Reproduce “Decoder Decomposition for the Analysis of the Latent Space of Nonlinear Autoencoders With Wind-Tunnel Experimental Data” on OpenResearch

About this paper

Yaxin Mo, Tullio Traverso, Luca Magri

A post-processing method termed decoder decomposition is introduced to enhance the interpretability of nonlinear autoencoder latent spaces by quantitatively linking latent variables to coherent flow structures. The method was validated on both simulated laminar and experimental turbulent flow data, demonstrating its ability to identify and filter latent variables representing specific physical phenomena like vortex shedding.

Key insights

• The choice of latent space dimension is critical for deriving physically interpretable and non-redundant representations, with inappropriate choices leading to artifacts or redundant information.
• For complex turbulent flows, the decoder's network capacity (size) can be as significant as the latent space dimension in achieving accurate reconstruction and interpretability.
• Standard autoencoders (AEs) can sometimes learn more parsimonious representations of flow dynamics compared to Mode-Decomposing Autoencoders (MD-AEs), which may not fully disentangle data modes.

Summary by alphaXiv.