Explainable ensemble machine learning for reliable concrete compressive strength prediction: a validation and uncertainty assessment framework

Obtaining a highly accurate estimation of concrete compressive strength is essential for safe and economical structural design. However, the highly nonlinear relationship between mixture constituents and strength development limits the capability of traditional regression and empirical models. This study develops machine learning (ML)-based predictive models using easily available mix design parameters: water, cement, sand, coarse aggregate, and curing age to estimate concrete compressive strength. Six ML models, including Random Forest (RF), XGBoost, CatBoost, LightGBM, Gradient Boosting, and stacking ensemble, were evaluated using an 80/20 train–test split and 10-fold cross-validation. The contribution of this study lies in establishing an engineering-oriented validation framework that integrates ensemble learning, independent external validation, uncertainty quantification, explainable analysis, and practical deployment using commonly available concrete mix parameters. The results revealed that the stacking ensemble achieved the best performance with R² = 0.9245, MAE = 2.8561, and RMSE = 4.3892, while Gradient Boosting showed the highest cross-validation stability (R² = 0.9244). External validation confirmed reasonable model generalization, although performance decreased due to differences in material characteristics and dataset distribution. These findings demonstrate that ensemble ML models can effectively capture nonlinear strength behavior and provide reliable predictions under similar concrete conditions.