Data-driven Modeling and Mixture Optimization of Self-Healing Engineered Cementitious Composites

In this study, a machine learning model is developed using an exper- imental dataset comprising 617 measurements related to the self-healing behavior of concrete specimens. The dataset includes key materials with self-healing properties that served as features, Fly Ash, Silica Fume, Lime Powder, Crack Width Before and Crack Width After, obtained from ex- perimental mixing and casting of various concrete mixtures prepared with a consistent mixing rate. Extreme Gradient Boosting was employed for predictive modeling. Data normalization was systematically applied dur- ing preprocessing. To evaluate the performance and stability of the models, a compre- hensive analysis of the learning curves and outliers was conducted. To assess the models’ ability to capture the underlying physics of materi- als, the monotonic dependency and SHAP values were examined. To determine feature importance, we integrated information with XGBoost theory. To optimize prediction accuracy across diverse mixture composi- tions, a hybrid approach integrating machine learning with metaheuristic optimization, specifically Differential Evolution (DE), was implemented. This strategy effectively reduces reliance on costly physical experiments by enabling comprehensive exploration of the complex, high-dimensional mixture-design space and generating optimized mixture composition rec- ommendations. All code and datasets are made available by the authors to facilitate reproducibility and further research.