Hydration heat prediction is important for mixture design, thermal control, and early-age cracking risk assessment for cement-based materials. The increasing use of supplementary cementitious materials (SCMs) and chemical admixtures in low-carbon concrete has made hydration behavior more complex, while conventional empirical or simplified analytical models have difficulty describing multicomponent heat release. Therefore, this study develops a general model for predicting cumulative hydration heat directly from mixture composition.

A database of 219 cumulative hydration heat curves and 13 048 data points was compiled from 33 published isothermal calorimetry studies, covering ordinary Portland cement and blends with fly ash, ground granulated blast-furnace slag, and silica fume. The input features included water-to-binder ratio, component fractions, oxide compositions, specific surface area, and loss on ignition. Hydration kinetic equations, including the Schindler-Folliard, Krstulovic-Dabic, and Knudsen equations, were embedded into neural networks as physical priors. The framework consisted of a physical-parameter prediction branch, a physical-equation layer, and a residual-correction branch, enabling end-to-end physics-driven prediction with data correction. Ten model variants were trained and compared under identical experimental settings.

SFQmaxDecompNet achieved the best overall performance, with a test R² of 0.962, a generalization gap of 0.018, and 94% of test curves reaching R²≥0.9. Its component-wise decomposition of Q_max better represented the different heat-release potentials of cement, FA, GGBS, and SF. KD-B ranked second (R² = 0.937), benefiting from its two-peak representation of early cement hydration and delayed SCM reactions. In contrast, KD-C had seven physical parameters but reached only R² = 0.868, showing that model performance depended more on the consistency between physical assumptions and data characteristics than on the number of parameters alone. PureNN reached R² = 0.874, lower than several physics-embedded models.

Data-efficiency tests showed that physical priors reduced the dependence on training data. When only 30% of the training curves (45 curves) were used, SFQmaxDecompNet still achieved R²=0.571, whereas PureNN degraded nearly completely. The physics branch dominated under limited data, while the residual branch provided local corrections when data were sufficient. Interpretability analyses further showed that Q_max represented the total heat potential, τ reflected the hydration time scale, and β described the concentration of heat release. Permutation importance and ALE analyses identified cement content as the dominant output feature, followed by FA, GGBS, FA-CaO, W/B, and SF. The temporal ALE trends of Q and dQ/dt were consistent with the delayed reactions of FA and GGBS and the early promoting effect of SF.

The proposed physics-driven and data-correction synergistic neural network provides accurate and interpretable hydration heat prediction of composite cementitious materials. Embedding hydration kinetic equations as inductive biases improves generalization, suppresses non-physical oscillations, and enhances robustness under small-sample conditions. This framework provides a useful route for mixture design, thermal analysis, and early-age crack-risk control in low-carbon cement-based materials.