Machine Learning Prediction of Concrete Compressive Strength: Model Comparison, CatBoost Optimization, and SHAP Interpretation
TL;DR
A comparative framework evaluating nine regression algorithms using the UCI Concrete Compressive Strength dataset, jointly integrating correlation-corrected statistical validation, multi-model Bayesian optimization, and domain-informed feature engineering with SHAP interpretation, rarely combined in prior concrete-strength studies.
Abstract
Accurate prediction of concrete compressive strength is vital for structural design, yet conventional testing is constrained by lengthy curing requirements. Machine learning offers an alternative by modeling non-linear mix-performance interactions. This study presents a comparative framework evaluating nine regression algorithms using the UCI Concrete Compressive Strength dataset (1,005 samples). Performance was assessed via 10x5 repeated cross-validation with 95% confidence intervals, and statistical significance was evaluated using a Linear Mixed-Effects Model with Holm-Bonferroni corrected pairwise t-tests. Tree-based ensembles outperformed linear approaches, with CatBoost yielding the highest baseline cross-validation R² of 0.931 (95% CI: 0.927 to 0.935). Subsequent Bayesian hyperparameter optimization via Optuna’s Tree-structured Parzen Estimator (400 trials) improved the final CatBoost model’s performance to a test of 0.943, RMSE of 4.142 MPa, and MAE of 2.616 MPa. SHAP analysis indicated that curing age, the water-to-binder ratio, and cement are the dominant predictors, while the model exhibited physically consistent behavior aligned with concrete hydration kinetics and Abrams' law. This work's key contribution is jointly integrating correlation-corrected statistical validation, multi-model Bayesian optimization, and domain-informed feature engineering with SHAP interpretation, rarely combined in prior concrete-strength studies. The framework offers an accurate, interpretable tool for preliminary concrete mix design.