Reliability analysis of the world's longest-span CFST arch bridge using an interval-focused active-learning surrogate model

Reliability analysis of long-span concrete-filled steel tubular (CFST) arch bridges remains challenging because structural stability is governed by multiple uncertain parameters, and conventional surrogate models may not provide balanced prediction accuracy across failure, critical, and safe response intervals. An interval-focused active-learning multilayer perceptron surrogate (IF-AL-MLP) is therefore developed for the global-stability reliability analysis of the Third Pingnan Bridge (560 m). A refined finite element model (FEM) is established to compute the stability index K . Under a common budget of 4000 FEM-evaluated training samples, the proposed model is compared with a conventional MLP and standard active Kriging-Monte Carlo simulation (AK-MCS) using the same 400 FEM test samples. The final surrogates are then applied to three independent sets of 20 million Monte Carlo samples. Although the failure-interval mean squared error (MSE) of the IF-AL-MLP is slightly higher than that of the conventional MLP, it reduces the critical- and safe-interval MSEs by 60.2% and 56.7%, respectively, and yields lower MSEs than AK-MCS in all three predefined intervals. All three models produce reliability indices exceeding the code requirement of 4.7. Their exact Poisson confidence intervals overlap substantially, indicating that the differences among the failure-probability estimates are not statistically significant, while the IF-AL-MLP yields identical failure counts across the three simulations. These results show that the proposed strategy provides balanced interval-wise regression accuracy and supports the same code-based reliability conclusion as the benchmark methods, without establishing a universal ranking of surrogate methods.