Student: Hannah Blumhoefer, graduate student in Aerospace Engineering, Iowa State University
Faculty Advisor: Anupam Sharma
Scientific Machine Learning Augmentation of Computational Fluid Dynamics
The primary objective of this research project is to improve the efficiency of computational fluid dynamics (CFD) modeling by integrating deep learning and optimizing computational expenses through enhanced shock detection and shock capture methods. These enhancements will advance CFD’s diagnostic and design capabilities in research and commercial use. These objectives motivate the focus of this project, which involves the development of both a modified convolution-based and a multigrid-based shock detector technique utilizing a hierarchy of grids with varying levels of fidelity. These detectors will be implemented into the Air Force Research Laboratory’s high-fidelity flow solver FDL3DI. The final task in this project will be extending both detectors into a multilayer perceptron (MLP) algorithm to differentiate between shockwave and turbulence conditions in 3D space.
Preliminary research in pursuit of this project has been centered on developing a strided convolution-based shock detection system. This endeavor involved creating and evaluating this system using the high-fidelity flow solver FDL3DI, made available by the Air Force Research Laboratory. To assess the effectiveness of the shock detection system, a straightforward test case, the 3D Explosion test, was selected. In this scenario, a spherical diaphragm separates a high-pressure, high-density region from a lower-pressure, lower-density region. At time zero, the diaphragm ruptures, causing an expanding shock wave to propagate in all directions. The Density Gradient Magnitude indicates the region that should be identified as a shock sometime after the diaphragm’s rupture. The strided convolution shock detector exhibits a more precise detection of the relevant shock, delineating a narrower flagged region, whereas the compressibility shock detector identifies a broader area extending beyond the critical shock point. Since shock regions necessitate reduced-order shock-capturing techniques to mitigate oscillations, these regions lose their high-order characteristics compared to non-flagged areas. The extended shock region identified by the compressibility shock detector implies that regions where reduced-order shock capturing is unnecessary will still undergo calculation using these techniques.