In recent years the construction of large datasets coupled with the fast development of Machine Learning (ML) techniques, is leading to new paradigms for the investigation of fluid mechanic phenomena. Fluid mechanics deals with complex dynamical systems, which are multi scale in space and time, leading to difficulties in making predictions. Numerical simulations of complex fluid dynamic phenomena are time and computational expensive, and data generated are high dimensional, making it unfeasible real-time flow predictions. For this reason, reduced order models are employed. In this framework, Autoencoders (AEs) are emerging as a powerful tool for the construction of non-linear models able to predict the evolution of complex non-linear phenomena in a very low-order representation. As non-linear extension of Proper Orthogonal Decomposition (POD) they are able to construct non-linear embedded representations of high-dimensional data into a low-dimensional latent space. New challenges are arising from the development of AE based algorithms: the interpretability of the latent representations, and the extraction of physical information from phenomena not well understood which can lead to the adoption of AEs as a powerful tool for the discovery of physical laws. This minisymposium aims at disseminating new ideas and applications of AEs for fluid mechanics (and more), addressing the efficiency of the training algorithm, the generalizability of the models, interpretability of the latent representations, physical laws retrieval and many others.