Explainable Ensemble Machine Learning for the Prediction and Optimization of Pozzolanic Concrete Compressive Strength

Pozzolanic concrete demonstrates intricate, highly nonlinear material interactions that pose significant challenges for the accurate prediction of compressive strength (CS). This study introduces a novel, interpretable ensemble machine learning (ML) framework for predicting CS based on 759 mixture records encompassing cement, aggregates, supplementary cementitious materials (pozzolans), water/binder (W/B), superplasticizer, water, and curing age. Descriptive analysis and ANOVA were used to identify key predictors, followed by an 80/20 train-test split with 10-fold cross-validation to ensure robust and generalizable modeling. To further enhance model reliability, 5% of outliers were removed using an isolation forest algorithm, after which data were normalized and ensemble hyperparameters optimized. Among the evaluated models, the extra trees algorithm with standard scaling demonstrated the most stable generalization, achieving a coefficient of determination (R²) of 0.978 and a root mean square error (RMSE) of 4.197 MPa on the test set, and R² = 0.966 (RMSE = 5.053 MPa) under 10-fold cross-validation. Feature importance, SHAP, and partial dependence analyses consistently demonstrated that W/B, curing age, and cement are the principal determinants of CS. Finally, multi-objective optimization generated high-strength, low-impact mixtures, confirming the framework's effectiveness as a transparent decision-support tool for performance- and sustainability-oriented pozzolanic concrete design. This study is novel in combining interpretable ensemble ML with multi-objective optimization to simultaneously achieve precise CS prediction and the formulation of sustainable, performance-optimized pozzolanic concrete mixtures.