Advanced hybrid machine learning models with explainable AI for predicting residual friction angle in clay soils

Area of Science:

• Geotechnical Engineering
• Machine Learning
• Artificial Intelligence

Background:

• Accurate estimation of residual strength friction angle in clay soils is crucial for geotechnical structure stability.
• Traditional methods are labor-intensive, time-consuming, and costly.
• Existing predictive methodologies have limitations.

Purpose of the Study:

• To explore advanced hybrid machine learning models for predicting the residual friction angle of clay soils.
• To address the critical gap in current predictive methodologies.
• To improve the accuracy and reliability of predictions for geotechnical applications.

Main Methods:

• Utilized a harmonized dataset of 400 global soil samples.
• Employed three hybrid machine learning models: Gradient Boosting Neural Network (GrowNet), Reinforcement Learning Gradient Boosting Machine (RL-GBM), and a Stacking Ensemble.
• Applied Explainable Artificial Intelligence (XAI) techniques (SHAP, LIME) for model transparency.

Main Results:

• GrowNet achieved the highest coefficient of determination (R² = 0.94) and lowest RMSE (1.87) and MAE (1.17) on the testing dataset.
• GrowNet demonstrated substantial improvement over traditional empirical correlations and previous machine learning approaches.
• XAI techniques identified Clay Fraction and Plasticity Index as the most influential input variables.

Conclusions:

• Integrating high-performing machine learning models with interpretability tools enhances residual friction angle prediction accuracy and reliability.
• The developed models offer practical value for geotechnical engineering applications, particularly in landslide-prone regions.
• GrowNet shows significant potential for improving the design and stability evaluation of geotechnical structures.