Accurately predicting the degree of hydration in cement systems blended with slag and fly ash is difficult when experimental data are limited. Empirical kinetic models typically need recalibration for each specific blend, while purely data-driven approaches often fail to maintain physical consistency. In this work, we present a physics-informed neural network for predicting isothermal hydration by incorporating the Avrami–Erofeev–Arrhenius ordinary differential equation into the training loss and using a composition-focused sub-network to relate blend proportions to blend-specific kinetic parameters. The model was trained and tested on 29,379 calorimetry data points from 77 multi-component cement systems gathered from two open-access datasets. For previously unseen test systems, it reached an $R^2$ of 0.9864 and a symmetric mean absolute percentage error (sMAPE) of 25.3%. Its overall sMAPE was lower than that of a neural network with the same architecture but without physics-based constraints (34.9%); stage-resolved analysis showed that the largest difference occurred during early hydration. Across eight random seeds, the physics constraint did not confer a data-efficiency or training-stability advantage over the same-architecture network. The distinguishing features of the PINN are instead its explicit ODE and monotonicity regularization and its composition-conditioned effective parameters. Although random forest regression produced lower pointwise error, it failed to maintain physically consistent hydration-rate behavior. Overall, the framework provides a composition-conditioned surrogate that incorporates kinetic regularization within the composition range investigated.