Automating Parameterization for Aerodynamic Shape Optimization
Aerodynamic shape optimization (ASO) is a key technique in aerodynamic design, aimed at enhancing an object's physical performance while adhering to specific constraints. ?Traditional ASO parameterization methods often require substantial manual tuning and are limited to surface deformations.
In a paper that was presented at the AIAA Aviation Forum'24 and that received an NDO best student paper award, we introduce the Deep Geometric Mapping (DeepGeo) model, a fully automatic neural-network-based parameterization method for complex geometries.
?DeepGeo exploits the universal approximation capability of deep networks to provide large shape deformation freedom with global shape smoothness, while achieving effective optimization in high-dimensional design spaces. Additionally, DeepGeo integrates volumetric mesh deformation, simplifying the ASO pipeline. By eliminating the need for extensive datasets and hyperparameter tuning, DeepGeo significantly reduces implementation complexity and cost.
In the example shown above, the objective is to minimize the drag coefficient for a given lift coefficient, which among other things means eliminating the shockwaves shown in yellow. The corresponding video can be downloaded from here. For more case studies, please refer to the paper. This the result of a collaboration between us and the groups of Micha?l Bauerheim from SupAero and Rhea Liem from HKUST.
CEO at Holbrook Aerospace
8 个月Mesh deformation is an extremely intensive form of aero-optimization. Can you imagine trying to optimize an airfoil by moving one or a couple of 200 different points at a time? Working with a mesh is like that only worse, because it's in 3D. The HAVF airfoil format is capable of describing quintuple variable profile/airfoil wings, including all chord and angle of attack values with ~50 less data points than even one typical 200 coordinate airfoil DAT file, let alone a 10,000 3D point mesh. Our complete library of existing airfoils is not only ready to be ingested by any learning algorithm, it's already being used that way by us to explore the actual design state space better than any existing mesh or point deformation algorithm can. Working with vectors is a winning decision. It can do heavier lifting with less compute effort, it can describe curves more accurately than a mesh ever can. Also, if you want to degrade the model back into a mesh for CFD, or a CAD it's totally possible. Automate-able even.