Conceptual Design Space Exploration of Hypersonic Aircraft Using Emulator Embedded Neural Networks

This study proposes a novel acquisition function for constraint boundary identification, with applications to hypersonic air vehicles. Hypersonic vehicles endure extreme thermal loads caused by aerodynamic heating, resulting in a strong coupling between structural performance and aerothermodynamics. Therefore, it is crucial to consider aerothermal-structural interactions from the early stage of conceptual design development. However, modeling coupled system behaviors requires simultaneous consideration of both aerodynamic and structural design variables, including shape variables of the outer mold line, structural component layouts, and material selection for different areas of the vehicle. This increases the dimensionality of the design trade space. As the dimensionality increases, it becomes more difficult to model the constraints with enough accuracy to identify the feasible region. Several active learning schemes have been proposed to accelerate identification of the composite feasible region that satisfies all constraints. Some of these require integrating the surrogate model over the entire design space with each acquisition function evaluation, but this scales poorly to high dimensional spaces. In this work, we propose an alternative acquisition function based on the Expected Magnitude of Incorrectness of a constraint. Each constraint is modeled with either Gaussian Process Regression or an ensemble of neural networks, either of which can provide the uncertainty information needed by the proposed acquisition function. The neural networks used are called Emulator Embedded Neural Networks, and can combine multiple fidelities of data even if they do not form a strict hierarchy. Our approach performs well on analytical functions when compared to current acquisition functions. Additionally, we demonstrate our method on a lift constraint problem for a hypersonic vehicle wing.