Predicting ultra high-performance concrete self-healing performance using hybrid models based on metaheuristic optimization techniques

Abstract

Ultra High-Performance Concrete (UHPC) has superior mechanical properties, including high compressive strength, tensile strain hardening behavior, and self-healing capacity. However, there has been limited focus on developing predictive models for UHPC's self-healing properties, despite extensive research in the aforesaid respect. While multi-physics modeling has made progress in predicting the coupled chemical, physical, and mechanical phenomena in cement-based materials, data-driven models, including Artificial Intelligence (AI) and Machine Learning (ML), are gaining popularity in predicting some concrete properties. In this study, a machine learning model was developed to predict UHPC's self-healing performance using three meta-heuristic algorithms, i.e., whales optimization algorithm (WOA), grey wolf optimization (GWO), and flower pollination algorithm (FPA), combined with extreme gradient boosting tree (Xgboost). The dataset used for the model was obtained from original experimental tests on UHPC's crack sealing performance under sustained through crack tensile stress and exposure to various aggressive environments for up to six months. The model's predictive performance was assessed using four mathematical indicators. The regression error characteristic (REC) and Taylor diagrams also showed the optimal models’ performance were found to be consistent and reliable across different optimization algorithms. SHapley Additive exPlanation (SHAP) results revealed that exposure time and crack width were most critical features for predicting self-healing performance. The study demonstrated the potential of using machine learning for predicting UHPC's self-healing performance and provided insights into the most critical factors affecting the process.