Investigation of lossy compression techniques to create training datasets for Reduced Order Models (ROMs) of rocket engines

High-fidelity combustion LES solvers can generate extremely large datasets which cannot be fully stored on disk. As a result, compromises are often made where the temporal frequency of disk output or the saved spatial grid resolution is greatly reduced from the native size. This is often undesirable when it comes to creating data-driven models from these simulations which attempt to learn efficient approximations while still capturing relevant dynamics. To address this, advanced lossy compression techniques are applied to the raw data before being saved to disk to greatly reduce the required storage size. To look at the effectiveness of this approach, two metrics are explored. First, the correlation of lossy compression error on reduced order model (ROM) prediction error is made. Second, the analysis of ROM improvement from increased training dataset size is conducted. The targeted application area of this work is the Rotating Detonation Rocket Engine (RDRE).