ECCOMAS 2024

Distances between Proper Orthogonal Decomposition Reduced Subspaces of Repeating Subdomains

  • Ruan, Shenhui (Karlsruhe Institute of Technology)
  • Yanez, Jorge (Karlsruhe Institute of Technology)
  • Class, Andreas (Karlsruhe Institute of Technology)

Please login to view abstract download link

Flow blockage is a high-frequency accident in generation-IV nuclear reactors. It is a relevant safety issue whose analysis involves extensive parametric studies. In such frames, carrying out conventional Computational Fluid Dynamic (CFD) simulations is prohibitively expensive, and the Reduced Order Model (ROM) appears as an efficient alternative. This paper aims to investigate the characteristics of the Reduced Basis (RB) as a primary stage to support the development of ROMs. We start our analysis by considering high-fidelity simulations for a set of blocking conditions. In our approach, the reactor core is subdivided into individual channels. The flow developing into the channel is constituted of numerous repeating patterns. We exploit that fact regarding each channel consisting of numerous repeating geometrical pieces. Those are analyzed individually. We adopted the POD method to extract dominant spatial flow modes in each subdomain, which assembly constitutes an RB. We aim to understand the similarities between local, piecewise RBs emanating from different obstructed conditions and locations. To carry out this goal, we study the distance defined in the Grassmannian manifolds between the subspaces generated by these RBs. Various distances with clearly differentiated physical meanings are briefly introduced, computed, and compared. We analyze the feasibility of utilizing them to measure the distinctions between RBs and their application scope. We conclude that Grassmann Distances is a feasible tool to measure the dissimilarities between POD RBs. The results indicate that RBs constructed from blocked cases can reasonably approximate the free flow without obstacles. Then, the possibility of creating a representative RB applicable to reproduce all cases with acceptable accuracy is investigated. The typical RB can be obtained from any individual or assembled snapshots, and they can approximate snapshots obtained from other cases with an L2-norm error in the order around 10-2. Furthermore, by comparing distances between various subspaces, we discovered that the existence of obstructions resulted in noticeable distinctions, which mainly affect the performance of the generic RB.