変分オートエンコーダを用いた翼形状創出における潜在空間の定量的分析

抄録

Research on shape generation using machine learning has been widely conducted, and two-dimensional laminar flow airfoils are treated as a benchmark problem. When learning airfoil shapes using variational autoencoders (VAEs), it is known that the results obtained by ordinary VAE (N-VAE) and hyperspherical VAE (S-VAE) differ significantly. The difference is attributed to the fact that the standard normal distribution is used as the prior in N-VAE and the vMF distribution is used in S-VAE, but quantitative comparison of the latent space of both has not been conducted. In this study, we quantitatively compared how the data are embedded in the latent space of both VAE models. It is shown that data with different trends are embedded near each other in N-VAE, while data with similar trends are embedded near each other in the latent space of S-VAE. The difference can be explained by the difference in KL divergence and data characteristics. The NACA airfoil data is used in the present study, and the dataset is not normally distributed, which is usually the case with other data in mechanical design. S-VAE is suitable in such a case.