Predicting compressive strength of concrete is crucial for optimized mix design for engineering properties and performance. While machine learning models are used for prediction, their effectiveness is often limited by challenges related to generalizability and scalability. This study developed knowledge-informed neural network (KINN) to predict compressive strength of concrete by integrating the empirical equation of concrete strength growth into hybrid loss function based on a large dataset from field projects. The performance of KINN is compared with other models including artificial neural networks (ANN), support vector regression (SVR), k-nearest neighbors (KNN), random forest (RF), eXtreme Gradient Boosting (XGBoost) and the empirical equation.The results demonstrate that KINN achieves superior predictive stability and physical consistency compared to ANN having identical architecture, while offering competitive accuracy. Compared with tree-based ensemble methods (XGBoost, RF), KINN presents a trade-off: slightly lower mean R² but substantially reduced variability and physically consistent predictions aligned with domain knowledge. Furthermore, the SHapley Additive exPlanations (SHAP) analysis reveals that knowledge-informed models not only capture the underlying physical principles but are also more interpretable for the effect of water-binder ratio. Partial dependence plots and Accumulated Local Effects (ALE) are employed to investigate the effects of curing ages, water-to-binder ratio and percentage of supplemental cementitious materials (SCM) on compressive strength, providing valuable insights for engineering decision making. The integration of empirical knowledge into machine learning models, as demonstrated by KINN, enables engineers to incorporate prior domain expertise into predictions of concrete performance.
