Supervised by Nguyen Anh K Doan, we investigated a way to compress the representation of small-scale turbulent flows by 150-fold using autoencoders, a Deep Learning model. By doing so, we explored the level to which spontaneous peaks in macroscopic flow properties (kinetic energy and diffusivity), so-called “extreme events”, could be foreseen.
Specifically, the use of a variational autoencoder helped to produce a latent space representation of a turbulent flow, from which a modularity maximization clustering technique is imposed to pin-point precursors to extreme events (outliers in macroscopic flow properties). Study combining Aerodynamics and Deep Learning techniques.
View Final Presentation