Physics-Conforming Turbulent Flow Simulations Compression Approach

With the growing size of turbulent flow simulations, data compression approaches become an utmost importance to analyze, visualize, or restart the simulations. Recently, in-situ autoencoder-based compression approaches were proposed and shown to be effective in producing dimensionality-reduced representations of flow simulations. However, these approaches tend to focus solely on training the model based on sample quality losses while not taking advantage of the physical properties of turbulent flows. In this paper, we show that training autoencoders with additional physics-informed regularizations, e.g., incompressibility and preservation of enstrophy, improves a baseline model without such regularizations in three ways: i) upon inspection of the trained compression filters of the neural network, we identify changes in the convolutions due to the inclusion of the physics-informed terms ii) the compressions prove to be more physics-conforming to homogeneous isotropic turbulences of different Reynolds numbers given that these adhere to both the divergence free condition and preservation of enstrophy without trading off reconstruction quality, and iii) as a performance byproduct, training shows to converge 4 times faster than the baseline model.