Search papers, labs, and topics across Lattice.
This paper introduces IG-GAN, a novel generative adversarial network designed to generate aerodynamic data by leveraging intrinsic geometry rather than traditional flat Euclidean space. By employing a generator that constructs aerodynamic data as a piecewise smooth manifold using B茅zier surfaces, IG-GAN significantly improves data representation and generation. Experimental results demonstrate that IG-GAN outperforms existing models, achieving a 97.41% reduction in predicted MSE for velocity on the Burgers' equation dataset and an 82.95% reduction in overall MSE for aerodynamic coefficients on the ONERA M6 aircraft dataset compared to the SSL-Transformer baseline.
IG-GAN slashes aerodynamic data generation errors by up to 97% by harnessing the power of intrinsic geometry.
Existing generative models learn data distributions in flat Euclidean space. However, most data in our real world are manifolds embedded in high dimensional Euclidean space. Therefore, we propose an intrinsic-geometry-based generative adversarial network (IG-GAN) for data generation in the field of aerodynamics. The generator of the IG-GAN represents aerodynamic data as a piecewise smooth manifold constructed by B茅zier surfaces, and the generator tries to learn the coefficients of each B茅zier surface to further combine multiple B茅zier surfaces into a smooth manifold automatically. The discriminator in the IG-GAN is a radial-basis-function based discriminator (RBF-D). Experimental results show that IG-GAN achieves lower predicted Mean Squared Errors (MSEs) than those of three baselines. Specifically, on the Burgers' equation dataset, IG-GAN reduces the predicted MSE of velocity u by 97.41% compared with state of the art SSL-Transformer. Additionally, on the ONERA M6 aircraft dataset, IG-GAN reduces the overall MSE of nine aerodynamic coefficients by 82.95% compared with SSL-Transformer.
