A hybrid stacking ensemble framework with domain-informed feature engineering for accurate prediction of concrete compressive strength

Accurate prediction of concrete compressive strength is essential for structural safety, mixture optimization, and efficient construction planning. Traditional experimental approaches are time-consuming, costly, and require extensive laboratory experiments. With the increasing availability of material datasets, machine learning techniques have emerged as effective alternatives for predicting concrete properties. However, many existing models rely on single learning algorithms and often fail to capture the complex nonlinear interactions between concrete mixture components. This study proposes a hybrid machine learning framework for predicting concrete compressive strength by integrating domain-informed feature engineering, nonlinear interaction modeling, automated hyperparameter optimization, and stacking-based ensemble learning. Physically meaningful features such as water-to-binder ratio, water-to-cement ratio, binder content, and aggregate ratio are first constructed to incorporate engineering knowledge into the modeling process. Nonlinear relationships between mixture variables are further captured using polynomial interaction expansion. Three optimized gradient boosting algorithms, namely XGBoost, LightGBM, and CatBoost, are employed as base learners and their predictions are combined using an ElasticNet meta-learner within a stacking ensemble architecture. Hyperparameter tuning for each model is performed using the Optuna optimization framework. The proposed model is evaluated on a dataset containing 1030 concrete mixture samples with eight input variables. Experimental results demonstrate that the stacking ensemble achieves strong predictive performance with a coefficient of determination ( $$R^2$$ ) of 0.936, a mean absolute error (MAE) of 2.695 MPa, and a root mean squared error (RMSE) of 4.071 MPa. These results indicate that the proposed framework effectively captures the complex relationships governing concrete strength development. Beyond academic contributions, the proposed approach provides practical benefits for the construction industry by enabling rapid evaluation of concrete mixture designs, reducing reliance on costly laboratory testing, and supporting data-driven decision-making in material optimization and quality control. The findings highlight the potential of advanced ensemble learning techniques for improving predictive modeling in civil engineering materials research.