Gradient-guided graph convolutional multi-mesh frameworks for aircraft aerodynamics modelling

Massegur Sampietro, David and Da Ronch, Andrea (2024) Gradient-guided graph convolutional multi-mesh frameworks for aircraft aerodynamics modelling. In AIAA Scitech 2024 Forum. American Institute of Aeronautics and Astronautics.. (doi:10.2514/6.2024-0456).

Abstract

Modern deep learning technologies are increasingly common for reduced-order modelling in aerospace applications. Their outstanding inference capability in highly nonlinear systems make them particularly suited to aircraft aerodynamic analyses, wherein high-fidelity computational fluid dynamics remains prohibitive to date. A problem with data-driven methods is the strong dependency on the training data adequately representing the design landscape in order for the model to correctly capture the physics of the system. The computing cost required to generate this essential data often results impractical as well. Embedding physics knowledge into the data-driven framework becomes attractive to aid the model with learning the correct physics, an important branch of research known as physics-informed machine learning. However, physics terms directly applicable to the aircraft surface are not straightforward. We propose leveraging the solution gradients, as these result computationally effective to capture the surface flow features. This work presents a hybrid gradient-guided geometric-deep-learning based framework for surface aerodynamic field predictions. The work builds upon a previous contribution to the AIAA Aviation Forum 2023, which introduced a geometric-deep-learning multi-resolution framework for aerodynamic modelling on large and unstructured surface meshes, typical in real-world computational fluid-dynamic analyses. To further enhance prediction accuracy of the data-driven framework in under-sampled design spaces, here we investigate various hybrid gradient-guided data-driven schemes. The prediction performance for each scheme is demonstrated on the NASA Common Research Model wing/body aircraft configuration. Our study identified two distinct gradient-based methods that consistently outperformed the regular data-driven implementation. Differentiation of the model outputs added to the training loss was found with the best prediction accuracy but at expense of training computational cost. On the other hand, a masking approach to emphasise regions with large gradients was less of an increment but more computationally efficient to train.

This record has no associated files available for download.

Contributors

Author: David Massegur Sampietro

Author: Andrea Da Ronch

Download statistics