Reproduce “Towards aerodynamic surrogate modeling based on $β$-variational autoencoders” on OpenResearch

About this paper

Víctor Francés-Belda, Alberto Solera-Rico, Javier Nieto-Centenero, Esther Andrés, Carlos Sanmiguel Vila, Rodrigo Castellanos

The research by V. Francés-Belda et al. introduces a robust surrogate modeling framework for predicting 3D transonic wing pressure distributions, integrating Principal Component Analysis with "beta"-Variational Autoencoders and Gaussian Process Regression, enhanced by a novel decoder fine-tuning process. This approach achieves high prediction accuracy while significantly reducing computational costs and memory footprint.

Key insights

• PCA pre-processing significantly reduces the "beta"-VAE's parameter count and memory footprint while stabilizing performance and enhancing latent space structure.
• A latent space dimension of two is sufficient to capture the main physical variations related to Mach number and angle of attack, confirming physical interpretability.
• The novel decoder fine-tuning process substantially improves prediction accuracy and, crucially, makes the model largely independent of the "beta" regularization parameter, allowing for disentanglement without sacrificing fidelity.

Summary by alphaXiv.