Abstract
<title>Abstract</title> <p>Reduced-order modeling of chaotic fluid flows using autoencoders is a promising approach. However, conventional variational autoencoders often suffer from posterior collapse and a trade-off between reconstruction accuracy and latent regularization. To overcome these limitations, we employ an Information Maximizing Variational Autoencoder (InfoVAE), whose more flexible objective function promotes informative latent representations while preserving reconstruction quality. In this paper, we combine the Information Maximizing Variational Autoencoder (InfoVAE) with an Easy-Attention-based Transformer to develop an efficient reduced-order modeling framework for fluid flow prediction. The proposed approach was evaluated on numerical datasets of two-dimensional viscous flows under both periodic and chaotic conditions. The learned disentangled latent representations provide an interpretable reduced-order model, with features that are consistent with the dominant modes identified by Proper Orthogonal Decomposition (POD). By applying InfoVAE to extract disentangled representations, an interpretable flow model was achieved. The features emerging from this model are similar to those observed in proper orthogonal decomposition.</p>