As universal function approximators, deep neural networks have the potential of being the surrogate solver of the Navier-Stokes (NS) equations. This was recently demonstrated via the Physics Informed Neural Network (PINN) on aneurysm flows by Sun et al. However, PINNs are specific to the geometry of the flow domain and require slow training for each new geometric scenario encountered. Here, we present an alternative approach, where a deep learning (DL) side network is cascaded to a PINN domain network for the pre-training of varied geometric cases, which has the potential to enhance network robustness and decrease training complexity.