Autoencoders for Aerodynamic Predictions

Autoencoders (AEs) are unsupervised Machine Learning (ML) algorithms which perform non-linear data compression. They are widely employed for dimensionality reduction, feature extraction, and image processing tasks. The encoder is capable of achieve extreme data compression, mapping high dimensional data in just a few latent variables; while the decoder performs the inverse transformation restoring the original dimensionality. Recently, the literature has shown a growing interest in AEs also in the field of Fluid Mechanics with different algorithms developed for the non-linear modal decomposition and the prediction of new unseen flow fields. The talk will focus on recent research activities conducted through a collaboration between the University of Naples Federico II and Stanford University, exploring AEs for accurate flow predictions around configurations of aerodynamic interest. In addition to ensuring accurate flow predictions, is it possible to identify and quantify different sources of uncertainty in these predictions? This is crucial for assessing the limitations of AEs as generative models. Another important open question in this field concerns the interpretability of the embedded representations generated by AEs. Specifically, can AEs be employed to extract meaningful insights into the underlying physics of a phenomenon? These central questions will be explored during the presentation, showing the recent advancements made by our research group. The talk will conclude with a brief discussion on how to effectively manage multi-fidelity data for training AEs.