Explainable Machine Learning for Predicting Concrete Compressive Strength from Mixture Composition and Curing Age

Abstract

The prediction of the concrete compressive strength based on composition of concrete mixtures and curing age can be used to support the preliminary assessment of concrete materials, but it is important to carefully evaluate the accuracy of the model and its explanations. In this study the five regression models were compared with UCI Concrete Compressive Strength dataset. So, 1005 observations were left after the removal of 25 duplicate observations. The composition groups were set aside in such a way that they do not overlap between train, val and test sets. Gradient Boosting was chosen for explanatory analysis because it had the lowest grouped cross validation error. The test performance gave R² = 0.878, RMS error = 6.207 MPa, and MAE = 4.249 MPa. This value of 5.797 MPa was obtained for the Extra Trees model, showing that differences existed between the models in the different evaluation stages. Permutation importance determined curing age, cement, and water to be the top three predictors. The partial dependence revealed a linear increase in predictions as age increased and a downwards trend in prediction as water content increased. The results indicate preliminary strength estimates and the need for independent estimation and experimental verification of proposed mixture adjustments.