Predicting the Shear Strength of Granular Waste Materials Using Machine Learning

Knowing the shear strength of soil is imperative for geotechnical design as shear failure, combined with excessive deformations, is the predominant failure mechanism within a loading environment. However, determining the shear strength in the laboratory is often laborious and hence costly. Moreover, a granular material’s behavior is complex which can compromise the accuracy and robustness of predictive models developed through traditional methods. This is exacerbated when considering non-traditional waste materials such as steel furnace slag, coal wash, and scrap rubber due to their increased nonlinearity and variability. Consequently, previous relationships and models proposed are often self-contained and break down when extrapolated beyond specific loading conditions or material types. In this study, predictive models for the peak friction angle (ϕ^′_peak) of various granular mixtures (waste and non-waste) were developed using two nonlinear machine learning (ML) techniques, namely, artificial neural network (ANN) and second-order multivariable regression (MR). Five key parameters were chosen to represent the mixture type (rubber content, median particle size), its physical properties (initial void ratio, dry unit weight), and the loading condition (effective confining pressure) using 154 consolidated drained triaxial test data samples. Although MR performed satisfactorily on both the original and secondary datasets, ANN combined with Bayesian regularisation was superior with R² of 0.96 and 0.82 for both phases, respectively. Hence, ANN is an attractive modelling technique as it is capable of capturing nonlinear relationships for various granular mixtures (i.e., waste and non-waste, with and without rubber) to predict shear strength without the need for laboratory testing.