A practical machine learning-based approach for predicting 1-D vertical swelling potential of expansive soils

Several lightly loaded geotechnical and transportation infrastructures such as residential buildings, pipelines, roads, and railways have significant swelling potential challenges when they are placed on or within expansive soils. Reliable measurements of swelling potential of expansive soils are possible using conventional oedometer tests; however, their use in conventional practice is limited because they are time-consuming and costly. Several empirical equations have been proposed in the literature to alleviate these limitations; however, their applicability is limited for region-specific soils for which they have been developed. To overcome these limitations, in this study three machine learning-based prediction models were developed using a comprehensive global database of 173 expansive soils. The models, developed using Multivariate Adaptive Regression Splines and Multilayer Perceptron algorithms, show strong performance on the compiled dataset, with coefficients of determination (R²) of 0.887 or higher. Among them is a simplified model expressed as an explicit equation that requires clay fraction, dry density, plasticity index, specific gravity, vertical load, and water content information that performs well with an R² of 0.964. Most importantly, the model provides reasonable estimations of several case studies from various regions of the world. In summary, the model serves as a reliable tool for estimating the in-situ swelling potential of expansive soils. Finally, this study results are promising for proposing heave mitigation strategies and to develop rational design procedures and maintenance measures for lightly loaded geotechnical and transportation infrastructure.