DID YOU KNOW? DTIC has over 3.5 million final reports on DoD funded research, development, test, and evaluation activities available to our registered users. Click HERE
to register or log in.
Double Cone Flow Field Reconstruction Between Mach 4 and 12 Using Machine Learning Techniques
[Technical Report, Master's Thesis]
AIR FORCE INSTITUTE OF TECHNOLOGY WRIGHT-PATTERSON AFB OH
Pagination or Media Count:
Analysis of hypersonic vehicle designs are costly due, in part, to the physical phenomena unique to the hypersonic flight regime. It is desirable to consider as many of these phenomena as possible early in vehicle design when computational resources are limited. Reduced order models can provide insight into these phenomena at a low cost by leveraging previous results. The utility of machine learning models for predicting the pressure field around a hypersonic weapon in flight is investigated. A parameterized double cone model is simulated at hypersonic speeds using a steady state RANS solver, Kestrel. The resulting pressure fields are used to train two neural network NN models, the U-Net and the Multiscale Network, as well as two metamodels, K-Nearest Neighbors and Regression Kriging. The NN models are designed to extract flow field relationships using distinct methodologies the U-Net utilizes auto-encoding while the Multiscale Network utilizes a sequential refinement scheme. All models predict the pressure values on a uniform Cartesian grid of much smaller resolution than the unstructured mesh required for CFD simulation. The accuracy, computational complexity, and versatility of the NN are compared against the meta models. Additionally, the ability for each method to accurately predict shock interactions or impingement with downstream vehicle geometry is examined. Such closed form ML models can provide advantages over traditional CFD solutions as they do not require any meshing of the computational domain and can quickly generate flow field predictions - on the order of seconds. The NN models were found have lacking yet robust performance on this dataset. Additionally, the NN models were shown to be effortlessly applicable to arbitrary geometries that cannot be described using the existing geometric parameterization.
[A, Approved For Public Release]