Physics-based active-learning surrogate model for rapid safety assessment of temporary steel trestle bridges under construction vehicle loads

Abstract

The safety status of temporary steel trestle bridge will deteriorate rapidly due to severe riverbed erosion caused by natural disasters such as floods and storm surges, rapid degradation of supporting structures, and the superposition of heavy vehicle loads, which can easily lead to collapse accidents. This paper develops a physics-based Gaussian-process regression (GPR) framework with active learning for rapid safety assessment of temporary steel trestle bridges. A parameterized Euler-Bernoulli beam model on elastic supports is used to calculate structural responses under moving construction vehicles. Seven uncertain parameters are considered: vehicle gross mass, vehicle speed, local scour depth, support stiffness factor, deck damage ratio, lane eccentricity, and braking force ratio. The resulting active-learning GPR (AL-GPR) surrogate selects new simulations near the safety limit state and uses the simulation budget efficiently. On an independent 500-sample test set, the AL-GPR model achieved a coefficient of determination (R 2 ) of 0.997, a root mean square error (RMSE) of 0.017, and a safe/unsafe classification accuracy of 0.996. Compared with passive sampling under the same simulation budget, active learning reduced the classification error by 33.33%. The results show that local scour depth, vehicle mass, and support stiffness are the dominant risk drivers, and that the trained surrogate can generate speed-load safety maps for real-time construction traffic management.