AI-based sub-grid scale closure with Super Resolution Network for Large Eddy Simulations in the human larynx
Please login to view abstract download link
Turbulence is a key phenomenon in fluid mechanics and is therefore the subject of intense research. Direct Numerical Simulation (DNS) in Computational Fluid Dynamics (CFD) provides the most accurate insight into flow behavior, but demands substantial computational resources. In contrast, Large Eddy Simulation (LES) provides a reliable characterization of flow behavior while avoiding high computational costs; however, in this approach, some information, such as sub-grid-scale turbulence quantities, is lost. This work aims to bridge the gap between LES and DNS by reconstructing lost information using a Super-Resolution Network (SRN) technique [1, 2] and by developing an AI-based alternative to the turbulence-closure problem. The SRN makes use of the open-source library AI4HPC [2] for a particular use case of laryngeal flow [3] from OVGU. Available DNS results are used as a reference to train the SRN in a two-step strategy (see Fig.~\ref{fig:srn}). First, a Gaussian filter of a specific kernel size (corresponding to the LES grid downsampling) is applied to the DNS velocity field to obtain the training dataset using Convolutional Defiltering Model (CDM) [1, 2]. Subsequently, the filtered DNS is used to train a 3D U-Net architecture to reconstruct the sub-grid scales, while LES results are used for further validation of the SRN model. Further details on the flow fine structure and reconstructed DNS data using the SRN will be discussed in the full paper. References [1] Fukami K., Fukagata K., and Taira K. arXiv preprint arXiv:2301.10937, (2023). [2] Sarma R., Inanc E., et al. Frontiers in High Performance Computing, 2, 1444337, (2024). [3] Abdelsamie A., Voß S., Berg P., et al. Computers & Fluids, 255, 105819, (2023).
