Enhancing Swelling Pressure Prediction Through Integrated Machine Learning Methods

Abstract

Swelling soils, characterized by their high capacity to change volume due to variations in water content, pose significant challenges in civil engineering, including structural deformations, cracks in infrastructure, and foundation instabilities. These issues lead to high maintenance costs and safety risks for structures. To address these challenges, this study proposes an integrated approach combining machine learning and feature selection techniques to predict the swelling pressure (SP, kPa) of soils. Using a dataset comprising seven explanatory variables (liquid limit (LL, %), plasticity index (PI, %), shrinkage limit (WS, %), particle size fractions (Ff2mm, Ff80µm, Ff2µm), and methylene blue index (VBS, g/100 g)), four feature selection methods were evaluated: Least Absolute Shrinkage and Selection Operator (Lasso), Random Forest, Gradient Boosting, and the F-statistic test. These methods identified the five most influential variables for training an artificial neural network (ANN). The results show that Random Forest achieved the best predictive performance (MSE = 785.31, R² = 0.9706, MAE = 21.39), followed by the F-test (MSE = 1784.99, R² = 0.9332, MAE = 30.66), Gradient Boosting (MSE = 2497.38, R² = 0.9066, MAE = 40.20), and Lasso CV (MSE = 2605.78, R² = 0.9035, MAE = 40.29). Analysis of the variable distributions revealed complex nonlinear relationships between geotechnical parameters and swelling pressure. This multi-method approach demonstrates the effectiveness of machine learning techniques for accurate and cost-effective prediction of geotechnical properties, offering a promising alternative to traditional, time-consuming, and costly oedometer tests.