An Artificial Intelligence-Based Data-Driven Method for Predicting Soil Shear Strength

Accurate prediction of soil shear strength is critical for safe and cost-effective geotechnical design. This study investigates the application of four machine learning (ML) models—Random Forest (RF), Decision Tree (DT), Artificial Neural Network (ANN), and Support Vector Regression (SVR)—to predict the shear strength of soils from Bahir Dar city using laboratory-obtained geotechnical data. A total of 298 soil samples and 13 geotechnical parameters were collected from depths of 0.13–35 m, encompassing both disturbed and undisturbed conditions. The dataset was divided into training (80%) and testing (20%) sets, and models were trained with optimized hyperparameters. The RF model achieved the highest accuracy (R2 = 0.9992, RMSE = 0.0983), followed by DT (R2 = 0.9974, RMSE = 0.1812). ANN and SVR showed lower predictive accuracy, with SVR demonstrating the largest maximum errors. Predicted vs. actual plots, kernel density estimates, and absolute error per sample analysis confirmed that tree-based models provide the most reliable predictions, while ANN and SVR exhibited sporadic large deviations. SHAP analysis revealed that Cohesion, Clay content, and Plasticity Index are the most influential factors in predicting shear strength. The results demonstrate that ensemble tree-based ML models offer a robust and accurate tool for geotechnical prediction, capturing complex nonlinear relationships in soil behavior.