While Finite Element Analysis (FEA) possesses high accuracy in mimicking real-world structural responses, the accuracy in the simulations of existing structures can greatly be affected by fluctuations of actual structural parameters such as material properties and boundary conditions. In response to the active trend of the Digital Twin that requires precise digital representations of the existing physical structures especially for long-term Structural Health Monitoring (SHM), this study hence illustrates Finite Element Model Updating (FEMU) of a short-span prestressed concrete girder bridge using multi-restart Covariance Matrix Adaptation Evolution Strategy (CMA-ES) with the help of Neural Networks (NNs) and Sequential Niche Technique (SNT). In this study, a highly detailed FE model of the bridge is first developed and validated using measured dynamic responses. A series of parametric studies are then conducted to identify a set of sufficiently influential updating parameters. Next, NN models are developed to be surrogate models that predict the high-dimensional and nonlinear relationships between the updating parameters and FE-simulated dynamic responses within the optimization routine. This significantly enhances the efficiency of the FEMU by dramatically accelerating the process. As a result, an updated FE model yielding dynamic responses that better align with the measurements is obtained. This can serve as a solid digital model for future uses in SHM.