Abstract
<jats:p>Accurate prediction of residual flexural strengths is essential for the structural design and performance assessment of steel fibre reinforced concrete (SFRC). While machine learning models have recently demonstrated strong predictive capability, most of the existing approaches rely on heuristic feature selection, unconstrained architectures, and random validation splits that may overestimate generalization capacity. This study proposes a Hard-constrained Mechanics-Informed Neural Network framework for predicting SFRC residual strengths across independent experimental studies. A multi-study database comprising 860 observations from 18 experimental assemblies was analysed to identify statistically dominant and physically interpretable predictors. Based on exploratory statistical analysis and micromechanical assessment, a hierarchical shared-backbone neural network architecture was developed. Physical admissibility was enforced structurally through non-negativity constraints, monotonic dependence on effective fibre bridging capacity, and bounded signed transitions between CMOD stages. Model performance was evaluated using five-fold group-based cross-validation, where entire studies were excluded from training to rigorously assess generalization. The proposed architecture achieved predictive performance comparable to or exceeding conventional neural networks and ensemble models, while eliminating physically inadmissible predictions and reducing fold-to-fold variability. The results demonstrate that embedding mechanical principles directly into neural network architecture enhances robustness and interpretability without compromising accuracy. The proposed hard-constrained Physics-Informed Neural Networks (PINN) framework provides a transferable methodology for physics-guided machine learning in structural materials modelling.</jats:p>