Hybrid machine learning framework for predicting pozzolanic reactivity: Integration of R3 test and GAN-augmented data

Abstract

The substitution of Portland cement with supplementary cementitious materials (SCMs) is essential for lowing CO₂ emission in the construction industry. However, the accurate assessment of pozzolanic reactivity is challenging due to time-consuming traditional tests and intrinsic material variability. To address these issues, this study developed an integrated machine learning model to predict pozzolanic reactivity using raw material properties, leveraging the internationally validated R³ test. To overcome data limitations, a conditional generative adversarial network (CTGAN) generated 430 synthetic data points for augmentation. The trained regression model achieved a high predictive accuracy (R² > 0.94) and was further validated against 43 original data points (R² = 0.85), confirming robustness. SHAP analysis identified amorphous content and Dv50-related features as key predictors, further improving model interpretability and reliability. This study provides a reliable regression model for R³ hydration heat estimation, highlighting its potential to streamline SCM assessment and enhance construction materials quality control.