Bound water evolution and machine-learning-assisted early hydration degree estimation of cement paste based on hyperspectral imaging

Abstract

The early hydration degree of cementitious materials affects their initial properties. This makes accurate hydration degree estimation necessary for a more precise reference for initial applications, particularly for various slag-blended cementitious materials. Hydration degree is closely related to bound water evolution. Accordingly, this study used the D-dry ignition method and near-infrared hyperspectral imaging (NIR-HSI) to establish the relationship between bound water evolution and spectral reflectance at different stages. The reflectance values in the 900–1700 nm wavelength generally increased as water was removed. By contrast, the 1390–1430 nm range exhibited a logarithmic decrease over time owing to the increase in hydration products. A linear relationship was observed between the bound water content and average reflectance within this wavelength range. In addition, an exponential relationship was found between the hydration rate and average reflectance difference. Several machine learning models, including the random forest (RF), backpropagation neural network (BPNN)–rectified linear unit (ReLU), gradient boosting decision tree (GBDT)–Huber, and radial basis function (RBF)–support vector machine (SVM), were used to estimate the hydration degree of the cement paste. The RF model demonstrated superior estimation accuracy and the most robust performance among all models. This study provides a new perspective for precisely estimating the hydration degree of cementitious materials within the context of data-driven science.