Explainable machine learning for predicting the splitting tensile strength of SCM-blended concrete using optimized Gaussian process regression

This study develops an explainable machine-learning framework to predict the splitting tensile strength of concrete incorporating supplementary cementitious materials (SCMs), which are widely used to reduce reliance on Portland cement. The analysis used 395 valid mixtures extracted from a global SCM-blended concrete database. Cement, water, coarse aggregate, fine aggregate, fly ash, silica fume, ground granulated blast-furnace slag, and superplasticizer were used as predictors. Predictor relevance was assessed using mRMR, F-test, RReliefF, and Random Forest importance. A broad set of regression algorithms was evaluated in MATLAB Regression Learner using 5-fold cross-validation. Rational Quadratic Gaussian Process Regression (GPR) was the best baseline model, with an RMSE of 0.51 MPa, MAE of 0.34 MPa, MAPE of 10.02%, and R² of 0.85. The GPR model was then optimized using Bayesian optimization, random search, and grid search. All three strategies produced comparable predictive accuracy. Bayesian optimization required the shortest training time and was therefore selected for the final model. The Bayesian-optimized GPR achieved an RMSE of 0.41898 MPa, MSE of 0.17554 MPa², MAE of 0.29095 MPa, MAPE of 8.9%, and R² of 0.89. Relative to the baseline model, RMSE and MSE decreased by 17.85% and 32.48%, respectively. Partial dependence plots and SHAP analysis identified water and cement as the dominant predictors. Higher water content generally reduced predicted tensile strength, whereas higher cement content increased it. The findings demonstrate that optimized and explainable GPR provides a reliable data-driven framework for predicting the tensile performance of SCM-blended concrete and supporting performance-informed design of more sustainable cementitious materials.