"Adaptive Selection of Decomposed Function Information Sources for Rapi" by Atticus Beachy, Harok Bae et al. 10.2514/1.J064590">
 

Adaptive Selection of Decomposed Function Information Sources for Rapid Neural Networks

Document Type

Article

Publication Date

6-2025

Abstract

A new approach is proposed of adaptively selecting emulators for emulator embedded neural networks. Emulators typically take the form of physics-based low-fidelity models. If querying a low-fidelity model incurs significant computational costs, an emulator can still be constructed by training a surrogate model, such as a Gaussian process model or neural network, on data gathered from the low-fidelity model. These emulators are embedded into the neural network architecture and play an important role in boosting neural network accuracy with limited training data. However, in some practical situations finding a suitable low-fidelity emulator model is challenging. Use of an improper emulator can fail to improve learning performance, and in some cases no viable emulator is available. To address this technical gap, this study proposes using the decomposed functions of the actual physics-based model as emulators. Global sensitivity analysis is performed to compute the Sobol indices of the decomposed functions, which reveal their contribution ranks to the system response of interest. Based on the contribution ranks, the decomposed functions are embedded as emulators in an adaptive and iterative process. The proposed method is demonstrated with fundamental analytical examples. Additionally, a representative hypersonic vehicle design problem is included as a practical engineering example.

Comments

This article is published by AIAA as cited, and is accessible by subscription or purchase using the DOI link on this page.

Source Publication

AIAA Journal (ISSN 0001-1452 | eISSN 1533-385X)

This document is currently not available here.

Share

COinS