Investigations into the application of generative deep learning to aerodynamic shape parameterisation

Abstract

The design of aerodynamic shapes consists of two distinct processes: the selection of a design candidate and corresponding representation of that shape, and the subsequent performance analysis of that candidate against a set of pre-defined criteria. Both aspects of this procedure can be cumbersome, with increases in efficiency and speed an enduring goal for designers. Recent advances in computational power, alongside an abundance of data, have led to researchers turning to data-driven methods in artificial intelligence and deep learning in their search for incremental gains within the design optimisation pipeline. Convolutional deep learning architectures in particular have shown very promising results in image processing across a range of applications, which together with the increasing availability of imaging and scan data of designed parts may offer image-based alternatives to existing tools within the design pipeline. Possibilities include shape parameterisation and dimensionality reduction utilising generative frameworks, low-fidelity performance analysis in the form of deep-learning-based surrogate models, or a combination of the two as an inverse design solution. This thesis explores the utility of generative deep learning models - both adversarial and auto-encoding in nature - as a parameterisation tool for the design and optimisation of aerodynamic shapes. In addition to model selection, a key component of image-based learning is the selection of an appropriate representation of the given geometry. As such, the viability of a selection of spatially-informed shape representation approaches are also investigated in this work. The thesis begins with a review and exploration of the two-dimensional design case - in particular in the context of aerofoil design - before proceeding to the three-dimensional problem where the potential for performance improvements is greater. Relative to alternative deep learning approaches the presented image-based methods show strong performance, and in addition are readily scalable to higher resolutions, more challenging geometries and the incorporation of flow field data.

More information

Identifiers

ORCID for Joshua Thomas Bamford: orcid.org/0000-0001-7416-5703

ORCID for Andy Keane: orcid.org/0000-0001-7993-1569

ORCID for David Toal: orcid.org/0000-0002-2203-0302

Catalogue record

Export record