Forecasting the dynamic behavior of curved thin-walled box-girder bridges subjected to high-speed train loads is computationally demanding yet crucial for evaluating structural integrity. This study seeks to assess the Dynamic Load Factor (DLF) through a comprehensive framework that ensures high precision while markedly decreasing computing demands. A detailed coupled dynamic model was created, including a 38-degree-of-freedom train system, a three-layer slab track, and a thin-walled box-girder bridge. The system addressed rail irregularities through power spectral density (PSD) and wheel-rail interactions based on Hertzian contact theory. The resultant data trained a feedforward artificial neural network (ANN) employing the Levenberg-Marquardt algorithm. The model utilized span length and bridge damping as input variables to forecast essential structural responses, such as shear force, bending, torsion, and distortion. The ANN model attained exceptional prediction accuracy, with regression coefficients (R) above 0.99 and negligible mean squared error (MSE) across all datasets. The results illustrate the model's capacity to precisely identify intricate nonlinear correlations between structural factors and dynamic outputs. This research presents a unique method for the rapid evaluation of bridge behavior by combining a robust finite element framework with a data-driven surrogate model and particularly suitable for real-time structural optimization in high-speed rail engineering.