A physics-informed cascaded learning framework for end-to-end stress–strain prediction of recycled aggregate concrete under axial compression

The reliable prediction of the full constitutive response in recycled aggregate concrete is frequently compromised by the high stochasticity of material properties and the unavailability of peak parameters during the structural design phase. This study develops a two-stage sequential learning framework for the end-to-end reconstruction of stress–strain trajectories without requiring predefined peak inputs. The architecture integrates gradient boosting surrogate models for peak stress and strain estimation with a physics-informed long short-term memory network for full-curve generation. By embedding failure constraints based on energy conservation alongside monotonicity and boundary priors into the learning process, the framework effectively suppresses non-physical behavior in the softening regime. Comparative evaluations demonstrate that the proposed system exhibits superior robustness and morphological fidelity, achieving a curve-level coefficient of determination of 0.923 and a mean dynamic time warping distance of 27.91. The framework remains stable under ten percent data contamination and utilizes quantile regression to provide quantified prediction intervals that address predictive uncertainty. Furthermore, analysis of the generated curves identifies the enhanced deformation capacity of high-replacement recycled aggregates can compensate for strength loss in high-porosity matrices. The energy toughness ratio is primarily governed by the aggregate replacement ratio rather than matrix strength, providing a standardized measure for the brittle-to-ductile transition. This research offers a computationally efficient numerical tool for characterizing sustainable concrete, ensuring that predicted responses are both mathematically optimized and thermodynamically consistent with fundamental failure mechanisms.