Abstract: Finite element models derived from construction drawings often fail to capture in-service bridge responses, and updating complex bridges is frequently hindered by parameter coupling and poor static-dynamic joint convergence. This study proposes physics-data dual-driven stepwise and unified dynamic model updating strategies for complex bridges. A physics-guided, data-driven framework is developed to enrich sparse samples and build a physics-embedded PINN-LSTM surrogate, reducing the prediction error by 82.67% relative to the purely data-driven baseline. Comparative results show that the stepwise “static-first, dynamic-next” strategy improves the static accuracy (error: 14.90%to1.90%) but yields limited frequency fitting, whereas the unified static-dynamic feature fusion strategy markedly enhances the dynamic fidelity, decreasing the first three modal frequency errors from 8.05%, 9.11%, and 9.90% to 0.02%, 0.70%, and 0.88%, respectively, and providing a calibrated baseline model for subsequent nonlinear dynamic analysis. An integrated visualization software tool is further developed to facilitate engineering applications.