A surrogate modeling approach for rapid moving-load prediction of steel box girder bridges within a unified COMSOL framework
Different software platforms may influence the surrogate model’s comparison and evaluation results. To overcome this challenge, this paper develops a unified surrogate modeling framework to predict the structural responses of a steel box girder bridge under moving-load analysis, ensuring consistency in both data generation and model assessment. Numerical simulations are conducted in MIDAS Civil to generate datasets of bridge’s deflection and combined stresses under variations in uniformly distributed and concentrated loads, as well as their simultaneous changes. Based on identical datasets, deep neural network, polynomial chaos expansion, and Gaussian process, are constructed and implemented within COMSOL for a consistent comparison. The models’ performance is quantified using the mean absolute error, root mean square error, and coefficient of determination. Bridge structural responses under moving-load analysis are effectively captured by all three surrogate models. The Gaussian process model performs best for single-parameter variations, while the polynomial chaos expansion model is more robust for multi-parameter cases. Stress responses are more sensitive to prediction errors than deflection. The proposed framework serves as a valuable benchmark for surrogate modeling-based prediction of bridge structural responses.