Efficient prediction of California bearing ratio in solid waste-cement-stabilized soil using improved hybrid extreme gradient boosting model

The California Bearing Ratio (CBR) is a critical parameter for evaluating the suitability of solid waste-cement stabilized soil (SW-CSS) in road construction projects. However, traditional CBR measurement methods are labor-intensive and often overlook various experimental factors, leading to reduced accuracy. Machine learning models offers a promising alternative. But current models for predicting CBR in soil typically rely on default hyperparameters, which results in less precise predictions. To address this limitation. The study integrates the basic chemical composition of solid waste into the model's inputs and proposes a hybrid model that enhances prediction accuracy, offering a more reliable and flexible solution. Firstly, a database containing 704 samples was established. Secondly, an extreme gradient boosting (XGB) model tuned with multi-strategy improved hiking optimization algorithm (MSHOA-XGB) was proposed. Thirdly, eight machine learning models, including five single models and three hybrid models, were developed, trained, and tested on this database. Their generalization performance was rigorously evaluated. Finally, feature variable impacts were examined using shapley additive explanations and partial dependence plots methods. The results indicated that the MSHOA-XGB model, offers the most accurate predictions of CBR among the models tested. Furthermore, XGB model was observed that the model performs robustly even without metaheuristic tuning, underscoring its practical applicability. The analysis identified fine grain and cement content as key variables influencing CBR, with calcium oxide content in solid waste exerting the most significant positive effect.