Abstract

Design space exploration and optimization of hypersonic vehicles is costly due to the difficulty of producing high fidelity CFD simulations in the hypersonic domain. Reduced order modeling can allow for design optimization at a fraction of the computational cost. This paper investigates the amount of training data required to produce an accurate Reduced Order Model. A PCA based ROM is developed and compared to a Sequential Fusion DeepONet by comparing model prediction accuracy, training time, and prediction time, across dataset sizes from 25 samples to 1024 samples. These models are applied to 2D hypersonic CFD simulations of the Orion reentry vehicle at Mach numbers ranging from 10 to 30 and AoA from 0 to 35°. For this application, it is found that the DeepONet predicts flowfields with errors between 1.9% at 1024 training samples and 20.2% with 25 samples. The PCA ROM predicts flowfields with errors between 2.7% and 22% across the same range of samples. A clear tradeoff between accuracy and dataset size is demonstrated for both PCA and DeepONet models. The PCA ROM trains in 95 seconds and predicts in 0.7 seconds when using 576 samples, and nears real time evaluation using 25 samples. The DeepONet based ROM trains in 35.3 hours and evaluates in 12.6 seconds when using 576 samples. The large difference in training and evaluation times is due to the difference in the number of matrix operations for each model. DeepONets require many matrix calculations in sequence leading to much longer training and prediction times. The DeepONet is mesh agnostic and does not require a common grid whereas the PCA ROM has an added mesh interpolation step.

Degree

MS

College and Department

Ira A. Fulton College of Engineering; Mechanical Engineering

Rights

https://lib.byu.edu/about/copyright/

Date Submitted

2026-08-03

Document Type

Thesis

Keywords

surrogate modeling, hypersonics, deep learning

Language

english

Download

Included in

Engineering Commons

Share

COinS