Transfer learning-enhanced physics-informed neural networks for efficient simulation of river flood in complex large-scale river systems

Abstract Physics-informed neural networks (PINNs) have shown considerable potential for solving hydraulic problems by integrating governing physical laws into neural network optimization. However, their application to real-world river flood remains challenging because of the high computational cost associated with nonlinear and non-convex PINN optimization. Conventional PINNs are generally initialized without task-specific hydraulic knowledge, each new hydraulic scenario requires optimization from an uninformed initial state, resulting in repeated training and substantial computational expense. This study proposes a transfer learning method that can significantly improve the prediction accuracy and training efficiency of PINNs for complex river floods under varying flood events. A transfer learning-enhanced PINN (TL-PINN) framework is developed, in which a data-driven neural network is first trained on a source flood event. The optimized parameters are then transferred to initialize the PINN for target flood events, followed by optimization using a composite physics-informed loss function. The proposed framework is evaluated on a real-world river reach with complex channel geometry and highly unsteady flood processes through three numerical experiments. TL-PINN is compared with three conventionally initialized PINNs using PyTorch Default Initialization, Orthogonal Initialization and Xavier Normal Initialization. TL-PINN consistently maintains superior performance under limited labeled-data conditions and under more severe and complex flood inflows. The results indicate that the transferred parameters provide an effective initialization, leading to improved prediction accuracy and optimization efficiency. These findings demonstrate that transfer learning might significantly enhance the real-world applicability of PINNs for efficient and accurate river flood simulations under varying flood events.