Stochastic Multi-Scale Reconstruction of Turbulent Rotating Flows with Generative Models (Online)

Abstract

Turbulence reconstruction poses significant challenges in a wide range of fields, including geophysics, astronomy, and even the natural and social sciences. The complexity of these challenges is largely due to the non-trivial geometrical and statistical properties observed over decades of time and spatial scales. Recent advances in machine learning, such as generative adversarial networks (GANs), have shown notable advantages over classical methods in addressing these challenges[1,2]. In addition, the success of generative diffusion models (DMs), particularly in computer vision, has opened up new avenues for tackling turbulence problems. These models use Markovian processes that progressively add and remove noise scale by scale, which naturally aligns with the multiscale nature of turbulence. In this presentation we discuss a conditional DM tailored for turbulence reconstruction tasks. The inherent stochasticity of DM provides a probabilistic set of predictions based on known measureme...