Abstract: Broadband ground motion simulation is a critical issue in engineering seismology, as traditional methods often face issues of spectral mismatch and energy phase conflicts when combining low-frequency physics-based modeling with high-frequency stochastic components. This paper introduces a hybrid method that integrates artificial neural networks (ANN) with spectral element method (SEM). The ANN is trained on the Strong Motion Flatfile of China to capture the nonlinear mapping of short-period response spectra, while the SEM is employed to simulate low-frequency ground motions. High-frequency stochastic components are optimized using scaling factors, and energy alignment is applied to synchronize low- and high-frequency time histories, ensuring stable broadband simulation results. Taking the Yangbi M_S6.4 earthquake as a case study, a finite fault model derived from inversion and a refined 3D velocity structure model are used to generate low-frequency time histories for monitors. The broadband method is then applied to produce corresponding broadband simulation time histories. The results demonstrate that the simulated broadband acceleration time histories exhibit strong consistency with observed records, providing reliable and realistic inputs for seismic hazard analysis.