Thermo-mechanical performance and robust design of polypropylene waste–modified concrete using physics-guided machine learning

The application of polymer-based waste to concrete is one promising way to improve sustainability, however, the properties of this type of concrete are still not well understood under high-temperature conditions. This study proposes a combined experimental and physics-informed computational framework to support the robust design of concrete modified with shredded woven polypropylene (SWPP) under elevated temperatures, and a total of 330 specimens were prepared using 11 concrete mixtures with SWPP replacement ratios of 0% to 50%. The specimens were tested at five temperatures (24, 200, 300, 400 and 600 °C), and the temperature-dependent degradation of mechanical performance, including compressive strength (CS), flexural strength (FS) and splitting tensile strength (STS), were measured. Unlike traditional experiments or purely data-driven approaches, the proposed framework incorporates leakage-safe machine learning, regime-aware data augmentation and reliability-based optimization within a unified framework, and a piecewise-aware data augmentation strategy is proposed to preserve physically meaningful thermal degradation regimes. To build the surrogate models, ensemble surrogate models were trained independently for each mechanical response, and to ensure the robustness of the ensemble surrogates, we employed four different methods to validate the accuracy and reliability of the trained surrogates. The first method is based on a frozen real-only test, while the second approach involves conducting ablation studies to further verify the accuracy of the trained surrogates. The third method involves residual diagnostics to provide insights into the error distribution and the potential outliers in the training set, and the last method includes calibration checks to evaluate the reliability of the trained surrogate models. Finally, we incorporate the Monte Carlo-based uncertainty propagation into the surrogate models to generate reliability-oriented design maps that aim to maximize the performance stability instead of maximizing the nominal strength. It is shown that the thermo-mechanical degradation rate increases significantly with increasing temperature above 300 °C, and tensile-related properties (FS and STS) are more sensitive to both SWPP content and thermal damage compared with compressive strength. Explainability analyses confirm physically consistent correlations between temperature, polymer content, and strength loss, and the surrogate-based response surfaces also demonstrate a residual-strength plateau at high temperatures, indicating a shift toward matrix-dominated failure behavior. Microstructural observations are consistent with experimental findings, revealing a deteriorated ITZ, polymer degradation, and increased matrix porosity at high temperatures, and the proposed framework provides a scalable, interpretable, and reliability-oriented tool for designing sustainable fire-resistant cementitious composites, especially under limited data.