Research On High-Efficiency Solutions For Large-Scale Flow To Vibration Problems

In the high temperature,high pressure and highly radioactive operating environment of nuclear reactors,long-term coolant washout affects fuel assemblies.The circulating flow of coolant causes small vibrations,which in turn cause micro-motion wear and high circumferential fatigue,both of which mechanisms have an impact on the structural integrity of the fuel assembly under normal operating and accident conditions.In order to control the structural vibration response caused by flow induced vibration,the structural design must be optimised and the flow rate adjusted to control the vibration response within acceptable limits.Numerical calculations of fuel rod rheological vibrations are therefore an integral part of safety analysis.As the scale of the grid increases,high-precision numerical simulation has become increasingly indispensable for investigating fuel rod flow-induced vibration problems.However,this comes with a growing complexity of problem-solving and computational resources needed,leading to traditional computer clusters and serial methods being unable to meet the demand for solutions.This study investigates a parallel solving method for large-scale flow-induced vibration based on a domestic heterogeneous supercomputing architecture.To address the problem of large-scale flow-induced vibration in the entire core fuel assembly,this thesis proposes an optimized numerical simulation parallel method based on the heterogeneous platform.Furthermore,to tackle the problem of high grid dependence in traditional finite element methods,this study explores the method of embedding physical knowledge neural networks and achieves a domain decomposition-based embedded physical knowledge neural network model to solve flow-induced vibration problems.The main contributions of this thesis are:(1)In this thesis,the New Mark integration method based on finite element tearing and interconnecting method(FETI)is used to solve large-scale flow-induced vibration problems.The FETI method implements region division and parallel solution,and the New Mark method is used for time-step updating.In order to solve the problem of uneven distribution of the number of Lagrange multipliers after the second-level decomposition introduced by the FETI method and the computational performance difference of the heterogeneous computing core,this thesis proposes a parallel optimisation method for numerical simulation of fluidic vibrations based on a heterogeneous platform(FVFETI),including a graph bisection algorithm for domain boundary balancing and a dynamic load balancing optimisation strategy.At the same time,multi-stream pipeline processing is enabled to overlap data reading and computation to improve processor utilisation efficiency.(2)This thesis proposes a domain decomposition-based embedded physical knowledge neural networks(FVDDPINNs)model for solving large-scale fluid vibration problems,which faces problems such as grid dependence and large computational effort.The model uses neural networks to analyse and extract the solid vibration laws of largescale fluid vibration problems,and limits and corrects the network output by embedding the physical equations of the flow-induced vibration problems as well as the initial conditions to ensure the reliability and accuracy of the model.The domain decomposition method is also used to achieve parallel solution and meshless prediction of large-scale fluid vibration problems and to improve the training efficiency of the neural network.(3)In this thesis,high-fidelity simulations of large-scale fluidic vibration problems are carried out in a high-fidelity analysis software for reactor structural mechanics simulation(HARSA).The proposed FVFETI method enables the numerical simulation of a 10-billion-cell-scale grid of core fuel rod assemblies,with solutions at the minute level and scalable to 16,000 AMD GPUs.FVDDPINNS has higher prediction accuracy and more precise accuracy than traditional physics informed neural networks in dealing with large-scale fluidic vibration problems.The two methods proposed in this thesis are suitable for the numerical solution of very large scale fluidic vibration problems.