Vector quantization variational autoencoder for turbulent flow images

This study explores, for the first time, the application of vector-quantized variational autoencoders to reconstruct and analyze flow images from fluid dynamics simulations across varying resolutions. The method demonstrates effective reconstruction performance, achieving meaningful quantitative results, such as mean squared error, peak signal-to-noise ratio, and structural similarity index measure. These metrics indicate that the method effectively captures and preserves key flow patterns relative to reference images. However, experiments with latent space interpolation reveal that direct linear interpolation between coarse and fine latents produces lower-quality medium-resolution reconstructions than the reference data, highlighting the need for additional conditioning or constraints. Despite this limitation, the interpolation results outperform naive pixel-based methods, suggesting that the technique encodes meaningful and structured latent representations of flow images. The potential applications of the proposed method in fluid dynamics are significant. By conditioning latent representations with specific flow parameters, the technique can synthesize flow images for novel fluid dynamics scenarios, including intermediate states or parameterized datasets. This capability opens avenues for generating flow states that may not be achievable through direct simulations or experiments, such as scenarios involving dynamically changing boundary conditions–a situation typical in atmospheric, oceanographic, or engineered flows. Moreover, generating such synthetic data could considerably expand the accessible parameter space while reducing the computational cost of running extensive simulations or experiments. These features position the proposed method as a promising tool for compressing, reconstructing, and synthesizing flow data for fluid dynamics research and applications.