Prediction on the Resilience Modulus of Roadbeds Based on Coupled Bayesian with XGBoost Models

Accurately and efficiently estimating the rebound modulus of the roadbed is essential for ensuring the safety of the highway roadbed system and achieving reliable structural response design. To address this, the study utilized sample data from 12 different roadbeds, consisting of over 2800 sets of actual rebound modulus data, and employed Bayesian analysis for prediction. Initial considerations encompassed multiple parameters, such as mechanics, soil properties, and environmental factors, leading to the selection of six influencing factors including weighted plasticity index, dry density, shear stress, deviatoric stress, moisture content, and freeze-thaw cycles as input parameters. Subsequently, the dataset underwent preprocessing using the entropy weight method and data normalization techniques to acquire the weight coefficients of various influencing factors, providing a more accurate reflection of their impact on the rebound modulus. Based on this, the XGBoost algorithm was utilized for predicting the roadbed’s rebound modulus. Due to the numerous hyperparameters of the XGBoost algorithm, the study combined the Bayesian optimization (BO) algorithm to adjust these hyperparameters. Finally, a sensitivity analysis was conducted to explore the relative importance of each input variable. The results revealed that in comparison to XGBoost based on grid search, the BO-XGBoost model could accurately predict the subgrade rebound modulus within a shorter time frame, with a correlation coefficient R2 of 0.997 and an RMSE of 1.53. Additionally, when compared to artificial neural network models, support vector machine models, and traditional LGP and Kim empirical models, the BO-XGBoost model demonstrated superior performance.