Effects of mixture design and cooling rate on the low-temperature TSRST responses of asphalt mixtures: Experimental evaluation and artificial neural network modelling


Creative Commons License

Qadir A., Guler M., Karakaya Y.

CONSTRUCTION AND BUILDING MATERIALS, ss.148751, 2026 (SCI-Expanded, Scopus)

Özet

This study evaluated seven material, mixture design, specimen, and testing factors within a Thermal Stress Restrained Specimen Test (TSRST) framework. A 72-combination fractional factorial design comprising 72 compacted slabs and two separately tested beam specimens from each slab produced 144 TSRST results for fracture strength (FS), fracture temperature (FT), relaxation stress (RS), and transition temperature (TT). Artificial neural network (ANN) models were fitted to the treatment combination means and evaluated using repeated nested cross-validation, grouped validation, and benchmark models. Aggregate type and binder content were the only factors significantly associated with all four TSRST responses. Limestone mixtures exhibited higher FS and RS and more negative FT and TT than basalt mixtures; however, because air void content differed substantially between aggregate groups, these results should be interpreted as differences between the investigated mixture systems rather than aggregate mineralogy alone. Mixtures prepared at their mixture-specific optimum binder content generally exhibited a more favourable balance of TSRST responses than mixtures deviating from the optimum. Styrene–butadiene–styrene (SBS) modification was associated with FS and FT but did not show consistent improvement across responses, whereas cooling rate was not significantly associated with any of the responses within the investigated 5 to 20 ◦ C/h range. ANN performance was moderate for FS and RS, but limited for FT and TT, and benchmark comparisons showed that the ANN had no consistent predictive advantage over ridge or random forest regression. Overall, the study integrates fracture and pre-fracture TSRST assessment with preliminary ANN-based prediction within the experimental domain investigated.