Machine Learning-Based Prediction of Well Logs Guided by Rock Physics and Its Interpretation

Area of Science:

• Geophysics
• Machine Learning
• Petroleum Engineering

Background:

• Traditional well log refinement uses rock physics models with limitations.
• Machine learning (ML) offers an alternative but often lacks explicit interpretability.
• Integrating ML with rock physics principles is a novel approach for well log analysis.

Purpose of the Study:

• To predict porosity and clay volume fraction using four ML algorithms.
• To assess the alignment of ML predictions with rock physics principles using SHAP analysis.
• To explore the interpretability of ML models in well log analysis.

Main Methods:

• Employed Random Forests (RF), Gradient Boosting Decision Trees (GBDT), Multilayer Perceptrons (MLP), and Linear Regression (LR).
• Guided feature engineering and interpretation using rock physics principles (Gardner and Larionov relations).
• Utilized SHapley Additive exPlanations (SHAP) to analyze algorithm behavior and interpret predictions.

Main Results:

• Satisfactory predictions for porosity and clay volume fraction were achieved.
• SHAP analysis revealed consistent patterns across all four ML algorithms, aligning with rock physics principles.
• ML algorithms and SHAP analysis demonstrated adherence to rock physics cause-effect relationships, even when critical input features were omitted.

Conclusions:

• The integration of ML with rock physics principles enhances the interpretability of well log analysis.
• SHAP analysis provides a bridge between mathematical predictions and philosophical understanding of ML models.
• This approach moves beyond the traditional 'black box' nature of ML, offering deeper insights into geoscientific data.