Prediction and variable importance analysis for small-strain stiffness of soil based on ensemble learning with Bayesian optimization

Abstract

Small-strain stiffness is a fundamental soil property described by the normalized shear modulus (G/Gmax) and maximum shear modulus (Gmax). Though numerous experiments and empirical models were developed to predict the small-strain stiffness based on soil status and properties, accuracy was still a challenging issue to achieve due to the various soil types and complex conditions. This study developed an ensemble machine learning approach using Random Forest and XGBOOST models to improve prediction accuracy. The experimental equipment and database development process were described. The structure of the XGBOOST model using a Bayesian Optimization was proposed to optimize hyperparameters. The results of the comparison indicated that the ensemble machine learning approach had a higher accuracy of prediction than the empirical models. The variable importance analysis identified water content as one of the key factors influencing the prediction of G/Gmax and Gmax. For the prediction of G/Gmax, the void ratio might not be the primary soil property influencing the prediction for clay and sandy soils. Finally, a design procedure based on the ensemble machine learning approach with regression analysis using theoretical formulas was proposed to estimate the small-strain stiffness of soil for the convenience of design engineers.