Portfolio-scale seismic fragility of RC bridge columns with series-distributed neural networks

This paper proposes a novel series-distributed artificial neural network framework for rapidly constructing seismic fragility curves of reinforced-concrete (RC) bridge columns at markedly reduced computational cost. Three coupled surrogate models are trained on datasets generated from nonlinear time-history and pushover analyses of RC piers with randomly sampled geometric and material properties subjected to hazard-consistent ground motions. The first network learns correlations among a reduced set of efficient ground-motion intensity measures (IMs), the second predicts drift demand from IMs and modelling parameters, and the third provides drift capacities for multiple damage states directly from capacity-curve information, thereby incorporating epistemic uncertainty in structural capacity. The trained surrogates are embedded in a Monte Carlo simulation scheme to estimate, in a largely non-parametric manner, the probability that drift demand exceeds capacity at each IM level. A case study on a portfolio of simply supported bridges in the Da Nang area, including selected bridges along National Highway 1A, demonstrates that the framework reproduces benchmark fragility curves from nonlinear analyses while achieving substantial reductions in analysis time. The results highlight systematic differences between rectangular and circular piers and quantify the impact of relaxing internal lognormal assumptions relative to traditional cloud-based fragility derivation. The proposed approach is implementation-ready, which relies on standard structural and ground-motion descriptors, delivers conventional fragility parameters, and is readily scalable to portfolio- and network-level seismic risk assessments and screening.