Advancing in creep index of soil prediction: A groundbreaking machine learning approach with Multivariate Adaptive Regression Splines

The significance of accurately predicting the creep index coefficient for assessing the long-term settlement of soil is critical in geotechnical engineering. However, current empirical methods for determining the creep coefficient often lack precision, highlighting the need for a more accurate predictive model. This study employs the Multivariate Adaptive Regression Splines (MARS) model to predict the creep index in clay, a critical parameter in geotechnical engineering. The dataset was divided into training and testing subsets. Grid search hyperparameter tuning was applied to optimize the MARS model, as well as a black-box model (Support Vector Machine, SVM) and a white-box model (Lasso), with five-fold cross-validation used to assess their performance. MARS demonstrated superior predictive accuracy, as evidenced by the mean and median R² and RMSE values obtained from the five-fold cross-validation results. The optimized MARS model was then applied to the test set, achieving excellent predictive accuracy. Finally, the model's performance was compared to previously developed machine learning models and empirical equations across the entire dataset. The MARS model outperformed all others based on RMSE, R², MAE, and KGE metrics, highlighting its robustness and reliability in predicting the clay creep index.