Revealing the nature of Ultra-High-Performance concrete using computational intelligence

Abstract

Ultra-High-Performance Concrete (UHPC) is an advanced cementitious material with exceptional strength and durability, widely used in high-performance structural applications. However, predicting its compressive strength remains a challenge due to the complex nonlinear interactions among its mix constituents. This study employs machine learning (ML) models—Adaptive Boosting (AdaBoost), Gradient Boosting Machine (GBM), and Extreme Gradient Boosting (XGBoost)—to develop predictive models for UHPC compressive strength. To further enhance predictive accuracy, four advanced metaheuristic optimization algorithms were integrated —Grey Wolf Optimizer (GWO), Whale Optimization Algorithm (WOA), Giant Trevally Optimizer (GTO), and Mountain Gazelle Optimizer (MGO)—to fine-tune hyperparameters of the ML models. A dataset of 810 UHPC mix samples with 15 input variables was used to train and evaluate the models. Among the tested approaches, the MGO-optimized XGBoost (MGO-XGB) model achieved the highest accuracy, with an R² of 0.9966 in training and 0.9839 in testing, along with the lowest root mean square error (RMSE) and mean absolute error (MAE). The results demonstrate that integrating metaheuristic optimization significantly improves ML model performance, with MGO-XGB emerging as the best predictor. Additionally, SHAP analysis identified key influencing factors, including silica fume content, superplasticizer dosage, and curing age, which play a critical role in UHPC strength development. The findings indicate that ML-assisted optimization can reduce reliance on extensive experimental testing, offering a cost-effective and efficient approach for UHPC mix design and quality control. To enhance practical usability, a graphical user interface (GUI) was developed, allowing engineers and researchers to input mix parameters and obtain immediate strength predictions. This study contributes to data-driven advancements in concrete technology, enabling more efficient and sustainable UHPC design and construction. • Revealing of material’s nature. • Hybrid ML models and metaheuristic optimization were used for UHPC strength prediction. • XGBoost optimized with MGO demonstrated the best predictive performance. • SHAP analysis identified key mix components influencing UHPC strength. • A user-friendly GUI was developed for real-time strength estimation.

Topics & Concepts