A novel prediction model for the resilient modulus of unbound aggregate bases using eXtreme Gradient Boosting (XGBoost) and Bayesian optimisation (BO) is presented in this study. First, a local dataset containing 260 samples of some common tests, including sieve analysis, Atterberg limits tests, compaction tests, and repeated load triaxial tests of unbound aggregate materials, is collected. Then, the proposed BO-XGBoost model is trained on the collected dataset. Multiple evaluation metrics, including Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and coefficient of determination (R2), were employed to evaluate the performance of the BO-XGBoost model. The model's accuracy and generalisation capabilities were evaluated by computing these metrics for both the training and testing datasets. The findings of this study, clearly showed that the combined BO-XGBoost model outperformed other state-of-the-art machine learning or even deep learning models, in predicting the resilient modulus of unbound aggregate base. Finally, the permutation method was used to conduct a feature importance analysis and determine the impact of different input features on the resilient modulus of the unbound aggregate base. This latter was highly influenced by stress, moisture content, and liquidity limit, as demonstrated by the results of the feature importance analysis.

Resilient modulus, unbound aggregates, XGBoost, Bayesian optimisation