Prediction of the Swelling Properties of Expansive Soils Using Machine Learning Methods

Abstract

Expansive soils swell significantly when they absorb water and exhibit high swelling pressures if their volume changes are restrained. This characteristic behavior of all expansive soils is mainly attributed to the presence of hydrophilic clay minerals such as montmorillonite and illite. Geotechnical engineers worldwide face considerable challenges in providing reliable approaches for construction and design of geo-infrastructures constructed with or within expansive soils. Common engineering problems caused by these soils include cracked and uneven pavements, uplifted and tilted foundations, slope failures, and the sliding of retaining walls. Over the past several decades, annual global economic losses associated with the damage and repair costs of geotechnical infrastructures built within or with expansive soils have been estimated at several billion US dollars. To better understand and address these challenges, researchers have developed various experimental methods over the past five decades to determine key swelling properties of expansive soils that include: swelling and compression indices, swelling pressure, and swelling potential. These laboratory methods, particularly oedometer tests, are reliable and reproducible. However, they require extensive equipment and trained personnel, and the testing process often takes several weeks to complete, making them both time-consuming and expensive. As a result, they are not always practical for use in routine geotechnical investigations. Although several empirical and semi-empirical equations have been proposed to predict the swelling behavior of expansive soils, their applications are often limited due to regional differences in expansive soil characteristics and a lack of general applicability. Therefore, there is a need for more simple, economical, reproducible and reliable prediction methods that can be used universally for different types of expansive soils based on basic soil properties that can be obtained from conventional laboratory tests. In this study, to achieve this key objective, simple models / equations are developed for key geotechnical properties related to the swelling and compressibility behavior of expansive soils with the aid of advanced Machine Learning (ML) techniques. Comprehensive databases were compiled from published literature, covering a wide range of soil types and geological conditions around the world. Three comprehensive ML algorithms—Multilayer Perceptron (MLP), Support Vector Machine (SVM), and Extreme Learning Machine (ELM)—were applied, along with regression analysis, to predict the swelling index and swelling pressure based on fundamental soil properties; namely, liquid limit, plasticity index, dry density, and initial suction. The proposed models demonstrated reliable predictive capabilities across diverse datasets. Additionally, Multivariate Adaptive Regression Splines (MARS) and MLP algorithms were employed to develop prediction models for estimating the swelling potential of expansive soils. An extensive dataset was used to support model training and validation. Sensitivity analysis was applied to refine the input variables, ensuring an efficient and accurate model structure. Simplified equations were derived from the optimized models to facilitate their application in practice. The validation of these models was conducted through comprehensive comparative analyses, including evaluations between different ML models and against traditional empirical equations. The results demonstrated significant improvements in prediction accuracy and generalization, offering a practical approach for assessing swelling potential. Furthermore, in this study ML-based models are presented for predicting the compression index of expansive soils, a critical parameter for evaluating compressibility and settlement behavior in foundation design. A database of 238 expansive soil samples gathered from different countries of the world was established, representing diverse geotechnical conditions. Sensitivity and correlation analyses identified liquid limit, plasticity index, initial void ratio, and dry density as the most influential input factors. Two primary ML algorithms, Newton-Raphson-Based Optimizer Enhanced Extreme Gradient Boosting (NRBO-XGBoost) and MLP, were used to develop reliable prediction models. While the NRBO-XGBoost model exhibited superior performance in terms of accuracy and generalization, the MLP model was further simplified into an empirical equation, making it more convenient for preliminary design applications. To further evaluate the accuracy and reliability of the proposed models and equations, extensive validation studies were carried out. Published case studies results from the literature were used to assess the prediction accuracy of the models for four key expansive soil parameters: swelling index, swelling pressure, swelling potential, and compression index. The results of these analyses showed strong agreement between the predicted and measured values across different soil types and geographical regions. These case studies confirmed the applicability of the models for capturing the behavior of expansive soils under varying conditions. In addition, this study integrates the proposed models to jointly simulate the volume change behavior of expansive soils under constant-volume and consolidation-swell conditions. By applying the swelling index, swelling pressure, swelling potential, and compression index models together, the overall volume change behavior was predicted and validated using laboratory data from multiple expansive soils. The results demonstrate the consistency and compatibility of the models when applied in combination, reinforcing their practical value in engineering practice applications. In summary, this study presents a valuable framework that integrates conventional soil mechanics principles with advanced ML algorithms to improve the prediction of key volumetric parameters of expansive soils. The proposed models and simplified equations enable reliable estimation of swelling and compressibility characteristics based on basic soil properties. Their consistency and compatibility, demonstrated through extensive validations on extensive laboratory test results and case studies, highlight their practical value in conventional design and geotechnical engineering applications across various regions around the world affected by expansive soils.