability in reservoir pay zones based on core analyses

Area of Science:

• Petrophysics and Reservoir Engineering
• Machine Learning Applications in Geoscience
• Computational Intelligence for Earth Sciences

Background:

• Reservoir rock permeability is a critical petrophysical property governing fluid flow.
• Accurate permeability prediction is essential for effective hydrocarbon reservoir management.
• Traditional laboratory methods for permeability determination are often time-consuming and resource-intensive.

Purpose of the Study:

• To develop and evaluate intelligent computer-based models for predicting reservoir permeability.
• To compare the performance of novel models (Decision Tree, Bagging Tree, Extra Trees) against established techniques (Random Forest, SVR, MVR).
• To identify the most effective machine learning approach for accurate permeability forecasting.

Main Methods:

• Utilized a dataset of 197 data points from a heterogeneous petroleum reservoir in the Jeanne d'Arc Basin.
• Collected data included laboratory-derived permeability (K), oil saturation, water saturation, grain density, porosity (φ), and depth.
• Employed statistical metrics (R², MSE, MAE, RMSE, MAPE, maxE, minE) to assess model performance and ranked feature importance.

Main Results:

• Extra Trees model achieved the highest accuracy with an R² of 0.976, surpassing Random Forest (0.961) and Bagging Tree (0.964).
• Multiple Variable Regression (MVR) proved unsuitable for permeability prediction, with all machine learning models significantly outperforming it.
• Porosity (φ) and water saturation were identified as the most influential input parameters for permeability modeling.

Conclusions:

• Machine learning models, particularly Extra Trees, offer a highly accurate and efficient alternative to traditional laboratory methods for permeability prediction.
• The study highlights the limitations of MVR and the superior predictive power of ensemble machine learning techniques in reservoir characterization.
• The developed models enable rapid and reliable permeability estimation using core attributes, reducing reliance on costly and time-consuming lab work.