Identification and Interpretability Analysis of Shale Reservoir Types: Insights from LightGBM and SHAP Algorithms

Area of Science:

• Geosciences
• Petroleum Engineering
• Machine Learning

Background:

• Conventional methods struggle with complex nonlinear relationships in deep shale reservoir characterization.
• Accurate prediction of shale reservoir types is crucial for effective resource evaluation and development.

Purpose of the Study:

• To develop and optimize a machine learning model for identifying deep shale reservoir types.
• To compare the performance of classification-based versus regression-based schemes using LightGBM.
• To evaluate the importance of different logging curves in reservoir type identification.

Main Methods:

• Utilized the Light Gradient Boosting Machine (LightGBM) algorithm for reservoir type identification.
• Implemented SHapley Additive exPlanations (SHAP) for quantitative feature importance analysis.
• Compared classification-based and regression-based modeling schemes.

Main Results:

• The classification-based LightGBM scheme achieved weighted precision of 90.5% and weighted recall of 90.4%, outperforming the regression-based scheme (85.9% and 86.1%).
• Key logging curves influencing reservoir type identification include compensated density (DEN), gamma ray (GR), compensated neutron (CNL), and acoustic transit time (AC).
• SHAP analysis revealed complex nonlinear relationships between logging responses and reservoir classification outcomes.

Conclusions:

• The optimized LightGBM model provides an efficient and accurate method for deep shale reservoir type identification and grading.
• Machine learning, particularly LightGBM, offers significant advantages over traditional methods for complex reservoir characterization.
• This approach offers a novel framework for the comprehensive grading evaluation of deep shale reservoirs.