Prediction of soil shear strength using hybrid machine learning approaches for performance and interpretability analysis

Abstract Shear strength of soil plays an essential role in geotechnical properties affecting construction stability. Conventional laboratory testing for determining soil shear strength is often expensive and time-consuming. Therefore, machine learning (ML) methods were employed to predict soil shear strength using geotechnical parameters compiled from previously published literature based on investigations conducted at the Le Trong Tan Geleximco project in western Hanoi, Vietnam. In total, there were 202 samples analyzed, including parameters such as depth, sand percent, loam percent, clay percent, moisture content, wet density, dry density, void ratio, liquid limit, plastic limit, plasticity index, and liquidity index. Four predictive models have been developed and analyzed, including Multiple Linear Regression (MLR), Support Vector Machine (SVM), Random Forest (RF), and Multi-Expression Programming (MEP). The model performance was evaluated using several parameters like coefficient of determination (R 2 ), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), residual analysis, and Taylor diagrams. RF showed the best predictive performance, with training R 2 = 0.9527 and testing R 2 = 0.8578, along with the lowest prediction error. The SVM model performed impressively in terms of prediction, while the MEP model demonstrated satisfactory accuracy along with the added benefit of mathematical equation formulation. Conversely, the MLR model was relatively less accurate because of its inability to deal with nonlinear relations. Moreover, the SHAP analysis revealed that liquid index, moisture content, and plasticity index were the most critical factors influencing the prediction of soil shear strength. This research has established that modern machine learning models like RF and SVM prove to be efficient at modeling the nonlinear nature of soil characteristics.