Predicting uniaxial tensile strength of expansive soil with ensemble learning methods

Tensile stress is a major parameter controlling the desiccation cracking, which may incur the tensile failure of expansive soil slopes. In this study, a series of experiments was performed and a dataset of 125 records was initially recorded with 6 features viz., dry density (DD), water content (W), matric suction (MS), unconfined compressive strength (UCS), failure compressive strain (FCS) and the failure tensile strain (FTS) and 1 label, uniaxial tensile strength (UTS). Three classical ML methods and three ensemble methods viz., random forests, extreme gradient boosting, and the stacked generalization, were utilized to capture the relationships between the label and the features. The experimental results show that the tensile strength of the expansive soils is subject to the coupling effects of the UCS, the W and the FTS. And their relative importance to the uniaxial tensile strength can be discerned by the top performing models as UCS (40.18%) > W (36.46%) > FTS (23.36%). The stacked generalization can effectively integrate the merits of the other 5 ML methods, and, hence, provides the most accurate prediction of UTS for the given dataset.