Abstract
This study presents a novel approach to accurately predict the settlement of shallow foundations using advanced machine learning techniques while assessing the influence of key variables. Four machine learning models Gradient Boosting (GB), Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN) are enhanced with Particle Swarm Optimization (PSO) for hyperparameter tuning, resulting in hybrid models GB-PSO, RF-PSO, SVM-PSO, and KNN-PSO. The experimental dataset comprises 189 samples, and model performance is rigorously evaluated through K-Fold Cross-Validation alongside R², RMSE, MAE, and MAPE metrics. The results indicate that PSO tuning does not consistently improve the prediction accuracy, with the original models, particularly GB and RF, outperforming their PSO-optimized counterparts. Sensitivity analysis via Shapley Additive Explanation (SHAP) highlights average Standard Penetration Test blow count (SPT) and footing width (B) as the most influential variables, with footing embedment ratio (D(f)/B) and net applied pressure (q) also significantly impacting settlement predictions. The study offers a new Excel tool based on the GB model, facilitating practical applications for civil engineers, and providing a dependable, user-friendly tool to predict shallow foundation settlement.