Machine-Learning-Based uncertainty and sensitivity analysis of Reinforced-Concrete slabs subjected to fire

Reinforced concrete (RC) slabs are integral parts of building structures and provide compartmentation func-tionality when subjected to fire. However, the fire resistance of RC slabs is affected by numerous factors that are mostly inherent uncertainties. Therefore, uncertainty and sensitivity analyses were conducted to examine the influence and significance of the potential uncertain parameters on the fire resistance of RC slabs. To this end, a set of RC slab samples with various design parameters were generated and different fire scenarios were considered using parametric fire curves. The slab samples were randomly coupled with the fire scenario samples. For each of the slab-fire pairs, finite-element simulations were conducted, and three specific fire durations were identified to represent the slab fire resistance corresponding to the failure criteria on the steel temperature (SP-I), unexposed surface temperature (SP-II), and mid-span deflection (SP-III). As a consequence, a database was established by collecting the calculated fire resistance for all the considered slab-fire samples. The prediction models of the slab fire resistance in association with uncertain parameters were developed by four commonly used ML algorithms, including linear regression, random forest, gradient boosting decision tree, and extreme gradient boosting. Among the developed ML-based prediction models, the extreme gradient boosting model was proven to have superior predictive accuracy. Therefore, it was further utilized for uncertainty analysis by considering 14 uncertain parameters from fire scenarios, material strengths, geometric dimensions, and external loads. The uncertainty analysis results showed that the considered uncertain parameters cause significant vari-ability in the fire resistance of slabs, and the coefficients of variation were 11.5%, 13.1%, and 20.6% for the fire durations related to SP-I, SP-II, and SP-III, respectively. Moreover, the SHapley Additive exPlanations method was used to examine the sensitivity of the considered uncertain parameters. It was found that the parameters related to fire scenario were more influential to the fire resistance of slabs than the slab design parameters. The opening factor, slab thickness, and fire load density had noticeable effects on the fire resistance of SP-I, SP-II, and SP-III, respectively, whereas the uncertainties from material strengths and convection conditions had negligible effects.