10.2514/6.2026-5023">
 

Gradient Enhanced Emulator Embedded Neural Networks for Active Learning of Hypersonic Constraint Boundaries

Document Type

Conference Proceeding

Publication Date

7-7-2026

Abstract

This research proposes advances in the Emulator Embedded Neural Network (E2NN) method previously developed by the authors. This development enables E2NN to fit not only the training data, but known gradients at the training points. This is useful for aerospace applications, where the adjoint method enables simultaneous calculation of gradients with respect to all design parameters. The higher the dimensionality of the problem, the greater the benefit of including gradients. This makes the method particularly suited to hypersonic problems, which must often consider many variables simultaneously. Additionally, E2NN naturally handles multiple fidelities of data, further improving sample efficiency. In this paper, we demonstrate gradient-enhanced E2NN on a variety of active learning problems for constraint boundary identification. E2NN is found to compare favorably with Gradient-Enhanced Gaussian Process Regression (GPR). The E2NN ensembles use the Expected Magnitude of Incorrectness (EMI) acquisition function previously developed by the authors, while Gradient-Enhanced GPR is tested on $6$ different acquisition functions. Multiple analytical problems are tested, as well as a practical lift constraint problem for a hypersonic vehicle wing. In this final example, we find gradients of lift using NASA's FUN3D CFD code.

Comments

The full conference paper is available from AIAA via subscription or purchase, using the DOI link below.

Session: Computational Methods II

Source Publication

27th AIAA International Space Planes and Hypersonic Systems and Technologies Conference, 7-10 July 2026, Naples, Italy

This document is currently not available here.

Share

COinS