Mineral-Phase-Informed Machine Learning for Interpreting and Regulating pH-Monitored Wet-Carbonation Responses of Carbonatable Binders

Abstract

Carbonatable binders offer a promising route for CO2 utilization and low-carbon cementitious materials, but their carbonation behavior is complicated by the asynchronous responses of coexisting mineral phases. This study developed a mineral-phase-informed machine-learning framework to interpret and regulate the time-resolved wet carbonation response of multimineral carbonatable binders. Using a dataset comprising 163 experimental samples, an artificial neural network was used to predict mineral-phase composition from synthesis parameters, while an XGBoost model predicted the pH-decrease-rate curve from mineral-phase composition and normalized reaction time. Model interpretation indicated that in γ-C2S-dominated binders, C3S2 was associated with an enhanced intermediate-stage response, whereas β-C2S was associated with a delayed and more sustained response. These patterns were qualitatively supported by pH and conductivity measurements of high-purity mineral phases. Inverse optimization further translated the model-inferred phase-response relationship into binder design. Compared with the nonoptimized sample, optimized binders showed higher compressive strength, increasing from 55.78 to 61.04–66.33 MPa, and slightly higher mean CO2 uptake, increasing from 156.36 to 161.81–164.27 kg/t. These results demonstrate that mineral-phase-informed machine learning can guide the design of carbonatable binders with tailored carbonation responses and improved engineering performance.