Application of machine learning methods for predicting swelling pressure of expansive clays

Swelling pressure is a critical property of expansive clays that can induce significant structural damage due to moisture-driven volume changes. This study presents a machine learning-based framework to predict the swelling pressure of expansive clays using easily obtainable geotechnical parameters, including liquid limit, plastic limit, dry unit weight, specific gravity, fine content and free swell index. Six models, namely, Random Forest Regressor (RFR), k-Nearest Neighbour (kNN), Support Vector Regressor (SVR), Artificial Neural Network (ANN), XGBoost (XGB) and XGBoost with Particle Swarm Optimisation (XGB-PSO) were systematically evaluated using a dataset of 325 clay specimens. Model performance was assessed on training (70%), testing (20%) and validation (10%) datasets using multiple statistical measures, including R2, RMSE, MSE, MAD, MAPE and MAE. The hybrid XGB-PSO model consistently outperformed all other models, achieving the highest R2 values (0.97 for training, 0.96 for testing and validation) and the lowest error metrics. Sensitivity analysis using SHAP revealed dry unit weight and free swell index as the most influential predictors. Independent assessment revealed that the predicted swelling pressure using the XGB-PSO model matched the experimental value with an error of 1.2%. These results highlight the potential of hybrid machine learning approaches to provide reliable predictions of swelling pressure.