Fresh state and strength performance evaluation of slag-based alkali-activated concrete using soft-computing methods

In this study, machine learning prediction models for the slump (SL) and compressive strength (CS) of alkaliactivated concrete (AAC) were developed. Extreme gradient boosting (XGB) as an ensemble and support vector machine (SVM) as individual methods were chosen. To evaluate the performance of the models, the Taylor diagram, k-fold validation, and statistical tests were performed. Moreover, to determine the significance of features, a SHapley Additive exPlanations (SHAP) study was carried out. XGB outperformed SVM considerably in predicting the SL and CS of AAC. XGB outperformed SVM in terms of R2 (0.94 for SL and 0.97 for CS), which was 0.86 and 0.88, respectively. Precursor content had the greatest effect on the SL of AAC, followed by blast furnace slag ratio, test time, SiO2/Na2O, and quantities of NaOH, aggregate, and water, according to the results of the SHAP study. The SHAP investigation revealed that curing time had the greatest effect on the CS of AAC, followed by SiO2/Na2O, NaOH quantity, precursor content, aggregate quantity, blast furnace slag ratio, and water quantity. It was determined that curing time, SiO2/Na2O, and blast furnace slag had beneficial effects; precursor content and aggregate quantity had adverse effects, while water had both beneficial and deleterious effects.