Aerodynamic sensing for hypersonics via scientific machine learning

Julie Pham, Bryan Morreale, Noel Clemens, Karen Willcox

Abstract:

This work develops a scientific machine learning (SciML) approach to infer aerodynamic
parameters from measurements of the structural strain field induced by aerodynamic pressure
loads. The SciML approach provides the computational underpinning for a novel sensing
strategy for hypersonic vehicles to estimate distributed surface pressure loads, providing valuable
information for guidance, navigation, and control. The feasibility of the idea is first explored
by formulating the inference problem as a deterministic optimization problem. Numerical
studies show the effects of sensor sparsity and measurement noise on the accuracy of inferred
aerodynamic parameters. Optimal decision trees are then proposed as an interpretable, rapid
inverse mapping from the strain measurements to the aerodynamic parameters. Two types
of decision trees, optimal classification trees and optimal regression trees, are assessed. The
performance of each optimal decision tree is evaluated for the estimation of the aerodynamic
parameters and reconstruction of the distributed surface pressure loads. Using synthetic strain
data with realistic levels of noise, the pressure loads can be predicted to within an average of 2%
relative error.