Clustering of State Vectors from CFD Simulations for Construction of Local Reduced-Order Models

A novel method for clustering computational fluid dynamics (CFD) solution vectors with the goal of building efficient local reduced-order models (ROMs) is proposed. In this method, the clustering process is accelerated through the introduction of triangle-inequality bounds to avoid unnecessary computation during K-Means, and the accuracy of the resulting ROM simulation is improved through the introduction of a semi-supervised clustering algorithm. The performance benefits of the proposed method are demonstrated using two data sets generated from CFD simulations. For the larger of the two datasets the proposed method was able to cluster 183 gigabytes of data (1252 state vectors) into five clusters in under two minutes using 512 CPUs.