Physics-constrained machine learning for constitutive modelling of high-strength concrete across strain rates

This study presents a physics-constrained extreme gradient boosting framework (PhysicsXGB) for the reliable constitutive modelling of high-strength concrete across strain rates. While data-driven models show promise for capturing complex material responses, their application in mechanics is often limited by potential physical inconsistencies and opaque decision-making, particularly in critical regimes like post-peak softening. To bridge this gap, the proposed framework embeds constitutive knowledge through mechanics-aware feature engineering and enforces monotonic damage evolution via a dual-objective loss function that balances statistical accuracy with physics compliance. Furthermore, it introduces a phase-stratified interpretability system that extends Shapley additive explanation analysis into a mechanics-aware diagnostic tool, quantifying feature importance across elastic, pre-peak, peak, and softening regimes. Results demonstrate that PhysicsXGB not only achieves high predictive accuracy ( R ² > 0.99) but, more importantly, ensures exceptional physical consistency, reducing error magnitudes by over 80% compared to a physics-agnostic extreme gradient boost model. The framework reliably reproduces rate-dependent stiffness, strength, and damage-driven degradation, effectively closing key research gaps in post-peak prediction and model interpretability. It thus establishes a framework that unifies data-driven learning with constitutive principles for transparent and trustworthy material modelling.