Sensitivity analysis of slope stability based on eXtreme gradient boosting and SHapley Additive exPlanations: An...

Area of Science:

• Geotechnical Engineering
• Artificial Intelligence
• Data Science

Background:

• Slope instability poses significant risks, necessitating robust analysis methods in geotechnical engineering.
• Traditional slope stability analysis methods are often complex and time-consuming.
• Existing AI-based approaches identify factor importance but lack quantified correlation analysis.

Purpose of the Study:

• To perform sensitivity analysis on slope stability factors using XGBoost and SHAP.
• To quantify the correlation between key geotechnical parameters and slope stability.
• To develop a data-driven approach for more accurate AI interpretation in slope stability assessment.

Main Methods:

• Sensitivity analysis employing XGBoost and SHAP algorithms.
• Validation of SHAP results using GeoStudio software simulations.
• Implementation of a priori data-driven approach to enhance AI interpretation.

Main Results:

• Identified slope height and angle of internal friction as the most and least influential parameters, respectively.
• GeoStudio simulations revealed negative correlations for slope height, slope angle, unit weight, and pore water pressure coefficient; positive correlations for cohesion and angle of internal friction.
• The data-driven approach provided more reliable critical values for destabilization (e.g., cohesion: 18 Kpa, slope angle: 28°, internal friction angle: 32°, slope height: 30m, pore water pressure coefficient: 0.28) compared to raw AI interpretation.

Conclusions:

• XGBoost and SHAP offer valuable tools for sensitivity analysis in slope stability.
• A priori data-driven methods are crucial for overcoming AI limitations in understanding real-world geotechnical mechanisms.
• The proposed approach yields more accurate and reliable results, especially with limited or low-quality data.