The production of Portland cement, as the main constituent of concrete, has considerable environmental and economic impacts. Partial replacement of the cement with pozzolanic materials such as fly ash (FA) and silica fume (SF) is an effective approach to reduce these effects and enhance the durability of concrete. However, the mechanical behavior of pozzolanic concretes is unpredictable and nonlinear due to complex chemical reactions and interdependent relationships among mix variables. This study aimed to develop and compare three white-box soft computing models, including Group Method of Data Handling (GMDH), Gene Expression Programming (GEP), and Response Surface Methodology (RSM), to predict the compressive strength of concrete containing FA and SF. In addition, sensitivity analysis was conducted to assess the influence of each input parameter on the model output. For this purpose, a dataset of 143 laboratory data points with varying mix compositions, including water (W), fly ash (FA), silica fume (SF), superplasticizer (HRWRA), fine aggregate (ssa), coarse aggregate (CA), sample age (AS), and total cementitious material (TCM), was used to model training and evaluation. The results indicated that the GMDH model outperformed the other approaches, showing very high accuracy and correlation. This model was able to predict the compressive strength of concrete with a correlation coefficient (R) of 0.918, a root mean square error (RMSE) of 10.57, a mean absolute error (MAE) of 8.20, and a mean absolute percentage error (MAPE) of 21.9%. The GEP model also provided satisfactory performance with R = 0.85, RMSE = 12.50, and MAPE = 25.8%, although its accuracy was lower than that of GMDH. Despite the relatively high correlation coefficient (R = 0.91), the RSM model had lower accuracy due to high prediction error (RMSE = 21.97). Sensitivity analysis revealed that W was identified as the most effective variable. After that, SF had the most positive role in improving the compressive strength and fine-grained structure of concrete. In contrast, the effects of FA, HRWRA, and TCM were minor and less variable. Overall, the GMDH model demonstrated not only high prediction accuracy but also an interpretable and computationally efficient structure, making it a robust tool for analyzing and forecasting the behavior of pozzolanic concretes. This study, for the first time, presents a comprehensive comparative evaluation of three white-box modeling approaches coupled with a detailed sensitivity analysis, offering valuable insights into the role of key mix parameters in developing sustainable concretes and reducing cement dependency.
Tavakkolian et al. (Tue,) studied this question.