Physics-informed neural networks (PINNs), built upon general neural networks, integrate physical information into the loss function, achieving remarkable results in solving partial differential equations (PDEs). Meanwhile, PINNs face challenges in solving complex interface problems, such as the Navier–Stokes/Darcy interface model considered in this paper. To address these challenges, we extend the general PINNs to propose subregion–identified physics-informed neural networks (SI-PINNs), which consist of three different neural networks for different subregions (Navier–Stokes, Darcy, and interface) with separated datasets. By identifying and classifying the regional features of the neural networks and datasets, the complexity of the physical information that SI-PINNs need to process is effectively split, while the networks also take all physical information into account during the training. Additionally, this splitting feature substantially reduces the training costs while enhancing or maintaining the accuracy of the general PINNs for the target model. Numerical experiments are performed to validate the effectiveness of the SI-PINN method.