Hybrid machine learning model to predict the mechanical properties of ultra-high-performance concrete (UHPC) with experimental validation

Abstract

Ultra-high-performance concrete (UHPC) incorporating waste cementitious materials has become widely used due to its extraordinary mechanical strength and durability. Adding such waste also addresses the environmental sustainability aspect of the materials, making them a potential alternative. This study explores using Random Forest (RF) and XGBoost (XGB) as the primary model. Further, metaheuristic algorithms like the Pelican optimization algorithm (POA) and Walrus optimization algorithm (WOA) should be used to tune the hyperparameters of the primary model. This study shows that the XGB-POA is highly accurate, exceeding R² of 0.96 in the testing set. Additionally, ten-fold cross-validation ensures the model’s robustness by mitigating the overfitting issues. Similarly, other employed models, like XGB-WOA, RF-POA, and RF-WOA, also exhibited better training and testing set results. Moreover, this study is subjected to Shapley’s Additive Explanation (SHAP) analysis to explore the model’s explainable behaviour. The study reveals that the XGB-POA is the best-performing model, identifying age, fiber content, cement, and SF dosage as the most influential features in the development of UHPC. Experimental data sets that showcase more than 95% accuracy are used to validate the model performance. These insights help to understand the relationships of features involved with comprehensive assessments of UHPC for adopting sustainable practices.