Embedded parameter information in conditional generative adversarial networks for compressor airfoil design
Embedded parameter information in conditional generative adversarial networks for compressor airfoil design
The use of conditional generative adversarial networks has become popular with the advancement of computer power, and in particular graphics processing units. Large hardware and software companies such as Google, IBM, and Facebook have been experimenting in this field for over a decade for facial recognition, image classification and pattern recognition, and deep fake facial and object design. One essential key to their success is access to large quantities of tagged and classified images, with millions of images typically being used. In contrast, relatively few similar advances have been seen in the engineering sector, mainly because engineering analyses that produce suitable images are often very expensive processes that absorb a considerable amount of effort to generate. In addition, feeding synthetically generated image data back into traditional engineering workflows is not entirely straightforward. In this paper we show how adversarial networks can be used if order 10
4 images are available and focused on the problem in hand. In particular we show how such images can then be augmented with histograms and glyphs to enhance the image content with pictorial representations of numerical data. This is shown to significantly assist the network training process on our data when used in contexts where numerical data categorizing individual images are available and numerical performance measures of products must be predicted. Crucially, it allows for follow-on analyses without the need to use artificially generated flow fields.
6753-6762
Keane, Andy
26d7fa33-5415-4910-89d8-fb3620413def
Voutchkov, Ivan
16640210-6d07-49cc-aebd-28bf89c7ac27
Keane, Andy
26d7fa33-5415-4910-89d8-fb3620413def
Voutchkov, Ivan
16640210-6d07-49cc-aebd-28bf89c7ac27
Keane, Andy and Voutchkov, Ivan
(2022)
Embedded parameter information in conditional generative adversarial networks for compressor airfoil design.
AIAA Journal, 60 (12), .
(doi:10.2514/1.J061544).
Abstract
The use of conditional generative adversarial networks has become popular with the advancement of computer power, and in particular graphics processing units. Large hardware and software companies such as Google, IBM, and Facebook have been experimenting in this field for over a decade for facial recognition, image classification and pattern recognition, and deep fake facial and object design. One essential key to their success is access to large quantities of tagged and classified images, with millions of images typically being used. In contrast, relatively few similar advances have been seen in the engineering sector, mainly because engineering analyses that produce suitable images are often very expensive processes that absorb a considerable amount of effort to generate. In addition, feeding synthetically generated image data back into traditional engineering workflows is not entirely straightforward. In this paper we show how adversarial networks can be used if order 10
4 images are available and focused on the problem in hand. In particular we show how such images can then be augmented with histograms and glyphs to enhance the image content with pictorial representations of numerical data. This is shown to significantly assist the network training process on our data when used in contexts where numerical data categorizing individual images are available and numerical performance measures of products must be predicted. Crucially, it allows for follow-on analyses without the need to use artificially generated flow fields.
This record has no associated files available for download.
More information
e-pub ahead of print date: 4 August 2022
Additional Information:
Funding Information:
The authors would like to thank Rolls-Royce plc for providing approval to publish this work. The research leading to these results has received funding from Innovate UK, under the Collaboration Across Business Boundaries (COLIBRI) project (Ref no. 113296).
Publisher Copyright:
© 2022 by Rolls-Royce plc. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission.
Identifiers
Local EPrints ID: 473373
URI: http://eprints.soton.ac.uk/id/eprint/473373
ISSN: 0001-1452
PURE UUID: 1c9d0de7-73b4-4662-9c4e-d58fa9cf899a
Catalogue record
Date deposited: 17 Jan 2023 17:35
Last modified: 17 Mar 2024 02:43
Export record
Altmetrics
Download statistics
Downloads from ePrints over the past year. Other digital versions may also be available to download e.g. from the publisher's website.
View more statistics
