Physics-Informed Neural Network Surrogate for Predicting Hydration Heat Evolution Profiles of Portland Limestone Cement (PLC)

Accurate prediction of the heat-evolution profile is particularly important for Portland-limestone cement (PLC) because limestone does not simply dilute the binder but fundamentally alters the hydration kinetics and reaction pathways. Therefore, relying solely on the final cumulative heat is insufficient; instead, the full time-resolved heat-flow curve is essential for understanding and modeling PLC behavior. This study introduces a data-assisted physics-informed neural network (DA-PINN) that fuses calorimetry with mechanistic kinetics to learn continuous cumulative heat of hydration Q ( t , α ). The model encodes rate, capacity, and onset as smooth α -dependent functions. Across eight PLC mixtures (0–40% limestone), DA-PINN closely tracks both induction and acceleration stages and shows strong interpolation performance under leave-one-out cross-validation (LOOCV). Additional comparisons with both physics-free multiple layer perceptron (pure MLP) and Knudsen-type kinetic baseline further clarify the benefit of embedding physically constrained kinetics into the learning process. The scope of the present study is deliberately restricted to PLC systems; transfer to chemically distinct binder systems is left for future work and would require separate governing assumptions and independent validation. Within this calibrated PLC domain, DA-PINN provides a practical, interpretable surrogate for reconstructing full hydration-heat curves from sparse calorimetry data and for accelerating low-carbon PLC mixture screening, which is urgently needed. • DA-PINN reconstructs PLC hydration curves from sparse calorimetry. • LOOCV confirms interpolation across 0-40% limestone replacement. • RBF kinetic functions enable physically guided composition interpolation. • Pure MLP and Knudsen baselines clarify DA-PINN advantages.