Multi-input PINN Surrogate Model for Digital Twin of Lead-cooled Fast Reactor Main Pump

The main coolant pump (MCP) of lead-cooled fast reactor (LFR) faces significant corrosion challenges due to the high-temperature, high-density coolant flow and heat transfer within complex geometric domains. Consequently, detailed attention to the internal flow field of the pump is crucial during the design process. Currently, the design of MCP for LFR heavily relies on extensive numerical simulations, requiring the generation of high-quality meshes and the discretization of governing equations for computation. This approach results in high computational costs and prolonged design cycles. To address the issues, a physics-informed neural network (PINN) surrogate model with a multi-input structure was developed and applied it for rapid prediction of the flow field of the MPC in LFR under different operating conditions, to be used in the state tracking module of a digital twin system. The surrogate model was proposed for overcoming the limitations of traditional numerical simulations, such as high computational complexity and large data demands, while enhancing model response speed and reducing storage requirements. First, the flow field within an axial lead-bismuth eutectic (LBE) pump was modeled and the loss function for training the multi-input PINN surrogate model based on the governing equations of the pump was derived. Subsequently, the accuracy of the surrogate model was validated by comparing its reconstruction of the flow field under several specific operating conditions with the results obtained from traditional computational fluid dynamics (CFD) simulations. Additionally, the computational time of both approaches was compared. The results show that the model demonstrates good ability in identifying flow field details near the rotating blades which suffer from the most corrosive environment in the whole pump, with a relative error to CFD calculation results of generally less than 20%, a prediction time of approximately 0.273 seconds per operating condition, and a prediction speed improvement of 15 000 times. The training time is equivalent to that of 2 to 3 CFD simulations for different operating conditions. Furthermore, the model parameters for predicting the flow field across the entire flow passage of the MCP under arbitrary operating conditions require only 286 kB, whereas the flow field data generated by CFD for a single operating condition amount to 413 MB. The conclusion indicates that the surrogate model can effectively fit nonlinear training data, with high prediction accuracy and generalization capability, providing a rapid and efficient flow field prediction tool for the design of the MCP of LFRs, which can be used as the state tracking module in the corresponding digital twin software.