Machine learning–driven strength prediction and optimal waste glass powder dosage for concrete exposed to high temperatures

This study explores the synergistic effects of the addition of waste glass powder (WGP) and elevated temperatures on the mechanical behavior of concrete, and proposes a machine-learning-based framework for mixture design optimization. Six concrete mixtures containing 0–50% WGP were tested for compressive, flexural, and splitting tensile strengths over the temperature range 24–600 °C, and the results consistently showed an optimal WGP dosage of 10–20% at room temperature and a pronounced degradation in these strengths above 400 °C. To model the strength evolution across the entire WGP-temperature space, a physics-aware data augmentation method and a multi-output Random Forest surrogate model are developed and validated using mixture-wise three-fold cross-validation, parity plots, Y-randomization tests, SHAP-based interpretability analysis, and bootstrap-based uncertainty quantification. The validated surrogate is then combined with NSGA-II to determine WGP dosages that simultaneously maximize room-temperature strength, high-temperature residual capacity, and material efficiency. Both prospective within-domain experimental validation showed close agreement between the surrogate predictions and independently measured responses. Contour maps and optimization results also consistently show that the 10–20% WGP dosage offers the best trade-off between initial strength and thermal resilience. The results demonstrate the two-fold role of WGP as a micro-filler in improving ambient performance and as a thermally sensitive phase governing degradation mechanisms at elevated temperatures. Therefore, this work presents a validated ML-based framework and practical design charts for optimizing WGP-modified concrete for fire-exposed or high-temperature applications. SEM analyses of the glass powder-reinforced concrete sample exposed to 600 °C showed an increase in cracks and voids in the microstructure; EDX results indicated Ca enrichment with the dominance of O (51.49%), Ca (20.78%), and Si (8.17%); FT-IR spectrum showed the disappearance of the ~ 3400 and ~ 1650 cm⁻¹ bands and the weakening of the ~ 950–1000 cm⁻¹ band, confirming C–S–H degradation; and XRD analysis revealed a decrease in amorphous humps and the prominence of crystalline peaks, indicating that the structure became more crystalline and brittle at high temperature. TGA results indicated that room-temperature-cured samples exhibited higher mass loss and progressive degradation at low temperatures, whereas samples pre-exposed to 600 °C exhibited reduced mass loss and enhanced thermal stability, owing to prior dehydration and microstructural reorganization.