Analyzing the influence of particle size distribution on the maximum shear modulus of soil with an interpretable machine learning framework and laboratory test dataset

Abstract

The maximum shear modulus (G_max) is a key parameter used to characterize the dynamic properties of soils. In this research, a dataset was systematically collected and constructed through literature review. It comprises 2782 instances of G_max values and their influencing factors for various soil types, aimed at examining the effect of particle size distribution on the G_max. The eXtreme Gradient Boosting (XGBoost) algorithm was employed to develop the predictive model for G_max, followed by the enhancement of model's performance through Bayesian Optimization (BO) algorithm. After comparison with other empirical models, the BO-XGBoost model was selected as the best model. Finally, the prediction of BO-XGBoost was interpreted using the SHapley Additive exPlanations (SHAP) framework in order to overcome the black box problem of traditional machine learning methods. The results suggest that SHAP effectively extracts critical information from the data when data labels are appropriately configured, thereby augmenting the reliability of the prediction outcomes. Globally, the feature importance ranking and the direction of correlations between input features and the output variable align with the prior knowledge. Locally, however, the importance ranking of features for individual samples may deviate from the global trend. Meanwhile, the influence of identical input features can vary across different samples.