Predictive Modelling of Strength of Self-Compacting Concrete using different Machine Learning techniques

Abstract

This study presents predictive modelling of 28-day compressive strength of self-compacting concrete (SCC) in a comparative manner using different machine learning techniques. The models were constructed using three algorithms: Gene Expression Programming (GEP), Multi Expression Programming (MEP), and Extreme Gradient Boosting (XGB). A total of 231 data sets obtained from internationally published literature were used for model development. The dataset comprised seven input parameters measured in kilograms per cubic meter, including cement content, fly ash and silica fume quantities, coarse and fine aggregate content, water content, and superplasticizer dosage. To assess the accuracy and predictive capabilities of the models, various error metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), coefficient of correlation (R), coefficient of determination (𝑅2), as well as performance indicators like performance index (𝜌) and objective function (OF) were employed. Statistical analysis demonstrated that all models performed well and exhibited strong generalization capabilities. While the MEP and GEP models provided empirical equations as output, the XGB model did not yield an equation. The accuracy of the models was further evaluated through external validation criteria and comparison with statistical regression models. The comparative analysis of the developed models revealed that the XGB model outperformed both the MEP and GEP models, exhibiting the highest level of performance. Subsequently, additional analysis was conducted on the XGB model, including sensitivity and parametric analyses, which offered further insights into the relationships between input and output variables. Furthermore, a user-friendly interface was developed to facilitate easy prediction using all three models.