Abstract :
[en] Specialized resistance models for headed stud connectors in recycled aggregate concrete (RAC) are lacking. Although the load-bearing mechanisms of headed studs in natural aggregate concrete (NAC) and RAC are similar, the reliability of using NAC-based models to predict stud resistance in RAC slabs remains uncertain, particularly given the inferior mechanical properties of RAC, such as its lower modulus of elasticity and higher variability. These limitations hinder the wider application of RAC in steel–concrete composite structures. In this context, the present study evaluated existing empirical models as well as machine learning (ML) models, which are often expected to provide more accurate predictions than traditional descriptive formulations. In addition, a novel polynomial chaos expansion (PCE) model was developed, enabling explicit quantification of uncertainty propagation from the input random variables. Using the available tests on headed studs in RAC, reliability analyses per Eurocode 0 were performed, and required partial factors for design resistance were determined. The preliminary results, based on this limited dataset (27 tests), showed that a higher partial factor, approximately 1.94 instead of 1.25, would be needed to achieve the target reliability for headed stud connections using RAC, corresponding to about a 36% reduction in design resistance, even with the same concrete strength class. Among empirical models, Eurocode 4 gives the highest design resistance. Although both the PCE and ML models are developed from the same NAC dataset, the PCE model demonstrates superior performance, which may be attributed to its lower sensitivity to data heterogeneity. Nevertheless, a well-fitted PCE model still requires a sufficient number of data points from well-designed tests that adequately cover the variable domains. Notably, the PCE model produces design resistance 16% higher than that predicted by Eurocode 4. These findings suggest that PCE may offer a promising approach for developing resistance models, leveraging both data-driven advantages and the underlying physical mechanisms.
