Application of machine learning in seismic fragility assessment of bridges with SMA-restrained rocking columns

Abstract

This paper evaluates the seismic fragility of a two-span reinforced concrete (RC) bridge with shape memory alloy (SMA)-restrained rocking (SRR) columns through machine learning (ML) techniques. SRR columns incorporate a combination of replaceable superelastic NiTi (SMA) links and mild steel energy-dissipating links to achieve self-centering and energy dissipation, respectively, while their rocking joints are protected against compressive concrete damage through steel jacketing. To produce seismic fragility functions, initially, multi-parameter probabilistic seismic demand models (PSDMs) are generated for various engineering demand parameters through five different ML techniques (including neural network) and considering various sources of uncertainty, and the most accurate PSDMs are selected. The selected PSDMs are then interpreted using four different methods to investigate the effects of two key SRR column design parameters (self-centering coefficient and SMA link initial strain) and ambient temperature on the seismic performance of SRR columns. Subsequently, using neural networks, the PSDMs developed earlier, and appropriate capacity models, multi-parameter fragility functions are developed for various bridge damage states. After examining the effects of the two SRR column design parameters on the seismic fragility of the bridge, its seismic fragility is compared with those of the same bridge with monolithic RC and posttensioned (PT) rocking columns. It is shown that, in general, increasing the initial strain of the SMA links and decreasing the self-centering coefficient as possible (i.e., without compromising the self-centering) reduce the overall bridge damage. In addition, even considering the ambient temperature’s uncertainty, SRR columns are proven, at least, as effective as PT columns in mitigating the seismic damage of the bridges of monolithic RC columns.