Degree Type

Thesis

Date of Award

2018

Degree Name

Master of Science

Department

Aerospace Engineering

Major

Aerospace Engineering

First Advisor

Thomas Ward

Abstract

Generating and parsing through large amounts of wind tunnel,

ight test, or computational

uid dynamics (CFD) data can prove to be expensive. This makes, for example, the optimization

of aerothermal hypersonic components, which may contain a large number of independent variables,

challenging. Having a surrogate model to quickly and accurately approximate the data can help

with the optimal design process. A lower order model can be used instead of or in conjunction

with a higher order model to model a system with less computational eort. Typically, additional

assumptions are made to make a lower order model. These have the benet of being faster to execute

or simpler to solve but come at the cost of reduced accuracy. Here, an eort is made to construct

lower order/surrogate models, that are built and operate with exactly the same assumptions as the

higher order model, in the form of machine learning and deep learning-based surrogate models. In

tandem with this eort, the goal is to have a function evaluation time signicantly smaller than the

higher order method. This can increase the number of variables the designer can consider during the

optimal design process or reduce the time required to design a component. Potential methodologies

using machine learning (ML) and deep learning as surrogate models are used to evaluate the

ow

eld and subsequently the performance of a 2-D hypersonic compression ramp. Gradient Boosting,

Neural Networks, and Random Forest regressors are used as the surrogate models. With the

ow

eld predicted, performance characteristics can be calculated such as: lift, drag, moment, pressure

distribution, shear, heat transfer, etc. Thus, multiple phenomena can be considered during the

design process. The nine

ow parameters being predicted are: pressure, total pressure, density,

dynamic viscosity, Mach number, temperature, total temperature, and velocities in the principal

axes. CFD simulations using a commercial package are used to generate the training data for the

regressors. The predictors are measured on how accurately the simulation data can be predicted.

Copyright Owner

Nathan Hemming

Language

en

File Format

application/pdf

File Size

353 pages

Share

COinS