Quantification of GGBS hydration using deep learning - A comparison with SEM-EDS mapping, PONKCS XRD and isothermal calorimetry methods

Quantifying degree of hydration (DoH) of ground granulated blast-furnace slag in blended cement paste has long been a challenge due to its amorphous nature. In this study, Convolutional Neural Networks (CNNs) were trained on over 37,000 image pairs to automatically segment GGBS particle from backscattered electron (BSE) images and estimate its DoH. The results were compared with three conventional methods: EDS mapping, PONKCS XRD, and isothermal calorimetry. With minimal human involvement, the trained models consistently estimated GGBS DoH across varying reaction times, image qualities, cement types, substitution rates, and GGBS sources. EDS mapping, though capable with appropriate normalization, was time-consuming and sensitive to element distribution of raw GGBS and hydration products. PONKCS XRD enabled quantification of amorphous GGBS, aligning well with image analysis. Isothermal calorimetry can also track GGBS DoH if calibrated, but is less suitable for long-term reaction (DoH > 50 %) due to weak signals and extensive equipment demand.