AeroINR: Meta-learning for Efficient Generation of Aerodynamic Geometries

Tom Bamford, David Toal, Andy J. Keane

Abstract: Effective optimisation of aerodynamic shapes requires high-quality parameterisation of candidate geometries. In recent years, the increasing availability and applicability of data - through increasing computational power, GPUs, cloud storage and AI - has motivated the development of data-driven approaches to the parameterisation problem, particularly those that can process the image-based data coming from scanned design parts. In this paper a novel approach to aerodynamic shape parameterisation is proposed, which leverages meta-learning in a generative deep learning framework. The solution put forward - AeroINR - aims to learn continuous neural representations as surrogates of the discrete field data used for shape representation in image-based applications. This approach transforms the learning problem to that of the surrogate model weight distribution of candidate geometries, rather than grid-based field values directly, which can reduce the number of variables describing each geometry by an order of magnitude or more. Benchmarking is carried out against three state-of-the-art deep-learning based aerofoil parameterisations, with AeroINR shown to outperform these models in two of the three metrics considered. Ablation study results show the robustness of this approach to generative framework and choice of discrete field representation.

External IDs:dblp:conf/pkdd/BamfordTK24