Machine-learning-based performance prediction for partially encased steel-concrete composite columns after fire exposure

The partially encased steel-concrete composite (PEC) members inherit the advantages of easy connection and good durability from the conventional steel and concrete structural members respectively, and are being widely employed in modern engineering structures. Due to the partially exposed characteristics of the steel sections, the performance of such members under and after elevated temperature exposures is a major concern among researchers and engineers. Therefore, this work aims to examine the post-fire residual bearing capacity of the PEC columns via a series of machine-learning-based methods. A finite element model was developed and employed to enrich the analysis database since the experimental data available is limited. Based on the reasonably sampled data in an amount of approximately 400 groups and the numerical tests with seven machine-learning algorithms for options, good precision in the prediction of the post-fire performance was achieved with such limited sample sizes, which could be employed for the fast and accurate post-fire performance assessment of the PEC columns. Moreover, design equations were also proposed in an efficient mode, i.e., aided by the developed machine-learning-based (ML) prediction model for fast parameter sensitivity identification and fitted using the numerical output directly, which attains the analytical results and conclusions in this work can also be blended into the traditional post-fire assessment/design approaches.