To reduce the carbon footprint of the Portland cement (PC) industry, the prevailing practice is to partially replace the PC in concrete with supplementary cementitious materials (SCMs). Each SCM, owing to its distinctive chemical composition and molecular structure, affects hydration kinetics and microstructural evolution of cementitious binders uniquely. Current computational models cannot produce reliable predictions of hydration kinetics of complex [PC + SCM] binders. In the past two decades, the combination of Big data and machine learning (ML) has emerged as a promising tool to produce predictions for material properties. This study employs ML models to produce predictions of hydration kinetics of PC replaced by various SCMs at different replacement levels. However, ML cannot produce highly reliable predictions of hydration kinetics of [PC + SCM] binders because it is hard for ML to completely learn highly nonlinear correlations from a small database. To enhance prediction accuracy, we introduce two methods – Fourier transformation and phase boundary nucleation model – to reduce the degree of complexity of the database, which allows ML to produce highly reliable predictions. Furthermore, thermodynamic constraints derived from thermodynamic criteria are applied to inform and guide the ML model.