Predicting the shear strength parameter of cohesionless soil using machine learning techniques

This study uses sophisticated machine learning algorithms to predict the angle of internal friction Phi of silty sand (SM) from SPT-N value and percentage Fines as the independent variables, based on the Unified Soil Classification System (USCS) in the context of Bangladesh. A hefty dataset encompassing a thousand boreholes from various parts of Bangladesh was used for this study. Three regression methods were utilized to develop an integrated correlation model: Multiple Linear Regression (MLR), Deep Neural network (DNN), and XGboost. Evaluation metrics: Mean Absolute Error (MAE), R-squared (R2), and Root Mean Squared Error (RMSE), were exercised to determine which model performed the best. In addition, sensitivity analysis was done to find out the importance of parameters in predicting Phi for the chosen model. The evaluation metrics showed that XGboost had outperformed the other models, providing a more precise prediction of Phi for the silty sand of Bangladesh's soil. Unlike the previous widely used empirical models, this model uses percentage Fines as one of the predictors for Phi , and the soil data is native to the country's soil. In most of the projects in Bangladesh, standard penetration tests (SPT) are the only means to find all the engineering properties of soil due to limited budgets, tight schedules, a lack of concern, and the inability to perform laboratory tests. This demonstrates that this model's implementation could result in more reliability in the field of geotechnical engineering during the feasibility/preliminary design stage for rapid prediction of friction angle based on the SPT-N values and percentage Fines in the context of Bangladesh.