A case study of resilient modulus prediction leveraging an explainable metaheuristic-based XGBoost

The resilient modulus (M_R) of pavement subgrade soils is an index describing the structural response of flexible pavement foundations. Commonly, M_R under different conditions of confining pressures and deviatoric stresses are tested by cyclic triaxial compressive experiments. However, such experiments are elaborate, expensive, and time-consuming, so developing more flexible and efficient approaches is imperative. This study investigates the potential application of a tree-based model termed extreme gradient boosting (XGBoost) on predicting M_R. First, a dataset containing 891 samples of repeated load triaxial tests and the characteristics of subgrade soil is collected. Then, an XGBoost model, combined with a feature selection technique (Exhaustive Feature Selector (EFS)) and an optimization algorithm (Jellyfish Swarm Optimizer (JSO)), is trained on the collected dataset. EFS is used to identify the most suitable combinations of factors for M_R prediction and JSO is applied to determine the hyper-parameters of the XGBoost model, which aim to establish a robust XGBoost model with the best predictive capacity. Lastly, this study employs two advanced model interpretation techniques to identify the predominant factors affecting M_R prediction based on the established XGBoost model. The results indicated that the EFS approach can effectively ascertain the best combination of factors for M_R prediction; the JSO algorithm can effectively capture the optimal hyper-parameters of the XGBoost model; the resultant XGBoost model achieved a favorable capacity for M_R prediction. Moreover, three primary factors affecting M_R prediction are unveiled, which are the degree of soil saturation (S_r), confining stress (σ₃), and plasticity index (PI).