Design and Application of Surrogate Models for Hypersonic Inlet Physics

In this work we utilize computational models for a hypersonic vehicle inlet to build low-data surrogate models which predict air pressure and speed at the throat of an inlet. We also assess the viability of using these surrogates to generate sample inlets in lieu of these computational models. Gaussian Process Regression (GPR) using several different kernels is utilized to train surrogates. Deterministic, uniform random, and Latin Hypercube Sampling (LHS) sampling procedures are also utilized to generate the training sets for these models. By testing a variety of surrogates with different kernels and training sets, we found that surrogate modeling in general was effective for creating accurate representations of the computational models. The best performing surrogate model using a reasonably low amount of data was a GPR using a rational quadratic kernel trained on n = 60 sample points generated by Latin Hypercube sampling. This model had an error of less than .5% for estimating air pressure at the inlet throat, and an error of approximately 2% for estimating air speed at the inlet throat, making the surrogates less effective but much quicker than than the true computational model. Most sample inlets created using surrogate models had four out of five target angles estimated within 1 degree of the optimal design, but also had one angle within 10 degrees of the optimal design, suggesting that there are issues with applying the models into inlet design