Machine learning models to estimate stress wave velocities of cohesionless soils during triaxial compression influenced by particle characteristics

Abstract

Stress wave velocities, i.e. shear and compression wave velocities and hence small strain stiffness, are important parameters required for dynamic analysis and designing various geotechnical structures. Measuring wave velocities has been challenging, let alone developing a model for their prediction. There exist few models in the literature; however, their use is limited and can be used mostly in specific cases, for example, isotropic conditions. This paper uses the dataset developed in the author's previous study and aims to develop a model where the wave velocities can be predicted accurately. The dataset contains wave velocities evolution of three different materials: Toyoura sand, river sand and silica sand at three different initial relative densities sheared by increasing the major principal stress. The three different materials allowed the study of different particle characteristics, including sphericity, convexity, aspect ratio and roughness, in predicting and modelling the wave velocities. This paper utilizes three different machine learning (ML) models viz. artificial neural network (ANN), support vector regression (SVR) and gradient boosting regression (GBR) to model the observed behaviour of wave velocities evolution. All three models were equally efficient in modelling the evolution of both shear (Vs) and compression wave (Vp) velocities. The predicted Vs and Vp using the three different machine learning models are also compared with a literature model available for isotropic stress conditions and showed that the ML models developed in this study are superior by incorporating the effect of grain characteristics which is missing from the literature model. The relationship between the prediction of the wave velocities and input parameters is highlighted using feature importance and Pearson's coefficient. At the end of this study, ANN is used to provide a set of nonlinear equations in the matrix form for Vs and Vp, which can be used by anyone and does not require any prior knowledge of ML or coding.