Predicting the impact of adding metakaolin on the splitting strength of concrete using ensemble ML classification and symbolic regression techniques -A comparative study Provisionally Accepted

Cesar Garcia^1* Alexis Ivan Andrade Valle¹ Angel Alberto Silva Conde¹ Nestor Ulloa² Alireza Bahrami^3* Kennedy C. Onyelowe^{4, 5*} Ahmed M. Ebid⁶ Shadi Hanandeh⁷

• ¹National University of Chimborazo, Ecuador
• ²Escuela Superior Politécnica del Chimborazo, Ecuador
• ³Department of Building Engineering, Energy Systems and Sustainability Science, Faculty of Engineering and Sustainable Development, University of Gavle, Sweden
• ⁴Michael Okpara University of Agriculture, Nigeria
• ⁵University of Peloponnese, Greece
• ⁶Future University in Egypt, Egypt
• ⁷Al-Balqa Applied University, Jordan

The mechanical characteristics of concrete are crucial factors in structural design standards especially in concrete technology. Employing reliable prediction models for concrete's mechanical properties can reduce the number of necessary laboratory trials, checks and experiments to obtain valuable representative design data, thus saving both time and resources. Metakaolin (MK) is commonly utilized as a supplementary replacement for Portland cement in sustainable concrete production due to its technical and environmental benefits towards net-zero goals of the United Nations Sustainable Development Goals (UNSDGs). In this research work, 204 data entries from concrete mixes produced with the addition of metakaolin (MK) were collected and analyzed using eight (8) ensemble machine learning tools and one (1) symbolic regression technique. The application of multiple machine learning protocols such as the ensemble group and the symbolic regression techniques have not been presented in any previous research work on the modeling of splitting tensile strength of MK mixed concrete.The data was partitioned and applied according to standard conditions. Lastly, some selected performance evaluation indices were used to test the models' accuracy in predicting the splitting strength (Fsp) of the studied MK-mixed concrete. At the end, results show that the k-nearest neighbor (KNN)outperformed the other techniques in the ensemble group with the following indices; SSE of 4% and 1%, MAE of 0.1 and 0.2 MPa, MSE of 0, RMSE of 0.1 and 0.2 MPa, Error of 0.04 and 0.04%, Accuracy of 0.96 and 0.96 and R2 of 0.98 and 0.98 for the training and validation models, respectively. This is followed closely by the support vector machine (SVM) with the following indices; SSE of 7 and 3%, MAE of 0.2 and 0.2 MPa, MSE of 0.0 and 0.1 MPa, RMSE of 0.2 and 0.3 MPa, Error of 0.05 and 0.06%, Accuracy of 0.95 and 0.94, and R2 of 0.96 and 0.95, for the training and validation models, respectively.