Recurrent geometric deep learning for aerodynamic prediction of the future fighter demonstrator in dynamic manoeuvres

Massegur Sampietro, David and Da Ronch, Andrea (2024) Recurrent geometric deep learning for aerodynamic prediction of the future fighter demonstrator in dynamic manoeuvres. In AIAA Aviation Forum and Ascend 2024. American Institute of Aeronautics and Astronautics.. (doi:10.2514/6.2024-4067).

Abstract

NATO/STO Research Task Group AVT-351 addresses investigation of reduced order modelling techniques for computationally efficient aerodynamic predictions of aircraft and ships. As part of the AVT-351 working group, this work presents a geometric deep learning based framework for reduced-order aerodynamics modelling of the Future Fighter Demonstrator aircraft, subject to pitching motion. Geometric deep learning is the umbrella of neural networks algorithms especially designed for irregular domains, typical in computational fluid dynamics of air and sea vehicles. For simulation of dynamic manoeuvres, the geometric deep learning framework is embedded in a time-marching scheme, involving a recurrent neural network to limit the long-term error accumulation. The model also features a hierarchical multi-mesh scheme to increase the computational efficiency and capture flow phenomena of different spatial scales, similar to the multi-grid algorithms for the solution of partial differential equations. To assess the performance of this framework, a quasi-steady model is also implemented, by means of a model architecture built in blocks to facilitate the conversion and transfer of parameters among the diverse models involved. It was found that the recurrent model was able to predict the unsteady flow fields more accurately than the quasi-steady alternative, due to accounting for the past flow condition at each time-step prediction. In addition, the computational gain to complete new forced-motion simulations was found over 99.9% with either reduced order model, concluding that most of the computational requirement is for the generation of the preliminary data.

This record has no associated files available for download.

Contributors

Author: David Massegur Sampietro

Author: Andrea Da Ronch

Download statistics