Solving Inverse Problems in Hypersonic Boundary Layers Using Operator Networks
Abstract
At hypersonic speeds, the aero-thermo-dynamics of very thin boundary layers on in the flight vehicle have significant impact on performance. The boundary-layer flow is extremely sensitive to environmental disturbances which can lead to laminar-to-turbulence transition and, as a result, excessive heating and drag on the vehicle. Since the hydrodynamic rates in this regime are similar to the rates of dissociation and re-association of the air, non-equilibrium chemical effects must be taken into account. The extreme flow conditions and speeds, small spatial scales and complex multi-scale physics present significant challenges for laboratory and in-flight measurements. Data assimilation techniques can augment measurements data beyond the original sensor resolution. Viewed through the lens of data assimilation, measurements are not mere records of a specific quantity at a particular location and time; instead, they become an encoding of the preceding flow dynamics. Data assimilation then decodes the sensor information: this inverse problem involves a nonlinear optimization where the unknown flow parameters that optimally reproduce the measurements are sought. The optimization is computationally very expensive because it involves repeated simulations for every estimate of the unknown parameters. The proposed effort targets this challenge in high-speed flight using state-of-the-art machine learning techniques, specifically deep operator networks. These networks can accurately represent operators, or functionals. Once trained, they provide an efficient and accurate model of the dynamics. Operator networks will be adopted to accelerate data assimilation by (a) providing a fast evaluation of the forward model and (b) to replace the entire data assimilation procedure. Accuracy of the networks will be established for both forward and inverse problems, for high-speed chemical non-equilibrium boundary layers. The capacity of the operator networks to make predictions beyond the training range will be assessed, and techniques to improve their generalization will be examined.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 21, 2022
- Source ID
- FA95502110345XX0
Entities
People
- Tamer A. Zaki
Organizations
- Air Force Office of Scientific Research
- Johns Hopkins University
- United States Air Force