This study focuses on the complex dynamic interaction between high-speed trains and curved, thin-walled box-girder bridges, where traditional computational approaches are typically insufficient for lengthy parametric simulations. A thorough numerical model was developed by combining a 38-degree-of-freedom train system and a slab track structure with a thin-walled, curved box-beam finite element model. The resulting coupled dynamic simulation data was utilized to train two different Artificial Neural Network (ANN) models that predicted essential structural reactions based on two major input variables, train velocity and subgrade suspension system parameters. The parametric analysis yielded substantial results for dynamic behaviour of the bridge. The model demonstrated that increasing train velocity from 150 km/h to 400 km/h significantly increases all stress resultants such as shear force, bending moment and torsional moment. Changing the subgrade suspension parameters, on the other hand, led in only an insignificant change in the bridge's dynamic response, implying a minor impact on global behaviour. Both ANN frameworks underwent intensive validation, displaying remarkable predictive performance. Diagnostic plots repeatedly revealed near-perfect correlation coefficients (R nearing 1), confirming the robustness of the Levenberg-Marquardt training procedure, which achieved rapid, stable convergence. Furthermore, the Error Histogram analysis confirmed that the network produces unbiased prediction residuals, indicating that the model is reliable and robust for generalization. The successful implementation of the ANN framework offers structural engineers with an accurate, computationally efficient tool for performing quick safety assessments and design optimization on high-speed rail infrastructure.