Predicting aerodynamics with hierarchical geometric deep learning: the GCN-MM-AE framework

Abstract

Persistent problem in industry:A major problem in aerospace and automotive industries is that the computational cost of high-fidelity computational fluid dynamics in full-scale applications is prohibitive. The disproportion of computing capacity against demand will persist for the foreseeable future. Consequently, a new paradigm is sought for more efficient and accurate simulations, aligned to the frenetic engineering design processes, leading to a breakthrough in engineering progress and more profitable businesses.State-of-the-art:Significant advances in artificial-intelligence technologies, namely neural networks, are leading to major progress in distinct disciplines. Neural-network techniques are attractive for reduced-order modelling of complex and nonlinear systems. However, the deployment of these modern technologies in the aerospace community to address relevant tasks has been more elusive. This represents the motivation for our research work: develop artificial-intelligence based frameworks to overcome the current limitations from classical tools.Research Focus:This PhD research proposes novel deep-learning based frameworks for efficient aerodynamics and multiphysics reduced order models. Critical challenges around aerodynamic predictions of aircraft components are the high dimensionality and the unstructured nature of the discretised computational domains, as well as the limited data availability due to the computational burden. Applicability of the framework is demonstrated in relevant applications, including aerodynamics of transonic flight envelopes, aircraft manoeuvring, aeroelastic flutter search, ice accretion and shape design optimisation. The promising findings presented here lay the groundwork for an efficient, general-purpose multiphysics and optimisation framework.

More information

Identifiers

ORCID for David Massegur Sampietro: orcid.org/0000-0001-6586-5097

ORCID for Andrea Da Ronch: orcid.org/0000-0001-7428-6935

Catalogue record

Export record