Capturing Extreme Events in Turbulence using an Extreme Variational Autoencoder

Turbulent flows are characterized by intense generation of turbulent kinetic energy through nonlinear physical processes which cascade from the large- to small-scale structures in a forward energy cascade, which is chaotic in nature, and statistically intermittent. Using a recently developed extreme variational autoencoder (XVAE), the turbulent flow fields are replicated to a high order of accuracy. In this extended abstract, we demonstrate XVAE as a powerful alternative to the classical Proper Orthogonal Decomposition (POD) technique for reconstructing large-eddy-simulation (LES) data for scalar temperatures from a buoyant turbulent field at a high Reynolds number of $10^{10}$.