Predictive modeling for durability characteristics of blended cement concrete utilizing machine learning algorithms

Abstract

Chloride penetration and carbonation resistance are critical durability attributes that assess concrete's ability to withstand challenging environmental conditions. However, determining these parameters requires time-consuming and resource-intensive physical experiments. Accordingly, this study employed gene expression programming (GEP) and multi-expression programming (MEP) to develop a robust model for predicting these parameters, providing mathematical equations for their estimation. Additionally, the study to develop a graphical user interface that would allow for predictions based solely on input values, thereby eliminating the need for extensive physical testing. To thoroughly assess the effectiveness of the proposed GEP and MEP models, a range of statistical metrics were employed, including the coefficient of determination (R²), adjusted R², root mean square error (RMSE), mean absolute error (MAE), and root mean square error to observation’s standard deviation ratio (RSR), along with engineering indices like the a10-index and a20-index. Both GEP and MEP models consistently demonstrated outstanding performance across all statistical indicators for both carbonation rate and chloride penetration. The GEP model showed high precision in modeling chloride penetration with an R² of 0.954, MAE of 0.252, and RMSE of 1.050, and for carbonation rate with an R² of 0.99, MAE of 0.230, and RMSE of 1.100. Similarly, the MEP model performed well, achieving an R² of 0.913, MAE of 0.489, and RMSE of 1.434 for chloride penetration, and an R² of 0.985, MAE of 0.560, and RMSE of 1.440 for carbonation rate. In addition, the SHapley Additive exPlanation (SHAP) method was employed to comprehend the model estimations. In predicting chloride penetration, cement to water ratio (C/B) emerged as the most impactful feature, followed by silica fume to binder ratio (SF/B) and water to binder ratio (W/B) in terms of importance. For carbonation rate, W/B stood out as the most influential, with C/B and fly ash to binder ratio (FA/B) being the subsequent key factors. These intuitions are further supported by partial dependence plots (PDPs). Furthermore, the SHAP summary plots distinctly reveal the relationships between the various parameters and the estimated characteristics.

Keywords

• Blended cement concrete
• Machine learning
• Gene expression programming
• Multiple expression programming