Mineralogical Properties and Reactivity of Natural Pozzolans: A Machine Learning Perspective

ABSTRACT

This study integrates experimental testing, domain knowledge, and interpretable machine learning (ML) to improve the understanding of the reactivity of natural pozzolans (NPs). A total of 33 NPs of diverse origins and mineralogical compositions were characterized. Physically meaningful descriptors were derived from chemical and mineralogical composition to distinguish reactive from inert phases, obtaining reactive Al ₂ O ₃ and reactive SiO ₂ contents that more accurately represent the components participating in pozzolanic reactions. Reactivity was quantified through the R ³ test, where the heat release at 7 days was experimentally measured, and ultimate heat release ( H _ult ) as well as reaction rate ( r ) were determined by first‐order kinetic fitting. Bayesian ridge regression and random forest regression models were trained and validated using leave‐one‐group‐out cross‐validation (LOGO‐CV) to ensure robustness and prevent overfitting. Shapley additive explanation analysis revealed reactive Al ₂ O ₃ , fineness, and reactive SiO ₂ as the primary drivers of NP overall reactivity (heat release 7d, R ²  = 0.65), whereas amorphous content ( R ²  = −0.22) and bulk chemistry descriptors ( R ²  = −0.05) showed no explanatory capability. While the proposed descriptors enabled reliable correlation with the heat release at 7 days and H _ult , they showed no correlation with r , suggesting that they capture overall reactive potential rather than early‐age reaction kinetics. The proposed feature grouping strategy, prioritizing phase‐resolved reactive descriptors over bulk chemistry or total amorphous content, reduced spurious correlations and improved interpretability in the investigated dataset, which mainly consisted of volcaniclastic NPs. The results demonstrate the value of integrating domain knowledge with interpretable ML for physically informed descriptor development.