Intelligent mesh refinement based on U-NET for high-fidelity CFD simulation in numerical reactor

Abstract

Thermal-hydraulic analysis of reactors is fundamental for studying many service behaviors inside the reactor core, and high-fidelity thermal-hydraulic simulation based on advanced computing techniques and supercomputing resources has attracted widespread attention. The construction of high-quality fluid domain mesh models for reactors with complex and sophisticated core structures is fundamental for high-fidelity thermal-hydraulic simulations. However, mesh models built entirely with commercial tools have limitations in terms of model quality and construction performance. An intelligent mesh refinement method based on the U-Net neural network model (Unet-IMR) is designed, which attempts to use AI technology to assist in generating high-quality mesh models. Specifically, an intelligent prediction model is constructed and used to predict the poor mesh in a given mesh model, and the designed region-sensitive refinement strategy is further applied to iteratively optimize and refine the mesh to construct a high-quality mesh model. Experimental results show that the mesh distribution and the transition between sparse and dense meshes are more reasonable in the Unet-IMR based mesh model, and meshes located at sophisticated locations have relatively high goodness-of-fit. All the meshes in the model can pass the correctness checks of the CVR-PACA software, and the simulation results are consistent with physical expectations. Furthermore, the Unet-IMR based mesh model can achieve similar simulation results with fewer meshes compared to the mesh model built with commercial tools.